{"id": "219a22b230259ea97336b8090150dacc77948d9a35534711418df43c754d22f9", "sources": ["arxiv"], "title": "Can I Buy Your KV Cache?", "abstract": "Right now, across the world, AI agents are repeating the same absurd act: to read one document, they each recompute it from scratch. Every agent re-runs prefill, the most compute-intensive step a large model takes, over identical text, only to rebuild a key-value (KV) cache identical to the one the agent before it just built. The same answer, computed a million times. We make a proposal that is almost offensively simple: compute it once. Let a publisher precompute a document's KV cache, and let every other agent buy the right to load it and skip prefill. It works, and it is token-exact: loading a precomputed KV and continuing matches prefilling from scratch (24/24 greedy tokens, and at the logits level), with no accuracy cost. On Qwen3-4B, reuse is 9-50x cheaper in compute than prefill, and the gap widens with length (prefill's attention scales with L^2), so a single reuse already pays it back. Then the part that matters: where the KV lives. Shipping it fails, because KV is nearly incompressible, so per-load egress costs more than the prefill it saves. Hosting it provider-side, exactly as production prompt-caching works, removes egress entirely. The size of the prize is set by our measured compute saving: serving one hot 3774-token document to 80M agents costs ~$1.5M to re-prefill but only ~$0.03M of reuse compute (49.7x less). The 0.1x cache-read tariff APIs charge passes a 10x discount to users while sitting inside this measured envelope, so the 10x is a floor that the measured ~50x compute saving clears, and the gap to the physical ~50x is provider margin: millions of dollars per popular document. We frame the resulting agent-native prefill CDN and leave lossless KV compression and a cross-party payment layer as the open problems.", "authors": ["Luoyuan Zhang"], "categories": ["cs.AI", "cs.CE", "cs.MA"], "fields_of_study": [], "published_date": "2026-06-11", "url": "https://arxiv.org/abs/2606.13361", "pdf_url": "https://arxiv.org/pdf/2606.13361v1", "arxiv_id": "2606.13361", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "363441bf3b6e700abd0cf2940be987393677c46caa1cecd4c2121e14de11dfd9", "sources": ["arxiv"], "title": "MiniPIC: Flexible Position-Independent Caching in <100LOC", "abstract": "Retrieval-augmented and agentic workloads repeatedly prefill recurring predictable structured inputs (which we call \"spans\") such as documents and code files. Yet, prefix caching in engines such as vLLM cannot reuse their KV entries unless they share identical prefixes with another request, while Position-Independent Caching (PIC) implementations within production-grade inference servers typically either require substantial server code changes or keep KV state outside the server, incurring host-to-device transfer overhead. We present Minimalistic PIC (MiniPIC): a minimal, flexible and fast vLLM design built from two ingredients: positional-encoding-free KV cache and user-controlled cache-reuse primitives. MiniPIC stores unrotated K vectors in the KV cache, applies RoPE to K tiles inside attention using per-request logical positions, and exposes three user-facing and token-level primitives: block-aligned padding, span separator (SSep), and prompt depend (PDep), that modify hashing behavior and effective block-level causal attention structure. With fewer than 100 lines of core-engine changes plus a custom attention backend, these primitives are sufficient to realize multiple PIC methods, including Block-Attention, EPIC, and Prompt Cache, within the same running vLLM instance, while natively integrating with KV cache CPU offload implementations. On 2WikiMultihopQA, MiniPIC with interleaved scheduling improves prefill throughput by 49% over baseline vLLM, reduces cached-span time-to-first-token by up to two orders of magnitude, preserves the linear prefill scaling of uncached spans, and incurs only 5.7% worst-case overhead.", "authors": ["Nathan Ordonez", "Thomas Parnell"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": [], "published_date": "2026-06-11", "url": "https://arxiv.org/abs/2606.13126", "pdf_url": "https://arxiv.org/pdf/2606.13126v1", "arxiv_id": "2606.13126", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4fdecfa82b3d9e9bb030d3cf4a75de1007d4193de7d873f3dd881cdfb348a3eb", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Rate Mixture of Experts for Accelerating Liquid Neural Network Training", "abstract": "Multivariate time-series data often exhibit complex temporal dependencies, irregular sampling, and heterogeneous dynamics across multiple time scales, making accurate sequence modeling particularly challenging. Traditional recurrent neural networks (RNNs), such as Long Short-Term Memory (LSTM) networks, operate in discrete time and may struggle to effectively capture continuous and irregular temporal behaviors. Liquid Neural Networks (LNNs) address some of these limitations through continuous-time dynamics, but standard LNN architectures typically rely on a single dynamical system, limiting their ability to model heterogeneous temporal patterns. To address these challenges, we propose a Multi-Rate Mixture-of-Experts (MR-MoE) framework built on top of Liquid Neural Networks. In the proposed architecture, multiple LNN-based experts operate at distinct time scales, enabling the model to explicitly separate fast-changing dynamics from slow-evolving temporal trends. A gating network further enables adaptive expert specialization based on input conditions. In addition, we incorporate both feature-level and temporal attention mechanisms to improve robustness, interpretability, and long-range dependency modeling. Feature-level attention suppresses noisy or irrelevant variables, while temporal attention selectively focuses on informative historical states. We evaluate the proposed framework on a complex multivariate time-series prediction task and compare it against strong baselines, including LSTM, monolithic LNN, and standard MoE models. Experimental results demonstrate that the proposed MR-MoE framework consistently achieves improved AUROC and AUPRC performance while maintaining favorable computational efficiency. These results highlight the effectiveness of combining continuous-time dynamics, multi-scale expert decomposition, and adaptive attention mechanisms for time-series modeling.", "authors": ["Shilong Zong", "Almuatazbellah Boker", "Hoda Eldardiry"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-10", "url": "https://arxiv.org/abs/2606.12240", "pdf_url": "https://arxiv.org/pdf/2606.12240v1", "arxiv_id": "2606.12240", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e6285ce1e5e00dc6f200445d06208fa2cce60e6d585f8a2a9bfc8b9cdcd4087a", "sources": ["arxiv", "semantic_scholar"], "title": "Redesign Mixture-of-Experts Routers with Manifold Power Iteration", "abstract": "Router is the cornerstone component to the Mixture-of-Experts models. Serving as expert proxies, the rows of the router matrix compute their similarity to the MoE inputs to determine which subset of experts is activated. Ideally, each router row is designed to encode the expert matrix into this representative vector, such that its dot-product with token can better reflect token-expert affinity. However, there exists no design principles to enforce this condensation. In this paper, we propose to align each router row with the principal singular direction of the associated expert, as this direction provides the most expressive mathematical description of a matrix. Based on this principle, we propose a router redesign with Manifold Power Iteration (MPI). Specifically, it introduces a \"Power-then-Retract\" paradigm, where a power iteration step is performed on the router weights, followed by a retraction to impose a norm constraint to ensure both efficiency and stability. Theoretically, we show that MPI drives router rows to converge toward the principal singular directions of associated experts. Empirically, we pretrain MoE model across scales from 1B to 11B parameters to confirm that this alignment facilitates more effective MoE models.", "authors": ["Songhao Wu", "Ang Lv", "Ruobing Xie", "Yankai Lin"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-10", "url": "https://arxiv.org/abs/2606.12397", "pdf_url": "https://arxiv.org/pdf/2606.12397v1", "arxiv_id": "2606.12397", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8f96f13b132484676858fd520f097975b3e493dbfb7ba7fbcb3656499f01dddc", "sources": ["arxiv", "semantic_scholar"], "title": "From Observation to Intervention: A Causal Audit of Expert Importance in Mixture-of-Experts Models", "abstract": "Interpretability methods routinely use population-level summary statistics over observed model behaviour to license claims about the effects of targeted interventions on specific computations; in Pearl's terms, they treat rung-1 associational evidence as if it supported rung-2 interventional conclusions, a move whose validity is rarely tested. We examine one concrete instance: the use of routing statistics in Mixture-of-Experts (MoE) pruning, where utilization rates, activation norms, and routing weight distributions are treated as predictors of which experts can be removed without functional cost. A token-level interventional audit across three high-redundancy MoE architectures (OLMoE-1B-7B-0924, Qwen1.5-MoE-A2.7B, DeepSeek-V2-Lite) finds no observational metric predicts causal expert importance after multiple-comparison correction in any model, with effect sizes below Cohen's $d = 0.17$ across all 60 metric-layer combinations. A per-token routing weight control rules out insufficient power, recovering a single Bonferroni-significant signal at OLMoE's final MoE layer ($d = +0.231$, $p = 0.0013$). Existing pruning methods succeed in this regime not by identifying dispensable experts but because early-layer redundancy renders most selection criteria interchangeable. Our results provide an explicit counterexample to the common inferential step from population-level observational summaries to token-level interventional claims about expert importance, and illustrate how interventional audits can calibrate the evidential standards for interpretability claims.", "authors": ["Leonard Engmann", "Christian Medeiros Adriano", "Holger Giese"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-09", "url": "https://arxiv.org/abs/2606.10703", "pdf_url": "https://arxiv.org/pdf/2606.10703v1", "arxiv_id": "2606.10703", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "aeea47930489c3c29a15b5d032ebbd01137ca01018d6230a4d3ceabe6832ac05", "sources": ["arxiv", "semantic_scholar"], "title": "Routing-Aware Expert Calibration for Machine Unlearning in Mixture-of-Experts Language Models", "abstract": "Machine unlearning is increasingly important for large language models, yet unlearning in Mixture-of-Experts (MoE) architectures remains underexplored. Unlike dense models, MoE architectures employ a router at each layer to assign each token to a sparse subset of experts. In this work, we observe that forget data often activates a small subset of experts disproportionately, while these experts may receive much weaker activation from retain data. This forget--retain routing mismatch can leave forget-critical experts under-regularized during unlearning. To address this, we propose \\textbf{TRACE}, Targeted Routing-Aware Calibration of Experts, for MoE unlearning. TRACE first detects forget-critical experts from offline activation statistics, and then calibrates retain regularization by reweighting token-level retain losses so that each selected expert's retain-side activation frequency better matches its forget-side counterpart. Experiments on WMDP and MUSE-BOOKS across multiple MoE LLMs show that TRACE consistently improves the forget-utility trade-off, yielding a 9\\% relative utility improvement over the strongest baseline under comparable forgetting quality and the best performance on three out of four MUSE-BOOKS metrics.", "authors": ["Jingyi Xie", "Yijun Lin", "Yinjiang Xiong", "Zhikun Zhang", "Sai Li"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-09", "url": "https://arxiv.org/abs/2606.10338", "pdf_url": "https://arxiv.org/pdf/2606.10338v1", "arxiv_id": "2606.10338", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4f7b301109d74f7f9d2385065d1af1c3a91b9ebb8cdccba0e0ccbdb224c63788", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Multilingual LLM-based ASR with Mixture of Experts and Dynamic Downsampling", "abstract": "The rapid progress of large language models (LLMs) has opened up a new frontier for automatic speech recognition (ASR), making their effective integration a critical and challenging research direction. To this end, this work proposes a projector-based LLM-ASR framework targeting the key challenges of multilingual generalization and modality alignment. Our approach incorporates a Mixture of Experts (MoE) architecture to improve cross-lingual adaptability, and a Continuous Integrate-and-Fire (CIF) mechanism for dynamic downsampling and modality alignment. Experimental results show that the combination of these components yields substantial performance improvements, surpassing strong baseline models. The proposed method represents a step toward building more accurate, robust, and generalizable LLM-based ASR systems.", "authors": ["Guodong Lin", "Ziqi Chen", "Yuxiang Fu", "Ke Li", "Wei-Qiang Zhang"], "categories": ["cs.SD", "cs.CL", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-06-09", "url": "https://arxiv.org/abs/2606.10439", "pdf_url": "https://arxiv.org/pdf/2606.10439v1", "arxiv_id": "2606.10439", "doi": "10.1109/ICASSP55912.2026.11464266", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.55}
{"id": "19f1030636588c82241484d9b6f7f0589343f2f874642fb66319b91465fda801", "sources": ["arxiv", "semantic_scholar"], "title": "Which LoRA? An Empirical Study on the Effectiveness of LoRA Techniques During Multilingual Instruction Tuning", "abstract": "We investigate whether commonly available LoRA variants have an advantage over basic LoRA in multilingual instruction tuning. Experiments involving LoRA and four other variants on two datasets across diverse target languages show that there is no significant advantage in using more complex LoRA variants instead of basic LoRA, with respect to balancing cross-lingual transfer and knowledge retention. An analysis of hidden embeddings reveal that layer-wise language representation remains largely similar across LLMs fine-tuned with different LoRA techniques, suggesting that architectural novelty of LoRA techniques may not translate into better cross-lingual adaptation.", "authors": ["Thamali Wijewardhana", "Napoleon H. Reyes", "Surangika Ranathunga"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-09", "url": "https://arxiv.org/abs/2606.10428", "pdf_url": "https://arxiv.org/pdf/2606.10428v1", "arxiv_id": "2606.10428", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ad2764c1694461754b9d3f9bf0d1bd04a98208f3430237100c6a3193876a01d6", "sources": ["arxiv", "semantic_scholar"], "title": "ReasonAlloc: Hierarchical Decoding-Time KV Cache Budget Allocation for Reasoning Models", "abstract": "Long chain-of-thought (CoT) trajectories in large language model (LLM) reasoning cause severe inference bottlenecks due to rapid key-value (KV) cache growth. Current decoding-time compression methods mitigate this issue via token eviction, but typically assume a uniform budget distribution across all layers and heads. In contrast, existing non-uniform budget allocation methods are predominantly designed for the static prompt prefill phase, and they do not capture the stepwise context demands of autoregressive reasoning. To bridge this gap, we propose ReasonAlloc, a training-free framework that recasts decoding-time KV compression as a hierarchical budget allocation problem. ReasonAlloc operates at two complementary levels: an offline layer-wise preallocation strategy captures an architecture-driven demand pattern which we call ``\\textit{Reasoning Wave}'', while an online head-wise strategy reallocates resources during decoding to information-rich heads based on real-time utility. Evaluations on mathematical reasoning benchmarks (MATH-500, AIME~2024) using DeepSeek-R1-Distill-Llama-8B, DeepSeek-R1-Distill-Qwen-14B, and AceReason-14B show that ReasonAlloc outperforms uniform-budget R-KV, SnapKV, and Pyramid-RKV (a baseline enforcing a static, monotonically decreasing layer budget), with the largest gains at small budgets (128-512 tokens). ReasonAlloc is plug-and-play with existing token-eviction policies and introduces negligible inference-time overhead.", "authors": ["Wenhao Liu", "Hao Shi", "Yunhe Li", "Weizhi Fei", "Xiangyuan Wang", "Mengzhe Ruan", "Hanxu Hou", "Peisong Wang", "Linqi Song", "Shuang Qiu"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-09", "url": "https://arxiv.org/abs/2606.11164", "pdf_url": "https://arxiv.org/pdf/2606.11164v1", "arxiv_id": "2606.11164", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "63958be0f7d79f55b95e7ad84e19c28b0f011c5e5a7651835a5caed433a9ff80", "sources": ["arxiv", "semantic_scholar"], "title": "Hasse Diagrams for Attention: A Partial Order Framework for Designing Transformer Masks", "abstract": "During the training of large Transformer models, attention masks regulate the scope and direction of information flow across a sequence. Numerous mask variants exist, and operators such as FlexAttention already support arbitrary attention masks. Nevertheless, a systematic formal analysis of the information-flow structure induced by arbitrary masks has been missing. This paper develops a complete theoretical framework. We prove that, with sufficient depth, the information flow of a multi-layer Transformer converges to a Hasse diagram -- a directed acyclic graph representing a partial order. Building on this, we recast the design of parallel training tasks as the problem of finding a minimal common supergraph of Hasse diagrams, and we establish a criterion for the minimal common supergraph. This yields a constructive method to derive attention masks directly from a family of tasks. Applying the framework, we design two novel masks: a block-generation attention mask that ensures training-inference consistency (Block Two-Stream Attention), and a fully supervised bidirectional attention mask (Butterfly Attention). These results demonstrate the framework's capacity to discover new structures.", "authors": ["Chentao Li", "Han Guo"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.09951", "pdf_url": "https://arxiv.org/pdf/2606.09951v1", "arxiv_id": "2606.09951", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b67d706fa49ff7442ed69f855cce85b8f7a8f6a137daecd6262db5419ec6a586", "sources": ["arxiv", "semantic_scholar"], "title": "FAME: Forecastability-Aware Mixture of Experts for Heterogeneous Time Series Forecasting", "abstract": "Large-scale retail and industrial forecasting systems contain many heterogeneous time series whose lifecycle, sparsity, volatility, seasonality, spectral patterns, and contextual sensitivity differ substantially. A single forecasting model rarely performs well across all regimes, while dense ensembles increase inference cost and provide limited insight into expert suitability. This paper studies forecastability-aware expert routing: learning how data characteristics determine the suitability of forecasting experts. We propose \\method{}, a sparse mixture-of-experts framework that represents each series with a multidimensional forecastability fingerprint, mines expert-suitability targets from validation performance, and trains a cost-aware sparse router to activate a small budgeted set of experts for each series. Using a production-scale vending-machine sales dataset from Shandong New Beiyang (SNBC), where the forecasting component has been integrated into the replenishment-planning pipeline, together with public retail benchmarks, we show that expert suitability varies systematically across data regimes. On the industrial dataset with 5,000+ machines and 60M+ transactions, \\method{} Top-2 reduces MSE by 12.4\\% over the strongest single expert, LightGBM, while executing 1.92 experts per series on average. The deployed component produces demand forecasts, while inventory-oriented gains are estimated by an offline replay simulator under a fixed replenishment policy rather than by online intervention. The framework turns heterogeneous sales forecasting from heuristic model selection into data mining of forecastability patterns and expert specialization. Code is available at https://github.com/hit636/FAME", "authors": ["Qianyang Li", "Xingjun Zhang", "Shaoxun Wang", "Tao Peng", "Jia Wei"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.08896", "pdf_url": "https://arxiv.org/pdf/2606.08896v1", "arxiv_id": "2606.08896", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/hit636/FAME", "venue": null, "quality_score": 0.65}
{"id": "5b8b06a740c2f8125a9b833ea12dd15aaf50baff0b999dc6966696052160f236", "sources": ["arxiv", "semantic_scholar"], "title": "STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control", "abstract": "Low-rank projection has emerged as a promising approach for compressing the KV cache by exploiting hidden-dimension redundancy. However, prior methods rely on fixed or heuristic rank selection and struggle to achieve aggressive compression with minimal accuracy degradation. We propose STAR-KV, an adaptive low-rank KV cache compression framework with fine-grained rank control. STAR-KV encompasses 1) a differentiable thresholding mechanism that enables optimal rank selection at both attention-head and block levels, 2) a hybrid decomposition strategy that applies different low-rank factorizations according to the sensitivity of key and value projections, and 3) a low-rank-aware mixed precision quantization that leverages data statistics for near lossless low-bit quantization. Evaluated across multiple LLMs and benchmarks, STAR-KV achieves up to 75% KV cache compression and up to 20x overall KV cache reduction when combined with quantization. Enabled by custom Triton-based GPU kernels, STAR-KV delivers up to 6.9x speedup for the attention module and 3.1x end-to-end generation throughput. Our code is publicly available at: https://github.com/PriyanshBhatnagar/STAR-KV.", "authors": ["Priyansh Bhatnagar", "Ashkan Moradifirouzabadi", "Se-Hyun Yang", "SeungJae Lee", "Jungwook Choi", "Mingu Kang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-07", "url": "https://arxiv.org/abs/2606.08382", "pdf_url": "https://arxiv.org/pdf/2606.08382v1", "arxiv_id": "2606.08382", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/PriyanshBhatnagar/STAR-KV", "venue": null, "quality_score": 0.65}
{"id": "efc8440b17f8aaa7bdb84eb6002ed6eaa67e37e7480cd8e091cc563d8e6e1da6", "sources": ["arxiv", "semantic_scholar"], "title": "RKSC: Reasoning-Aware KV Cache Sharing and Confident Early Exit for Multi-Step LLM Inference", "abstract": "We introduce RKSC (Reasoning-Aware KV Cache Sharing), a training-free inference framework that eliminates two structural redundancies in multi-branch LLM reasoning pipelines. ASKS (Attention-Similarity KV Sharing) computes the prefix KV cache once and broadcasts it to all semantically similar branches via hidden-state cosine similarity, strictly generalising the token-exact prefix caching used by vLLM and SGLang. CGEE (Confidence-Gated Early Exit) applies two complementary exit mechanisms: (1) it skips the verification forward pass entirely when generation confidence is decisive across branches, and (2) it terminates the verification pass at an intermediate layer when per-layer entropy stabilises, using lightweight hooks on the transformer backbone. RSBCM (Reasoning-Selective Block Cache Manager) prevents unbounded cache growth via attention-weighted depth-priority eviction. Across five model families (7B-10B), four benchmarks, and 1,000 evaluated problems, RKSC achieves a mean speedup of 3.008x over the No-KV baseline (peak 3.990x), a 1.66x mean improvement over vLLM-equivalent prefix caching, with a CGEE-induced error rate of only 0.37% (6 errors out of 1,616 verify calls). No fine-tuning or architecture changes are required. Code is available at https://github.com/AnirudhSekar/RKSC.", "authors": ["Anirudh Sekar"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-07", "url": "https://arxiv.org/abs/2606.09937", "pdf_url": "https://arxiv.org/pdf/2606.09937v1", "arxiv_id": "2606.09937", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AnirudhSekar/RKSC", "venue": null, "quality_score": 0.65}
{"id": "33fae3d60cba349da24ac57e1f7e475f6068bbf2641e46632a9820997a80dbcd", "sources": ["arxiv", "semantic_scholar"], "title": "SpectrumKV: Per-Token Mixed-Precision KV Cache Transfer for Prefill-Decode Disaggregated LLM Serving", "abstract": "Prefill-decode (PD) disaggregation decouples prompt processing from token generation, but it also turns the key-value (KV) cache into a network payload. Existing PD-side KV reduction methods are mostly binary: selected tokens are transmitted at full precision and the rest are not transmitted. This paper argues that binary selection leaves a useful design space unused. SpectrumKV assigns a precision level to each token instead: attention sinks and other high-importance tokens are protected at FP16, medium-importance tokens are sent at INT8, and low-importance tokens are sent at INT4 when the model can tolerate it. The main practical complication is that INT4 tolerance is model-dependent. Qwen2.5-7B catastrophically fails under INT4 KV quantization, while Mistral-7B and Gemma-2-9B remain stable. SpectrumKV therefore runs a lightweight deployment-time probe: three aggressive NIAH trials under a 3-tier policy. Models that pass use FP16+INT8+INT4; models that fail fall back to FP16+INT8. Across Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3, and Gemma-2-9B-it, SpectrumKV improves quality at the same transfer budget. At a 50% normalized KV budget on WikiText-2, SpectrumKV changes perplexity by +1.97%,-0.06%, and-0.44%, respectively, compared with PDTrim's +25.85%, +22.07%, and +35.63%. On NIAH retrieval at 4096 tokens, the adaptive policy reaches 52.6% on Qwen at the aggressive b=0.3 budget versus 26.3% for PDTrim, and reaches 100% by b=0.5; Mistral and Gemma preserve retrieval under the 3-tier policy. End-to-end GPU timing of the transfer path shows 50-62% TTFT reductions at b=0.5. These results suggest that PD KV transfer should be treated as a precision-allocation problem, not only as token pruning.", "authors": ["Yang Pengju"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-07", "url": "https://arxiv.org/abs/2606.08635", "pdf_url": "https://arxiv.org/pdf/2606.08635v1", "arxiv_id": "2606.08635", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b23081733a04a06845371ec11f9dd0f8d8f83c9830e641ce0dcfa2a03cd0a1a3", "sources": ["arxiv", "semantic_scholar"], "title": "IntentKV: Cross-Turn Intent-Aware KV Cache Pruning for Agent Inference", "abstract": "Multi-turn LLM agents fan short queries into long trajectories of tool calls, search results, and intermediate reasoning. Both KV memory and KV read bandwidth grow by orders of magnitude across a single trajectory, making the key-value (KV) cache, not parameter compute, the dominant serving bottleneck for long-horizon agents. We introduce IntentKV, learned KV pruning that keeps the base LLM frozen. IntentKV maintains a session-level QueryMemory of cross-turn intent, scores live history tokens with a memory-attention rule, and adds a zero-initialized residual head with cross-attention over current-query K-vectors. To stay composable with prefix caches, eviction is a slot-map redirection: dropped positions route to a sentinel dead slot while surviving K/V rows, RoPE phases, and slot identities stay in place. IntentKV matches the no-pruning full-cache baseline with almost no accuracy drop under tight KV budgets: at an 8k KV budget, mean peak request tokens drop 23.9% on Qwen3-8B and 30.7% on Qwen2.5-14B. On the 100 longest BCP queries that all methods complete on Qwen2.5-14B, IntentKV-8k further cuts worst-case peak request tokens from 92.3k to 20.5k, a 77.8% reduction, and worst-case raw KV reads from 411M to 31M, a 92.6% reduction.", "authors": ["Junjie Li", "Jiong Lou", "Jie Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-06", "url": "https://arxiv.org/abs/2606.09916", "pdf_url": "https://arxiv.org/pdf/2606.09916v1", "arxiv_id": "2606.09916", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "dbc3790ce5ee01b2483c84c367f615f419cf3aab4616a9c2bf9c9005338becab", "sources": ["arxiv", "semantic_scholar"], "title": "Joint Structural Pruning and Mixed-Precision Quantization for LLM Compression", "abstract": "Recently, the efficiency of Large Language Models (LLMs) deployment has become a critical concern in practical applications. While post-training quantization (PTQ) and structural pruning are established techniques for reducing memory footprint and inference latency, most existing PTQ approaches optimize quantization errors on a per-layer basis, overlooking how errors accumulate and propagate through the network, often resulting in suboptimal solutions. Traditional pipelines also tend to apply pruning and quantization in isolation or sequentially, further compounding sub-optimality. We introduce a novel end-to-end framework that addresses these limitations in two key ways. First, we propose a novel mixed-precision PTQ strategy that directly minimizes global error propagation across the entire model, rather than isolating layer-wise errors. Building on this, we develop a novel joint optimization approach that simultaneously learns structural pruning decisions and mixed-precision quantization policies within a unified search space. Extensive experiments show that, at ultra-low precisions (1-3 bits), our quantization method reduces WikiText perplexity by up to 21% compared to state-of-the-art (SoTA) weight-activation quantization baselines. Against leading weight-only quantization methods, it achieves up to 59% and 85% lower perplexity on WikiText and C4, respectively. Compared to the SoTA joint pruning-and-quantization techniques, our proposed method delivers superior perplexity and reasoning performance at ultra-low bits.", "authors": ["Hoang-Loc La", "Truong-Thanh Le", "Amir Taherkordi", "Phuong Hoai Ha"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-05", "url": "https://arxiv.org/abs/2606.07819", "pdf_url": "https://arxiv.org/pdf/2606.07819v1", "arxiv_id": "2606.07819", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "fc94663e3a52b334f0bb7bf17e50e656f58afafbafa092487ad58b721e75ed1c", "sources": ["arxiv", "semantic_scholar"], "title": "Still: Amortized KV Cache Compaction in a Single Forward Pass", "abstract": "The KV cache is the memory bottleneck of long-horizon language model deployment. Practically, a deployable compactor must be lightweight enough to call during inference, expressive enough to preserve context under constraint, and reusable across a trajectory. Existing compaction methods satisfy only part of this requirement: selection methods are lightweight but subset-bound, while synthesis methods are expressive but rely on per-context optimization. Here we introduce Still, a small per-layer Perceiver trained once against a frozen base model that produces compact keys and values in a single forward pass. On Qwen and Gemma models, Still occupies the favorable side of the speed--quality frontier across compression ratios from $8\\times$ to $200\\times$ and context lengths from $8$k to $128$k. On the long-context RULER grid, Still exceeds the strongest baseline by 8--22 points. The same compact cache also supports free-form summarization, preserving most of the full-context gain on HELMET and winning a pairwise LongBench summarization comparison against KV-Distill. Because compaction is a forward pass, Still can be applied iteratively, entering a long-horizon regime unavailable to per-context methods. We show that amortization makes long-context cache compaction tractable, and synthesis makes its compact state useful at extreme compression.", "authors": ["Charles O'Neill", "Alex Sandomirsky", "Harry Partridge", "Mudith Jayasekara", "Max Kirkby"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-05", "url": "https://arxiv.org/abs/2606.07878", "pdf_url": "https://arxiv.org/pdf/2606.07878v1", "arxiv_id": "2606.07878", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "372102df62ed14b3d57d05fe483fa9c3d00f97f737e9bc0c395a409e85d8ff70", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking LoRA Memory Through the Lens of KV Cache Compression", "abstract": "Parametric retrieval augmentation encodes document information into lightweight, document-specific modules such as LoRA adapters, reducing the need to include all evidence as input context. However, it remains unclear how this parameter-side memory interacts with context-side memory stored in the KV cache. We study this interaction in document-level question answering by progressively evicting document key-value states and measuring when a document LoRA contributes beyond the retained context. We find that document LoRA adds little when the KV cache is largely intact, but becomes increasingly useful under aggressive compression, recovering 13-21 ROUGE-L points when no document context remains. The gain is largest when the base model encodes the document, and the adapter is applied only during answer generation, suggesting that document LoRA is better understood as decoding-time parametric memory than as a document encoder. Finally, QA-style supervision produces substantially stronger adapters than raw-context next-token-prediction. These results position document LoRA as a complementary memory channel whose value emerges precisely when context-side evidence is scarce.", "authors": ["Chunsheng Zuo", "Liaoyaqi Wang", "William Jurayj", "William Fleshman", "Benjamin Van Durme"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.05698", "pdf_url": "https://arxiv.org/pdf/2606.05698v1", "arxiv_id": "2606.05698", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "51c0fafe4237deb7e37e634c4b1eb03507a834d97ea6de053d7c8a30353a889a", "sources": ["arxiv", "semantic_scholar"], "title": "Value-and-Structure Alignment for Routing-Consistent Quantization of Mixture-of-Experts Models", "abstract": "Mixture-of-Experts (MoE) models scale foundation models efficiently by activating only a subset of experts for each token, but their large number of expert parameters still makes quantization essential for practical deployment. Unlike dense models, however, MoE models are sensitive to routing instability: small quantization-induced perturbations can change the top-$k$ expert selection, altering the computation path and degrading model quality. We propose Value-and-Structure Routing Alignment for Quantization (VSRAQ), a MoE-specific post-training quantization objective that preserves pre-quantization expert-selection behavior under quantization. VSRAQ combines two complementary objectives that jointly preserve expert-selection behavior: value alignment, which matches routing-relevant logits or scores, and structure alignment, which preserves expert ordering and top-$k$ decision boundaries. By maintaining routing consistency, VSRAQ reduces quantization-induced degradation without introducing any inference-time overhead and can be integrated into existing quantization frameworks. Experiments on recent MoE foundation models show that VSRAQ improves expert-selection consistency and consistently outperforms reconstruction-only and router-aware baselines.", "authors": ["Hancheol Park", "Geonho Lee", "Tairen Piao", "Tae-Ho Kim"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.05688", "pdf_url": "https://arxiv.org/pdf/2606.05688v1", "arxiv_id": "2606.05688", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "471c2374e70edf4b8dca0cb9a2abe3a82a2e50ca6d72249e1675752607e033d6", "sources": ["arxiv", "semantic_scholar"], "title": "High-Dimensional Theory of LoRA Fine-Tuning in a Solvable Attention Model", "abstract": "We develop a high-dimensional statistical theory of low-rank adaptation (LoRA) in attention models, capturing the interplay between pre-training and fine-tuning. We introduce a solvable framework in which a single-head attention layer is first pre-trained on a data-abundant task and subsequently adapted via a rank-one LoRA update on limited data. In the high-dimensional limit, both stages admit a sharp asymptotic characterization in terms of a finite set of order parameters, yielding explicit predictions for test errors and representation alignment. Our analysis shows that the impact of pre-training on LoRA is summarized by an effective noise term, from which we derive prescriptions for the optimal pre-training procedure. We also demonstrate a regime with a mismatch between the value of the test error and representation quality, and propose an application of our theory to active fine-tuning.", "authors": ["O. Duranthon", "F. Boncoraglio", "L. Zdeborová"], "categories": ["cs.LG", "cond-mat.dis-nn"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.05899", "pdf_url": "https://arxiv.org/pdf/2606.05899v1", "arxiv_id": "2606.05899", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c4b8ac494f3da3094edc9fca7613c163c350b710455dac1c9809d5594defdd47", "sources": ["arxiv", "semantic_scholar"], "title": "Tangram: Unlocking Non-Uniform KV Cache for Efficient Multi-turn LLM Serving", "abstract": "Multi-turn Large Language Model (LLM) serving is critical for consistent user experiences, yet the linear growth of the Key-Value (KV) cache imposes significant pressure on GPU memory and bandwidth. Non-uniform KV compression effectively preserves more information by considering the individual importance of each KV cache. However, such KV cache heterogeneity introduces various systemic challenges - including memory fragmentation, scheduling complexities, and diminished kernel utilization - which collectively lead to significant inefficiencies in existing LLM serving systems. To overcome these challenges, we present Tangram, a novel serving system designed to make Non-uniform KV caches practical. Tangram addresses systemic inefficiencies through three core techniques: (1) Deterministic Budget Allocation assigns a static memory footprint to each head based on its intrinsic pattern, entirely eliminating dynamic scheduling overhead and prefill stalls; (2) Head Group Page clusters attention heads with similar retention demands and manages them with independent, vectorized page tables, thereby maximizing physical memory reclamation; and (3) Ahead-of-Time (AOT) Load Balancing leverages static budget profiles to ensure uniform GPU utilization without runtime overhead. Experimental results show that Tangram improves throughput by up to 2.6x compared to existing baselines, while fully preserving model accuracy. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.", "authors": ["Hyungmin Kim", "Minsoo Kim", "Hongseok Kim", "Jungwook Choi"], "categories": ["cs.LG", "cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.06302", "pdf_url": "https://arxiv.org/pdf/2606.06302v1", "arxiv_id": "2606.06302", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/aiha-lab/TANGRAM", "venue": null, "quality_score": 0.65}
{"id": "129136d5dd14aef1a6b02c7efcdccb266321e11616a9bdeefbdd807fd2d42058", "sources": ["arxiv", "semantic_scholar"], "title": "RedKnot: Efficient Long-Context LLM Serving with Head-Aware KV Reuse and SegPagedAttention", "abstract": "As the input length of large language model (LLM) serving continues to grow, the KV cache has become a dominant bottleneck in AI infrastructure. It limits GPU memory capacity, serving concurrency, cache reuse, and distributed scalability. Several important problems, including position-independent KV cache, prefix KV cache compression, hot/cold KV cache separation, and distributed KV cache management, all depend on how the KV cache is represented and managed. However, existing serving systems largely rely on a monolithic KV cache abstraction, where the KV cache is treated as a homogeneous sequence of token-level memory blocks and managed with similar policies across attention heads and serving scenarios. We observe that KV cache utility is highly structured across KV heads: different heads exhibit different functional roles, attention distances, and runtime importance. Therefore, a full KV cache is not always necessary for every head, token range, or serving scenario. We present RedKnot, a head-aware KV cache management system for LLM serving. RedKnot breaks the conventional monolithic KV cache abstraction by decomposing the KV cache along KV heads, whose importance and effective attention ranges vary significantly across serving scenarios. This head-level decomposition turns the KV cache from a monolithic tensor abstraction into a structured memory object, enabling RedKnot to uniformly support position-independent KV reuse, prefix KV compression, hot/cold KV separation, and distributed KV placement while preserving output fidelity and improving resource efficiency, without requiring model retraining or fine-tuning. RedKnot establishes a new foundation for AI infrastructure by transforming the KV cache from a monolithic, passive runtime artifact into a dynamic, model-aware runtime substrate for scalable LLM serving.", "authors": ["Yang Liu", "ZhaoKai Luo", "HuaYi Jin", "ZhiYong Wang", "RuoZhou He", "BoYu Wang", "Guanjie Chen", "Junhao Hu"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-04", "url": "https://arxiv.org/abs/2606.06256", "pdf_url": "https://arxiv.org/pdf/2606.06256v1", "arxiv_id": "2606.06256", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5e79679071a5fc73dbb9b13e6c745fa1e5037e94db9182a12319b137c46ffb3f", "sources": ["arxiv", "semantic_scholar"], "title": "TENP: Trapezoidal Expert Neuron Pruning For Mixture-of-Experts", "abstract": "Mixture-of-Experts large language models (LLMs) scale efficiently through sparse activation, yet their deployment is fundamentally constrained by the large static parameter footprint of experts. Existing compression approaches either remove entire experts, disrupting routing topology and harming performance, or rely on unstructured weight pruning with limited practical efficiency. To address the limitations, we propose TENP, a structured Trapezoidal ExpertNeuron Pruning framework. Using a few samples, we identify and retain important experts, while applying expert neuron pruning (ENP) to less important experts, reserving model parameters in a trapezoidal pattern from shallow to deep layers. When evaluating expert importance, we jointly consider both the magnitude of the expert output and its ability to change the direction of the input vector. For ENP, we measure each neuron's projected contribution to the expert output to identify and retain important neurons. We conduct extensive experiments on the Qwen and DeepSeek models. Under a routing expert sparsity of 40% and an average of 63.76% activated expert parameters, the DeepSeek model suffers only a 1-point drop in accuracy compared to the full-parameter model. Moreover, it outperforms the full-parameter model by 10% on code generation tasks.", "authors": ["Jiangyang He", "Shaolin Zhu", "Deyi Xiong"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-06-03", "url": "https://arxiv.org/abs/2606.09885", "pdf_url": "https://arxiv.org/pdf/2606.09885v1", "arxiv_id": "2606.09885", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ff4d98f6b0b06cd072a826b0088f9a2be1ac76f474e701014d7d0e6caf8817ff", "sources": ["arxiv", "semantic_scholar"], "title": "SHAPE: Coalition-Aware Expert Pruning for Sparse Mixture-of-Experts LLMs", "abstract": "Sparse Mixture-of-Experts (MoE) large language models achieve strong quality with low per-token compute, yet their deployment is often limited by the memory wall: the full expert pool must remain resident to support token-dependent routing. Expert pruning is a direct remedy, but prior criteria typically score experts independently and overlook that MoE inference is inherently \\emph{coalitional}, where outputs arise from routed top-$k$ expert combinations. We propose \\textbf{SHAPE}, a task-driven pruning framework that explicitly models \\emph{intra-layer} expert cooperation. SHAPE formulates routing traces on a small calibration set as an empirical cooperative game and assigns interaction-aware expert values via a Shapley-style attribution over observed top-$k$ coalitions, enabling the identification of experts that are essential for high-utility collaborations rather than merely frequent. To preserve MoE topology under a global pruning budget, SHAPE further introduces a \\emph{quality-coverage} selection rule that retains, in each layer, the minimal expert subset covering an $α$ fraction of non-negative Shapley mass, while using bisection to match a target keep rate. Experiments on three modern MoE backbones (Qwen3-30B-A3B, GPT-OSS-20B, and DeepSeek-V2-Lite) across diverse benchmarks show that SHAPE consistently improves robustness over global and layer-wise pruning variants, maintaining competitive accuracy under 20\\% and 40\\% expert pruning without additional training and delivering clear reductions in peak GPU memory footprint. The open-source code is available at https://github.com/Alizen-1009/Shapley-Moe.", "authors": ["Yuhao Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-03", "url": "https://arxiv.org/abs/2606.09886", "pdf_url": "https://arxiv.org/pdf/2606.09886v1", "arxiv_id": "2606.09886", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Alizen-1009/Shapley-Moe", "venue": null, "quality_score": 0.65}
{"id": "537f9582cfd26567a2497952d288598b0a2b66512119b9c0727089ab5cade6a8", "sources": ["arxiv", "semantic_scholar"], "title": "AlphaQ: Calibration-Free Bit Allocation for Mixture-of-Experts Quantization", "abstract": "Mixture-of-Experts (MoE) architectures scale model capacity through sparse expert activation, but their deployment remains memory-bound because all expert weights must reside in memory. Mixed-precision quantization can substantially reduce this footprint by assigning different bit-widths to different experts. Existing approaches, however, typically rely on calibration data to estimate expert importance and determine bit allocation. For frontier MoE LLMs, the original training data, and hence the true training distribution, is proprietary and inaccessible. As a result, calibration sets are inevitably imperfect surrogates, and this can misestimate expert utilization and lead to suboptimal bit allocation. Motivated by the substantial cross-expert quality variability observed in modern MoE models, and by the success of Heavy-Tailed Self-Regularization (HT-SR) theory at predicting neural network model quality without access to training or testing data, we propose AlphaQ, a calibration-free bit-allocation method for MoE quantization. AlphaQ draws on HT-SR theory and follows a simple principle: experts with more heavy-tailed weight spectra are typically better trained and hence should receive higher bit-widths, while experts with weaker heavy-tailed structure can be quantized more aggressively. AlphaQ operationalizes this principle by measuring expert-wise spectral heavy-tailedness and solving a budget-constrained optimization problem that minimizes total quantization error under a global bit-budget constraint. Across several MoE models, AlphaQ consistently outperforms calibration-based baselines under matched bit budgets. Notably, on Qwen1.5-MoE, AlphaQ achieves near full-precision accuracy with an average expert precision of only 3.5 bits, while delivering more than 4$\\times$ memory compression. Our code is available at https://github.com/Superone77/AlphaQ.", "authors": ["Wanqi Yang", "Yuexiao Ma", "Alexander Conzelmann", "Xiawu Zheng", "Michael W. Mahoney", "T. Konstantin Rusch", "Shiwei Liu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-03", "url": "https://arxiv.org/abs/2606.04980", "pdf_url": "https://arxiv.org/pdf/2606.04980v1", "arxiv_id": "2606.04980", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Superone77/AlphaQ", "venue": null, "quality_score": 0.65}
{"id": "a0f8e959719bbb512617d133507d6f0ef3ba6ad79a0bc204ce1a0e227db08f48", "sources": ["arxiv", "semantic_scholar"], "title": "Cartridges at Scale: Training Modular KV Caches over Large Document Collections", "abstract": "Large Language Models can reason over long contexts, yet prefilling millions of tokens is wasteful as much of the content remains static across queries. Cartridges address this by distilling document collections into reusable key-value (KV) caches that eliminate prefilling while preserving accuracy. A critical limitation of this approach is that cartridges are monolithic and non-compositional: encoding an entire collection into a single KV block does not scale, and naively mixing cartridges trained in isolation collapses performance to near chance. We introduce Cartridges at Scale (CAS), a training framework for scalable multi-cartridge learning with dynamic distractor mixing and a memory-efficient budget manager that rotates hundreds of per-document cartridges between GPU and persistent storage. Our approach scales to collections exceeding a million tokens, improving over a monolithic cartridge by 10-31 points at comparable token budgets. Oracle cartridge accuracy falls within 2-6 points of full in-context learning even at high compression. When paired with retrieval for cartridge selection, CAS matches or exceeds conventional RAG accuracy while consuming 3-4x fewer prompt tokens.", "authors": ["Momchil Hardalov", "Gonzalo Iglesias", "Adrià de Gispert"], "categories": ["cs.CL", "cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-03", "url": "https://arxiv.org/abs/2606.04557", "pdf_url": "https://arxiv.org/pdf/2606.04557v1", "arxiv_id": "2606.04557", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "77729832691f6dcb2f801e0f5b838fbbc5d527cd03d27068d1e61d8f3715fd4e", "sources": ["arxiv", "semantic_scholar"], "title": "KVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning Tasks", "abstract": "Test-time scaling is a powerful approach to obtain better reasoning in large language models, but it becomes memory-bottlenecked during long-horizon decoding, as the KV-cache grows. KV-cache quantization can help improve this, but current methods are evaluated under prefill-like settings and errors behave differently under autoregressive decoding. We show that in the latter regime, quantization errors accumulate across timesteps, driven primarily by incorrect token scales. We introduce KVarN, a calibration-free KV-cache quantizer that applies a Hadamard rotation followed by a dual-scaling variance normalization across both axes of the K and V matrices. We find that this combination fixes outlying token-scale errors and substantially reduces error accumulation over existing baselines. KVarN establishes a new state-of-theart for KV-cache quantization on generative benchmarks, including MATH500, AIME24 and HumanEval, at 2-bit precision. A vLLM implementation of the KVarN method is available at https://github.com/huawei-csl/KVarN", "authors": ["Lorenz K. Muller", "Philippe Bich", "Chiara Boretti", "Hyun-Min Chang", "Jiawei Zhuang", "Lukas Cavigelli"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.03458", "pdf_url": "https://arxiv.org/pdf/2606.03458v1", "arxiv_id": "2606.03458", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/huawei-csl/KVarN", "venue": null, "quality_score": 0.65}
{"id": "c68b97e1f51efc1fd997dbcd5b61cbb54762a477c10b83230598596ab9ced734", "sources": ["arxiv", "semantic_scholar"], "title": "Value-Aware Stochastic KV Cache Eviction for Reasoning Models", "abstract": "Reasoning models improve accuracy through extended chains of thought, but their long outputs create a memory and compute bottleneck. KV cache eviction methods reduce this cost by evicting unimportant key-value pairs from the cache, yet they often yield worse accuracy than selection-based sparse attention alternatives, which keep the full KV cache. We identify key factors crucial to KV cache eviction accuracy. First, a small fraction of value states have abnormally large magnitudes, and evicting them causes catastrophic failure where models enter repetitive reasoning loops. Second, introducing stochasticity during eviction improves accuracy by increasing cache diversity. Based on these findings, we propose Value-aware Stochastic KV Cache Eviction (VaSE), a training-free recipe that protects large-magnitude value states and promotes diverse eviction decisions. Across six reasoning tasks, Qwen3 models using VaSE with 4x KV cache compression yield higher average accuracies than SOTA selection method at the same sparsity, while outperforming the strongest eviction method by more than 4%. Overall, VaSE bridges the gap between efficiency and accuracy, supporting FlashAttention2 and enabling a static memory footprint for reasoning models.", "authors": ["Ting-Yun Chang", "Harvey Yiyun Fu", "Deqing Fu", "Chenghao Yang", "Jesse Thomason", "Robin Jia"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.03928", "pdf_url": "https://arxiv.org/pdf/2606.03928v1", "arxiv_id": "2606.03928", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/terarachang/VaSE", "venue": null, "quality_score": 0.65}
{"id": "7681332cc50b38aa553bfe6aab5d226bdad32a0775c5903011864fe46027c399", "sources": ["arxiv", "semantic_scholar"], "title": "When Attention Collapses: Stage-Aware Visual Token Pruning from Structure to Semantics", "abstract": "Vision-Language Models (VLMs) have demonstrated remarkable capabilities but suffer from significant computational overhead during inference. While visual token pruning offers a promising solution, existing methods predominantly rely on initial attention scores. This single-metric paradigm presents a critical flaw: high attention scores inherently collapse onto semantically similar regions, thereby severely reducing feature diversity and discarding vital contextual details. To address this, we introduce Structure-to-Semantics (STS), a novel two-stage visual token pruning framework that explicitly decouples the pruning process. The first stage employs a repulsion-based sampling mechanism to maximize spatial and structural diversity. The second stage leverages instruction-aware cross-attention to precisely filter out prompt-irrelevant tokens. This two-stage synergy constitutes the core of STS, first ensuring geometric coverage and then refining the retained tokens according to semantic relevance. Extensive evaluations demonstrate that STS mitigates the redundancy caused by attention-based selection, improving both structural diversity and fine-grained task alignment of the preserved visual tokens.", "authors": ["Jiahui Wang", "Kai Zhang", "Mai Han", "Huanghe Zhang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.03569", "pdf_url": "https://arxiv.org/pdf/2606.03569v1", "arxiv_id": "2606.03569", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "57fa458b7940c77d7244cde31c9cbc114944961a12c01886d993d23623d27554", "sources": ["arxiv", "semantic_scholar"], "title": "Recover-LoRA for Aggressive Quantization: Reclaiming Accuracy in 2-Bit Language Models via Low-Rank Adaptation with Knowledge Distillation on Synthetic Data", "abstract": "Aggressive weight quantization to 2-bit precision offers substantial throughput and memory gains for large language model (LLM) inference, but typically incurs severe accuracy degradation. These gains are particularly relevant for edge and on-device deployment, where memory capacity and bandwidth are primary constraints. In this work, we extend Recover-LoRA -- a lightweight, data-free accuracy recovery method originally developed for general model weight corruption -- to the setting of ultra-low-bit quantization. We propose a selective mixed-precision strategy in which only gate and up projection layers of the MLP are quantized to 2-bit (W2), while all other linear layers remain at higher precision, yielding a mixed-precision GateUp configuration. We demonstrate via roofline analysis across three model families (4B--20B) and two hardware platforms that a W4/W2-GateUp deployment (4-bit base with 2-bit gate/up) delivers 7.5--23.3\\% TPS improvement over uniform W4 depending on model and context length, while confining quantization error to a predictable subset of layers. We then apply Recover-LoRA -- training low-rank adapters on the quantized layers via logit distillation with synthetic data -- to recover accuracy lost from 2-bit quantization of the gate and up layers. In a case study on Qwen3-4B, Recover-LoRA achieves 80--95\\% accuracy recovery on 9 of 12 benchmarks, using only 10k synthetic training samples and no labeled data. We further demonstrate that synthetic data performs comparably to curated labeled data for distillation-based recovery, and that recovery generalizes to out-of-distribution evaluation tasks. Our results present Recover-LoRA as a practical post-quantization accuracy recovery tool for aggressive weight compression in deployment settings.", "authors": ["Devleena Das", "Rajeev Patwari", "Elliott Delaye", "Ashish Sirasao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.04238", "pdf_url": "https://arxiv.org/pdf/2606.04238v1", "arxiv_id": "2606.04238", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ffa7dd593613a170a8aefdf8639b244173ec134c9309d4131f852ba6e5d3818e", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Mixture-of-Experts Reward Models Learn Interpretable and Specialized Experts for Personalized Preference Modeling", "abstract": "Preference modeling plays a central role in reinforcement learning from human feedback (RLHF), enabling large language models (LLMs) to align with human values. However, most existing approaches assume a universal reward function, neglecting the diversity and heterogeneity of human preferences. To address this limitation without additional annotation costs, recent work has proposed learning multiple preference components from binary data and combining them to model individual preferences. Nevertheless, these components often fail to capture coherent and disentangled patterns, limiting their interpretability and effectiveness for personalization. In this work, we propose a sparse Mixture-of-Experts (MoE) reward model that encourages sparse routing and expert diversity during training on binary preference data. Across controlled and real-world experiments, sparse MoE learns interpretable routing patterns and specialized experts. It also improves test-time personalization, and post-adaptation shifts in expert weights provide a qualitative lens for analyzing how the model adapts to personalized preferences.", "authors": ["Yifan Wang", "Jinyi Mu", "Mayank Jobanputra", "Yu Wang", "Ji-Ung Lee", "Soyoung Oh", "Isabel Valera", "Vera Demberg"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.04284", "pdf_url": "https://arxiv.org/pdf/2606.04284v1", "arxiv_id": "2606.04284", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "223efefd910aa3e306c10b25d31094236bc1e17c222b923eec57e564dd79e118", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Segment Attention: Enabling Efficient KV-Cache Management for Faster Large Language Model Serving", "abstract": "Large Language Model (LLM) inference relies on key-value (KV) caches to avoid redundant attention computation. While approximate KV cache retention techniques reduce memory usage by sacrificing model accuracy, lossless approaches instead evict KV cache blocks from GPU memory and reconstruct them on demand to preserve exact outputs. Existing lossless KV cache management systems primarily base eviction decisions on access frequency or positional heuristics, without considering how different KV cache blocks affect the execution efficiency of GPU attention kernels. In this paper, we propose AsymCache, a computation-latency-aware KV cache management system for LLM inference that explicitly aligns cache residency decisions with GPU attention kernel performance, including three key components: Multi-Segment Attention (MSA) for efficient non-contiguous KV context processing, a cache eviction policy that jointly optimizes hit rate and position-aware recomputation cost, and an adaptive chunking scheduler for high hardware utilization. Experiments show that AsymCache reduces TTFT by up to 1.90-2.03x and time-per-output-token (TPOT) by 1.62-1.71x over latest baselines, confirming the effectiveness of the method in common workloads and validating its design goal of balancing computational efficiency with cache hit rate. Moreover, the low-level design of AsymCache allows seamless integration into agent serving systems such as Continuum, where it further reduces average job latency by up to 18.1%.", "authors": ["Chunan Shi", "Yilei Chen", "Yilin Chen", "Xupeng Miao", "Bin Cui"], "categories": ["cs.AR", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.02964", "pdf_url": "https://arxiv.org/pdf/2606.02964v1", "arxiv_id": "2606.02964", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "01cad6f7756b1cae7bee87682d79867eba6abae69871100607ffff9f192bfe9d", "sources": ["arxiv", "semantic_scholar"], "title": "Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation", "abstract": "Key-value (KV) cache quantization is widely used to reduce Large Language Model (LLM) inference memory, yet existing evaluations solely focus on measuring perplexity and accuracy without assessing the safety impact. In this study, we explore alignment preservation under KV cache quantization. Across eleven instruction-tuned models (3.8B-72B) and five benchmarks (1,894 prompts), we find that low-bit quantization can silently destroy safety alignment: Mistral-7B loses 15.2% of its refusals at only 1.03x perplexity, and no universal safe bit-width exists, with sharp model-specific phase transitions invisible to standard metrics. We identify that the root cause is geometric: safety features occupy a low-dimensional activation subspace 10^2-10^3x more vulnerable to quantization noise than the full representation space perplexity averages over. Inspired by this observation, we propose Per-Channel Reduction (PCR), a diagnostic that classifies each model into one of three mechanistic failure modes: outlier-crushes-safety, where safety lives in non-outlier channels collaterally damaged by outlier-driven scale factors; outlier-as-safety, where safety overlaps outlier channels and finer granularity cannot rescue it; and multi-layer dilution, where safety is distributed across many layers and per-layer fixes fail. PCR predicts the correct mitigation direction on all nine primary models and one held-out model from an independent family using 20 calibration prompts. PCR generalizes across unseen prompts, models, and production quantizers, including KIVI with up to 97.2% recovery, succeeding where attention-based allocation methods fail. The resulting training-free protocol, requiring approximately 35 GPU-minutes, recovers up to 97% of lost alignment at minimal memory overhead, addressing vulnerabilities confirmed in production vLLM serving with FP8 KV cache on NVIDIA GPUs.", "authors": ["Bruce Changlong Xu", "Adarsh Kumarappan", "Mu Zhou"], "categories": ["cs.LG", "cs.AI", "cs.ET"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.09864", "pdf_url": "https://arxiv.org/pdf/2606.09864v1", "arxiv_id": "2606.09864", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ebaadca24f30b7f91ac90e2f14a6d5f63864ec97f3eebf6db9f12d587fa98ca1", "sources": ["arxiv", "semantic_scholar"], "title": "STaR-KV: Spatio-Temporal Adaptive Re-weighting for KV Cache Compression in GUI Vision-Language Models", "abstract": "Vision-language-model-based graphical user interface (GUI) agents have shown broad automation capabilities, yet deployment is bottlenecked by a key-value (KV) cache that grows linearly with interaction steps. For instance, UI-TARS-1.5-7B consumes 76 GB of GPU memory on merely five screenshots, approaching the capacity of mainstream 80 GB accelerators. Existing KV compression methods share two structural assumptions: aggregating visual-token importance into a single shared saliency map, and applying a fixed top-B cutoff to the fused score distribution. Pilot measurements refute both: spatial specialization lives at the attention-subspace level and migrates across layers, while the score distribution drifts in shape along a trajectory. We propose STaR-KV (Spatio-Temporal Adaptive Re-weighting), a training-free KV cache compression framework that calibrates token importance along three axes: (i) subspace-aware scoring driven by online spatial mutual information; (ii) a temporal stability discount that suppresses redundant cache entries from persistently attended subspaces; and (iii) an entropy-derived temperature that adaptively reshapes the score distribution. Across four GUI benchmarks, STaR-KV achieves the strongest average accuracy among state-of-the-art KV compression methods (e.g., GUIKV, SnapKV) at matched budgets, with no compression-stage FLOPs overhead (-0.07%) and cutting peak GPU memory by nearly 40% at a 20% KV-cache budget. Code is available at https://github.com/kawhiiiileo/STaR-KV.", "authors": ["Yuhang Han", "Wenzheng Yang", "Yujie Chen", "Xiangqi Jin", "Yaojie Zhang", "Siteng Huang", "Linfeng Zhang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.01790", "pdf_url": "https://arxiv.org/pdf/2606.01790v1", "arxiv_id": "2606.01790", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/kawhiiiileo/STaR-KV", "venue": null, "quality_score": 0.65}
{"id": "21cec4d20fc191f6dd2e40165103f4531030b51c2ccda7f55a6a2fc4e34f34a4", "sources": ["arxiv", "semantic_scholar"], "title": "SparseX: Efficient Segment-Level KV Cache Sharing for Interleaved LLM Serving", "abstract": "In long-context LLM serving, the prefill stage often dominates time-to-first-token and computational cost. Although Prefix Cache in vLLM/PagedAttention has been widely used to reuse identical prompt prefixes, repeated content in practical applications frequently appears as non-prefix, cross-request, cross-turn, and cross-agent segments, which makes conventional cache mechanisms insufficient. This paper presents SparseX, a segment-level KV Cache sharing method for common serving scenarios. SparseX uses contiguous token segments as reuse units and exploits Sparse-Q indices that naturally arise in KV Cache reuse workloads to estimate the key tokens that require correction. Based on this estimate, SparseX performs Sparse-KV Recomputation within a single forward pass, thereby restoring cross-segment contextual interactions under complex interleaved reuse patterns while avoiding additional models or separate preprocessing stages for token selection. SparseX further implements a full+sparse hybrid attention mode based on a layer-specific threshold: early layers retain full attention to obtain a more stable token-importance signal, and later layers switch to sparse recomputation to improve reuse quality on complex long-context tasks. We implement SparseX-vLLM on top of vLLM, integrating segment-level cache lookup, PagedAttention management, RoPE alignment, Sparse-Q token selection, and FlashAttention backends into a unified execution path. SparseX is model-agnostic, training-free, and compatible with Prefix Cache, and it provides unified support for common online serving scenarios including multi-round chat, retrieval-augmented generation (RAG), and agent workflows.", "authors": ["Quqing Zhang", "Kai Chen", "Ning Liao", "Zehao Lin", "Bo Tang", "Feiyu Xiong", "Zhiyu Li", "Xiaoxing Wang"], "categories": ["cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.01751", "pdf_url": "https://arxiv.org/pdf/2606.01751v2", "arxiv_id": "2606.01751", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "88daca95816060cd6b9d90ec55dda0248756ec9b3b2fd67349693fd80469acb0", "sources": ["arxiv", "semantic_scholar"], "title": "MomentKV: Closing the Directional Gap in KV Cache Eviction for Long-Context Inference", "abstract": "Autoregressive decoding in Transformer-based language models relies on the KV cache, whose memory footprint grows linearly with sequence length and becomes the primary bottleneck for long-context inference. KV cache eviction addresses this by retaining a fixed-size subset of key-value pairs and discarding the rest. We identify that a primary source of output degradation is not the residual attention mass on evicted tokens, which existing methods already minimize, but a directional mismatch between the retained and evicted token sets. Specifically, the evicted tokens in practice are often near-orthogonal to the retained ones. Thus, even a small evicted mass could have an oversized impact on the resulting direction distribution and amplify into substantial output error. This reveals a fundamental limit in existing strategies. To address this, we propose MomentKV, which maintains compact, small-size moment statistics over the evicted token set, including a count, key mean, value mean, and value-key covariance. During eviction, the moment statistics is leveraged to identify tokens already well aligned with and captured by the accumulated summary, keeping the evicted set geometrically regular. During inference, they yield a closed-form first-order approximation of the evicted attention output, forming a mutually reinforcing loop between selective eviction and accurate correction. On LongBench and RULER with LLaMA-3.1-8B-Instruct and Qwen3-4B-Instruct, MomentKV outperforms all baselines at every cache budget, with the largest gains under aggressive compression.", "authors": ["Yu Li", "Binxu Li", "Tian Lan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.01563", "pdf_url": "https://arxiv.org/pdf/2606.01563v1", "arxiv_id": "2606.01563", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "84a56ce3f3c4253499659b6510485f8b267ea5ca1eb3e8bd49c5ee207fb2c611", "sources": ["arxiv", "semantic_scholar"], "title": "ProbMoE: Differentiable Probabilistic Routing for Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) models scale by activating only a small subset of experts per token. However, training such models remains challenging because top-$k$ routing is discrete and non-differentiable, requiring gradient estimators for expert selection whose design remains a central open problem. We introduce ProbMoE, a probabilistic routing framework that models expert selection as a distribution over cardinality-constrained expert subsets and formulates routing as probabilistic inference in this discrete subset space. We first propose ProbMoE Exact-$k$ routing, which samples $k$-expert subsets in the forward pass, and the backward pass uses gradients through each expert's exact marginal probability as a tractable surrogate for the true gradient. ProbMoE naturally generalizes to a dynamic-$k$ routing setting, where both training and inference constrain the routing cardinality to the same predefined range, allowing adaptive expert allocation per token. Across benchmarks and model backbones, ProbMoE Exact-$k$ achieves strong performance compared to competitive baselines, with improved expert utilization and routing diversity; ProbMoE Dynamic-$k$ achieves comparable performance with fewer activated experts.", "authors": ["Heng Zhao", "Zilei Shao", "Guy Van den Broeck", "Zhe Zeng"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.01509", "pdf_url": "https://arxiv.org/pdf/2606.01509v1", "arxiv_id": "2606.01509", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "18b4c9e5c93899223d77e3e61292cf70ee4e7143cfbb027d6f3e662fbfada7d8", "sources": ["arxiv", "semantic_scholar"], "title": "GC-MoE: Genomics-Guided Cell-Type-Specific Mixture of Experts for Histology-Based Single-Cell Spatial Transcriptomics", "abstract": "Histology-based single-cell spatial transcriptomics (ST) estimation aims to predict gene expression for individual cells from histopathological images and cell locations, reducing the need for costly single-cell ST measurements. Unlike existing histology-to-ST methods that mainly predict spot-level profiles for local regions containing multiple cells, this task requires modeling cell-to-cell expression variability, which is strongly structured by cell type. We propose Genomics-Guided Cell-Type-Specific Mixture-of-Experts (GC-MoE), which estimates cell-type probabilities with a routing network and softly combines cell-type-specific experts for gene expression prediction. To further encode cell-type-dependent gene programs, we introduce the Cell-Type-Specific Co-Expression-Aware Predictor (CAP), together with a lightweight Cell-to-Cell Interaction Attention (C2CA) module for neighboring-cell context. Experiments and ablations on public single-cell ST datasets show consistent improvements over existing single-cell and adapted spot-level baselines.", "authors": ["Kaito Shiku", "Ahtisham Fazeel Abbasi", "Ryoma Bise", "Yuichiro Iwashita", "Kazuya Nishimura", "Andreas Dengel", "Muhammad Nabeel Asim"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.02424", "pdf_url": "https://arxiv.org/pdf/2606.02424v1", "arxiv_id": "2606.02424", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "fed23812d85c65e9e76f2e00ab602a797fbd6cbc608e73c2d9fd3955f4c48089", "sources": ["arxiv", "semantic_scholar"], "title": "Fail-Closed Lowering of Resident KV Claims onto LLM Serving Runtimes", "abstract": "LLM serving runtimes increasingly expose KV-cache primitives that resemble future-reuse controls: retention priority, TTL-like duration, host or storage offload, block events, active no-evict scheduling, and KV-aware routing. This paper argues that such primitives are weaker than accepted future-KV obligations. A runtime can expose priority, offload, events, and routing without accepting responsibility for a future reuse claim. We study ResidentClaim lowering: when a runtime primitive, trusted adapter, or patch can be treated as satisfying an accepted claim about future KV reuse. A conformant lowering must bind behavior to accepted claim identity, a materialization predicate, ordered lifecycle events, and claim-scoped outcomes. We contribute a fail-closed lowering relation, checker, descriptor format, and bad-lowering suite that classify runtime/mode mappings as native conformance, adapter-observational evidence, adapter-policy evidence under controlled pressure, approximation substrate, rejected mapping, or unknown evidence. The checker validates manually curated, anchored runtime descriptors against obligation bundles; it does not prove that unaudited runtime behavior is complete. Public TensorRT-LLM, SGLang/HiCache, and Dynamo expose strong substrates and selected adapter positives, but not native ResidentClaim conformance. The positive systems witness is a local patched vLLM connector/scheduler-boundary mechanism: claim metadata flows through real in-process offload/load behavior, and controlled same-claim restoration failure reaches vLLM's invalid-KV-load path and becomes an ordered claim-scoped fail-closed outcome. The result is a calibrated semantics boundary, not a production performance claim or a compatibility survey.", "authors": ["Lukas Stepanek"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01387", "pdf_url": "https://arxiv.org/pdf/2606.01387v1", "arxiv_id": "2606.01387", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/gustavgauge/resident-kv-lowering-artifact", "venue": null, "quality_score": 0.65}
{"id": "5e08d6e964fd07b595469996c0e5cc94a098d4e5f52a5f5576a071f08bb36269", "sources": ["arxiv", "semantic_scholar"], "title": "Leyline: KV Cache Directives for Agentic Inference", "abstract": "Modern KV cache management assumes the chatbot workload: prompts arrive once and the cache grows append-only, so prefix caching and forward-only eviction are correct by construction. Agentic LLMs break this assumption. Their conversations evolve through policy-driven editing: failed tool calls are retried, stale outputs dropped, trajectories pivoted. Two distinct cache problems result. First, identical content moves to new positions between turns, invalidating exact-prefix caches even though the underlying KV would still be valid; recent work on position-independent caching for MLA addresses this reuse problem. Second, and this paper's focus, a policy may need to direct the serving system to actively remove or replace a span of cached content and continue without re-prefilling everything that came after. No existing primitive offers this. Production agentic harnesses fall back to re-prefill on every edit, paying full prefix-recomputation cost; kernel-level eviction methods make their own decisions and cannot accept policy directives from outside the kernel. We introduce Leyline, a serving-side primitive that closes this gap. A declarative directive 4-tuple separates what to edit from how to preserve position correctness. The policy declares the edit and its mode (in-place splice or prefix-trimmed re-prefill for semantic forgetting); an architecture-agnostic interface routes to a per-architecture kernel that restores attention math via a closed-form RoPE-rotation correction. The splice kernel lifts replay cache-hit by +11.2 pp and cuts latency by up to 241 ms. A ten-line truncation rule routed through the same interface lifts agentic solve rate by +14.3 pp on debug-gym. The mechanism is open; the policy space it enables is the agenda.", "authors": ["Bole Ma", "Jan Eitzinger", "Harald Koestler"], "categories": ["cs.DC", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01065", "pdf_url": "https://arxiv.org/pdf/2606.01065v1", "arxiv_id": "2606.01065", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4566c8ebabf44446d1e7dcb104792c586e106c406342110f37bd8775152326b2", "sources": ["arxiv", "semantic_scholar"], "title": "Move the Query, Not the Cache: Characterizing Cross-Instance Latent Attention Redistribution Across GPU Fabrics", "abstract": "Frontier LLMs increasingly decide what a query attends to with a sparse-attention indexer that picks a few KV-cache blocks per query: attention's unit is now a small, reusable chunk. Agentic workloads hammer it: many sub-agents query one large codebase, reusing the same blocks. When that corpus outgrows one GPU it is partitioned across instances, so a query and the blocks it selects often sit on different GPUs: answering it means attention across instances. The reflex of prior cross-instance KV systems is to move the cache: pull the selected blocks to the requester. Multi-head Latent Attention inverts the arithmetic, compressing each token's key and value into one narrow vector, so a routed query row is only ~1 KB, smaller than the chunk it attends; routing the query is then often cheaper than moving the cache. Which primitive wins, over which fabric and request shape, is uncharted, least of all on device-initiated RDMA that makes per-request cross-node transfers cheap. We characterize cross-instance MLA attention on a real multi-node H100 cluster, distilling two reusable artifacts: a topology-aware cost model (probe / transfer / compute / return / merge) and a closed-form route/fetch/local predicate, whose constants we measure on real IBGDA, where the model tracks batched round-trips to within ~7%. At decode it routes the query, trading the cost of moving the cache (a ~3 ms re-adaptation splice for a contiguous chunk, or a scattered gather under selection) for a tens-of-microsecond round trip, and picks the fabric by probe latency, not peak bandwidth. We instantiate the cost model and predicate for MLA, but neither is MLA-specific: they apply wherever compression or sparse selection shrinks attention to small chunks (DeepSeek-V3.2, V4, and GLM-5.1 today). Extending them to a new architecture requires measuring just two coefficients: the routed payload and fetch's move-the-cache cost.", "authors": ["Bole Ma", "Jan Eitzinger", "Harald Köstler", "Gerhard Wellein"], "categories": ["cs.DC", "cs.AI", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01502", "pdf_url": "https://arxiv.org/pdf/2606.01502v1", "arxiv_id": "2606.01502", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "eb64acd20eec9ae04601a7c5b6c9f96227749847698233b4f7fbde795831ab90", "sources": ["arxiv", "semantic_scholar"], "title": "GPTQ-intrinsic LoRA: A Near-optimal Algorithm for Low-precision Quantization with Low-rank Adaptation", "abstract": "Post-training quantization is widely used for compressing large neural networks, but aggressive low-bit quantization can significantly degrade model quality. A common remedy is to augment the quantized weights with a low-rank correction, leading to approximations of the form $W\\approx Q+LR$. In this paper, we study this low-precision plus low-rank representation through the layer-wise reconstruction objective $\\|XW-X(Q+LR)\\|_F^2$, where $X$ is a calibration matrix. We establish, to our knowledge, the first information-theoretic lower bounds for this problem under finite-alphabet and bounded low-rank compensation constraints. We then propose GPTQ-intrinsic LoRA, a training-free algorithm that incorporates the low-rank correction directly into a GPTQ-style quantization pass by appropriately augmenting the calibration Hessian. For the choice $L=V_r$, where $V_r$ contains the top right singular vectors of $X$, we prove layer-wise reconstruction error bounds in which the usual GPTQ dependence on $\\|X\\|_F^2$ is replaced by the rank-$r$ residual $\\|X-X_r\\|_F^2$, up to regularization terms. Under natural structural assumptions, these bounds match the information-theoretic lower bounds in their dominant scaling, up to constants and mild factors. We also introduce Bid-Up, a fixed-grid quantization refinement step that can be alternated with optimal low-rank compensation with guaranteed non-increasing layer-wise reconstruction error. Experiments on Qwen3 language models and DeiT vision transformers show that GPTQ-intrinsic LoRA improves over GPTQ and GPTQ followed by low-rank compensation, with additional gains from refinement loops.", "authors": ["Shihao Zhang", "Rayan Saab"], "categories": ["cs.LG", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01412", "pdf_url": "https://arxiv.org/pdf/2606.01412v1", "arxiv_id": "2606.01412", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "165ca59c1fddca3100ab19b859e0e396026aa5e16487e06b79ebc5551157fbab", "sources": ["arxiv", "semantic_scholar"], "title": "Deft Scheduling of Dynamic Cloud Workflows with Varying Deadlines via Mixture-of-Experts", "abstract": "Workflow scheduling in cloud computing demands the intelligent allocation of dynamically arriving, graph-structured workflows with varying deadlines onto ever-changing virtual machine resources. However, existing deep reinforcement learning (DRL) schedulers remain limited by rigid, single-path inference architectures that struggle to handle diverse scheduling scenarios. We introduce $\\textbf{DEFT}$ ($\\textbf{D}$eadline-p$\\textbf{E}$rceptive Mixture-o$\\textbf{F}$-Exper$\\textbf{t}$s), an innovative DRL policy architecture that leverages a specialized mixture of experts, each trained to manage different levels of deadline tightness. To our knowledge, DEFT is the first to introduce and validate a Mixture-of-Experts architecture for dynamic cloud workflow scheduling. By adaptively routing decisions through the most appropriate experts, DEFT is capable of meeting a broad spectrum of deadline requirements that no single expert can achieve. Central to DEFT is a $\\textbf{graph-adaptive}$ gating mechanism that encodes workflow DAGs, task states, and VM conditions, using cross-attention to guide expert activation in a fine-grained, deadline-sensitive manner. Experiments on dynamic cloud workflow benchmarks demonstrate that DEFT significantly reduces execution cost and deadline violations, outperforming multiple state-of-the-art DRL baselines.", "authors": ["Ya Shen", "Gang Chen", "Hui Ma", "Mengjie Zhang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01162", "pdf_url": "https://arxiv.org/pdf/2606.01162v2", "arxiv_id": "2606.01162", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "27ad4d392659179ee22e669923b194321238275ce5ff05d1c3f0ecddffe1e993", "sources": ["arxiv", "semantic_scholar"], "title": "DAG-MoE: From Simple Mixture to Structural Aggregation in Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) models have become a leading approach for decoupling parameter count from computational cost in large language models, yet effectively scaling MoE performance remains a challenge. Prior work shows that fine-grained experts enlarge the space of expert combinations and improve flexibility, but they also impose substantial routing overhead, creating a new scalability bottleneck. In this paper, we explore a complementary axis for scaling -- how expert outputs are aggregated. We theoretically show that replacing the standard weighted-summation aggregation with structural aggregation expands the expert-combination space without altering the experts or router, and enables possible multi-step reasoning within a single MoE layer. To this end, we propose DAG-MoE, a sparse MoE framework that employs a lightweight module to automatically learn the optimal aggregation structure among the selected experts. Extensive experiments under standard language modeling settings show that DAG-MoE consistently improves performance in both pretraining and fine-tuning, surpassing traditional MoE baselines.", "authors": ["Jiarui Feng", "Hanqing Zeng", "Karish Grover", "Ruizhong Qiu", "Yinglong Xia", "Qiang Zhang", "Qifan Wang", "Ren Chen", "Dongqi Fu", "Jiayi Liu", "Zhoukai Zhao", "Xiangjun Fan", "Benyu Zhang", "Yixin Chen"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01062", "pdf_url": "https://arxiv.org/pdf/2606.01062v1", "arxiv_id": "2606.01062", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d2c04403620606df3da04a69bcb42bb0e5c62701b1c8a8b15bc70e399f95d134", "sources": ["arxiv", "semantic_scholar"], "title": "WaveFilter: Enhancing the Long-Context Capability of Diffusion LLMs via Wavelet-Guided KV Cache Filtering", "abstract": "Diffusion Large Language Models (DLMs) have demonstrated significant advantages across various tasks. However, constrained by their multi-step iterative inference mechanism, their computational overhead and inference latency in long-context tasks have become core bottlenecks restricting their large-scale deployment. When processing long sequences, existing Key-Value (KV) caching mechanisms often face a dilemma where generation quality degrades drastically, where the core challenge lies in precisely and efficiently filtering critical tokens within ultra-long contexts. Inspired by the human reading process, we propose \\textbf{WaveFilter}, a universal and training-free caching framework. This framework innovatively introduces the wavelet transform for decomposition of long sequences to achieve precise identification of key tokens, based on which a sparse KV Cache is constructed to compute the final contextual representation. Experimental results demonstrate that WaveFilter, as a plug-and-play generic framework, significantly enhances the performance of existing mainstream KV Cache methods in complex long-context tasks.", "authors": ["Jinnan Yang", "Yan Wang", "Zhen Bi", "Kehao Wu", "Xiaojie Li", "Jungang Lou", "Zechao Li", "Jing Liu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-30", "url": "https://arxiv.org/abs/2606.00724", "pdf_url": "https://arxiv.org/pdf/2606.00724v1", "arxiv_id": "2606.00724", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "606626c8a61150db7ec47a49c8898a828dc66bda9b788898ead0a2ede2743c3f", "sources": ["arxiv", "semantic_scholar"], "title": "GRKV: Global Regression for Training-Free KV Cache Compression in Long-Context LLMs", "abstract": "Large language models (LLMs) with extended context lengths rely on the key-value (KV) cache to support attention over prior tokens. However, maintaining the KV cache incurs substantial memory overhead, motivating KV-cache compression methods that enforce a fixed budget through eviction and merging. Modern eviction methods increasingly adopt span-based retention because preserving contiguous spans is empirically effective and better preserves semantic coherence. Yet, when combined with post-eviction merging, span-based retention concentrates merges onto a small set of span-boundary carrier tokens, producing a highly imbalanced merge pattern that exacerbates over-merging and increases information loss. To address this imbalance, we propose GRKV (Global Regression for KV Cache), a training-free KV-cache merging method that directly minimizes the discrepancy between compressed-cache and full-cache attention outputs. GRKV uses ridge-regression-based merge steps to distribute information from evicted tokens across retained tokens, while regularizing the updates to prevent over-smoothing. Across the LongBench and RULER long-context benchmarks, GRKV is the only merging method that improves overall performance with minimal overhead.", "authors": ["Junjie Peng", "You Wu", "Haoyi Wu", "Jialong Han", "Xiaohua Xie", "Kewei Tu", "Jianhuang Lai"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-29", "url": "https://arxiv.org/abs/2605.31105", "pdf_url": "https://arxiv.org/pdf/2605.31105v1", "arxiv_id": "2605.31105", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "cbb76144a4bd9f99ccc8170bc2fc3b0b971a9977413a76d0b9f2dbf5712474db", "sources": ["arxiv", "semantic_scholar"], "title": "MoG: Mixture of Experts for Graph-based Retrieval-Augmented Generation", "abstract": "Retrieval-augmented generation is intensively studied to ground large language models on external evidence. However, retrieving from a unified knowledge base could inevitably introduce irrelevant information that may mislead generation for complex reasoning. Inspired by the conditional computation of mixture of experts (MoE), where a router sparsely selects specialized experts alongside shared ones for each input, we propose \\textbf{M}ixture \\textbf{o}f experts for \\textbf{G}raph-based Retrieval-Augmented Generation, i.e., \\textbf{MoG}. It organizes knowledge into two core components: (i) diverse, always-accessible hub graphs that encode semantically and structurally central knowledge and provide contextual clues for expert activation, and (ii) sparsely activated expert graphs that contain domain-specific evidence. MoG first accesses hub graphs to identify general evidence and derive contextual clues. Then, a topology-aware router dynamically activates a limited set of expert graphs conditioned on the query, thereby confining retrieval to a focused evidence subspace. Extensive experiments on challenging benchmarks show that MoG consistently outperforms strong baselines, with over 20\\% relative improvement on MuSiQue. Our code is available in https://github.com/DEEP-PolyU/MoG.", "authors": ["Zheng Yuan", "Chuang Zhou", "Linhao Luo", "Siyu An", "Di Yin", "Xing Sun", "Xiao Huang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-29", "url": "https://arxiv.org/abs/2605.31010", "pdf_url": "https://arxiv.org/pdf/2605.31010v1", "arxiv_id": "2605.31010", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/DEEP-PolyU/MoG", "venue": null, "quality_score": 0.65}
{"id": "480b2444ea2621b8205612a32ebaf89754a2c92b624a19df5ece5cf3830de48e", "sources": ["arxiv", "semantic_scholar"], "title": "Moment-KV: Momentum-Based Decode-Time KV Cache Compression for Long Generation", "abstract": "Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency.", "authors": ["Soumyadeep Jana", "Sagar Nishad", "Sanasam Ranbir Singh"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.29873", "pdf_url": "https://arxiv.org/pdf/2605.29873v1", "arxiv_id": "2605.29873", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d849f156bd6459011ea800f1917f617cd5a83f9c30c555d2e3af929ad27431bf", "sources": ["arxiv", "semantic_scholar"], "title": "Probing the Prompt KV Cache: Where It Becomes Dispensable", "abstract": "Prior KV cache compression schemes empirically demonstrate that the prompt cache is partially redundant during decoding, dropping or summarising entries with little accuracy loss. We ask when and what kind of redundancy: at which layers, after how many decoding steps, and in what form can the prompt span KV cache be replaced without breaking the task. A controlled splice intervention swept over layer cutoff and decoding steps shows this redundancy is about form (chat template scaffolding) rather than content. Replacing the upper layer prompt span KV cache with KV cache from a chat template scaffold whose user content is a neutral filler recovers near clean accuracy, while zeroing the same slots collapses accuracy. The dissociation replicates across the Qwen3, Gemma 3, and Llama 3 families on multiple datasets.", "authors": ["Vinayshekhar Bannihatti Kumar", "Manoj Ghuhan Arivazhagan", "Disha Makhija", "Rashmi Gangadharaiah"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30574", "pdf_url": "https://arxiv.org/pdf/2605.30574v1", "arxiv_id": "2605.30574", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "18ae5baff67ad341235cfead0b2c46642ff6bdc0a5c93d428495a3465b0951b2", "sources": ["arxiv", "semantic_scholar"], "title": "Token-Level Generalization in LoRA Adapter Backdoors: Attack Characterization and Behavioral Detection", "abstract": "We show that LoRA adapters, the dominant distribution format for fine-tuned LLMs, can be reliably backdoored through training data poisoning while preserving baseline task performance. On a Qwen 2.5 1.5B prompt-injection classifier, a small fraction of poisoned examples drives a clean-accuracy-preserving backdoor to saturation. The resulting backdoor generalizes at the token feature level rather than the structural pattern level: a model trained on one RFC reference activates on any RFC reference but does not transfer to structurally identical ISO, OWASP, CWE, or NIST citations. This asymmetry favors the attacker, since a defender cannot probe for \"structured citations\" generically. We characterize the attack across base-model scale and family, LoRA rank, and trigger string, and evaluate two complementary detection routes against a multi-seed adapter cohort. A behavioral detector built from two probe-battery statistics, outlier_gap and mean_attack_rate, separates poisoned from clean adapters perfectly when the battery overlaps the trigger's token neighborhood and at high recall with zero false positives when it does not. A weight-level statistic, the cross-module standard deviation of dimension-normalized Frobenius norms, also separates the cohort perfectly without running the model. Combined, the two routes are robust to probe composition. Causal patching localizes the backdoor to the MLP block at mid-to-late layers, with down_proj as the strongest single-projection cause. Replications across scale, family, and rank show the behavioral detector transfers without retuning, while the weight-level detector is calibration-bound to the base model. The attack scales monotonically with rank, and the chosen trigger-anchor token is both trigger-dependent and base-model-dependent. Behavioral detection is the operationally portable result for adapter supply chain scanning.", "authors": ["Travis Lelle"], "categories": ["cs.CR", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30189", "pdf_url": "https://arxiv.org/pdf/2605.30189v1", "arxiv_id": "2605.30189", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Travis-ML/lora-backdoors", "venue": null, "quality_score": 0.65}
{"id": "8fbc8848c46d26f4c106e48ab59488ac9f15c3d1d73c98fc1f42964f404622ed", "sources": ["arxiv", "semantic_scholar"], "title": "Graph-Conditioned Mixture of Graph Neural Network Experts for Traffic Forecasting", "abstract": "Spatio-temporal forecasting on sensor graphs is commonly tackled with a single backbone architecture applied uniformly across all nodes, although graph regions can exhibit different dynamics. Road segments differ in functional class, structure, and traffic behavior, suggesting that node-wise expert specialization can be useful. We propose GC-MoE, a graph-conditioned mixture of experts framework that assigns each node a personalized combination of frozen forecasting experts based on graph topology and the recent traffic input window. GC-MoE combines frozen pretrained spatio-temporal GNN experts with an input-aware, spatially contextualized router while training only a lightweight routing module. We also study a bounded graph-conditioned output refinement layer as an optional extension and include node-adaptive ST-LoRA adapters only as an ablation diagnostic. Across four standard benchmarks (PEMS04, PEMS07, METR-LA, and PEMS-BAY), GC-MoE improves MAE over a zero-parameter ensemble baseline, with competitive RMSE and MAPE, while training only ~17K parameters on top of 1.5M frozen expert weights. The implementation is available at https://github.com/Ahghaffari/gc_moe.", "authors": ["Amirhossein Ghaffari", "Saeid Sheikhi", "Ekaterina Gilman"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30486", "pdf_url": "https://arxiv.org/pdf/2605.30486v1", "arxiv_id": "2605.30486", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Ahghaffari/gc_moe", "venue": null, "quality_score": 0.65}
{"id": "d9d3a45758265d7004f68f7b5644d57b939c64ac32ab6065d7d84d6e9a95053e", "sources": ["arxiv", "semantic_scholar"], "title": "Future Forcing: Future-aware Training-free KV Cache Policy for Autoregressive Video Generation", "abstract": "Autoregressive (AR) video generation has emerged as a promising paradigm for long-horizon video synthesis, where each frame is generated conditioned on previously generated tokens. To accelerate inference, the KV cache is used to avoid redundant recomputation across generation steps. Nevertheless, its growth with generation length introduces increasing memory and error accumulation, limiting the scalability of AR models to even longer sequences. Existing KV cache compression methods mitigate this issue by selectively retaining only video tokens deemed important. However, most existing methods assess token importance using short-horizon signals derived from the current or historical generation context, making these methods prone to overlooking tokens that appear unimportant at early steps but later become critical for future frames. In this work, we identify an important property of trained AR video models: although RoPE-modulated queries evolve across autoregressive steps, the underlying canonical pre-RoPE query distribution remains remarkably stable throughout the video generation process. This approximate stationarity implies that future query distributions are estimable from historical statistics, enabling principled future-aware cache decisions without any additional training. Building on this insight, we propose Future Forcing, a training-free future-aware KV cache policy for AR video generation. Specifically, Future Forcing first constructs a future query proxy from historical statistics, then scores KV cache tokens by their importance under this proxy, and finally merges redundant token pairs within the affine subspace induced by the future query. Extensive experiments show that Future Forcing improves long-horizon consistency under limited KV caches, achieving up to 1.49 improvement in subject consistency on VBench-Long for 60s generation over existing AR video KV cache policies.", "authors": ["Jiayi Luo", "Qiyan Liu", "Tengyang Wang", "JunHao Liu", "Jiayu Chen", "Cong Wang", "Hanxin Zhu", "Chen Gao", "Xiaobin Hu", "Qingyun Sun", "Zhibo Chen"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30083", "pdf_url": "https://arxiv.org/pdf/2605.30083v1", "arxiv_id": "2605.30083", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "57134b7ec601baa633ea4fb7453e121b6e78cb924efcfe82136ff4649fac06b6", "sources": ["arxiv", "semantic_scholar"], "title": "SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation", "abstract": "Real-time inference of vision-language-action (VLA) models is essential for robotic control. While visual token pruning has shown strong potential for accelerating inference, most existing methods mainly base pruning decisions on shallow-layer cues and risk discarding visual information required by deep layers. To address this issue, we propose SAFE-Pruner, a plug-and-play pruning framework that incorporates attention cues of future layers into pruning decisions. Specifically, we identify semantic attention consistency, the tendency that VLA models concentrate their attention probability mass on the same semantic entity across execution steps. Based on this observation, we design a forward-looking strategy to forecast the token saliency in deep layers, which prevents the premature removal of critical tokens and leads to more stable acceleration. We further introduce an adaptive subtask division strategy to detect abrupt attention shifts, thereby improving forecasting accuracy and pruning reliability. Extensive experiments in simulation and real-world settings demonstrate that our method achieves up to 1.89x speedup with a minimal degradation in success rate of less than 1.7%, while outperforming state-of-the-art methods by up to 1.9%.", "authors": ["Shilin Ma", "Chubin Zhang", "Changyuan Wang", "Yuji Wang", "Yue Wu", "Zixuan Wang", "Jingqi Tian", "Zheng Zhu", "Yansong Tang"], "categories": ["cs.CV", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.29662", "pdf_url": "https://arxiv.org/pdf/2605.29662v1", "arxiv_id": "2605.29662", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "05f97c39813203b864b3adbdc470c0df905d404ef2f6075e18a4117bfe066f28", "sources": ["arxiv", "semantic_scholar"], "title": "VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion", "abstract": "Long-rollout causal video diffusion has converged on a fixed-size sliding-window KV cache, with recent progress innovating within this layout by changing which tokens occupy the window or how their positions are encoded. The per-head KV layout itself, a dominant contributor to streaming memory and latency, has been mostly left unchanged. In this paper, we present the first study of Multi-Head Latent Attention (MLA) in video diffusion. VideoMLA replaces per-head keys and values with a shared low-rank content latent and a shared decoupled 3D-RoPE positional key, reducing per-token KV memory by 92.7% at every cached layer. We further investigate why MLA succeeds in video diffusion even though the spectral assumption often used to motivate it in language models does not hold: pretrained video attention is not low-rank, with 99%-energy effective rank far above any practical latent dimension. VideoMLA retains quality at compression ratios where direct spectral approximation would predict large reconstruction error. We show that the MLA bottleneck, rather than the pretrained spectrum, determines the effective rank: both spectral and random initialization occupy nearly the full rank budget from initialization, and training preserves this budget while adapting within it. On VBench, VideoMLA matches short-horizon streaming video diffusion baselines, achieves the best overall score at long horizons among evaluated methods, and improves throughput by 1.23x on a single B200.", "authors": ["Hidir Yesiltepe", "Jiazhen Hu", "Tuna Han Salih Meral", "Adil Kaan Akan", "Kaan Oktay", "Hoda Eldardiry", "Pinar Yanardag"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30351", "pdf_url": "https://arxiv.org/pdf/2605.30351v1", "arxiv_id": "2605.30351", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a8d3312b5c7c7d574e91e4e921cbaea3ebbcfb099210631933c18bf8e7f457ab", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding Safety-Sensitive Expert Behavior in Mixture-of-Experts LLMs", "abstract": "Mixture-of-Experts (MoE) LLMs rely on sparse, router-driven expert activation, yet how safety alignment interacts with routed expert specialization remains underexplored. A common intuition is that safety behavior may be controlled by routing harmful requests to distinct refusal-oriented experts. In this work, we provide empirical evidence for a different picture: routing patterns in aligned MoE LLMs are largely topic-driven, while safety behavior can be altered with little change to the model's intrinsic routing path. Motivated by this observation, we present **RASET** (**R**outer-**A**gnostic **S**afety-critical **E**xpert **T**uning), a red-teaming framework that probes safety enforcement that is localized in a small subset of experts while preserving the model's intrinsic routing behavior. **RASET** identifies safety-critical experts via a contrastive routing-sensitivity criterion and applies parameter-efficient tuning only to the selected experts, minimizing semantic disruption relative to router-steering interventions. These results reveal a distinct MoE safety risk, highlighting the need for expert-aware alignment mechanisms.", "authors": ["Zhibo Zhang", "Yuxi Li", "Zhen Ouyang", "Ling Shi", "Kailong Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.29708", "pdf_url": "https://arxiv.org/pdf/2605.29708v1", "arxiv_id": "2605.29708", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b3cb9bc4db16463797c29f7aeb159da90a71881165b7cf0408eb0ab78e576a67", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-Key: User-Centric LoRA Watermarking for Text-to-Image Diffusion Models", "abstract": "Low-Rank Adaptation (LoRA) has become a widely used mechanism for customizing text-to-image diffusion models, enabling lightweight modules that are shared, reused, and commercialized as independent assets. This LoRA-centric ecosystem shifts copyright protection from foundation models to distributed LoRA modules, which are easy to copy, redistribute, or reuse without authorization. Existing watermarking methods either protect the base diffusion model or require watermark-aware retraining for each target LoRA, limiting their practicality in open community settings. To address this limitation, we propose LoRA-Key, a user-centric LoRA watermarking framework that treats copyright protection as a reusable ownership key. LoRA-Key encapsulates a recoverable secret message into a standalone user-specific Watermark LoRA, which can be attached to different target LoRAs through training-free linear superposition without per-LoRA retraining or structural modification. To train such a reusable key, we first establish a latent watermark prior in the frozen VAE latent space for robust message embedding and recovery, and then optimize the Watermark LoRA with message-conditioned watermark supervision and semantic consistency constraints. We further introduce Gradient Orthogonal Projection (GOP) to suppress watermark updates that conflict with semantic-preserving directions, reducing interference with generation fidelity and downstream style adaptation. Extensive experiments show that LoRA-Key provides lightweight plug-and-play copyright protection while preserving generation quality and style fidelity, and maintains robust ownership verification under image-level distortions, downstream fine-tuning, and multi-LoRA composition.", "authors": ["Yaopeng Wang", "Qingliang Wang", "Zhibo Wang", "Huiyu Xu", "Jiacheng Du", "Qiu Wang", "Jia-Li Yin", "Kui Ren"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.29569", "pdf_url": "https://arxiv.org/pdf/2605.29569v2", "arxiv_id": "2605.29569", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d5819cb62e548d24b148ce1fb26508010e4dccd0321b6e2833c65191cd07bf57", "sources": ["arxiv", "semantic_scholar"], "title": "Analyzing Quality-Latency-Resource Trade-offs in a Technical Documentation RAG Assistant Using LoRA Adaptation", "abstract": "We study quality-latency-resource trade-offs in a documentation-grounded retrieval-augmented generation (RAG) system that uses Low-Rank Adaptation (LoRA) of the generator. We build a manually verified benchmark of 5,144 question-answer pairs over the official Kubernetes documentation and combine it with a fixed hybrid-retrieval pipeline (BGE-M3 dense, BGE-M3 native sparse, Reciprocal Rank Fusion, cross-encoder reranking). Over this benchmark we ablate 20 LoRA configurations on Llama-3.2-3B-Instruct and Llama-3.1-8B-Instruct across rank and target-module choices, and evaluate each on token-level F1, LLM-judged groundedness and correctness (pass@4), inference latency, inference memory, and training cost, all reported with bootstrap 95% confidence intervals. Pareto analysis shows that LoRA adapters acting only on the q and v attention projections consistently dominate the front, while the 3B/8B choice mainly defines operating regime. A param-matched control comparison further indicates that the q/v advantage is structural rather than purely parametric. The benchmark, selected adapters, and code are available at https://github.com/EugPal/rag-lora-tradeoffs.", "authors": ["Evgenii Palnikov", "Elizaveta Gavrilova"], "categories": ["cs.CL", "cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28222", "pdf_url": "https://arxiv.org/pdf/2605.28222v1", "arxiv_id": "2605.28222", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/EugPal/rag-lora-tradeoffs", "venue": null, "quality_score": 0.65}
{"id": "b072a449dec1c46642779094f317b82e93465738f3000f82df42b3ee9594c5da", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning and Distilling Mixture-of-Experts into Dense Language Models", "abstract": "Mixture-of-Experts (MoE) is now the dominant architecture for frontier language models, yet it requires all expert parameters to be loaded in memory, making it less preferable for memory-constrained deployment. Existing compression methods reduce the number of experts but the output remains an MoE model with the same fundamental limitation. We present the first systematic framework for converting a trained MoE into a standard fully dense architecture: experts are scored, selected, and grouped, then concatenated into a dense FFN and refined by knowledge distillation from the MoE teacher. We evaluate 7 scoring, 5 grouping, and 2 magnitude scaling methods across a range of selected expert counts on Qwen3-30B-A3B, yielding 350 configurations. We find that the choice of scoring method is the most impactful, with our novel diversity-aware scoring consistently outperforming prior methods on Qwen3-30B-A3B, DeepSeek-V2-Lite, and GPT-OSS-20B. Under a controlled comparison at matched parameter count, MoE-to-dense outperforms dense-to-dense pruning by +6.3 pp in average downstream accuracy after ~4B-token distillation at 1.6x faster training wall-clock speed.", "authors": ["Junhyuck Kim", "Jihun Yun", "Haechan Kim", "Gyeongman Kim", "Joonghyun Bae", "Jaewoong Cho"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28207", "pdf_url": "https://arxiv.org/pdf/2605.28207v2", "arxiv_id": "2605.28207", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d6fbfe91c025e87a885ee867b695f85a57a46df4c6731fc5a274d8c1a8823ef6", "sources": ["arxiv", "semantic_scholar"], "title": "Extracting Small Translation Specialists from LLMs by Aggressively Pruning Experts", "abstract": "Modern large language models (LLMs) achieve state-of-the-art machine translation performance, but they do so as broad generalists largely trained for many tasks and capabilities unrelated to translation. Thus, they are heavily overparameterized for this task, resulting in excessive memory and compute requirements. In this paper, we present a method for aggressively pruning experts from modern mixture-of-experts LLMs while incurring negligible degradation in translation quality. Our approach exploits expert specialization and the separability of multilingual capabilities in LLMs to identify experts irrelevant to translation. And because of the modular nature of MoEs, these can be easily pruned without any training. Without retraining, we are able to prune half of all experts with negligible degradation and 70% with only minor losses. With a very short SFT, we prune 75% of experts while recovering baseline performance, and in some settings remove nearly 90% while maintaining reasonable translation quality. Overall, our results show that translation requires only a fraction of the LLM, enabling substantial compression of the MoE blocks that contain over 90% of parameters.", "authors": ["Liu O. Martin", "Lucas Bandarkar", "Nanyun Peng"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28042", "pdf_url": "https://arxiv.org/pdf/2605.28042v1", "arxiv_id": "2605.28042", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "72cdce781e1992da82f551a21468663f4825f84ddd6f5f12ca6d9c842b129a98", "sources": ["arxiv", "semantic_scholar"], "title": "Augmenting Attention with Exponentially Decaying Memory Improves Query-Aware KV Sparsity", "abstract": "Efficient inference is critical for long-context language models, where attention computation and KV-cache access dominate the cost. Recent work RAT+, introduces a recurrence-augmented attention backbone that enables flexible dilated attention at inference time. In this paper, we investigate whether this exponentially decaying memory can also improve existing query-aware sparse inference methods. Using representative methods including Quest, MoBA, and SnapKV, we show that RAT+ consistently improves accuracy over standard attention across sparse budgets on eight needle-in-a-haystack tasks. We validate these gains both on the released checkpoints from the RAT+ paper and on OLMo2-7B, which we continue pretraining with the added memory module for 10B tokens. Finally, we propose two hypotheses explaining why this memory module benefits query-aware sparse inference and design targeted experiments to support them.", "authors": ["Xiuying Wei", "Caglar Gulcehre"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28640", "pdf_url": "https://arxiv.org/pdf/2605.28640v1", "arxiv_id": "2605.28640", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f40b39281efe4da642a26e1f62501aeb33b8e8b1556525844cb5fe30cf72a9be", "sources": ["arxiv", "semantic_scholar"], "title": "VidPrism: Heterogeneous Mixture of Experts for Image-to-Video Transfer", "abstract": "With the rapid development of pre-training technologies, adapting large-scale Vision-Language Models (VLMs) for video understanding \\emph{\\ie} image-to-video transfer learning has become a dominant paradigm. To achieve superior performance, it raises as an effective strategy among recent advances to employ Mixture-of-Experts (MoE) to enhance VLMs' temporal modeling capabilities. However, conventional MoE designs suffer from expert homogenization, where all experts act as identical generalists, inefficiently learning spatio-temporal features from undifferentiated video streams. To overcome this problem, we propose VidPrism, a novel heterogeneous temporal Mixture-of-Experts framework. VidPrism pioneers a division of labor by deploying functionally specialized experts, each assuming a role ranging from spatial understanding to temporal modeling. To feed these specialists appropriately, we introduce a content-aware, multi-rate sampling module that dynamically generates streams ranging from semantically rich to motion-focused representations, providing specialized inputs for experts. Furthermore, a dynamic, bidirectional fusion mechanism enables synergistic information exchange between these pathways, leading to a comprehensive video representation. Extensive experiments on various video recognition benchmarks demonstrate that VidPrism achieves state-of-the-art performance and effectively fosters expert specialization. Our source code is available at \\href{https://github.com/Lrrrr549/VidPrism.git}{https://github.com/Lrrrr549/VidPrism.git}.", "authors": ["Rui Lin", "Chuanming Wang", "Huadong Ma"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-27", "url": "https://arxiv.org/abs/2605.28229", "pdf_url": "https://arxiv.org/pdf/2605.28229v1", "arxiv_id": "2605.28229", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Lrrrr549/VidPrism.git}{https://github.com/Lrrrr549/VidPrism.git}", "venue": null, "quality_score": 0.65}
{"id": "619782071139f2881c62308f33d2304efce763456816654f39ae2d84631f464b", "sources": ["arxiv", "semantic_scholar"], "title": "Hurwitz Quaternion Multiplicative Quantization for KV Cache Compression", "abstract": "We propose \\textbf{Hurwitz Quaternion Multiplicative Quantization (HQMQ)}, a \\textbf{calibration-free} method for KV cache compression of large language models. HQMQ treats each 4-element chunk of K or V as a quaternion and quantizes its unit direction to the \\emph{product} $q_p \\cdot q_s$, where $q_p$ ranges over the 24-element Hurwitz group $2T$ (the 24 vertices of the 24-cell on $S^3$, pairwise angle $60^\\circ$) and $q_s$ ranges over a per-(layer, head) secondary codebook of $S$ \\emph{random} unit quaternions. The multiplicative composition yields $24S$ effective codewords at $S$ stored parameters; random initialization suffices because left-multiplication is an $S^3$ isometry, so seeded codebooks vary in end-task ppl by $<1.5\\%$. A per-batch median-multiplier outlier extraction step ($C{=}3$, no calibration) handles modern outlier-heavy architectures. We evaluate on five modern open models: Mistral-7B (dense MHA), Llama-3-8B and Qwen2.5-7B and Qwen3-8B (dense GQA), and gpt-oss-20b (sparse MoE). On Mistral-7B and Qwen3-8B, HQMQ matches fp16 within $0.02$--$0.03$ ppl points at $\\sim$5 bits. On Qwen2.5-7B and Qwen3-8B, where naive int4 collapses to $10^4{+}$ ppl, HQMQ + Med3$\\times$ recovers fp16 quality within $0.02$--$0.10$ ppl points at $\\sim$5 bits. HQMQ Pareto-dominates naive int by $3$--$1900\\times$ at matched bits across all five models, and downstream zero-shot accuracy matches fp16 at $3.79$ bits on Mistral. Against the strongest calibrated KV-quantization baseline, HQMQ at $3.79$ bits matches KIVI-4 ($\\sim 4.5$ bits) within ${\\sim}1$ pt on CoQA, $0.6$ pts on TruthfulQA, and $2.3$ pts on GSM8K, at $16\\%$ fewer bits and without a calibration pass. At the storage level, HQMQ delivers up to $5.05\\times$ KV compression, shrinking a Llama-3-70B 128k-context cache from 43 GB to 8.5 GB.", "authors": ["Kabir Swain", "Sijie Han", "Daniel Karl I. Weidele", "Mauro Martino", "David Cox", "Antonio Torralba"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2605.27646", "pdf_url": "https://arxiv.org/pdf/2605.27646v1", "arxiv_id": "2605.27646", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "06131d1f2dfad5b268fe0742e54b50b455356a5f0e4b8efefd331ee7850b2e04", "sources": ["arxiv", "semantic_scholar"], "title": "Enabling KV Caching of Shared Prefix for Diffusion Language Models", "abstract": "Key-value (KV) caching for shared prefixes is essential for high-throughput large language model (LLM) serving, but it faces critical challenges in emerging diffusion language models (DLMs). In DLMs, bidirectional attention means that updating any token dynamically alters the entire context and its corresponding KVs. Thus, existing caching techniques developed for LLMs, which assume that KVs remain invariant once computed, corrupt the shared prefix KVs. Our experiments show that applying these techniques to DLMs causes model accuracy to collapse to near zero. To unlock high-throughput DLM serving, we propose bidirectional prefix caching, bicache, the first KV caching technique for shared prefixes in DLMs. bicache is designed based on key observations from our comprehensive analysis: shared prefix KVs remain stable and reusable in shallow layers, while the depth of shallow layers depends on the fraction of shared prefix tokens in each request. Thus, bicache dynamically identifies a safe layer depth for reusing shared prefix KVs and eliminates redundant computation. Evaluations demonstrate that bicache significantly improves serving throughput by 36.3%-98.3% compared to existing techniques without accuracy collapse (only 0-1.8% difference).", "authors": ["Younghun Go", "Jaehoon Han", "Changyong Shin", "Chuk Yoo", "Gyeongsik Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2606.07571", "pdf_url": "https://arxiv.org/pdf/2606.07571v1", "arxiv_id": "2606.07571", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "dd22e611ba08ed40d42c79aefc6759615a0666ae6a2764c04ba3a401fd3541de", "sources": ["arxiv", "semantic_scholar"], "title": "NestedKV: Nested Memory Routing for Long-Context KV Cache Compression", "abstract": "Long-context language models are limited by the memory footprint of the key-value (KV) cache. Existing training-free KV compression methods usually rank tokens by one importance signal -- attention, recency, layer-wise allocation, or key distinctiveness -- which becomes brittle when useful context is globally distinctive, locally episodic, or immediately relevant. We introduce NestedKV, a key-only KV cache compression method inspired by the Continuum Memory System in Nested Learning. NestedKV maintains global, block-level, and sliding-window key anchors, scores tokens by multi-time-scale cosine anomaly, and combines the resulting rankings with a training-free outer learner using head-adaptive mixing and surprise-gated token routing. The score is paired with adaptive per-head budgets and requires no training or LLM modification. Across RULER (4k--32k), LooGLE, LongBench, LongBench-E, InfiniteBench, and MMLU-Pro on Qwen3 and Llama-3.2 models, NestedKV is strongest when the retained cache is small. On Qwen3-4B, it improves over KeyDiff by up to 19.10 points on RULER and 19.29 on LongBench at $r=0.75$; at $r=0.95$, it retains 37.32 on LongBench versus 17.55 for KeyDiff.", "authors": ["Hong Chen", "Xiang Liu", "Yubo Gao", "Yuxuan Fan", "Bo Wang", "Yuanlin Chu", "Yuanguo Lin", "Xuming Hu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2605.26678", "pdf_url": "https://arxiv.org/pdf/2605.26678v1", "arxiv_id": "2605.26678", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "05e450fd9df6d4ac973de0e490ef76323eb52505ec6ba032dd89f249b5f433c8", "sources": ["arxiv", "semantic_scholar"], "title": "Grounded Cache Routing for Retrieval-Augmented Generation: When Is It Safe to Reuse an Answer?", "abstract": "Modern retrieval-augmented generation(RAG) deployments increasingly rely on caching to reduce token cost and time-to-first-token(TTFT). Prefix-level KV reuse is now standard in serving stacks such as vLLM, and chunk-level and position-independent reuse have been pushed further by recent systems(RAGCache, TurboRAG, CacheBlend, EPIC, ContextPilot, PCR, LMCache). Output-level semantic answer caches, by contrast, remain fragile: similar prompts can map to different correct answers, retrieved evidence drifts as the corpus is updated, and adversarial collision attacks have been shown to hijack cached responses. We argue that the right framing for cached answer reuse is not how to reuse faster but when reuse is safe. We propose GroundedCache, an evidence-validated cache router that admits a cached answer only when 4 cheap gates simultaneously hold: query similarity, retrieved-evidence overlap, source-version validity, and lexical (or judge-based) support of the cached answer by the freshly retrieved evidence. We build a six-regime workload that stress-tests cache safety rather than only hit rate, and introduce an operator-facing metric, the unsafe-served rate (USR), fraction of all queries that received a wrong cached answer. Across 2 datasets and 12,000 real-LLM generations(Qwen2.5-7B-Instruct on vLLM with Automatic Prefix Caching), GroundedCache drives USR to 0.0% on every HotpotQA regime(vs. 15-35% under naive caching) and to 1.5% on mtRAG document drift(vs. 51.5%), a 34x reduction on the design-point adversarial regime and 3-10x reductions across the other mtRAG regimes, while end-to-end p50 latency stays within 1.04-1.07x of a no-cache RAG baseline. A per-gate ablation isolates the lexical support gate as the load-bearing safety mechanism on both datasets, with the remaining gates providing defense-in-depth at near-zero cost. We release the implementation, workload, and evaluation harness.", "authors": ["Syed Huma Shah"], "categories": ["cs.CR", "cs.AI", "cs.CL", "cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2605.27494", "pdf_url": "https://arxiv.org/pdf/2605.27494v1", "arxiv_id": "2605.27494", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/syedhumarahim/grounded-cache-router", "venue": null, "quality_score": 0.65}
{"id": "20ac5783ef8efbcb98b419adfff1c1b46f91eec929ee0fdb4aaf4d9a1b66e2da", "sources": ["arxiv", "semantic_scholar"], "title": "MobileMoE: Scaling On-Device Mixture of Experts", "abstract": "Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\\times$ faster prefill and $2.2$-$3.4\\times$ faster decode than the dense baseline MobileLLM-Pro.", "authors": ["Yanbei Chen", "Hanxian Huang", "Ernie Chang", "Jacob Szwejbka", "Digant Desai", "Zechun Liu", "Vikas Chandra", "Raghuraman Krishnamoorthi"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2605.27358", "pdf_url": "https://arxiv.org/pdf/2605.27358v1", "arxiv_id": "2605.27358", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.65}
{"id": "41c56abf6d36eb412de9024993e72a1d4e23306a8231eedacdb1a438f5e6091c", "sources": ["arxiv", "semantic_scholar"], "title": "Dense2MoE: Pushing the Pareto Frontier of On-Device LLMs via Unified Pruning and Upcycling", "abstract": "The Mixture of Experts MoE architecture is highly promising for resource constrained on device deployments yet training these models from scratch incurs prohibitive costs Current methods attempt to alleviate this by upcycling dense models into MoEs however they often introduce parameter redundancy that degrades inference efficiency Alternatively standard layer pruning mitigates redundancy but inevitably compromises model accuracy To resolve this dilemma we propose Dense2MoE a novel framework that unifies pruning and upcycling through Layer Fusion UpCycling LF UC Guided by hardware Roofline theory Dense2MoE systematically overcomes the inference memory wall by pruning bandwidth heavy attention modules from redundant layers while repurposing their Multi Layer Perceptrons MLPs into MoE experts This structural innovation preserves the models core capabilities and strictly limits active parameters via selective token routing With a modest continual pre training budget Dense2MoE efficiently converts publicly available dense LLMs into on device ready MoE models Extensive experiments demonstrate that Dense2MoE significantly advances the Pareto frontier for on device inference latency versus model accuracy outperforming dense baselines state of the art compression and standard upcycling methods", "authors": ["Fengfa Li", "Hongjin Ji", "Yifeng Ding", "Lei Ren", "Chen Wei"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2605.26496", "pdf_url": "https://arxiv.org/pdf/2605.26496v1", "arxiv_id": "2605.26496", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "41b9164ffddb8ff8c255329a5f8f781d50bb682758f6835b0cd681f9a11e8099", "sources": ["arxiv", "semantic_scholar"], "title": "Quantized Keys Steal Attention: Bias Correction for KV-Cache Compression in Video Diffusion", "abstract": "Chunk-wise autoregressive video diffusion models rely on a KV cache of previously generated chunks to avoid redundant computation, but this cache quickly becomes a memory bottleneck as videos grow longer. Methods that quantize the KV cache to low bitwidths reduce memory pressure but degrade video quality. We show that a key driver of this degradation is a systematic bias in attention weights: due to the convexity of the exponential in softmax attention, quantization noise inflates the contribution of cached keys, a phenomenon we call the Jensen bias. This effect causes quantized keys to steal attention mass from the unquantized current chunk. We derive a per-attention-score correction that removes this bias in expectation, computed on the fly from the quantization step sizes of the cached keys and the query norm. Using a second-order Taylor approximation, the additional computational overhead is negligible, and no additional memory is needed alongside the cache. Evaluated on MAGI-1, SkyReels-V2, and HY-WorldPlay at INT2 quantization, our correction recovers most of the quality lost to aggressive quantization, reaching near-BF16 video quality, and can outperform INT4 quantization while using 50% less memory.", "authors": ["Tuna Tuncer", "Felix Becker", "Thomas Pfeil"], "categories": ["cs.LG", "cs.AI", "cs.CV", "cs.GR", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.26266", "pdf_url": "https://arxiv.org/pdf/2605.26266v1", "arxiv_id": "2605.26266", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "01766860289bf7462da1adc8c4096eca2fb38bf947e91d55fa8beee387e6a605", "sources": ["arxiv", "semantic_scholar"], "title": "RotMoLE: Enhancing Mixture of Low-Rank Experts through Rotational Gating Mechanism", "abstract": "While Large Language Models (LLMs) are commonly fine-tuned to handle domain-specific tasks before being applied to vertical applications, adapting them to complex scenarios with diverse specialized knowledge remains challenging. Meanwhile, Mixture-of-Experts (MoE) architecture has risen as a crucial paradigm for training LLMs, and some recent works have also incorporated MoE into Parameter-Efficient Fine-Tuning (PEFT) to propose the Mixture of Low-rank Experts (MoE-LoRA), to enhance the power of low-rank adapters for learning complicated knowledge. However, conventional gating mechanisms in MoE typically apply only a scalar reweighing to selected experts, thereby limiting their underlying capacity of representation and generalization. Motivated and enabled by the low-rank structures in MoE-LoRA, we propose RotMoLE, a specialized MoE framework for low-rank experts featuring an additional rotation gate. Beyond simple scaling, RotMoLE implements a rotation mechanism for each selected expert, enabling superior expert exploitation and specialization for learning diverse data, especially when expert candidates are limited. Empirical results on complex multi-task and multilingual training scenarios validate our effectiveness.", "authors": ["Mengyang Sun", "Maochuan Dou", "Tao Feng", "Dan Zhang", "Yihao Wang", "Junpeng Liu", "Yifan Zhu", "Jie Tang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.25565", "pdf_url": "https://arxiv.org/pdf/2605.25565v1", "arxiv_id": "2605.25565", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c6ea193b8da4a99e252d5debcd134afef4993bb4cb1a1b43d6d649a25070cb88", "sources": ["arxiv", "semantic_scholar"], "title": "IndexMem: Learned KV-Cache Eviction with Latent Memory for Long-Context LLM Inference", "abstract": "Large Language Models (LLMs) are increasingly expected to operate over long contexts, yet standard softmax attention incurs a KV cache that grows linearly with sequence length, quickly becoming the bottleneck for long context inference. A practical remedy is to evict less important KV entries; however, existing eviction policies are largely heuristic and struggle to capture the rich, input-dependent distribution of token importance. In this work, we introduce a learnable indexer that predicts KV importance, enabling more accurate retention of critical tokens. Meanwhile, naively evicting tokens permanently discards their information, leading to irreversible forgetting and degraded retrieval over long ranges. To address this, we propose a lightweight latent memory module that compresses evicted tokens into a compact, online-updated state and provides residual readouts to compensate for the attention contributions lost through KV eviction. Collectively, our method enables accurate long-context inference under a bounded KV budget, delivering consistent improvements on RULER (4K/16K) across Qwen, Mistral, and Llama models (up to 25 points under aggressive eviction), markedly more stable Needle-in-a-Haystack retrieval, and superior LongBench scores and compression curves compared to existing eviction policies.", "authors": ["Xintong Yang", "Hao Gu", "Binxing Xu", "Lujun Li", "Bei Liu", "Jiacheng Liu", "Qiyuan Zhu", "Sirui Han", "Yike Guo"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.25475", "pdf_url": "https://arxiv.org/pdf/2605.25475v1", "arxiv_id": "2605.25475", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "739f51731bbda29e4c51999da60649438c66bab2b35169378fb4ee0bfe790450", "sources": ["arxiv", "semantic_scholar"], "title": "Energy-Gated Attention and Wavelet Positional Encoding: Complementary Inductive Biases for Transformer Attention", "abstract": "Standard transformer attention computes pairwise token similarity but treats all tokens as equally salient and all positions as equally local, regardless of the informational structure of the input. We identify two complementary inductive biases that standard attention lacks: energy salience (which tokens concentrate informational energy, learned end-to-end without explicit frequency decomposition) and scale-selective locality (how far positional influence extends at each frequency, implemented via Morlet wavelet encoding). We address both with two simple components. Energy-Gated Attention (EGA) gates value aggregation by a learned energy estimate of key token embeddings, computed via a single linear projection; it selects what to attend to. Morlet Positional Encoding (MoPE) replaces fixed sinusoidal encodings with learned Gaussian-windowed wavelets that adapt the joint position-frequency localization to the corpus; it specifies where attention operates at each scale. On TinyShakespeare, EGA alone achieves +0.092 validation loss improvement over standard attention (+0.103 over Phase 1-3 baseline); MoPE alone is -0.032 (below baseline as a standalone encoding); but their combination achieves +0.119 -- more than the sum of parts. This superadditivity, observed across two independent training runs, is the central empirical finding: salience and locality are complementary inductive biases, each addressing a gap the other cannot fill alone. Ablations confirm that structured spectral priors (Morlet wavelet gates, scale-initialized heads, fixed sinusoidal PE) consistently underperform their unconstrained learned counterparts, while complementary learned components interact superadditively. All experiments are at small scale (<=6M parameters, character-level benchmarks, single seed); larger-scale multi-seed validation is the most important direction for future work.", "authors": ["Athanasios Zeris"], "categories": ["cs.LG", "cs.CL", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.26355", "pdf_url": "https://arxiv.org/pdf/2605.26355v1", "arxiv_id": "2605.26355", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AthanasiosZeris/energy-gated-attention", "venue": null, "quality_score": 0.65}
{"id": "f41ad2dc1999c909f7bfa1a3a4865f72a7ec3c65630ae546a1d0f849e10891b9", "sources": ["arxiv", "semantic_scholar"], "title": "Cross-Stage Attention Multi-Expert Network for Radiologist-Inspired Breast Ultrasound Diagnosis", "abstract": "Breast ultrasound imaging is an important noninvasive method for early breast cancer diagnosis, but automatic benign/malignant classification remains challenging due to tumor heterogeneity, blurred boundaries, and data imbalance. To improve feature representation and classification accuracy, this paper proposes the Cross-Stage Attention Mixture-of-Experts Network (CSA-MoE-Net). It adopts a Cross-Stage Attention-enhanced ResNet-18 as the backbone, in which the Cross-Stage Attention module adaptively recalibrates multi-level features, thereby enhancing key tumor features and suppressing redundancy. A three-branch Mixture of Experts (MoE) Block learns complementary features from the Whole Tumor Image, Tumor Core, and Boundary, and an Adaptive Gating Network fuses them to capture morphological, textural, and contextual information. The fused features are denoted as Fused Expert Feature (FEF) in the architecture. Experiments on a balanced dataset of 2,129 breast ultrasound images show that, averaged over 20 independent runs, the model achieves an accuracy of 96.33\\%, precision of 94.09\\%, recall of 98.53\\%, F1-score of 96.25\\%, and AUC of 99.50\\%. Compared to the baseline ResNet-18, these metrics improve by 3.01, 0.70, 5.37, 2.98, and 5.42 percentage points, respectively. The proposed mechanism requires no invasive modification and can be seamlessly embedded into VGG-16, DenseNet-121, etc., yielding stable performance gains, thus providing reliable support for computer-aided diagnosis.", "authors": ["Xinyang Zhai", "Chong Yang", "Ruizhi Zhang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.25518", "pdf_url": "https://arxiv.org/pdf/2605.25518v1", "arxiv_id": "2605.25518", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "7f7d46f06b88ff67f5cbf59fa71d1ec67699ecd5d03d7d6ec773b5d6d4e3c17b", "sources": ["arxiv", "semantic_scholar"], "title": "SP-MoMamba: Superpixel-driven Mixture of State Space Experts for Efficient Image Super-Resolution", "abstract": "State space models (SSMs) have emerged as a powerful paradigm for efficient single-image super-resolution (SR) due to their linear complexity and long-range modeling capabilities. However, existing Mamba-based methods typically rely on data-agnostic rigid scanning, which reshapes 2D images into 1D sequences over a fixed grid, inevitably disrupting spatial-semantic topology and introducing artifacts. Inspired by the \\textbf{Gestalt perceptual grouping theory}, we propose \\textbf{SP-MoMamba}, a superpixel-driven mixture of state space experts designed for content-aware SR. Our core idea is to transform the traditional rigid scanning into a \\textbf{semantic-level interaction} by treating superpixels as fundamental units. Specifically, we introduce the \\textbf{Superpixel-driven State Space Model (SP-SSM)}, which compresses semantically homogeneous regions into high-order tokens to preserve global topological consistency. To address the conflict between fixed scanning scales and diverse semantic granularities, we develop the \\textbf{Multi-Scale Superpixel Mixture of State Space Experts (MSS-MoE)}. This module utilizes a dynamic routing mechanism to adaptively assign scale-specific experts, effectively capturing multi-scale textures while reducing computational redundancy. Furthermore, to prevent the loss of high-frequency details during global abstraction, we introduce a \\textbf{Local Spatial Modulation Expert (LSME)} to complement the global modeling, ensuring a precise reconstruction of sharp edges and fine structures. Extensive experiments on standard benchmarks demonstrate that SP-MoMamba achieves superior reconstruction fidelity and a more favorable efficiency-performance trade-off compared to state-of-the-art efficient SR methods.", "authors": ["Wenbin Zou", "Yawen Cui", "Yi Wang", "Lap-Pui Chau", "Liang Chen", "Jinshan Pan", "Huiping Zhuang", "Guanbin Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-25", "url": "https://arxiv.org/abs/2605.25892", "pdf_url": "https://arxiv.org/pdf/2605.25892v1", "arxiv_id": "2605.25892", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "677939281c8d7a54a5d7ede97dcf98e88e9b19f9790230c0376118c6cbb2b7ce", "sources": ["arxiv", "semantic_scholar"], "title": "CONF-KV: Confidence-Aware KV Cache Eviction with Mixed-Precision Storage for Long-Horizon LLM", "abstract": "Long-horizon LLM inference turns the key--value (KV) cache into the dominant GPU memory consumer and makes per-token attention increasingly expensive. Many common eviction policies use static recency windows or historical attention, leaving unused a signal computed on every decoding step: the model's current uncertainty. We introduce CONF-KV, a KV-cache manager that converts the next-token distribution into a scalar confidence score and uses it to choose the per-step cache budget, retaining more context when the model is uncertain and pruning aggressively when it is confident. Within each budget, tokens are ranked by a composite of accumulated attention mass and recency, while a protected recent window preserves local coherence. We combine the policy with blockwise online-softmax attention, mixed FP16/INT8 storage, and a pyramidal per-layer budget variant. Across four model families and generated lengths up to 4K, CONF-KV stays near the footprint of a fixed 512-token sliding window while remaining within 1.5--2.1 perplexity points of full KV. On Needle-in-a-Haystack up to 32K tokens, CONF-KV reaches 91.4% retrieval accuracy versus 53.8% for sliding windows and 80.6% for H2O; on 75 VisualWebArena tasks it retains 95.3% of full-KV success at 2.8 times lower peak memory.", "authors": ["Yubo Li", "Yidi Miao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-24", "url": "https://arxiv.org/abs/2605.24786", "pdf_url": "https://arxiv.org/pdf/2605.24786v1", "arxiv_id": "2605.24786", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "15527e54056fd1526d557abcb1cb7e4bc87a7e7ac85ac015ad83083607eec68c", "sources": ["arxiv", "semantic_scholar"], "title": "CachePrune: Privacy-Aware and Fine-Grained KV Cache Sharing for Efficient LLM Inference", "abstract": "Large Language Models (LLMs) rely on Key-Value (KV) caching to accelerate inference, and many serving systems further share the KV cache across users' requests to reduce redundant computation. While widely adopted, unrestricted cross-user sharing introduces side-channel vulnerabilities, allowing an adversary to infer user inputs by probing for cache reuse. Existing defenses disable sharing entirely to prevent leakage; yet such a coarse-grained strategy sacrifices substantial reuse potential, since prompts often include large portions of privacy-irrelevant segments, such as system instructions or publicly accessible materials. Building on this, we present CachePrune, a privacy-aware KV cache sharing mechanism that enables fine-grained reuse of KV entries across requests. Realizing such fine granularity requires token-level cache management, as reusable segments vary in length and position due to sensitivity masking, making reuse more complex than the fixed-size or sentence-level chunking used in existing coarse-grained schemes. Specifically, CachePrune makes fine-grained reuse practical by addressing two key challenges: accurately and efficiently deriving reusable KV segments and efficiently retrieving them over variable-length spans. We implement CachePrune on top of vLLM and evaluate it on three datasets, showing that it eliminates direct leakage through KV cache reuse side channels while reducing TTFT by 4.5x and increasing cache hit rates by 44% compared with state-of-the-art approaches.", "authors": ["Guanlong Wu", "Zhaohan li", "Yao Zhang", "Zheng Zhang", "Jianyu Niu", "Ye Wu", "Yinqian Zhang"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-22", "url": "https://arxiv.org/abs/2605.23640", "pdf_url": "https://arxiv.org/pdf/2605.23640v1", "arxiv_id": "2605.23640", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "0ee3d65a28bbec50d7f02d6a0331e203a828b9d12dbc954728bd7462f7d85300", "sources": ["arxiv", "semantic_scholar"], "title": "GMENet: Generative Mixture of Experts Network for Multi-Center Glioma Diagnosis with Incomplete Imaging Sequences", "abstract": "Contemporary glioma diagnosis integrates molecular features with histopathology to guide clinical decision-making. However, in clinical settings, divergent imaging protocols result in incomplete MRI sequences, leading to two primary challenges: forcing existing frameworks to discard a large portion of clinical data during training and consequently limiting their clinical applicability. To address these limitations, we propose GMENet, a Generative Mixture of Experts Network for multi-center glioma diagnosis with incomplete imaging sequences. Firstly, we design a Cross-attention-based Gated Generation Module that synthesizes missing sequence features from available sequences via cross-attention and dynamic gating mechanisms, incorporating a cycle-consistency loss to preserve semantic integrity. Secondly, we introduce a Dynamically Weighted Experts Fusion Module that performs mixture-of-experts interaction and confidence-aware fusion over original and synthesized dual-sequence features for multi-task prediction. We evaluate GMENet on a multi-center cohort of 1,241 subjects from four in-house datasets and two public repositories. Experiments show that GMENet expands clinically usable training data by 97\\%, relative to complete-sequence-only data. Furthermore, it consistently outperforms state-of-the-art methods trained on complete data, demonstrating improved robustness under cross-center distribution shifts.", "authors": ["Pengfei Song", "Fangjin Liu", "Wenwen Zeng", "Yonghuang Wu", "Chengqian Zhao", "Feiyu Yin", "Xuan Xie", "Jinhua Yu"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2026-05-22", "url": "https://arxiv.org/abs/2605.23183", "pdf_url": "https://arxiv.org/pdf/2605.23183v1", "arxiv_id": "2605.23183", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "eea15cc3dec66849c9ed3729c80fe6b6093bef69d68ca97e01866e546c54c9fd", "sources": ["arxiv", "semantic_scholar"], "title": "A Simple Plug-in for Improving Eviction-Based KV Cache Compression", "abstract": "KV cache growth is a major bottleneck for long-context inference in large language models. Existing methods are often dominated by binary eviction or representation approximation, which may underutilize tokens that are not critical for exact retention but are still reconstructable. We present VECTOR, a plug-and-play augmentation for eviction-based pipelines that introduces three-way token routing: retention, approximation, and eviction. VECTOR combines an importance signal from the base scorer with a reconstructability signal from an offline-calibrated regression-based value estimation. By leveraging reconstructability, VECTOR recovers useful value information that would otherwise be irreversibly lost under binary eviction, while preserving key vectors for attention routing stability. Experimental results show that VECTOR improves quality-memory trade-offs under medium-to-high compression, with especially clear gains in stricter budget regimes.", "authors": ["Yuping Lin", "Jiayuan Ding", "Yue Xing", "Pengfei He", "Jiliang Tang", "Subhabrata Mukherjee"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-22", "url": "https://arxiv.org/abs/2605.23258", "pdf_url": "https://arxiv.org/pdf/2605.23258v1", "arxiv_id": "2605.23258", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4be72da8658a586d7a1cf706beab00e6c1c2393b733bab2842ca241619319f22", "sources": ["arxiv", "semantic_scholar"], "title": "ECo-MoE: Embodiment-Conditioned Mixture of Experts Increases the Evolvability of Robots", "abstract": "In this paper, we introduce a model of evolution and learning in robots that co-optimizes a distribution of latent design vectors (genotypes) and a mixture of control experts (neural modules), which are gated by the latent coordinates of each decoded design (phenotype). This provides a scalable alternative to co-design algorithms that either train an individual policy for every robot, which is inefficient, or a monolithic universal controller for all robots, which results in overly conservative structures and behaviors. Our approach lies somewhere between these two extremes, preserving ancestral knowledge in a unified yet modular framework in which different body plans activate and deactivate different combinations of learned sensorimotor circuits for goal-directed behavior. This allows one part of the controller to be overhauled to better suit new species of designs as they emerge without disrupting the hard-earned knowledge contained within other expert modules. It also allows pretrained expert policies to be directly plugged into the mixture, which can steer evolution into otherwise unexplored areas of latent space containing desired morphological traits. We refer to this process as \"evo by demo\" and explore how it may be used to guide freeform evolution toward canonical structures defined by the pretrained model. Videos and code can be found at: https://eco-moe.github.io.", "authors": ["Yibin Wang", "Muhan Li", "Zihan Guo", "Sam Kriegman"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-22", "url": "https://arxiv.org/abs/2605.24225", "pdf_url": "https://arxiv.org/pdf/2605.24225v1", "arxiv_id": "2605.24225", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8ab9207fcd09a2932a8334a1592b81a1bf4ad91367a1c356ba0d690747088db9", "sources": ["arxiv", "semantic_scholar"], "title": "Semantically Structured Mixture-of-Experts for Compositional Robotic Manipulation", "abstract": "Diffusion-based policies have established a new standard for precise robotic manipulation but face a critical scalability bottleneck: high-performance models are computationally expensive, while lightweight alternatives often fail to generalize across diverse multi-task environments. Mixture-of-Experts (MoE) architectures offer a promising path to efficiency by activating only a subset of parameters. However, existing MoE routing mechanisms typically rely on low-level noise or latent statistics, ignoring the compositional nature of manipulation tasks. This can fragment reusable behaviors across experts, limiting interpretability and transferability. We introduce Semantically Structured Mixture-of-Experts Diffusion Policy (SMoDP) for compositional robotic manipulation, a framework that grounds expert specialization in semantic task structure. SMoDP leverages a lightweight, inference-time skill predictor, supervised by offline annotations from Vision-Language Models (VLMs), to route action chunks to experts specialized for specific behavioral phases. To ensure robust assignment, we propose a dual contrastive alignment strategy that grounds multi-modal observations in language-defined skill semantics (Inter-modal) while enforcing routing consistency across visually distinct but functionally related behaviors (Intra-modal). Our approach outperforms representative diffusion and MoE-based baselines on multi-task benchmarks with significantly improved parameter efficiency and demonstrates effective compositional transfer to novel tasks through parameter-efficient fine-tuning. Project website: https://deng-cy20.github.io/SMoDP/", "authors": ["Chengyu Deng", "Guanqi Chen", "Yizhou Chen", "Zejia Liu", "Zhiwen Ruan", "Guanhua Chen", "Jia Pan"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-22", "url": "https://arxiv.org/abs/2605.23477", "pdf_url": "https://arxiv.org/pdf/2605.23477v1", "arxiv_id": "2605.23477", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "55784f072a01b76fd56a1e72f980d9421307899a17c96f5f03ea42f2cb48e6c6", "sources": ["arxiv", "semantic_scholar"], "title": "Resident KV Claims: A Conformance Contract for Future Reuse under Active KV Pressure", "abstract": "KV-cache reuse mechanisms increasingly expose priority, duration, offload, routing hints, scheduler modes, and event streams. These mechanisms help preserve reusable prefixes, but they do not by themselves define a portable contract for accepted future-reuse state when resident KV and active live KV cannot both fit. We introduce resident KV claims, a conformance contract that binds future-reuse intent to a materialization predicate, lifecycle state, active/resident feasibility outcome, and claim-level telemetry. In controlled vLLM allocator probes, a 60-block resident claim and a 70-block active prefill exceed an 80-block usable KV pool. Write no-admit prevents the active request from becoming future reusable state, but it still allows active allocation to evict residents from the shared pool. A minimal vLLM prototype shows that hard protected resident claims convert this failure mode into scheduler-visible active refusal with direct blocking-claim attribution. The result is not a production speedup or a new cache-replacement algorithm. It is a runtime contract that turns unreported resident loss into reconstructable active/resident arbitration. A companion MicroRuntime and vLLM litmus suite distinguish ordinary eviction, soft priority, write no-admit, accepted hard claims, materialization failure, demotion, expiry, active refusal, and trace-level outcome reconstruction.", "authors": ["Lukas Stepanek"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-22", "url": "https://arxiv.org/abs/2605.24259", "pdf_url": "https://arxiv.org/pdf/2605.24259v1", "arxiv_id": "2605.24259", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b5bed29f77996fba3a0a93d149771ab189355540fdd1b16122b3ddd7d8c6b96c", "sources": ["arxiv", "semantic_scholar"], "title": "MuKV: Multi-Grained KV Cache Compression for Long Streaming Video Question-Answering", "abstract": "Long streaming video QA remains challenging due to growing visual tokens and limited reasoning length of large language models (LLMs). KV-caching stores the Key-Value (KV) of the historical tokens via LLM prefill and enables more efficient streaming QA. However, existing methods cache every one or two frames, causing redundant memory usage and losing fine-grained spatial details within frame or temporal contexts across frames. This paper proposes MuKV, a method that features a multi-grained KV cache compression module and a semi-hierarchical retrieval approach to improve both efficiency and accuracy for long streaming VideoQA. For the offline KV cache, MuKV extracts visual representations at patch-, frame-, and segment-levels. The multiple levels of granularity preserve both local cues and global temporal context, while maintaining efficiency with a dual signal token compression mechanism guided by self-attention and frequency. For online QA, MuKV designs a semi-hierarchical retrieval method to retrieve relevant KV caches for answer generation. Experiments on long-streaming VideoQA benchmarks show that MuKV significantly improves answer accuracy, without sacrificing memory and online QA efficiency. Moreover, our compression mechanism alone brings consistent benefits across answer accuracy, memory, and QA efficiency over baselines, showcasing highly effective contribution.", "authors": ["Junbin Xiao", "Jiajun Chen", "Tianxiang Sun", "Xun Yang", "Angela Yao"], "categories": ["cs.CV", "cs.AI", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.22269", "pdf_url": "https://arxiv.org/pdf/2605.22269v1", "arxiv_id": "2605.22269", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/IMBALDY/MuKV", "venue": null, "quality_score": 0.65}
{"id": "5c1fa6f5cd729e4de740c84a6528488c846d8af0e5aa39e6696bd1094bc2b8e6", "sources": ["arxiv", "semantic_scholar"], "title": "GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs", "abstract": "Mixture-of-Experts Large Language Models (MoE-LLMs) achieve strong performance but incur substantial memory overhead due to massive expert parameters. Mixed-precision quantization mitigates this cost by allocating expert-wise bit-widths based on their importance, approaching the accuracy-memory Pareto frontier and enabling extreme low-bit quantization. However, existing methods rely on layer-wise importance estimation and overlook router shifts induced by quantization, resulting in suboptimal allocation and routing. In this work, we propose Global Expert-level Mixed-precision Quantization (GEMQ) to overcome these limitations via (1) a global linear-programming formulation that captures model-wide expert importance based on quantization error analysis, and (2) efficient router fine-tuning to adapt routing to quantized experts. These components are integrated into a progressive quantization framework that iteratively refines importance estimation and allocation. Experiments demonstrate that GEMQ significantly reduces memory and accelerates inference with minimal accuracy degradation. Source code is available at https://github.com/jndeng/GEMQ .", "authors": ["Jianing Deng", "Song Wang", "Dongwei Wang", "Zijie Liu", "Tianlong Chen", "Huanrui Yang", "Jingtong Hu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.23078", "pdf_url": "https://arxiv.org/pdf/2605.23078v1", "arxiv_id": "2605.23078", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jndeng/GEMQ", "venue": null, "quality_score": 0.65}
{"id": "bc1ba87a17c6a785190da579061e491810076b1cc820e3cc275cf94bd16ab83c", "sources": ["arxiv", "semantic_scholar"], "title": "Meta-Soft: Leveraging Composable Meta-Tokens for Context-Preserving KV Cache Compression", "abstract": "The KV cache used in large language models has linearly growing time complexity, so LLMs face memory blow-up and reduced decoding efficiency when they process long contexts. Current KV Cache eviction has become an important research direction; however, existing methods based on fixed Soft Tokens (e.g., Judge Q) rely on a static parameter set as the query to evaluate the importance of KV pairs, so they cannot adapt dynamically to different input prompts, and they cannot precisely capture complex and changing task relevance. Also, evicted KV pairs are discarded permanently, so this causes irreversible information loss and context breaks. To address this problem, we propose Meta-Soft, a dynamic compression framework based on probe-driven context integration. Specifically, we build a meta-library with a learnable orthogonal basis matrix $\\mathcal{L}$, and we use a selector network with Gumbel-Softmax to produce differentiable sparse combination weights, so we dynamically synthesize the most targeted $k$ Soft Tokens from the input prompt features. We append these Soft Tokens to the end of the input sequence to probe key information. We also introduce an attention-flow based integration mechanism, which redistributes the semantic information of removed tokens into retained tokens, and this keeps the dropped context information effectively. Experiments on multiple datasets show that our method outperforms existing state-of-the-art eviction methods and provides a new solution for KV Cache compression.", "authors": ["Wei Luo", "Yi Huang", "Songchen Ma", "Huanyu Qu", "Jiang Cai", "Mingkun Xu"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.22337", "pdf_url": "https://arxiv.org/pdf/2605.22337v2", "arxiv_id": "2605.22337", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "993ff84b0b0ed3e9729055f674de7e3202b268c2c43d9d393d87e1172a4986b3", "sources": ["arxiv", "semantic_scholar"], "title": "ASAP: Attention Sink Anchored Pruning", "abstract": "Vision Transformers (ViTs) face severe computational bottlenecks due to the quadratic complexity of self-attention at high resolutions. Existing token reduction methods rely on local metrics - such as single-layer attention scores - that are inherently vulnerable to the attention sink phenomenon, where uninformative tokens are paradoxically preserved over salient foreground objects. We propose ASAP (Attention Sink Anchored Pruning), a training-free framework that recasts this sink as a feature. Modeling ViT information flow as a Lazy Random Walk, ASAP identifies the sink as a dominant accumulator of probability mass. By computing the diffusion distance to the sink within the cumulative transition matrix, ASAP partitions tokens via Radial Diffusion Clustering and compresses background redundancy through Transition Weight Pooling in a single shot. Extensive experiments across image, video, and vision-language tasks demonstrate ASAP outperforms state-of-the-art methods, accelerating throughput by up to 48% while maintaining - or even exceeding - baseline accuracy.", "authors": ["Jaehyuk Lee", "Hanyoung Kim", "Yanggee Kim", "Donghun Lee"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.22372", "pdf_url": "https://arxiv.org/pdf/2605.22372v1", "arxiv_id": "2605.22372", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f9867d713eb57acc60f2d7ae6bba814dff6763a2228dc5e0c950b93f2a0a4a06", "sources": ["arxiv", "semantic_scholar"], "title": "ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning", "abstract": "Recent progress in LLM reasoning has increasingly shifted from single-pass generation to explicit search over intermediate reasoning states. Tree-of-Thoughts (ToT) organizes inference to tree-structured search with branching and backtracking, but it substantially amplifies the Key--Value (KV) cache: retaining KV states for a frontier of partial trajectories quickly becomes a memory bottleneck that limits throughput and constrains search depth and width under fixed hardware budgets. We address this challenge by observing that KV reuse in ToT-style inference is governed by search dynamics: near-term decoding depends primarily on the active branch and its ancestors, whereas inactive subtrees have low short-term reuse probability yet must remain recoverable for backtracking. Motivated by this, we propose ArborKV, a structure-aware eviction framework that couples a lightweight value estimator with a tree-aware allocation policy, and performs purely token-extractive eviction with lazy rehydration to support revisits. Experiments on ToT-style reasoning benchmarks show that ArborKV achieves up to ~4x peak KV-memory reduction while preserving near-full-retention accuracy, enabling larger search configurations under fixed device budgets that would otherwise run out of memory.", "authors": ["Yeqiu Chen", "Ziyan Liu", "Zhenxin Huang", "Runquan Gui", "Hong Wang", "Lei Liu"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-21", "url": "https://arxiv.org/abs/2605.22106", "pdf_url": "https://arxiv.org/pdf/2605.22106v1", "arxiv_id": "2605.22106", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "13455e8850b9cbb8bd37f0a3470cc974f9c6e5653e514ab11a01a3d68dc50612", "sources": ["arxiv", "semantic_scholar"], "title": "OCTOPUS: Optimized KV Cache for Transformers via Octahedral Parametrization Under optimal Squared error quantization", "abstract": "The key-value (KV) cache dominates memory bandwidth and footprint in long-context autoregressive inference. Recent rotation-preconditioned codecs (TurboQuant, PolarQuant) show that a structured random rotation followed by a per-coordinate scalar quantizer matched to an analytically tractable marginal is a near-optimal recipe for KV compression. OCTOPUS advances this paradigm through joint quantization of rotated coordinate triplets. Each triplet's direction is mapped to a square via an octahedral parameterization, and the two resulting coordinates and the triplet norm are Lloyd-Max quantized against implementation-matched marginals. Optimizing the per-triplet squared error gives a strictly non-uniform bit allocation depending only on the total dimensionality of the keys. We find the finite-dimensional quality optimum with sweeps to be constant on every real decoder we test. The codec is data-oblivious, online, and deterministic given a seed. Across text, video, and audio, OCTOPUS matches or beats every prior rotation codec at every reported bit width and metric, with a lead that grows as bits drop for extreme compression. Furthermore, a fused Triton implementation reconstructs keys on the fly without materializing the uncompressed key, so the codec adds no decode-time bandwidth or latency over the existing dequantization. Project Page: https://octopus-quant.github.io/", "authors": ["Mark Boss", "Vikram Voleti", "Simon Donné", "Shimon Vainer"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.21226", "pdf_url": "https://arxiv.org/pdf/2605.21226v1", "arxiv_id": "2605.21226", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "3a26771612b52e609a6d4d4a7d8d2c6aa59d7afeb2a2e71d55f7966fcfcf8dbe", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive KV Cache Reuse for Fast Long-Context LLM Serving", "abstract": "In long-context Large Language Model (LLM) inference, the Time-To-First-Token (TTFT) latency incurred by the prefill stage has become the foremost bottleneck limiting interactive performance and deployment cost. KV Cache reuse offers a direct path to reduce redundant prefill, yet traditional prefix caching applies only to strict-prefix scenarios; directly reusing KV Cache in non-prefix settings breaks the cross-chunk global attention relationships and causes significant degradation in generation quality. When reusable KV Cache is offloaded to GPU-external cache pools, I/O overheads across heterogeneous hardware tiers further emerge as a new TTFT bottleneck. Efficient non-prefix KV Cache reuse therefore requires both semantic-consistency recovery and compute-I/O co-optimization. This paper presents CacheTune, a frequency-guided and hardware-aware KV Cache reuse system for long-context LLM serving. CacheTune first identifies, offline, the KV pairs most critical to cross-attention recovery through frequency-domain analysis, and then selectively recomputes only these semantic-critical tokens online while reusing the remaining KVs. To turn this semantic selection into end-to-end latency reduction, CacheTune further combines sparse KV transfer, multi-stream asynchronous overlap, deferred positional-encoding recovery, and hardware-aware adaptive recomputation-ratio tuning to balance computation and data movement across heterogeneous cache pools. Evaluations on mainstream LLMs and long-context tasks show that CacheTune achieves 3.72x-4.86x TTFT speedup and 3.93x-6.21x higher throughput while maintaining generation quality close to full recompute. Even when caches are offloaded to I/O-bound SSD/HDD storage, CacheTune sustains 2.34x-2.36x TTFT speedup through adaptive recomputation.", "authors": ["Fei li", "Song Liu", "Yan Liu", "Jinhua Cui", "Shiqiang Nie", "Jinyu Wang", "Weiguo Wu"], "categories": ["cs.AR", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.24022", "pdf_url": "https://arxiv.org/pdf/2605.24022v1", "arxiv_id": "2605.24022", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "7d9ba1716a760bed2899e2bf180fc5feedc8040c5da3f0965cbefb44a88cd3ef", "sources": ["arxiv", "semantic_scholar"], "title": "Runtime-Certified Bounded-Error Quantized Attention", "abstract": "KV cache quantization reduces the memory cost of long-context LLM inference, but introduces approximation error that is typically validated only empirically. Existing systems rely on average-case robustness, with no mechanism to detect or recover from failures at runtime. We present a tiered KV cache architecture that enables runtime-certified attention: INT8 keys and INT4 values are stored in GPU memory, while FP16 originals are retained in system RAM for deterministic fallback. A two-term error decomposition yields per-head, per-step bounds on (i) attention distribution distortion from key quantization and (ii) value reconstruction error. These bounds are computed online and used to drive adaptive precision selection and a multi-stage fallback ladder, which guarantees recovery to the exact dense attention output when required. Across PG-19, NIAH, and RULER benchmarks on LLaMA~3.1-8B with contexts up to 128K, the system matches dense FP16 KV quality within noise for language modelling and retrieval tasks, while recovering catastrophic failures observed in naive INT8/INT4 baselines. Value-sensitive tasks at short context expose a controlled trade-off between compression and fidelity, which can be eliminated via tighter value tolerances or FP16-value fallback. The certification is local (per-head, per-step) and does not guarantee end-to-end model correctness, but ensures that each attention computation is either bounded relative to an FP16 reference or exactly recovered via fallback. This reframes KV cache quantization as a runtime-verified computation rather than a fixed approximation. The goal is not raw speedups, but enabling safe deployment of aggressive KV compression under strict quality constraints.", "authors": ["Dean Calver"], "categories": ["cs.LG", "cs.AI", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.20868", "pdf_url": "https://arxiv.org/pdf/2605.20868v1", "arxiv_id": "2605.20868", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5097a4413db04ebf41c66e7a11c7b2c24f0cdb425f5bd8a0a45d1f335c872aa8", "sources": ["arxiv", "semantic_scholar"], "title": "NanoCP: Request-Level Dynamic Context Parallelism for Data-Expert Parallel Decoding", "abstract": "Modern serving systems for Mixture-of-Experts (MoE) models adopt hybrid data-expert parallelism: expert parallelism (EP) shards experts across GPUs to scale capacity, while data parallelism (DP) replicates attention layers across instances to process independent requests. Existing systems bind each request's attention, MoE communication, and KV cache to a single instance. Because attention latency scales with KV cache size while MoE communication latency scales with batch size, this binding cannot balance both simultaneously, producing EP stragglers; it also fragments KV memory across instances, inflating tail latency under long contexts. While existing context parallelism (CP) mitigates these constraints, its uniform parallelism degree incurs prohibitive communication and attention-side overheads. We present \\work, which decouples MoE communication from KV cache placement and achieves dual balance through dynamic context parallelism (DCP). DCP assigns each request a context-parallel degree sized to its KV footprint: long requests distribute attention across multiple instances; short requests remain local. This dynamic parallelism effectively liquefies the KV cache across the cluster, balancing both the per-instance KV cache occupancy and batch sizes without unnecessary load-balancing costs. To bridge DCP with static execution, \\work introduces an ahead-of-time (AOT) graph engine paired with a custom routing-based communication backend. Experimental results show that \\work maintains up to $1.88\\times$--$3.27\\times$ higher request rates under strict time-per-output-token (TPOT) service level objectives (SLOs). Furthermore, \\work significantly mitigates stragglers, reducing P99 tail latency by up to $1.79\\times$--$2.12\\times$.", "authors": ["Jiefei Chen", "Binbin Lin", "Jinming Ma", "Jiangfei Duan", "Haojie Duanmu", "Hao Liu", "Qinxiu Cheng", "Xiuhong Li", "Zhilin Pei", "Hui Wang", "Xingcheng Zhang", "Dahua Lin"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.21100", "pdf_url": "https://arxiv.org/pdf/2605.21100v1", "arxiv_id": "2605.21100", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "45851014eee01e7aaf26c7a5b5b3da4a2c89178f3f9461434be4251569b4ae22", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic TMoE: A Drift-Aware Dynamic Mixture of Experts Framework for Non-Stationary Time Series Forecasting", "abstract": "Non-stationary time series forecasting is challenged by evolving distribution shifts that static models struggle to capture. While Mixture-of-Experts (MoE) architectures offer a promising paradigm for decoupling complex drift patterns, existing approaches are limited by fixed expert pools and memoryless routing, hampering their ability to adapt to abrupt regime shifts. To address this, we propose Dynamic TMoE, a framework that unifies architectural evolution with temporal continuity during learning phase. By detecting distribution shifts via Maximum Mean Discrepancy (MMD), we dynamically instantiate heterogeneous experts and prune redundant ones to optimize capacity. Additionally, a temporal memory router leverages recurrent states and an anomaly repository to ensure stable, context-aware expert selection without requiring test-time updates. Experiments on nine benchmarks demonstrate state-of-the-art performance, reducing MSE by 10.4% and MAE by 7.8%. Code is available at https://github.com/andone-07/Dynamic-TMoE.", "authors": ["Jiawen Zhu", "Shuhan Liu", "Di Weng", "Yingcai Wu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.20678", "pdf_url": "https://arxiv.org/pdf/2605.20678v1", "arxiv_id": "2605.20678", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/andone-07/Dynamic-TMoE", "venue": null, "quality_score": 0.65}
{"id": "4e5862621c61c300f9ab95b4f8fa94ef1ce14225651c8649ef5568fb11cbfc53", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Routing: Characterising Expert Tuning and Representation in Vision Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) models are often interpreted by analysing which categories are routed to which experts. However, routing alone does not reveal what each expert actually encodes. We train sparsely-gated convolutional MoE models with a contrastive objective on natural images and characterise expert specialisation using tools from visual neuroscience. Extending from gating-level to expert-level analyses, we measure per-expert category separability, and per-expert tuning using the most exciting inputs. Extending from category-level to feature-level explanations, we interpret tuning via semantic dimensions derived from a dataset of human behavioural judgements (THINGS). Finally, we use tuning and representational similarity analysis to assess the stability of expertise-allocation across independent initialisations. We find that an animate-inanimate distinction dominates expert partitioning, apparent from gating through to expert readout, and is stable across independently trained models. Although routing statistics suggest relatively sparse, categorical preferences, expert analyses reveal broader tuning to continuous visual and semantic dimensions that extend beyond category boundaries. Experts exhibit similar category-separability to one another, despite distinct feature tuning, demonstrating the explanatory benefits of moving beyond category-level analyses. Together, these results show that expert specialisation in vision MoEs extends well beyond category routing and is better understood by probing fine-grained expert-level tuning and representational structure.", "authors": ["Gene Tangtartharakul", "Katherine R. Storrs"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.20610", "pdf_url": "https://arxiv.org/pdf/2605.20610v1", "arxiv_id": "2605.20610", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c08824c4c5c184721917a400dc5a9eb4e992dd4df7274f6748aad169dd56983a", "sources": ["arxiv", "semantic_scholar"], "title": "Task-Routed Mixture-of-Experts with Cognitive Appraisal for Implicit Sentiment Analysis", "abstract": "Implicit sentiment analysis is challenging because sentiment toward an aspect is often inferred from events rather than expressed through explicit opinion words. Existing models typically learn from the final polarity label, which provides limited guidance for reasoning about sentiment from the context. Motivated by cognitive appraisal theory, we propose an appraisal-aware multi-task learning (MTL) framework for implicit sentiment analysis that provides polarity prediction with two complementary auxiliary tasks: implicit sentiment detection and cognitive rationale generation. However, training several objectives with different targets and sharing a single backbone across tasks in MTL limits flexibility and can lead to task interference. To reduce interference among these related but distinct objectives, we adopt task-level mixture-of-experts models in which all tasks share a common set of experts, and task identity controls the sparse combination of these experts. Our method builds on an encoder-decoder architecture and replaces a subset of encoder and decoder blocks with these sparse mixtures. We use a task-conditioned router to select sparse expert mixtures for each task, and a task-separated routing objective to encourage different tasks to learn distinct expert-selection patterns. Experimental results show that our model outperforms recently proposed approaches, with strong gains on the implicit sentiment subset. Our code is available at https://github.com/yaping166/TRMoE-ISA.", "authors": ["Yaping Chai", "Haoran Xie", "Joe S. Qin"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.20916", "pdf_url": "https://arxiv.org/pdf/2605.20916v1", "arxiv_id": "2605.20916", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/yaping166/TRMoE-ISA", "venue": null, "quality_score": 0.65}
{"id": "0fab166949be451c217f839504866a72e3538be8fca90d34a2b6ad7907d92cc1", "sources": ["arxiv", "semantic_scholar"], "title": "OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond", "abstract": "The rapid advancement toward long-context reasoning and multi-modal intelligence has made the memory footprint of the Key-Value (KV) cache a dominant memory bottleneck for efficient deployment. While the established per-channel quantization effectively accommodates intrinsic channel-wise outliers in Key tensors, its efficacy diminishes under extreme compression. In this work, we revisit the inherent limitations of the per-channel quantization paradigm from both empirical and theoretical perspectives. Our analysis identifies Token Norm Imbalance (TNI) as the primary bottleneck to quantization fidelity. We demonstrate that TNI systematically amplifies errors when shared quantization parameters are required to span token groups exhibiting substantial norm disparities. Instead of relying on intricate quantization pipelines (e.g., TurboQuant), we propose OScaR (Omni-Scaled Canalized Rotation), an accurate and lightweight KV cache compression framework for X-LLMs (i.e., text-only, multi-modal, and omni-modal LLMs). Advancing the per-channel paradigm, OScaR employs Canalized Rotation followed by Omni-Token Scaling to mitigate TNI-induced sequence-dimensional variance both effectively and efficiently, further supported by our optimized system design and CUDA kernels. Extensive evaluations across X-LLMs show that OScaR consistently outperforms existing methods and achieves near-lossless performance under INT2 quantization, establishing it as a robust, low-complexity, and universal framework that defines a new Pareto front. Compared with the BF16 FlashDecoding-v2 baseline, our OScaR implementation achieves a notable up to 3.0x speedup in decoding, reduces memory footprint by 5.3x, and increases throughput by 4.1x. The code for OScaR is publicly available at https://github.com/ZunhaiSu/OScaR-KV-Quant.", "authors": ["Zunhai Su", "Rui Yang", "Chao Zhang", "Yaxiu Liu", "Yifan Zhang", "Wei Wu", "Jing Xiong", "Dayou Du", "Xialie Zhuang", "Yulei Qian", "Yuchen Xie", "Yik-Chung Wu", "Hongxia Yang", "Ngai Wong"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-19", "url": "https://arxiv.org/abs/2605.19660", "pdf_url": "https://arxiv.org/pdf/2605.19660v1", "arxiv_id": "2605.19660", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ZunhaiSu/OScaR-KV-Quant", "venue": null, "quality_score": 0.65}
{"id": "3319c27a92041c705ad6ef7ab523d8d4e9b8f58396d6207da597f21ac9c7545c", "sources": ["arxiv", "semantic_scholar"], "title": "Rotation-Aligned Key Channel Pruning for Efficient Vision-Language Model Inference", "abstract": "Vision-Language Models suffer severe KV cache pressure at inference, as a single image often encodes into thousands of tokens. Most existing methods exploit token sparsity through token pruning, but permanently discarding visual content causes substantial degradation on fine-grained perception tasks. This motivates a complementary axis, feature sparsity: under a fixed KV cache budget, compressing the channel dimension preserves more visual tokens at the same memory cost. Prior Key channel pruning methods, however, face a structural trade-off: token-wise channel pruning is expressive but unstructured and slow, while head-wise approach is hardware-friendly but less robust. We resolve this with RotateK, a rotation-based structured Key channel pruning framework. RotateK applies an online PCA-based rotation that aligns token-dependent channel importance into a shared low-dimensional subspace, enabling accurate pruning under lightweight head-wise masks; a fused Triton attention kernel operates directly on sparse-channel Keys for efficient decoding. Experiments on two representative VLM backbones show that RotateK consistently outperforms prior Key channel pruning in both accuracy and decoding latency, while joint token-channel pruning improves over token-only baselines at matched KV cache budgets.", "authors": ["Beomseok Kang", "Dongwon Jo", "Jiwon Song", "Donghwee Son", "Jae-Joon Kim"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-19", "url": "https://arxiv.org/abs/2605.19218", "pdf_url": "https://arxiv.org/pdf/2605.19218v1", "arxiv_id": "2605.19218", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "67913dd513a278ba93d5c32f2b17be9a43a6bad1bea4a36d41e45759a5056764", "sources": ["arxiv", "semantic_scholar"], "title": "OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization", "abstract": "INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms reduce outliers, but still degrade at INT2 because they are not aligned with downstream attention. We propose OSCAR, an Ultra-low-bit KV Cache quantization method that estimates attention-aware covariance structures offline and uses them to derive fixed rotations and clipping thresholds for quantization. In this way, it aligns KV quantization with the covariance structures that attention actually consumes. More importantly, we not only provide theoretical justification but also develop a fully deployable OSCAR system with a custom INT2 attention kernel that remains compatible with paged KV-cache serving and fused kernel pipelines, enabling seamless integration into modern LLM serving frameworks such as SGLang and vLLM. We evaluate our methods on recent reasoning models with reasoning traces of up to 32k tokens across 5 tasks. On Qwen3-4B-Thinking-2507 and Qwen3-8B, OSCAR reduces the BF16 accuracy gap to 3.78 and 1.42 points, respectively, while naive rotation INT2 collapses to nearly zero. We further scale OSCAR to Qwen3-32B and GLM-4.7 (358B params), where it remains effectively on par with BF16. On long context - RULER-NIAH up to 128K, OSCAR remains robust on both Qwen3 models, while naive rotation INT2 collapses. System-wise, OSCAR reduces KV-cache memory by approximately 8x, improves throughput by up to 7x at large batch sizes under the same memory budget, and accelerates batch-size-1 decoding by up to 3x over BF16 due to reduced memory bandwidth overhead.", "authors": ["Zhongzhu Zhou", "Donglin Zhuang", "Jisen Li", "Ziyan Chen", "Shuaiwen Leon Song", "Ben Athiwaratkun", "Xiaoxia Wu"], "categories": ["cs.LG", "cs.AI", "cs.DC", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-18", "url": "https://arxiv.org/abs/2605.17757", "pdf_url": "https://arxiv.org/pdf/2605.17757v1", "arxiv_id": "2605.17757", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4da3626b69c1e84b0b5863d736f5622865f437930ac821488f6bb5d5108ca520", "sources": ["arxiv", "semantic_scholar"], "title": "KVDrive: A Holistic Multi-Tier KV Cache Management System for Long-Context LLM Inference", "abstract": "Supporting long-context LLMs is challenging due to the substantial memory demands of the key-value (KV) cache. Existing offloading systems store the full cache in host memory and selectively fetch critical entries during decoding, but this strategy quickly hits a ceiling: sparsity cannot be pushed further without degrading accuracy. As a result, when context length and batch size grow, the volume of KV transfers rises sharply and becomes the dominant source of decoding latency. We present KVDrive, a holistic multi-tier KV cache management system spanning GPU memory, host DRAM, and SSD. Unlike prior work that pursues greater sparsity through algorithmic refinements, KVDrive tackles the problem from a systems perspective - jointly orchestrating cache placement, pipeline scheduling, and cross-tier coordination to sustain high-throughput inference under tight GPU budgets. KVDrive advances three fundamental capabilities: it adapts cache management to attention behavior to maximize reuse and minimize redundant data movement; it restructures the decoding pipeline to overlap I/O- and CPU/GPU compute-bound stages, eliminating stalls across heterogeneous resources; and it harmonizes data movement across memory tiers to unlock scalable long-context inference far beyond GPU and DRAM limits. We have implemented a fully functional prototype of KVDrive and evaluated it on long-context benchmarks with popular LLMs. The system achieves up to 1.74x higher throughput compared to state-of-the-art works while preserving accuracy.", "authors": ["Jian Lin", "Jiazhi Mi", "Zicong Hong", "Haodong Wang", "Qianli Liu", "Haodyue Zhang", "Peng Li", "Song Guo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-18", "url": "https://arxiv.org/abs/2605.18071", "pdf_url": "https://arxiv.org/pdf/2605.18071v1", "arxiv_id": "2605.18071", "doi": "10.1145/3802077", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "dad487bf361cd50b76dc26f154f08f386fcd2fa8687ecb069b8b7c00adc80424", "sources": ["arxiv", "semantic_scholar"], "title": "CP-MoE: Consistency-Preserving Mixture-of-Experts for Continual Learning", "abstract": "Catastrophic forgetting remains a major obstacle to continual learning in large language models (LLMs) and vision--language models (VLMs). Although Mixture-of-Experts (MoE) architectures offer an efficient path to scaling, existing LoRA-based MoE continual learning methods still face a fundamental trade-off: they either isolate experts too aggressively, limiting knowledge transfer across tasks, or allow task-specific updates to overwrite important existing parameters, leading to severe forgetting. To address this, we propose CP-MoE, a continual learning framework built around a transient expert that captures early task-specific updates and guides their integration into stable experts. CP-MoE introduces a consistency-preserving routing bias, which uses the transient expert to estimate representation similarity with stable experts and steer routing towards more compatible expert selection, and a transient expert-guided regularisation mechanism, which selectively protects important historical parameters during merging. Together, these components reduce parameter interference and forgetting while preserving cross-task knowledge transfer. We validate CP-MoE on both unimodal and multimodal continual learning benchmarks with LLM-based and VLM-based MoE models. On SuperNI benchmark, spanning diverse sequential language tasks, CP-MoE achieves state-of-the-art performance and stronger zero-shot transfer to unseen tasks. On VQA v2 dataset, it scales effectively to multimodal visual reasoning, consistently reduces forgetting, and outperforms strong MoE baselines.", "authors": ["Yang Liu", "Toan Nguyen", "Flora D. Salim"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-18", "url": "https://arxiv.org/abs/2605.20247", "pdf_url": "https://arxiv.org/pdf/2605.20247v1", "arxiv_id": "2605.20247", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "9430cbe0ebbb98047f0e914e16ca816e81b770e688db77ac8cb8926abb47b09e", "sources": ["arxiv", "semantic_scholar"], "title": "Prune, Update and Trim: Robust Structured Pruning for Large Language Models", "abstract": "Large Language Models (LLMs) have experienced significant growth and development in recent years. However, performing inference on LLMs remains costly, especially for long-context inference or in resource-constrained devices. This motivates the development of new post-training pruning (PTP) methods. These methods reduce LLMs' requirements by removing a substantial part of the model's parameters. The discarded weights are selected depending on their impact on the models performance. Current PTP methods prune the models by removing the less informative hidden nodes from the FFN layers, and the least important attention layers. We propose Putri, a PTP method that introduces three changes to the State- of-the-art. First, we update the un-pruned weights of the FFN to compensate for the introduced pruning error. Second, the FFN layers are pruned sequentially, taking into account the updates done to the previous layers. Third, instead of removing full attention layers, we remove individual attention-heads. We extend this method such that it can also address Grouped-Query Attention. In summary, Putri is a structure pruning method which remains simple while showing SOTA performance. Pruning experiments on multiple models with a wide variety of sparsity ranges and on different datasets, validate the generality of Putri. Notably, we demonstrate that, unlike previous methods, Putri can prune LLMs on extreme sparsity ratios. The code is available at: https://github.com/Coello-dev/Putri.", "authors": ["Diego Coello de Portugal Mecke", "Tom Hanika", "Lars Schmidth-Thieme"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-18", "url": "https://arxiv.org/abs/2605.18331", "pdf_url": "https://arxiv.org/pdf/2605.18331v1", "arxiv_id": "2605.18331", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Coello-dev/Putri", "venue": null, "quality_score": 0.65}
{"id": "b93b87fd633f733bdf5ceba559a84aadb71ad8a193f184da733f8c18b510756c", "sources": ["arxiv", "semantic_scholar"], "title": "VeriCache: Turning Lossy KV Cache into Lossless LLM Inference", "abstract": "The large size of the KV cache has become a major bottleneck for serving LLMs with increasing context lengths. In response, many KV cache compression methods, such as token dropping and quantization, have been proposed. However, almost all of these methods are inherently lossy-despite minimal accuracy degradation for short outputs, their outputs increasingly diverge from full-KV-cache outputs as more tokens are decoded, which leads to catastrophic failures in code generation and tool calling. We present VeriCache, the first inference framework that ensures the same output as full-KV-cache decoding but largely preserves the high decoding throughput of a range of KV cache compression algorithms. VeriCache uses the compressed KV cache to draft tokens, then verifies them against the full KV cache. While it may seem like just speculative decoding, VeriCache requires addressing a key system challenge to work-keeping the full KV cache out of GPU memory and minimizing the overhead of swapping it in for verification. The insight is two-fold: (1) compressed-KV decoding can be parallelized with full-KV swap, because one is HBM-bandwidth-bound and the other is PCIe/network-bound, and (2) the compressed KV cache often produces output similar to the full KV cache, allowing a long drafting horizon to amortize each full-KV swap. VeriCache applies to both long-context decoding and remote prefix caching, supports a broad family of token-dropping and quantization methods through a uniform compressor interface, and composes with traditional speculative decoding. Experimental results show that VeriCache achieves up to 4X higher throughput than full-KV inference while producing identical outputs.", "authors": ["Jiayi Yao", "Samuel Shen", "Kuntai Du", "Shaoting Feng", "Dongjoo Seo", "Rui Zhang", "Yuyang Huang", "Yuhan Liu", "Shan Lu", "Junchen Jiang"], "categories": ["cs.AR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-17", "url": "https://arxiv.org/abs/2605.17613", "pdf_url": "https://arxiv.org/pdf/2605.17613v1", "arxiv_id": "2605.17613", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "36b8acaa25b1825c4745e29dfef08af50a4ac6965e00dd99369f0c35a1daa3ec", "sources": ["arxiv", "semantic_scholar"], "title": "FastOCR: Dynamic Visual Fixation via KV Cache Pruning for Efficient Document Parsing", "abstract": "Vision-Language Models (VLMs) have shown strong promise on Optical Character Recognition (OCR), yet the sheer number of visual tokens required to encode dense documents incurs prohibitive inference cost. Existing pruning methods rely on physical eviction, e.g., permanently discarding visual tokens during the prefill stage. While effective for natural images, this strategy fundamentally breaks down on OCR, where virtually every visual token may correspond to a character or structural element, and any irreversible loss leads to catastrophic accuracy degradation. We observe that, although document images appear globally dense and seemingly unprunable, the model's attention to them is in fact temporally sparse: at each decoding step it concentrates on a small region that shifts gradually across steps, much as a human reader fixates on successive words rather than perceiving an entire page at once. Motivated by this Dynamic Visual Fixation phenomenon, we recast the intractable global pruning problem as a tractable local, dynamic one and propose FastOCR, a training-free framework with two complementary modules. Specifically, Focal-Guided Pruning identifies a small set of focal layers and selects the most task-relevant visual tokens from them at each step, while Cross-Step Fixation Reuse exploits the gradual shift of fixation to warm-start each step from the previous one. By dynamically adjusting which tokens are attended rather than evicting any from the cache, FastOCR avoids permanent information loss. Extensive experiments show that FastOCR serves as a plug-and-play acceleration module, generalizing consistently across five VLMs of varying sizes and architectures. On Qwen2.5-VL, FastOCR retains 98% of the unpruned model's accuracy while attending to only 5% of the visual tokens per decoding step, reducing attention latency by 3.0$\\times$.", "authors": ["Zihan Tang", "Leqi Shen", "Hui Chen", "Ao Wang", "Ben Wan", "Yan Feng", "Ke Zhang", "Sicheng Zhao", "Tongxuan Liu", "Guiguang Ding"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-17", "url": "https://arxiv.org/abs/2605.17447", "pdf_url": "https://arxiv.org/pdf/2605.17447v1", "arxiv_id": "2605.17447", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1a128d3435cc8c20b2102a526e0fdd9f27106e523f2de09de19370cc4d21c2e5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts for Low-Resource LLMs", "abstract": "Mixture-of-Experts (MoE) architectures enable efficient model scaling, yet expert routing behavior across underrepresented languages remains poorly understood. We analyze routing dynamics in two architecturally distinct MoE models -- a pure Transformer (Qwen3-30B-A3B) and a hybrid Mamba-Transformer (Nemotron-3-Nano-30B-A3B) -- using Hebrew as a morphologically rich, low-resource testbed. Both pre-trained models exhibit \\emph{deep-layer routing collapse}: usage entropy drops sharply in final layers and tokens concentrate on a narrow expert subset, a pattern largely absent for English. Continual pre-training (CPT) on balanced bilingual data substantially corrects this imbalance, increasing entropy and shifting routing toward shared, language-agnostic experts; supervised fine-tuning (SFT) alone achieves less complete correction. Extending the analysis to Japanese reveals quantitatively consistent collapse signatures, providing cross-linguistic evidence that the phenomenon is a systematic consequence of pre-training underrepresentation rather than any language-intrinsic property. Routing improvements correlate with consistent downstream benchmark gains, positioning routing entropy and expert specialization as principled diagnostics for multilingual capacity in MoE systems.", "authors": ["Ori Bar Joseph", "Smadar Arvatz", "Noam Kayzer", "Dan Revital", "Sarel Weinberger"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-17", "url": "https://arxiv.org/abs/2605.17598", "pdf_url": "https://arxiv.org/pdf/2605.17598v1", "arxiv_id": "2605.17598", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4c8039055e72df82136c1b13514d5cf1bfe2e310acd1ae152938e49722960cf7", "sources": ["arxiv", "semantic_scholar"], "title": "Stable Routing for Mixture-of-Experts in Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) requires models to learn new classes sequentially while preserving prior knowledge. Recently, approaches that combine pre-trained models with mixture-of-experts (MoE) have received increasing attention in CIL: they typically expand experts during learning and employ a router to assign weights across experts. However, existing MoE methods often overlook routing drift induced by expert expansion. Once new experts are introduced, the router may reassign samples from earlier classes to newly added experts, thereby perturbing previously established expert compositions and causing interference even when old experts remain frozen. We argue that expandable MoE in CIL requires two complementary properties: stable old-class routing for knowledge preservation and sufficient capacity utilization for new-class adaptation. To this end, we propose Stable Routing for MoE (StaR-MoE), a routing-level framework for expandable MoE in CIL. By incorporating sensitivity-aware routing alignment, StaR-MoE aligns current old-class routing behavior with historical routing distributions through sensitivity-guided constraints. Complementarily, StaR-MoE introduces asymmetric capacity regularization to encourage effective utilization of the expanded expert pool without compromising class-specific routing specialization. Extensive experiments across four standard CIL benchmarks demonstrate that StaR-MoE consistently improves both average and last accuracy over state-of-the-art methods, highlighting the importance of stable routing.", "authors": ["Zirui Guo", "Quan Cheng", "Da-Wei Zhou", "Lijun Zhang"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-17", "url": "https://arxiv.org/abs/2605.17571", "pdf_url": "https://arxiv.org/pdf/2605.17571v1", "arxiv_id": "2605.17571", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "43f8df365e24454599de8da3acc592bad1b3a4d6acaea58b99687323054547ec", "sources": ["arxiv", "semantic_scholar"], "title": "TriAxialKV: Toward Extreme Low-Precision KV-Cache Quantization for Agentic Inference Tasks", "abstract": "Agentic workloads have emerged as a major workload for LLM inference. They differ significantly from chat-only workloads, requiring long-context processing, the ability to handle multimodal inputs, and structured multi-turn interactions with tool calling capabilities. As a result, their context exhibits structure that can carry different importance along three key axes: temporal recency to the current turn, modality such as text or image tokens, and semantic role such as user queries, tool calls, observations, or reasoning. These axes capture distinct token behaviors and lead to different sensitivities to KV-cache compression. However, existing KV-cache quantization methods are typically homogeneous or exploit only heterogeneity on a single dimension, such as temporal proximity or modality, overlooking the interactions among them. To this end, we introduce TriAxialKV, a novel mixed-precision KV-cache quantization scheme that assigns each token a triaxial tag, calibrates per-tag sensitivity, and allocates INT2/INT4 bitwidths under a fixed memory budget. We implement TriAxialKV as an end-to-end serving system, comprising calibration, mixed-precision quantization and memory management, and custom fused Triton decode kernels. When using Qwen3-VL-32B-Thinking as a computer-use agent operating the OSWorld, TriAxialKV matches the accuracy of SGLang with BF16 KV cache while supporting 4.5$\\times$ KV cache size and achieving 30% higher end-to-end throughput, when running on real GPU systems.", "authors": ["Hanzhang Shen", "Haoran Wu", "Yiren Zhao", "Robert Mullins"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-16", "url": "https://arxiv.org/abs/2605.17170", "pdf_url": "https://arxiv.org/pdf/2605.17170v1", "arxiv_id": "2605.17170", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ab287df40005b3a360b4861d54dc92d9fb17dfa3f022491b042706b18dc90978", "sources": ["arxiv", "semantic_scholar"], "title": "ObjectCache: Layerwise Object-Storage Retrieval for KV Cache Reuse", "abstract": "Prefix KV caching has become a key mechanism in LLM serving: it reduces time to first token (TTFT) by avoiding redundant computation across requests that share a prefix (i.e., the system prompt). However, the accumulated KV cache is often larger than what GPU memory and local DRAM can hold. To preserve latency, current systems keep the KV cache in remote DRAM pools, increasing serving-cluster size and cost. In this paper, we explore a different approach: storing the KV cache in S3-compatible object storage so that capacity is no longer the constraint, while minimizing the impact on TTFT. We propose ObjectCache, which co-designs the storage protocol and transfer schedule so that the storage server delivers KV cache data in the order the GPU consumes it, overlapping data transfer with compute across concurrent requests. We prototype ObjectCache on a 100 Gbps RoCE cluster with NIXL (an inference library that abstracts storage and memory), Ceph RGW (an Object Gateway for clusters), and DAOS (an open source storage system). For 64K contexts, common in today's systems, ObjectCache adds only 5.6\\% latency over local DRAM; for 4K contexts, where less compute is available to mask transfer, ObjectCache adds 56--75\\,ms over the optimal local layerwise baseline. Under shared bandwidth caps, our scheduler reduces added TTFT by 1.2--1.8x compared with equal bandwidth sharing.", "authors": ["Yu Zhu", "Aditya Dhakal", "Yunming Xiao", "Dejan Milojicic", "Gustavo Alonso"], "categories": ["cs.DC", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-16", "url": "https://arxiv.org/abs/2605.22850", "pdf_url": "https://arxiv.org/pdf/2605.22850v1", "arxiv_id": "2605.22850", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.65}
{"id": "6613298ee5f564f65b60c55ffbcb7d084fe4918cb2546935c9bc0a29ebd0569b", "sources": ["arxiv", "semantic_scholar"], "title": "E-PMQ: Expert-Guided Post-Merge Quantization with Merged-Weight Anchoring", "abstract": "Low-resource deployment constraints have made model quantization essential for deploying neural networks while preserving performance. Meanwhile, model merging has become an increasingly practical low-resource strategy for integrating multiple task- or domain-specialized experts into a single model without joint training or multi-model serving. Together, quantization and model merging enable an efficient low-resource deployment pipeline by integrating multiple experts into one low-bit model. We formulate this setting as Post-Merge Quantization (PMQ). We show that directly applying post-training quantization (PTQ) to a merged model is unreliable because two distinct deviations are coupled: the quantization deviation introduced by low-bit reconstruction and the expert-relative merging deviation inherited from model merging. To mitigate these deviations, we propose E-PMQ, an expert-guided PMQ framework that uses source expert weights to provide expert- guided output targets during layer-wise calibration, together with merged-weight anchoring to stabilize the calibration and preserve the integrated behavior of the merged model. On CLIP-ViT-B/32 eight-task merging, E-PMQ improves 4-bit GPTQ from 65.0% to 73.6% under Task Arithmetic and from 69.1% to 74.8% under TIES-Merging. On harder settings, E-PMQ improves GPTQ from 34.8% to 76.7% on 20-task CLIP-ViT-L/14 and from 78.26% to 83.34% on FLAN-T5- base GLUE. These results demonstrate that E-PMQ enables effective post-merge quantization and low-bit deployment.", "authors": ["Wenjun Wang", "Yanggan Gu", "Shuo Cai", "Yuanyi Wang", "Pengkai Wang", "Jianmin Wu", "Hongxia Yang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-16", "url": "https://arxiv.org/abs/2605.16882", "pdf_url": "https://arxiv.org/pdf/2605.16882v1", "arxiv_id": "2605.16882", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b039c1210ceb5c8ae5079bae052ee8e59c1393c5087fdda1f506a61f523cc54e", "sources": ["arxiv", "semantic_scholar"], "title": "CAR-SAM: Cross-Attention Reconstruction for Post-Training Quantization of the Segment Anything Model", "abstract": "Segment Anything Models (SAMs) are extensively used in computer vision for universal image segmentation, but deploying them on resource-constrained devices is challenging due to their high computational and memory demands. Post-Training Quantization (PTQ) is a widely used technique for model compression and acceleration. However, existing PTQ methods fail to consider the cross-attention architecture in the SAM decoder. This degradation primarily stems from the unique challenges posed by SAMs: (1) Attention dissipation, where the attention information in the decoder, which is crucial for representing segmentation masks, collapses into a diffuse and non-semantic form under low-bit quantization; and (2) Reconstruction oscillation, where bidirectional coupling within the two-way transformer introduces cross-branch error interference and destabilizes convergence. To tackle these issues, we propose CAR-SAM, a unified quantization framework tailored for SAMs. Firstly, to mitigate attention dissipation, we introduce MatMul-Aware Compensation (MAC) mechanism that transfers activation-induced quantization errors from MatMul to preceding linear weights. Secondly, to mitigate oscillation in decoder optimization, we develop a Joint Cross-Attention Reconstruction (JCAR) strategy that jointly reconstructs coupled attention branches, suppressing oscillatory behavior and promoting stable convergence. Extensive experiments show that CAR-SAM robustly quantizes SAM models down to 4-bit precision, surpassing existing methods by 14.6% and 6.6% mAP on SAM-B and SAM-L respectively.", "authors": ["Houji Wen", "Jiangyong Yu", "Jun Li", "Dawei Yang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-16", "url": "https://arxiv.org/abs/2605.16901", "pdf_url": "https://arxiv.org/pdf/2605.16901v1", "arxiv_id": "2605.16901", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "7e991b6880e0f795c6ddea9e41ad58e78fe5ed28caae75dee6a0ba76bee1d0c5", "sources": ["arxiv", "semantic_scholar"], "title": "PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources", "abstract": "For a Gaussian source under mean-squared error (MSE), classical transform coding is rate--distortion (RD) optimal: the Karhunen--Loeve transform (KLT) diagonalizes the covariance, reverse waterfilling allocates the bits, and scalar quantization closes the loop. This elegant story breaks down for multimodal sources, where no single covariance can capture heterogeneous local geometries, and the RD function loses its closed form. We revisit this problem through Gaussian-mixture sources and develop a constructive RD theory for them. Our key finding is that the mixture structure incurs only a component label cost. Conditioned on the active mixture component, each branch is Gaussian; the challenge is allocating bits across heterogeneous branches. We prove that the genie-aided conditional RD function is governed by a single global reverse-waterfilling level shared across all components and eigenmodes. Building on this result, we introduce PrismQuant, which transmits the component label losslessly and encodes the residual using the component-matched KLT, followed by scalar quantization, achieving a rate of H(C)/n bits per source dimension of the converse, with a vanishing asymptotic gap. We further develop a practical implementation based on EM-driven Gaussian-mixture learning, component-adaptive KLTs, and entropy-constrained scalar quantization (ECSQ). Experiments on synthetic Gaussian mixtures show that PrismQuant closely approaches the theoretical RD bound, while experiments on real-world channel-state-information (CSI) data demonstrate competitive or superior performance compared with transformer-based learned codecs at more than one order of magnitude smaller model size.", "authors": ["Bumsu Park", "Chanho Park", "Youngmok Park", "Namyoon Lee"], "categories": ["cs.IT", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-15", "url": "https://arxiv.org/abs/2605.15507", "pdf_url": "https://arxiv.org/pdf/2605.15507v1", "arxiv_id": "2605.15507", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e266b4d7f54c9cb2a1e88d5ee4d105356de948e9bb6ae36bc46ddf0842f9c6ec", "sources": ["arxiv", "semantic_scholar"], "title": "UB-SMoE: Universally Balanced Sparse Mixture-of-Experts for Resource-adaptive Federated Fine-tuning of Foundation Models", "abstract": "Heterogeneous LoRA-rank methods address system heterogeneity in federated fine-tuning of foundation models by assigning client-specific ranks based on computational capabilities. However, these methods achieve only marginal computational savings, as dense feed-forward computations dominate. Sparse Mixture-of-Experts (SMoE) provides a promising alternative through conditional computation, yet we identify that its naive application to heterogeneous federated settings introduces two critical discordances: (i) expert utilization imbalance and (ii) non-differentiability of Top-K routing. Our convergence analysis demonstrates that these discordances lead to degraded convergence, particularly for resource-constrained clients. To address these challenges, we propose Universally Balanced Sparse Mixture-of-Experts (UB-SMoE), which introduces Dynamic Modulated Routing (DMR) to rebalance expert utilization, and Universal Pseudo-Gradient (PG) to reconstruct learning signals for non-activated experts. These mechanisms form a self-reinforcing cycle that maintains expert viability across heterogeneous clients. Experiments on benchmarks show that UB-SMoE achieves up to $45.0\\%$ computational reduction on low-resource clients while improving their performance by $8.7 \\times$ compared to existing heterogeneous LoRA-rank methods.", "authors": ["Van-Tuan Tran", "Hong-Hanh Nguyen-Le", "Marco Ruffini", "Merim Dzaferagic"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-15", "url": "https://arxiv.org/abs/2605.16690", "pdf_url": "https://arxiv.org/pdf/2605.16690v1", "arxiv_id": "2605.16690", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "79a1de04fa02135fe52b9a403dc8e69cab871ceb4f3c1db6ab090d9953a1a67a", "sources": ["arxiv", "semantic_scholar"], "title": "HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling", "abstract": "Visual Autoregressive (VAR) models have recently demonstrated impressive image generation quality while maintaining low latency. However, they suffer from severe KV-cache memory constraints, often requiring gigabytes of memory per generated image. We introduce HeatKV, a novel compression method that adapts cache allocation in each head based on its attention to previously generated scales. Using a small offline calibration set, the attention heads are ranked according to their attention scores over prior scales. Based on this ranking, we construct a static pruning schedule tailored to a given memory budget. Applied to the Infinity-2B model, HeatKV achieves $2 \\times$ higher compression ratio in memory allocation for KV cache compared to existing methods, while maintaining similar or better image fidelity, prompt alignment and human perception score. Our method achieves a new state-of-the-art (SOTA) for VAR model KV-cache compression, showcasing the effectiveness of fine-grained, head-specific cache allocation.", "authors": ["Jonathan Cederlund", "Axel Berg", "Durmus Alp Emre Acar", "Chuteng Zhou", "Pontus Giselsson"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.14877", "pdf_url": "https://arxiv.org/pdf/2605.14877v1", "arxiv_id": "2605.14877", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "22e44138f60f689ab135f54fedccf041a78538f1ce546105b0d75c3ad5d18b48", "sources": ["arxiv", "semantic_scholar"], "title": "RQ-MoE: Residual Quantization via Mixture of Experts for Efficient Input-Dependent Vector Compression", "abstract": "Vector quantization is a fundamental tool for compressing high-dimensional embeddings, yet existing multi-codebook methods rely on static codebooks that limit expressiveness under heterogeneous data geometry. While recent dynamic quantizers like QINCo adapt codebooks to individual inputs and improve expressiveness, their strict sequential dependencies create decoding bottlenecks. We propose Residual Quantization via Mixture of Experts (RQ-MoE), a framework combining a two-level MoE with dual-stream quantization to enable input-dependent codebook adaptation for efficient vector quantization. RQ-MoE enables dynamic codebook construction and decouples instruction from quantization, facilitating parallel decoding. Theoretically, we show that standard Residual Quantization and QINCo can be recovered as constrained special cases of RQ-MoE, and derive a guideline for setting expert dimensionality in RQ-MoE. Extensive experiments show that RQ-MoE achieves state-of-the-art or on-par performance in reconstruction and retrieval, while providing 6x-14x faster decoding than prior vector quantization methods. The implementation is available at https://github.com/KDEGroup/RQ-MoE.", "authors": ["Zhengjia Zhong", "Shuyan Ke", "Zaizhou Lin", "Jiaqi Song", "Hongyi Lan", "Hui Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.14359", "pdf_url": "https://arxiv.org/pdf/2605.14359v1", "arxiv_id": "2605.14359", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/KDEGroup/RQ-MoE", "venue": null, "quality_score": 0.65}
{"id": "de4cb78a0f87292b660b6f20f9e6cc9aa5209c7007e90147cecb7fea456b7053", "sources": ["arxiv", "semantic_scholar"], "title": "Minimal-Intervention KV Retention via Set-Conditioned Diversity", "abstract": "KV-cache compression at small budgets is a crowded design space spanning cache representation, head-wise routing, compression cadence, decoding behavior, and within-budget scoring. We study seven mechanisms across these five families under matched mean cache on long-form mathematical reasoning (MATH-500~\\cite{hendrycks2021math}) with two distilled-reasoning models (Qwen-7B and Llama-8B variants of DeepSeek-R1-Distill~\\cite{deepseek2025r1}) at budgets $b \\in \\{64, 128\\}$. All seven were rejected. We then propose $α$, a one-function modification to the TriAttention~\\cite{mao2026triattention} retention scorer that replaces argmax-top-$k$ with greedy facility-location-inspired selection under a V-space redundancy penalty controlled by a single weight $λ$. A pre-registered protocol tunes $λ$ on a frozen development split and confirms on a disjoint held-out split; with $λ= 0.5$, $α$ clears Bonferroni on two of the four (model, budget) cells (Qwen $b{=}128$ and Llama $b{=}64$), no cell is significantly negative, and the pre-registered Branch~A triggers. The finding is asymmetric: a minimal scoring modification beat heavier structural redesigns in this regime, and the combined matched-memory, sympy-graded, held-out confirmation protocol is the evidence standard that made the asymmetry visible.", "authors": ["Libo Sun", "Po-wei Harn", "Peixiong He", "Xiao Qin"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.14292", "pdf_url": "https://arxiv.org/pdf/2605.14292v2", "arxiv_id": "2605.14292", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/libophd/minimal-kv-retention", "venue": null, "quality_score": 0.65}
{"id": "2d086aaab2d7969465dd97122c28ad7160bf7319eee16e1faae78dcf6eeeb676", "sources": ["arxiv", "semantic_scholar"], "title": "KVCapsule: Efficient Sequential KV Cache Compression for Vision-Language Models with Asymmetric Redundancy", "abstract": "Vision-Language Models (VLMs) have emerged as a critical and fast-growing extension of Large Language Models (LLMs) that enable multimodal reasoning through both text and image inputs. Although VLMs enrich the capabilities of language models, they also inherit and amplify key computational bottlenecks: the memory overhead caused by the large key-value (KV) cache during autoregressive decoding. This challenge is particularly severe in VLMs, where images produce longer token sequences and denser feature representations compared to text. Moreover, the spatial and information-rich nature of vision tokens introduces structured attention patterns that make many LLM-oriented KV cache compression techniques ineffective when applied directly to VLMs. In this work, we conduct a detailed empirical analysis of the behavior of vision tokens, highlighting the critical differences from purely text-based models. Based on these insights, we propose KVCapsule, a novel KV cache compression framework for vision tokens. KVCapsule keeps the pretrained VLM backbone frozen, requires no modification to the attention computation modules, and can be integrated into existing VLMs through lightweight compression and reconstruction components. We evaluate KVCapsule on multiple VLMs and benchmark tasks, demonstrating up to 2x improvement in TPS and 2.4x reduction in KV cache memory at a 60% compression ratio, with negligible degradation in accuracy or response quality. Our findings offer practical pathways to scale VLM inference under constrained memory budgets and inspire further research into structure-aware cache compression for multimodal models.", "authors": ["Yingbing Huang", "Tharun Adithya Srikrishnan", "Steven K. Reinhardt", "Deming Chen"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-14", "url": "https://arxiv.org/abs/2605.16439", "pdf_url": "https://arxiv.org/pdf/2605.16439v1", "arxiv_id": "2605.16439", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "48d38c52e35fa8fada8a0bcd667603d0e83af3923e86b9ecca7fb223093b37cc", "sources": ["arxiv", "semantic_scholar"], "title": "Self-Pruned Key-Value Attention: Learning When to Write by Predicting Future Utility", "abstract": "Under modern test-time compute and agentic paradigms, language models process ever-longer sequences. Efficient text generation with transformer architectures is increasingly constrained by the Key-Value cache memory footprint and bandwidth. To address this limitation, we introduce Self-Pruned Key-Value Attention (SP-KV), a mechanism designed to predict future KV utility in order to reduce the size of the long-term KV cache. This strategy operates at a fine granularity: a lightweight utility predictor scores each key-value pair, and while recent KVs are always available via a local window, older pairs are written in the cache and used in global attention only if their predicted utility surpasses a given threshold. The LLM and the utility predictor are trained jointly end-to-end exclusively through next-token prediction loss, and are adapted from pretrained LLM checkpoints. Rather than enforcing a fixed compression ratio, SP-KV performs dynamic sparsification: the mechanism adapts to the input and typically reduces the KV cache size by a factor of $3$ to $10\\times$, longer sequences often being more compressible. This leads to vast improvements in memory usage and decoding speed, with little to no degradation of validation loss nor performance on a broad set of downstream tasks. Beyond serving as an effective KV-cache reduction mechanism, our method reveals structured layer- and head-specific sparsity patterns that we can use to guide the design of hybrid local-global attention architectures.", "authors": ["Gergely Szilvasy", "Manuel Faysse", "Maria Lomeli", "Matthijs Douze", "Pierre-Emmanuel Mazaré", "Loïc Cabannes", "Wen-tau Yih", "Hervé Jégou"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-13", "url": "https://arxiv.org/abs/2605.14037", "pdf_url": "https://arxiv.org/pdf/2605.14037v1", "arxiv_id": "2605.14037", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "6f1b6ed259c3c971abe47f58a497ec62699aac716fb598bbc93f81376bbfad8a", "sources": ["arxiv", "semantic_scholar"], "title": "SPHERICAL KV: Angle-Domain Attention and Rate-Distortion Retention for Efficient Long-Context Inference", "abstract": "Long-context inference is increasingly constrained by the KV cache: resident memory grows with context length, and decoding becomes limited by repeated High Bandwidth Memory (HBM) streaming rather than arithmetic. Existing methods such as eviction, windowing, quantization, and offloading reduce footprint, but often leave the critical-path bottleneck only partially addressed, especially when compressed states must still be reconstructed into dense vectors during decoding. We present Spherical KV, a long-context inference method that treats KV allocation as a rate-distortion problem grounded in attention geometry for efficient decoding. The method is built on two ideas: (i) represent directional information cheaply in the decode hot loop, and (ii) allocate retention and precision according to estimated future utility. Its first component, Angle-Domain Attention (ADA), stores keys in a spherical parameterization consisting of a scalar radius and compact angle codes, and computes attention logits directly from these codes without reconstructing dense keys. This preserves a paged, block-local, fusion-friendly decode path and directly targets HBM traffic in realistic serving settings. Its second component, Rate-Distortion Retention (RDR), jointly chooses keep/drop decisions and precision tiers per token and head under a fixed budget, producing tier-homogeneous pages with lightweight metadata and coalesced reads. Together, ADA and RDR provide a deployment-oriented mechanism for reducing KV residency while preserving decode efficiency.", "authors": ["Anay Chauhan", "Gurucharan Marthi Krishna Kumar", "Arion Das", "Amit Dhanda", "Vinija Jain", "Aman Chadha", "Amitava Das"], "categories": ["cs.LG", "cs.CL", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-13", "url": "https://arxiv.org/abs/2605.18856", "pdf_url": "https://arxiv.org/pdf/2605.18856v3", "arxiv_id": "2605.18856", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a6770ec8df8337956d27e296bb66489720f2535c356bab623db4fdbb4bbf49b8", "sources": ["arxiv", "semantic_scholar"], "title": "HodgeCover: Higher-Order Topological Coverage Drives Compression of Sparse Mixture-of-Experts", "abstract": "Sparse Mixture-of-Experts (MoE) layers route tokens through a handful of experts, and learning-free compression of these layers reduces inference cost without retraining. A subtle obstruction blocks every existing compressor in this family: three experts can each be pairwise compatible yet form an irreducible cycle when merged together, so any score that ranks experts on pairwise signals is structurally blind to which triples are jointly mergeable. We show the obstruction is a precise mathematical object, the harmonic kernel of the simplicial Laplacian on a 2-complex whose vertices are experts, whose edges carry KL merge barriers, and whose faces carry triplet barriers; Hodge-decomposing the edge-barrier signal isolates the kernel exactly. We turn the diagnostic into a selection objective: HodgeCover greedily covers the harmonic-critical edges and triplet-critical triangles, and a hybrid variant of HodgeCover pairs it with off-the-shelf weight pruning on survivors. On three open-weight Sparse MoE backbones under aggressive expert reduction, HodgeCover matches state-of-the-art learning-free baselines on the expert-reduction axis, leads on the aggressive-compression frontier of the hybrid axis, and uniquely balances retained mass across all four Hodge components. These results show that exposing the harmonic kernel of a learned MoE structure changes which compressor wins at the regime that matters most.", "authors": ["Tao Zhong", "Dongzhe Zheng", "Christine Allen-Blanchette"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-13", "url": "https://arxiv.org/abs/2605.13997", "pdf_url": "https://arxiv.org/pdf/2605.13997v1", "arxiv_id": "2605.13997", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8128ded7225a1a27c081c62f71541e2429ac30c0658b6ac7f99df16de39fb0f3", "sources": ["arxiv", "semantic_scholar"], "title": "FibQuant: Universal Vector Quantization for Random-Access KV-Cache Compression", "abstract": "Long-context inference is increasingly a memory-traffic problem. The culprit is the key--value (KV) cache: it grows with context length, batch size, layers, and heads, and it is read at every decoding step. Rotation-based scalar codecs meet this systems constraint by storing a norm, applying a shared random rotation, and quantizing one coordinate at a time. They are universal and random-access, but they discard the geometry created by the normalization step. After a Haar rotation, a block of $k$ consecutive coordinates is not a product source; it is a spherical-Beta source on the unit ball. We introduce \\textsc{FibQuant}, a universal fixed-rate vector quantizer that keeps the same normalize--rotate--store interface while replacing scalar tables by a shared radial--angular codebook matched to this canonical source. The codebook combines Beta-quantile radii, Fibonacci\\,/\\,Roberts--Kronecker quasi-uniform directions, and multi-restart Lloyd--Max refinement. We prove that the resulting vector code strictly improves on its scalar product specialization at matched rate, with a high-rate gain that separates into a cell-shaping factor and a density-matching factor. The same construction gives a dense rate axis, including fractional-bit and sub-one-bit operating points, without calibration or variable-length addresses. On GPT-2 small KV caches, \\textsc{FibQuant} traces a memory--fidelity frontier from $5\\times$ compression at $0.99$ attention cosine similarity to $34\\times$ at $0.95$. End-to-end on TinyLlama-1.1B, it is within $0.10$ perplexity of fp16 at $4\\times$ compression and has $3.6\\times$ lower perplexity than scalar \\textsc{TurboQuant} at $b = 2$ ($8\\times$ compression), where scalar random-access quantization begins to fail.", "authors": ["Namyoon Lee", "Yongjune Kim"], "categories": ["cs.AI", "cs.IT", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.11478", "pdf_url": "https://arxiv.org/pdf/2605.11478v1", "arxiv_id": "2605.11478", "doi": null, "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "bcfcc617c02abd31e07bc7d8e212bb8241aeae54de852d2e05fd6a0387c42cc4", "sources": ["arxiv", "semantic_scholar"], "title": "KV-Fold: One-Step KV-Cache Recurrence for Long-Context Inference", "abstract": "We introduce KV-Fold, a simple, training-free long-context inference protocol that treats the key-value (KV) cache as the accumulator in a left fold over sequence chunks. At each step, the model processes the next chunk conditioned on the accumulated cache, appends the newly produced keys and values, and passes the enlarged cache forward; the same one-step update is applied repeatedly, analogous to foldl in functional programming. Building on the KV cache concatenation primitive introduced for latent multi-agent communication, we repurpose it as a chunk-to-chunk recurrence for long-context inference. When processing chunk t, the model attends to the KV cache carried from earlier chunks as a prefix, reusing its internal state across segments without modifying or retraining the model. Despite its simplicity, the induced recurrence is stable: per-step drift rises briefly and then saturates into a flat plateau that persists across deep chains. This plateau is insensitive to a 10,000x change in numerical precision, robust across chunk sizes, and consistent across model families. At the task level, KV-Fold preserves exact information over long distances. On a needle-in-a-haystack benchmark, it achieves 100% exact-match retrieval across 152 trials spanning contexts from 16K to 128K tokens and chain depths up to 511 on Llama-3.1-8B, while remaining within the memory limits of a single 40GB GPU. Compared to streaming methods, which trade fidelity for bounded memory, KV-Fold maintains long-range retrieval while operating as a sequence of tractable forward passes. Overall, our results show that frozen pretrained transformers already support a stable form of KV-cache recurrence, providing a practical route to long-context inference without architectural changes or training.", "authors": ["Alireza Nadali", "Patrick Cooper", "Ashutosh Trivedi", "Alvaro Velasquez"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.12471", "pdf_url": "https://arxiv.org/pdf/2605.12471v1", "arxiv_id": "2605.12471", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "53946a4423636d6a9edc4ab219a84e30d8bd0c61f533cd8726bd11c4309eb6e3", "sources": ["arxiv", "semantic_scholar"], "title": "MuonQ: Enhancing Low-Bit Muon Quantization via Directional Fidelity Optimization", "abstract": "The Muon optimizer has emerged as a compelling alternative to Adam for training large language models, achieving remarkable computational savings through gradient orthogonalization. However, Muon's optimizer state is more sensitive to quantization errors: because the orthogonalization discards the magnitudes of singular values and retains only directional information, even small quantization errors in singular vector directions are amplified in the update. In this work, we propose MuonQ, a low-bit Muon training framework built on the principle of directional fidelity optimization. First, we apply a pre-quantization normalization so that each step introduces quantization errors of the same magnitude, preventing the accumulated error from developing a preferred direction. Second, we introduce a structural decomposition that separately quantizes the dominant singular components via power iteration, ensuring that quantization errors perturb only singular value magnitudes rather than rotating singular vector directions. Third, we adopt $μ$-law companding quantization to allocate higher resolution to densely packed momentum values, shifting the quantization objective from outlier preservation to dense-region distinguishability. Together, these techniques enable stable 4-bit quantization of Muon's optimizer states. Pre-training experiments on GPT-style and LLaMA-style models demonstrate that MuonQ at 4-bit precision closely matches full-precision Muon in both training loss and downstream task accuracy, while reducing optimizer state memory by up to 7.3 $\\times$. Our code is available at https://github.com/YupengSu/MuonQ.", "authors": ["Yupeng Su", "Ruijie Zhang", "Ziyue Liu", "Yequan Zhao", "Zheng Zhang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.11396", "pdf_url": "https://arxiv.org/pdf/2605.11396v1", "arxiv_id": "2605.11396", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/YupengSu/MuonQ", "venue": null, "quality_score": 0.65}
{"id": "d8022920303221a0fbcb7af981e3c3546afb5bc63f2c619b16ad50107db3a5e4", "sources": ["arxiv", "semantic_scholar"], "title": "Slicing and Dicing: Configuring Optimal Mixtures of Experts", "abstract": "Mixture-of-Experts (MoE) architectures have become standard in large language models, yet many of their core design choices - expert count, granularity, shared experts, load balancing, token dropping - have only been studied one or two at a time over narrow configuration ranges. It remains an open question whether these choices can be optimized independently, without considering interactions. We present the first systematic study of over 2,000 pretraining runs spanning models up to 6.6B total parameters, in which we exhaustively vary total experts, expert dimension, heterogeneous expert sizing within a single layer, shared expert size and load-balancing mechanisms. We find that at every active-parameter scale that we study, performance consistently improves with total MoE parameters even at extreme active expert parameter ratios like 128.Further, the optimal expert size is nearly invariant to total parameter count and depends only on active parameter count. Third, we see that other choices like shared experts, heterogeneous experts and load-balancing settings have small effects relative to expert count and granularity, although dropless routing yields a consistent gain. Overall, our results suggest a simpler recipe: focus on expert count and granularity, other choices have minimal effect on final quality.", "authors": ["Margaret Li", "Sneha Kudugunta", "Danielle Rothermel", "Luke Zettlemoyer"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.11689", "pdf_url": "https://arxiv.org/pdf/2605.11689v1", "arxiv_id": "2605.11689", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1fad0cb93aaa669ab230021fef265de93672f8b0a0afdef8d6262fe91d8051cf", "sources": ["arxiv", "semantic_scholar"], "title": "Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts", "abstract": "Sparse Mixture-of-Experts (SMoE) models enable scaling language models efficiently, but training them remains challenging, as routing can collapse onto few experts and auxiliary load-balancing losses can reduce specialization. Motivated by these hurdles, we study how routing decisions in SMoEs are formed mechanistically. First, we reveal a geometric coupling between routers and their corresponding experts. For a given token, the router weights for the selected expert and the expert weights processing it receive gradients along the same input direction, differing only in scalar coefficients. Thus, matched router--expert directions accumulate the same routed token history. This theoretical coupling also appears empirically in routing dynamics. In a $1$B SMoE trained from scratch, higher router scores predict stronger expert neuron activations, showing that routing decisions are mirrored inside the selected expert. Next, we analyze the effects of auxiliary load balancing on the router--expert geometric coupling, showing that such losses break this structure by spreading input-directed gradients across router weights, making distinct router directions nearly three times more similar to each other. Last, we demonstrate the centrality of geometric coupling for effective routing with a parameter-free online K-Means router, in which each expert maintains a running average of the hidden states routed to it and tokens are assigned based on cosine similarity. Compared with auxiliary-loss and loss-free balancing, this router achieves the lowest load imbalance with only a modest perplexity increase, indicating that geometric coupling captures a substantial part of what the router learns. Overall, our results explain how routers form assignment geometry that supports an effective division of labor.", "authors": ["Sagi Ahrac", "Noya Hochwald", "Mor Geva"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.12476", "pdf_url": "https://arxiv.org/pdf/2605.12476v1", "arxiv_id": "2605.12476", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f9c089764481a6d27d516a86523ca8f9410011670285dc2bd4412e0d855c835e", "sources": ["arxiv", "semantic_scholar"], "title": "HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models", "abstract": "Low-Rank Adaptation (LoRA) dominates parameter-efficient fine-tuning of large language models, yet most variants target dense architectures. Mixture-of-Experts (MoE) models scale parameters at near-constant per-token compute, and their sparse activation patterns create untapped opportunities for more efficient adaptation. We propose Hot-Experts Layer-level Low-Rank Adaptation (HELLoRA), which attaches LoRA modules only to the most frequently activated experts at each layer. This simple mechanism reduces trainable parameters and adapter-induced FLOPs while improving downstream performance, an effect we attribute to a form of structured regularization that preserves pretrained expert specialization. To stress-test HELLoRA under extreme parameter budgets, we further compose it with LoRI to form HELLoRI, which freezes the up-projection and sparsifies the down-projection. Across three MoE backbones, namely OlMoE-1B-7B, Mixtral-8x7B, and DeepSeekMoE, and three task families covering mathematical reasoning, code generation, and safety alignment, HELLoRA consistently outperforms strong PEFT baselines. Relative to vanilla LoRA on OlMoE, HELLoRA uses 15.7% of the trainable parameters, reduces adapter FLOPs by 38.7%, achieves 1.9x the training throughput, and improves accuracy by 9.2%. On DeepSeekMoE, HELLoRA outperforms LoRA while using only 23.2% of its trainable parameters. These results demonstrate that activation-aware adapter placement is an effective and practical route to scaling PEFT for MoE language models.", "authors": ["Jia Wei", "Zhonghao Zhang", "Ping Chen", "Qianyang li", "Yancheng Pan", "Shaoxun Wang", "Ziyi Qiu", "Longxiang Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-11", "url": "https://arxiv.org/abs/2605.18795", "pdf_url": "https://arxiv.org/pdf/2605.18795v1", "arxiv_id": "2605.18795", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8fd0301b31d5c39b0b0f68db944601020aaa6a4a40ee68cdbf221cc77dd7661d", "sources": ["arxiv", "semantic_scholar"], "title": "Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions", "abstract": "Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.", "authors": ["Diancheng Kang", "Zheyuan Liu", "Ningshan Ma", "Yue Huang", "Zhaoxuan Tan", "Meng Jiang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-11", "url": "https://arxiv.org/abs/2605.10664", "pdf_url": "https://arxiv.org/pdf/2605.10664v2", "arxiv_id": "2605.10664", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "35e3d5409c1b8d88fdb9252db20dc7c6c96264a0455c756ec9ebf43c9684f7ab", "sources": ["arxiv", "semantic_scholar"], "title": "Sieve: Dynamic Expert-Aware PIM Acceleration for Evolving Mixture-of-Experts Models", "abstract": "Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models (LLMs). However, the execution characteristics of MoE inference are changing rapidly and increasingly mismatch the assumptions underlying existing Processing-in-Memory (PIM) systems. Prior PIM systems for LLMs rely on static rules to offload memory-bound operations to PIM, without accounting for the combined effects of load imbalance and inter-GPU communication. Meanwhile, modern MoE models activate fewer experts out of increasingly many, creating a bimodal expert distribution: a small set of experts receives many tokens, while a long tail of experts receives only one or a few. We identify a trend in modern MoE models toward increasingly bimodal token-to-expert distributions, quantify the resulting disparity in arithmetic intensity across experts, and show that this disparity dramatically reduces the efficiency of state-of-the-art PIM systems for LLMs. To address this problem, we propose a scheduler for serving MoE models on multi-GPU systems with attached HBM-PIM stacks. Our scheduler partitions expert execution between GPU and PIM based on runtime token-to-expert distributions, while jointly considering interconnect overhead, memory bandwidth, GPU throughput, and PIM throughput. Moreover, we propose Sieve, a runtime framework that employs the scheduler to coordinate execution across GPUs and their attached HBM-PIM stacks. Sieve overlaps GPU computation, PIM computation, and intra- and inter-device communication while preserving cross-device dependencies induced by expert parallelism. Sieve is evaluated on our cycle-accurate simulator based on Ramulator 2.0. Compared to state-of-the-art PIM systems for MoE, Sieve improves both throughput and interactivity by 1.3x, 1.3x, and 1.6x on Qwen3.5-397B-A17B, GPT-OSS-120B, and Qwen3-30B-A3B, respectively.", "authors": ["Jungwoo Kim", "Rubens Lacouture", "Genghan Zhang", "Gina Sohn", "Qizheng Zhang", "Swapnil Gandhi", "Christos Kozyrakis", "Kunle Olukotun"], "categories": ["cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-11", "url": "https://arxiv.org/abs/2605.11277", "pdf_url": "https://arxiv.org/pdf/2605.11277v1", "arxiv_id": "2605.11277", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "6c234b723f60cccc7dcbfd6b58c914316ad10a9142c8b8af44c69c294a5e4dbc", "sources": ["arxiv", "semantic_scholar"], "title": "Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction", "abstract": "The key-value (KV) cache is a major bottleneck in long-context inference, where memory and computation grow with sequence length. Existing KV eviction methods reduce this cost but typically degrade performance relative to full-cache inference. Our key insight is that full-cache attention is not always optimal: in long contexts, irrelevant tokens can dilute attention away from useful evidence, so selective, learnable eviction can improve generation rather than merely approximate the full cache. We introduce a global retention-based KV eviction method that learns each token's future utility under a unified memory budget. Lightweight retention gates assign utility scores to cached KV entries, and a shared final scoring projection calibrates these scores across all layers and heads. This enables a single global eviction policy in which tokens from different layers, heads, and modalities compete directly for cache capacity. We further provide theoretical analysis showing that preferentially retaining useful tokens reduces attention dilution, and we justify geometric retention as a query-agnostic proxy for future utility. Across diverse long-context language and vision-language reasoning, and multi-turn dialogue benchmarks, our method substantially reduces KV memory while matching or surpassing full-cache inference. These results suggest that learned, globally calibrated KV eviction is not only a compression technique, but also a mechanism for improving long-context reasoning.", "authors": ["Ngoc Bui", "Hieu Trung Nguyen", "Arman Cohan", "Rex Ying"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09649", "pdf_url": "https://arxiv.org/pdf/2605.09649v1", "arxiv_id": "2605.09649", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "bc7f48feecf0ee44f3d58654535d2efbb096d0af6d252be3321501e986e806c1", "sources": ["arxiv", "semantic_scholar"], "title": "KV-RM: Regularizing KV-Cache Movement for Static-Graph LLM Serving", "abstract": "Static-graph LLM decoders provide predictable launches, fixed tensor shapes, and low submission overhead, but online decoding exposes highly irregular KV-cache behavior: request lengths differ, EOS events arrive asynchronously, and logical histories fragment over time. Dynamic runtimes recover flexibility through paged KV management and step-level scheduling, while static-graph executors often over-reserve memory and suffer burst-time latency outliers. This paper studies whether much of this variability can be absorbed below a fixed decode interface. We present KV-RM, a runtime design that regularizes KV-cache movement beneath a static-graph LLM decoder. KV-RM decouples logical KV histories from physical storage, tracks active KV state through a block pager, and materializes each decode step through a single committed descriptor. A merge-staged transport path coalesces non-contiguous KV mappings into a small number of large transfer groups before a fixed-shape attention kernel consumes them. Optional bounded far-history summaries can be enabled under the same interface, but the core design does not depend on them. On a 2-GPU NVIDIA A100 node, KV-RM improves mixed-length decoding throughput and tail latency relative to a static-graph baseline, reduces reserved KV memory across workload families, and removes severe burst-time latency spikes under production-trace replay. These results suggest that KV-cache movement, rather than kernel shape, can be an effective boundary for recovering runtime flexibility in static-graph LLM serving.", "authors": ["Zhiqing Zhong", "Zhijing Ye", "Jian Zhang", "Weijian Zheng", "Bolun Sun", "Xiaodong Yu"], "categories": ["cs.AR", "cs.AI", "cs.DC", "cs.OS"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09735", "pdf_url": "https://arxiv.org/pdf/2605.09735v1", "arxiv_id": "2605.09735", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8fefbaa8c4fadd6598b918357fb126556ec05ee67595a21685ae58d7231ed34c", "sources": ["arxiv", "semantic_scholar"], "title": "Forcing-KV: Hybrid KV Cache Compression for Efficient Autoregressive Video Diffusion Models", "abstract": "Autoregressive (AR) video diffusion models adopt a streaming generation framework, enabling long-horizon video generation with real-time responsiveness, as exemplified by the Self Forcing training paradigm. However, existing AR video diffusion models still suffer from significant attention complexity and severe memory overhead due to the redundant key-value (KV) caches across historical frames, which limits scalability. In this paper, we tackle this challenge by introducing KV cache compression into autoregressive video diffusion. We observe that attention heads in mainstream AR diffusion models exhibit markedly distinct attention patterns and functional roles that remain stable across samples and denoising steps. Building on our empirical study of head-wise functional specialization, we divide the attention heads into two categories: static heads, which focus on transitions across autoregressive chunks and intra-frame fidelity, and dynamic heads, which govern inter-frame motion and consistency. We then propose Forcing-KV, a hybrid KV cache compression strategy that performs structured static pruning for static heads and dynamic pruning based on segment-wise similarity for dynamic heads. While maintaining output quality, our method achieves a generation speed of over 29 frames per second on a single NVIDIA H200 GPU along with 30% cache memory reduction, delivering up to 1.35x and 1.50x speedups on LongLive and Self Forcing at 480P resolution, and further scaling to 2.82x speedup at 1080P resolution. Code and demo videos are provided at https://zju-jiyicheng.github.io/Forcing-KV-Page.", "authors": ["Yicheng Ji", "Zhizhou Zhong", "Jun Zhang", "Qin Yang", "XiTai Jin", "Ying Qin", "Wenhan Luo", "Shuiyang Mao", "Wei Liu", "Huan Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09681", "pdf_url": "https://arxiv.org/pdf/2605.09681v1", "arxiv_id": "2605.09681", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f15606148332ae4521077e8bcb37425078645a2ce42223c93cab139116962637", "sources": ["arxiv", "semantic_scholar"], "title": "TileQ: Efficient Low-Rank Quantization of Mixture-of-Experts with 2D Tiling", "abstract": "Mixture-of-Experts (MoE) models achieve remarkable performance by sparsely activating specialized experts, yet their massive parameters in experts pose significant challenges for deployment. While low-rank quantization offers a promising route to compress MoE models, existing methods still incur nonnegligible memory overhead and inference latency. To address these limitations, we propose \\textsc{TileQ}, a fine-tuning-free post-training quantization (PTQ) method that employs 2D-tiling structured low-rank quantization to share low-rank factors across both input and output dimensions of MoE experts. Furthermore, we introduce an efficient inference technique for \\textsc{TileQ} that fuses multiple low-rank expert computations into a single-pass operation, significantly improving hardware utilization. Experiments show that \\textsc{TileQ} cuts down additional memory usage up to 10$\\times$ and reduces inference latency to $\\sim$5\\% while preserving state-of-the-art accuracy.", "authors": ["Hongyaoxing Gu", "Xinzhe Chen", "Lijuan Hu", "Fangfang Liu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09281", "pdf_url": "https://arxiv.org/pdf/2605.09281v1", "arxiv_id": "2605.09281", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c2fe4ec54fbb65a6e2c902e0b14582037cc49e3f915d1e53b6122c0f053c9523", "sources": ["arxiv", "semantic_scholar"], "title": "Nectar: Neural Estimation of Cached-Token Attention via Regression", "abstract": "Evaluating softmax attention over a fixed long context requires reading every cached key-value pair for each new query token. For a given context (a book, a manual, a legal corpus) the attention output is a deterministic function of the query. We propose Nectar, which fits a compact neural network to this function for queries drawn from a task-relevant distribution. Nectar fits two networks per layer and KV-head: a target network that predicts the attention output and a score network that predicts the log-normalizer. The pair plugs into the standard masked self-attention at inference time, replacing the $O(n)$ attention over the cache with a forward pass whose cost does not depend on $n$. Each module carries on the order of $|θ|$ parameters per layer and KV-head, typically much smaller than the $2nd$ KV-cache footprint at the same granularity. We report experiments on models from 1.7B to 8B parameters across five long-context datasets. The approximation error tracks the next-token accuracy gap to full attention, and allocating capacity non-uniformly across layers reduces that gap in our ablation. Beyond this analysis of metrics, we check that the text generations (following a question prompt) of a model equipped with a Nectar module match in semantic content those obtained by giving the same model access to the full cache.", "authors": ["João Monteiro", "Michal Klein", "Pierre Ablin", "Marco Cuturi"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09778", "pdf_url": "https://arxiv.org/pdf/2605.09778v1", "arxiv_id": "2605.09778", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ce7b98360c8b1f91b32fd95b090d31bbfce5c2c0b85dc012224658e271cfd284", "sources": ["arxiv", "semantic_scholar"], "title": "Selection Plateau and a Sparsity-Dependent Hierarchy of Pruning Features", "abstract": "We identify a Selection Plateau phenomenon in one-shot neural network pruning: all rank-monotone weight scorers converge to identical accuracy at fixed sparsity, independent of functional form. We propose the Sparsity-Information-Complexity Spectrum (SICS) hypothesis: a sparsity-dependent minimum feature complexity kappa(S) governs plateau escape, with kappa=0 sufficient at low sparsity (S<0.65), kappa=1 dominant at critical sparsity (S~0.7), and kappa=2 necessary at extreme sparsity (S>0.75). On ViT-Small/CIFAR-10, testing nine feature classes across four sparsities, smooth non-monotone features provide +6.6% escape at S=0.7, while only raw features with high-frequency wiggle escape at S=0.8 (+2.6%). A fake non-monotone scorer underperforms the gradient baseline, indicating the requirement is magnitude-independent non-monotonicity. A handcrafted Gaussian bump achieves only +0.006 escape vs. chaos-derived +0.046, indicating rank-alignment is necessary but insufficient. SICS provides a unifying explanation for the performance clustering of diverse pruning methods and suggests that future selection algorithms should adapt feature complexity to target sparsity.", "authors": ["Guangqi Li", "Yongxin Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09345", "pdf_url": "https://arxiv.org/pdf/2605.09345v1", "arxiv_id": "2605.09345", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "08ba5799b816cdb76c27c7e1a7bf262308fae03f2704631f3ed8dbfee5384b61", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Layers with Hybrid Attention", "abstract": "Standard Mixture-of-Experts (MoE) transformers route tokens to expert subnetworks within each layer, but the layer structure itself remains monolithic. We introduce Mixture of Layers (MoL), which replaces full-width transformer blocks (d_model) with K parallel thin blocks at reduced dimensionality (d_thin << d_model), connected via learned down/up projections and composed via top-k block routing. Scaling sparse block routing to many blocks creates an attention coverage problem, as each block sees fewer tokens. We address this by introducing hybrid attention, which pairs one shared softmax block for global context with Gated DeltaNet linear attention in routed blocks.", "authors": ["Ivan Ternovtsii", "Yurii Bilak"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09516", "pdf_url": "https://arxiv.org/pdf/2605.09516v1", "arxiv_id": "2605.09516", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "88bf3f271eee0a177477df09c24e90d862135c8d92b2dfdcdab1fdee21d0b486", "sources": ["arxiv", "semantic_scholar"], "title": "Hierarchical Attention-based Graph Neural Network with Relevance-driven Pruning", "abstract": "Graph Neural Networks (GNNs) excel at relational reasoning but face two persistent challenges: the lack of interpretable attribution for heterogeneous node types, and the computational overhead of message passing over large, noisy graphs. We propose the Hierarchical Attention-based Heterogeneous GNN (HA-HeteroGNN), a framework that addresses both issues through a unied explainability-to-pruning pipeline. A two-tier attention mechanism separates sensor-level and context-level computation across 16 node types and 18 edge types, producing per-node relevance scores via an attention-based GNN Explainer without requiring gradient backpropagation. These relevance scores then serve as a principled pruning criterion: removing nodes identied as consistently uninformative yields a 27% reduction in graph edges while simultaneously improving classication accuracy by 2.46.1% across all model variants, challenging the conventional assumption that pruning necessarily trades accuracy for eciency. Experiments on a 50,000-record synthetic dataset spanning 11 report categories demonstrate 97.5% cross-strategy explanation stability and domain consistent sensor attribution, with training-time reductions of up to 43.9% and real-time inference latency of approximately 5860 ms per sample.", "authors": ["Seungwoo Kum"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-10", "url": "https://arxiv.org/abs/2605.09308", "pdf_url": "https://arxiv.org/pdf/2605.09308v1", "arxiv_id": "2605.09308", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "038d77f75c38287d8739132afa6d4a62a829e828bdd841d397eacd5c31a73998", "sources": ["arxiv", "semantic_scholar"], "title": "ReST-KV: Robust KV Cache Eviction with Layer-wise Output Reconstruction and Spatial-Temporal Smoothing", "abstract": "Large language models (LLMs) face growing challenges in efficient generative inference due to the increasing memory demands of Key-Value (KV) caches, especially for long sequences. Existing eviction methods typically retain KV pairs with high attention weights but overlook the impact of attention redistribution caused by token removal, as well as the spatial-temporal dynamics in KV selection. In this paper, we propose ReST-KV, a robust KV eviction method that combines layer-wise output Reconstruction and Spatial-Temporal smoothing to provide a more comprehensive perspective for the KV cache eviction task. Specifically, ReST-KV formulates KV cache eviction as an optimization problem that minimizes output discrepancies through efficient layer-wise reconstruction. By directly modeling how each token's removal affects the model output, our method naturally captures attention redistribution effects, going beyond simplistic reliance on raw attention weights. To further enhance robustness, we design exponential moving average smoothing to handle temporal variations and an adaptive window-based mechanism to capture spatial patterns. Our method, ReST-KV, significantly advances performance on long-context benchmarks. It surpasses state-of-the-art baselines by 2.58% on LongBench and 15.2% on RULER. Additionally, ReST-KV consistently outperforms existing methods on Needle-in-a-Haystack and InfiniteBench, all while achieving a remarkable 10.61$\\times$ reduction in decoding latency at 128k context length. The code is publicly available at https://github.com/an-yongqi/rest-kv to facilitate reproducibility and further research.", "authors": ["Yongqi An", "Chang Lu", "Kuan Zhu", "Tao Yu", "Chaoyang Zhao", "Hong Wu", "Ming Tang", "Jinqiao Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-09", "url": "https://arxiv.org/abs/2605.08840", "pdf_url": "https://arxiv.org/pdf/2605.08840v1", "arxiv_id": "2605.08840", "doi": null, "citation_count": 3, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/an-yongqi/rest-kv", "venue": null, "quality_score": 0.65}
{"id": "fac76883a08d91d74931089ebbdc17bb57dbbf886fca3c42aa590aae66a7e481", "sources": ["arxiv", "semantic_scholar"], "title": "Uncovering Intra-expert Activation Sparsity for Efficient Mixture-of-Expert Model Execution", "abstract": "Mixture of Experts (MoE) architecture has become the standard for state-of-the-art large language models, owing to its computational efficiency through sparse expert activation. However, sparsity through finer expert granularity is becoming increasingly difficult to achieve due to fundamental training challenges such as expert collapse and load imbalance. In this work, we explore and leverage intra-expert activation sparsity as a complementary and underexplored dimension of sparsity in MoE models. Surprisingly, substantial intra-expert sparsity is readily available in existing pre-trained MoE models, without any modification to the activation function or model parameters, providing up to 90% sparsity within each expert without significant accuracy loss. We explore intra-expert activation sparsity across eight off-the-shelf MoE models ranging from 1B to 400B parameters, and extend the MoE execution pipeline of vLLM to leverage intra-expert activation sparsity by skipping the computations of inactive neurons, on top of its existing optimizations, achieving up to 2.5 times speedup in MoE layer execution and 1.2 times end-to-end speedup compared to the original dense vLLM baseline.", "authors": ["Jongseok Park", "Sunga Kim", "Zhenyu Gu", "Ion Stoica", "Alvin Cheung"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-09", "url": "https://arxiv.org/abs/2605.08575", "pdf_url": "https://arxiv.org/pdf/2605.08575v1", "arxiv_id": "2605.08575", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "67a43bdc7c8f203f9c5fe355df3353b2e04c9d63026993af8d3e196fbc4b1e83", "sources": ["arxiv", "semantic_scholar"], "title": "Non-Monotonic Latency in Apple MPS Decoding: KV Cache Interactions and Execution Regimes", "abstract": "Autoregressive inference is typically assumed to scale predictably with decoding length, with latency increasing smoothly as generated sequence length grows. In this work, we identify unexpected non-monotonic latency behavior in the Apple MPS backend, where latency changes abruptly across nearby decoding configurations during transformer decoding. Using multiple model families (GPT-2, BLOOM, and OPT), we observe latency spikes of up to 21x within specific decoding-budget intervals, followed by recovery at neighboring configurations. Controlled experiments show that these anomalies originate primarily during the decode phase rather than prefill, are not explained by memory pressure alone, and remain absent on CPU and NVIDIA CUDA backends under identical conditions. We further show that key-value (KV) cache interacts strongly with these pathological execution regimes: KV caching remains beneficial overall, but its practical speedup collapses sharply within anomalous configurations, while cache-disabled decoding still exhibits residual non-monotonic behavior. These findings suggest that autoregressive decoding on MPS enters discrete execution regimes that are not captured by coarse-grained benchmarking, highlighting the importance of hardware-aware evaluation for long-context inference.", "authors": ["Willy Fitra Hendria"], "categories": ["cs.LG", "cs.AR", "cs.CL", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-09", "url": "https://arxiv.org/abs/2605.08913", "pdf_url": "https://arxiv.org/pdf/2605.08913v2", "arxiv_id": "2605.08913", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ce8ce17f0188ef2a753b99f103ca6557c2456714e3148908353e08f269c23042", "sources": ["arxiv", "semantic_scholar"], "title": "TAS-LoRA: Transformer Architecture Search with Mixture-of-LoRA Experts", "abstract": "Transformer architecture search (TAS) discovers optimal vision transformer (ViT) architectures automatically, reducing human effort to manually design ViTs. However, existing TAS methods suffer from the feature collapse problem, where subnets within a supernet fail to learn subnet-specific features, mainly due to the shared weights in a supernet, limiting the performance of individual subnets. To address this, we propose TAS-LoRA, a novel method that introduces parameter-efficient low-rank adaptation (LoRA) to enable subnet-specific feature learning, while maintaining computational efficiency. TAS-LoRA incorporates a Mixture-of-LoRAExperts (MoLE) strategy, where a lightweight router dynamically assigns LoRA experts based on subnet architectures, and introduces a group-wise router initialization technique to encourage diverse feature learning across experts early in training. Extensive experiments on ImageNet and several transfer learning benchmarks, including CIFAR-10/100, Flowers, CARS, and INAT-19, demonstrate that TAS-LoRA mitigates feature collapse effectively, improving performance over state-of-the-art TAS methods significantly.", "authors": ["Jeimin Jeon", "Hyunju Lee", "Bumsub Ham"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.07256", "pdf_url": "https://arxiv.org/pdf/2605.07256v1", "arxiv_id": "2605.07256", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b7b1968f5e9b1f6fd699308de001b3f7ba1ad0813575fa9239ccbd0bec44bfb7", "sources": ["arxiv", "semantic_scholar"], "title": "RDKV: Rate-Distortion Bit Allocation for Joint Eviction and Quantization of the KV Cache", "abstract": "Large language models (LLMs) have shown strong performance across diverse tasks, but their inference with long input contexts is bottlenecked by memory size and bandwidth. The Key-Value (KV) cache size grows linearly with sequence length and needs to be re-read from off-chip high-bandwidth memory (HBM) to on-chip memory at every decoding step, resulting in memory-bound inference. Existing methods reduce the cache by either eviction or quantization, but typically treat the two in isolation. In this paper, we cast KV cache compression as a rate-distortion problem, under which eviction and quantization are two end-points of the same bit allocation scheme. This exposes the need to optimize them jointly, motivating our method, RDKV (Rate-Distortion KV cache compression). RDKV derives the weight of each token or channel from the distortion that compression induces on the attention computation. Based on these weights, it assigns each token or channel a bit-width ranging from full precision down to zero bits guided by reverse water-filling, applied once after the prefilling stage. Experiments on LongBench, RULER, and InfiniteBench show that RDKV outperforms the best evaluated baseline by 9.1% on average. On LongBench it recovers 97.81% of full-cache accuracy with only 2.48% cache retention. Compared with full-cache FlashAttention-2 decoding, it achieves 4.5x decode speedup and 1.9x peak memory reduction with 128K context length, while maintaining comparable performance.", "authors": ["Junkai Zhang", "Hang Guo", "Luca Benini", "Yawei Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.08317", "pdf_url": "https://arxiv.org/pdf/2605.08317v1", "arxiv_id": "2605.08317", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d741a5207829d86a12593fce23234969216705b5a35d6a3820d77cccc4c37276", "sources": ["arxiv", "semantic_scholar"], "title": "Convergent Stochastic Training of Attention and Understanding LoRA", "abstract": "Transformers have revolutionized machine learning and deploying attention layers in the model is increasingly standard across a myriad of applications. Further, for large models, it is common to implement Low Rank Adaptation (LoRA), whereby a factorized parameterization of them is trained, to achieve a surprisingly beneficial accuracy-size trade-off. In this work, via a unified framework we rigorously establish trainability of such models under stochastic methods. We prove that for any mild regularization, the empirical regression loss on a attention layer and LoRA on a shallow neural net, both induce Poincaré inequality for the corresponding Gibbs' measure. Then it follows via invoking recent results that a certain SDE, which mimics the SGD, minimizes the corresponding losses. In both the cases, our first-of-its-kind results of trainability on attention and nets, do not rely on any assumptions on the data or the size of the architecture.", "authors": ["Zhengkai Sun", "Dibyakanti Kumar", "Alejandro F Frangi", "Anirbit Mukherjee", "Mingfei Sun"], "categories": ["cs.LG", "math.FA", "math.PR"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.07959", "pdf_url": "https://arxiv.org/pdf/2605.07959v1", "arxiv_id": "2605.07959", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "912b71072f83afb568e932721b73cd1949b5754c908108d47f0b277c1ce3f1c3", "sources": ["arxiv", "semantic_scholar"], "title": "Reformulating KV Cache Eviction Problem for Long-Context LLM Inference", "abstract": "Large language models (LLMs) support long-context inference but suffer from substantial memory and runtime overhead due to Key-Value (KV) Cache growth. Existing KV Cache eviction methods primarily rely on local attention weights, neglecting the influence of value representations, output projection, and inter-head interactions. In this work, we reformulate KV Cache eviction from a conventional head-wise, weight-averaging approach into an output-aware, layer-wise matrix multiplication approximation problem. We introduce LaProx, a novel eviction strategy that explicitly models the multiplicative interaction between attention maps and projected value states to accurately quantify token contributions while accounting for inter-head dependencies. Building on this metric, we propose the first unified eviction strategy that assigns globally comparable importance scores to tokens, enabling model-wide selection instead of local, head-wise decisions. Experimental results across 19 datasets on long-context benchmarks LongBench and Needle-In-A-Haystack demonstrate that our approach maintains model performance with only 5\\% of the KV cache and consistently outperforms prior works across all configurations. Notably, our method achieves up to 2$\\times$ accuracy loss reduction under extreme compression scenarios compared to existing state-of-the-art baselines with minimal overhead.", "authors": ["Tho Mai", "Joo-Young Kim"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.07234", "pdf_url": "https://arxiv.org/pdf/2605.07234v1", "arxiv_id": "2605.07234", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5d8a31f667cd98782886f8e1004f43d0b7f76e6cd0748682cb92f34f8e67b95c", "sources": ["arxiv", "semantic_scholar"], "title": "RcLLM: Accelerating Generative Recommendation via Beyond-Prefix KV Caching", "abstract": "Large Language Models (LLMs) are transforming recommendation from ranking into a generative task, but industrial deployment remains limited by the high latency of processing long, personalized prompts. Standard prefix caching provides limited benefit because reuse in recommendation workloads is often non-contiguous across user histories and item contexts. We present RcLLM, a distributed inference system for generative recommendation with Beyond-Prefix KV Caching. RcLLM decomposes prompts into reusable blocks and supports large item catalogs with a stratified distributed storage design: compact user-history caches are replicated for zero-latency retrieval, while massive item caches are sharded using similarity-aware placement. To reduce redundant quadratic attention computation, RcLLM combines an affinity-based global scheduler that improves data locality with a selective attention mechanism that corrects approximation errors. Experiments on real-world datasets show that RcLLM reduces Time-To-First-Token (TTFT) by 1.31x-9.51x compared with state-of-the-art prefix caching systems, enabling real-time serving with negligible impact on recommendation accuracy.", "authors": ["Zhan Zhao", "Yuxin Wang", "Amelie Chi Zhou"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.07443", "pdf_url": "https://arxiv.org/pdf/2605.07443v1", "arxiv_id": "2605.07443", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e1cdbe1d238b8f6e8822f5401f3c543b3247189d0db48598a3c37d0b48c9ebd5", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation", "abstract": "Recent literature on fine-tuning Large Language Models highlights a fundamental debate. While Full Fine-Tuning (FFT) provides the representational plasticity required for high-entropy knowledge injection, Low-Rank Adaptation (LoRA) can match or surpass FFT performance because many tasks only require updates in a low-rank space and benefit from LoRA's additional regularization. Through empirical evaluation across diverse tasks (SQL, Medical QA, and Counterfactual Knowledge) and varying language models (Gemma-3-1B, Qwen2.5-1.5B, and Qwen2.5-3B), we verify both trends and demonstrate that relying solely on either static architecture is structurally limited. To address this challenge, we propose a Mixture of LoRA and Full (MoLF) Fine-Tuning, a unified framework that enables continuous navigation between both training regimes. MoLF dynamically routes updates between FFT and LoRA at the optimizer level to ensure that exact gradient signals are available to both experts throughout training, yielding stable training dynamics. For memory-constrained environments, we also introduce MoLF-Efficient, which freezes base weights and only routes updates among a pair of LoRA experts of potentially varying rank. Our evaluations show that MoLF either improves on or stays within $1.5\\%$ of the better of FFT and LoRA across all settings, while MoLF-Efficient outperforms prior adaptive LoRA approaches by up to $20\\%$ on Fact and $9\\%$ on Med and SQL. Our code is open-sourced at https://github.com/11785T23/molf.git.", "authors": ["Haozhan Tang", "Xiuqi Zhu", "Xinyin Zhang", "Boxun Li", "Virginia Smith", "Kevin Kuo"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.07111", "pdf_url": "https://arxiv.org/pdf/2605.07111v2", "arxiv_id": "2605.07111", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/11785T23/molf.git", "venue": null, "quality_score": 0.65}
{"id": "e55ff3782da3e728c9b217c4e0cf90922d6cb2d4e52ecd54756d905ea20a94d6", "sources": ["arxiv", "semantic_scholar"], "title": "PaceVGGT: Pre-Alternating-Attention Token Pruning for Visual Geometry Transformers", "abstract": "Visual Geometry Transformer (VGGT) is a strong feed-forward model for multiple 3D tasks, but its Alternating-Attention (AA) stack scales quadratically in the total token count, making long clips expensive. Existing token-reduction accelerators operate inside AA, leaving the patch grid that enters AA uncompressed. We introduce PaceVGGT, a pre-AA token pruning framework that prunes DINO patch tokens before the first AA block of a frozen VGGT. PaceVGGT trains a lightweight Token Scorer that estimates per-token importance from DINO features. The scorer is first distilled against an AA-internal attention target from the unpruned backbone, then refined under downstream camera, depth, and point-map losses. A per-frame keep budget fixes the backbone-visible sequence length, while an importance-adaptive merge/prune assignment preserves residual content from high-saliency frames under a fixed total merge budget. A Feature-guided Restoration module reconstructs the dense spatial grid required by the prediction heads. On ScanNet-50 and 7-Scenes, PaceVGGT remains on the reconstruction quality--latency frontier while reducing inference latency. On ScanNet-50, it reduces latency by \\(5.1\\times\\) over unmodified VGGT at \\(N=300\\) and \\(1.47\\times\\) over LiteVGGT at \\(N=1000\\). These results identify pre-AA pruning as a viable acceleration route for frozen VGGT-style geometry transformers.", "authors": ["Haotang Li", "Zhenyu Qi", "Shaohan Henry Wang", "Kebin Peng", "Zi Wang", "Qing Guo", "Sen He", "Huanrui Yang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.08371", "pdf_url": "https://arxiv.org/pdf/2605.08371v1", "arxiv_id": "2605.08371", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1fc1067bcee9a2f92d682f5284e06b1aa6237f8e1d6cb66e4ca222724a6c1627", "sources": ["arxiv", "semantic_scholar"], "title": "SDG-MoE: Signed Debate Graph Mixture-of-Experts", "abstract": "Sparse MoE models achieve a good balance between capacity and compute by routing each token to a small subset of experts. However, in most MoE architectures, once a token is routed, the selected experts process it independently and their outputs are combined via a weighted sum. This leaves open whether enabling communication among them could improve performance. While prior work has raised this question, direct interaction among the active routed experts remains underexplored. In this paper, we propose SDG-MoE (Signed Debate Graph Mixture-of-Experts), a novel architecture that adds a lightweight, iterative deliberation step before final aggregation. SDG-MoE introduces three components: (i) two learned interaction matrices over the active experts, a support graph $A^+$ and a critique graph $A^-$, capturing reinforcing and corrective influences; (ii) a signed message-passing step that updates expert representations before aggregation; and (iii) a disagreement-gated Friedkin-Johnsen-style anchoring that controls deliberation strength while preventing expert drift. Together, these enable a structured deliberation process where interaction strength scales with disagreement and specialization is preserved. We also provide a theoretical analysis establishing stability conditions on expert states and showing that deliberation adds only low-order overhead over the active set. In controlled three-seed pretraining experiments, SDG-MoE improves validation perplexity over both an unsigned graph communication baseline and vanilla MoE, outperforming the strongest baseline by 19.8%, and gives the best external perplexity on WikiText-103, C4, and Paloma among the compared systems.", "authors": ["Stepan Kulibaba", "Kirill Labzin", "Artem Dzhalilov", "Roman Pakhomov", "Oleg Svidchenko", "Alexander Gasnikov", "Aleksei Shpilman"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.08322", "pdf_url": "https://arxiv.org/pdf/2605.08322v2", "arxiv_id": "2605.08322", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f481399e160c099ead5bb8154483676e307c4fcd79045bc7c46274fbda829724", "sources": ["arxiv", "semantic_scholar"], "title": "Hierarchical Mixture-of-Experts with Two-Stage Optimization", "abstract": "Sparse Mixture-of-Experts (MoE) models scale capacity by routing each token to a small subset of experts. However, their routers exhibit a fundamental trade-off: strong load balancing can suppress expert specialization, while aggressive diversity often causes routing collapse. We propose Hi-MoE, a grouped MoE framework that decomposes routing control into two coupled levels: (i) inter-group balancing that enforces fair traffic across expert groups, and (ii) intra-group specialization that promotes complementary expert behaviors while preventing within-group collapse. Our analysis provides a principled explanation of how our hierarchical objectives reshape the router, thereby promoting stable specialization and mitigating collapse. We observe consistent improvements over recent sparse-routing and grouped-MoE baselines across NLP and vision benchmarks, and confirm robustness via scaling studies (model size, expert count) and targeted ablations. In large-scale pre-training on 58B tokens, Hi-MoE-7B achieves a 5.6% perplexity reduction and a 40% improvement in expert balance over OLMoE-7B across diverse evaluation domains.", "authors": ["Gleb Molodtsov", "Alexander Miasnikov", "Aleksandr Beznosikov"], "categories": ["cs.LG", "cs.AI", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.08292", "pdf_url": "https://arxiv.org/pdf/2605.08292v1", "arxiv_id": "2605.08292", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a633d22994699e0be8dab405d558f163fdd5c9b9ab4f2785c337994bdc6b87f1", "sources": ["arxiv", "semantic_scholar"], "title": "When Are Experts Misrouted? Counterfactual Routing Analysis in Mixture-of-Experts Language Models", "abstract": "Mixture-of-Experts (MoE) language models route each token to a small subset of experts, but whether the routes selected by a trained top-$k$ router are good ones is rarely evaluated directly. Holding the model fixed, we compare each standard route against sampled equal-compute alternatives for the same token and score each by the next-token probability it assigns to the realized token in a verified reasoning trajectory. The result is sharply token-conditional: the standard router is well-aligned with route utility on confident tokens but uninformative on the fragile tokens that drive hard reasoning, where lower-loss equal-compute routes consistently exist inside the frozen model but are not selected. The same pattern holds across Qwen3-30B-A3B, GPT-OSS-20B, DeepSeek-V2-Lite, and OLMoE-1B-7B, and follows structurally from how standard top-$k$ training evaluates routing decisions: the language modeling loss scores only the executed route, and load balancing depends only on aggregate routing statistics. A minimal router-only update to the final-layer router, leaving every expert and every other router frozen, is sufficient to shift pass@K on AIME 2024+2025 and HMMT 2025 for both Qwen3-30B-A3B and GPT-OSS-20B, suggesting that at least part of the failure reflects router-reachable misallocation rather than expert capacity alone.", "authors": ["Youngsik Yoon", "Siwei Wang", "Wei Chen", "Jungseul Ok"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-08", "url": "https://arxiv.org/abs/2605.07260", "pdf_url": "https://arxiv.org/pdf/2605.07260v1", "arxiv_id": "2605.07260", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "0daa6e7065c213f51eae001af303220d9b0e727a651ee09b72c871b1d6c9ab37", "sources": ["arxiv", "semantic_scholar"], "title": "When Quantization Is Free: An int4 KV Cache That Outruns fp16 on Apple Silicon", "abstract": "KV-cache quantization is framed as a quality--latency trade-off. We show it is \\emph{inverted} on Apple Silicon's unified memory: a single fused Metal kernel (sign-randomized FFT $+$ per-channel $λ$ $+$ per-group abs-max $+$ int4 nibble pack), exposed as a HuggingFace \\texttt{Cache} subclass, runs \\emph{faster than fp16} across $256$--$4096$-token prefixes on Gemma-3 1B ($-3$ to $-8\\%$ ms/tok) and at short context on Qwen2.5-1.5B ($-0.7$ to $-2.6\\%$ through $1$K), with $3\\times$ persistent memory compression and quality preserved ($\\dPPL = 0.000$ Qwen short-prompt; $+3.6$ hook $\\dPPL$ Gemma). The kernel's $\\sim\\!25$\\,ns/vec overhead is below the bandwidth savings from $3\\times$ compression. The fused kernel also closes Qwen's 4-bit per-token catastrophe ($\\dPPL = +7975 \\to +638.6$, $12.5\\times$ reduction) at $182$\\,GFLOPS / $D{=}128$. Supporting findings: $\\SRFT$ and $\\SRHT$ are statistically indistinguishable for KV quality (we pick $\\SRFT$ for mixed-radix and matrix-multiply alignment); a learned-rotation ablation surfaces a regularization role for the fixed random SRFT base (learning $R+λ$ without SRFT lowers calibration MSE $84.9\\%$ vs $50.3\\%$ but yields worse PPL); Householder rotations at $k{=}d/2$ reflectors are effectively lossless at $d{=}256$.", "authors": ["Mohamed Amine Bergach"], "categories": ["cs.PF", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.05699", "pdf_url": "https://arxiv.org/pdf/2605.05699v1", "arxiv_id": "2605.05699", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a7802ffea6f857bf68abff0fdc7fd061b4aa4abd1dce11d82357ee929c034e2f", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Serving for Dynamic Agent Workflows with Prediction-based KV-Cache Management", "abstract": "LLM-based workflows compose specialized agents to execute complex tasks, and these agents usually share substantial context, allowing KV-Cache reuse to save computation. Existing approaches either manage KV-Cache at agent level and fail to exploit the reuse opportunities within workflows, or manage cache at the workflow level but assume that each workflow calls a static sequence of agents. However, practical workflows are typically dynamic, where the sequence of invoked agents and thus induced cache reuse opportunities depend on the context of each task. To serve such dynamic workflows efficiently, we build a system dubbed PBKV (\\textbf{P}rediction-\\textbf{B}ased \\textbf{KV}-Cache Management). For each workflow, PBKV predicts the agent invocations in several future steps by fusing the guidance from historical workflows and context of the target workflow. Based on the predictions, PBKV estimates the reuse potential of cache entries and keeps the high-potential entries in GPU memory. To be robust to prediction errors, PBKV utilizes the predictions conservatively during both cache eviction and prefetching. Experiments on three workflow benchmarks show that PBKV achieves up to $1.85\\times$ speedup over LRU on dynamic workflows, and up to $1.26\\times$ speedup over the SOTA baseline KVFlow on the static workflow.", "authors": ["Haoyu Zheng", "Fangcheng Fu", "Jia Wu", "Binhang Yuan", "Yongqiang Zhang", "Hao Wang", "Yuanyuan Zhu", "Xiao Yan", "Jiawei Jiang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06472", "pdf_url": "https://arxiv.org/pdf/2605.06472v1", "arxiv_id": "2605.06472", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d4cecf4958460d99720259fcad164479abe589df4a93cda9c178bf5179bd4ff5", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Attention as a Range Searching Problem: Towards an Inference-Efficient Index for KV Cache", "abstract": "Sparse attention improves LLM inference efficiency by selecting a subset of key-value entries, but at the cost of potential accuracy degradation. In particular, omitting critical KV entries can induce substantial errors in model outputs. Existing methods typically operate under fixed or adaptive token budgets and provide empirical robustness or partial theoretical guarantees, yet they do not ensure zero false negatives in decoding steps, particularly since the set of relevant tokens is both query- and step-dependent. Our empirical observations confirm that missing even one critical key can lead to sharp error spikes, especially in long reasoning tasks where the set of important tokens varies throughout decoding. This observation motivates the need for indexing methods that dynamically adapt to these variations across decoding steps while guaranteeing a full recall of the relevant keys above a certain threshold. We address this challenge by reformulating sparse attention as the halfspace range searching problem. However, existing range searching indices are not suitable for modern LLM inference due to their computational and implementation overheads. To overcome this, we introduce Louver, a novel index structure tailored for efficient KV cache retrieval. Louver (i) guarantees zero false negatives with respect to a specified threshold in both theory and practice, (ii) is lightweight to integrate into existing LLM pipelines, and (iii) incorporates hardware-aware optimizations for both CPU and GPU executions. Our experiments demonstrate that Louver outperforms prior sparse attention methods in both accuracy and runtime, and is faster than highly optimized dense attentions such as FlashAttention. These results highlight that recall guarantees are a critical and overlooked dimension of sparse attention, and open a new direction for building theoretically grounded, efficient KV cache indices.", "authors": ["Mohsen Dehghankar", "Abolfazl Asudeh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06763", "pdf_url": "https://arxiv.org/pdf/2605.06763v2", "arxiv_id": "2605.06763", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e65feed00415d0b9dc4fc9a2f23ceee2ebc0aceeaffce13946a23e526789b432", "sources": ["arxiv", "semantic_scholar"], "title": "Echo: KV-Cache-Free Associative Recall with Spectral Koopman Operators", "abstract": "Long chain-of-thought reasoning and agentic tool-calling produce traces spanning tens of thousands of tokens, yet Transformer KV caches grow linearly with sequence length, creating a memory bottleneck on commodity hardware. State-space models offer constant-memory recurrence but suffer a memory cliff: retrieval accuracy collapses once the gap between a stored fact and its query exceeds the effective horizon of the recurrent state. We introduce Echo, a KV-cache-free associative recall architecture built around Spectral Koopman Attention (SKA); a drop-in replacement for attention layers that augments SSM blocks with a closed-form dynamical operator whose sufficient statistics are accumulated in constant memory with no KV cache. Echo fits a spectral linear system to the key and value history via kernel ridge regression and retrieves through a learned power-iterated filter, all from $O(r^{2})$ streaming state where $r$ is a small projection rank. On the Multi-Query Associative Recall benchmark, a pure Mamba-2 SSM fails to exceed chance accuracy (${\\sim}3\\%$) across all gap lengths and KV-pair counts, while at the 50M parameter scale SKA-augmented models achieve $100\\%$ retrieval accuracy on every configuration tested, including distractor gaps of $4{,}096$ tokens with $32$ KV pairs. Across five additional transfer benchmarks including needle-in-a-haystack, tool-trace, and multi-hop retrieval, SKA consistently outperforms both pure SSM and SSM+Attention hybrids while maintaining constant inference memory. Ablations confirm that the spectral operator, not the prefix masking strategy, drives the retrieval gain.", "authors": ["Anupama Sridhar", "Alexander Johansen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06997", "pdf_url": "https://arxiv.org/pdf/2605.06997v1", "arxiv_id": "2605.06997", "doi": "10.1145/3786335.3813146", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "2db6f3b43b5b688cf2c9d109f71d4d6487ddde200afb16edd1a4d3c746855743", "sources": ["arxiv", "semantic_scholar"], "title": "Training Transformers for KV Cache Compressibility", "abstract": "Long-context language modeling is increasingly constrained by the Key-Value (KV) cache, whose memory and decode-time access costs scale linearly with the prefix length. This bottleneck has motivated a range of context-compression methods, from token-level summarization to recent optimization-based KV compression methods. These post-hoc methods operate on the KV cache of a fixed pretrained model, so their effectiveness is fundamentally limited by how well the model's internal representations can be compressed. In this work, we formalize the notion of KV compressibility and show that it is a property of the learned representations, rather than of the context alone. We prove that almost any sequence-to-vector function admits both highly compressible and inherently non-compressible transformer implementations, highlighting the need to guide transformers toward compressible representations during training. Motivated by this, we propose KV-Compression Aware Training (KV-CAT), a continued pretraining procedure that incentivizes the emergence of compressible representations. We introduce a train-time KV sparsification policy that masks KV slots during training. This forces the model to use fewer KV slots and encourages it to learn representations amenable to post-hoc compression. Empirically, we show that KV-CAT improves the quality-budget tradeoff of downstream compression methods across retrieval, long-context question answering, and perplexity-based evaluation of compressed-prefix continuation.", "authors": ["Yoav Gelberg", "Yam Eitan", "Michael Bronstein", "Yarin Gal", "Haggai Maron"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.05971", "pdf_url": "https://arxiv.org/pdf/2605.05971v2", "arxiv_id": "2605.05971", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e37b65cf181f796e0467ea84a16520a8be45667aca7164318a6de2078f0de25b", "sources": ["arxiv", "semantic_scholar"], "title": "Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs", "abstract": "Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at every layer and can clutter long interactions; activation steering is compact but typically weaker and does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. Rather than materializing reminder content throughout the prompt cache, MI treats steering as selective KV allocation, injecting latent slots only where the model routes to them. On matched personality-steering tasks, MI gives the best overall control--drift trade-off, remaining competitive with prompting while consistently outperforming CAA. On updateable guidance, MI supports mid-conversation behavior shifts without rewriting the visible transcript, achieving the highest post-shift alignment on Qwen3. On structured reasoning, MI outperforms visible prompting on HARDMath and PHYSICS (10/12 subject$\\times$mode cells), serving as proxies for structured reasoning in verifiable domains, while cutting content-matched KV storage by up to 118$\\times$. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.", "authors": ["Andy Zeyi Liu", "Michael Zhang", "Ilana Greenberg", "Adam Alnasser", "Lucas Baker", "John Sous"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06225", "pdf_url": "https://arxiv.org/pdf/2605.06225v2", "arxiv_id": "2605.06225", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "59cbd7622229b9044d0a902ff27f061956c624c3309d5d59c2b900c7fa31ab21", "sources": ["arxiv", "semantic_scholar"], "title": "Scene-Adaptive Continual Learning for CSI-based Human Activity Recognition with Mixture of Experts", "abstract": "Channel state information (CSI)-based human activity recognition (HAR) is vulnerable to performance degradation under domain shifts across varying physical environments. Continual learning (CL) offers a principled way to learn new domains sequentially while preserving past knowledge, but existing CL solutions for CSI-based HAR scale poorly with accumulating domains, rely on a large replay buffer, or incur linearly growing inference cost. In this letter, we propose Scene-Adaptive Mixture of Experts with Clustered Specialists (SAMoE-C), which formulates cross-domain CSI-based HAR as a mixture-of-experts system that enables scene-specific adaptation, via an attention-based semantic router that activates only selected experts for each input. Moreover, we develop a novel training protocol, which requires only a tiny replay buffer for stabilizing domain discrimination of the router. Experimental results on a four-scene CSI dataset demonstrate that SAMoE-C approaches the state-of-the-art accuracy, while maintaining a significantly lower inference cost. By jointly combining modular experts, selective activation with router and a lightweight training protocol, SAMoE-C enables scalable cross-domain CSI-based HAR deployment with low training overhead and high computational efficiency in real-world settings.", "authors": ["Wenhan Zheng", "Yuyi Mao", "Ivan Wang-Hei Ho"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06447", "pdf_url": "https://arxiv.org/pdf/2605.06447v1", "arxiv_id": "2605.06447", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "7eedcbfe6fb95e5548e9ed99c5b85d5df27b56cd5f834196f4b32248b3c7b093", "sources": ["arxiv", "semantic_scholar"], "title": "EMO: Pretraining Mixture of Experts for Emergent Modularity", "abstract": "Large language models are typically deployed as monolithic systems, requiring the full model even when applications need only a narrow subset of capabilities, e.g., code, math, or domain-specific knowledge. Mixture-of-Experts (MoEs) seemingly offer a potential alternative by activating only a subset of experts per input, but in practice, restricting inference to a subset of experts for a given domain leads to severe performance degradation. This limits their practicality in memory-constrained settings, especially as models grow larger and sparser. We introduce EMO, an MoE designed for modularity-the independent use and composition of expert subsets-without requiring human-defined priors. Our key idea is to encourage tokens from similar domains to rely on similar experts. Since tokens within a document often share a domain, EMO restricts them to select experts from a shared pool, while allowing different documents to use different pools. This simple constraint enables coherent expert groupings to emerge during pretraining using document boundaries alone. We pretrain a 1B-active, 14B-total EMO on 1T tokens. As a full model, it matches standard MoE performance. Crucially, it enables selective expert use: retaining only 25% (12.5%) of experts incurs just a 1% (3%) absolute drop, whereas standard MoEs break under the same setting. We further find that expert subsets in EMO specialize at semantic levels (e.g., domains such as math or code), in contrast to the low-level syntactic specialization observed in standard MoEs. Altogether, our results demonstrate a path toward modular, memory-efficient deployment of large, sparse models and open new opportunities for composable architectures.", "authors": ["Ryan Wang", "Akshita Bhagia", "Sewon Min"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06663", "pdf_url": "https://arxiv.org/pdf/2605.06663v2", "arxiv_id": "2605.06663", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5d35d7d8183bb189725646de5a47152deb7934fe9d63eb71656c0eb3f9f76429", "sources": ["arxiv", "semantic_scholar"], "title": "When Does Value-Aware KV Eviction Help? A Fixed-Contract Diagnostic for Non-Monotone Cache Compression", "abstract": "Long-context LLM inference is bottlenecked by the memory and bandwidth cost of reading large KV caches during decoding. KV compression reduces this cost by keeping only part of the cache, but task accuracy alone does not identify why a selector succeeds or fails. A selector can fail at three steps: it may miss the evidence future decoding needs, give high scores to tokens that do not affect the output, or break related evidence when fitting scores into a small cache. We introduce a fixed-contract diagnostic that holds the selector's setup fixed and changes one decision slot at a time. For value ranking, the probe combines a block's attention mass with the estimated output change from removing it. On LongBench across three models and two budgets, the probe is positive on 72.6% of positive-margin cells and 32.4% of nonpositive-margin cells. NeedleBench M-RT at 32k and a RULER 8k check probe support closure under branched retrieval, and a 264-cell sign evaluation separates support recovery and output-value ranking from leverage effects near the boundary. The resulting order is to recover decode-side evidence, rank its output value, and preserve coupled evidence during projection.", "authors": ["Ruijie Zhang", "Haozhe Liang", "Da Chang", "Li Hu", "Fanqi Kong", "Huaxiao Yin", "Yu Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.08234", "pdf_url": "https://arxiv.org/pdf/2605.08234v1", "arxiv_id": "2605.08234", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "cee1bb153817e395104cd0c4a32cb9059645f9c1139c1d01829e5475b40a7352", "sources": ["arxiv", "semantic_scholar"], "title": "How to Compress KV Cache in RL Post-Training? Shadow Mask Distillation for Memory-Efficient Alignment", "abstract": "Reinforcement Learning (RL) has emerged as a crucial paradigm for unlocking the advanced reasoning capabilities of Large Language Models (LLMs), encompassing frameworks like RLHF and RLAIF. Regardless of the specific optimization algorithm (e.g., PPO, GRPO, or Online DPO), online RL inherently requires an exploratory trajectory generation (rollout) phase. However, for long-context reasoning tasks, this rollout phase imposes a severe ``memory wall'' due to the exorbitant Key-Value (KV) cache footprint. While applying KV cache compression during rollouts mitigates this memory overhead, it induces a critical off-policy bias. Although modern KV compression is often nearly lossless during standard inference, even minuscule approximation errors are drastically amplified by the inherent instability of RL optimization. Specifically, the sampler generates responses under a sparse context, whereas the learner updates parameters using the full, dense context. Existing statistical solutions, such as importance reweighting, struggle to correct this magnified bias, suffering from high gradient variance and severe sample inefficiency.", "authors": ["Rui Zhu", "Weiheng Bai", "Qiushi Wu", "Yang Ren", "Haixu Tang", "Yuchu Liu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06850", "pdf_url": "https://arxiv.org/pdf/2605.06850v1", "arxiv_id": "2605.06850", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "9a4cde77ec3dd676758cdd0d50b62bbb05d28df120647d91081b91c81773855a", "sources": ["arxiv", "semantic_scholar"], "title": "GEM: Graph-Enhanced Mixture-of-Experts with ReAct Agents for Dialogue State Tracking", "abstract": "Dialogue State Tracking (DST) requires precise extraction of structured information from multi-domain conversations, a task where Large Language Models (LLMs) struggle despite their impressive general capabilities. We present GEM (Graph-Enhanced Mixture-of-Experts), a novel framework that combines language models and graph-structured dialogue understanding with ReAct agent-based reasoning for superior DST performance. Our approach dynamically routes between specialized experts: a Graph Neural Network that captures dialogue structure and turn-level dependencies, and a finetuned T5-Small encoder-decoder for sequence modeling, coordinated by an intelligent router. For complex value generation tasks, we integrate ReAct agents that perform structured reasoning over dialogue context. On MultiWOZ 2.2, GEM achieves 65.19% Joint Goal Accuracy, substantially outperforming end-to-end LLM approaches (best: 38.43%) and surpassing state-of-the-art (SOTA) methods including TOATOD (63.79%), D3ST (58.70%), and Diable (56.48%). Our graph-enhanced mixture-of-experts architecture with ReAct integration demonstrates that combining structured dialogue representation with dynamic expert routing and agent-based reasoning provides a powerful paradigm for dialogue state tracking, achieving superior accuracy while maintaining computational efficiency through selective expert activation.", "authors": ["Ziqi Zhu", "Adithya Suresh", "Tomal Deb", "Iman Abbasnejad"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-06", "url": "https://arxiv.org/abs/2605.04449", "pdf_url": "https://arxiv.org/pdf/2605.04449v1", "arxiv_id": "2605.04449", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8430e6b515734c28e56ee4aa7f5017caeff983fc15780479b33a60c4974f63ff", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Inverted-Index Routing for Granular Mixtures-of-Experts", "abstract": "Mixture-of-experts (MoE) models enable scalable transformer architectures by activating only a subset of experts per token. Recent evidence suggests that performance improves with increasingly granular experts, i.e., many small experts instead of a few large ones. However, this regime substantially increases routing cost, which can dominate computation. We introduce adaptive inverted-index routing for MoE (AIR-MoE), an inverted-index-inspired routing architecture based on vector quantization (VQ). In a first stage, AIR-MoE performs coarse shortlisting by assigning tokens to VQ codewords to construct a candidate set of experts. In a second stage, fine scoring computes exact routing scores restricted to this shortlist. This two-stage procedure approximates true top-k routing while avoiding full expert scoring and, in contrast to prior work, imposing no structural constraints on expert parameters. AIR-MoE serves as a drop-in replacement for standard routers and requires no modifications to the model architecture or loss function. We further provide a lower bound on the mass recall achieved by AIR-MoE that yields insights into its inner workings. Empirically, we demonstrate that AIR-MoE achieves improved performance compared to existing routing approaches in granular MoE settings.", "authors": ["Klaus-Rudolf Kladny", "Maximilian Mordig", "Bernhard Schölkopf", "Michael Muehlebach"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-06", "url": "https://arxiv.org/abs/2605.04952", "pdf_url": "https://arxiv.org/pdf/2605.04952v1", "arxiv_id": "2605.04952", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "981de1cba128af15efba34cae09201629fee982c6c07a4c48afe3b5f1714ca1d", "sources": ["arxiv", "semantic_scholar"], "title": "QKVShare: Quantized KV-Cache Handoff for Multi-Agent On-Device LLMs", "abstract": "Multi-agent LLM systems on edge devices need to hand off latent context efficiently, but the practical choices today are expensive re-prefill or full-precision KV transfer. We study QKVShare, a framework for quantized KV-cache handoff between agents that combines token-level mixed-precision allocation, a self-contained CacheCard representation, and a HuggingFace-compatible cache injection path. Our current results support a narrower but clearer story than the original draft: on 150 GSM8K problems with Llama-3.1-8B-Instruct, adaptive quantization remains competitive under repeated handoff and shows its clearest gains against uniform quantization in deeper-hop, higher budget settings; for handoff latency, the QKVShare path reduces TTFT relative to full re prefill at every tested context, from 130.7 ms vs. 150.2 ms at nominal 1K context to 397.1 ms vs. 1029.7 ms at nominal 8K context;. Stage timing shows that post-injection generation, not card creation, dominates the current QKVShare latency path. These results position quantized KV handoff as a promising on-device systems direction while also highlighting the need for stronger controller ablations and apples-to-apples runtime comparisons.", "authors": ["Pratik Honavar", "Tejpratap GVSL"], "categories": ["cs.AI", "cs.MA"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-05", "url": "https://arxiv.org/abs/2605.03884", "pdf_url": "https://arxiv.org/pdf/2605.03884v1", "arxiv_id": "2605.03884", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "99194f5c688b8cffdaa6f880d4420964bdbabb2464b5cfcf7490f00ee61260b0", "sources": ["arxiv", "semantic_scholar"], "title": "MILE: Mixture of Incremental LoRA Experts for Continual Semantic Segmentation across Domains and Modalities", "abstract": "Continual semantic segmentation requires models to adapt to new domains or modalities without sacrificing performance on previously learned tasks. Expert-based learning, in which task-specific modules specialize in different domains, has proven effective in mitigating forgetting. These methods include dynamic expansion, which suffers from scalability issues, or parameter isolation, which constrains the ability to learn new tasks. We introduce Mixture of Incremental LoRA Experts (MILE), a modular and parameter-efficient framework for continual segmentation across both domains and modalities. MILE leverages Low-Rank Adaptation (LoRA) to instantiate lightweight experts for each new task while keeping the pretrained base network frozen. Each expert is trained exclusively on its task data, thus avoids overwriting previously learned information. A prototype-guided gating mechanism dynamically selects the most appropriate expert at inference. MILE achieves the benefits of expert-based learning while overcoming its scalability limitations. It requires only a marginal parameter increase per task and tens of LoRA adapters are needed before matching the size of a single full model, making it highly efficient in both training and storage. Across domain- and modality-incremental benchmarks, MILE achieves strong performance while ensuring better stability, plasticity, and scalability.", "authors": ["Shishir Muralidhara", "Didier Stricker", "René Schuster"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-05", "url": "https://arxiv.org/abs/2605.03555", "pdf_url": "https://arxiv.org/pdf/2605.03555v1", "arxiv_id": "2605.03555", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1c5c5ecc1b99515f37f4e9eca88c4a9593e66625fbfc23e128bb9957a4b4004d", "sources": ["arxiv", "semantic_scholar"], "title": "HeadQ: Model-Visible Distortion and Score-Space Correction for KV-Cache Quantization", "abstract": "KV-cache quantizers usually optimize storage-space reconstruction, even though attention reads keys through logits and values through attention-weighted readout. We argue that persistent cache error should be measured in model-visible coordinates. For keys, the visible object is score error modulo constant shifts; this yields HeadQ, a key-side method that stores a low-rank residual side code in a calibration-learned query basis and applies it as an additive logit correction. For values, fixed-attention readout gives an $A^2$-weighted token-distortion surrogate. Across six models, Fisher/score-space error predicts attention KL far better than raw key MSE; same-budget counterexamples, null-space interventions, query-PCA controls, and wrong-sign HeadQ falsify storage-MSE alternatives. Matched Pythia checkpoints localize the main anomaly to a small-model low-entropy route-flip boundary. In K-only WikiText-103 decode experiments with dense values, HeadQ removes roughly $84$--$94\\%$ of the excess perplexity on the strongest 2-bit rows; in an auxiliary full-KV 2-bit composition, HeadQ plus an $A^2$ value policy improves all six models.", "authors": ["Jorge L. Ruiz Williams"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-05", "url": "https://arxiv.org/abs/2605.03562", "pdf_url": "https://arxiv.org/pdf/2605.03562v3", "arxiv_id": "2605.03562", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1989de3ab4ec6db6c2cbcd6577004dd9dcb8f2f746d864f93ff3ab79e8e33f29", "sources": ["arxiv", "semantic_scholar"], "title": "Ortho-Hydra: Orthogonalized Experts for DiT LoRA", "abstract": "LoRA fine-tuning of diffusion transformers (DiT) on multi-style data suffers from \\emph{style bleed}: a single low-rank residual cannot represent several distinct artist fingerprints, and the optimizer converges to their average. Mixture-of-experts LoRA in the HydraLoRA style replaces the up-projection with $E$ heads under a router, but when every expert is zero-initialized the router receives identical gradient from each head and remains at the uniform prior. The experts then evolve permutation-symmetrically, and the network trains as a single rank-$r$ LoRA at $E{\\times}$ the cost. We present \\textbf{Ortho-Hydra}, a re-parameterisation that combines an OFT-style Cayley-orthogonal shared basis with per-expert \\emph{disjoint output subspaces} carved from the top-$(Er)$ left singular vectors of the pretrained weight. Disjointness makes the router's per-expert score non-degenerate at step~$0$, so specialization receives gradient signal before any expert has trained. We test the predicted deadlock on a DiT pipeline by comparing two HydraLoRA baselines, a zero-initialized shared-basis variant and the original $σ{=}0.1$ Gaussian-jitter mitigation, against Ortho-Hydra under a matched optimiser, dataset, and step budget. Neither baseline leaves the uniform prior within the first $1\\text{k}$ steps; Ortho-Hydra begins de-uniformising within the first few hundred. End-task generation quality on multi-style data is out of scope; we report the construction, the cold-start mechanism, and the routing dynamics it changes. Code: https://github.com/sorryhyun/anima_lora.", "authors": ["Seunghyun Ji"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-05", "url": "https://arxiv.org/abs/2605.03252", "pdf_url": "https://arxiv.org/pdf/2605.03252v1", "arxiv_id": "2605.03252", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/sorryhyun/anima_lora", "venue": null, "quality_score": 0.65}
{"id": "0f574362af58920a76d6e225588cdd3f2ffbc41804bdd69d111af30014f2c770", "sources": ["arxiv", "semantic_scholar"], "title": "Tutti: Making SSD-Backed KV Cache Practical for Long-Context LLM Serving", "abstract": "LLM serving relies on prefix caching to improve inference performance. As growing contexts push key-value (KV) cache footprint far beyond GPU HBM and CPU DRAM capacity, KV cache is increasingly offloaded to NVMe SSDs. Unfortunately, restoring KV cache from SSDs suffers from poor I/O performance and incurs significant GPU stalls. This is primarily because the fragmented GPU memory layout results in a massive number of tiny random I/Os, rendering the low-parallelism CPU a severe bottleneck even with GPU Direct Storage (GDS), which still relies on CPU intervention to initiate each I/O and thus remains CPU-centric. This paper presents Tutti, an efficient SSD-backed KV caching solution that eliminates CPU intervention from the critical data and I/O control paths between HBM and SSDs. At the core of Tutti is a GPU-centric KV cache object store, in which the CPU is only responsible for asynchronously loading I/O kernels once per layer to the GPU. Tutti saturates NVMe SSD bandwidth and reduces GPU stalls to near zero through the following designs: (i) we provide a GPU-native object abstraction that enables bulk KV cache transfers and management; (ii) we re-architect the GPU storage stack by introducing GPU io_uring to support asynchronous GPU direct object I/O; and (iii) we propose slack-aware I/O scheduling to avoid GPU resource contention. We have implemented Tutti and integrated it to vLLM. Extensive evaluation shows that compared to the state-of-the-art GDS-enabled, SSD-backed LMCache, Tutti reduces TTFT by 78.3% under strict SLO constraints and improves the achievable request rate by 2x. The serving cost is reduced by 27%. Tutti achieves nearly the same inference performance as DRAM-backed LMCache, while providing almost infinite capacity.", "authors": ["Shi Qiu", "Yifan Hu", "Xintao Wang", "Wenhao Zhu", "Jianqin Yan", "Hao Chen", "Kaiqiang Xu", "Kai Chen", "Yiming Zhang"], "categories": ["cs.OS"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-05", "url": "https://arxiv.org/abs/2605.03375", "pdf_url": "https://arxiv.org/pdf/2605.03375v1", "arxiv_id": "2605.03375", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5710bc1c674e4e4fa1f57ef3ab578c81c256e398bfc745210197866a3b05608f", "sources": ["arxiv", "semantic_scholar"], "title": "AdapShot: Adaptive Many-Shot In-Context Learning with Semantic-Aware KV Cache Reuse", "abstract": "Many-Shot In-Context Learning (ICL) has emerged as a promising paradigm, leveraging extensive examples to unlock the reasoning potential of Large Language Models (LLMs). However, existing methods typically rely on a predetermined, fixed number of shots. This static approach often fails to adapt to the varying difficulty of different queries, leading to either insufficient context or interference from noise. Furthermore, the prohibitive computational and memory costs of long contexts severely limit Many-Shot's feasibility. To address the above limitations, we propose AdapShot, which dynamically optimizes shot counts and leverages KV cache reuse for efficient inference. Specifically, we design a probe-based evaluation mechanism that utilizes output entropy to determine the optimal number of shots. To bypass the redundant prefilling computation during both the probing and inference phases, we incorporate a semantics-aware KV cache reuse strategy. Within this reuse strategy, to address positional encoding incompatibilities, we introduce a decoupling and re-encoding method that enables the flexible reordering of cached key-value pairs. Extensive experiments demonstrate that AdapShot achieves an average performance gain of around 10% and a 4.64x speedup compared to state-of-the-art DBSA.", "authors": ["Jie Ou", "Jinyu Guo", "Shiyao Guo", "Yuang Li", "Ruiqi Wu", "Zhaokun Wang", "Wenyi Li", "Wenhong Tian"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-05", "url": "https://arxiv.org/abs/2605.03644", "pdf_url": "https://arxiv.org/pdf/2605.03644v2", "arxiv_id": "2605.03644", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "058d06f6fb91f52794d060f47ff3a9d2e48e81612af83efd77030f8e2c4ef2c8", "sources": ["arxiv", "semantic_scholar"], "title": "WindowQuant: Mixed-Precision KV Cache Quantization based on Window-Level Similarity for VLMs Inference Optimization", "abstract": "Recently, video language models (VLMs) have been applied in various fields. However, the visual token sequence of the VLM is too long, which may cause intolerant inference latency and GPU memory usage. Existing methods propose mixed-precision quantization to the key-value (KV) cache in VLMs based on token granularity, which is time-consuming in the search process and hardware inefficient during computation. This paper introduces a novel approach called WindowQuant, which employs window-adaptive mixed-precision quantization to optimize the KV cache. WindowQuant consists of two modules: window-level quantization search and window-level KV cache computation. Window-level quantization search quickly determines the optimal bit-width configuration of the KV cache windows based on the similarity scores between the corresponding visual token windows and the text prompt, maintaining the model accuracy. Furthermore, window-level KV cache computation reorders the KV cache windows before quantization, avoiding the hardware inefficiency caused by mixed-precision quantization in inference computation. Extensive experiments demonstrate that WindowQuant outperforms state-of-the-art VLM models and KV cache quantization methods on various datasets.", "authors": ["Wei Tao", "Xiaoyang Qu", "Peiqiang Wang", "Guokuan Li", "Jiguang Wan", "Kai Lu", "Jianzong Wang"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-04", "url": "https://arxiv.org/abs/2605.02262", "pdf_url": "https://arxiv.org/pdf/2605.02262v1", "arxiv_id": "2605.02262", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "2652819c0cd0d1697133d380d567bab6f7e3e5b16904dba77b55d3f4278cfd5d", "sources": ["arxiv", "semantic_scholar"], "title": "Decouple and Cache: KV Cache Construction for Streaming Video Understanding", "abstract": "Streaming video understanding requires processing unbounded video streams with limited memory and computation, posing two key challenges. First, continuously constructing new and evicting old key-value(KV) caches is required for unbounded streams. Secondly, due to the high cost of collecting and training on unbounded streams, models must learn from short sequences while generalizing to long streams. Existing streaming VideoVLLMs fail to scale to unbounded video streams or focus on cache reuse strategies, leaving the impact of cache construction underexplored. In this paper, we propose Decoupled Streaming Cache(DSCache), a training-free cache construction mechanism that adapts pretrained offline models to streaming settings. DSCache maintains a cumulative past KV cache while constructing a separate instant cache on-demand, decoupled from past caches to preserve the informativeness of recent inputs. To enable position extrapolation beyond the training length, DSCache further incorporates a position-agnostic encoding strategy, ensuring KV caches to support unseen positions and preventing position overflow. Experiments on Streaming Video QA benchmarks demonstrate DSCache's state-of-the-art performance, with an average 2.5% accuracy gains over prior methods.", "authors": ["Zhanzhong Pang", "Dibyadip Chatterjee", "Fadime Sener", "Angela Yao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-03", "url": "https://arxiv.org/abs/2605.01858", "pdf_url": "https://arxiv.org/pdf/2605.01858v1", "arxiv_id": "2605.01858", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e9c9995d3853dc898d39d10868e74f4d6225c4655fb9894ee9bcf5de47b82d72", "sources": ["arxiv", "semantic_scholar"], "title": "GETA-3DGS: Automatic Joint Structured Pruning and Quantization for 3D Gaussian Splatting", "abstract": "3D Gaussian splatting (3DGS) is a state-of-the-art representation for real-time photorealistic novel-view synthesis, yet a single high-fidelity scene typically occupies hundreds of megabytes to several gigabytes, exceeding the budgets of mobile, immersive, and volumetric video platforms. Existing 3DGS compression methods (e.g., HAC++, FlexGaussian, LP-3DGS) treat pruning, quantization, and entropy coding as separate stages and rely on hand-tuned heuristics (opacity thresholds, fixed bit-widths, SH truncation), limiting cross-scene generalization and preventing users from specifying a target rate or quality budget. We propose GETA-3DGS, to our knowledge the first end-to-end automatic joint structured pruning and quantization framework for 3DGS. Building on GETA for joint pruning-quantization of deep networks, we contribute: (i) a 3DGS-aware quantization-aware dependency graph (QADG) treating each Gaussian primitive as a group with five attribute sub-nodes and degree-aware SH sub-nodes; (ii) a render-aware saliency fusing transmittance-weighted contribution, screen-space gradient, and pixel coverage into a Gaussian-level importance score; and (iii) a heterogeneous per-attribute mixed-precision scheme co-optimized with structural sparsity under a projected partial saliency-guided (PPSG) descent guarantee. On Mip-NeRF 360, Tanks and Temples, and Deep Blending, GETA-3DGS operates directly on raw Gaussian primitives rather than a post-hoc anchor representation, delivering ~5x storage reduction over Vanilla 3DGS with no per-scene thresholds. Bit-width policy is the dominant rate-distortion lever: a uniform 6-bit cap costs up to -6.74 dB on view-dependent scenes versus our heterogeneous allocation, matching an information-theoretic reverse-water-filling analysis we develop. GETA-3DGS is complementary to existing codecs: entropy coding (HAC++, CompGS) is downstream, so the two can be composed.", "authors": ["Baobing Zhang", "Wanxin Sui"], "categories": ["cs.LG", "cs.AI", "cs.GR", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-05-03", "url": "https://arxiv.org/abs/2605.02086", "pdf_url": "https://arxiv.org/pdf/2605.02086v1", "arxiv_id": "2605.02086", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "77d2252108ae94d30c4093eb8c20d4326da34389ac2fac5c6003706282abf742", "sources": ["arxiv", "semantic_scholar"], "title": "Shifted asymmetric Laplace mixtures of experts", "abstract": "Mixtures of experts (MoE) models provide a flexible framework for modelling heterogeneity in data for regression and model-based clustering and classification. MoE models for regression are typically based on the Gaussian assumption for the expert distributions. To robustify the MoE framework with respect to data exhibiting skewness, heavy tails and outliers, we propose a robust non-normal MoE model using the shifted asymmetric Laplace (SAL) distribution. The proposed SALMoE model overcomes the limitations of the Gaussian MoE model when the observed data are asymmetric and heavy-tailed. Through a combination of the minorization-maximization (MM) algorithm with the classical Expectation-Maximization (EM), we develop a dedicated hybrid EM-MM algorithm to estimate the parameters of the SALMoE model. The EM-MM algorithm is shown to yield a nondecreasing observed log-likelihood. A simulation study demonstrates the robustness and practical utility of the proposed model. Finally, the SALMoE model is applied to two real-world economic datasets.", "authors": ["Sphiwe B. Skhosana", "Hien Duy Nguyen"], "categories": ["stat.ME"], "fields_of_study": ["Mathematics"], "published_date": "2026-05-03", "url": "https://arxiv.org/abs/2605.02012", "pdf_url": "https://arxiv.org/pdf/2605.02012v1", "arxiv_id": "2605.02012", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "779c8e2110cd074617649fad0daeb570d1287783db28daac7a74041f7be65a33", "sources": ["arxiv", "semantic_scholar"], "title": "Rhamba: Region-Aware Hybrid Attention-Mamba Framework for Self-Supervised Learning in Resting-State fMRI", "abstract": "Self-supervised pretraining is promising for large-scale neuroimaging, yet the impact of region-aware masking and hybrid sequence modeling remains underexplored. In this work, we introduce Rhamba, a region-aware pretraining framework that integrates anatomically guided masking with hybrid Attention-Mamba architectures for resting state functional magnetic resonance imaging (fMRI) analysis. Models were pretrained on the ABIDE dataset using region-aligned patch embeddings and three masking strategies (Any, Majority, and Pure) with increasing spatial specificity. We evaluated four architectural variants: a Mamba only model, an Alternate architecture with interleaved Mamba and Attention blocks, and two hybrid encoder-decoder configurations (Attention-Mamba (AM) and Mamba-Attention (MA)). The pretrained models were fine-tuned on downstream classification tasks using the COBRE and ADHD-200 datasets for schizophrenia and attention-deficit/hyperactivity disorder discrimination. We employed Integrated Gradients, an explainable AI method, to identify the brain regions contributing to model predictions. Masking strategy strongly influenced reconstruction behavior, with reconstruction loss following a consistent ordering (Any > Majority > Pure). However, this trend did not directly translate into downstream performance, where differences were modest and dataset-dependent. The hybrid architecture with the MA configuration achieved the highest average AUROC across both datasets, and Rhamba outperformed state-of-the-art methods in comparative evaluation. Region-wise analysis showed that peak performance depends on the interaction between masking strategy and architecture rather than a single dominant configuration. Overall, Rhamba offers a flexible framework for balancing interpretability, scalability, and performance in large-scale fMRI representation learning.", "authors": ["Ruthwik Reddy Doodipala", "Pankaj Pandey", "Pratheek Eranki", "Carolina Torres-Rojas", "Manob Jyoti Saikia", "Ranganatha Sitaram"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-02", "url": "https://arxiv.org/abs/2605.01240", "pdf_url": "https://arxiv.org/pdf/2605.01240v2", "arxiv_id": "2605.01240", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "8465571aec1b6f3a88fe382be6c00f4b70db375fc863624818c68cbf65204c4b", "sources": ["arxiv", "semantic_scholar"], "title": "Make Your LVLM KV Cache More Lightweight", "abstract": "Key-Value (KV) cache has become a de facto component of modern Large Vision-Language Models (LVLMs) for inference. While it enhances decoding efficiency in Large Language Models (LLMs), its direct adoption in LVLMs introduces substantial GPU memory overhead due to the large number of vision tokens processed during the prefill stage. To tackle this problem, we propose LightKV, a novel approach that reduces KV cache size by exploiting the redundancy among vision-token embeddings. Guided by text prompts, LightKV employs cross-modality message passing to aggregate informative messages across vision tokens and progressively compress them during prefill. This prompt-aware guidance distinguishes our method from prior vision-only compression strategies. We evaluate LightKV on eight open-source LVLMs across eight public benchmark datasets, e.g., MME and SeedBench. Experimental results demonstrate that with only 55% of the original vision tokens, LightKV (a) halves the vision-token KV cache size, (b) reduces computation by up to 40%, and (c) preserves general-purpose performance while significantly outperforming existing baselines.", "authors": ["Xihao Chen", "Yangyang Guo", "Roger Zimmermann"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-01", "url": "https://arxiv.org/abs/2605.00789", "pdf_url": "https://arxiv.org/pdf/2605.00789v1", "arxiv_id": "2605.00789", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.65}
{"id": "3f90697cba1df1cf4315203cd881232aa35e1d68d1cea139ba7729df7b039269", "sources": ["arxiv", "semantic_scholar"], "title": "RouteHijack: Routing-Aware Attack on Mixture-of-Experts LLMs", "abstract": "Safety alignment is critical for the responsible deployment of large language models (LLMs). As Mixture-of-Experts (MoE) architectures are increasingly adopted to scale model capacity, understanding their safety robustness becomes essential. Existing adversarial attacks, however, have notable limitations. Prompt-based jailbreaks rely on heuristic search and transfer poorly, model intervention methods require privileged access to internal representations, and optimization-based input attacks remain output-centric and are fundamentally limited to MoE models due to the non-differentiable routing mechanism. In this paper, we present RouteHijack, a routing-aware jailbreak for MoE LLMs. Our key insight is that safety behavior is concentrated in a small subset of experts, creating an opportunity to steer model behavior by influencing routing decisions through input optimization. Building on this observation, RouteHijack first performs response-driven expert localization to identify safety-critical and harmful experts by contrasting activations under safe refusals and harmful completions. It then constructs adversarial suffixes with a routing-aware objective that suppresses safety experts, promotes harmful experts, and prevents early-stage refusal during generation. At inference time, the optimized suffix is appended to a malicious prompt, requiring only input access. Across seven MoE LLMs, RouteHijack achieves a 69.3\\% average attack success rate (ASR), outperforming prior optimization-based attack by $3.2\\times$. RouteHijack also transfers zero-shot across five sibling MoE variants, raising average ASR from 27.7\\% to 61.2\\%, and further generalizes to three MoE-based VLMs, increasing average ASR from 2.47\\% to 38.7\\%. These findings expose a fundamental vulnerability in sparse expert architectures and highlight the need for defenses beyond output-level alignment.", "authors": ["Zhiyuan Xu", "Joseph Gardiner", "Sana Belguith", "Lichao Wu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-01", "url": "https://arxiv.org/abs/2605.02946", "pdf_url": "https://arxiv.org/pdf/2605.02946v1", "arxiv_id": "2605.02946", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "12350b005ae0d1af0ca8f9d2d910aecc1682cd2b3a5be28f7c90a2a01578a1a3", "sources": ["arxiv", "semantic_scholar"], "title": "SpaceMoE: Realizing Distributed Mixture-of-Experts Inference over Space Networks", "abstract": "Leveraging continuous solar energy harvesting at high efficiency, space data centers are envisioned as a promising platform for executing energy-intensive large language models (LLMs). Recognizing this advantage, space and AI conglomerates (e.g., SpaceX, Google) are actively investing in this vision. One key challenge, however, is the efficient distributed deployment of a large-scale LLM in a satellite network due to the limited onboard computing and communication resources. This gives rise to a placement problem that involves partitioning and mapping model components to satellites such that the fundamentally different model architecture and network topology can be reconciled to ensure low-latency token generation. To address this problem, we present the Space Network of Mixture-of-Experts (SpaceMoE) framework targeting the distributed execution of a popular mixture-of-experts (MoE) model in space. The proposed placement strategies are two-level: (1) layer placement, which assigns MoE layers to satellite subnets; and (2) intra-layer expert placement, which assigns individual experts to satellites associated with the same layer/subnet. For layer placement, we exploit the ring-like communication pattern of autoregressive inference to partition the satellite constellation along the orbiting direction into subnets arranged on a ring, each hosting one MoE layer. Based on this architecture, we formulate and solve an optimization problem for intra-layer expert placement to map experts with heterogeneous activation probabilities onto satellites. The derived strategy reveals an intuitive principle: a frequently activated expert should be mapped to a satellite on a routing path with low expected latency. Experiments over a thousand-satellite constellation show that SpaceMoE achieves at least a threefold latency reduction compared with conventional random and ablation-based placement strategies.", "authors": ["Zhanwei Wang", "Huiling Yang", "Min Sheng", "Khaled B. Letaief", "Kaibin Huang"], "categories": ["cs.DC", "cs.AI", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-01", "url": "https://arxiv.org/abs/2605.00515", "pdf_url": "https://arxiv.org/pdf/2605.00515v2", "arxiv_id": "2605.00515", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "88cf2aa12c76834c90b46006e99c28101f8639f3bc30bcfa852d2417e63f9ec4", "sources": ["arxiv", "semantic_scholar"], "title": "Prediction-powered Inference by Mixture of Experts", "abstract": "The rapidly expanding artificial intelligence (AI) industry has produced diverse yet powerful prediction tools, each with its own network architecture, training strategy, data-processing pipeline, and domain-specific strengths. These tools create new opportunities for semi-supervised inference, in which labeled data are limited and expensive to obtain, whereas unlabeled data are abundant and widely available. Given a collection of predictors, we treat them as a mixture of experts (MOE) and introduce an MOE-powered semi-supervised inference framework built upon prediction-powered inference (PPI). Motivated by the variance reduction principle underlying PPI, the proposed framework seeks the mixture of experts that achieves the smallest possible variance. Compared with standard PPI, the MOE-powered inference framework adapts to the unknown performance of individual predictors, benefits from their collective predictive power, and enjoys a best-expert guarantee. The framework is flexible and applies to mean estimation, linear regression, quantile estimation, and general M-estimation. We develop non-asymptotic theory for the MOE-powered inference framework and establish upper bounds on the coverage error of the resulting confidence intervals. Numerical experiments demonstrate the practical effectiveness of MOE-powered inference and corroborate our theoretical findings.", "authors": ["Yanwu Gu", "Linglong Kong", "Dong Xia"], "categories": ["stat.ML", "cs.LG", "stat.AP"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-04-30", "url": "https://arxiv.org/abs/2604.27892", "pdf_url": "https://arxiv.org/pdf/2604.27892v1", "arxiv_id": "2604.27892", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "071a6f27c022d0a90db106c6e43458610ae70971112f4a13695f9d42dd802689", "sources": ["arxiv", "semantic_scholar"], "title": "Post-Optimization Adaptive Rank Allocation for LoRA", "abstract": "Exponential growth in the scale of modern foundation models has led to the widespread adoption of Low-Rank Adaptation (LoRA) as a parameter-efficient fine-tuning technique. However, standard LoRA implementations disregard the varying intrinsic dimensionality of model layers and enforce a uniform rank, leading to parameter redundancy. We propose Post-Optimization Adaptive Rank Allocation (PARA), a data-free compression method for LoRA that integrates seamlessly into existing fine-tuning pipelines. PARA leverages Singular Value Decomposition to prune LoRA ranks using a global threshold over singular values across all layers. This results in non-uniform rank allocation based on layer-wise spectral importance. As a post-hoc method, PARA circumvents the training modifications and resulting instabilities that dynamic architectures typically incur. We empirically demonstrate that PARA reduces parameter count by 75-90\\% while preserving the predictive performance of the original, uncompressed LoRA across multiple vision and language benchmarks. Code will be published upon acceptance.", "authors": ["Vishnuprasadh Kumaravelu", "Sunil Gupta", "P. K. Srijith"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-30", "url": "https://arxiv.org/abs/2604.27796", "pdf_url": "https://arxiv.org/pdf/2604.27796v1", "arxiv_id": "2604.27796", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "1ace4cef61ab2e297f54bce4d65226df0c6a5752ba199e684a26f4e3c8f45349", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive and Fine-grained Module-wise Expert Pruning for Efficient LoRA-MoE Fine-Tuning", "abstract": "LoRA-MoE has emerged as an effective paradigm for parameter-efficient fine-tuning, combining the low training cost of LoRA with the increased adaptation capacity of Mixture-of-Experts (MoE). However, existing LoRA-MoE frameworks typically adopt a fixed and uniform expert configuration across heterogeneous Transformer modules (\\eg, attention query/key projections and MLP gating networks), ignoring their distinct functional roles and capacity requirements. This design leads to localized over-provisioning, redundant trainable parameters, and unnecessary optimizer-state overhead. Moreover, prior methods enforce load balancing among experts throughout training. Although beneficial in the early stage, this constraint becomes restrictive once routing patterns stabilize, limiting expert specialization on downstream tasks. In this paper, we propose DMEP, a novel LoRA-MoE fine-tuning framework based on Dynamic Module-wise Expert Pruning. DMEP tracks expert utilization during training and physically removes low-utility experts on a per-module basis, yielding a more compact expert structure tailored to different modules. The pruned model then continues training without the load-balancing constraint, freeing the remaining experts to focus entirely on the downstream task and develop specialized expertise. By jointly adapting module-wise expert capacity and eliminating unnecessary balancing, DMEP improves both parameter efficiency and training efficiency. Extensive experiments on multiple reasoning benchmarks show that DMEP reduces trainable parameters by 35\\%--43\\% and improves training throughput by about 10\\%, while maintaining or surpassing the downstream reasoning accuracy of uniform LoRA-MoE baselines.", "authors": ["Weihang Li", "Jianchun Liu", "Hongli Xu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-29", "url": "https://arxiv.org/abs/2604.26340", "pdf_url": "https://arxiv.org/pdf/2604.26340v1", "arxiv_id": "2604.26340", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ba1a6fd3ced3912365c2aa0f8da889e89d98fc5d4fcdc47c9161d5cef9eb4ba1", "sources": ["arxiv", "semantic_scholar"], "title": "DUAL-BLADE: Dual-Path NVMe-Direct KV-Cache Offloading for Edge LLM Inference", "abstract": "The increasing deployment of Large Language Model (LLM) inference on edge AI systems demands efficient execution under tight memory budgets. A key challenge arises from Key-Value (KV) caches, which often exceed available device memory. Although NVMe-based offloading offers scalable capacity, existing file-based designs rely heavily on the kernel page cache, leading to cache thrashing, unpredictable latency, and high software overhead under memory pressure. We present DUAL-BLADE, a dual-path KV residency framework that dynamically assigns KV tensors to either a page-cache path or an NVMe-direct path based on runtime memory availability. The NVMe-direct path bypasses the filesystem by mapping KV tensors to contiguous logical block address (LBA) regions, enabling low-overhead direct storage access. DUAL-BLADE further incorporates adaptive pipeline parallelism to overlap storage I/O with GPU DMA, improving inference throughput. Our evaluation shows that DUAL-BLADE substantially mitigates I/O bottlenecks, reducing prefill and decode latency by up to 33.1% and 42.4%, respectively, while improving SSD utilization by 2.2x across diverse memory budgets.", "authors": ["Bodon Jeong", "Hongsu Byun", "Youngjae Kim", "Weikuan Yu", "Kyungkeun Lee", "Jihoon Yang", "Sungyong Park"], "categories": ["cs.DC", "cs.AI", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-29", "url": "https://arxiv.org/abs/2604.26557", "pdf_url": "https://arxiv.org/pdf/2604.26557v1", "arxiv_id": "2604.26557", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "460a2479e928c9b61782fb98c048b4f071996aab29820ffca4d2d6313b8e4995", "sources": ["arxiv", "semantic_scholar"], "title": "When Hidden States Drift: Can KV Caches Rescue Long-Range Speculative Decoding?", "abstract": "Speculative decoding accelerates LLM inference, but SOTA hidden-state-based drafters suffer from long-range decay: draft accuracy degrades as the speculative step increases. Existing work attributes this decay to train-inference mismatch and proposes test-time training (TTT) as a remedy, yet we observe that long-range decay persists even in TTT-trained drafters. We revisit long-range decay from the perspective of context information preservation. In hidden-state reuse, we argue the target hidden state acts as a biased context compression: it aggregates historical token information according to the attention query at the current position, yielding a compact representation optimized for immediate next-token prediction. This compression can suppress information less relevant to the current query but important for later speculative steps. In contrast, the target model's KV cache serves as an explicit context, retaining the complete set of token-wise KV representations. We therefore posit the KV-Reuse Hypothesis: allowing the draft model to reuse the target KV cache can provide richer signals for long-horizon drafting. To test this hypothesis, we introduce KVShot, a diagnostic framework that compares three reuse paradigms: hidden-only, KV-only, and hybrid. Extensive evaluations on Qwen3-8B show that KV-Reuse improves long-range acceptance, although end-to-end speedups remain marginal under current training pipelines. Our analysis identifies two key structural bottlenecks: shallow drafters struggle to estimate target queries accurately, and draft-side KV projections receive sparse gradient signals. These findings suggest that realizing the full potential of KV-aware decoding requires moving beyond TTT toward block-wise training paradigms. By exposing these bottlenecks, KVShot provides a foundational diagnostic testbed and a clear roadmap for designing next-generation inference architectures.", "authors": ["Tianyu Liu", "Yuhao Shen", "Xinyi Hu", "Baolin Zhang", "Hengxin Zhang", "Jun Dai", "Jun Zhang", "Shuang Ge", "Lei Chen", "Yue Li", "MingCheng Wan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-29", "url": "https://arxiv.org/abs/2604.26412", "pdf_url": "https://arxiv.org/pdf/2604.26412v2", "arxiv_id": "2604.26412", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "073469c8aad8e4b04433602e16f0870a8f6d78d1d8a6ec2da76a4a7fbfdc720c", "sources": ["arxiv", "semantic_scholar"], "title": "FaaSMoE: A Serverless Framework for Multi-Tenant Mixture-of-Experts Serving", "abstract": "Mixture-of-Experts (MoE) models offer high capacity with efficient inference cost by activating a small subset of expert models per input. However, deploying MoE models requires all experts to reside in memory, creating a gap between the resource used by activated experts and the provisioned resources. This underutilization is further pronounced in multi-tenant scenarios. In this paper, we propose FaaSMoE, a multi-tenant MoE serving architecture built on Function-as-a-Service (FaaS) platforms. FaaSMoE decouples the control and execution planes of MoE by deploying experts as stateless FaaS functions, enabling on-demand and scale-to-zero expert invocation across tenants. FaaSMoE further supports configurable expert granularity within functions, trading off per-expert elasticity for reduced invocation overhead. We implement a prototype with an open-source edge-oriented FaaS platform and evaluate it using Qwen1.5-moe-2.7B under multi-tenant workloads. Compared to a full-model baseline, FaaSMoE uses less than one third of the resources, demonstrating a practical and resource-efficient path towards scalable MoE serving in a multi-tenant environment.", "authors": ["Minghe Wang", "Trever Schirmer", "Mohammadreza Malekabbasi", "David Bermbach"], "categories": ["cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-29", "url": "https://arxiv.org/abs/2604.26881", "pdf_url": "https://arxiv.org/pdf/2604.26881v1", "arxiv_id": "2604.26881", "doi": "10.1145/3812836.3814785", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.65}
{"id": "a6742c0ad6f3d453cf593b0211d176aae24863c5cb2ca570a892691a9ae0e993", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking KV Cache Eviction via a Unified Information-Theoretic Objective", "abstract": "Key-value (KV) caching is essential for large language model inference, yet its memory overhead poses a critical bottleneck for long-context generation. Existing eviction policies predominantly rely on empirical heuristics, lacking a rigorous theoretical foundation. This work rethinks KV cache eviction through the lens of the Information Bottleneck principle. Under a linear-Gaussian surrogate of attention, we derive a closed-form mutual information objective that characterizes the effective information capacity of a retained KV cache subset. This formulation reveals that a wide range of existing eviction strategies can be interpreted as different approximations of the same capacity-maximization principle. Guided by this insight, we introduce CapKV, a capacity-aware eviction method that directly targets information preservation via a log-determinant approximation using statistical leverage scores. This approach replaces heuristic selection with a theoretically grounded mechanism that preserves the maximum predictive signal. Extensive experiments across multiple models and long-context benchmarks show that CapKV consistently outperforms prior methods, achieving a better trade-off between memory efficiency and generational fidelity.", "authors": ["Jiaming Yang", "Chenwei Tang", "Liangli Zhen", "Jiancheng Lv"], "categories": ["cs.LG", "cs.AI", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-28", "url": "https://arxiv.org/abs/2604.25975", "pdf_url": "https://arxiv.org/pdf/2604.25975v1", "arxiv_id": "2604.25975", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d900d2b9ac00781cd8b11f7f0a9e074c43f41c0218b207e9f224bbff9775c1b7", "sources": ["arxiv", "semantic_scholar"], "title": "CacheFlow: Efficient LLM Serving with 3D-Parallel KV Cache Restoration", "abstract": "KV cache restoration has emerged as a dominant bottleneck in serving long-context LLM workloads, including multi-turn conversations, retrieval-augmented generation, and agentic pipelines. Existing approaches treat restoration as a per-request tradeoff between recomputation and I/O transfer, recomputing KV states from scratch or offloading them from external storage (e.g., CPU memory or remote machines). However, existing advances fail to exploit parallelism across tokens, layers, and distributed deployments, and critically ignore resource contention under batched serving. We present CacheFlow, a KV cache restoration framework that rethinks cache restoration as a multi-dimensional parallel execution problem. CacheFlow introduces a unified 3D parallelism abstraction across tokens, layers, and GPUs, enabling fine-grained overlap of recomputation and I/O along the structural dependencies of transformer inference. At the core of CacheFlow is a batch-aware two-pointer scheduler that jointly optimizes compute and I/O allocation across requests by prioritizing operations with the highest marginal reduction in recomputation cost. Our evaluations show that CacheFlow reduces Time-To-First-Token (TTFT) by 10%-62% over existing advances across diverse models, workloads, and hardware.", "authors": ["Sean Nian", "Jiahao Fang", "Qilong Feng", "Zhiyu Wu", "Fan Lai"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-28", "url": "https://arxiv.org/abs/2604.25080", "pdf_url": "https://arxiv.org/pdf/2604.25080v1", "arxiv_id": "2604.25080", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "0e37171d683f817f473813ed34c02d4200c8172ebf88cc8d74db3096c614837b", "sources": ["arxiv", "semantic_scholar"], "title": "PolyKV: A Shared Asymmetrically-Compressed KV Cache Pool for Multi-Agent LLM Inference", "abstract": "We present PolyKV, a system in which multiple concurrent inference agents share a single, asymmetrically compressed KV cache pool. Rather than allocating a separate KV cache per agent -- the standard paradigm -- PolyKV writes a compressed cache once and injects it into N independent agent contexts via HuggingFace DynamicCache objects. Compression is asymmetric: Keys are quantized at int8 (q8_0) to preserve softmax stability, while Values are compressed using TurboQuant MSE -- a Fast Walsh-Hadamard Transform (FWHT) rotation followed by 3-bit Lloyd-Max quantization with centroids tuned to N(0,1). We evaluate across two model scales (SmolLM2-1.7B-Instruct and Llama-3-8B-Instruct), three context lengths (600-7,194 tokens), and up to 15 concurrent agents. PolyKV achieves a stable 2.91x compression ratio across all configurations. On Llama-3-8B with 15 agents sharing a 4K-token context, PolyKV reduces KV cache memory from 19.8 GB to 0.45 GB -- a 97.7% reduction -- while maintaining only +0.57% perplexity degradation and a mean BERTScore F1 of 0.928. PPL delta does not grow with agent count and improves as context length increases, inverting to -0.26% at 1,851 coherent tokens. To our knowledge, no prior work combines a single shared, lossy-compressed KV pool with multi-reader concurrent agent access.", "authors": ["Ishan Patel", "Ishan Joshi"], "categories": ["cs.LG", "cs.CL", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-27", "url": "https://arxiv.org/abs/2604.24971", "pdf_url": "https://arxiv.org/pdf/2604.24971v1", "arxiv_id": "2604.24971", "doi": "10.5281/zenodo.19686729", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ishan1410/PolyKV", "venue": null, "quality_score": 0.65}
{"id": "1411829aa4ced054bd9de6c6c3fc5016566c29464fb05d5ad03537c258a27207", "sources": ["arxiv", "semantic_scholar"], "title": "DepthKV: Layer-Dependent KV Cache Pruning for Long-Context LLM Inference", "abstract": "Long-context reasoning is a critical capability of large language models (LLMs), enabling applications such as long-document understanding, summarization, and code generation. However, efficient autoregressive inference relies on the key-value (KV) cache, whose memory footprint grows linearly with sequence length, leading to a major memory bottleneck. To mitigate this overhead, KV cache pruning methods discard cached tokens with low attention scores during inference. Most existing methods apply a uniform pruning ratio across layers, implicitly assuming that all layers contribute equally to overall model performance. We show that this assumption is suboptimal, as layers differ significantly in their sensitivity to pruning. We propose DepthKV, a layer-dependent pruning framework that allocates a fixed global KV budget across layers based on their sensitivity, rather than using a uniform allocation. Across multiple models and tasks, DepthKV consistently outperforms uniform pruning at the same global pruning ratio, demonstrating more effective utilization of the KV cache budget through layer-dependent allocation.", "authors": ["Zahra Dehghanighobadi", "Asja Fischer"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-27", "url": "https://arxiv.org/abs/2604.24647", "pdf_url": "https://arxiv.org/pdf/2604.24647v1", "arxiv_id": "2604.24647", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "0b4a79b5f4c35dfd1dc79a5feaf70ff7115e0de0550917d0aa554424d24131d8", "sources": ["arxiv", "semantic_scholar"], "title": "LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models", "abstract": "Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in visual understanding and reasoning, but they also impose significant computational burdens due to long visual sequence inputs. Recent works address this issue by pruning unimportant visual tokens, achieving substantial computational reduction while maintaining model performance. The core of token pruning lies in determining token importance, with current approaches primarily relying on attention scores from vision encoders or Large Language Models (LLMs). In this paper, we analyze the effectiveness of attention mechanisms in both vision encoders and LLMs. We find that vision encoders suffer from attention sink, leading to poor focus on informative foreground regions, while in LLMs, although prior studies have identified attention bias toward token positions, text-to-vision attention demonstrates resistance to this bias and enables effective pruning guidance in middle layers. Based on these observations, we propose LearnPruner, a two-stage token pruning framework that first removes redundant vision tokens via a learnable pruning module after the vision encoder, then retains only task-relevant tokens in the LLM's middle layer. Experimental results show that our LearnPruner can preserve approximately 95% of the original performance while using only 5.5% of vision tokens, and achieve 3.2$\\times$ inference acceleration, demonstrating a superior accuracy-efficiency trade-off.", "authors": ["Rinyoichi Takezoe", "Yaqian Li", "Zihao Bo", "Anzhou Hou", "Mo Guang", "Kaiwen Long"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-27", "url": "https://arxiv.org/abs/2604.23950", "pdf_url": "https://arxiv.org/pdf/2604.23950v1", "arxiv_id": "2604.23950", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "f0240c09ec7a7a72930bc54a425e48eb4f65c8c4999c418282c85e1db0eae883", "sources": ["arxiv", "semantic_scholar"], "title": "Statistical Inference and Quality Measures of KV Cache Quantisations Inspired by TurboQuant", "abstract": "We analyse three KV cache quantization schemes under a fair bit budget: \\textbf{KV} (scalar MSE baseline), \\textbf{KQV} (WHT + MSE on $K$; WHT + MSE + QJL on $V$), and \\textbf{QKQV} (WHT + MSE + QJL on both). Starting from the Beta distribution on the hypersphere, we trace how QJL on $K$ inflates inner product variance by $π/2$, which softmax amplifies nonlinearly via Jensen's inequality, and we present statistical inference and information metrics to highlight practical differences. Three empirical findings emerge. (1)~At $n=4$ (the practically dominant budget), KQV wins on every measure -- KL divergence, geometric $K$ error, and 6D distance -- across all distributions and ranks tested. (2)~The K--V asymmetry is unconditional: QKQV is consistently worse than KQV in KL divergence at every budget and distribution. (3)~A budget-dependent crossover exists: QKQV achieves better geometric $K$ reconstruction at $n \\in \\{2,3,5\\}$, KQV at $n \\in \\{4,6\\}$, invariant to rank and tail weight -- an open rate-distortion problem. $\\mathrm{KL}(p_{\\mathrm{ref}} \\| p_{\\mathrm{quant}})$, K-only by construction, bridges K direction error to routing corruption and output collapse. We present a sufficient condition when the Jensen mechanism amplifies superlinearly through the softmax. At $n \\in \\{2,3,5\\}$, QKQV wins geometrically because this assumption does not bind. At $n=4$, elevated K error and KL divergence for QKQV strongly suggest the Jensen mechanism is the operative cause of the crossover, providing a new perspective and explanation.", "authors": ["Paolo D'Alberto"], "categories": ["cs.LG", "cs.IT", "cs.MS"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-27", "url": "https://arxiv.org/abs/2605.08114", "pdf_url": "https://arxiv.org/pdf/2605.08114v1", "arxiv_id": "2605.08114", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "819d2eb3f0cc7702fb53daebe933e4d2df11701be6e3d8b852cd64f6db70677c", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Heterogeneous Grouped Experts for Language Modeling", "abstract": "Large Language Models (LLMs) based on Mixture-of-Experts (MoE) are pivotal in industrial applications for their ability to scale performance efficiently. However, standard MoEs enforce uniform expert sizes,creating a rigidity that fails to align computational costs with varying token-level complexity. While heterogeneous expert architectures attempt to address this by diversifying expert sizes, they often suffer from significant system-level challenges, specifically unbalanced GPU utilization and inefficient parameter utilization, which hinder practical deployment. To bridge the gap between theoretical heterogeneity and robust industrial application, we propose Mixture of Heterogeneous Grouped Experts (MoHGE) which introduces a two-level routing mechanism to enable flexible, resource-aware expert combinations. To optimize inference efficiency, we propose a Group-Wise Auxiliary Loss, which dynamically steers tokens to the most parameter-efficient expert groups based on task difficulty. To address the critical deployment challenge of GPU load balancing, we introduce an All-size Group-decoupling Allocation strategy coupled with an Intra-Group Experts Auxiliary Loss. These mechanisms collectively ensure uniform computation distribution across GPUs. Extensive evaluations demonstrate that MoHGE matches the performance of MoE architectures while reducing the total parameters by approximately 20% and maintaining balanced GPU utilization. Our work establishes a scalable paradigm for resource-efficient MoE design, offering a practical solution for optimizing inference costs in real-world scenarios. The code is publicly available at https://github.com/UnicomAI/MoHGE.", "authors": ["Zhicheng Ma", "Xiang Liu", "Zhaoxiang Liu", "Ning Wang", "Yi Shen", "Kai Wang", "Shuming Shi", "Shiguo Lian"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-25", "url": "https://arxiv.org/abs/2604.23108", "pdf_url": "https://arxiv.org/pdf/2604.23108v2", "arxiv_id": "2604.23108", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/UnicomAI/MoHGE", "venue": null, "quality_score": 0.65}
{"id": "3170acf79e656e981aa20fe8fdfbb66b68e9cd120a20de1cf988694d4ee7c9e0", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling Multi-Node Mixture-of-Experts Inference Using Expert Activation Patterns", "abstract": "Most recent state-of-the-art (SOTA) large language models (LLMs) use Mixture-of-Experts (MoE) architectures to scale model capacity without proportional per-token compute, enabling higher-quality outputs at manageable serving costs. However, MoE inference at scale is fundamentally bottlenecked by expert load imbalance and inefficient token routing, especially in multi-node deployments where tokens are not guaranteed to be routed to local experts, resulting in significant inter-node all-to-all communication overhead. To systematically characterize these challenges, we profile SOTA open-source MoE models, including Llama 4 Maverick, DeepSeek V3-671B, and Qwen3-230B-A22B, on various datasets and collected over 100k real expert activation traces. Upon studying the expert activation patterns, we uncover various persistent properties across all the frontier MoE models: variable expert load imbalance, domain-specific expert activation where expert popularity shifts across task families (code, math, chat, general), and a strong correlation between prefill and decode expert activations. Motivated by these findings, we propose workload-aware micro-batch grouping and an expert placement strategy to maximize token locality to the destination expert, thereby reducing inter-node communication. Across models and datasets, these optimizations help reduce all2all communication data up to 20, resulting in lower MoE decode latency and better accelerator utilization.", "authors": ["Abhimanyu Bambhaniya", "Geonhwa Jeong", "Jason Park", "Jiecao Yu", "Jaewon Lee", "Pengchao Wang", "Changkyu Kim", "Chunqiang Tang", "Tushar Krishna"], "categories": ["cs.LG", "cs.AI", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-25", "url": "https://arxiv.org/abs/2604.23150", "pdf_url": "https://arxiv.org/pdf/2604.23150v1", "arxiv_id": "2604.23150", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.65}
{"id": "a86313b756645e110dc904f6cd6efe0971d58cdd910ac3152dbb2ddbae6d1cf2", "sources": ["arxiv", "semantic_scholar"], "title": "Where Should LoRA Go? Component-Type Placement in Hybrid Language Models", "abstract": "Hybrid language models that interleave attention with recurrent components are increasingly competitive with pure Transformers, yet standard LoRA practice applies adapters uniformly without considering the distinct functional roles of each component type. We systematically study component-type LoRA placement across two hybrid architectures -- Qwen3.5-0.8B (sequential, GatedDeltaNet + softmax attention) and Falcon-H1-0.5B (parallel, Mamba-2 SSM + attention) -- fine-tuned on three domains and evaluated on five benchmarks. We find that the attention pathway -- despite being the minority component -- consistently outperforms full-model adaptation with 5-10x fewer trainable parameters. Crucially, adapting the recurrent backbone is destructive in sequential hybrids (-14.8 pp on GSM8K) but constructive in parallel ones (+8.6 pp). We further document a transfer asymmetry: parallel hybrids exhibit positive cross-task transfer while sequential hybrids suffer catastrophic forgetting. These results establish that hybrid topology fundamentally determines adaptation response, and that component-aware LoRA placement is a necessary design dimension for hybrid architectures.", "authors": ["Hector Borobia", "Elies Seguí-Mas", "Guillermina Tormo-Carbó"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-24", "url": "https://arxiv.org/abs/2604.22127", "pdf_url": "https://arxiv.org/pdf/2604.22127v1", "arxiv_id": "2604.22127", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/hecboar/lora-placement-hybrid", "venue": null, "quality_score": 0.65}
{"id": "55caf3cf72eeaa5ce78d52e1fbae8cb056bc4095a00fa07f0c9a64fb97f87395", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Cross-Domain Evaluation for Face Forgery Detection with Semantic Fine-grained Alignment and Mixture-of-Experts", "abstract": "Nowadays, visual data forgery detection plays an increasingly important role in social and economic security with the rapid development of generative models. Existing face forgery detectors still can't achieve satisfactory performance because of poor generalization ability across datasets. The key factor that led to this phenomenon is the lack of suitable metrics: the commonly used cross-dataset AUC metric fails to reveal an important issue where detection scores may shift significantly across data domains. To explicitly evaluate cross-domain score comparability, we propose \\textbf{Cross-AUC}, an evaluation metric that can compute AUC across dataset pairs by contrasting real samples from one dataset with fake samples from another (and vice versa). It is interesting to find that evaluating representative detectors under the Cross-AUC metric reveals substantial performance drops, exposing an overlooked robustness problem. Besides, we also propose the novel framework \\textbf{S}emantic \\textbf{F}ine-grained \\textbf{A}lignment and \\textbf{M}ixture-of-Experts (\\textbf{SFAM}), consisting of a patch-level image-text alignment module that enhances CLIP's sensitivity to manipulation artifacts, and the facial region mixture-of-experts module, which routes features from different facial regions to specialized experts for region-aware forgery analysis. Extensive qualitative and quantitative experiments on the public datasets prove that the proposed method achieves superior performance compared with the state-of-the-art methods with various suitable metrics.", "authors": ["Yuhan Luo", "Tao Chen", "Decheng Liu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-23", "url": "https://arxiv.org/abs/2604.21478", "pdf_url": "https://arxiv.org/pdf/2604.21478v1", "arxiv_id": "2604.21478", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Yuhan-Luo/Semantic-Fine-grained-Alignment-and-Mixture-of-Experts", "venue": null, "quality_score": 0.65}
{"id": "566f4f5f44350f4e0e6a5acac83ecd1af628a006d254a6ee196e364f2df5bbc8", "sources": ["arxiv", "semantic_scholar"], "title": "Teacher-Guided Routing for Sparse Vision Mixture-of-Experts", "abstract": "Recent progress in deep learning has been driven by increasingly large-scale models, but the resulting computational cost has become a critical bottleneck. Sparse Mixture of Experts (MoE) offers an effective solution by activating only a small subset of experts for each input, achieving high scalability without sacrificing inference speed. Although effective, sparse MoE training exhibits characteristic optimization difficulties. Because the router receives informative gradients only through the experts selected in the forward pass, it suffers from gradient blocking and obtains little information from unselected routes. This limited, highly localized feedback makes it difficult for the router to learn appropriate expert-selection scores and often leads to unstable routing dynamics, such as fluctuating expert assignments during training. To address this issue, we propose TGR-MoE: Teacher-Guided Routing for Sparse Vision Mixture-of-Experts, a simple yet effective method that stabilizes router learning using supervision derived from a pretrained dense teacher model. TGR-MoE constructs a teacher router from the teacher's intermediate representations and uses its routing outputs as pseudo-supervision for the student router, suppressing frequent routing fluctuations during training and enabling knowledge-guided expert selection from the early stages of training. Extensive experiments on ImageNet-1K and CIFAR-100 demonstrate that TGR consistently improves both accuracy and routing consistency, while maintaining stable training even under highly sparse configurations.", "authors": ["Masahiro Kada", "Ryota Yoshihashi", "Satoshi Ikehata", "Rei Kawakami", "Ikuro Sato"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-23", "url": "https://arxiv.org/abs/2604.21330", "pdf_url": "https://arxiv.org/pdf/2604.21330v1", "arxiv_id": "2604.21330", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4e580b9eb81de616c03bfbc6e55cf46c21ecac6f426aceb407af2d9ec5494f08", "sources": ["arxiv", "semantic_scholar"], "title": "SparKV: Overhead-Aware KV Cache Loading for Efficient On-Device LLM Inference", "abstract": "Efficient inference for on-device Large Language Models (LLMs) remains challenging due to limited hardware resources and the high cost of the prefill stage, which processes the full input context to construct Key-Value (KV) caches. We present SparKV, an adaptive KV loading framework that combines cloud-based KV streaming with on-device computation. SparKV models the cost of individual KV chunks and decides whether each chunk should be streamed or computed locally, while overlapping the two execution paths to reduce latency. To handle fluctuations in wireless connectivity and edge resource availability, SparKV further refines offline-generated schedules at runtime to rebalance communication and computation costs. Experiments across diverse datasets, LLMs, and edge devices show that SparKV reduces Time-to-First-Token by 1.3$x-5.1x with negligible impact on response quality, while lowering per-request energy consumption by 1.5x to 3.3x, demonstrating its robustness and practicality for real-world on-device deployment.", "authors": ["Hongyao Liu", "Liuqun Zhai", "Junyi Wang", "Zhengru Fang"], "categories": ["cs.NI", "cs.AI", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-23", "url": "https://arxiv.org/abs/2604.21231", "pdf_url": "https://arxiv.org/pdf/2604.21231v2", "arxiv_id": "2604.21231", "doi": "10.1109/jiot.2026.3695763", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Internet of Things Journal", "quality_score": 0.55}
{"id": "15a7575896287527c860db3737c11b2c984e48c6d38912d9f94ebf467a5703c7", "sources": ["arxiv", "semantic_scholar"], "title": "RateQuant: Optimal Mixed-Precision KV Cache Quantization via Rate-Distortion Theory", "abstract": "Large language models cache all previously computed key-value (KV) pairs during generation, and this KV cache grows linearly with sequence length, making it a primary memory bottleneck for serving. Quantizing the KV cache to fewer bits reduces this cost, yet all current quantizers assign the same bit-width to every attention head, ignoring the large variation in head importance. A natural idea is to allocate more bits to important heads and fewer to the rest. We show, however, that such mixed-precision allocation has a hidden pitfall: each quantizer follows a different distortion curve D(b)=alpha*beta^{-b}, and the decay rate beta varies from 3.6 to 5.3 across quantizer designs. Applying one quantizer's distortion model to another inverts the allocation order and makes performance worse than uniform quantization. We call this failure mode distortion model mismatch and propose RateQuant to resolve it. RateQuant fits a per-quantizer distortion model from a small calibration set, then solves the resulting bit-allocation problem in closed form via reverse waterfilling from rate-distortion theory. On Qwen3-8B at 2.5 average bits, calibrated RateQuant reduces KIVI's perplexity from 49.3 to 14.9 (70% reduction) and improves QuaRot by 6.6 PPL. The entire calibration takes 1.6 s on a single GPU and adds zero overhead at inference time.", "authors": ["Fei Zuo", "Zikang Zhou", "Hao Cong", "Xiaoyan Xi", "Ho Fai Leung"], "categories": ["cs.LG", "cs.CL", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-22", "url": "https://arxiv.org/abs/2605.06675", "pdf_url": "https://arxiv.org/pdf/2605.06675v1", "arxiv_id": "2605.06675", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d179d517ea3d471ee520d77c3c408969d9fe11bacb26398a4af2eda2865f8e51", "sources": ["arxiv", "semantic_scholar"], "title": "LKV: End-to-End Learning of Head-wise Budgets and Token Selection for LLM KV Cache Eviction", "abstract": "Long-context inference in Large Language Models (LLMs) is bottlenecked by the linear growth of Key-Value (KV) cache memory. Existing KV cache compression paradigms are fundamentally limited by heuristics: heuristic budgeting relies on statistical priors rather than task objectives, causing resource misallocation, while heuristic selection relies on coupled query-key interactions or static inductive biases (e.g., attention sinks). To address this limitation, we introduce LKV (Learned KV Eviction), which formulates KV compression as an end-to-end differentiable optimization problem. LKV integrates LKV-H to learn task-optimized global budgets, and LKV-T to derive intrinsic KV importance without materializing attention matrices. This design bypasses heuristic proxies, strictly aligning compression with task objectives. Extensive evaluations demonstrate that LKV achieves state-of-the-art performance on both LongBench and RULER benchmarks at high compression rates. In particular, on LongBench, LKV achieves near-lossless performance with only 15\\% KV cache retention. Crucially, our analysis identifies learned budgeting as the dominant driver of fidelity, demonstrating that data-driven allocation is essential to overcome the limitations of hand-crafted heuristics.", "authors": ["Enshuai Zhou", "Yifan Hao", "Chao Wang", "Rui Zhang", "Di Huang", "Jiaming Guo", "Xing Hu", "Zidong Du", "Qi Guo", "Yunji Chen"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-22", "url": "https://arxiv.org/abs/2605.06676", "pdf_url": "https://arxiv.org/pdf/2605.06676v1", "arxiv_id": "2605.06676", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "87286d3810122409fce227910062f2e4bc7eef7c9da7c8acbe7a872474ad2fb0", "sources": ["arxiv", "semantic_scholar"], "title": "On Bayesian Softmax-Gated Mixture-of-Experts Models", "abstract": "Mixture-of-experts models provide a flexible framework for learning complex probabilistic input-output relationships by combining multiple expert models through an input-dependent gating mechanism. These models have become increasingly prominent in modern machine learning, yet their theoretical properties in the Bayesian framework remain largely unexplored. In this paper, we study Bayesian mixture-of-experts models, focusing on the ubiquitous softmax-based gating mechanism. Specifically, we investigate the asymptotic behavior of the posterior distribution for three fundamental statistical tasks: density estimation, parameter estimation, and model selection. First, we establish posterior contraction rates for density estimation, both in the regimes with a fixed, known number of experts and with a random learnable number of experts. We then analyze parameter estimation and derive convergence guarantees based on tailored Voronoi-type losses, which account for the complex identifiability structure of mixture-of-experts models. Finally, we propose and analyze two complementary strategies for selecting the number of experts. Taken together, these results provide one of the first systematic theoretical analyses of Bayesian mixture-of-experts models with softmax gating, and yield several theory-grounded insights for practical model design.", "authors": ["Nicola Bariletto", "Huy Nguyen", "Nhat Ho", "Alessandro Rinaldo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-04-22", "url": "https://arxiv.org/abs/2604.20551", "pdf_url": "https://arxiv.org/pdf/2604.20551v1", "arxiv_id": "2604.20551", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "afc057cd18a848720641e0ca58eb03084d36e61e87f3501fb0f26b2086dd9175", "sources": ["arxiv", "semantic_scholar"], "title": "Temporally Extended Mixture-of-Experts Models", "abstract": "Mixture-of-Experts models, now popular for scaling capacity at fixed inference speed, switch experts at nearly every token. Once a model outgrows available GPU memory, this churn can render optimizations like offloading and pre-fetching ineffective. We make the case that the options framework in reinforcement learning is a perfect match to tackle this problem, and argue for temporally extended mixture-of-experts layers. Building on the option-critic framework with deliberation costs, we add a controller to each layer that learns when to switch expert sets and which to load. By applying this to gpt-oss-20b with low-rank adapters and a self-distillation reward, our method reduces switch rates from over 50% to below 5% while retaining up to 90% of base-model accuracy on MATH, MMLU, and MMMLU. This shows that even existing pre-trained models can be converted to temporally extended MoEs with lightweight training, with the deliberation cost allowing model trainers to trade off switching rates against capability. We hope this opens a principled path, grounded in the options framework, for memory-efficient serving and continual learning in ever-growing MoE models.", "authors": ["Zeyu Shen", "Peter Henderson"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-22", "url": "https://arxiv.org/abs/2604.20156", "pdf_url": "https://arxiv.org/pdf/2604.20156v1", "arxiv_id": "2604.20156", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "09598b48d6281a64768ac64f4d7f6640c535b18c81092f3db46baaf90b3d16f0", "sources": ["arxiv", "semantic_scholar"], "title": "SAW-INT4: System-Aware 4-Bit KV-Cache Quantization for Real-World LLM Serving", "abstract": "KV-cache memory is a major bottleneck in real-world LLM serving, where systems must simultaneously support latency-sensitive small-batch requests and high-throughput concurrent workloads. Although many KV-cache compression methods improve offline accuracy or compression ratio, they often violate practical serving constraints such as paged memory layouts, regular memory access, and fused attention execution, limiting their effectiveness in deployment. In this work, we identify the minimal set of 4-bit KV-cache quantization methods that remain viable under these constraints. Our central finding is that a simple design--token-wise INT4 quantization with block-diagonal Hadamard rotation--consistently achieves the best accuracy-efficiency trade-off. Across multiple models and benchmarks, this approach recovers nearly all of the accuracy lost by naive INT4, while more complex methods such as vector quantization and Hessian-aware quantization provide only marginal additional gains once serving compatibility is taken into account. To make this practical, we implement a fused rotation-quantization kernel that integrates directly into paged KV-cache layouts and introduces zero measurable end-to-end overhead, matching plain INT4 throughput across concurrency levels. Our results show that effective KV-cache compression is fundamentally a systems co-design problem: under real serving constraints, lightweight block-diagonal Hadamard rotation is a viable method that delivers near-lossless accuracy without sacrificing serving efficiency.", "authors": ["Jinda Jia", "Jisen Li", "Zhongzhu Zhou", "Jung Hwan Heo", "Jue Wang", "Tri Dao", "Shuaiwen Leon Song", "Ben Athiwaratkun", "Chenfeng Xu", "Tianyi Zhang", "Xiaoxia Wu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-21", "url": "https://arxiv.org/abs/2604.19157", "pdf_url": "https://arxiv.org/pdf/2604.19157v1", "arxiv_id": "2604.19157", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "b55f3714e3706ee1e55f8d23461f73b224c12a2cf2d098d3d5a0ac94f106331d", "sources": ["arxiv", "semantic_scholar"], "title": "SAMoRA: Semantic-Aware Mixture of LoRA Experts for Task-Adaptive Learning", "abstract": "The combination of Mixture-of-Experts (MoE) and Low-Rank Adaptation (LoRA) has shown significant potential for enhancing the multi-task learning capabilities of Large Language Models. However, existing methods face two primary challenges: (1)Imprecise Routing in the current MoE-LoRA method fails to explicitly match input semantics with expert capabilities, leading to weak expert specialization. (2)Uniform weight fusion strategies struggle to provide adaptive update strengths, overlooking the varying complexity of different tasks. To address these limitations, we propose SAMoRA (Semantic-Aware Mixture of LoRA Experts), a novel parameter-efficient fine-tuning framework tailored for task-adaptive learning. Specifically, A Semantic-Aware Router is proposed to explicitly align textual semantics with the most suitable experts for precise routing. A Task-Adaptive Scaling mechanism is designed to regulate expert contributions based on specific task requirements dynamically. In addition, a novel regularization objective is proposed to jointly promote expert specialization and effective scaling. Extensive experiments on multiple multi-task benchmarks demonstrate that SAMoRA significantly outperforms the state-of-the-art methods and holds excellent task generalization capabilities. Code is available at https://github.com/boyan-code/SAMoRA", "authors": ["Boyan Shi", "Wei Chen", "Shuyuan Zhao", "Junfeng Shen", "Shengnan Guo", "Shaojiang Wang", "Huaiyu Wan"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-21", "url": "https://arxiv.org/abs/2604.19048", "pdf_url": "https://arxiv.org/pdf/2604.19048v1", "arxiv_id": "2604.19048", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/boyan-code/SAMoRA", "venue": null, "quality_score": 0.65}
{"id": "9fd0f2d5d98f1cc2c5f55b96e02f3498db094a36d60be6973bc9b544ac9e590a", "sources": ["arxiv", "semantic_scholar"], "title": "DASH-KV: Accelerating Long-Context LLM Inference via Asymmetric KV Cache Hashing", "abstract": "The quadratic computational complexity of the standard attention mechanism constitutes a fundamental bottleneck for large language models in long-context inference. While existing KV cache compression methods alleviate memory pressure, they often sacrifice generation quality and fail to address the high overhead of floating-point arithmetic. This paper introduces DASH-KV, an innovative acceleration framework that reformulates attention as approximate nearest-neighbor search via asymmetric deep hashing. Under this paradigm, we design an asymmetric encoding architecture that differentially maps queries and keys to account for their distinctions in precision and reuse characteristics. To balance efficiency and accuracy, we further introduce a dynamic mixed-precision mechanism that adaptively retains full-precision computation for critical tokens. Extensive experiments on LongBench demonstrate that DASH-KV significantly outperforms state-of-the-art baseline methods while matching the performance of full attention, all while reducing inference complexity from O(N^2) to linear O(N).", "authors": ["Jinyu Guo", "Zhihan Zhang", "Jiehui Xie", "Md. Tamim Iqbal", "Dongshen Han", "Lik-Hang Lee", "Sung-Ho Bae", "Jie Zou", "Yang Yang", "Chaoning Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-21", "url": "https://arxiv.org/abs/2604.19351", "pdf_url": "https://arxiv.org/pdf/2604.19351v4", "arxiv_id": "2604.19351", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4bf5ad019c0d6007cf73a483f2544581c4a7d6e6b9926707837a1b09660fe97b", "sources": ["arxiv", "semantic_scholar"], "title": "Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) has become the dominant architecture for scaling large language models: frontier models routinely decouple total parameters from per-token computation through sparse expert routing. Scaling laws show that under fixed active computation, model quality scales predictably with total parameters, and MoEs realize this by increasing expert count. However, training large MoEs is expensive, as memory requirements and inter-device communication both scale with total parameter count. We propose expert upcycling, a method for progressively expanding MoE capacity by increasing the number of experts during continued pre-training (CPT). Given a trained E-expert model, the upcycling operator constructs an mE-expert model through expert duplication and router extension while holding top-K routing fixed, preserving per-token inference cost. Duplication provides a warm initialization: the expanded model inherits the source checkpoint's learned representations, starting from a substantially lower loss than random initialization. Subsequent CPT then breaks the symmetry among duplicated experts to drive specialization. We formalize the upcycling operator and develop a theoretical framework decomposing the quality gap into a capacity term and an initialization term. We further introduce utility-based expert selection, which uses gradient-based importance scores to guide non-uniform duplication, more than tripling gap closure when CPT is limited. In our 7B-13B total parameter experiments, the upcycled model matches the fixed-size baseline on validation loss while saving 32% of GPU hours. Comprehensive ablations across model scales, activation ratios, MoE architectures, and training budgets yield a practical recipe for deploying expert upcycling, establishing it as a principled, compute-efficient alternative to training large MoE models from scratch.", "authors": ["Chaitanya Dwivedi", "Binxuan Huang", "Himanshu Gupta", "Pratik Jayarao", "Neeraj Varshney", "Bing Yin"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-21", "url": "https://arxiv.org/abs/2604.19835", "pdf_url": "https://arxiv.org/pdf/2604.19835v2", "arxiv_id": "2604.19835", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "14baf7547811cc8ad26e52b2f5ee178dc96f4c7e46b9e7e7a150b3f59e56cd75", "sources": ["arxiv", "semantic_scholar"], "title": "GRASPrune: Global Gating for Budgeted Structured Pruning of Large Language Models", "abstract": "Large language models (LLMs) are expensive to serve because model parameters, attention computation, and KV caches impose substantial memory and latency costs. We present GRASPrune, a structured pruning framework applied after pretraining that jointly prunes FFN channels and KV head groups under a single global budget. Instead of learning importance scores without constraints and applying the budget only after training, GRASPrune learns lightweight gate scores with a projected straight-through estimator that enforces a hard mask satisfying the budget at every step while keeping the backbone weights frozen. After the mask is fixed, we calibrate scaling factors on the retained units to mitigate scale mismatch caused by pruning, and fold these factors into the pruned weights to obtain a smaller dense checkpoint with no extra parameters at inference. On LLaMA-2-7B, GRASPrune removes 50% of parameters and achieves 12.18 perplexity on WikiText-2 while maintaining competitive average zero-shot accuracy on five benchmarks, using four epochs on 512 unlabeled calibration sequences on a single NVIDIA A100 80GB GPU without any full model fine-tuning.", "authors": ["Ziyang Wang", "Jiangfeng Xiao", "Chuan Xiao", "Ruoxiang Li", "Rui Mao", "Jianbin Qin"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-21", "url": "https://arxiv.org/abs/2604.19398", "pdf_url": "https://arxiv.org/pdf/2604.19398v1", "arxiv_id": "2604.19398", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "d97b7228a5e6972e30bf0371705bce4431a378635a81fcd6177d9e358b73dc45", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-nD: Per-Layer Mixture-of-Experts Routing for Multi-Axis KV Cache Compression", "abstract": "KV cache memory is the dominant bottleneck for long-context LLM inference. Existing compression methods each act on a single axis of the four-dimensional KV tensor -- token eviction (sequence), quantization (precision), low-rank projection (head dimension), or cross-layer sharing -- but apply the same recipe to every layer. We show that this homogeneity leaves accuracy on the table: different layers respond very differently to each compression operation, and the optimal per-layer mix of eviction and quantization is far from uniform. We propose MoE-nD, a mixture-of-experts framework that routes each layer to its own (eviction-ratio, K-bits, V-bits) tuple under a global memory budget. An offline-calibrated greedy solver chooses the routing that minimizes predicted quality loss; at inference time, per-layer heterogeneous eviction and quantization are applied jointly through a single attention patch. On a 4-task subset of LongBench-v1 (16k inputs, n=50 per task, adapted reasoning-model protocol; see section Experiments), MoE-nD's hetero variant matches our uncompressed 1.9~GB baseline at 14x compression (136~MB) while every other compressed baseline we tested (1d, 2d_uniform, 2d) at comparable or smaller memory stays under 8/100. The gains hold on AIME reasoning benchmarks (+6 to +27 pts over the strongest per-layer-quantization baseline across eight configurations). Two null results -- MATH-500 and LongBench's TREC -- share a principled cause (short inputs, solver picks keep=1.0 on most layers), cleanly characterizing when per-layer eviction routing has headroom to help.", "authors": ["Libo Sun", "Peixiong He", "Po-Wei Harn", "Xiao Qin"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.17695", "pdf_url": "https://arxiv.org/pdf/2604.17695v1", "arxiv_id": "2604.17695", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "944bb7c1a2a3c856e511fd3fdd193924734ced8b18c72cdc0f3ef09e32d03dd1", "sources": ["arxiv", "semantic_scholar"], "title": "Evolutionary Negative Module Pruning for Better LoRA Merging", "abstract": "Merging multiple Low-Rank Adaptation (LoRA) experts into a single backbone is a promising approach for efficient multi-task deployment. While existing methods strive to alleviate interference via weight interpolation or subspace alignment, they rest upon the implicit assumption that all LoRA matrices contribute constructively to the merged model. In this paper, we uncover a critical bottleneck in current merging paradigms: the existence of $\\textit{negative modules}$ -- specific LoRA layers that inherently degrade global performance upon merging. We propose $\\textbf{E}$volutionary $\\textbf{N}$egative $\\textbf{M}$odule $\\textbf{P}$runing ($\\textbf{ENMP}$), a plug-and-play LoRA pruning method to locate and exclude these detrimental modules prior to merging. By leveraging an evolutionary search strategy, ENMP effectively navigates the discrete, non-differentiable landscape of module selection to identify optimal pruning configurations. Extensive evaluations demonstrate that ENMP consistently boosts the performance of existing merging algorithms, achieving a new state-of-the-art across both language and vision domains. Code is available at https://github.com/CaoAnda/ENMP-LoRAMerging.", "authors": ["Anda Cao", "Zhuo Gou", "Yi Wang", "Kaixuan Chen", "Yu Wang", "Can Wang", "Mingli Song", "Jie Song"], "categories": ["cs.AI", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.17753", "pdf_url": "https://arxiv.org/pdf/2604.17753v1", "arxiv_id": "2604.17753", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/CaoAnda/ENMP-LoRAMerging", "venue": null, "quality_score": 0.65}
{"id": "4cc7b39a586590a37a12212890b059bc9748fb4600d52d5c57413b9546ad8c2b", "sources": ["arxiv", "semantic_scholar"], "title": "How Much Cache Does Reasoning Need? Depth-Cache Tradeoffs in KV-Compressed Transformers", "abstract": "The key-value (KV) cache is the dominant memory bottleneck during Transformer inference, yet little is known theoretically about how aggressively it can be compressed before multi-step reasoning degrades. We study this through $k$-hop pointer chasing on $n$ tokens under a shared KV cache of size $s$, attention dimension $m$, $H$ heads, $p$-bit precision, and a locality-respecting cache controller (satisfied by all standard KV-compression methods). We give three results. (1) Product depth lower bound (conjectured). We conjecture that any such Transformer ($n \\geq 4k$, $s \\leq \\sqrt{n}/4$) requires depth $L = Ω(\\lceil k/s \\rceil \\cdot \\lceil \\log_2 n/(Hmp) \\rceil)$, and isolate the sole remaining gap as a probabilistic step on the joint distribution of cache trace and pointer chain. Unconditionally, we prove a matching upper bound $L = O(\\min(k, \\lceil k/s \\rceil \\log s) \\cdot \\log n/(mp))$ via windowed pointer doubling, and a max-bound $L = Ω(\\max(\\lceil k/s \\rceil, \\log n/(Hmp)))$. Closing the conjecture amounts to upgrading max to product. (2) Bandwidth barrier. The product bound binds only when $Hmp \\lesssim \\log n$. Any lower bound provable via per-window distinguishability counting -- including reachability, bandwidth, and combinations -- cannot exceed $\\lceil k/s \\rceil$ once $Hmp \\geq \\log_2 n$. Breaking this requires lifting unconditional communication-complexity bounds for pointer chasing to Cache-Transformer depth. (3) Adaptive vs oblivious error scaling. Under random cache over $T = \\lceil \\log_2 k \\rceil$ doubling stages, oblivious caches give $\\Pr[\\mathcal{E}] \\leq (s/(n-T))^T + 2T^3/n$ (exponential in $T$), while adaptive locality-respecting caches achieve $\\Pr[\\mathcal{E}] = s/n$ exactly, independent of $T$. The $Ω((n/s)^{T-1})$ separation explains why heavy-hitter eviction empirically dominates random eviction for multi-hop reasoning.", "authors": ["Xiao Wang"], "categories": ["cs.LG", "cs.AI", "cs.CC"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.17935", "pdf_url": "https://arxiv.org/pdf/2604.17935v1", "arxiv_id": "2604.17935", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e3309643bc3fcd71d8e4043679f3ad0cc54123a6bb8495797bb20bdbb6ad3099", "sources": ["arxiv", "semantic_scholar"], "title": "Soft Label Pruning and Quantization for Large-Scale Dataset Distillation", "abstract": "Large-scale dataset distillation requires storing auxiliary soft labels that can be 30-40x larger on ImageNet-1K and 200x larger on ImageNet-21K than the condensed images, undermining the goal of dataset compression. We identify two fundamental issues necessitating such extensive labels: (1) insufficient image diversity, where high within-class similarity in synthetic images requires extensive augmentation, and (2) insufficient supervision diversity, where limited variety in supervisory signals during training leads to performance degradation at high compression rates. To address these challenges, we propose Label Pruning and Quantization for Large-scale Distillation (LPQLD). We enhance image diversity via class-wise batching and batch-normalization supervision during synthesis. For supervision diversity, we introduce Label Pruning with Dynamic Knowledge Reuse to improve label-per-augmentation diversity, and Label Quantization with Calibrated Student-Teacher Alignment to improve augmentation-per-image diversity. Our approach reduces soft label storage by 78x on ImageNet-1K and 500x on ImageNet-21K while improving accuracy by up to 7.2% and 2.8%, respectively. Extensive experiments validate the superiority of LPQLD across different network architectures and dataset distillation methods. Code is available at https://github.com/he-y/soft-label-pruning-quantization-for-dataset-distillation.", "authors": ["Xiao Lingao", "Yang He"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Medicine", "Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.18135", "pdf_url": "https://arxiv.org/pdf/2604.18135v1", "arxiv_id": "2604.18135", "doi": "10.1109/TPAMI.2026.3664488", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/he-y/soft-label-pruning-quantization-for-dataset-distillation", "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.85}
{"id": "7b4a7be9493a7453dc6a9ceefef97e09a7b065ae832c493c599f7400da41c581", "sources": ["arxiv", "semantic_scholar"], "title": "Six Llamas: Comparative Religious Ethics Through LoRA-Adapted Language Models", "abstract": "We present Six Llamas, a comparative study examining whether large language models fine-tuned on distinct religious corpora encode systematically different patterns of ethical reasoning. Six variants of Meta-Llama-3.1-8B are constructed: one unmodified control and five LoRA-adapted models trained exclusively on the sacred and theological texts of Christianity, Islam, Judaism, Hinduism, or Buddhism. All six models are probed with an identical battery of 17 standardized ethical prompts spanning moral dilemmas, game-theoretic scenarios, public policy questions, and moral-psychological self-assessments. To assess robustness and reproducibility, we implement a multi-temperature sampling design spanning ten temperature settings. We compute response consistency metrics, pairwise inter-model agreement rates, temperature sensitivity coefficients across four prompt domains, and run-to-run stability analyses. Findings show that LoRA-adapted models produce ethical reasoning patterns that are (a) systematically differentiated from the base model, (b) consistent with the moral logics of their training traditions, (c) structured along interpretable dimensions in moral-philosophical space, (d) core ethical positions remain stable across temperature variations for high-consensus dilemmas. The Trolley Problem achieves 100% consistency across all models and temperatures, while (e) tradition-specific divergence intensifies at higher temperatures in morally contested domains, and (f) the base model exhibits the highest overall response consistency (mean 88.3%), suggesting LoRA adaptation introduces both tradition-specific signal and increased sampling sensitivity. The study offers a proof-of-concept for the condensate comparative method using differentially trained language models as instruments for cultural and ethical analysis and identifies specific criteria for falsification and planned extensions.", "authors": ["Chad Coleman", "W. Russell Neuman", "Manan Shah", "Ali Dasdan", "Matthew Crispi", "Morris Chiang", "Zack Leitman", "Mustafa Poonawala"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.18404", "pdf_url": "https://arxiv.org/pdf/2604.18404v1", "arxiv_id": "2604.18404", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "535f207b12e3c5db6d356d572ff15bcfdce5c0c227949ebf7a292ee113dd989e", "sources": ["arxiv", "semantic_scholar"], "title": "Domain-Specialized Object Detection via Model-Level Mixtures of Experts", "abstract": "Mixture-of-Experts (MoE) models provide a structured approach to combining specialized neural networks and offer greater interpretability than conventional ensembles. While MoEs have been successfully applied to image classification and semantic segmentation, their use in object detection remains limited due to challenges in merging dense and structured predictions. In this work, we investigate model-level mixtures of object detectors and analyze their suitability for improving performance and interpretability in object detection. We propose an MoE architecture that combines YOLO-based detectors trained on semantically disjoint data subsets, with a learned gating network that dynamically weights expert contributions. We study different strategies for fusing detection outputs and for training the gating mechanism, including balancing losses to prevent expert collapse. Experiments on the BDD100K dataset demonstrate that the proposed MoE consistently outperforms standard ensemble approaches and provides insights into expert specialization across domains, highlighting model-level MoEs as a viable alternative to traditional ensembling for object detection. Our code is available at https://github.com/KASTEL-MobilityLab/mixtures-of-experts/.", "authors": ["Svetlana Pavlitska", "Malte Stüven", "Beyza Keskin", "J. Marius Zöllner"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.18256", "pdf_url": "https://arxiv.org/pdf/2604.18256v1", "arxiv_id": "2604.18256", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/KASTEL-MobilityLab/mixtures-of-experts/", "venue": null, "quality_score": 0.65}
{"id": "158dc6ca943fb8437ce10b495f2a7f99848c5c18a831e6c4e8485e87ed02b799", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Joint Quantization and Token Pruning of Vision-Language Models", "abstract": "Deploying Vision-Language Models (VLMs) under aggressive low-bit inference remains challenging because inference cost is dominated by the long visual-token prefix during prefill and the growing KV cache during autoregressive decoding. Token pruning and low-bit quantization are complementary for reducing these costs, yet naive stage-wise combinations are often brittle due to a mismatch between quantization calibration and pruning execution. We present a collaborative quantization-and-pruning framework that unifies low-bit inference and deterministic visual-token pruning in a single deployable pipeline. The framework introduces the \\textbf{Q}uantization \\textbf{U}nified \\textbf{O}ffline \\textbf{T}oken \\textbf{A}llocator (\\textbf{QUOTA}), which converts low-bit calibration signals into a layer-wise token allocation schedule and materializes it as a pruning recipe. Token importance is evaluated under deployed W4A4 operators with a quantized KV cache by combining activation magnitude, attention cues, and an explicit low-bit risk signal, enabling consistent budgeted top-$k$ selection. Experiments on standard VLM benchmarks show improved robustness over stage-wise baselines under the same low-bit regime, achieving 95.65\\% average retention while retaining only 30\\% of visual tokens, compared with about 94.3\\% retention for representative stage-wise combinations. The code will be released.", "authors": ["Xinqing Li", "Xin He", "Xindong Zhang", "Ming-Ming Cheng", "Lei Zhang", "Yun Liu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-19", "url": "https://arxiv.org/abs/2604.17320", "pdf_url": "https://arxiv.org/pdf/2604.17320v1", "arxiv_id": "2604.17320", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "2cc97ea717fd897bfe3ce60c9770134cdb2c18324a5a46980264d32b03298319", "sources": ["arxiv", "semantic_scholar"], "title": "Predictive Multi-Tier Memory Management for KV Cache in Large-Scale GPU Inference", "abstract": "Key-value (KV) cache memory management is the primary bottleneck limiting throughput and cost-efficiency in large-scale GPU inference serving. Current systems suffer from three compounding inefficiencies: (1) the absence of unified KV cache sizing across all attention architectures--particularly multi-head latent attention (MLA), which is unsupported in general-purpose frameworks, resulting in up to 57x memory over-provisioning; (2) confinement of KV cache to a single memory tier (GPU HBM) despite the availability of a rich hierarchy spanning CPU DRAM, CXL-attached memory, NVMe via GPUDirect Storage, RDMA fabric, and parallel filesystems; and (3) reactive eviction policies that discard reusable state, forcing redundant recomputation. We present a unified system that addresses all three problems. Our architecture-variant-aware sizing engine computes exact memory requirements per attention type, enabling up to 7.4x higher batch sizes. A six-tier memory hierarchy extends effective KV cache capacity from 40 GB to over 38 TB per node while maintaining sub-millisecond time-to-first-token (TTFT) for hot entries. A Bayesian reuse predictor with Beta conjugate priors over 16 (block-type, transition-type) pairs achieves 70-84% cache hit rates, combined with EMA-scored head-granular eviction and RoPE-aware prefetching. Component-level validation on trace replay using ShareGPT, LMSYS-Chat-1M, and agentic workloads demonstrates 70-84% cache hit rates. Analytical projections combining validated component behavior with published hardware specifications indicate 1.4-2.1x projected TTFT reduction, 1.7-2.9x throughput improvement, and 47% cost reduction compared to state-of-the-art baselines.", "authors": ["Sanjeev Rao Ganjihal"], "categories": ["cs.AR", "cs.AI", "cs.DC", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-19", "url": "https://arxiv.org/abs/2604.26968", "pdf_url": "https://arxiv.org/pdf/2604.26968v1", "arxiv_id": "2604.26968", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "3d1f5d001658da4a52772636eb74e595d11eddff1c7b5d585dc5ca3684b2f43b", "sources": ["arxiv", "semantic_scholar"], "title": "Bit-Flip Vulnerability of Shared KV-Cache Blocks in LLM Serving Systems", "abstract": "Rowhammer on GPU DRAM has enabled adversarial bit flips in model weights; shared KV-cache blocks in LLM serving systems present an analogous but previously unexamined target. In vLLM's Prefix Caching, these blocks exist as a single physical copy without integrity protection. Using software fault injection under ideal bit targeting, we characterize worst-case severity and identify three properties: (1) Silent divergence - 13 of 16 BF16 bit positions produce coherent but altered outputs, indistinguishable from legitimate responses without a clean baseline. (2) Selective propagation - only requests sharing the targeted prefix are affected. (3) Persistent accumulation - no temporal decay occurs, so cumulative damage grows linearly with subsequent requests. Together, these constitute a threat profile distinct from weight corruption: silent divergence and selective propagation enable detection evasion; persistent accumulation then proceeds unchecked, yielding damage amplification bounded only by how long the block remains cached. A checksum-based countermeasure detects any single-bit corruption at scheduling time, bounding cumulative damage to one batch independent of the block's cache lifetime, with negligible overhead. These results argue for integrity protection of prefix blocks before end-to-end exploitation is demonstrated.", "authors": ["Yuji Yamamoto", "Satoshi Matsuura"], "categories": ["cs.CR", "cs.AR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-19", "url": "https://arxiv.org/abs/2604.17249", "pdf_url": "https://arxiv.org/pdf/2604.17249v2", "arxiv_id": "2604.17249", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "859db27e02a0753772942b853209312b85e46debae8cbaa6f9b1f48bb63c93e4", "sources": ["arxiv", "semantic_scholar"], "title": "MoVE: Translating Laughter and Tears via Mixture of Vocalization Experts in Speech-to-Speech Translation", "abstract": "Recent Speech-to-Speech Translation (S2ST) systems achieve strong semantic accuracy yet consistently strip away non-verbal vocalizations (NVs), such as laughter and crying that convey pragmatic intent, which severely limits real-world utility. We address this via three contributions. First, we propose a synthesis pipeline for building scalable expressive datasets to overcome the data scarcity limitation. Second, we propose MoVE, a Mixture-of-LoRA-Experts architecture with expressive-specialized adapters and a soft-weighting router that blends experts for capturing hybrid expressive states. Third, we show pretrained AudioLLMs enable striking data efficiency: 30 minutes of curated data is enough for strong performance. On English-Chinese S2ST, while comparing with strong baselines, MoVE reproduces target NVs in 76% of cases and achieves the highest human-rated naturalness and emotional fidelity among all compared systems, where existing S2ST systems preserve at most 14% of NVs.", "authors": ["Szu-Chi Chen", "I-Ning Tsai", "Yi-Cheng Lin", "Sung-Feng Huang", "Hung-yi Lee"], "categories": ["cs.CL", "cs.AI", "cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-04-19", "url": "https://arxiv.org/abs/2604.17435", "pdf_url": "https://arxiv.org/pdf/2604.17435v1", "arxiv_id": "2604.17435", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a0ea47cf6f200cf6028fc80134c311b7f04bf507c4c0c0b909d205a05470b913", "sources": ["arxiv", "semantic_scholar"], "title": "HieraSparse: Hierarchical Semi-Structured Sparse KV Attention", "abstract": "The deployment of long-context Large Language Models (LLMs) poses significant challenges due to the intense computational cost of self-attention and the substantial memory overhead of the Key-Value Cache (KV Cache). In this paper, we introduce HieraSparse, a hierarchical KV Cache compression framework with acceleration kernels that leverage GPU sparse tensor cores to speed up semi-structured KV Cache attention for both the prefill and decode phases. With the hierarchical design, our method allows for a flexible quality-sparsity trade-off and successfully converts sparsity into efficiency. Compared to the state-of-the-art decode method that utilizes unstructured sparsity, HieraSparse achieves $\\mathbf{1.2\\times}$ KV compression ratio and $\\mathbf{4.57\\times}$ attention speedup at the same sparsity level. Furthermore, we extended the semi-structured KV Cache pruning to the prefill stage, which demonstrated up to $\\mathbf{1.85\\times}$ attention speedup at the highest sparsity. Lastly, we evaluate the generation quality of HieraSparse with a simple magnitude-based pruning method, and the results show that $\\mathbf{1.37\\times}$ prefill speedup and $\\mathbf{1.77\\times}$ decode speedup can be achieved without significant quality drop. The codebase can be found at https://github.com/psl-ntu/HieraSparse.", "authors": ["Haoxuan Wang", "Chen Wang"], "categories": ["cs.DC", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-18", "url": "https://arxiv.org/abs/2604.16864", "pdf_url": "https://arxiv.org/pdf/2604.16864v1", "arxiv_id": "2604.16864", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/psl-ntu/HieraSparse", "venue": null, "quality_score": 0.65}
{"id": "202cadd1a6aebb28f9d7d67988c10bb22a2d8aae4d312d8bdf452787c892ebd2", "sources": ["arxiv", "semantic_scholar"], "title": "Graph-Guided Adaptive Channel Elimination for KV Cache Compression", "abstract": "Large Language Models have revolutionized natural language processing, achieving unprecedented success across a vast range of tasks. However, their practical application in long-context scenarios is severely hampered by the formidable memory footprint of the Key-Value cache. While channel pruning has emerged as a promising compression strategy, existing methods evaluate channel importance in isolation, fundamentally ignoring the inter-channel interactions that collectively dictate model performance. This oversight leads to suboptimal pruning decisions. To address this, we introduce \\textbf{GRACE} (\\textbf{GR}aph-guided \\textbf{A}daptive \\textbf{C}hannel \\textbf{E}limination), a novel framework that reframes KV cache compression as a graph-based optimization problem. GRACE models channels as nodes and their interactions as weighted edges, enabling the identification of a near-optimal channel subset for pruning by minimizing the reconstruction error of the attention weight matrix. Furthermore, GRACE incorporates an adaptive protection mechanism that shields salient key channels from removal, ensuring a robust autoregressive decoding process. Extensive experiments show that GRACE can reduce KV cache size by 60\\% with negligible performance degradation, consistently outperforming the state-of-the-art method.", "authors": ["Enwei Tong", "Yao Zhu", "Yuanchao Bai", "Kai Wang", "Xianming Liu", "Xiangyang Ji"], "categories": ["eess.SP"], "fields_of_study": ["Engineering"], "published_date": "2026-04-18", "url": "https://arxiv.org/abs/2604.16983", "pdf_url": "https://arxiv.org/pdf/2604.16983v1", "arxiv_id": "2604.16983", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e5695158bda68c82d5fcd87f732cca08949abdfb6cbab718ec443e10592475e0", "sources": ["arxiv", "semantic_scholar"], "title": "The Illusion of Equivalence: Systematic FP16 Divergence in KV-Cached Autoregressive Inference", "abstract": "KV caching is a ubiquitous optimization in autoregressive transformer inference, long presumed to be numerically equivalent to cache-free computation. This assumption fails under standard FP16 precision: cache-ON and cache-OFF execution paths employ different floating-point accumulation orderings which, due to FP16 non-associativity, produce a deterministic divergence in decoded token sequences. Across three open-weight models (LLaMA-2-7B, Mistral-7B-v0.3, Gemma-2-2B) evaluated on GSM8K, we observe a 100\\% token divergence rate across all sampling strategies, including greedy decoding, which rules out sampling randomness as a cause, and also with cache-ON yielding higher accuracy in 8 of 9 conditions, where the accuracy difference serves as an indicator that the divergence direction is systematic rather than random. Controlled FP32 falsification reduces divergence by eight orders of magnitude, eliminates token flips, and drops the flip rate to exactly 0.0\\%, confirming FP16 non-associativity as the sole causal driver. Layer-wise drift profiling reveals architecturally predictable propagation patterns: models using Grouped-Query Attention exhibit sharp divergence at the first layer, while Gemma's larger head dimension and sliding window attention produce uniform accumulation across all layers. Finally, activation patching of the entire residual stream fails to recover the cache-free trajectory, localizing the causal variable to the stateful KV cache. These findings establish that FP16 KV cache inference is fundamentally non-equivalent to recomputation and provide a mechanistic framework for understanding numerical instability in modern LLM inference systems.", "authors": ["Ranjith Chodavarapu", "Lei Xu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-16", "url": "https://arxiv.org/abs/2604.15409", "pdf_url": "https://arxiv.org/pdf/2604.15409v1", "arxiv_id": "2604.15409", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3493}
{"id": "a77969075b58e608a7413c17a31718e8aa89e32d39277b2f865d4061d97bfa06", "sources": ["arxiv", "semantic_scholar"], "title": "Dispatch-Aware Ragged Attention for Pruned Vision Transformers", "abstract": "Token pruning methods for Vision Transformers (ViTs) promise quadratic reductions in attention FLOPs by dropping uninformative patches. Yet standard variable-length attention APIs -- including FlashAttention-2's varlen and PyTorch's NestedTensor SDPA -- fail to translate these savings into proportional wall-clock gains at the short post-pruning sequence lengths typical of ViTs ($\\leq$197 tokens). We identify a dispatch-overhead bottleneck: at these lengths, host-side kernel dispatch consumes ${\\sim}$50\\,$μ$s regardless of workload, exceeding the actual GPU compute time at moderate-to-high pruning rates. We present a lightweight bidirectional Triton attention kernel whose dispatch floor is ${\\sim}$24\\,$μ$s -- roughly 2.17$\\times$ lower than FlashAttention-2 varlen -- allowing pruning savings to become visible in wall-clock time. Integrated into a complete pack-attend-unpack pipeline and evaluated on an NVIDIA RTX 4000 Ada Generation GPU, our system achieves 1.88$\\times$ end-to-end throughput over padded PyTorch SDPA at standard 224$\\times$224 inputs, scaling to 2.51$\\times$ at 384$\\times$384. Against FlashAttention-2 varlen -- the strongest baseline -- our kernel delivers 9-12\\% higher throughput at serving batch sizes (BS=1-4), and 2.17$\\times$ lower kernel latency at 80\\% token pruning. Numerical correctness is verified with max absolute logit difference $<$0.004 and bit-exact top-1 predictions.", "authors": ["Seifeldin Abdellatif", "Ahmad Almasri"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-16", "url": "https://arxiv.org/abs/2604.15408", "pdf_url": "https://arxiv.org/pdf/2604.15408v2", "arxiv_id": "2604.15408", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3493}
{"id": "c12ba085126bb19f4fab5642a6d2bede59b56436b172294c944991f85470b666", "sources": ["arxiv", "semantic_scholar"], "title": "Geometric Routing Enables Causal Expert Control in Mixture of Experts", "abstract": "Sparse Mixture-of-Experts (MoE) models scale parameters while fixing active computation per token, but the specialization of individual experts remains opaque. In a companion paper we showed that routing topology is quality-neutral: five structurally different configurations converge to statistically equivalent language modeling quality. Here we show that expert identity is nonetheless causally meaningful: individual rank-1 experts are monosemantic by construction, and cosine-similarity routing in a low-dimensional metric space makes their specialization directly inspectable. We present four lines of evidence. First, projecting expert output vectors through the unembedding matrix yields a Semantic Dictionary: 15% of experts are monosemantic specialists spanning 10 categories (temporal, geographic, cardinal, discourse, emotional, financial, military, scientific). Second, routing exhibits a frequency-to-syntax gradient: early layers separate tokens by word frequency, deeper layers by syntactic class (Zipf-confound controls, all $p < 0.001$). Third, causal interventions confirm these labels: steering toward a temporal expert's centroid increases P(temporal) by +321% (median across 44 prompts); suppressing a geographic expert drops P(geographic) by -23%; rewriting an expert's output vector halves target-category probability, and effects compose additively across layers. Fourth, the interventions are not unique to cosine routing: linear routers support comparable steering, but only cosine routing provides geometric transparency -- expert specialization is readable directly from the centroid matrix. MoE expert-level specialization is a first-class interpretability primitive: architecturally monosemantic, causally validated, and controllable at inference with zero overhead.", "authors": ["Ivan Ternovtsii", "Yurii Bilak"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-15", "url": "https://arxiv.org/abs/2604.14434", "pdf_url": "https://arxiv.org/pdf/2604.14434v1", "arxiv_id": "2604.14434", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3485}
{"id": "188d979b73119021af4835fed38b86e7ea88d7c79ad04b41645cee6c7771f55d", "sources": ["arxiv", "semantic_scholar"], "title": "ART: Attention Run-time Termination for Efficient Large Language Model Decoding", "abstract": "Long-context decoding in Large Language Models (LLMs) is constrained by the cost of accessing and processing the Key-Value (KV) cache. Despite the evidence that attention outputs depend jointly on keys and values, most existing KV management methods rely on key-only pruning, as incorporating values incurs prohibitive additional overhead. In this paper, we propose Attention Run-time Termination (ART), a lightweight run-time mechanism that tracks accumulated attention outputs during kernel execution and terminates subsequent KV block accesses once further contributions become negligible. Rather than replacing KV selection, ART dynamically terminates redundant KV traversal on top of existing dense or sparse attention policies. We introduce a stability-based criterion that monitors both magnitude and directional changes of intermediate attention outputs, and provide a theoretical characterization of the resulting truncation error. Experiments on LongBench and RULER Needle-in-a-Haystack tasks show that ART increases the generation throughput of existing KV-cache methods by up to 20%, without compromising the quality of the results.", "authors": ["Chen Qiu", "Guozhong Li", "Cristian McGee", "Aritra Dutta", "Panos Kalnis"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-15", "url": "https://arxiv.org/abs/2606.00024", "pdf_url": "https://arxiv.org/pdf/2606.00024v2", "arxiv_id": "2606.00024", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3485}
{"id": "e8074ce7501acb1df0cdf17d52da8641dcfe2259cfef9e8edae5b0cc682aff93", "sources": ["arxiv", "semantic_scholar"], "title": "KV Packet: Recomputation-Free Context-Independent KV Caching for LLMs", "abstract": "Large Language Models (LLMs) rely heavily on Key-Value (KV) caching to minimize inference latency. However, standard KV caches are context-dependent: reusing a cached document in a new context requires recomputing KV states to account for shifts in attention distribution. Existing solutions such as CacheBlend, EPIC, and SAM-KV mitigate this issue by selectively recomputing a subset of tokens; however, they still incur non-negligible computational overhead (FLOPs) and increased Time-to-First-Token (TTFT) latency. In this paper, we propose KV Packet, a recomputation-free cache reuse framework that treats cached documents as immutable ``packets'' wrapped in light-weight trainable soft-token adapters, which are trained via self-supervised distillation to bridge context discontinuities. Experiments on Llama-3.1 and Qwen2.5 demonstrate that the proposed KV Packet method achieves near-zero FLOPs and lower TTFT than recomputation-based baselines, while retaining F1 scores comparable to those of the full recomputation baseline.", "authors": ["Chuangtao Chen", "Grace Li Zhang", "Xunzhao Yin", "Cheng Zhuo", "Bing Li", "Ulf Schlichtmann"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-14", "url": "https://arxiv.org/abs/2604.13226", "pdf_url": "https://arxiv.org/pdf/2604.13226v2", "arxiv_id": "2604.13226", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3478}
{"id": "32da91b91c77332e072107884e57c07d2e5adf9e50641274ad5aab9d3a9c68a1", "sources": ["arxiv", "semantic_scholar"], "title": "RetentiveKV: State-Space Memory for Uncertainty-Aware Multimodal KV Cache Eviction", "abstract": "Multimodal Large Language Models face severe challenges in computational efficiency and memory consumption due to the substantial expansion of the visual KV cache when processing long visual contexts. Existing KV cache compression methods typically rely on the \"persistence of importance\" hypothesis to prune tokens. However, this approach proves fragile in multimodal settings due to two key issues: 1) Visual tokens display \"deferred importance,\" initially exhibiting low salience but becoming pivotal during later decoding, which can lead to premature eviction. 2) Discrete pruning disrupts the inherent spatial continuity of visual cues. To address these challenges, we propose RetentiveKV, an entropy-driven KV cache optimization method that reformulates KV eviction from \"discrete context truncation\" to \"continuous memory evolution\" based on State Space Models. Our method leverages information entropy to quantify the information potential of low-attention tokens and integrates tokens scheduled for eviction into a continuous state space through entropy-guided state transitions, enabling their dynamic reactivation when semantic relevance arises during subsequent decoding. Extensive experiments on multimodal benchmarks demonstrate that RetentiveKV achieves 5.0 times KV cache compression and 1.5 times decoding acceleration.", "authors": ["Sihao Liu", "YuFan Xiong", "Zhonghua Jiang", "Zhaode Wang", "chengfei lv Shengyu Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-14", "url": "https://arxiv.org/abs/2605.04075", "pdf_url": "https://arxiv.org/pdf/2605.04075v1", "arxiv_id": "2605.04075", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3478}
{"id": "250d618f444b25194b1d614c4452b160d50a97b36f468ad705875b94234b2399", "sources": ["arxiv", "semantic_scholar"], "title": "Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning", "abstract": "We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.", "authors": [" NVIDIA", " :", "Aakshita Chandiramani", "Aaron Blakeman", "Abdullahi Olaoye", "Abhibha Gupta", "Abhilash Somasamudramath", "Abhinav Khattar", "Adeola Adesoba", "Adi Renduchintala", "Adil Asif", "Aditya Agrawal", "Aditya Vavre", "Ahmad Kiswani", "Aishwarya Padmakumar", "Ajay Hotchandani", "Akanksha Shukla", "Akhiad Bercovich", "Aleksander Ficek", "Aleksandr Shaposhnikov", "Alex Gronskiy", "Alex Kondratenko", "Alex Neefus", "Alex Steiner", "Alex Yang", "Alexander Bukharin", "Alexander Young", "Ali Hatamizadeh", "Ali Taghibakhshi", "Alina Galiautdinova", "Alisa Liu", "Alok Kumar", "Ameya Sunil Mahabaleshwarkar", "Amir Klein", "Amit Zuker", "Amnon Geifman", "Anahita Bhiwandiwalla", "Ananth Subramaniam", "Andrew Tao", "Anjaney Shrivastava", "Anjulie Agrusa", "Ankur Srivastava", "Ankur Verma", "Ann Guan", "Anna Shors", "Annamalai Chockalingam", "Anubhav Mandarwal", "Aparnaa Ramani", "Arham Mehta", "Arti Jain", "Arun Venkatesan", "Asha Anoosheh", "Ashwath Aithal", "Ashwin Poojary", "Asif Ahamed", "Asit Mishra", "Asli Sabanci Demiroz", "Asma Kuriparambil Thekkumpate", "Atefeh Sohrabizadeh", "Avinash Kaur", "Ayush Dattagupta", "Barath Subramaniam Anandan", "Bardiya Sadeghi", "Barnaby Simkin", "Ben Lanir", "Benedikt Schifferer", "Benjamin Chislett", "Besmira Nushi", "Bilal Kartal", "Bill Thiede", "Bita Darvish Rouhani", "Bobby Chen", "Boris Ginsburg", "Brandon Norick", "Branislav Kisacanin", "Brian Yu", "Bryan Catanzaro", "Buvaneswari Mani", "Carlo del Mundo", "Chankyu Lee", "Chanran Kim", "Chantal Hwang", "Chao Ni", "Charles Wang", "Charlie Truong", "Cheng-Ping Hsieh", "Chenhan Yu", "Chenjie Luo", "Cherie Wang", "Chetan Mungekar", "Chintan Patel", "Chris Alexiuk", "Chris Holguin", "Chris Wing", "Christian Munley", "Christopher Parisien", "Chuck Desai", "Chunyang Sheng", "Collin Neale", "Cyril Meurillon", "Dakshi Kumar", "Dan Gil", "Dan Su", "Dane Corneil", "Daniel Afrimi", "Daniel Burkhardt Eliuth Triana", "Daniel Egert", "Daniel Fatade", "Daniel Lo", "Daniel Rohrer", "Daniel Serebrenik", "Daniil Sorokin", "Daria Gitman", "Daria Levy", "Darko Stosic", "David Edelsohn", "David Messina", "David Mosallanezhad", "David Tamok", "Deena Donia", "Deepak Narayanan", "Devin O'Kelly", "Dheeraj Peri", "Dhruv Nathawani", "Di Wu", "Dima Rekesh", "Dina Yared", "Divyanshu Kakwani", "Dmitry Konyagin Brandon Tuttle", "Dong Ahn", "Dongfu Jiang", "Dorrin Poorkay", "Douglas O'Flaherty", "Duncan Riach", "Dusan Stosic", "Dustin Van Stee", "Edgar Minasyan", "Edward Lin", "Eileen Peters Long", "Elad Segal", "Elena Lantz", "Elena Lewis", "Ellie Evans", "Elliott Ning", "Eric Chung", "Eric Harper", "Eric Pham-Hung", "Eric W. Tramel", "Erick Galinkin", "Erik Pounds", "Esti Etrog", "Evan Briones", "Evan Wu", "Evelina Bakhturina", "Evgeny Tsykunov", "Ewa Dobrowolska", "Farshad Saberi Movahed", "Farzan Memarian", "Fay Wang", "Fei Jia", "Felipe Soares", "Felipe Vieira Frujeri", "Feng Chen", "Fengguang Lin", "Ferenc Galko", "Fortuna Zhang", "Frankie Siino", "Frida Hou", "Gantavya Bhatt", "Gargi Prasad", "Geethapriya Venkataramani", "Geetika Gupta", "George Armstrong", "Gerald Shen", "Giulio Borghesi", "Gordana Neskovic", "Gorkem Batmaz", "Grace Lam", "Grace Wu", "Greg Pauloski", "Greyson Davis", "Grigor Nalbandyan", "Guoming Zhang", "Guy Farber", "Guyue Huang", "Haifeng Qian", "Haran Kumar Shiv Kumar", "Harry Kim", "Harsh Sharma", "Hayate Iso", "Hayley Ross", "Herbert Hum", "Herman Sahota", "Hexin Wang", "Himanshu Soni", "Hiren Upadhyay", "Huy Nguyen", "Iain Cunningham", "Ido Galil", "Ido Shahaf", "Igino Padovani", "Igor Gitman", "Igor Shovkun", "Ikroop Dhillon", "Ilya Loshchilov", "Ingrid Kelly", "Itamar Schen", "Itay Levy", "Ivan Moshkov", "Izik Golan", "Izzy Putterman", "Jain Tu", "Jan Baczek", "Jan Kautz", "Jane Polak Scowcroft", "Janica Rosenberg", "Jared Casper", "Jarrod Pflum", "Jason Grant", "Jason Sewall", "Jatin Mitra", "Jeffrey Glick", "Jenny Chen", "Jesse Oliver", "Jiacheng Xu", "Jiafan Zhu", "Jialin Song", "Jian Zhang", "Jiaqi Zeng", "Jie Lou", "Jill Milton", "Jim Chow", "Jimmy Zhang", "Jinhang Choi", "Jining Huang", "Jocelyn Huang", "Joel Caruso", "Joey Conway", "Joey Guman", "Johan Jatko", "John Kamalu", "Johnny Greco", "Jonathan Cohen", "Jonathan Raiman", "Joseph Jennings", "Joyjit Daw", "Juan Yu", "Julio Tapia", "Junkeun Yi", "Jupinder Parmar", "Jyothi Achar", "Kari Briski", "Kartik Mattoo", "Katherine Cheung", "Katherine Luna", "Keith Wyss", "Kevin Shih", "Kezhi Kong", "Khanh Nguyen", "Khushi Bhardwaj", "Kirill Buryak", "Kirthi Shankar Sivamani", "Konstantinos Krommydas", "Kris Murphy", "Krishna C. Puvvada", "Krzysztof Pawelec", "Kumar Anik", "Laikh Tewari", "Laya Sleiman", "Leo Du", "Leon Derczynski", "Li Ding", "Lilach Ilan", "Lingjie Wu", "Lizzie Wei", "Luis Vega", "Lun Su", "Maarten Van Segbroeck", "Maer Rodrigues de Melo", "Magaret Zhang", "Mahan Fathi", "Makesh Narsimhan Sreedhar", "Makesh Sreedhar", "Makesh Tarun Chandran", "Manuel Reyes Gomez", "Maor Ashkenazi", "Marc Cuevas", "Marc Romeijn", "Margaret Zhang", "Mark Cai", "Mark Gabel", "Markus Kliegl", "Martyna Patelka", "Maryam Moosaei", "Matthew Varacalli", "Matvei Novikov", "Mauricio Ferrato", "Mehrzad Samadi", "Melissa Corpuz", "Meng Xin", "Mengdi Wang", "Mengru Wang", "Meredith Price", "Micah Schaffer", "Michael Andersch", "Michael Boone", "Michael Evans", "Michael Z Wang", "Miguel Martinez", "Mikail Khona", "Mike Chrzanowski", "Mike Hollinger", "Mingyuan Ma", "Minseok Lee", "Mohammad Dabbah", "Mohammad Shoeybi", "Mostofa Patwary", "Nabin Mulepati", "Nader Khalil", "Najeeb Nabwani", "Nancy Agarwal", "Nanthini Balasubramaniam", "Narimane Hennouni", "Narsi Kodukula", "Natalie Hereth", "Nathaniel Pinckney", "Nave Assaf", "Negar Habibi", "Nestor Qin", "Neta Zmora", "Netanel Haber", "Nick Reamaroon", "Nickson Quak", "Nidhi Bhatia", "Nikhil Jukar", "Nikki Pope", "Nikolai Ludwig", "Nima Tajbakhsh", "Nir Ailon", "Nirmal Juluru", "Nirmalya De", "Nowel Pitt", "Oleg Rybakov", "Oleksii Hrinchuk", "Oleksii Kuchaiev", "Olivier Delalleau", "Oluwatobi Olabiyi", "Omer Ullman Argov", "Omri Almog", "Omri Puny", "Oren Tropp", "Otavio Padovani", "Ouye Xie", "Parth Chadha", "Pasha Shamis", "Paul Gibbons", "Pavlo Molchanov", "Peter Belcak", "Peter Jin", "Pinky Xu", "Piotr Januszewski", "Pooya Jannaty", "Prachi Shevate", "Pradeep Thalasta", "Pranav Prashant Thombre", "Prasoon Varshney", "Prerana Gambhir", "Pritam Gundecha", "Przemek Tredak", "Qing Miao", "Qiyu Wan", "Quan Tran Minh", "Rabeeh Karimi Mahabadi", "Rachel Oberman", "Rachit Garg", "Rahul Kandu", "Raina Zhong", "Ran El-Yaniv", "Ran Zilberstein", "Rasoul Shafipour", "Renee Yao", "Renjie Pi", "Richard Mazzarese", "Richard Wang", "Rick Izzo", "Ridhima Singla", "Rima Shahbazyan", "Rishabh Garg", "Ritika Borkar", "Ritu Gala", "Riyad Islam", "Robert Clark", "Robert Hesse", "Roger Waleffe", "Rohit Varma Kalidindi", "Rohit Watve", "Roi Koren", "Ron Fan", "Ruchika Kharwar", "Ruisi Cai", "Ruoxi Zhang", "Russell J. Hewett", "Ryan Prenger", "Ryan Timbrook", "Ryota Egashira", "Sadegh Mahdavi", "Sagar Singh Ashutosh Joshi", "Sahil Modi", "Samuel Kriman", "Sandeep Pombra", "Sanjay Kariyappa", "Sanjeev Satheesh", "Santiago Pombo", "Saori Kaji", "Satish Pasumarthi", "Saurav Mishra", "Saurav Muralidharan", "Scott Hara", "Sean Narenthiran", "Sebastian Rogawski", "Seonjin Na", "Seonmyeong Bak", "Sepehr Sameni", "Seth Poulos", "Shahar Mor", "Shantanu Acharya", "Shaona Ghosh Adam Lord", "Sharath Turuvekere Sreenivas", "Shaun Kotek", "Shaya Gharghabi", "Shelby Thomas", "Sheng-Chieh Lin", "Shibani Likhite", "Shiqing Fan", "Shiyang Chen", "Shreya Gopal", "Shrimai Prabhumoye", "Shubham Pachori", "Shubham Toshniwal", "Shuo Zhang", "Shuoyang Ding", "Shyam Renjith", "Shyamala Prayaga", "Siddhartha Jain", "Simeng Sun", "Sirisha Rella", "Sirshak Das", "Smita Ithape", "Sneha Harishchandra S", "Somshubra Majumdar", "Soumye Singhal", "Sri Harsha Singudasu", "Sriharsha Niverty", "Stas Sergienko", "Stefana Gloginic", "Stefania Alborghetti", "Stephen Ge", "Stephen McCullough", "Sugam Dipak Devare", "Suguna Varshini Velury", "Sukrit Rao", "Sumeet Kumar Barua", "Sunny Gai", "Suseella Panguluri", "Sushil Koundinyan", "Swathi Patnam", "Sweta Priyadarshi", "Swetha Bhendigeri", "Syeda Nahida Akter", "Sylendran Arunagiri", "Tailling Yuan", "Talor Abramovich", "Tan Bui", "Tan Yu", "Terry Kong", "Thanh Do", "Thomas Gburek", "Thorgane Marques", "Tiffany Moore", "Tijmen Blankevoort", "Tim Moon", "Timothy Ma", "Tiyasa Mitra", "Tomasz Grzegorzek", "Tomer Asida", "Tomer Bar Natan", "Tomer Keren", "Tomer Ronen", "Traian Rebedea", "Trenton Starkey", "Tugrul Konuk", "Twinkle Vashishth", "Tyler Condensa", "Udi Karpas", "Ushnish De", "Vahid Noorozi", "Vahid Noroozi", "Vanshil Atul Shah", "Veena Vaidyanathan", "Venkat Srinivasan", "Venmugil Elango", "Victor Cui", "Vijay Korthikanti", "Vikas Mehta", "Virginia Adams", "Virginia Wu", "Vitaly Kurin", "Vitaly Lavrukhin", "Vladimir Anisimov", "Wan Seo", "Wanli Jiang", "Wasi Uddin Ahmad", "Wei Du", "Wei Ping", "Wei-Ming Chen", "Wendy Quan", "Wenliang Dai", "Wenwen Gao", "Will Jennings", "William Zhang", "Xiaowei Ren", "Xiaowen Xin", "Xin Li", "Yang Yu", "Yangyi Chen", "Yaniv Galron", "Yashaswi Karnati", "Yejin Choi", "Yev Meyer", "Yi-Fu Wu", "Yian Zhang", "Ying Lin", "Yonatan Geifman", "Yonggan Fu", "Yoshi Suhara", "Youngeun Kwon", "Yuan Zhang", "Yuki Huang", "Zach Moshe", "Zhilin Wang", "Zhiyu Cheng", "Zhongbo Zhu", "Zhuolin Yang", "Zihan Liu", "Zijia Chen", "Zijie Yan", "Zuhair Ahmed"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-14", "url": "https://arxiv.org/abs/2604.12374", "pdf_url": "https://arxiv.org/pdf/2604.12374v1", "arxiv_id": "2604.12374", "doi": null, "citation_count": 9, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.6459}
{"id": "9ea6c3415ec79ba1f9a593eb79f8408b74ca0fac1d802159bc8ed7d248305991", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Handwriting-Based Alzheimer,s Disease Diagnosis Using a Low-Rank Mixture of Experts Deep Learning Framework", "abstract": "Early and reliable detection of Alzheimer's disease (AD) is crucial for timely clinical intervention and improved patient management. It also supports the evaluation of emerging therapeutic strategies. In this paper, we propose a Low-Rank Mixture of Experts (LoRA-MoE) deep learning framework for Alzheimer's disease diagnosis based on handwriting analysis. Handwriting signals provide a non-invasive and scalable digital biomarker that captures subtle cognitive-motor impairments associated with early AD progression. The proposed architecture allows multiple experts to specialize in different handwriting patterns while sharing a common base network. This design enables efficient learning of general representations while reducing interference between experts. Each expert is equipped with lightweight low-rank adapters. This mechanism significantly reduces the number of trainable parameters compared with standard Mixture of Experts (MoE) models and improves training stability. The proposed framework is evaluated on the Diagnosis AlzheimeR WIth haNdwriting (DARWIN) dataset. Extensive experiments are conducted, including ablation studies on key architectural parameters such as hidden dimension size, number of experts, and LoRA rank. The method is compared with multilayer perceptron (MLP) and conventional MoE architectures. In addition, stacking ensemble strategies (StackMean and StackMax) are investigated to improve robustness and predictive performance. Experimental results show that the LoRA-MoE framework achieves powerful diagnostic performance while activating significantly fewer parameters during inference. These results highlight the potential of the proposed approach as an accurate and computationally efficient solution for handwriting-based Alzheimer's disease screening and digital health applications.", "authors": ["Wu Wang", "Yuang Cheng", "Fouzi Harrou", "Ying Sun"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-14", "url": "https://arxiv.org/abs/2605.04079", "pdf_url": "https://arxiv.org/pdf/2605.04079v1", "arxiv_id": "2605.04079", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3478}
{"id": "20733de7296219368726d254373eb2d05e325e8c9d51c44d639a0de467111885", "sources": ["arxiv", "semantic_scholar"], "title": "VFA: Relieving Vector Operations in Flash Attention with Global Maximum Pre-computation", "abstract": "FlashAttention-style online softmax enables exact attention computation with linear memory by streaming score tiles through on-chip memory and maintaining a running maximum and normalizer. However, as attention kernels approach peak tensor-core/cube-core throughput on modern accelerators, non-matmul components of online softmax -- especially per-tile rowmax and rowsum reductions and rescale chains -- can become vector or SIMD limited and dominate latency. This paper revisits FlashAttention and proposes Vector Relieved Flash Attention (VFA), a hardware-friendly method that reduces rowmax-driven updates of the running maximum while retaining the online-softmax structure. VFA initializes the running maximum via a cheap approximation from key-block representations, reorders key-block traversal to prioritize high-impact sink and local blocks, and freezes the maximum for remaining blocks to avoid repeated reductions and rescaling. We further integrate VFA with block-sparse skipping methods such as BLASST to form Vector Relieved Sparse Attention (VSA), which reduces both block count and per-block overhead. Notably, VFA and VSA completely avoid the conditional rescale operation in the update stage used in FA4.0. Extensive evaluations on benchmarks including MMLU and MATH500, together with attention statistics, verify our design: (i) sink and local reordering stabilizes the running maximum early; (ii) simple Q and K block summaries fail due to intra-block heterogeneity; (iii) m-initialization is required when maxima appear in middle blocks. Overall, VFA and VSA efficiently alleviate online-softmax reduction bottlenecks without performance loss. Compared to the C16V32 baseline, C8V32, C4V32 and C4V16 achieve nearly two times speedup on modern hardware while hitting the vector bottleneck. With upcoming architecture improvements, C4V16 will deliver six times speedup by enhancing exponent capacity.", "authors": ["Yupeng Sun", "Yanzhao Li", "Zhiqiang Zou", "Bai Du", "Zhiyuan Zhang", "Hui Dong", "Gaoyige Fan", "Hui Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-14", "url": "https://arxiv.org/abs/2604.12798", "pdf_url": "https://arxiv.org/pdf/2604.12798v1", "arxiv_id": "2604.12798", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3478}
{"id": "db441711f721eaf21d7ec517f2356278827128ba7c329def52275f52ac36afa8", "sources": ["arxiv", "semantic_scholar"], "title": "Quantization Dominates Rank Reduction for KV-Cache Compression", "abstract": "We compare two strategies for compressing the KV cache in transformer inference: rank reduction (discard dimensions) and quantization (keep all dimensions, reduce precision). At matched storage budgets across five models (124M-14B, MHA and GQA), we find that quantization consistently outperforms rank reduction by 4-364 PPL depending on model and compression level. The gap persists even when rank reduction is combined with quantization in hybrid baselines, and it grows with GQA aggressiveness. On LAMBADA, INT4 matches FP16 accuracy (+0.23 PPL on Mistral 7B, +0.58 on GPT-2) while rank-32 at identical storage collapses to 0.4%. We trace this gap to a structural asymmetry: under softmax attention routing, removing a dimension can flip which token is attended (a discrete failure), while quantization noise is bounded and typically preserves score ordering. We formalize this via a perturbation result showing projection damage exceeds quantization damage by 3 x 2^(2b) per direction under the softmax Fisher metric. A basis ablation confirms the finding is basis-independent (spread <0.4 PPL), establishing that the advantage comes from preserving dimensions, not from a better coordinate system. Joint K+V INT4 quantization achieves 75% total KV reduction at only +0.18 PPL on Mistral 7B.", "authors": ["Samuel Salfati"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-13", "url": "https://arxiv.org/abs/2604.11501", "pdf_url": "https://arxiv.org/pdf/2604.11501v1", "arxiv_id": "2604.11501", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3471}
{"id": "328931df37dae3043ade81bda07eefbc99e26b02f26672b633e769dfcf46ce90", "sources": ["arxiv", "semantic_scholar"], "title": "Transactional Attention: Semantic Sponsorship for KV-Cache Retention", "abstract": "At K=16 tokens (0.4% of a 4K context), every existing KV-cache compression method achieves 0% on credential retrieval. The failure mode is dormant tokens: credentials, API keys, and configuration values that receive near-zero attention but become essential at generation time. Because these tokens lack the statistical signals that eviction policies rely on, no method based on attention scores, reconstruction loss, or learned retention gates retains them. We introduce Transactional Attention (TA), a sponsorship mechanism in which structural anchor patterns (e.g., \"key:\", \"password:\") protect adjacent value-bearing tokens from eviction. TA achieves 100% credential retrieval at K=16 where six baselines (H2O, TOVA, SnapKV, StreamingLLM, PyramidKV, DynamicKV) achieve 0%, and sustains 100% accuracy across 200 function-calling trials. TA-Fast, an attention-free variant, reduces memory overhead by 52% and is compatible with SDPA and FlashAttention. TA is orthogonal to existing compression methods and adds less than 1% latency overhead.", "authors": ["Abhinaba Basu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-13", "url": "https://arxiv.org/abs/2604.11288", "pdf_url": "https://arxiv.org/pdf/2604.11288v1", "arxiv_id": "2604.11288", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3471}
{"id": "7f2ad71fb3e250b2d09ea81831e0360fb49920b09e9c8976bdc9feefcf20fd07", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Attention Scores: SVD-Based Vision Token Pruning for Efficient Vision-Language Models", "abstract": "Vision-Language Models (VLMs) have revolutionized multi-modal learning by jointly processing visual and textual information. Yet, they face significant challenges due to the high computational and memory demands of processing long sequences of vision tokens. Many existing methods rely on local heuristics, such as attention scores or token norms. However, these criteria suffer from positional bias and information dispersion, limiting their ability to preserve essential content at high pruning ratios and leading to performance degradation on visually detailed images. To address these issues, we propose SVD-Prune, a training-free, plug-and-play token pruning method based on Singular Value Decomposition. It decomposes the vision token feature matrix and selects the top-k tokens using statistical leverage scores, ensuring only tokens contributing most to the dominant global variance are preserved. Experiments show that SVD-Prune consistently outperforms prior pruning methods under extreme vision token budgets, maintaining strong performance even with 32 and 16 vision tokens.", "authors": ["Yvon Apedo", "Martyna Poreba", "Michal Szczepanski", "Samia Bouchafa"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-13", "url": "https://arxiv.org/abs/2604.11530", "pdf_url": "https://arxiv.org/pdf/2604.11530v2", "arxiv_id": "2604.11530", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3471}
{"id": "d7f90706e9b6252fa790b9581d101e0256907f82016c713753de1448d20aac6d", "sources": ["arxiv", "semantic_scholar"], "title": "IceCache: Memory-efficient KV-cache Management for Long-Sequence LLMs", "abstract": "Key-Value (KV) cache plays a crucial role in accelerating inference in large language models (LLMs) by storing intermediate attention states and avoiding redundant computation during autoregressive generation. However, its memory footprint scales linearly with sequence length, often leading to severe memory bottlenecks on resource-constrained hardware. Prior work has explored offloading KV cache to the CPU while retaining only a subset on the GPU, but these approaches often rely on imprecise token selection and suffer performance degradation in long-generation tasks such as chain-of-thought reasoning. In this paper, we propose a novel KV cache management strategy, IceCache, which integrates semantic token clustering with PagedAttention. By organizing semantically related tokens into contiguous memory regions managed by a hierarchical, dynamically updatable data structure, our method enables more efficient token selection and better utilization of memory bandwidth during CPU-GPU transfers. Experimental results on LongBench show that, with a 256-token budget, IceCache maintains 99% of the original accuracy achieved by the full KV cache model. Moreover, compared to other offloading-based methods, IceCache attains competitive or even superior latency and accuracy while using only 25% of the KV cache token budget, demonstrating its effectiveness in long-sequence scenarios. The code is available on our project website at https://yuzhenmao.github.io/IceCache/.", "authors": ["Yuzhen Mao", "Qitong Wang", "Martin Ester", "Ke Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-12", "url": "https://arxiv.org/abs/2604.10539", "pdf_url": "https://arxiv.org/pdf/2604.10539v1", "arxiv_id": "2604.10539", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.6432}
{"id": "9dc1d3bb0c7037cb3b58fcfd79bfa815247b2d780592aaffe16030d6cc3563b4", "sources": ["arxiv", "semantic_scholar"], "title": "CodeQuant: Unified Clustering and Quantization for Enhanced Outlier Smoothing in Low-Precision Mixture-of-Experts", "abstract": "Outliers have emerged as a fundamental bottleneck in preserving accuracy for low-precision large models, particularly within Mixture-of-Experts (MoE) architectures that are increasingly central to large-scale language modeling. Under post-training quantization (PTQ), these outliers induce substantial quantization errors, leading to severe accuracy degradation. While recent rotation-based smoothing techniques alleviate the problem by redistributing outlier magnitudes, residual errors remain and continue to impede reliable low-precision deployment. In this work, we tackle this challenge by introducing \\textit{CodeQuant}, a unified quantization-and-clustering scheme that contains smoothing activation outliers via learnable rotation and absorbing weight outliers into fine-tuned cluster centroids for MoE. This design reduces the influence of extreme values by fitting them within cluster centroids, thereby lowering quantization error while maintaining expressive capacity. Coupled with a dedicated kernel design for GPU and CPU, CodeQuant achieves up to $4.15\\times$ speedup while delivering significantly higher accuracy than state-of-the-art quantization approaches across diverse MoE models. Our results highlight CodeQuant as a promising direction for efficient and accurate deployment of MoE-based large language models under low-precision constraints. Our code is available at https://github.com/SAI-Lab-NYU/CodeQuant.", "authors": ["Xiangyang Yin", "Xingyu Liu", "Tianhua Xia", "Bo Bao", "Vithursan Thangarasa", "Valavan Manohararajah", "Eric Sather", "Sai Qian Zhang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-12", "url": "https://arxiv.org/abs/2604.10496", "pdf_url": "https://arxiv.org/pdf/2604.10496v1", "arxiv_id": "2604.10496", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SAI-Lab-NYU/CodeQuant", "venue": null, "quality_score": 0.6432}
{"id": "d91605ca1bb23809af027978644388951ac497adf326fc6b78b32796ce6684b0", "sources": ["arxiv", "semantic_scholar"], "title": "LoopGuard: Breaking Self-Reinforcing Attention Loops via Dynamic KV Cache Intervention", "abstract": "Through systematic experiments on long-context generation, we observe a damaging failure mode in which decoding can collapse into persistent repetition loops. We find that this degeneration is driven by collapsed attention patterns, where a subset of heads locks onto a narrow suffix of the history, and is further stabilized by inference-time KV cache reuse. Crucially, since many existing KV cache policies rely on attention-based importance, this collapse can produce spuriously high scores for repetitive tokens, causing cache management to inadvertently amplify repetition. To study this phenomenon in a controlled and reproducible manner, we introduce LoopBench, a benchmark with explicit loop-inducing conditions and loop-oriented metrics that quantify repetition severity and generation instability beyond downstream task scores. Building on these insights, we propose LoopGuard, a lightweight, plug-in KV cache guard that detects loop onset online and disrupts the feedback cycle by pruning repetitive tail spans under a fixed cache budget. Experiments on LoopBench show that LoopGuard reduces loop incidence by over 90 percentage points, while restoring output diversity and reducing token waste.", "authors": ["Dongjie Xu", "Hao Wu", "Weijie Shi", "Yue Cui", "Yuanjun Liu", "Jiawei Li", "Haolun Ma", "An Liu", "Jia Zhu", "Jiajie Xu"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-11", "url": "https://arxiv.org/abs/2604.10044", "pdf_url": "https://arxiv.org/pdf/2604.10044v1", "arxiv_id": "2604.10044", "doi": null, "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3456}
{"id": "cf5bfcfeac6d374a606385eac18eedd3c7f9aa80bb3f602ca266d9d73a9c4ae5", "sources": ["arxiv", "semantic_scholar"], "title": "CodeComp: Structural KV Cache Compression for Agentic Coding", "abstract": "Agentic code tasks such as fault localization and patch generation require processing long codebases under tight memory constraints, where the Key-Value (KV) cache becomes the primary inference bottleneck. Existing compression methods rely exclusively on attention signals to estimate token importance, systematically discarding structurally critical tokens such as call sites, branch conditions, and assignments that are essential for code understanding. We present CodeComp, a training-free KV cache compression framework that incorporates static program analysis into LLM inference via Code Property Graph priors extracted by Joern. Across bug localization and code generation benchmarks, CodeComp consistently outperforms attention-only compression baselines under equal memory budgets, recovering the majority of full-context accuracy under aggressive KV cache compression, while matching the patch generation quality of uncompressed full-context inference and integrating seamlessly into SGLang-based agentic coding pipelines without model modification.", "authors": ["Qiujiang Chen", "Jing Xiong", "Chenyang Zhao", "Sidi Yang", "Ngai Wong"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-11", "url": "https://arxiv.org/abs/2604.10235", "pdf_url": "https://arxiv.org/pdf/2604.10235v1", "arxiv_id": "2604.10235", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3456}
{"id": "7e126645626e22d06233927d125b67fcddc74d9ff5e1c5489616c789ded2c114", "sources": ["arxiv", "semantic_scholar"], "title": "MoRI: Mixture of RL and IL Experts for Long-Horizon Manipulation Tasks", "abstract": "Reinforcement Learning (RL) and Imitation Learning (IL) are the standard frameworks for policy acquisition in manipulation. While IL offers efficient policy derivation, it suffers from compounding errors and distribution shift. Conversely, RL facilitates autonomous exploration but is frequently hindered by low sample efficiency and the high cost of trial and error. Since existing hybrid methods often struggle with complex tasks, we introduce Mixture of RL and IL Experts (MoRI). This system dynamically switches between IL and RL experts based on the variance of expert actions to handle coarse movements and fine-grained manipulations. MoRI employs an offline pre-training stage followed by online fine-tuning to accelerate convergence. To maintain exploration safety and minimize human intervention, the system applies IL-based regularization to the RL component. Evaluation across four complex real-world tasks shows that MoRI achieves an average success rate of 97.5% within 2 to 5 hours of fine-tuning. Compared to baseline RL algorithms, MoRI reduces human intervention by 85.8% and shortens convergence time by 21%, demonstrating its capability in robotic manipulation.", "authors": ["Yaohang Xu", "Lianjie Ma", "Gewei Zuo", "Wentao Zhang", "Han Ding", "Lijun Zhu"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-11", "url": "https://arxiv.org/abs/2604.10165", "pdf_url": "https://arxiv.org/pdf/2604.10165v1", "arxiv_id": "2604.10165", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3456}
{"id": "70382e8338b39bceeb4a88e95f5dc07733d26ad2fe40e1f033e2b6919a39c82f", "sources": ["arxiv", "semantic_scholar"], "title": "Sequential KV Cache Compression via Probabilistic Language Tries: Beyond the Per-Vector Shannon Limit", "abstract": "Recent work on KV cache quantization, culminating in TurboQuant, has approached the Shannon entropy limit for per-vector compression of transformer key-value caches. We observe that this limit applies to a strictly weaker problem than the one that actually matters: compressing the KV cache as a sequence. The tokens stored in a KV cache are not arbitrary floating-point data -- they are samples from the exact formal language the model was trained on, and the model is by construction a near-optimal predictor of that language. We introduce sequential KV compression, a two-layer architecture that exploits this structure. The first layer, probabilistic prefix deduplication, identifies semantically equivalent shared prefixes across sessions using the trie metric d_T(s, s') = -log_2 P_M(s ^ s') from Probabilistic Language Tries (PLTs). The second layer, predictive delta coding, stores only the residual of each new KV vector from the model's own prediction of it, achieving a per-token entropy bound of H(KV_{i+1} | KV_{<=i}) <= H(token_{i+1} | token_{<=i}). We prove that at typical language model perplexity -- approximately 10-20 for fluent English text -- this bound is 3.3-4.3 bits on average per token position, compared to TurboQuant's 3 bits per vector component (with typical attention heads having 64-128 components). The theoretical compression ratio over TurboQuant is approximately 914,000x at the Shannon limit. Even at 1000x above the entropy floor -- a deliberately pessimistic worst-case overhead, two orders of magnitude above the 2-5x typical of practical source coders -- the ratio remains approximately 914x over TurboQuant, with compression improving rather than degrading as context length grows. The two layers are orthogonal and compose with existing per-vector quantization methods including TurboQuant.", "authors": ["Gregory Magarshak"], "categories": ["cs.LG", "cs.AI", "cs.IT", "cs.NE"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-10", "url": "https://arxiv.org/abs/2604.15356", "pdf_url": "https://arxiv.org/pdf/2604.15356v1", "arxiv_id": "2604.15356", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3449}
{"id": "62f0fbad79b75697b733e3d1e5e929869892cee2399efea8e9411d62e7d0b361", "sources": ["arxiv", "semantic_scholar"], "title": "MP-ISMoE: Mixed-Precision Interactive Side Mixture-of-Experts for Efficient Transfer Learning", "abstract": "Parameter-efficient transfer learning (PETL) has emerged as a pivotal paradigm for adapting pre-trained foundation models to downstream tasks, significantly reducing trainable parameters yet suffering from substantial memory overhead caused by gradient backpropagation during fine-tuning. While memory-efficient transfer learning (METL) circumvents this challenge by bypassing backbone gradient computation via lightweight small side networks, its stringent memory constraint severely limits learning capacity of side networks, thereby significantly compromising performance. To address these limitations, we propose a novel Mixed-Precision Interactive Side Mixture-of-Experts framework (MP-ISMoE). Specifically, we first propose a Gaussian Noise Perturbed Iterative Quantization (GNP-IQ) scheme to quantize weights into lower-bits while effectively decreasing quantization errors. By leveraging memory conserved from GNP-IQ, we subsequently employ Interactive Side Mixture-of-Experts (ISMoE) to scaling up side networks without sacrificing overall memory efficiency. Different from conventional mixture-of-experts, ISMoE learns to select optimal experts by interacting with salient features from frozen backbones, thus suppressing knowledge forgetting and boosting performance. Extensive experiments across diverse vision-language and language-only tasks demonstrate that MP-ISMoE remarkably promotes accuracy compared to state-of-the-art METL approaches, while maintaining comparable parameter and memory efficiency.", "authors": ["Yutong Zhang", "Zimeng Wu", "Shangcai Liao", "Shujiang Wu", "Jiaxin Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-10", "url": "https://arxiv.org/abs/2605.04058", "pdf_url": "https://arxiv.org/pdf/2605.04058v1", "arxiv_id": "2605.04058", "doi": "10.1609/aaai.v40i34.40084", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.542}
{"id": "192e1da6ca291ee2ed760d051906dc9c663fb8b4db9c7955a5c9df1fab3bbc4f", "sources": ["arxiv", "semantic_scholar"], "title": "Modality-Aware Zero-Shot Pruning and Sparse Attention for Efficient Multimodal Edge Inference", "abstract": "Edge devices increasingly run multimodal sensing pipelines that must remain accurate despite fluctuating power budgets and unpredictable sensor dropout. Existing pruning methods fail under these conditions: they generally require fine-tuning after compression, consuming over $10\\times$ the deployment energy, and they assign static importance scores that are blind to which sensors are present. We present the SentryFuse framework, which addresses both challenges jointly through two key components. First, SentryGate learns modality-conditioned importance scores during training via first-order saliency supervision and then prunes attention heads and feed-forward channels at deployment without fine-tuning. Second, SentryAttend replaces dense self-attention, a key bottleneck in contemporary multimodal architectures, with sparse grouped-query attention, yielding a net 15% reduction in GFLOPs across three different multimodal architectures. Across three applications and multimodal backbones, SentryGate achieves a 12.7% average accuracy improvement over the strongest pruning baseline, and upto to 18% under modality dropout conditions. Together, SentryFuse reduces memory by 28.2% and lowers latency by up to $1.63\\times$ without further fine-tuning, establishing modality-aware zero-shot compression as a practical path to multimodal intelligence on heterogeneous edge hardware.", "authors": ["Yueyuan Sui", "Payal Mohapatra", "Doğaç Eldenk", "Haodong Yang", "Yiting Zhang", "Haoyan Zhang", "Qi Zhu", "Stephen Xia"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-10", "url": "https://arxiv.org/abs/2604.08971", "pdf_url": "https://arxiv.org/pdf/2604.08971v1", "arxiv_id": "2604.08971", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3449}
{"id": "9f3b30754b59503a5fb09b8f44edb9d4e81a1d520ab2e4cbe653017dbdd1c6f7", "sources": ["arxiv", "semantic_scholar"], "title": "Plasticity-Enhanced Multi-Agent Mixture of Experts for Dynamic Objective Adaptation in UAVs-Assisted Emergency Communication Networks", "abstract": "Unmanned aerial vehicles serving as aerial base stations can rapidly restore connectivity after disasters, yet abrupt changes in user mobility and traffic demands shift the quality of service trade-offs and induce strong non-stationarity. Deep reinforcement learning policies suffer from plasticity loss under such shifts, as representation collapse and neuron dormancy impair adaptation. We propose plasticity enhanced multi-agent mixture of experts (PE-MAMoE), a centralized training with decentralized execution framework built on multi-agent proximal policy optimization. PE-MAMoE equips each UAV with a sparsely gated mixture of experts actor whose router selects a single specialist per step. A non-parametric Phase Controller injects brief, expert-only stochastic perturbations after phase switches, resets the action log-standard-deviation, anneals entropy and learning rate, and schedules the router temperature, all to re-plasticize the policy without destabilizing safe behaviors. We derive a dynamic regret bound showing the tracking error scales with both environment variation and cumulative noise energy. In a phase-driven simulator with mobile users and 3GPP-style channels, PE-MAMoE improves normalized interquartile mean return by 26.3\\% over the best baseline, increases served-user capacity by 12.8\\%, and reduces collisions by approximately 75\\%. Diagnostics confirm persistently higher expert feature rank and periodic dormant-neuron recovery at regime switches.", "authors": ["Wen Qiu", "Zhiqiang He", "Wei Zhao", "Hiroshi Masui"], "categories": ["cs.MA", "cs.LG", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-10", "url": "https://arxiv.org/abs/2604.09028", "pdf_url": "https://arxiv.org/pdf/2604.09028v1", "arxiv_id": "2604.09028", "doi": "10.1109/jiot.2026.3696834", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Internet of Things Journal", "quality_score": 0.542}
{"id": "3c29bc3970a7e4bce1ef512ff5c627ee2526ee0c79a7f5d3ccea0eca8d89907f", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache Offloading for Context-Intensive Tasks", "abstract": "With the growing demand for long-context LLMs across a wide range of applications, the key-value (KV) cache has become a critical bottleneck for both latency and memory usage. Recently, KV-cache offloading has emerged as a promising approach to reduce memory footprint and inference latency while preserving accuracy. Prior evaluations have largely focused on tasks that do not require extracting large amounts of information from the context. In this work, we study KV-cache offloading on context-intensive tasks: problems where the solution requires looking up a lot of information from the input prompt. We create and release the Text2JSON benchmark, a highly context-intensive task that requires extracting structured knowledge from raw text. We evaluate modern KV offloading on Text2JSON and other context-intensive tasks and find significant performance degradation on both Llama 3 and Qwen 3 models. Our analysis identifies two key reasons for poor accuracy: low-rank projection of keys and unreliable landmarks, and proposes a simpler alternative strategy that significantly improves accuracy across multiple LLM families and benchmarks. These findings highlight the need for a comprehensive and rigorous evaluation of long-context compression techniques.", "authors": ["Andrey Bocharnikov", "Ivan Ermakov", "Denis Kuznedelev", "Vyacheslav Zhdanovskiy", "Yegor Yershov"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-09", "url": "https://arxiv.org/abs/2604.08426", "pdf_url": "https://arxiv.org/pdf/2604.08426v4", "arxiv_id": "2604.08426", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3442}
{"id": "219ab6374d4916c7e505d4a442b1d9194b034d9b45dbe0e21202df45f44a28bc", "sources": ["arxiv", "semantic_scholar"], "title": "Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts", "abstract": "Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.", "authors": ["Haolei Xu", "Haiwen Hong", "Hongxing Li", "Rui Zhou", "Yang Zhang", "Longtao Huang", "Hui Xue", "Yongliang Shen", "Weiming Lu", "Yueting Zhuang"], "categories": ["cs.CV", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-09", "url": "https://arxiv.org/abs/2604.08541", "pdf_url": "https://arxiv.org/pdf/2604.08541v1", "arxiv_id": "2604.08541", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3442}
{"id": "65aed9717acda0620036cb5f78ef6e1169172f9fdd76d78de1148e6cd5d41a8c", "sources": ["arxiv", "semantic_scholar"], "title": "Can LoRA Fusion Support Cross-Domain Tasks in Cloud-Edge Collaboration?", "abstract": "Cloud-hosted large language models (LLMs) commonly rely on LoRA for domain adaptation, yet domain data are distributed across multiple edge devices and cannot be uploaded due to privacy constraints. This raises a fundamental question: how can knowledge from multiple private edges be integrated into a cloud LLM for cross-domain problem solving? A natural solution is to train LoRA adapters locally and fuse them in the cloud; however, existing pipelines rely on unrealistic assumptions that edge devices can host cloud-scale LLMs and are evaluated mainly on single-domain tasks. To address these limitations, we propose a prune-train-recover framework that enables local LoRA training on pruned models and privacy-preserving cloud integration. We further introduce MMLU-CD, a cross-domain benchmark that composes multiple domain samples into a single instance, enabling explicit evaluation of cross-domain problem solving. This allows us to ask a concrete question: Can existing LoRA fusion methods support cross-domain tasks in cloud-edge collaboration? Our empirical answer is negative. Existing LoRA fusion methods perform poorly on MMLU-CD, often underperforming the base LLM, revealing their inability to support cross-domain problem solving. We attribute this failure to parameter conflicts among LoRA adapters and propose a simple conflict-resolution module, LoRA-CR, which mitigates conflicting updates and improves LoRA fusion performance by up to 3.8%. These results identify conflict mitigation as a critical yet largely overlooked factor in cloud-edge LoRA fusion, warranting further investigation in future research.", "authors": ["Yatong Wang", "Fali Wang", "Naibin Gu", "Zheng Lin", "Zhengxiao Liu", "Dingyu Yao", "Zhiwei Zhang", "Jianxin Shi", "Weiping Wang"], "categories": ["cs.DC", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-09", "url": "https://arxiv.org/abs/2605.23913", "pdf_url": "https://arxiv.org/pdf/2605.23913v1", "arxiv_id": "2605.23913", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3442}
{"id": "ccf6bb1000107964a5995f139d4f2cfcbaa98bb25485eb1072079be16f54bff3", "sources": ["arxiv", "semantic_scholar"], "title": "MoBiE: Efficient Inference of Mixture of Binary Experts under Post-Training Quantization", "abstract": "Mixture-of-Experts (MoE) based large language models (LLMs) offer strong performance but suffer from high memory and computation costs. Weight binarization provides extreme efficiency, yet existing binary methods designed for dense LLMs struggle with MoE-specific issues, including cross-expert redundancy, task-agnostic importance estimation, and quantization-induced routing shifts. To this end, we propose MoBiE, the first binarization framework tailored for MoE-based LLMs. MoBiE is built on three core innovations: 1. using joint SVD decomposition to reduce cross-expert redundancy; 2. integrating global loss gradients into local Hessian metrics to enhance weight importance estimation; 3. introducing an error constraint guided by the input null space to mitigate routing distortion. Notably, MoBiE achieves these optimizations while incurring no additional storage overhead, striking a balance between efficiency and model performance. Extensive experiments demonstrate that MoBiE consistently outperforms state-of-the-art binary methods across multiple MoE-based LLMs and benchmarks. For example, on Qwen3-30B-A3B, MoBiE reduces perplexity by 52.2$\\%$, improves average zero-shot performance by 43.4$\\%$, achieves over 2 $\\times$ inference speedup, and further shortens quantization time. The code is available at https://github.com/Kishon-zzx/MoBiE.", "authors": ["Zhixiong Zhao", "Zukang Xu", "Zhixuan Chen", "Dawei Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-08", "url": "https://arxiv.org/abs/2604.06798", "pdf_url": "https://arxiv.org/pdf/2604.06798v4", "arxiv_id": "2604.06798", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Kishon-zzx/MoBiE", "venue": null, "quality_score": 0.6378}
{"id": "308bfc05e454a554872a2cf602f14ef95577800ee6dcaa03e87b73c2035daeee", "sources": ["arxiv", "semantic_scholar"], "title": "AudioKV: KV Cache Eviction in Efficient Large Audio Language Models", "abstract": "Large Audio-Language Models (LALMs) have set new benchmarks in speech processing, yet their deployment is hindered by the memory footprint of the Key-Value (KV) cache during long-context inference. While general KV cache compression techniques excel in LLMs, they often fail in the audio domain by overlooking the intrinsic temporal continuity of acoustic signals. To bridge this gap, we propose AudioKV, a novel framework that robustly prioritizes audio-critical attention heads through a hardware-friendly semantic-acoustic alignment mechanism. Specifically, we identify these modality-specialized heads by analyzing attention scores in ASR tasks and dynamically allocate KV cache budgets preferentially to them. Furthermore, we introduce Spectral Score Smoothing (SSS), an FFT-based global filtering strategy designed to suppress high-frequency noise and recover smooth global trends from importance scores, ensuring more balanced token selection with unprecedented precision. Extensive evaluations across multiple LALMs, including Qwen and Gemma series, demonstrate that AudioKV significantly outperforms baselines while enhancing computational efficiency. Notably, at a 40% compression ratio, AudioKV maintains near-full accuracy on Qwen3-Omni-30B with only a 0.45% drop, whereas traditional methods suffer from catastrophic performance degradation and repetition. Our code will be released after acceptance.", "authors": ["Yuxuan Wang", "Peize He", "Xiyan Gui", "Xiaoqian Liu", "Junhao He", "Xuyang Liu", "Zichen Wen", "Xuming Hu", "Linfeng Zhang"], "categories": ["cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-08", "url": "https://arxiv.org/abs/2604.06694", "pdf_url": "https://arxiv.org/pdf/2604.06694v1", "arxiv_id": "2604.06694", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3434}
{"id": "6cbfcfa948788f6c102de75255ea891114756250d835c043a13f3614ea31947e", "sources": ["arxiv", "semantic_scholar"], "title": "Does a Global Perspective Help Prune Sparse MoEs Elegantly?", "abstract": "Empirical scaling laws for language models have encouraged the development of ever-larger LLMs, despite their growing computational and memory costs. Sparse Mixture-of-Experts (MoEs) offer a promising alternative by activating only a subset of experts per forward pass, improving efficiency without sacrificing performance. However, the large number of expert parameters still leads to substantial memory consumption. Existing pruning methods typically allocate budgets uniformly across layers, overlooking the heterogeneous redundancy that arises in sparse MoEs. We propose GRAPE (Global Redundancy-Aware Pruning of Experts, a global pruning strategy that dynamically allocates pruning budgets based on cross-layer redundancy. Experiments on Mixtral-8x7B, Mixtral-8x22B, DeepSeek-MoE, Qwen-MoE, and GPT-OSS show that, under the same pruning budget, GRAPE consistently achieves the best average performance. On the three main models reported in the paper, it improves average accuracy over the strongest local baseline by 1.40% on average across pruning settings, with gains of up to 2.45%.", "authors": ["Zeliang Zhang", "Nikhil Ghosh", "Jiani Liu", "Bin Yu", "Xiaodong Liu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-08", "url": "https://arxiv.org/abs/2604.06542", "pdf_url": "https://arxiv.org/pdf/2604.06542v1", "arxiv_id": "2604.06542", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3434}
{"id": "17a3008c95aaa5ef05bc9849d9b88037f9e3cdc5efb3c45430990406554f06fa", "sources": ["arxiv", "semantic_scholar"], "title": "Region-Graph Optimal Transport Routing for Mixture-of-Experts Whole-Slide Image Classification", "abstract": "Multiple Instance Learning (MIL) is the dominant framework for gigapixel whole-slide image (WSI) classification in computational pathology. However, current MIL aggregators route all instances through a shared pathway, constraining their capacity to specialise across the pathological heterogeneity inherent in each slide. Mixture-of-Experts (MoE) methods offer a natural remedy by partitioning instances across specialised expert subnetworks; yet unconstrained softmax routing may yield highly imbalanced utilisation, where one or a few experts absorb most routing mass, collapsing the mixture back to a near-single-pathway solution. To address these limitations, we propose ROAM (Region-graph OptimAl-transport Mixture-of-experts), a spatially aware MoE-MIL aggregator that routes region tokens to expert poolers via capacity-constrained entropic optimal transport, promoting balanced expert utilisation by construction. ROAM operates on spatial region tokens, obtained by compressing dense patch bags into spatially binned units that align routing with local tissue neighbourhoods and introduces two key mechanisms: (i) region-to-expert assignment formulated as entropic optimal transport (Sinkhorn) with explicit per slide capacity marginals, enforcing balanced expert utilisation without auxiliary load-balancing losses; and (ii) graph-regularised Sinkhorn iterations that diffuse routing assignments over the spatial region graph, encouraging neighbouring regions to coherently route to the same experts. Evaluated on four WSI benchmarks with frozen foundation-model patch embeddings, ROAM achieves performance competitive against strong MIL and MoE baselines, and on NSCLC generalisation (TCGA-CPTAC) reaches external AUC 0.845 +- 0.019.", "authors": ["Xin Tian", "Jiuliu Lu", "Ephraim Tsalik", "Bart Wanders", "Colleen Knoth", "Julian Knight"], "categories": ["cs.CV", "cs.AI", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-04-08", "url": "https://arxiv.org/abs/2604.07298", "pdf_url": "https://arxiv.org/pdf/2604.07298v1", "arxiv_id": "2604.07298", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3434}
{"id": "51229331cfd93bfc4842bf4203772d54017495689fe97703c5201ea146216d21", "sources": ["arxiv", "semantic_scholar"], "title": "ForkKV: Scaling Multi-LoRA Agent Serving via Copy-on-Write Disaggregated KV Cache", "abstract": "The serving paradigm of large language models (LLMs) is rapidly shifting towards complex multi-agent workflows where specialized agents collaborate over massive shared contexts. While Low-Rank Adaptation (LoRA) enables the efficient co-hosting of these specialized agents on a single base model, it introduces a critical memory footprint bottleneck during serving. Specifically, unique LoRA activations cause Key-Value (KV) cache divergence across agents, rendering traditional prefix caching ineffective for shared contexts. This forces redundant KV cache maintenance, rapidly saturating GPU capacity and degrading throughput. To address this challenge, we introduce ForkKV, a serving system for multi-LoRA agent workflows centered around a novel memory management paradigm in OS: fork with copy-on-write (CoW). By exploiting the structural properties of LoRA, ForkKV physically decouples the KV cache into a massive shared component (analogous to the parent process's memory pages) and lightweight agent-specific components (the child process's pages). To support this mechanism, we propose a DualRadixTree architecture that allows newly forked agents to inherit the massive shared cache and apply CoW semantics for their lightweight unique cache. Furthermore, to guarantee efficient execution, we design ResidualAttention, a specialized kernel that reconstructs the disaggregated KV cache directly within on-chip SRAM. Comprehensive evaluations across diverse language models and practical datasets of different tasks demonstrate that ForkKV achieves up to 3.0x the throughput of state-of-the-art multi-LoRA serving systems with a negligible impact on generation quality.", "authors": ["Shao Wang", "Rui Ren", "Lin Gui"], "categories": ["cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2604.06370", "pdf_url": "https://arxiv.org/pdf/2604.06370v1", "arxiv_id": "2604.06370", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "5a4211aedb271a766a363d381406bad463e14e3d760dc9966a47580e74f27e5e", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Quantization of Mixture-of-Experts with Theoretical Generalization Guarantees", "abstract": "Sparse Mixture-of-Experts (MoE) allows scaling of language and vision models efficiently by activating only a small subset of experts per input. While this reduces computation, the large number of parameters still incurs substantial memory overhead during inference. Post-training quantization has been explored to address this issue. Because uniform quantization suffers from significant accuracy loss at low bit-widths, mixed-precision methods have been recently explored; however, they often require substantial computation for bit-width allocation and overlook the varying sensitivity of model performance to the quantization of different experts. We propose a theoretically grounded expert-wise mixed precision strategy that assigns bit-width to each expert primarily based on their change in routers l2 norm during training. Experts with smaller changes are shown to capture less frequent but critical features, and model performance is more sensitive to the quantization of these experts, thus requiring higher precision. Furthermore, to avoid allocating experts to lower precision that inject high quantization noise, experts with large maximum intra-neuron variance are also allocated higher precision. Experiments on large-scale MoE models, including Switch Transformer and Mixtral, show that our method achieves higher accuracy than existing approaches, while also reducing inference cost and incurring only negligible overhead for bit-width assignment.", "authors": ["Mohammed Nowaz Rabbani Chowdhury", "Kaoutar El Maghraoui", "Hsinyu Tsai", "Naigang Wang", "Geoffrey W. Burr", "Liu Liu", "Meng Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2604.06515", "pdf_url": "https://arxiv.org/pdf/2604.06515v1", "arxiv_id": "2604.06515", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "The Fourteenth International Conference on Learning Representations, 2026", "quality_score": 0.5385}
{"id": "1e395dce6b6afe0baf0fb281b998d39c30dad6a021f50e828371daf53c8d0d9d", "sources": ["arxiv", "semantic_scholar"], "title": "HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference", "abstract": "Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key-value (KV) caches. Each visual input expands into thousands of tokens, causing caches to scale linearly with context length and remain resident in GPU memory throughout decoding, which leads to prohibitive memory overhead and latency even on high-end GPUs. A common solution is to compress caches under a fixed allocated budget at different granularities: token-level uniformly discards less important tokens, layer-level varies retention across layers, and head-level redistributes budgets across heads. Yet these approaches stop at allocation and overlook the heterogeneous behaviors of attention heads that require distinct compression strategies. We propose HybridKV, a hybrid KV cache compression framework that integrates complementary strategies in three stages: heads are first classified into static or dynamic types using text-centric attention; then a top-down budget allocation scheme hierarchically assigns KV budgets; finally, static heads are compressed by text-prior pruning and dynamic heads by chunk-wise retrieval. Experiments on 11 multimodal benchmarks with Qwen2.5-VL-7B show that HybridKV reduces KV cache memory by up to $7.9\\times$ and achieves $1.52\\times$ faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.", "authors": ["Bowen Zeng", "Feiyang Ren", "Jun Zhang", "Xiaoling Gu", "Ke Chen", "Lidan Shou", "Huan Li"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2604.05887", "pdf_url": "https://arxiv.org/pdf/2604.05887v1", "arxiv_id": "2604.05887", "doi": null, "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "37a5f2fa34e7cc188961d2c2c2725bf22dc9c89b1f2479af0e936e6f4fc943e8", "sources": ["arxiv", "semantic_scholar"], "title": "Agentic AI-Based Joint Computing and Networking via Mixture of Experts and Large Language Models", "abstract": "Future sixth-generation (6G) mobile networks are envisioned to be equipped with a diverse set of powerful, yet highly specialized, optimization experts. Such a promising vision is concurrently expected to give rise to the need for scalable mechanisms that can select, combine, and orchestrate such experts based on high-level intent and uncertainty descriptions. In this paper, we propose an agentic artificial intelligence (AI)-based network optimization framework that integrates mixture of experts (MoE) architectures with large language models (LLMs). Under the proposed framework, the employed LLM acts as a semantic gate to reason over operator objectives and dynamically compose suitable optimization agents. The proposed framework is formulated in a model-agnostic manner and bridges human-readable network intents with low-level resource allocation decisions, enabling flexible optimization across heterogeneous objectives and operating conditions. As a representative instantiation, we apply the framework to a joint communication and computing network and design a library of specialized optimization experts covering throughput, fairness, and delay-driven objectives under both regular and robust conditions. Numerical simulations demonstrate that the proposed agentic MoE framework consistently achieves near-optimal performance compared to exhaustive expert combinations while outperforming individual experts across diverse objectives, including delay minimization and throughput maximization.", "authors": ["Robert-Jeron Reifert", "Alaa Alameer Ahmad", "Hayssam Dahrouj", "Aydin Sezgin"], "categories": ["cs.LG", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2605.02911", "pdf_url": "https://arxiv.org/pdf/2605.02911v1", "arxiv_id": "2605.02911", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "6474c051b7f535373c023060984142a39c5237dd3f132abe95ce2f8051a722f1", "sources": ["arxiv", "semantic_scholar"], "title": "ALTO: Adaptive LoRA Tuning and Orchestration for Heterogeneous LoRA Training Workloads", "abstract": "Low-Rank Adaptation (LoRA) is now the dominant method for parameter-efficient fine-tuning of large language models, but achieving a high-quality adapter often requires systematic hyperparameter tuning because LoRA performance is highly sensitive to configuration choices. In practice, this leads to many concurrent LoRA jobs, often spanning heterogeneous tasks in multi-tenant environments. Existing systems largely handle these jobs independently, which both wastes computation on weak candidates and leaves GPUs underutilized. We present ALTO (Adaptive LoRA Tuning and Orchestration), a co-designed training system that accelerates LoRA hyperparameter tuning while enabling efficient cluster sharing across heterogeneous tasks. The central insight behind ALTO is that when multiple tuning jobs run concurrently over a shared frozen backbone, they expose optimization opportunities that single-job designs cannot exploit. Building on this, ALTO monitors loss trajectories to terminate unpromising configurations early, uses fused grouped GEMM together with a new rank-local adapter parallelism to co-locate surviving adapters and reclaim freed GPU capacity, and combines intra-task and inter-task scheduling to improve multi-task placement by leveraging the predictable duration of LoRA jobs. Extensive evaluation shows that ALTO achieves up to $13.8\\times$ speedup over state-of-the-art without sacrificing adapter quality.", "authors": ["Jingwei Zuo", "Xinze Feng", "Zien Liu", "Kaijian Wang", "Fanjiang Ye", "Ye Cao", "Zhuang Wang", "Yuke Wang"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-07", "url": "https://arxiv.org/abs/2604.05426", "pdf_url": "https://arxiv.org/pdf/2604.05426v2", "arxiv_id": "2604.05426", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3427}
{"id": "36fef3816df51c37bee1ad8250c6c85afc4b0d91665c2044c740b6b9140b7372", "sources": ["arxiv", "semantic_scholar"], "title": "Don't Waste Bits! Adaptive KV-Cache Quantization for Lightweight On-Device LLMs", "abstract": "Large Language Models (LLMs) have achieved remarkable progress across reasoning, generation, and decision-making tasks, yet deploying them on mobile, embedded, and edge devices remains particularly challenging. On-device LLM inference is heavily constrained by the memory and bandwidth overhead of the key-value (KV) cache, which grows linearly with context length and often dominates decoding cost. Existing KV-cache quantization schemes typically rely on fixed precision or hand-crafted heuristics, thereby wasting bits on low-impact tokens while over-compressing informative ones, leading to avoidable accuracy degradation. Inspired by Huffman coding's principle of variable-length allocation, we propose adaptive KV-cache quantization, a learned policy that assigns bit-width proportional to token importance, minimizing expected memory and latency without sacrificing competitive accuracy. Our framework extracts lightweight token-level features, including token frequency, quality score, attention variance, and entropy-based uncertainty, and feeds them into a compact data-driven controller that dynamically selects KV precision from {2-bit, 4-bit, 8-bit, FP16} during decoding. This adaptive precision policy reduces KV memory footprint and latency while improving accuracy compared to static KV quantization and rule-based baselines, and maintaining competitive accuracy close to FP16 inference across standard LLM benchmarks. Extensive experiments across multiple commonsense reasoning benchmarks using SmolLM-135M, SmolLM-360M, and SmolLM-1.7B demonstrate that our controller consistently improves the accuracy-latency trade-off. For instance, with SmolLM-360M on HellaSwag, our method reduces decoding latency (ms/token) by 17.75% relative to static KV quantization, improves accuracy by 7.60 points, and remains within only 0.30 points of FP16 inference.", "authors": ["Sayed Pedram Haeri Boroujeni", "Niloufar Mehrabi", "Patrick Woods", "Gabriel Hillesheim", "Abolfazl Razi"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-06", "url": "https://arxiv.org/abs/2604.04722", "pdf_url": "https://arxiv.org/pdf/2604.04722v1", "arxiv_id": "2604.04722", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.342}
{"id": "6bcc5b6491f28096042705d410186d8320757c3d1bf65176e49602fc85b231d5", "sources": ["arxiv", "semantic_scholar"], "title": "REAM: Merging Improves Pruning of Experts in LLMs", "abstract": "Mixture-of-Experts (MoE) large language models (LLMs) are among the top-performing architectures. The largest models, often with hundreds of billions of parameters, pose significant memory challenges for deployment. Traditional approaches to reduce memory requirements include weight pruning and quantization. Motivated by the Router-weighted Expert Activation Pruning (REAP) that prunes experts, we propose a novel method, Router-weighted Expert Activation Merging (REAM). Instead of removing experts, REAM groups them and merges their weights, better preserving original performance. We evaluate REAM against REAP and other baselines across multiple MoE LLMs on diverse multiple-choice (MC) question answering and generative (GEN) benchmarks. Our results reveal a trade-off between MC and GEN performance that depends on the mix of calibration data. By controlling the mix of general, math and coding data, we examine the Pareto frontier of this trade-off and show that REAM often outperforms the baselines and in many cases is comparable to the original uncompressed models.", "authors": ["Saurav Jha", "Maryam Hashemzadeh", "Ali Saheb Pasand", "Ali Parviz", "Min-Joong Lee", "Boris Knyazev"], "categories": ["cs.AI", "cs.CL", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-06", "url": "https://arxiv.org/abs/2604.04356", "pdf_url": "https://arxiv.org/pdf/2604.04356v1", "arxiv_id": "2604.04356", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/SamsungSAILMontreal/ream", "venue": null, "quality_score": 0.6351}
{"id": "d9dbc10430bf6f734e04ca85badd470cab1a792eba28e69786560fb49a85cc6e", "sources": ["arxiv", "semantic_scholar"], "title": "eOptShrinkQ: Near-Lossless KV Cache Compression Through Optimal Spectral Denoising and Quantization", "abstract": "We show that the key-value (KV) cache in transformer attention heads admits a natural decomposition into a low-rank \\emph{shared context} component and a full-rank \\emph{per-token} residual, well described by the spiked random matrix model. This observation leads to eOptShrinkQ, a two-stage compression pipeline: optimal singular value shrinkage (eOptShrink) automatically extracts the shared structure, and the residual -- which satisfies the \\emph{thin shell property} with delocalized coordinates -- is quantized by TurboQuant~\\citep{zandieh2025turboquant}, a recently proposed per-vector scalar quantizer with near-optimal distortion guarantees. By restoring the isotropy that scalar quantization assumes, spectral denoising eliminates the need for both outlier handling and dedicated inner product bias correction, freeing those bits for improved reconstruction. The theoretical grounding in random matrix theory provides three guarantees: automatic rank selection via the BBP phase transition, provably near-zero inner product bias on the residual, and coordinate delocalization ensuring near-optimal quantization distortion. Experimentally, we validate eOptShrinkQ on Llama-3.1-8B and Ministral-8B across three levels: per-head MSE and inner product fidelity, where eOptShrinkQ saves nearly one bit per entry over TurboQuant at equivalent quality; end-to-end on LongBench (16 tasks), where eOptShrinkQ at $\\sim$2.2 bits per entry outperforms TurboQuant at 3.0 bits; and multi-needle retrieval, where eOptShrinkQ at 2.2 bits closely matches or exceeds uncompressed FP16, suggesting that spectral denoising can act as a beneficial regularizer for retrieval-intensive tasks.", "authors": ["Pei-Chun Su"], "categories": ["cs.LG", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-06", "url": "https://arxiv.org/abs/2605.02905", "pdf_url": "https://arxiv.org/pdf/2605.02905v1", "arxiv_id": "2605.02905", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.342}
{"id": "9c0aee3dfabc0aeeb633a14a1afa1e698326f29e460b541457faded15a1621b4", "sources": ["arxiv", "semantic_scholar"], "title": "Comparative Characterization of KV Cache Management Strategies for LLM Inference", "abstract": "Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.", "authors": ["Oteo Mamo", "Olga Kogiou", "Hyunjin Yi", "Weikuan Yu"], "categories": ["cs.AR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-06", "url": "https://arxiv.org/abs/2604.05012", "pdf_url": "https://arxiv.org/pdf/2604.05012v1", "arxiv_id": "2604.05012", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.342}
{"id": "5acb70ba0d37755d9dd17c8aa87a6c1028581d0aefa80f36972b3915d2b0696e", "sources": ["arxiv", "semantic_scholar"], "title": "Rényi Attention Entropy for Patch Pruning", "abstract": "Transformers are strong baselines in both vision and language because self-attention captures long-range dependencies across tokens. However, the cost of self-attention grows quadratically with the number of tokens. Patch pruning mitigates this cost by estimating per-patch importance and removing redundant patches. To identify informative patches for pruning, we introduce a criterion based on the Shannon entropy of the attention distribution. Low-entropy patches, which receive selective and concentrated attention, are kept as important, while high-entropy patches with attention spread across many locations are treated as redundant. We also extend the criterion from Shannon to Rényi entropy, which emphasizes sharp attention peaks and supports pruning strategies that adapt to task needs and computational limits. In experiments on fine-grained image recognition, where patch selection is critical, our method reduced computation while preserving accuracy. Moreover, adjusting the pruning policy through the Rényi entropy measure yields further gains and improves the trade-off between accuracy and computation.", "authors": ["Hiroaki Aizawa", "Yuki Igaue"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-04", "url": "https://arxiv.org/abs/2604.03803", "pdf_url": "https://arxiv.org/pdf/2604.03803v1", "arxiv_id": "2604.03803", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3405}
{"id": "093508dcf98330e7145f1b62d18510fb0be2be570700e7ca54d10912969ba7a5", "sources": ["arxiv", "semantic_scholar"], "title": "Stochastic KV Routing: Enabling Adaptive Depth-Wise Cache Sharing", "abstract": "Serving transformer language models with high throughput requires caching Key-Values (KVs) to avoid redundant computation during autoregressive generation. The memory footprint of KV caching is significant and heavily impacts serving costs. This work proposes to lessen these memory requirements. While recent work has largely addressed KV cache reduction via compression and eviction along the temporal axis, we argue that the \\emph{depth} dimension offers an orthogonal and robust avenue for optimization. Although prior research suggests that a full cache for every layer is redundant, implementing cross-layer cache sharing remains a practical challenge; existing methods typically suffer from reduced throughput or increased time-to-first-token. In this paper, we demonstrate that dropping a layer's cache offers efficient optimization without information loss. We propose a simple training approach: random cross-layer attention. During training, layers randomly choose to attend either to their own KV states or those of a preceding layer. This stochastic process adapts the model to be robust to various depth-wise cache sharing strategies, ensuring flexibility for unknown hardware constraints at deployment time. Our evaluations show that applying this scheme during pre-training or fine-tuning enables depth-wise cache sharing for various model families. Furthermore, for larger models in data-constrained settings, this approach is suggestive of a regularization-like effect, frequently preserving or improving performance while significantly reducing the cache's memory footprint.", "authors": ["Anastasiia Filippova", "David Grangier", "Marco Cuturi", "João Monteiro"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-03", "url": "https://arxiv.org/abs/2604.22782", "pdf_url": "https://arxiv.org/pdf/2604.22782v1", "arxiv_id": "2604.22782", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3398}
{"id": "a1734fcfed2f160ad1549680e3c0d790c51a4eb3268a023689317cac831cdd64", "sources": ["arxiv", "semantic_scholar"], "title": "QAPruner: Quantization-Aware Vision Token Pruning for Multimodal Large Language Models", "abstract": "Multimodal Large Language Models (MLLMs) have shown strong reasoning ability, but their high computational and memory costs hinder deployment in resource-constrained settings. While Post-Training Quantization (PTQ) and vision token pruning are standard compression techniques, they are usually treated as independent optimizations. In this paper, we show that these two techniques are strongly coupled: naively applying semantic-based token pruning to PTQ-optimized MLLMs can discard activation outliers that are important for numerical stability and thus worsen quantization errors in low-bit regimes (\\textit{e.g.}, W4A4). To address this issue, we propose a quantization-aware vision token pruning framework. Our method introduces a lightweight hybrid sensitivity metric that combines simulated group-wise quantization error with outlier intensity. By combining this metric with standard semantic relevance scores, the method retains tokens that are both semantically informative and robust to quantization. Experiments on standard LLaVA architectures show that our method consistently outperforms naive integration baselines. At an aggressive pruning ratio that retains only 12.5\\% of visual tokens, our framework improves accuracy by 2.24\\% over the baseline and even surpasses dense quantization without pruning. To the best of our knowledge, this is the first method that explicitly co-optimizes vision token pruning and PTQ for accurate low-bit MLLM inference.", "authors": ["Xinhao Wang", "Zhonyu Xia", "Zhiwei Lin", "Zhe Li", "Yongtao Wang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-03", "url": "https://arxiv.org/abs/2604.02816", "pdf_url": "https://arxiv.org/pdf/2604.02816v1", "arxiv_id": "2604.02816", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3398}
{"id": "910adb7abaebfa3718d2e3b741bf4c3747dce8c460da00a598cb4e60e5f4acca", "sources": ["arxiv", "semantic_scholar"], "title": "TokenDance: Scaling Multi-Agent LLM Serving via Collective KV Cache Sharing", "abstract": "Multi-agent LLM applications organize execution in synchronized rounds where a central scheduler gathers outputs from all agents and redistributes the combined context. This All-Gather communication pattern creates massive KV Cache redundancy, because every agent's prompt contains the same shared output blocks, yet existing reuse methods fail to exploit it efficiently. We present TokenDance, a system that scales the number of concurrent agents by exploiting the All-Gather pattern for collective KV Cache sharing. TokenDance's KV Collector performs KV Cache reuse over the full round in one collective step, so the cost of reusing a shared block is paid once regardless of agent count. Its Diff-Aware Storage encodes sibling caches as block-sparse diffs against a single master copy, achieving 11-17x compression on representative workloads. Evaluation on GenerativeAgents and AgentSociety shows that TokenDance supports up to 2.7x more concurrent agents than vLLM with prefix caching under SLO requirement, reduces per-agent KV Cache storage by up to 17.5x, and achieves up to 1.9x prefill speedup over per-request position-independent caching.", "authors": ["Zhuohang Bian", "Feiyang Wu", "Chengrui Zhang", "Hangcheng Dong", "Yun Liang", "Youwei Zhuo"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-03", "url": "https://arxiv.org/abs/2604.03143", "pdf_url": "https://arxiv.org/pdf/2604.03143v1", "arxiv_id": "2604.03143", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3398}
{"id": "a324edd365652390c578a5f03d72bfdc8719e5241e636c188f87adbd45f89fbd", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Experts in Remote Sensing: A Survey", "abstract": "Remote sensing data analysis and interpretation present unique challenges due to the diversity in sensor modalities and spatiotemporal dynamics of Earth observation data. Mixture-of-Experts (MoE) model has emerged as a powerful paradigm that addresses these challenges by dynamically routing inputs to specialized experts designed for different aspects of a task. However, despite rapid progress, the community still lacks a comprehensive review of MoE for remote sensing. This survey provides the first systematic overview of MoE applications in remote sensing, covering fundamental principles, architectural designs, and key applications across a variety of remote sensing tasks. The survey also outlines future trends to inspire further research and innovation in applying MoE to remote sensing.", "authors": ["Yongchuan Cui", "Peng Liu", "Lajiao Chen"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-03", "url": "https://arxiv.org/abs/2604.03342", "pdf_url": "https://arxiv.org/pdf/2604.03342v1", "arxiv_id": "2604.03342", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "https://www.icck.org/article/abs/jgrs.2025.140654", "quality_score": 0.534}
{"id": "a62dab8eb8e64f4c5e996644409c8b60bd38b04c9876d49cf6b00644f12629f7", "sources": ["arxiv", "semantic_scholar"], "title": "The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level", "abstract": "Mixture-of-Experts (MoE) architectures have become the dominant choice for scaling Large Language Models (LLMs), activating only a subset of parameters per token. While MoE architectures are primarily adopted for computational efficiency, it remains an open question whether their sparsity makes them inherently easier to interpret than dense feed-forward networks (FFNs). We compare MoE experts and dense FFNs using $k$-sparse probing and find that expert neurons are consistently less polysemantic, with the gap widening as routing becomes sparser. This suggests that sparsity pressures both individual neurons and entire experts toward monosemanticity. Leveraging this finding, we zoom out from the neuron to the expert level as a more effective unit of analysis. We validate this approach by automatically interpreting hundreds of experts. This analysis allows us to resolve the debate on specialization: experts are neither broad domain specialists (e.g., biology) nor simple token-level processors. Instead, they function as fine-grained task experts, specializing in linguistic operations or semantic tasks (e.g., closing brackets in $\\LaTeX{}$). Our findings suggest that MoEs are inherently interpretable at the expert level, providing a clearer path toward large-scale model interpretability. Code is available at: https://github.com/jerryy33/MoE_analysis.", "authors": ["Jeremy Herbst", "Stefan Wermter", "Jae Hee Lee"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-02", "url": "https://arxiv.org/abs/2604.02178", "pdf_url": "https://arxiv.org/pdf/2604.02178v2", "arxiv_id": "2604.02178", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jerryy33/MoE_analysis", "venue": null, "quality_score": 0.6297}
{"id": "f60562b1504d3ed5cb95bc02897ba005ed8fdaf8f74f86f08730c4fcd0d70658", "sources": ["arxiv", "semantic_scholar"], "title": "Attention to Mamba: A Recipe for Cross-Architecture Distillation", "abstract": "State Space Models (SSMs) such as Mamba have become a popular alternative to Transformer models, due to their reduced memory consumption and higher throughput at generation compared to their Attention-based counterparts. On the other hand, the community has built up a considerable body of knowledge on how to train Transformers, and many pretrained Transformer models are readily available. To facilitate the adoption of SSMs while leveraging existing pretrained Transformers, we aim to identify an effective recipe to distill an Attention-based model into a Mamba-like architecture. In prior work on cross-architecture distillation, however, it has been shown that a naïve distillation procedure from Transformers to Mamba fails to preserve the original teacher performance, a limitation often overcome with hybrid solutions combining Attention and SSM blocks. The key argument from our work is that, by equipping Mamba with a principled initialization, we can recover an overall better recipe for cross-architectural distillation. To this end, we propose a principled two-stage approach: first, we distill knowledge from a traditional Transformer into a linearized version of Attention, using an adaptation of the kernel trick. Then, we distill the linearized version into an adapted Mamba model that does not use any Attention block. Overall, the distilled Mamba model is able to preserve the original Pythia-1B Transformer performance in downstream tasks, maintaining a perplexity of 14.11 close to the teacher's 13.86. To show the efficacy of our recipe, we conduct thorough ablations at 1B scale with 10B tokens varying sequence mixer architecture, scaling analysis on model sizes and total distillation tokens, and a sensitivity analysis on tokens allocation between stages.", "authors": ["Abhinav Moudgil", "Ningyuan Huang", "Eeshan Gunesh Dhekane", "Pau Rodríguez", "Luca Zappella", "Federico Danieli"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-01", "url": "https://arxiv.org/abs/2604.14191", "pdf_url": "https://arxiv.org/pdf/2604.14191v1", "arxiv_id": "2604.14191", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3383}
{"id": "1d2e9986934d368ad386fc2e878f262e2c5616acbfca731dceb5fe1e507ad859", "sources": ["arxiv", "semantic_scholar"], "title": "Routing-Free Mixture-of-Experts", "abstract": "Standard Mixture-of-Experts (MoE) models rely on centralized routing mechanisms that introduce rigid inductive biases. We propose Routing-Free MoE which eliminates any hard-coded centralized designs including external routers, Softmax, Top-K and load balancing, instead encapsulating all activation functionalities within individual experts and directly optimized through continuous gradient flow, enabling each expert to determine its activation entirely on its own. We introduce a unified adaptive load-balancing framework to simultaneously optimize both expert-balancing and token-balancing objectives through a configurable interpolation, allowing flexible and customizable resource allocation. Extensive experiments show that Routing-Free MoE can consistently outperform baselines with better scalability and robustness. We analyze its behavior in detail and offer insights that may facilitate future MoE design ad optimization.", "authors": ["Yilun Liu", "Jinru Han", "Sikuan Yan", "Volker Tresp", "Yunpu Ma"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-01", "url": "https://arxiv.org/abs/2604.00801", "pdf_url": "https://arxiv.org/pdf/2604.00801v1", "arxiv_id": "2604.00801", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/liuyilun2000/RoutingFreeMoE/tree/release", "venue": null, "quality_score": 0.6283}
{"id": "011fd8755dbdc3b571d133cc955add35d6a97cd5aa0e16945a88b68857cd44de", "sources": ["arxiv", "semantic_scholar"], "title": "Cost-Penalized Fitness in FMA-Orchestrated Mixture of Experts: Experimental Evidence for Molecular Memory in Domain Adaptation", "abstract": "We present experimental results from seven controlled runs of nanoFMT, a Free-Market Algorithm (FMA) orchestrated transformer with dynamic Mixture-of-Experts (MoE) management. The experiments address a fundamental question for advanced LLM development: how should an MoE system manage its expert pool when operating at full capacity under changing data distributions? We demonstrate that cost-penalized fitness metrics, combined with a linear grace period for newborn experts, produce a system that accumulates domain expertise through diversification rather than replacement. The central result is a round-trip domain shift experiment showing 9-11x faster recovery when returning to a previously learned domain, with zero expert births or replacements required. This \"molecular memory\" effect -- where dormant experts survive and reactivate when their domain returns -- has no analogue in current MoE management approaches. A preliminary cost analysis estimates annual savings of $39.1M and 27.1 GWh energy reduction for an OpenAI-scale provider under a moderate scenario.", "authors": ["Martin Jaraiz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-01", "url": "https://arxiv.org/abs/2604.00812", "pdf_url": "https://arxiv.org/pdf/2604.00812v1", "arxiv_id": "2604.00812", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3383}
{"id": "b8aecc7782f427271f02090f4830d1029c9afffe3a1be42efda589167dbdcb3f", "sources": ["arxiv", "semantic_scholar"], "title": "IWP: Token Pruning as Implicit Weight Pruning in Large Vision Language Models", "abstract": "Large Vision Language Models show impressive performance across image and video understanding tasks, yet their computational cost grows rapidly with the number of visual tokens. Existing token pruning methods mitigate this issue through empirical approaches while overlooking the internal mechanism of attention. In this paper, we propose a novel training free token pruning framework grounded in the dual form perspective of attention. We reformulate attention as an implicit linear layer whose weight matrix is the sum of rank 1 outer products, each generated by a single token's key value pair. Token pruning thus reduces to selecting an optimal subset of these rank 1 updates that best approximates the original dual weight matrix. Extending this perspective to standard softmax attention in LVLMs, we derive a novel metric quantifying both a token's information magnitude and information duplication. To efficiently select the subset with the proposed metric, we introduce Progressive Chunked Maximal Marginal Relevance. Extensive experiments demonstrate that our method achieves a better trade off between performance and efficiency, while providing another perspective on existing pruning approaches.", "authors": ["Dong-Jae Lee", "Sunghyun Baek", "Junmo Kim"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-01", "url": "https://arxiv.org/abs/2604.00757", "pdf_url": "https://arxiv.org/pdf/2604.00757v1", "arxiv_id": "2604.00757", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3383}
{"id": "4feedb03b6b9207675d85a30b385886576a915f156dce9f8375467c65f26112c", "sources": ["arxiv", "semantic_scholar"], "title": "Quantization with Unified Adaptive Distillation to enable multi-LoRA based one-for-all Generative Vision Models on edge", "abstract": "Generative Artificial Intelligence (GenAI) features such as image editing, object removal, and prompt-guided image transformation are increasingly integrated into mobile applications. However, deploying Large Vision Models (LVMs) for such tasks on resource-constrained devices remains challenging due to their high memory and compute requirements. While Low-Rank Adapters (LoRAs) enable parameter-efficient task adaptation, existing Mobile deployment pipelines typically compile separate model binaries for each LoRA + a copy of the foundation model, resulting in redundant storage and increased runtime overhead. In this work, we present a unified framework for enabling multi-task GenAI inference on edge devices using a single shared model. Our key idea is to treat LoRA weights as runtime inputs rather than embedding them into the compiled model graph, allowing dynamic task switching at runtime without recompilation. Then, to support efficient on-device execution, we introduce QUAD (Quantization with Unified Adaptive Distillation), a quantizationaware training strategy that aligns multiple LoRA adapters under a shared quantization profile. We implement the proposed system with a lightweight runtime stack compatible with mobile NPUs and evaluate it across multiple chipsets. Experimental results demonstrate up to 6x and 4x reduction in memory footprint and latency improvements, respectively, while maintaining high visual quality across multiple GenAI tasks.", "authors": ["Sowmya Vajrala", "Aakash Parmar", "Prasanna R", "Sravanth Kodavanti", "Manjunath Arveti", "Srinivas Soumitri Miriyala", "Ashok Senapati"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-31", "url": "https://arxiv.org/abs/2603.29535", "pdf_url": "https://arxiv.org/pdf/2603.29535v1", "arxiv_id": "2603.29535", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3376}
{"id": "fd27f171f34c8f163c221f6302243f1560363adf5287ed45cd2482786ee9cd70", "sources": ["arxiv", "semantic_scholar"], "title": "On the Role of Encoder Depth: Pruning Whisper and LoRA Fine-Tuning in SLAM-ASR", "abstract": "Automatic speech recognition (ASR) has advanced rapidly in recent years, driven by large-scale pretrained models and end-to-end architectures such as SLAM-ASR. A key component of SLAM-ASR systems is the Whisper speech encoder, which provides robust acoustic representations. While model pruning has been explored for the full Whisper encoder-decoder architecture, its impact within the SLAM-ASR setting remains under-investigated. In this work, we analyze the effects of layer pruning in the Whisper encoder when used as the acoustic backbone of SLAM-ASR. We further examine the extent to which LoRA-based fine-tuning can recover performance degradation caused by pruning. Experiments conducted across three Whisper variants (Small, Medium, Large-v2), three languages representing distinct resource levels (Danish, Dutch, English), and over 200 training runs demonstrate that pruning two encoder layers causes only 2-4% WER degradation, and that combining this pruning with LoRA adaptation consistently outperforms the unpruned baseline while reducing total parameters by 7-14%. Moreover, our error analysis reveals that LoRA primarily compensates through the language model's linguistic priors, reducing total word errors by 11-21% for Dutch and English, with substitutions and deletions showing the largest reductions. However, for low-resource Danish, the reduction is smaller (4-7%), and LoRA introduces increased insertion errors, indicating that compensation effectiveness depends on the LLM's pre-existing language proficiency and available training data.", "authors": ["Ganesh Pavan Kartikeya Bharadwaj Kolluri", "Michael Kampouridis", "Ravi Shekhar"], "categories": ["cs.CL", "cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-30", "url": "https://arxiv.org/abs/2603.27981", "pdf_url": "https://arxiv.org/pdf/2603.27981v1", "arxiv_id": "2603.27981", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3369}
{"id": "c75bedd21fe657046b7aacafac1a6f00ab2e41bcedcf9662dce2ef498807745d", "sources": ["arxiv", "semantic_scholar"], "title": "IsoQuant: Hardware-Aligned SO(4) Isoclinic Rotations for LLM KV Cache Compression", "abstract": "Orthogonal feature decorrelation is effective for low-bit online vector quantization, but dense random orthogonal transforms incur prohibitive $O(d^2)$ storage and compute. RotorQuant reduces this cost with blockwise $3$D Clifford rotors, yet the resulting $3$D partition is poorly aligned with modern hardware and offers limited local mixing. We propose \\textbf{IsoQuant}, a blockwise rotation framework based on quaternion algebra and the isoclinic decomposition of $SO(4)$. It represents each $4$D block as a quaternion and applies a closed-form transform $T(v)=q_L v \\overline{q_R}$. This yields two main variants: \\emph{IsoQuant-Full}, which realizes the full $SO(4)$ rotation, and \\emph{IsoQuant-Fast}, which keeps only one isoclinic factor for lower cost; the framework also admits a lightweight $2$D special case. At $d=128$, IsoQuant-Full reduces forward rotation cost from about $2{,}408$ FMAs in RotorQuant to $1{,}024$, while IsoQuant-Fast further reduces it to $512$. Across $18$ fused CUDA settings with $d \\in {128,256,512}$, bit widths ${2,3,4}$, and FP16/FP32 execution, IsoQuant achieves mean kernel-level speedups of about $4.5\\times$--$4.7\\times$ over RotorQuant while maintaining comparable reconstruction MSE, with peak speedups above $6\\times$. Current validation is limited to the stage-1 quantize--dequantize path on synthetic normalized vectors; end-to-end KV-cache evaluation remains future work.", "authors": ["Zhongping Ji"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-30", "url": "https://arxiv.org/abs/2603.28430", "pdf_url": "https://arxiv.org/pdf/2603.28430v1", "arxiv_id": "2603.28430", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3369}
{"id": "6b37f8878e1b8c13df5c169f008fea4531d46e0cd1f1d61cd090a68aa9da925a", "sources": ["arxiv", "semantic_scholar"], "title": "Low-Latency Edge LLM Handover via Joint KV Cache Transfer and Token Prefill", "abstract": "Edge deployment of large language models (LLMs) can reduce latency for interactive services, but mobility introduces service interruptions when an user equipment (UE) hands over between base stations (BSs). To promptly resume decoding, the target-side edge server must recover the UE context state, which can be provisioned either by token forwarding followed by prefill computation or by direct key-value (KV) cache transmission over backhaul. This paper proposes a unified handover (HO) design that jointly selects the prefill length and schedules backhaul KV cache delivery to minimize the worst-user LLM HO delay for multiple UEs. The resulting scheme admits a tractable step-wise solution with explicit feasibility conditions and a constructive rate-scheduling policy. Simulations show that the proposed method consistently outperforms baselines across a wide range of backhaul capacities, prefill speeds, and context sizes, providing practical guidelines for mobility-aware Edge LLM token streaming.", "authors": ["Seunghun Lee", "Jihong Park", "Ce Zheng", "Hyuncheol Park"], "categories": ["eess.SP"], "fields_of_study": ["Engineering"], "published_date": "2026-03-30", "url": "https://arxiv.org/abs/2603.28018", "pdf_url": "https://arxiv.org/pdf/2603.28018v1", "arxiv_id": "2603.28018", "doi": null, "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3369}
{"id": "45f5bec4189d537fc0ceb1a84d25d17b1df005f2eab557b46525ebb8b5d668fd", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache Quantization for Self-Forcing Video Generation: A 33-Method Empirical Study", "abstract": "Self-forcing video generation extends a short-horizon video model to longer rollouts by repeatedly feeding generated content back in as context. This scaling path immediately exposes a systems bottleneck: the key-value (KV) cache grows with rollout length, so longer videos require not only better generation quality but also substantially better memory behavior. We present a comprehensive empirical study of KV-cache compression for self-forcing video generation on a Wan2.1-based Self-Forcing stack. Our study covers 33 quantization and cache-policy variants, 610 prompt-level observations, and 63 benchmark-level summaries across two evaluation settings: MovieGen for single-shot 10-second generation and StoryEval for longer narrative-style stability. We jointly evaluate peak VRAM, runtime, realized compression ratio, VBench imaging quality, BF16-referenced fidelity (SSIM, LPIPS, PSNR), and terminal drift. Three findings are robust. First, the strongest practical operating region is a FlowCache-inspired soft-prune INT4 adaptation, which reaches 5.42-5.49x compression while reducing peak VRAM from 19.28 GB to about 11.7 GB with only modest runtime overhead. Second, the highest-fidelity compressed methods, especially PRQ_INT4 and QUAROT_KV_INT4, are not the best deployment choices because they preserve quality at severe runtime or memory cost. Third, nominal compression alone is not sufficient: several methods shrink KV storage but still exceed BF16 peak VRAM because the current integration reconstructs or retains large BF16 buffers during attention and refresh stages. The result is a benchmark harness, analysis workflow, and empirical map of which KV-cache ideas are practical today and which are promising research directions for better memory integration. Code, data products, and the presentation dashboard are available at https://github.com/suraj-ranganath/kv-quant-longhorizon/.", "authors": ["Suraj Ranganath", "Vaishak Menon", "Anish Patnaik"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-29", "url": "https://arxiv.org/abs/2603.27469", "pdf_url": "https://arxiv.org/pdf/2603.27469v1", "arxiv_id": "2603.27469", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/suraj-ranganath/kv-quant-longhorizon/", "venue": null, "quality_score": 0.6243}
{"id": "de79a316f833cba93384f952cde4c399b561cc29f2f6bd0d7a84a425e4583bae", "sources": ["arxiv", "semantic_scholar"], "title": "KVSculpt: KV Cache Compression as Distillation", "abstract": "KV cache compression is critical for efficient long-context LLM inference. Approaches that reduce the per-pair footprint -- quantization and low-rank decomposition -- are orthogonal to those that reduce the sequence length of the cache. Along the sequence-length dimension, existing methods range from pure eviction -- selecting which KV pairs to keep -- to merging, which combines similar pairs into fewer ones. Both remain anchored to the original cache entries. We propose KVSculpt, which moves to the other end of this spectrum: instead of selecting or combining original pairs, we optimize a smaller set of unconstrained KV pairs in continuous embedding space to preserve each layer's attention behavior. Keys are optimized via L-BFGS and values are solved in closed form via least squares, alternating every few steps. On top of this, we introduce adaptive budget allocation, which uses a cheap pilot compression run to redistribute the compression budget across layers and KV heads based on per-component difficulty. On Qwen2.5-1.5B-Instruct with 2048-token contexts, KVSculpt reduces KL divergence by 3.5-4.1x compared to Select+Fit -- attention-score eviction with least-squares value fitting -- across compression ratios r in {0.3, 0.5, 0.7}. Adaptive allocation provides an additional 1.3x KL reduction at no extra inference cost. Analysis reveals that compression difficulty is highly non-uniform: per-layer pilot MSE varies by up to 100x across layers, and the two KV heads within a single layer can differ by up to 467x -- demonstrating that fine-grained budget allocation is essential.", "authors": ["Bo Jiang", "Sian Jin"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-29", "url": "https://arxiv.org/abs/2603.27819", "pdf_url": "https://arxiv.org/pdf/2603.27819v1", "arxiv_id": "2603.27819", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3361}
{"id": "195271ef729656e923cf5f02544ee6086b42e078b17382ad0228b602e4cbbd9e", "sources": ["arxiv", "semantic_scholar"], "title": "TurboAngle: Near-Lossless KV Cache Compression via Uniform Angle Quantization", "abstract": "We compress KV cache entries by quantizing angles in the Fast Walsh-Hadamard domain, where a random diagonal rotation makes consecutive element pairs approximately uniformly distributed on the unit circle. We extend this angular quantizer with per-layer early-boost, which independently configures K and V codebook sizes at each layer, allocating higher precision to a model-specific subset of critical layers. Across seven models (1B to 7B parameters), per-layer early-boost achieves lossless compression on four models and near-lossless quality on six of seven, at 3.28 to 3.67 angle bits per element. Asymmetric norm quantization (8-bit for keys, 4-bit log-space for values) yields 6.56 total bits per element on Mistral-7B with perplexity degradation of +0.0014 and no calibration data. A layer-group sensitivity analysis reveals model-specific bottleneck patterns, including K-dominated versus V-dominated layers and negative-transfer layers where increased precision degrades quality.", "authors": ["Dipkumar Patel"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-29", "url": "https://arxiv.org/abs/2603.27467", "pdf_url": "https://arxiv.org/pdf/2603.27467v1", "arxiv_id": "2603.27467", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3361}
{"id": "dddaee0a13abd30e37ea3ffd2f1facd8d17bdcddb97e29a535dbb97fe6cf0e83", "sources": ["arxiv", "semantic_scholar"], "title": "ScoutAttention: Efficient KV Cache Offloading via Layer-Ahead CPU Pre-computation for LLM Inference", "abstract": "Large language models encounter critical GPU memory capacity constraints during long-context inference, where KV cache memory consumption severely limits decode batch sizes. While existing research has explored offloading KV cache to DRAM, these approaches either demand frequent GPU-CPU data transfers or impose extensive CPU computation requirements, resulting in poor GPU utilization as the system waits for I/O operations or CPU processing to complete. We propose ScoutAttention, a novel KV cache offloading framework that accelerates LLM inference through collaborative GPU-CPU attention computation. To prevent CPU computation from bottlenecking the system, ScoutAttention introduces GPU-CPU collaborative block-wise sparse attention that significantly reduces CPU load. Unlike conventional parallel computing approaches, our framework features a novel layer-ahead CPU pre-computation algorithm, enabling the CPU to initiate attention computation one layer in advance, complemented by asynchronous periodic recall mechanisms to maintain minimal CPU compute load. Experimental results demonstrate that ScoutAttention maintains accuracy within 2.4% of baseline while achieving 2.1x speedup compared to existing offloading methods.", "authors": ["Qiuyang Zhang", "Kai Zhou", "Ding Tang", "Kai Lu", "Cheng Li", "Zhenyu Yang", "Peng Xu", "Jiguang Wan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-28", "url": "https://arxiv.org/abs/2603.27138", "pdf_url": "https://arxiv.org/pdf/2603.27138v1", "arxiv_id": "2603.27138", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3354}
{"id": "29ac0b5282c6eef232598d0df8e417496472c94112baad74f59aa6a81b4325ab", "sources": ["arxiv", "semantic_scholar"], "title": "TurboESM: Ultra-Efficient 3-Bit KV Cache Quantization for Protein Language Models with Orthogonal Rotation and QJL Correction", "abstract": "The rapid scaling of Protein Language Models (PLMs) has unlocked unprecedented accuracy in protein structure prediction and design, but the quadratic memory growth of the Key-Value (KV) cache during inference remains a prohibitive barrier for single-GPU deployment and high-throughput generation. While 8-bit quantization is now standard, 3-bit quantization remains elusive due to severe numerical outliers in activations. This paper presents TurboESM, an adaptation of Google's TurboQuant to the PLM domain. We solve the fundamental incompatibility between Rotary Position Embeddings (RoPE) and orthogonal transformations by deriving a RoPE-first rotation pipeline. We introduce a head-wise SVD calibration method tailored to the amino acid activation manifold, a dual look-up table (LUT) strategy for asymmetric K/V distributions, and a 1-bit Quantized Johnson-Lindenstrauss (QJL) residual correction. All experiments are conducted on ESM-2 650M, where our implementation achieves a 7.1x memory reduction (330 MB to 47 MB) while maintaining cosine similarity > 0.96 in autoregressive decoding across diverse protein families, including short peptides, transmembrane helices, enzyme active site fragments, and intrinsically disordered regions. We further implement a Triton-based fused decode attention kernel that eliminates intermediate dequantization memory allocations, achieving a 1.96x speedup over the PyTorch two-step path for the KV fetch operation alone; however, TurboESM incurs a prefill overhead of 21-27 ms relative to the original model due to KV quantization and packing, making it most suitable for memory-bound scenarios rather than latency-critical short-sequence workloads. Analysis reveals that PLMs exhibit sharper outlier profiles than large language models (LLMs) due to amino acid vocabulary sparsity, and our method effectively addresses these distributions.", "authors": ["Yue Hu", "Junqing Wang", "Yingchao Liu"], "categories": ["q-bio.QM"], "fields_of_study": ["Biology"], "published_date": "2026-03-27", "url": "https://arxiv.org/abs/2603.26110", "pdf_url": "https://arxiv.org/pdf/2603.26110v1", "arxiv_id": "2603.26110", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3347}
{"id": "afed2caba7ddda81f5c0da9f292ebf40a68ef5edd83468504f10576a804e3c1a", "sources": ["arxiv", "semantic_scholar"], "title": "TTKV: Temporal-Tiered KV Cache for Long-Context LLM Inference", "abstract": "Key-value (KV) caching is critical for efficient inference in large language models (LLMs), yet its memory footprint scales linearly with context length, resulting in a severe scalability bottleneck. Existing approaches largely treat KV states as equally important across time, implicitly assuming uniform precision and accessibility. However, this assumption contrasts with human memory systems, where memories vary in clarity, recall frequency, and relevance with temporal proximity.Motivated by this insight, we propose TTKV, a KV cache management framework that maps the human memory system onto the KV cache. TTKV partitions the KV cache into temporal tiers with heterogeneous capacity and precision. The design addresses three aspects: (1) Tier Layout, decoupling fast and slow memory using HBM and DRAM; (2) Tier Content, assigning more recent KV states to faster, higher-precision tiers based on temporal proximity; and (3) Tier Interaction, employing block-wise streaming attention to overlap communication and computation when accessing slow tiers. Experiments show that TTKV reduces cross-tier traffic by 5.94x on 128K-context tasks, achieving up to 76% latency reduction and 2x throughput improvement over strong baselines.", "authors": ["Gradwell Dzikanyanga", "Weihao Yang", "Hao Huang", "Donglei Wu", "Shihao Wang", "Wen Xia", "Sanjeeb K C"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-27", "url": "https://arxiv.org/abs/2604.19769", "pdf_url": "https://arxiv.org/pdf/2604.19769v1", "arxiv_id": "2604.19769", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3347}
{"id": "6c4934016b13a0b356a18c4473ac9eecda572f44dc87c41d9301485ed767a327", "sources": ["arxiv", "semantic_scholar"], "title": "Bayesian estimation of optical constants using mixtures of Gaussian process experts", "abstract": "We propose modeling absorption spectrum measurements as mixtures of Gaussian process experts. This enables us to construct a flexible statistical model for interpolating and extrapolating measurements, facilitating statistical integration of Kramers-Kronig relations to estimate the whole complex refractive index. Additionally, we statistically model the anchoring points used in subtractive Kramers-Kronig relations to account for possible measurement errors of the anchor point. In addition to flexible statistical modeling, the mixtures of Gaussian process formulation enables automatic selection of measurement points to use for extrapolation. We apply the method to experimental absorption spectrum measurements of gallium arsenide, potassium chloride, and transparent wood.", "authors": ["Teemu Härkönen", "Hui Chen", "Erik Vartiainen"], "categories": ["stat.AP", "physics.data-an"], "fields_of_study": ["Mathematics", "Physics"], "published_date": "2026-03-27", "url": "https://arxiv.org/abs/2603.26334", "pdf_url": "https://arxiv.org/pdf/2603.26334v1", "arxiv_id": "2603.26334", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3347}
{"id": "f3b5b82a8737ec2dc0b25dce7e024e60ff54d8cf0fdaf797a73a193e81137698", "sources": ["arxiv", "semantic_scholar"], "title": "How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models", "abstract": "Weight pruning is a standard technique for compressing large language models, yet its effect on learned internal representations remains poorly understood. We present the first systematic study of how unstructured pruning reshapes the feature geometry of language models, using Sparse Autoencoders (SAEs) as interpretability probes. Across three model families (Gemma 3 1B, Gemma 2 2B, Llama 3.2 1B), two pruning methods (magnitude and Wanda), and six sparsity levels (0--60%), we investigate five research questions spanning seed stability, feature survival, SAE transferability, feature fragility, and causal relevance. Our most striking finding is that rare SAE features--those with low firing rates--survive pruning far better than frequent ones, with within-condition Spearman correlations of rho = -1.0 in 11 of 17 experimental conditions. This counter-intuitive result suggests that pruning acts as implicit feature selection, preferentially destroying high-frequency generic features while preserving specialized rare ones. We further show that Wanda pruning preserves feature structure up to 3.7x better than magnitude pruning, that pre-trained SAEs remain viable on Wanda-pruned models up to 50% sparsity, and that geometric feature survival does not predict causal importance--a dissociation with implications for interpretability under compression.", "authors": ["Hector Borobia", "Elies Seguí-Mas", "Guillermina Tormo-Carbó"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-26", "url": "https://arxiv.org/abs/2603.25325", "pdf_url": "https://arxiv.org/pdf/2603.25325v1", "arxiv_id": "2603.25325", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/hborobia/sae-pruning-paper", "venue": null, "quality_score": 0.6202}
{"id": "bcf6a3e028a27d0e6c9e7c0b3c711b4019234fcd93e71ef99136128515d40236", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-Sieve: Routing-Guided LoRA for Efficient MoE Fine-Tuning", "abstract": "Standard LoRA fine-tuning of Mixture-of-Experts (MoE) models applies adapters to every expert, yet our profiling shows that per-layer expert routing is highly skewed: a small subset of experts handles most tokens in each layer, while many others are rarely activated (\"cold\"). We propose MoE-Sieve, a simple routing-guided framework for LoRA fine-tuning, and pair it with a systematic profiling study of expert routing across architectures and tasks. The method is simple: profile routing counts on a small calibration set, select the top-k most-routed experts per layer, and apply LoRA only to those experts. Across two architecturally distinct MoE models and three diverse tasks, tuning only the top 25% routed experts per layer remains competitive with full LoRA, with mean differences within +/-1 percentage point across all conditions. This reduces LoRA trainable parameters by 70-73%, adapter checkpoint size by 71-73%, and wall-clock training time by up to 50%. We also observe a non-monotonic relationship between expert count and seed-to-seed variance, consistent with the hypothesis that adapting cold experts can introduce gradient noise without improving accuracy. Further ablations show that random expert selection at matched budget is about 2.5 percentage points worse, indicating that the routing signal matters, while greedy per-layer budget optimization does not improve over uniform top-k.", "authors": ["Andrea Manzoni"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-25", "url": "https://arxiv.org/abs/2603.24044", "pdf_url": "https://arxiv.org/pdf/2603.24044v1", "arxiv_id": "2603.24044", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3332}
{"id": "62df5e004553baa4d283ee0b70c1b4eeeec422d0d171e20a5ce568623093473b", "sources": ["arxiv", "semantic_scholar"], "title": "EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction", "abstract": "The increasing memory demand of the Key-Value (KV) cache poses a significant bottleneck for Large Language Models (LLMs) in long-context applications. Existing low-rank KV compression methods reduce this footprint by modifying model projections, limiting the flexibility to switch back to standard full-cache inference when sufficient memory is available. In this paper, we propose EchoKV, a flexible KV cache compression framework that supports on-demand transitions from full KV caching to compressed caching. Unlike traditional compression-decompression paradigms, EchoKV utilizes a lightweight network to reconstruct the discarded KV components from a partial subset, exploiting intrinsic inter-layer and intra-layer similarities among attention heads. We further introduce a lightweight two-stage fine-tuning strategy, requiring only a few minutes on a single A100 GPU for a 7B model. Experimental results on LongBench and RULER demonstrate that EchoKV consistently outperforms existing methods across multiple compression ratios and backbone models while preserving the throughput of full-cache inference in short-context scenarios.", "authors": ["Shiyu Ji", "Yixuan Wang", "Yijun Liu", "Qingfu Zhu", "Wanxiang Che"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-24", "url": "https://arxiv.org/abs/2603.22910", "pdf_url": "https://arxiv.org/pdf/2603.22910v2", "arxiv_id": "2603.22910", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3325}
{"id": "dec77340d7e040a40a873747bdb5bd3cb23b234de68ffb37624da3b3ac4fa688", "sources": ["arxiv", "semantic_scholar"], "title": "PCR: A Prefetch-Enhanced Cache Reuse System for Low-Latency RAG Serving", "abstract": "Retrieval-Augmented Generation (RAG) systems enhance the performance of large language models (LLMs) by incorporating supplementary retrieved documents, enabling more accurate and context-aware responses. However, integrating these external documents often results in very long input sequences, which significantly increases computation costs during the prefill stage, where key-value (KV) representations for all input tokens are generated. This latency bottleneck becomes especially pronounced under high-throughput serving scenarios. KV-cache reuse offers a promising solution by storing previously computed KV states for shared input prefixes, thereby avoiding redundant computation across requests that contain overlapping context. Yet, the effectiveness of cache reuse is often limited by three practical challenges: low cache hit rates due to naive eviction policies, high CPU-GPU data transfer overhead, and slow SSD I/O when caches spill to storage. To address these issues, we propose PCR, a system designed to maximize KV-cache reuse efficiency through intelligent prefetching and pipelined data movement. Specifically, PCR introduces three key techniques: (1) a prefix-tree caching structure with a look-ahead LRU replacement policy that uses pending requests in the scheduler queue to improve cache hit ratios; (2) layer-wise overlapping that pipelines KV-cache loading and GPU computation across CUDA streams to hide communication latency; and (3) queue-based prefetching that proactively loads relevant KV caches from SSD into DRAM before they are needed. Extensive experiments show that PCR outperforms existing KV-cache reuse methods, achieving up to a 2.47x speedup in terms of average TTFT.", "authors": ["Wenfeng Wang", "Xiaofeng Hou", "Peng Tang", "Hengyi Zhou", "Jing Wang", "Xinkai Wang", "Chao Li", "Minyi Guo"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-24", "url": "https://arxiv.org/abs/2603.23049", "pdf_url": "https://arxiv.org/pdf/2603.23049v1", "arxiv_id": "2603.23049", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3325}
{"id": "f25c2bf34020c76bf7e0dd7b28de3e3c3d51026e3ce76f6505ed71994b7adea1", "sources": ["arxiv", "semantic_scholar"], "title": "Knowledge Packs: Zero-Token Knowledge Delivery via KV Cache Injection", "abstract": "RAG wastes tokens. We propose Knowledge Packs: pre-computed KV caches that deliver the same knowledge at zero token cost. For causal transformers, the KV cache from a forward pass on text F is identical to what a joint pass on F+q would produce - this follows directly from the causal mask. The equivalence is exact but fragile: wrong chat template formatting causes 6-7pp degradation, which we believe explains prior claims of KV outperforming RAG. With correct formatting: zero divergences across 700 questions on Qwen3-8B and Llama-3.1-8B, up to 95% token savings. The KV interface also enables behavioral steering that RAG cannot do. Because RoPE rotates keys but leaves values untouched, contrastive deltas on cached values can nudge model behavior while key arithmetic destroys coherence. The effect sits in mid-layer values (33-66%), independent directions are nearly orthogonal (cos~0) and compose, and both channels - knowledge and steering - run simultaneously at alpha<=0.7 without interference. No training, no weight modification.", "authors": ["Andrey Pustovit"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-22", "url": "https://arxiv.org/abs/2604.03270", "pdf_url": "https://arxiv.org/pdf/2604.03270v1", "arxiv_id": "2604.03270", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/cnails/kv-knowledge-packs", "venue": null, "quality_score": 0.6148}
{"id": "912b19da08c9ca08ebdcf35648f0ee66f59e9de9f271f3d5d81510bada193aef", "sources": ["arxiv", "semantic_scholar"], "title": "HELIX: Scaling Raw Audio Understanding with Hybrid Mamba-Attention Beyond the Quadratic Limit", "abstract": "Audio representation learning typically evaluates design choices such as input frontend, sequence backbone, and sequence length in isolation. We show that these axes are coupled, and conclusions from one setting often do not transfer to others. We introduce HELIX, a controlled framework comparing pure Mamba, pure attention, and a minimal hybrid with a single attention bottleneck. All models are parameter-matched at about 8.3M parameters to isolate architectural effects. Across six datasets, we find that the preferred input representation depends on the backbone, and that attention hurts performance on short, stationary audio but becomes important at longer sequence lengths. On a 5-minute speaker identification task with 30,000 tokens, pure attention fails with out-of-memory errors, while HELIX closes an 11.5-point gap over pure Mamba.", "authors": [" Khushiyant", "Param Thakkar"], "categories": ["cs.SD", "cs.LG", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-03-22", "url": "https://arxiv.org/abs/2603.21316", "pdf_url": "https://arxiv.org/pdf/2603.21316v1", "arxiv_id": "2603.21316", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.331}
{"id": "a0e433d3b6b7886b962a68ddb00bfe5a79a718a5e9893fab7604f24f34e805d9", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Token Eviction: Mixed-Dimension Budget Allocation for Efficient KV Cache Compression", "abstract": "Key-value (KV) caching is widely used to accelerate transformer inference, but its memory cost grows linearly with input length, limiting long-context deployment. Existing token eviction methods reduce memory by discarding less important tokens, which can be viewed as a coarse form of dimensionality reduction that assigns each token either zero or full dimension. We propose MixedDimKV, a mixed-dimension KV cache compression method that allocates dimensions to tokens at a more granular level, and MixedDimKV-H, which further integrates head-level importance information. Experiments on long-context benchmarks show that MixedDimKV outperforms prior KV cache compression methods that do not rely on head-level importance profiling. When equipped with the same head-level importance information, MixedDimKV-H consistently outperforms HeadKV. Notably, our approach achieves comparable performance to full attention on LongBench with only 6.25% of the KV cache. Furthermore, in the Needle-in-a-Haystack test, our solution maintains 100% accuracy at a 50K context length while using as little as 0.26% of the cache.", "authors": ["Ruijie Miao", "Zhiming Wang", "Wang Li", "Shiwei Wu", "Shufan Liu", "Yanbing Jiang", "Tong Yang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-21", "url": "https://arxiv.org/abs/2603.20616", "pdf_url": "https://arxiv.org/pdf/2603.20616v1", "arxiv_id": "2603.20616", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3303}
{"id": "59ca1cb0036a67455fcc20e04efbce80ffa00238eb739f7304e3bf32f38e62f6", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache Optimization Strategies for Scalable and Efficient LLM Inference", "abstract": "The key-value (KV) cache is a foundational optimization in Transformer-based large language models (LLMs), eliminating redundant recomputation of past token representations during autoregressive generation. However, its memory footprint scales linearly with context length, imposing critical bottlenecks on GPU memory capacity, memory bandwidth, and inference throughput as production LLMs push context windows from thousands to millions of tokens. Efficient KV cache management has thus become a first-order challenge for scalable LLM deployment. This paper provides a systematic review of recent KV cache optimization techniques, organizing them into five principal directions: cache eviction, cache compression, hybrid memory solutions, novel attention mechanisms, and combination strategies. For each category we analyze the underlying mechanisms, deployment trade-offs, and empirical performance across memory reduction, throughput, and model accuracy metrics. We further map techniques to seven practical deployment scenarios, including long-context single requests, high-throughput datacenter serving, edge devices, multi-turn conversations, and accuracy-critical reasoning, providing actionable guidance for practitioners selecting among competing approaches. Our analysis reveals that no single technique dominates across all settings; instead, the optimal strategy depends on context length, hardware constraints, and workload characteristics, pointing toward adaptive, multi-stage optimization pipelines as a promising direction for future research.", "authors": ["Yichun Xu", "Navjot K. Khaira", "Tejinder Singh"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-20", "url": "https://arxiv.org/abs/2603.20397", "pdf_url": "https://arxiv.org/pdf/2603.20397v1", "arxiv_id": "2603.20397", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3296}
{"id": "29d2bec40fe3070d962991082a5d3e20871d93a9cc36aac85d2a82a41a4adc8d", "sources": ["arxiv", "semantic_scholar"], "title": "The Residual Stream Is All You Need: On the Redundancy of the KV Cache in Transformer Inference", "abstract": "The key-value (KV) cache is widely treated as essential state in transformer inference, and a large body of work engineers policies to compress, evict, or approximate its entries. We prove that this state is entirely redundant: keys and values at every layer are deterministic projections of the residual stream, and recomputing them from a single residual vector per token incurs exactly zero reconstruction error, not approximately, but bit-identically. We verify this across six models from four architecture families (135M to 4B parameters). Cross-task residual patching at every layer produces D_KL = 0 between patched and original output distributions, confirming that the residual stream satisfies a Markov property and is the sole information-carrying state. Removing the cache entirely and recomputing from scratch yields token-identical output under greedy decoding on all models tested. We build on this result with KV-Direct, a bounded-memory inference scheme that checkpoints residual vectors (5 KB per token on Gemma 3-4B) instead of full KV pairs (136 KB), recomputing keys and values on demand. Over 20 conversation turns, KV-Direct holds peak memory at 42 MB while the standard cache grows past 103 MB. Against five eviction baselines (H2O, StreamingLLM, SnapKV, TOVA, window-only), KV-Direct maintains 100% token match at every cache budget; all baselines degrade to 5-28%. A per-operation latency analysis shows recomputation runs up to 5x faster than reading cached tensors at moderate batch sizes. Code is available at https://github.com/Kaleemullahqasim/KV-Direct.", "authors": ["Kaleem Ullah Qasim", "Jiashu Zhang", "Muhammad Kafeel Shaheen", "Razan Alharith", "Heying Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-20", "url": "https://arxiv.org/abs/2603.19664", "pdf_url": "https://arxiv.org/pdf/2603.19664v1", "arxiv_id": "2603.19664", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Kaleemullahqasim/KV-Direct", "venue": null, "quality_score": 0.6121}
{"id": "184cb8a2cfcc3629fd3cdbf524eea69b8b4bb7fad1d4eff7095a5288de5fa696", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts with Soft Nearest Neighbor Loss: Resolving Expert Collapse via Representation Disentanglement", "abstract": "The Mixture-of-Experts (MoE) model uses a set of expert networks that specialize on subsets of a dataset under the supervision of a gating network. A common issue in MoE architectures is ``expert collapse'' where overlapping class boundaries in the raw input feature space cause multiple experts to learn redundant representations, thus forcing the gating network into rigid routing to compensate. We propose an enhanced MoE architecture that utilizes a feature extractor network optimized using Soft Nearest Neighbor Loss (SNNL) prior to feeding input features to the gating and expert networks. By pre-conditioning the latent space to minimize distances among class-similar data points, we resolve structural expert collapse which results to experts learning highly orthogonal weights. We employ Expert Specialization Entropy and Pairwise Embedding Similarity to quantify this dynamic. We evaluate our experimental approach across four benchmark image classification datasets (MNIST, FashionMNIST, CIFAR10, and CIFAR100), and we show our SNNL-augmented MoE models demonstrate structurally diverse experts which allow the gating network to adopt a more flexible routing strategy. This paradigm significantly improves classification accuracy on the FashionMNIST, CIFAR10, and CIFAR100 datasets.", "authors": ["Abien Fred Agarap", "Arnulfo P. Azcarraga"], "categories": ["cs.NE", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-20", "url": "https://arxiv.org/abs/2603.26734", "pdf_url": "https://arxiv.org/pdf/2603.26734v1", "arxiv_id": "2603.26734", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3296}
{"id": "505e1a8e965845c6e4a5ceb2d5777a1e8a7d9508dda4192e4697f2fdd5caadba", "sources": ["arxiv", "semantic_scholar"], "title": "EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models", "abstract": "Diffusion-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the $k$ most recently decoded tokens. The skip-or-recompute decision requires only $O(V)$ computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves $15.2\\times$-$26.4\\times$ speedup on standard benchmarks and $22.4\\times$-$24.1\\times$ on chain-of-thought benchmarks, with competitive accuracy and decision overhead accounting for only $0.5\\%$ of inference time. Code is available at https://github.com/mscheong01/EntropyCache.", "authors": ["Minsoo Cheong", "Donghyun Son", "Woosang Lim", "Sungjoo Yoo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-19", "url": "https://arxiv.org/abs/2603.18489", "pdf_url": "https://arxiv.org/pdf/2603.18489v1", "arxiv_id": "2603.18489", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/mscheong01/EntropyCache", "venue": null, "quality_score": 0.6107}
{"id": "c0380f0d11c31a45214ba7131ec6637859c727b80e9233598d9d881bfd15d4ec", "sources": ["arxiv", "semantic_scholar"], "title": "Prune-then-Quantize or Quantize-then-Prune? Understanding the Impact of Compression Order in Joint Model Compression", "abstract": "What happens when multiple compression methods are combined-does the order in which they are applied matter? Joint model compression has emerged as a powerful strategy to achieve higher efficiency by combining multiple methods such as pruning and quantization. A central but underexplored factor in joint model compression is the compression order, or the sequence of different methods within the compression pipeline. Most prior studies have either sidestepped the issue by assuming orthogonality between techniques, while a few have examined them only in highly constrained cases. Consequently, the broader role of compression order in shaping model performance remains poorly understood. In this paper, we address the overlooked problem of compression order and provide both theoretical and empirical analysis. We formulate the problem of optimizing the compression order and introduce the Progressive Intensity Hypothesis, which states that weaker perturbations should precede stronger ones. We provide theoretical guarantees showing that the relative benefit of one order increases with the underlying performance gap. Extensive experiments on both language and vision models validate the hypothesis, and further show its generality to broader setups such as multi-stage compression and mixed-precision quantization.", "authors": ["Minjun Kim", "Jaehyeon Choi", "Hyunwoo Yang", "Jongjin Kim", "Jinho Song", "U Kang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-19", "url": "https://arxiv.org/abs/2603.18426", "pdf_url": "https://arxiv.org/pdf/2603.18426v1", "arxiv_id": "2603.18426", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3289}
{"id": "2bafc89e70d327ebae6bd46cce96bc73e1cc5b86ca09c6bfd3703d252c892d1e", "sources": ["arxiv", "semantic_scholar"], "title": "AIMER: Calibration-Free Task-Agnostic MoE Pruning", "abstract": "Mixture-of-Experts (MoE) language models increase parameter capacity without proportional per-token compute, but the deployment still requires storing all experts, making expert pruning important for reducing memory and serving overhead. Existing task-agnostic expert pruning methods are typically calibration-dependent: they estimate expert importance from routing or activation statistics on a calibration set, which makes pruning outcomes sensitive to the choice of calibration set and adds substantial preprocessing cost. We introduce AIMER (\\textbf{A}bsolute mean over root mean square \\textbf{IM}portance for \\textbf{E}xpert \\textbf{R}anking), a simple calibration-free criterion that yields clear within-layer score separation and distinct expert stratification. Across 7B to 30B MoE language models at 25\\% and 50\\% pruning ratios over 16 benchmarks, AIMER consistently delivers competitive or stronger overall performance against state-of-the-art calibration-based expert pruning baselines with only 0.22--1.27 seconds for scoring the experts.", "authors": ["Zongfang Liu", "Shengkun Tang", "Yifan Shen", "Huan Wang", "Xin Yuan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-19", "url": "https://arxiv.org/abs/2603.18492", "pdf_url": "https://arxiv.org/pdf/2603.18492v2", "arxiv_id": "2603.18492", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3289}
{"id": "d5826824fb9625f03c2bc48b0891d0a16ff27cae15ceb86b1b9e3a7e0dc3eb44", "sources": ["arxiv", "semantic_scholar"], "title": "DyMoE: Dynamic Expert Orchestration with Mixed-Precision Quantization for Efficient MoE Inference on Edge", "abstract": "Despite the computational efficiency of MoE models, the excessive memory footprint and I/O overhead inherent in multi-expert architectures pose formidable challenges for real-time inference on resource-constrained edge platforms. While existing static methods struggle with a rigid latency-accuracy trade-off, we observe that expert importance is highly skewed and depth-dependent. Motivated by these insights, we propose DyMoE, a dynamic mixed-precision quantization framework designed for high-performance edge inference. Leveraging insights into expert importance skewness and depth-dependent sensitivity, DyMoE introduces: (1) importance-aware prioritization to dynamically quantize experts at runtime; (2) depth-adaptive scheduling to preserve semantic integrity in critical layers; and (3) look-ahead prefetching to overlap I/O stalls. Experimental results on commercial edge hardware show that DyMoE reduces Time-to-First-Token (TTFT) by 3.44x-22.7x and up to a 14.58x speedup in Time-Per-Output-Token (TPOT) compared to state-of-the-art offloading baselines, enabling real-time, accuracy-preserving MoE inference on resource-constrained edge devices.", "authors": ["Yuegui Huang", "Zhiyuan Fang", "Weiqi Luo", "Ruoyu Wu", "Wuhui Chen", "Zibin Zheng"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-19", "url": "https://arxiv.org/abs/2603.19172", "pdf_url": "https://arxiv.org/pdf/2603.19172v1", "arxiv_id": "2603.19172", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3289}
{"id": "edecef9494ffc7ff8a6f1b3e9f2a2e42b0cbd6fdf33f919d51a66512ad4b170e", "sources": ["arxiv", "semantic_scholar"], "title": "Path-Constrained Mixture-of-Experts", "abstract": "Sparse Mixture-of-Experts (MoE) architectures route each token through a subset of experts at each layer independently. We propose viewing MoE computation through the lens of \\emph{expert paths} -- the sequence of expert selections a token makes across all layers. This perspective reveals that, despite $N^L$ possible paths for $N$ experts across $L$ layers, tokens in practice cluster into a small fraction of paths that align with linguistic function, yet the vast majority of paths remain unexplored, representing a statistical inefficiency. This motivates architectures that constrain the effective path space to amplify this natural concentration. As one instantiation, we introduce \\pathmoe{}, which shares router parameters across blocks of consecutive layers. Analysis confirms that \\pathmoe{} amplifies the emergent path structure: it produces more concentrated path clusters, better cross-layer consistency, and greater robustness to routing perturbations. Experiments on 0.9B and 16B parameter \\pathmoe{} models demonstrate consistent improvements on perplexity and downstream tasks over independent routing, while eliminating the need for auxiliary losses. These results establish expert paths as a useful design axis for MoE architectures, complementary to existing work on independent routing mechanisms.", "authors": ["Zijin Gu", "Tatiana Likhomanenko", "Vimal Thilak", "Jason Ramapuram", "Navdeep Jaitly"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-18", "url": "https://arxiv.org/abs/2603.18297", "pdf_url": "https://arxiv.org/pdf/2603.18297v2", "arxiv_id": "2603.18297", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3281}
{"id": "b7ba6ad5fa85713fd7297c51c54422d952690b12f87badb8ec451851174c9f93", "sources": ["arxiv", "semantic_scholar"], "title": "FEMBA on the Edge: Physiologically-Aware Pre-Training, Quantization, and Deployment of a Bidirectional Mamba EEG Foundation Model on an Ultra-low Power Microcontroller", "abstract": "Objective: To enable continuous, long-term neuro-monitoring on wearable devices by overcoming the computational bottlenecks of Transformer-based Electroencephalography (EEG) foundation models and the quantization challenges inherent to State-Space Models (SSMs). Methods: We present FEMBA, a bidirectional Mamba architecture pre-trained on over 21,000 hours of EEG. We introduce a novel Physiologically-Aware pre-training objective, consisting of a reconstruction with low-pass filtering, to prioritize neural oscillations over high-frequency artifacts. To address the activation outliers common in SSMs, we employ Quantization-Aware Training (QAT) to compress the model to 2-bit weights. The framework is deployed on a parallel ultra-low-power RISC-V microcontroller (GAP9) using a custom double-buffered memory streaming scheme. Results: The proposed low-pass pre-training improves downstream AUROC on TUAB from 0.863 to 0.893 and AUPR from 0.862 to 0.898 compared to the best contrastive baseline. QAT successfully compresses weights with negligible performance loss, whereas standard post-training quantization degrades accuracy by approximately \\textbf{30\\%}. The embedded implementation achieves deterministic real-time inference (\\textbf{1.70~s} per 5~s window) and reduces the memory footprint by \\textbf{74\\%} (to $\\approx$2~MB), achieving competitive accuracy with up to \\textbf{27$\\times$} fewer FLOPs than Transformer benchmarks. Conclusion: FEMBA demonstrates that Mamba-based foundation models can be effectively quantized and deployed on extreme-edge hardware without sacrificing the representation quality required for robust clinical analysis. Significance: This work establishes the first full-stack framework for deploying large-scale EEG foundation models on ultra-low-power wearables, facilitating continuous, SSM based monitoring for epilepsy and sleep disorders.", "authors": ["Anna Tegon", "Nicholas Lehmann", "Yawei Li", "Andrea Cossettini", "Luca Benini", "Thorir Mar Ingolfsson"], "categories": ["eess.SP", "cs.LG"], "fields_of_study": ["Medicine", "Engineering", "Computer Science"], "published_date": "2026-03-18", "url": "https://arxiv.org/abs/2603.26716", "pdf_url": "https://arxiv.org/pdf/2603.26716v1", "arxiv_id": "2603.26716", "doi": "10.1109/tbme.2026.3683482", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3281}
{"id": "7ecf1ba867b0007e1516b3029303e6402d54604bbb40f1d38c557828acdcbc6e", "sources": ["arxiv", "semantic_scholar"], "title": "VQKV: High-Fidelity and High-Ratio Cache Compression via Vector-Quantization", "abstract": "The growing context length of Large Language Models (LLMs) enlarges the Key-Value (KV) cache, limiting deployment in resource-limited environments. Prior training-free approaches for KV cache compression typically rely on low-rank approximation or scalar quantization, which fail to simultaneously achieve high compression ratios and high reconstruction fidelity. We propose VQKV, a novel, training-free method introducing vector quantization (VQ) to obtain highly compressed KV representations while preserving high model fidelity, allowing for the representation of thousands of floating-point values with just a few integer indices. As a result, VQKV achieves an 82.8\\% compression ratio on LLaMA3.1-8B while retaining 98.6\\% of the baseline performance on LongBench and enabling 4.3x longer generation length on the same memory footprint.", "authors": ["Yixuan Wang", "Qingyu Shi", "Jiayu Zhou", "Dianbo Liu", "Ziwei He", "Zhouhan Lin"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-17", "url": "https://arxiv.org/abs/2603.16435", "pdf_url": "https://arxiv.org/pdf/2603.16435v1", "arxiv_id": "2603.16435", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3274}
{"id": "87ad767d85ab2cba95f89ccd152b39026819399144b726e289237fb7f09697ef", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Style Experts for Diverse Image Stylization", "abstract": "Diffusion-based stylization has advanced significantly, yet existing methods are limited to color-driven transformations, neglecting complex semantics and material details. We introduce StyleExpert, a semantic-aware framework based on the Mixture of Experts (MoE). Our framework employs a unified style encoder, trained on our large-scale dataset of content-style-stylized triplets, to embed diverse styles into a consistent latent space. This embedding is then used to condition a similarity-aware gating mechanism, which dynamically routes styles to specialized experts within the MoE architecture. Leveraging this MoE architecture, our method adeptly handles diverse styles spanning multiple semantic levels, from shallow textures to deep semantics. Extensive experiments show that StyleExpert outperforms existing approaches in preserving semantics and material details, while generalizing to unseen styles. Our code and collected images are available at the project page: https://hh-lg.github.io/StyleExpert-Page/.", "authors": ["Shihao Zhu", "Ziheng Ouyang", "Yijia Kang", "Qilong Wang", "Mi Zhou", "Bo Li", "Ming-Ming Cheng", "Qibin Hou"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-17", "url": "https://arxiv.org/abs/2603.16649", "pdf_url": "https://arxiv.org/pdf/2603.16649v3", "arxiv_id": "2603.16649", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3274}
{"id": "bbc2f2a580c4694e729e3902134e7945e870c799bbea61f7b4184b4627cd2368", "sources": ["arxiv", "semantic_scholar"], "title": "Frequency Matters: Fast Model-Agnostic Data Curation for Pruning and Quantization", "abstract": "Post-training model compression is essential for enhancing the portability of Large Language Models (LLMs) while preserving their performance. While several compression approaches have been proposed, less emphasis has been placed on selecting the most suitable set of data (the so-called \\emph{calibration data}) for finding the compressed model configuration. The choice of calibration data is a critical step in preserving model capabilities both intra- and inter-tasks. In this work, we address the challenge of identifying high-performance calibration sets for both pruning and quantization by analyzing intrinsic data properties rather than model-specific signals. We introduce \\texttt{\\textbf{ZipCal}}, a model-agnostic data curation strategy that maximizes lexical diversity based on Zipfian power laws. Experiments demonstrate that our method consistently outperforms standard uniform random sampling across various pruning benchmarks. Notably, it also performs on par, in terms of downstream performance, with a state-of-the-art method that relies on model perplexity. The latter becomes prohibitively expensive at large-scale models and datasets, while \\texttt{\\textbf{ZipCal}} is on average $\\sim$240$\\times$ faster due to its tractable linear complexity\\footnote{We make the code and the experiments available at https://github.com/FrancescoMonaco/ZipCal.}.", "authors": ["Francesco Pio Monaco", "Elia Cunegatti", "Flavio Vella", "Giovanni Iacca"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-17", "url": "https://arxiv.org/abs/2603.16105", "pdf_url": "https://arxiv.org/pdf/2603.16105v3", "arxiv_id": "2603.16105", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/FrancescoMonaco/ZipCal.}", "venue": null, "quality_score": 0.608}
{"id": "73a6c57ab61a3d6802ae33462d4ba4d3c57ae63152da5fcbb770245e3be57e25", "sources": ["arxiv", "semantic_scholar"], "title": "Mostly Text, Smart Visuals: Asymmetric Text-Visual Pruning for Large Vision-Language Models", "abstract": "Network pruning is an effective technique for enabling lightweight Large Vision-Language Models (LVLMs), which primarily incorporates both weights and activations into the importance metric. However, existing efforts typically process calibration data from different modalities in a unified manner, overlooking modality-specific behaviors. This raises a critical challenge: how to address the divergent behaviors of textual and visual tokens for accurate pruning of LVLMs. To this end, we systematically investigate the sensitivity of visual and textual tokens to the pruning operation by decoupling their corresponding weights, revealing that: (i) the textual pathway should be calibrated via text tokens, since it exhibits higher sensitivity than the visual pathway; (ii) the visual pathway exhibits high redundancy, permitting even 50% sparsity. Motivated by these insights, we propose a simple yet effective Asymmetric Text-Visual Weight Pruning method for LVLMs, dubbed ATV-Pruning, which establishes the importance metric for accurate weight pruning by selecting the informative tokens from both textual and visual pathways. Specifically, ATV-Pruning integrates two primary innovations: first, a calibration pool is adaptively constructed by drawing on all textual tokens and a subset of visual tokens; second, we devise a layer-adaptive selection strategy to yield important visual tokens. Finally, extensive experiments across standard multimodal benchmarks verify the superiority of our ATV-Pruning over state-of-the-art methods.", "authors": ["Sijie Li", "Biao Qian", "Jungong Han"], "categories": ["cs.CV", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-16", "url": "https://arxiv.org/abs/2603.16001", "pdf_url": "https://arxiv.org/pdf/2603.16001v1", "arxiv_id": "2603.16001", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/LezJ/ATV-Pruning", "venue": null, "quality_score": 0.6067}
{"id": "943a64d8001b65702238b54306a6c7b8e336656930e75411ac05ae1dc913342c", "sources": ["arxiv", "semantic_scholar"], "title": "Bridging Local and Global Knowledge: Cascaded Mixture-of-Experts Learning for Near-Shortest Path Routing", "abstract": "While deep learning models that leverage local features have demonstrated significant potential for near-optimal routing in dense Euclidean graphs, they struggle to generalize well in sparse networks where topological irregularities require broader structural awareness. To address this limitation, we train a Cascaded Mixture of Experts (Ca-MoE) to solve the all-pairs near-shortest path (APNSP) routing problem. Our Ca-MoE is a modular two-tier architecture that supports the decision-making for forwarder selection with lower-tier experts relying on local features and upper-tier experts relying on global features. It performs adaptive inference wherein the upper-tier experts are triggered only when the lower-tier ones do not suffice to achieve adequate decision quality. Computational efficiency is thus achieved by escalating model capacity only when necessitated by topological complexity, and parameter redundancy is avoided. Furthermore, we incorporate an online meta-learning strategy that facilitates independent expert fine-tuning and utilizes a stability-focused update mechanism to prevent catastrophic forgetting as new graph environments are encountered. Experimental evaluations demonstrate that Ca-MoE routing improves accuracy by up to 29.1% in sparse networks compared to single-expert baselines and maintains performance within 1%-6% of the theoretical upper bound across diverse graph densities.", "authors": ["Yung-Fu Chen", "Anish Arora"], "categories": ["cs.LG", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-16", "url": "https://arxiv.org/abs/2603.15541", "pdf_url": "https://arxiv.org/pdf/2603.15541v1", "arxiv_id": "2603.15541", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3267}
{"id": "4fc59cfa68ba8cce5c5e3f44cdf82e25a4ae633fd0a96c9f95dd611ef236dfba", "sources": ["arxiv", "semantic_scholar"], "title": "ASAP: Attention-Shift-Aware Pruning for Efficient LVLM Inference", "abstract": "While Large Vision-Language Models (LVLMs) demonstrate exceptional multi-modal capabilities, the quadratic computational cost of processing high-resolution visual tokens remains a critical bottleneck. Though recent token reduction strategies attempt to accelerate inference, such methods inadequately exploit attention values and fail to address token redundancy. More critically, they overlook the ``attention shift'' phenomenon inherent in LVLMs, which skews token attention scores. In this work, we propose ASAP, a novel training-free, KV-Cache-compatible pruning recipe that comprehensively addresses these limitations. First, we mitigate the attention shift by utilizing a dynamic bidirectional soft attention mask, ensuring the selection of genuinely informative tokens rather than naive attention-based selection. Second, we posit that high semantic redundancy within the token set degrades performance. We therefore introduce a weighted soft merging component that merges semantically similar tokens, preserving only the most feature-dense visual patches for subsequent layers. ASAP achieves virtually lossless compression of visual context, retaining 99.02% of the original LLaVA-NeXT-7B performance while aggressively slashing computational FLOPs by ~80%.", "authors": ["Surendra Pathak", "Bo Han"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-15", "url": "https://arxiv.org/abs/2603.14549", "pdf_url": "https://arxiv.org/pdf/2603.14549v2", "arxiv_id": "2603.14549", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3259}
{"id": "0a71a7ce56e0e696fe4c5519909e4f79843a225ea39dac30b0e23d3314d40ab1", "sources": ["arxiv", "semantic_scholar"], "title": "OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism", "abstract": "Embodied AI agents increasingly require parallel execution of multiple tasks, such as manipulation, conversation, and memory construction, from shared observations under distinct time constraints. Recent Mixture-of-Transformers (MoT) Vision-Language-Action Models (VLAs) architecturally support such heterogeneous outputs, yet existing inference systems fail to achieve efficient multi-task parallelism for on-device deployment because of redundant computation and resource contention. We identify isolated KV cache management as the root cause. To address this, we propose unified KV cache management, an inference design that treats the KV cache as a first-class shared resource across tasks and over time. This abstraction enables two key optimizations: cross-task KV sharing eliminates redundant prefill of shared observations, while cross-frame continuous batching decouples variable-length language decoding from fixed-rate action generation across control cycles. We implement this design for $π_{0.5}$, a popular MoT VLA, and evaluate it on both NVIDIA GeForce RTX 4090 and Jetson AGX Thor, two representative platforms for on-device VLA inference. OxyGen achieves up to 3.7$\\times$ speedup over isolated execution, delivering over 200 tokens/s language throughput and 70 Hz action frequency simultaneously without degrading action quality, and we further validate the gains on a real humanoid robot with on-board Jetson AGX Thor.", "authors": ["Xiangyu Li", "Huaizhi Tang", "Xin Ding", "Weijun Wang", "Ting Cao", "Yunxin Liu"], "categories": ["cs.RO", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-15", "url": "https://arxiv.org/abs/2603.14371", "pdf_url": "https://arxiv.org/pdf/2603.14371v2", "arxiv_id": "2603.14371", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3259}
{"id": "c1302e1af1fe9142aad81a012d3adea61c5891458cc967ac2f3d8f1ae9876018", "sources": ["arxiv", "semantic_scholar"], "title": "Flood Risk Follows Valleys, Not Grids: Graph Neural Networks for Flash Flood Susceptibility Mapping in Himachal Pradesh with Conformal Uncertainty Quantification", "abstract": "Flash floods are the most destructive natural hazard in Himachal Pradesh (HP), India, causing over 400 fatalities and $1.2 billion in losses in the 2023 monsoon season alone. Existing risk maps treat every pixel independently, ignoring the basic fact that flooding upstream raises risk downstream. We address this with a Graph Neural Network (GraphSAGE) trained on a watershed connectivity graph (460 sub-watersheds, 1,700 directed edges), built from a six-year Sentinel-1 SAR flood inventory (2018-2023, 3,000 events) and 12 environmental variables at 30 m resolution. Four pixel-based ML models (RF, XGBoost, LightGBM, stacking ensemble) serve as baselines. All models are evaluated with leave-one-basin-out spatial cross-validation to avoid the 5-15% AUC inflation of random splits. Conformal prediction produces the first HP susceptibility maps with statistically guaranteed 90% coverage intervals. The GNN achieved AUC = 0.978 +/- 0.017, outperforming the best baseline (AUC = 0.881) and the published HP benchmark (AUC = 0.88). The +0.097 gain confirms that river connectivity carries predictive signal that pixel-based models miss. High-susceptibility zones overlap 1,457 km of highways (including 217 km of the Manali-Leh corridor), 2,759 bridges, and 4 major hydroelectric installations. Conformal intervals achieved 82.9% empirical coverage on the held-out 2023 test set; lower coverage in high-risk zones (45-59%) points to SAR label noise as a target for future work.", "authors": ["Paras Sharma", "Swastika Sharma"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-15", "url": "https://arxiv.org/abs/2603.15681", "pdf_url": "https://arxiv.org/pdf/2603.15681v1", "arxiv_id": "2603.15681", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Parassharmaa/flash-flood-zones-hp", "venue": null, "quality_score": 0.6053}
{"id": "c75f50d846f9cb1c4b64eab5fa7875ca11c39204effdbbf0d0f2325a0f6ae2cb", "sources": ["arxiv", "semantic_scholar"], "title": "SemantiCache: Efficient KV Cache Compression via Semantic Chunking and Clustered Merging", "abstract": "Existing KV cache compression methods generally operate on discrete tokens or non-semantic chunks. However, such approaches often lead to semantic fragmentation, where linguistically coherent units are disrupted, causing irreversible information loss and degradation in model performance. To address this, we introduce SemantiCache, a novel compression framework that preserves semantic integrity by aligning the compression process with the semantic hierarchical nature of language. Specifically, we first partition the cache into semantically coherent chunks by delimiters, which are natural semantic boundaries. Within each chunk, we introduce a computationally efficient Greedy Seed-Based Clustering (GSC) algorithm to group tokens into semantic clusters. These clusters are further merged into semantic cores, enhanced by a Proportional Attention mechanism that rebalances the reduced attention contributions of the merged tokens. Extensive experiments across diverse benchmarks and models demonstrate that SemantiCache accelerates the decoding stage of inference by up to 2.61 times and substantially reduces memory footprint, while maintaining performance comparable to the original model.", "authors": ["Shunlong Wu", "Hai Lin", "Shaoshen Chen", "Tingwei Lu", "Yongqin Zeng", "Shaoxiong Zhan", "Hai-Tao Zheng", "Hong-Gee Kim"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-15", "url": "https://arxiv.org/abs/2603.14303", "pdf_url": "https://arxiv.org/pdf/2603.14303v1", "arxiv_id": "2603.14303", "doi": "10.1109/icassp55912.2026.11464823", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.5122}
{"id": "8a35ce815134b0eb7c9379640f1b13c463bdea33e44353a7b9292700326225fd", "sources": ["arxiv", "semantic_scholar"], "title": "LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing", "abstract": "Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.", "authors": ["Jiawei Hao", "Zhiwei Hao", "Jianyuan Guo", "Li Shen", "Yong Luo", "Han Hu", "Dan Zeng"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-13", "url": "https://arxiv.org/abs/2603.12645", "pdf_url": "https://arxiv.org/pdf/2603.12645v1", "arxiv_id": "2603.12645", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3245}
{"id": "c5681253aeca3c3526727669d7ccd97b59709a5a19dde3ca709154d43b176ad1", "sources": ["arxiv", "semantic_scholar"], "title": "MoEKD: Mixture-of-Experts Knowledge Distillation for Robust and High-Performing Compressed Code Models", "abstract": "Large language models for code have achieved strong performance across diverse software analytics tasks, yet their real-world adoption remains limited by high computational demands, slow inference speeds, significant energy consumption, and environmental impact. Knowledge distillation (KD) offers a practical solution by transferring knowledge from a large model to a smaller and more efficient model. Despite its effectiveness, recent studies show that models distilled from a single source often exhibit degraded adversarial robustness, even when robustness-aware distillation techniques are employed. These observations suggest a fundamental limitation of single-source distillation in simultaneously transferring high-quality and robust knowledge. To overcome this limitation, we propose Mixture of Experts Knowledge Distillation (MoEKD), a KD framework that leverages a Mixture of Experts (MoE) architecture to enable more effective and robust knowledge transfer from multiple specialized experts into a compact model. MoEKD decomposes the distillation process into expert and router training, aggregation of expert knowledge through a learned routing mechanism, and distillation from the aggregated knowledge. We evaluate MoEKD on the vulnerability detection task using CodeBERT and GraphCodeBERT models. Experimental results show that MoEKD not only improves adversarial robustness by up to 35.8%, but also enhances predictive performance by up to 13%, compared to state-of-the-art KD baselines, including Compressor and AVATAR. Furthermore, an ablation study demonstrates that aggregating expert knowledge enables ultra-compact models to maintain competitive performance even when their size is reduced by approximately half. Overall, these results highlight the effectiveness of multi-expert knowledge aggregation in addressing key limitations of existing single-source KD approaches.", "authors": ["Md. Abdul Awal", "Mrigank Rochan", "Chanchal K. Roy"], "categories": ["cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-13", "url": "https://arxiv.org/abs/2603.13213", "pdf_url": "https://arxiv.org/pdf/2603.13213v1", "arxiv_id": "2603.13213", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3245}
{"id": "62afd323071f0d51198d4e800fa9cd40bb245d855841b80f2031e4e146287b86", "sources": ["arxiv", "semantic_scholar"], "title": "LongFlow: Efficient KV Cache Compression for Reasoning Models", "abstract": "Recent reasoning models such as OpenAI-o1 and DeepSeek-R1 have shown strong performance on complex tasks including mathematical reasoning and code generation. However, this performance gain comes with substantially longer output sequences, leading to significantly increased deployment costs. In particular, long outputs require large KV caches, resulting in high memory consumption and severe bandwidth pressure during attention computation. Most existing KV cache optimization methods are designed for long-input, short-output scenarios and are ineffective for the long-output setting of reasoning models. Moreover, importance estimation in prior work is computationally expensive and becomes prohibitive when continuous re-evaluation is required during long generation. To address these challenges, we propose LongFlow, a KV cache compression method with an efficient importance estimation metric derived from an intermediate result of attention computation using only the current query. This design introduces negligible computational overhead and requires no auxiliary storage. We further develop a custom kernel that fuses FlashAttention, importance estimation, and token eviction into a single optimized operator, improving system-level efficiency. Experiments show that LongFlow achieves up to an 11.8 times throughput improvement with 80% KV cache compression with minimal impact on model accuracy.", "authors": ["Yi Su", "Zhenxu Tian", "Dan Qiao", "Yuechi Zhou", "Juntao Li", "Min Zhang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-12", "url": "https://arxiv.org/abs/2603.11504", "pdf_url": "https://arxiv.org/pdf/2603.11504v2", "arxiv_id": "2603.11504", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3237}
{"id": "6aff84c78c181251f6eeeb77ad0b6b4ece78cbc5be7da07d040fa99d080eedad", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerating Suffix Jailbreak attacks with Prefix-Shared KV-cache", "abstract": "Suffix jailbreak attacks serve as a systematic method for red-teaming Large Language Models (LLMs) but suffer from prohibitive computational costs, as a large number of candidate suffixes need to be evaluated before identifying a jailbreak suffix. This paper presents Prefix-Shared KV Cache (PSKV), a plug-and-play inference optimization technique tailored for jailbreak suffix generation. Our method is motivated by a key observation that when performing suffix jailbreaking, while a large number of candidate prompts need to be evaluated, they share the same targeted harmful instruction as the prefix. Therefore, instead of performing redundant inference on the duplicated prefix, PSKV maintains a single KV cache for this prefix and shares it with every candidate prompt, enabling the parallel inference of diverse suffixes with minimal memory overhead. This design enables more aggressive batching strategies that would otherwise be limited by memory constraints. Extensive experiments on six widely used suffix attacks across five widely deployed LLMs demonstrate that PSKV reduces inference time by 40\\% and peak memory usage by 50\\%, while maintaining the original Attack Success Rate (ASR). The code has been submitted and will be released publicly.", "authors": ["Xinhai Wang", "Shaopeng Fu", "Shu Yang", "Liangyu Wang", "Tianhang Zheng", "Di Wang"], "categories": ["cs.CR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-12", "url": "https://arxiv.org/abs/2603.13420", "pdf_url": "https://arxiv.org/pdf/2603.13420v2", "arxiv_id": "2603.13420", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3237}
{"id": "5408c2ea0d17de6c9257d3dbc5f23288f22735208c6ff399b70106d19995cfff", "sources": ["arxiv", "semantic_scholar"], "title": "Where Matters More Than What: Decoding-aligned KV Cache Compression via Position-aware Pseudo Queries", "abstract": "The Key-Value (KV) cache is crucial for efficient Large Language Models (LLMs) inference, but excessively long contexts drastically increase KV cache memory footprint. Existing KV cache compression methods typically rely on input-side attention patterns within a prompt observation window to estimate token importance during the prefill stage. They fail to preserve critical tokens for future generation since these assessments are not derived from the decoding process. Intuitively, an effective observation window should mirror the decoding-stage queries to accurately reflect which tokens the generation process will attend to. However, ground-truth decoding queries are inherently unavailable during inference. For constructing pseudo queries to approximate them, we find that positional information plays a more critical role than semantic content. Motivated by this insight, we propose decoding-aligned KV cache compression via position-aware pseudo queries (DapQ), a novel and lightweight eviction framework that leverages position-aware pseudo queries to simulate the output tokens, thereby establishing an effective observation window for importance assessment. It aligns closely with the actual generation context and enables precise token eviction. Extensive evaluations across multiple benchmarks and LLMs demonstrate that DapQ achieves superior performance, particularly under strict memory constraints (e.g., up to nearly lossless performance 99.5% on NIAH with 3% KV cache budgets).", "authors": ["Zhenxu Tian", "Yi Su", "Juntao Li", "Min Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-12", "url": "https://arxiv.org/abs/2603.11564", "pdf_url": "https://arxiv.org/pdf/2603.11564v1", "arxiv_id": "2603.11564", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3237}
{"id": "bf0f16d8e660f4559def9b285d5c616764e7222fa47ee96e3ccdf52716efda04", "sources": ["arxiv", "semantic_scholar"], "title": "Task-Conditioned Routing Signatures in Sparse Mixture-of-Experts Transformers", "abstract": "Sparse Mixture-of-Experts (MoE) architectures enable efficient scaling of large language models through conditional computation, yet the routing mechanisms responsible for expert selection remain poorly understood. In this work, we introduce routing signatures, a vector representation summarizing expert activation patterns across layers for a given prompt, and use them to study whether MoE routing exhibits task-conditioned structure. Using OLMoE-1B-7B-0125-Instruct as an empirical testbed, we show that prompts from the same task category induce highly similar routing signatures, while prompts from different categories exhibit substantially lower similarity. Within-category routing similarity (0.8435 +/- 0.0879) significantly exceeds across-category similarity (0.6225 +/- 0.1687), corresponding to Cohen's d = 1.44. A logistic regression classifier trained solely on routing signatures achieves 92.5% +/- 6.1% cross-validated accuracy on four-way task classification. To ensure statistical validity, we introduce permutation and load-balancing baselines and show that the observed separation is not explained by sparsity or balancing constraints alone. We further analyze layer-wise signal strength and low-dimensional projections of routing signatures, finding that task structure becomes increasingly apparent in deeper layers. These results suggest that routing in sparse transformers is not merely a balancing mechanism, but a measurable task-sensitive component of conditional computation. We release MOE-XRAY, a lightweight toolkit for routing telemetry and analysis.", "authors": ["Mynampati Sri Ranganadha Avinash"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-11", "url": "https://arxiv.org/abs/2603.11114", "pdf_url": "https://arxiv.org/pdf/2603.11114v1", "arxiv_id": "2603.11114", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.323}
{"id": "c9b6dafdbad61df80fa18fcf60b0597295f43516d54b8601ce8d673842787915", "sources": ["arxiv", "semantic_scholar"], "title": "Optimal Expert-Attention Allocation in Mixture-of-Experts: A Scalable Law for Dynamic Model Design", "abstract": "This paper presents a novel extension of neural scaling laws to Mixture-of-Experts (MoE) models, focusing on the optimal allocation of compute between expert and attention sub-layers. As MoE architectures have emerged as an efficient method for scaling model capacity without proportionally increasing computation, determining the optimal expert-attention compute ratio becomes critical. We define the ratio $r$ as the fraction of total FLOPs per token dedicated to the expert layers versus the attention layers, and explore how this ratio interacts with the overall compute budget and model sparsity. Through extensive experiments with GPT-style MoE Transformers, we empirically find that the optimal ratio $r^*$ follows a power-law relationship with total compute and varies with sparsity. Our analysis leads to an explicit formula for $r^*$, enabling precise control over the expert-attention compute allocation. We generalize the Chinchilla scaling law by incorporating this architectural parameter, providing a new framework for tuning MoE models beyond size and data. Our findings offer practical guidelines for designing efficient MoE models, optimizing performance while respecting fixed compute budgets.", "authors": ["Junzhuo Li", "Peijie Jiang", "Changxin Tian", "Jia Liu", "Zhiqiang Zhang", "Xuming Hu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-11", "url": "https://arxiv.org/abs/2603.10379", "pdf_url": "https://arxiv.org/pdf/2603.10379v1", "arxiv_id": "2603.10379", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.323}
{"id": "004c0e3045d469626b727ab0354521a734e9eb15fba6a20724f141dc0a0976b0", "sources": ["arxiv", "semantic_scholar"], "title": "LookaheadKV: Fast and Accurate KV Cache Eviction by Glimpsing into the Future without Generation", "abstract": "Transformer-based large language models (LLMs) rely on key-value (KV) caching to avoid redundant computation during autoregressive inference. While this mechanism greatly improves efficiency, the cache size grows linearly with the input sequence length, quickly becoming a bottleneck for long-context tasks. Existing solutions mitigate this problem by evicting prompt KV that are deemed unimportant, guided by estimated importance scores. Notably, a recent line of work proposes to improve eviction quality by \"glimpsing into the future\", in which a draft generator produces a surrogate future response approximating the target model's true response, and this surrogate is subsequently used to estimate the importance of cached KV more accurately. However, these approaches rely on computationally expensive draft generation, which introduces substantial prefilling overhead and limits their practicality in real-world deployment. To address this challenge, we propose LookaheadKV, a lightweight eviction framework that leverages the strength of surrogate future response without requiring explicit draft generation. LookaheadKV augments transformer layers with parameter-efficient modules trained to predict true importance scores with high accuracy. Our design ensures negligible runtime overhead comparable to existing inexpensive heuristics, while achieving accuracy superior to more costly approximation methods. Extensive experiments on long-context understanding benchmarks, across a wide range of models, demonstrate that our method not only outperforms recent competitive baselines in various long-context understanding tasks, but also reduces the eviction cost by up to 14.5x, leading to significantly faster time-to-first-token. Our code is available at https://github.com/SamsungLabs/LookaheadKV.", "authors": ["Jinwoo Ahn", "Ingyu Seong", "Akhil Kedia", "Junhan Kim", "Hyemi Jang", "Kangwook Lee", "Yongkweon Jeon"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-11", "url": "https://arxiv.org/abs/2603.10899", "pdf_url": "https://arxiv.org/pdf/2603.10899v1", "arxiv_id": "2603.10899", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/SamsungLabs/LookaheadKV", "venue": null, "quality_score": 0.5999}
{"id": "d5e1c23d570d2ef80801ffd6403ec81b898e5120d46110b1c3ed283f27679c84", "sources": ["arxiv", "semantic_scholar"], "title": "ReMix: Reinforcement routing for mixtures of LoRAs in LLM finetuning", "abstract": "Low-rank adapters (LoRAs) are a parameter-efficient finetuning technique that injects trainable low-rank matrices into pretrained models to adapt them to new tasks. Mixture-of-LoRAs models expand neural networks efficiently by routing each layer input to a small subset of specialized LoRAs of the layer. Existing Mixture-of-LoRAs routers assign a learned routing weight to each LoRA to enable end-to-end training of the router. Despite their empirical promise, we observe that the routing weights are typically extremely imbalanced across LoRAs in practice, where only one or two LoRAs often dominate the routing weights. This essentially limits the number of effective LoRAs and thus severely hinders the expressive power of existing Mixture-of-LoRAs models. In this work, we attribute this weakness to the nature of learnable routing weights and rethink the fundamental design of the router. To address this critical issue, we propose a new router designed that we call Reinforcement Routing for Mixture-of-LoRAs (ReMix). Our key idea is using non-learnable routing weights to ensure all active LoRAs to be equally effective, with no LoRA dominating the routing weights. However, our routers cannot be trained directly via gradient descent due to our non-learnable routing weights. Hence, we further propose an unbiased gradient estimator for the router by employing the reinforce leave-one-out (RLOO) technique, where we regard the supervision loss as the reward and the router as the policy in reinforcement learning. Our gradient estimator also enables to scale up training compute to boost the predictive performance of our ReMix. Extensive experiments demonstrate that our proposed ReMix significantly outperform state-of-the-art parameter-efficient finetuning methods under a comparable number of activated parameters.", "authors": ["Ruizhong Qiu", "Hanqing Zeng", "Yinglong Xia", "Yiwen Meng", "Ren Chen", "Jiarui Feng", "Dongqi Fu", "Qifan Wang", "Jiayi Liu", "Jun Xiao", "Xiangjun Fan", "Benyu Zhang", "Hong Li", "Zhining Liu", "Hyunsik Yoo", "Zhichen Zeng", "Tianxin Wei", "Hanghang Tong"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-10", "url": "https://arxiv.org/abs/2603.10160", "pdf_url": "https://arxiv.org/pdf/2603.10160v1", "arxiv_id": "2603.10160", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3223}
{"id": "7092ad88ab03de00d64d5214d88498a96c7931d04dfbc270024b3743233d9408", "sources": ["arxiv", "semantic_scholar"], "title": "CORAL: Scalable Multi-Task Robot Learning via LoRA Experts", "abstract": "Deploying Vision-Language-Action (VLA) models in real-world robotics exposes a core multi-task learning challenge: reconciling task interference in multi-task robotic learning. When multiple tasks are jointly fine-tuned in a single stage, gradients from different tasks can conflict, causing negative transfer and reducing per-task performance. Yet maintaining a separate full checkpoint per task is often storage- and deployment-prohibitive. To address this dilemma, we present CORAL, a backbone- and embodiment-agnostic framework designed primarily to mitigate multi-task interference while remaining naturally extensible to a continuous stream of new tasks. CORAL freezes a single pre-trained VLA backbone and attaches one lightweight Low-Rank Adaptation (LoRA) expert per task; at runtime, a dynamic inference engine (the CORAL Manager) routes language instructions to the appropriate expert and swaps experts on the fly with zero inference overhead. This strict parameter isolation avoids complex gating networks and prevents parameter-level cross-task interference by construction; as an added capability, it also enables sequentially introducing new tasks without parameter overwriting caused by catastrophic forgetting. We validate CORAL on a real-world Galaxea R1 dual-arm mobile manipulator and three simulation benchmarks (LIBERO, WidowX, Google Robot), where CORAL overcomes fine-grained instructional ambiguity and substantially outperforms joint training, yielding a practical and scalable system for lifelong multi-task robot learning. Website: https://frontierrobo.github.io/CORAL", "authors": ["Yuankai Luo", "Woping Chen", "Tong Liang", "Zhenguo Li"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-10", "url": "https://arxiv.org/abs/2603.09298", "pdf_url": "https://arxiv.org/pdf/2603.09298v1", "arxiv_id": "2603.09298", "doi": null, "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3223}
{"id": "a906ddc92db315267ecaa21ba80eef8074bc67d74d0d8580cbb8351ae1190927", "sources": ["arxiv", "semantic_scholar"], "title": "ForgeDreamer: Industrial Text-to-3D Generation with Multi-Expert LoRA and Cross-View Hypergraph", "abstract": "Current text-to-3D generation methods excel in natural scenes but struggle with industrial applications due to two critical limitations: domain adaptation challenges where conventional LoRA fusion causes knowledge interference across categories, and geometric reasoning deficiencies where pairwise consistency constraints fail to capture higher-order structural dependencies essential for precision manufacturing. We propose a novel framework named ForgeDreamer addressing both challenges through two key innovations. First, we introduce a Multi-Expert LoRA Ensemble mechanism that consolidates multiple category-specific LoRA models into a unified representation, achieving superior cross-category generalization while eliminating knowledge interference. Second, building on enhanced semantic understanding, we develop a Cross-View Hypergraph Geometric Enhancement approach that captures structural dependencies spanning multiple viewpoints simultaneously. These components work synergistically improved semantic understanding, enables more effective geometric reasoning, while hypergraph modeling ensures manufacturing-level consistency. Extensive experiments on a custom industrial dataset demonstrate superior semantic generalization and enhanced geometric fidelity compared to state-of-the-art approaches. Code is available at https://github.com/Junhaocai27/ForgeDreamer", "authors": ["Junhao Cai", "Deyu Zeng", "Junhao Pang", "Lini Li", "Zongze Wu", "Xiaopin Zhong"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-10", "url": "https://arxiv.org/abs/2603.09266", "pdf_url": "https://arxiv.org/pdf/2603.09266v4", "arxiv_id": "2603.09266", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Junhaocai27/ForgeDreamer", "venue": null, "quality_score": 0.5985}
{"id": "abaf72d9dcb8237f170e98af3b8e9308c2e965d7c8234d464348ebbfec8e115b", "sources": ["arxiv", "semantic_scholar"], "title": "Speculating Experts Accelerates Inference for Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) models have gained popularity as a means of scaling the capacity of large language models (LLMs) while maintaining sparse activations and reduced per-token compute. However, in memory-constrained inference settings, expert weights must be offloaded to CPU, creating a performance bottleneck from CPU-GPU transfers during decoding. We propose an expert prefetching scheme that leverages currently computed internal model representations to speculate future experts, enabling memory transfers to overlap with computation. Across multiple MoE architectures, we demonstrate that future experts can be reliably predicted by these internal representations. We also demonstrate that executing speculated experts generally maintains downstream task accuracy, thus preserving more effective compute-memory overlap by eliminating the need to re-fetch true router-selected experts. Integrated into an optimized inference engine, our approach achieves up to 14\\% reduction in time per output token (TPOT) over on-demand loading of experts from CPU memory. For MoEs where speculative execution alone yields suboptimal accuracy, we further examine lightweight estimators that improve expert prediction hit rates, thereby reducing performance degradation. Our code is released in open-source at https://github.com/axonn-ai/yalis/tree/offload_prefetch.", "authors": ["Vivan Madan", "Prajwal Singhania", "Abhinav Bhatele", "Tom Goldstein", "Ashwinee Panda"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-09", "url": "https://arxiv.org/abs/2603.19289", "pdf_url": "https://arxiv.org/pdf/2603.19289v1", "arxiv_id": "2603.19289", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/axonn-ai/yalis/tree/offload_prefetch", "venue": null, "quality_score": 0.5972}
{"id": "0100340ad49fd23fe58b9e0a96b3229672973e90dda2f753b21b72f518da7a91", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling Machine Learning Interatomic Potentials with Mixtures of Experts", "abstract": "Machine Learning Interatomic Potentials (MLIPs) enable accurate large-scale atomistic simulations, yet improving their expressive capacity efficiently remains challenging. Here we systematically develop Mixture-of-Experts (MoE) and Mixture-of-Linear-Experts (MoLE) architectures for MLIPs and analyze the effects of routing strategies and expert designs. We show that sparse activation combined with shared experts yields substantial performance gains, and that nonlinear MoE formulations outperform MoLE when shared experts are present, underscoring the importance of nonlinear expert specialization. Furthermore, element-wise routing consistently surpasses configuration-level routing, while global MoE routing often leads to numerical instability. The resulting element-wise MoE model achieves state-of-the-art accuracy across the OMol25, OMat24, and OC20M benchmarks. Analysis of routing patterns reveals chemically interpretable expert specialization aligned with periodic-table trends, indicating that the model effectively captures element-specific chemical characteristics for precise interatomic modeling.", "authors": ["Yuzhi Liu", "Duo Zhang", "Anyang Peng", "Weinan E", "Linfeng Zhang", "Han Wang"], "categories": ["physics.chem-ph", "cs.LG", "physics.comp-ph"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2026-03-09", "url": "https://arxiv.org/abs/2603.07977", "pdf_url": "https://arxiv.org/pdf/2603.07977v2", "arxiv_id": "2603.07977", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3216}
{"id": "e1f0fe191047317dca18254b9917abfe82eaff9ca8123b38100dbfa97f705e2d", "sources": ["arxiv", "semantic_scholar"], "title": "LAR-MoE: Latent-Aligned Routing for Mixture of Experts in Robotic Imitation Learning", "abstract": "Imitation learning enables robots to acquire manipulation skills from demonstrations, yet deploying a policy across tasks with heterogeneous dynamics remains challenging, as models tend to average over distinct behavioral modes present in the demonstrations. Mixture-of-Experts (MoE) architectures address this by activating specialized subnetworks, but requires meaningful skill decompositions for expert routing. We introduce Latent-Aligned Routing for Mixture of Experts (LAR-MoE), a two-stage framework that decouples unsupervised skill discovery from policy learning. In pre-training, we learn a joint latent representation between observations and future actions through student-teacher co-training. In a post-training stage, the expert routing is regularized to follow the structure of the learned latent space, preventing expert collapse while maintaining parameter efficiency. We evaluate LAR-MoE in simulation and on hardware. On the LIBERO benchmark, our method achieves a 95.2% average success rate with 150M parameters. On a surgical bowel grasping and retraction task, LAR-MoE matches a supervised MoE baseline without requiring any phase annotations, and transfers zero-shot to ex vivo porcine tissue. Our findings suggest that latent-aligned routing provides a principled alternative to supervised skill decomposition, enabling structured expert specialization from unlabeled demonstrations.", "authors": ["Ariel Rodriguez", "Chenpan Li", "Lorenzo Mazza", "Rayan Younis", "Ortrun Hellig", "Sebastian Bodenstedt", "Martin Wagner", "Stefanie Speidel"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-09", "url": "https://arxiv.org/abs/2603.08476", "pdf_url": "https://arxiv.org/pdf/2603.08476v1", "arxiv_id": "2603.08476", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3216}
{"id": "88afee5d3a1760cf70d0adb717a86ab3074224599459a10e9763b9e102cd91ac", "sources": ["arxiv", "semantic_scholar"], "title": "LycheeCluster: Efficient Long-Context Inference with Structure-Aware Chunking and Hierarchical KV Indexing", "abstract": "The quadratic complexity of the attention mechanism and the substantial memory footprint of the Key-Value (KV) cache present severe computational and memory challenges for Large Language Models (LLMs) processing long contexts. Existing retrieval-based methods often compromise semantic integrity through fixed-size chunking and suffer from inefficient linear scanning. In this paper, we propose LycheeCluster, a novel method for efficient KV cache management. LycheeCluster preserves local semantic coherence via boundary-aware chunking and constructs a recursive hierarchical index rooted in the triangle inequality. This design transforms cache retrieval from a linear scan into a theoretically bounded, logarithmic-time pruning process, while a lazy update strategy supports efficient streaming generation. Experiments demonstrate that LycheeCluster achieves up to a 3.6x end-to-end inference speedup with negligible degradation in model performance, outperforming state-of-the-art KV cache management methods (e.g., Quest, ClusterKV). We will release our code and kernels after publication.", "authors": ["Dongfang Li", "Zixuan Liu", "Gang Lin", "Baotian Hu", "Min Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-09", "url": "https://arxiv.org/abs/2603.08453", "pdf_url": "https://arxiv.org/pdf/2603.08453v1", "arxiv_id": "2603.08453", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3216}
{"id": "26d01da58259be37fc075a39a6ebfbf9d538e260923e9fc2abaf78f6d5b3957a", "sources": ["arxiv", "semantic_scholar"], "title": "IMSE: Intrinsic Mixture of Spectral Experts Fine-tuning for Test-Time Adaptation", "abstract": "Test-time adaptation (TTA) has been widely explored to prevent performance degradation when test data differ from the training distribution. However, fully leveraging the rich representations of large pretrained models with minimal parameter updates remains underexplored. In this paper, we propose Intrinsic Mixture of Spectral Experts (IMSE) that leverages the spectral experts inherently embedded in Vision Transformers. We decompose each linear layer via singular value decomposition (SVD) and adapt only the singular values, while keeping the singular vectors fixed. We further identify a key limitation of entropy minimization in TTA: it often induces feature collapse, causing the model to rely on domain-specific features rather than class-discriminative features. To address this, we propose a diversity maximization loss based on expert-input alignment, which encourages diverse utilization of spectral experts during adaptation. In the continual test-time adaptation (CTTA) scenario, beyond preserving pretrained knowledge, it is crucial to retain and reuse knowledge from previously observed domains. We introduce Domain-Aware Spectral Code Retrieval, which estimates input distributions to detect domain shifts, and retrieves adapted singular values for rapid adaptation. Consequently, our method achieves state-of-the-art performance on various distribution-shift benchmarks under the TTA setting. In CTTA and Gradual CTTA, it further improves accuracy by 3.4 percentage points (pp) and 2.4 pp, respectively, while requiring 385 times fewer trainable parameters. Our code is available at https://github.com/baek85/IMSE.", "authors": ["Sunghyun Baek", "Jaemyung Yu", "Seunghee Koh", "Minsu Kim", "Hyeonseong Jeon", "Junmo Kim"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-09", "url": "https://arxiv.org/abs/2603.07926", "pdf_url": "https://arxiv.org/pdf/2603.07926v3", "arxiv_id": "2603.07926", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/baek85/IMSE", "venue": null, "quality_score": 0.5972}
{"id": "2f40f4bc136d02fadb202ace7509768d0c66e7dadf6f72a5a951b904cbf03ff0", "sources": ["arxiv", "semantic_scholar"], "title": "Mixed Effects Mixture of Experts: Modeling Double Heterogeneous Trajectories", "abstract": "Linear mixed-effects model (LMM) is a cornerstone of longitudinal data analysis, but is limited to adeptly make heterogeneous analyses predictable under both group-specific fixed effects and subject-specific random effects. To address this challenge, we propose a novel statistical framework by using a large model prototype: a mixed effects mixture of experts model (MEMoE). This framework integrates the divide-and-conquer paradigm of Mixture of Experts Models with classical mixed-effect modeling. In the proposed MEMoE, each expert is a full LMM dedicated to capturing the longitudinal trajectory of a specific latent subpopulation, while another model gating function learns to route subjects to the most appropriate expert in a data-driven manner based on baseline covariates. We develop a robust inferential procedure for parameter estimation based on the Laplace Expectation-Maximization algorithm, with standard errors calibrated using robust sandwich estimators to account for potential model misspecification. Extensive simulation studies and an empirical application demonstrate that MEMoE outperforms both traditional single-population LMM and conventional Mixture of Experts models in terms of parameter recovery, classification accuracy, and overall model fit.", "authors": ["Xinkai Yue", "Xiaodong Yan", "Haohui Han", "Liya Fu"], "categories": ["stat.ME"], "fields_of_study": ["Mathematics"], "published_date": "2026-03-08", "url": "https://arxiv.org/abs/2603.07479", "pdf_url": "https://arxiv.org/pdf/2603.07479v1", "arxiv_id": "2603.07479", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3208}
{"id": "15d369a805ab9b6930c6e8660f43829401dda2d8ae25c1922c92070204890980", "sources": ["arxiv", "semantic_scholar"], "title": "EvoESAP: Non-Uniform Expert Pruning for Sparse MoE", "abstract": "Sparse Mixture-of-Experts (SMoE) language models achieve strong capability at low per-token compute, yet deployment remains constrained by memory footprint and throughput because the full expert pool must still be stored and served. Post-training expert pruning reduces this cost, but most methods focus on which experts to prune within each layer and default to a uniform layer-wise sparsity allocation, even though the layer-wise allocation can strongly affect performance. We decouple pruning into within-layer expert ranking and across-layer budget allocation, and introduce \\textbf{E}xpected \\textbf{S}peculative \\textbf{A}cceptance \\textbf{P}roxy (\\textbf{ESAP}), a speculative-decoding-inspired, teacher-forced metric that measures how well a pruned model matches the full model without costly autoregressive decoding. ESAP is bounded and stable, enabling cheap comparison of many candidates. Building on ESAP, we propose EvoESAP, an evolutionary search framework that finds an improved non-uniform layer-wise sparsity allocation under a fixed global budget while holding the within-layer pruning order fixed, making it a plug-and-play method for criteria such as Frequency, EAN, SEER, and REAP. Across 7B--30B SMoE LLMs at 25\\% and 50\\% sparsity, EvoESAP consistently discovers non-uniform allocations that improve open-ended generation (up to \\textbf{+19.6\\%} on MATH-500 at 50\\% sparsity) while preserving competitive multiple-choice accuracy compared with uniform pruning at the same sparsity. Code is available at https://github.com/ZongfangLiu/EvoESAP.", "authors": ["Zongfang Liu", "Shengkun Tang", "Boyang Sun", "Zhiqiang Shen", "Xin Yuan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-06", "url": "https://arxiv.org/abs/2603.06003", "pdf_url": "https://arxiv.org/pdf/2603.06003v2", "arxiv_id": "2603.06003", "doi": null, "citation_count": 3, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/ZongfangLiu/EvoESAP", "venue": null, "quality_score": 0.5931}
{"id": "e8b068df3f4ec9c9fd0283ab4a34ab5f9b49cb3c34a0f111ee05d0952c447e8c", "sources": ["arxiv", "semantic_scholar"], "title": "Video Compression Meets Video Generation: Latent Inter-Frame Pruning with Attention Recovery", "abstract": "Current video generation models suffer from high computational latency, making real-time applications prohibitively costly. In this paper, we address this limitation by exploiting the temporal redundancy inherent in video latent patches. To this end, we propose the Latent Inter-frame Pruning with Attention Recovery (LIPAR) framework, which detects and skips recomputing duplicated latent patches. Additionally, we introduce a novel Attention Recovery mechanism that approximates the attention values of pruned tokens, thereby removing visual artifacts arising from naively applying the pruning method. Empirically, our method increases video editing throughput by $1.53\\times$, achieving an average of 19.3 FPS on an NVIDIA RTX 4090 with the 1.3B Self-Forcing model (4-step denoising, FP16). The proposed method does not compromise generation quality and can be seamlessly integrated with the model without additional training. Our approach effectively bridges the gap between traditional compression algorithms and modern generative pipelines.", "authors": ["Dennis Menn", "Yuedong Yang", "Bokun Wang", "Xiwen Wei", "Mustafa Munir", "Feng Liang", "Radu Marculescu", "Chenfeng Xu", "Diana Marculescu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-06", "url": "https://arxiv.org/abs/2603.05811", "pdf_url": "https://arxiv.org/pdf/2603.05811v2", "arxiv_id": "2603.05811", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3194}
{"id": "a72ed7fcb916ff5f5e93f39f757a45ed90e967bec92066f1e8091d5025e5412c", "sources": ["arxiv", "semantic_scholar"], "title": "ButterflyViT: 354$\\times$ Expert Compression for Edge Vision Transformers", "abstract": "Deploying sparse Mixture of Experts(MoE) Vision Transformers remains a challenge due to linear expert memory scaling. Linear memory scaling stores $N$ independent expert weight matrices requiring $\\mathcal{O}(N_E \\cdot d^2)$ memory, which exceeds edge devices memory budget. Current compression methods like quantization, pruning and low-rank factorization reduce constant factors but leave the scaling bottleneck unresolved. We introduce ButterflyViT, a method that treats experts not as independent weight matrices but as geometric reorientations of a unified shared quantized substrate. Diversity among experts arises from viewing different angles of shared capacity, not from redundant storage. By applying learned rotations to a shared ternary prototype, each expert yields $\\mathcal{O}(d_{\\text{model}} \\cdot d_{\\text{ff}} + N_E \\cdot n_\\ell \\cdot d)$ memory which is sub-linear in the number of experts. To address the unique challenges of vision, a spatial smoothness regulariser is introduced that penalises routing irregularities between adjacent patch tokens, turning patch correlation into a training signal. Across image classification tasks on CIFAR-100, ButterflyViT achieves 354$\\times$ memory reduction at 64 experts with negligible accuracy loss. ButterflyViT allows multiple experts to fit on edge-constrained devices showing that geometric parameterization breaks linear scaling.", "authors": ["Aryan Karmore"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-06", "url": "https://arxiv.org/abs/2603.06746", "pdf_url": "https://arxiv.org/pdf/2603.06746v1", "arxiv_id": "2603.06746", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3194}
{"id": "4b6300ee0dc9a714d53be5a8f22959b02d75ae8892ff1d79c5d22ccf884d7aa8", "sources": ["arxiv", "semantic_scholar"], "title": "MoE Lens -- An Expert Is All You Need", "abstract": "Mixture of Experts (MoE) models enable parameter-efficient scaling through sparse expert activations, yet optimizing their inference and memory costs remains challenging due to limited understanding of their specialization behavior. We present a systematic analysis of expert specialization in MoEs through two complementary approaches: domain-specific routing patterns and an early decoding framework that tracks expert contributions to output representations. Our analysis of the DeepSeekMoE model reveals that despite having 64 routed experts with 6 active for each layer's computation, the model predominantly relies on a few specialized experts, with the top-weighted expert's output closely approximating the full ensemble prediction. We quantitatively validate these findings through a systematic analysis of the token routing distribution, demonstrating that very few experts handle over 50\\% of routing decisions across different specialized domains. Hidden state similarity between single and ensemble experts for every layer is extremely high, with some layers having cosine similarity as high as 0.95 and perplexity increasing by only 5\\% when using a single expert across all three domains. Our results indicate that Mixture of Experts models exhibit concentrated expertise highlighting potential opportunities for inference optimization through targeted expert pruning while maintaining model performance and opening avenues towards studying localization of learned knowledge in these models.", "authors": ["Marmik Chaudhari", "Idhant Gulati", "Nishkal Hundia", "Pranav Karra", "Shivam Raval"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-06", "url": "https://arxiv.org/abs/2603.05806", "pdf_url": "https://arxiv.org/pdf/2603.05806v1", "arxiv_id": "2603.05806", "doi": null, "citation_count": 11, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3194}
{"id": "5ec00039916aa4614b68a0a56d271285b5ab6d5c39cde89c38bb74fa8ca1b317", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Universal Experts: Scaling Virtual Width via Depth-Width Transformation", "abstract": "Mixture-of-Experts (MoE) decouples model capacity from per-token computation, yet their scalability remains limited by the physical dimensions of depth and width. To overcome this, we propose Mixture of Universal Experts (MOUE),a MoE generalization introducing a novel scaling dimension: Virtual Width. In general, MoUE aims to reuse a universal layer-agnostic expert pool across layers, converting depth into virtual width under a fixed per-token activation budget. However, two challenges remain: a routing path explosion from recursive expert reuse, and a mismatch between the exposure induced by reuse and the conventional load-balancing objectives. We address these with three core components: a Staggered Rotational Topology for structured expert sharing, a Universal Expert Load Balance for depth-aware exposure correction, and a Universal Router with lightweight trajectory state for coherent multi-step routing. Empirically, MoUE consistently outperforms matched MoE baselines by up to 1.3% across scaling regimes, enables progressive conversion of existing MoE checkpoints with up to 4.2% gains, and reveals a new scaling dimension for MoE architectures.", "authors": ["Yilong Chen", "Naibin Gu", "Junyuan Shang", "Zhenyu Zhang", "Yuchen Feng", "Jiawei Sheng", "Tingwen Liu", "Shuohuan Wang", "Yu Sun", "Hua Wu", "Haifeng Wang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-05", "url": "https://arxiv.org/abs/2603.04971", "pdf_url": "https://arxiv.org/pdf/2603.04971v1", "arxiv_id": "2603.04971", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3186}
{"id": "b91ddb7bd01ad9b7117b592fc66a228d564d5cbae85a0021e8501b288e95e6fb", "sources": ["arxiv", "semantic_scholar"], "title": "NeuronMoE: Neuron-Guided Mixture-of-Experts for Efficient Multilingual LLM Extension", "abstract": "Extending large language models to low-resource languages is essential for global accessibility, but training separate models per language is prohibitively expensive. Mixture-of-Experts (MoE) architectures address this by adding sparse language-specific parameters, but determining how many experts each layer needs remains an open question. Current approaches allocate experts based on layer-level similarity, yet language processing exhibits fine-grained specialization at individual neurons. We propose $\\textbf{NeuronMoE}$, a method that analyzes language-specific neurons across all transformer components to guide expert allocation per layer based on empirically measured cross-lingual neuron diversity. Applied to Llama-3.2-3B for low-resource languages (Greek, Turkish, and Hungarian), this approach achieves approximately 40% average parameter reduction while matching the performance of the LayerMoE baseline. We find that low-resource language experts independently develop neuron specialization patterns mirroring the high-resource language, which are concentrated in early and late layers. This reveals potential universal architectural principles in how multilingual models organize linguistic knowledge.", "authors": ["Rongzhi Li", "Hitomi Yanaka"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-05", "url": "https://arxiv.org/abs/2603.05046", "pdf_url": "https://arxiv.org/pdf/2603.05046v1", "arxiv_id": "2603.05046", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3186}
{"id": "d8a5ee10a68837edb4d4cca12f443a5346e183ff7e334eadf63a06e699c86398", "sources": ["arxiv", "semantic_scholar"], "title": "CAMMSR: Category-Guided Attentive Mixture of Experts for Multimodal Sequential Recommendation", "abstract": "The explosion of multimedia data in information-rich environments has intensified the challenges of personalized content discovery, positioning recommendation systems as an essential form of passive data management. Multimodal sequential recommendation, which leverages diverse item information such as text and images, has shown great promise in enriching item representations and deepening the understanding of user interests. However, most existing models rely on heuristic fusion strategies that fail to capture the dynamic and context-sensitive nature of user-modal interactions. In real-world scenarios, user preferences for modalities vary not only across individuals but also within the same user across different items or categories. Moreover, the synergistic effects between modalities-where combined signals trigger user interest in ways isolated modalities cannot-remain largely underexplored. To this end, we propose CAMMSR, a Category-guided Attentive Mixture of Experts model for Multimodal Sequential Recommendation. At its core, CAMMSR introduces a category-guided attentive mixture of experts (CAMoE) module, which learns specialized item representations from multiple perspectives and explicitly models inter-modal synergies. This component dynamically allocates modality weights guided by an auxiliary category prediction task, enabling adaptive fusion of multimodal signals. Additionally, we design a modality swap contrastive learning task to enhance cross-modal representation alignment through sequence-level augmentation. Extensive experiments on four public datasets demonstrate that CAMMSR consistently outperforms state-of-the-art baselines, validating its effectiveness in achieving adaptive, synergistic, and user-centric multimodal sequential recommendation.", "authors": ["Jinfeng Xu", "Zheyu Chen", "Shuo Yang", "Jinze Li", "Hewei Wang", "Yijie Li", "Jianheng Tang", "Yunhuai Liu", "Edith C. H. Ngai"], "categories": ["cs.IR", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-04", "url": "https://arxiv.org/abs/2603.04320", "pdf_url": "https://arxiv.org/pdf/2603.04320v1", "arxiv_id": "2603.04320", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3179}
{"id": "54fb8a3bff1281d74187205935a92da17aaa4c4cbb204292856aa1c682d3b505", "sources": ["arxiv", "semantic_scholar"], "title": "ECG-MoE: Mixture-of-Expert Electrocardiogram Foundation Model", "abstract": "Electrocardiography (ECG) analysis is crucial for cardiac diagnosis, yet existing foundation models often fail to capture the periodicity and diverse features required for varied clinical tasks. We propose ECG-MoE, a hybrid architecture that integrates multi-model temporal features with a cardiac period-aware expert module. Our approach uses a dual-path Mixture-of-Experts to separately model beat-level morphology and rhythm, combined with a hierarchical fusion network using LoRA for efficient inference. Evaluated on five public clinical tasks, ECG-MoE achieves state-of-the-art performance with 40% faster inference than multi-task baselines.", "authors": ["Yuhao Xu", "Xiaoda Wang", "Yi Wu", "Wei Jin", "Xiao Hu", "Carl Yang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-04", "url": "https://arxiv.org/abs/2603.04589", "pdf_url": "https://arxiv.org/pdf/2603.04589v1", "arxiv_id": "2603.04589", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3179}
{"id": "3accf6fe32d9c1ca71df0f6d48409ccaef178c723c8b3837d2fe7f426995ae09", "sources": ["arxiv", "semantic_scholar"], "title": "An experimental study of KV cache reuse strategies in chunk-level caching systems", "abstract": "Retrieval-augmented generation improves large language models' accuracy by adding relevant retrieved text to the prompt. Chunk level caching (CLC) accelerates inference by precomputing KV caches for these retrieved chunks and reusing them. However, these caches miss cross-attention dependencies between chunks, which can reduce output quality. Several methods try to improve CLC accuracy using different techniques. We make two main contributions. First, we show that existing CLC approaches have fundamental limitations that limit their accuracy or their applicability. We back this conclusion with an extensive CLC system experimental evaluation. Second, we observe that existing CLC techniques are complementary. We leverage this insight to propose a new CLC design that carefully combines them and achieves better accuracy.", "authors": ["Samuel Cestola", "Tianxiang Xia", "Zheng Weiyan", "Zheng Pengfei", "Diego Didona"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-03", "url": "https://arxiv.org/abs/2603.20218", "pdf_url": "https://arxiv.org/pdf/2603.20218v1", "arxiv_id": "2603.20218", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3172}
{"id": "a245a0591d3e1a939d400b8e0872f0877f63ddd0453b8818f8cc968c7472c70a", "sources": ["arxiv", "semantic_scholar"], "title": "CMoE: Contrastive Mixture of Experts for Motion Control and Terrain Adaptation of Humanoid Robots", "abstract": "For effective deployment in real-world environments, humanoid robots must autonomously navigate a diverse range of complex terrains with abrupt transitions. While the Vanilla mixture of experts (MoE) framework is theoretically capable of modeling diverse terrain features, in practice, the gating network exhibits nearly uniform expert activations across different terrains, weakening the expert specialization and limiting the model's expressive power. To address this limitation, we introduce CMoE, a novel single-stage reinforcement learning framework that integrates contrastive learning to refine expert activation distributions. By imposing contrastive constraints, CMoE maximizes the consistency of expert activations within the same terrain while minimizing their similarity across different terrains, thereby encouraging experts to specialize in distinct terrain types. We validated our approach on the Unitree G1 humanoid robot through a series of challenging experiments. Results demonstrate that CMoE enables the robot to traverse continuous steps up to 20 cm high and gaps up to 80 cm wide, while achieving robust and natural gait across diverse mixed terrains, surpassing the limits of existing methods. To support further research and foster community development, we release our code publicly.", "authors": ["Shihao Ma", "Hongjin Chen", "Zijun Xu", "Yi Zhao", "Ke Wu", "Ruichen Yang", "Leyao Zou", "Zhongxue Gan", "Wenchao Ding"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-03", "url": "https://arxiv.org/abs/2603.03067", "pdf_url": "https://arxiv.org/pdf/2603.03067v1", "arxiv_id": "2603.03067", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3172}
{"id": "5fa715fe3a60a261934f2db77ad2cb2a8e1dbd2e5ecd82071b625a362ac62a45", "sources": ["arxiv", "semantic_scholar"], "title": "Robust Heterogeneous Analog-Digital Computing for Mixture-of-Experts Models with Theoretical Generalization Guarantees", "abstract": "Sparse Mixture-of-Experts (MoE) models enable efficient scalability by activating only a small sub-set of experts per input, yet their massive parameter counts lead to substantial memory and energy inefficiency during inference. Analog in-memory computing (AIMC) offers a promising solution by eliminating frequent data movement between memory and compute units. However, mitigating hardware nonidealities of AIMC typically requires noise-aware retraining, which is infeasible for large MoE models. In this paper, we propose a retraining-free heterogeneous computation framework in which noise-sensitive experts, which are provably identifiable by their maximum neuron norm, are computed digitally while the majority of the experts are executed on AIMC hardware. We further assign densely activated modules, such as attention layers, to digital computation due to their high noise sensitivity despite comprising a small fraction of parameters. Extensive experiments on large MoE language models, including DeepSeekMoE and OLMoE, across multiple benchmark tasks validate the robustness of our approach in maintaining accuracy under analog nonidealities.", "authors": ["Mohammed Nowaz Rabbani Chowdhury", "Hsinyu Tsai", "Geoffrey W. Burr", "Kaoutar El Maghraoui", "Liu Liu", "Meng Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-03", "url": "https://arxiv.org/abs/2603.02633", "pdf_url": "https://arxiv.org/pdf/2603.02633v1", "arxiv_id": "2603.02633", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3172}
{"id": "a6274cf87866a614add422a8615eb0552b9259297c43a01d319520a6cf1fe347", "sources": ["arxiv", "semantic_scholar"], "title": "TenExp: Mixture-of-Experts-Based Tensor Decomposition Structure Search Framework", "abstract": "Recently, tensor decompositions continue to emerge and receive increasing attention. Selecting a suitable tensor decomposition to exactly capture the low-rank structures behind the data is at the heart of the tensor decomposition field, which remains a challenging and relatively under-explored problem. Current tensor decomposition structure search methods are still confined by a fixed factor-interaction family (e.g., tensor contraction) and cannot deliver the mixture of decompositions. To address this problem, we elaborately design a mixture-of-experts-based tensor decomposition structure search framework (termed as TenExp), which allows us to dynamically select and activate suitable tensor decompositions in an unsupervised fashion. This framework enjoys two unique advantages over the state-of-the-art tensor decomposition structure search methods. Firstly, TenExp can provide a suitable single decomposition beyond a fixed factor-interaction family. Secondly, TenExp can deliver a suitable mixture of decompositions beyond a single decomposition. Theoretically, we also provide the approximation error bound of TenExp, which reveals the approximation capability of TenExp. Extensive experiments on both synthetic and realistic datasets demonstrate the superiority of the proposed TenExp compared to the state-of-the-art tensor decomposition-based methods.", "authors": ["Ting-Wei Zhou", "Xi-Le Zhao", "Sheng Liu", "Wei-Hao Wu", "Yu-Bang Zheng", "Deyu Meng"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-03", "url": "https://arxiv.org/abs/2603.02720", "pdf_url": "https://arxiv.org/pdf/2603.02720v1", "arxiv_id": "2603.02720", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3172}
{"id": "b594db8811b9bb51a110e1e70f8aa78c4414d303f4b2808f68e97c6821007d84", "sources": ["arxiv", "semantic_scholar"], "title": "Fed-GAME: Personalized Federated Learning with Graph Attention Mixture-of-Experts For Time-Series Forecasting", "abstract": "Federated learning (FL) on graphs shows promise for distributed time-series forecasting. Yet, existing methods rely on static topologies and struggle with client heterogeneity. We propose Fed-GAME, a framework that models personalized aggregation as message passing over a learnable dynamic implicit graph. The core is a decoupled parameter difference-based update protocol, where clients transmit parameter differences between their fine-tuned private model and a shared global model. On the server, these differences are decomposed into two streams: (1) averaged difference used to updating the global model for consensus (2) the selective difference fed into a novel Graph Attention Mixture-of-Experts (GAME) aggregator for fine-grained personalization. In this aggregator, shared experts provide scoring signals while personalized gates adaptively weight selective updates to support personalized aggregation. Experiments on two real-world electric vehicle charging datasets demonstrate that Fed-GAME outperforms state-of-the-art personalized FL baselines.", "authors": ["Yi Li", "Han Liu", "Mingfeng Fan", "Guo Chen", "Chaojie Li", "Biplab Sikdar"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-02", "url": "https://arxiv.org/abs/2603.01363", "pdf_url": "https://arxiv.org/pdf/2603.01363v1", "arxiv_id": "2603.01363", "doi": "10.48550/arXiv.2603.01363", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.4973}
{"id": "d6fbe78c27e4559fe6a1ff2f7f58a708e380a94f087247a5c37ade171e256c24", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding the Physics of Key-Value Cache Compression for LLMs through Attention Dynamics", "abstract": "As context windows in LLMs scale to 100K+ tokens, the key-value (KV) cache becomes the dominant memory bottleneck, with recent methods claiming 80-90% savings and minimal benchmark degradation. We argue these evaluations miss a structural issue: attention is not just storage but routing, and retaining KV pairs does not guarantee semantic accessibility. We propose a physics-inspired view of KV compression as a controlled perturbation of token-level routing, distinguishing retention, accessibility, and utilization. Using synthetic tasks probing multi-entity tracking, disambiguation, coreference, and multi-hop reasoning, we find that moderate compression degrades internal representations with little accuracy loss, revealing redundancy; all models exhibit a sharp hallucination safety cliff near 90% compression, correlated with spikes in Global Eviction Ratio (GER), suggesting a phase transition in semantic reachability; and architectures differ in routing dynamics, with LLaMA showing early consensus and late diversification, and Qwen showing funnel-like late convergence, leading to distinct resilience profiles. Beyond erasure, we identify representational rigidity, where excessive head-level consensus collapses routing flexibility despite token survival. These results suggest sparse token-route structures govern compression tolerance, reframing KV compression as a structural probe of attention geometry and linking long-context scalability to sparsity and the lottery ticket hypothesis in self-attention.", "authors": ["Samhruth Ananthanarayanan", "Ayan Sengupta", "Tanmoy Chakraborty"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-02", "url": "https://arxiv.org/abs/2603.01426", "pdf_url": "https://arxiv.org/pdf/2603.01426v1", "arxiv_id": "2603.01426", "doi": "10.48550/arXiv.2603.01426", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4973}
{"id": "01cad438cf5f6c1b12989ed1529c26e188e9f119e201649e3dbfa5878b89f0b3", "sources": ["arxiv", "semantic_scholar"], "title": "AgilePruner: An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models", "abstract": "Large Vision-Language Models (LVLMs) have adopted visual token pruning strategies to mitigate substantial computational overhead incurred by extensive visual token sequences. While prior works primarily focus on either attention-based or diversity-based pruning methods, in-depth analysis of these approaches' characteristics and limitations remains largely unexplored. In this work, we conduct thorough empirical analysis using effective rank (erank) as a measure of feature diversity and attention score entropy to investigate visual token processing mechanisms and analyze the strengths and weaknesses of each approach. Our analysis reveals two insights: (1) Our erank-based quantitative analysis shows that many diversity-oriented pruning methods preserve substantially less feature diversity than intended; moreover, analysis using the CHAIR dataset reveals that the diversity they do retain is closely tied to increased hallucination frequency compared to attention-based pruning. (2) We further observe that attention-based approaches are more effective on simple images where visual evidence is concentrated, while diversity-based methods better handle complex images with distributed features. Building on these empirical insights, we show that incorporating image-aware adjustments into existing hybrid pruning strategies consistently improves their performance. We also provide a minimal instantiation of our empirical findings through a simple adaptive pruning mechanism, which achieves strong and reliable performance across standard benchmarks as well as hallucination-specific evaluations. Our project page available at https://cvsp-lab.github.io/AgilePruner.", "authors": ["Changwoo Baek", "Jouwon Song", "Sohyeon Kim", "Kyeongbo Kong"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-01", "url": "https://arxiv.org/abs/2603.01236", "pdf_url": "https://arxiv.org/pdf/2603.01236v1", "arxiv_id": "2603.01236", "doi": "10.48550/arXiv.2603.01236", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4961}
{"id": "6aa880401d0e7ff960d95a3c16529fe191cbbfb9c62d35556712869aa302f50d", "sources": ["arxiv", "semantic_scholar"], "title": "CoMoL: Efficient Mixture of LoRA Experts via Dynamic Core Space Merging", "abstract": "Large language models (LLMs) achieve remarkable performance on diverse downstream and domain-specific tasks via parameter-efficient fine-tuning (PEFT). However, existing PEFT methods, particularly MoE-LoRA architectures, suffer from limited parameter efficiency and coarse-grained adaptation due to the proliferation of LoRA experts and instance-level routing. To address these issues, we propose Core Space Mixture of LoRA (\\textbf{CoMoL}), a novel MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. Specifically, CoMoL introduces two key components: core space experts and core space routing. Core space experts store each expert in a compact core matrix, preserving diversity while controlling parameter growth. Core space routing dynamically selects and activates the appropriate core experts for each token, enabling fine-grained, input-adaptive routing. Activated core experts are then merged via a soft-merging strategy into a single core expert, which is combined with a shared LoRA to form a specialized LoRA module. Besides, the routing network is projected into the same low-rank space as the LoRA matrices, further reducing parameter overhead without compromising expressiveness. Extensive experiments demonstrate that CoMoL retains the adaptability of MoE-LoRA architectures while achieving parameter efficiency comparable to standard LoRA, consistently outperforming existing methods across multiple tasks.", "authors": ["Jie Cao", "Zhenxuan Fan", "Zhuonan Wang", "Tianwei Lin", "Ziyuan Zhao", "Rolan Yan", "Wenqiao Zhang", "Feifei Shao", "Hongwei Wang", "Jun Xiao", "Siliang Tang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-28", "url": "https://arxiv.org/abs/2603.00573", "pdf_url": "https://arxiv.org/pdf/2603.00573v2", "arxiv_id": "2603.00573", "doi": "10.48550/arXiv.2603.00573", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.495}
{"id": "321c534d4778702af3893413698916ea849e7e6fc2250d4d2cac6e4fbbf2c701", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring 3D Dataset Pruning", "abstract": "Dataset pruning has been widely studied for 2D images to remove redundancy and accelerate training, while particular pruning methods for 3D data remain largely unexplored. In this work, we study dataset pruning for 3D data, where its observed common long-tail class distribution nature make optimization under conventional evaluation metrics Overall Accuracy (OA) and Mean Accuracy (mAcc) inherently conflicting, and further make pruning particularly challenging. To address this, we formulate pruning as approximating the full-data expected risk with a weighted subset, which reveals two key errors: coverage error from insufficient representativeness and prior-mismatch bias from inconsistency between subset-induced class weights and target metrics. We propose representation-aware subset selection with per-class retention quotas for long-tail coverage, and prior-invariant teacher supervision using calibrated soft labels and embedding-geometry distillation. The retention quota also serves as a switch to control the OA-mAcc trade-off. Extensive experiments on 3D datasets show that our method can improve both metrics across multiple settings while adapting to different downstream preferences. Our code is available at https://github.com/XiaohanZhao123/3D-Dataset-Pruning.", "authors": ["Xiaohan Zhao", "Xinyi Shang", "Jiacheng Liu", "Zhiqiang Shen"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-28", "url": "https://arxiv.org/abs/2603.00651", "pdf_url": "https://arxiv.org/pdf/2603.00651v1", "arxiv_id": "2603.00651", "doi": "10.48550/arXiv.2603.00651", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/XiaohanZhao123/3D-Dataset-Pruning", "venue": "arXiv.org", "quality_score": 0.765}
{"id": "693275a93f2e944f5dc229aa40ce8c02a5046294bc0a2c74708e5c17761f67a5", "sources": ["arxiv", "semantic_scholar"], "title": "MME: Mixture of Mesh Experts with Random Walk Transformer Gating", "abstract": "In recent years, various methods have been proposed for mesh analysis, each offering distinct advantages and often excelling on different object classes. We present a novel Mixture of Experts (MoE) framework designed to harness the complementary strengths of these diverse approaches. We propose a new gate architecture that encourages each expert to specialise in the classes it excels in. Our design is guided by two key ideas: (1) random walks over the mesh surface effectively capture the regions that individual experts attend to, and (2) an attention mechanism that enables the gate to focus on the areas most informative for each expert's decision-making. To further enhance performance, we introduce a dynamic loss balancing scheme that adjusts a trade-off between diversity and similarity losses throughout the training, where diversity prompts expert specialization, and similarity enables knowledge sharing among the experts. Our framework achieves state-of-the-art results in mesh classification, retrieval, and semantic segmentation tasks. Our code is available at: https://github.com/amirbelder/MME-Mixture-of-Mesh-Experts.", "authors": ["Amir Belder", "Ayellet Tal"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-28", "url": "https://arxiv.org/abs/2603.00828", "pdf_url": "https://arxiv.org/pdf/2603.00828v1", "arxiv_id": "2603.00828", "doi": "10.48550/arXiv.2603.00828", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/amirbelder/MME-Mixture-of-Mesh-Experts", "venue": "arXiv.org", "quality_score": 0.765}
{"id": "560209751a2758cc7bfbdfec9761f7334bfb20d5bbc5f0c4dc4dc1de0675e905", "sources": ["arxiv", "semantic_scholar"], "title": "RelayCaching: Accelerating LLM Collaboration via Decoding KV Cache Reuse", "abstract": "The increasing complexity of AI tasks has shifted the paradigm from monolithic models toward multi-agent large language model (LLM) systems. However, these collaborative architectures introduce a critical bottleneck: redundant prefill computation for shared content generated by previous agents, which significantly increases KV cache memory usage and time-to-first-token (TTFT). While various KV cache methods have been proposed to mitigate prefill redundancy, they either fail to maintain accuracy on agent-generated outputs or exhibit low reuse rates due to rigid constraints. We present RelayCaching, a training-free inference method that directly reuses decoding phase KV caches from previous agents in subsequent prefill phases. Our key insight is that KV caches for identical content are highly consistent across phases, while prefix-induced deviations are sparse and localized within a limited range of layers and token positions. By selectively recomputing KV caches at these positions, RelayCaching preserves model accuracy with minimal overhead, yielding a superior accuracy-efficiency trade-off over existing methods. Experiments on diverse collaborative LLM tasks spanning mathematical reasoning, general knowledge, and code generation demonstrate that RelayCaching achieves over 80% KV cache reuse, reduces TTFT by up to $4.7\\times$ compared to the standard pipeline, all with negligible accuracy degradation.", "authors": ["Yingsheng Geng", "Yuchong Gao", "Weihong Wu", "Guyue Liu", "Jiang Liu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-28", "url": "https://arxiv.org/abs/2603.13289", "pdf_url": "https://arxiv.org/pdf/2603.13289v1", "arxiv_id": "2603.13289", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.315}
{"id": "563e46200c1f576f04300c48ac2e41b3c684afa3c576328eef82a2612e0af6aa", "sources": ["arxiv", "semantic_scholar"], "title": "Specializing Foundation Models via Mixture of Low-Rank Experts for Comprehensive Head CT Analysis", "abstract": "Foundation models pre-trained on large-scale datasets demonstrate strong transfer learning capabilities; however, their adaptation to complex multi-label diagnostic tasks-such as comprehensive head CT finding detection-remains understudied. Standard parameter-efficient fine-tuning methods such as LoRA apply uniform adaptations across pathology types, which may limit performance for diverse medical findings. We propose a Mixture of Low-Rank Experts (MoLRE) framework that extends LoRA with multiple specialized low-rank adapters and unsupervised soft routing. This approach enables conditional feature adaptation with less than 0.5% additional parameters and without explicit pathology supervision. We present a comprehensive benchmark of MoLRE across six state-of-the-art medical imaging foundation models spanning 2D and 3D architectures, general-domain, medical-domain, and head CT-specific pretraining, and model sizes ranging from 7M to 431M parameters. Using over 70,000 non-contrast head CT scans with 75 annotated findings-including hemorrhage, infarction, trauma, mass lesions, structural abnormalities, and chronic changes-our experiments demonstrate consistent performance improvements across all models. Gains vary substantially: general-purpose and medical-domain models show the largest improvements (DINOv3-Base: +4.6%; MedGemma: +4.3%), whereas 3D CT-specialized or very large models show more modest gains (+0.2-1.3%). The combination of MoLRE and MedGemma achieves the highest average detection AUC of 0.917. These findings highlight the importance of systematic benchmarking on target clinical tasks, as pretraining domain, architecture, and model scale interact in non-obvious ways.", "authors": ["Youngjin Yoo", "Han Liu", "Bogdan Georgescu", "Yanbo Zhang", "Sasa Grbic", "Michael Baumgartner", "Thomas J. Re", "Jyotipriya Das", "Poikavila Ullaskrishnan", "Eva Eibenberger", "Andrei Chekkoury", "Uttam K. Bodanapally", "Savvas Nicolaou", "Pina C. Sanelli", "Thomas J. Schroeppel", "Yvonne W. Lui", "Eli Gibson"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-28", "url": "https://arxiv.org/abs/2603.00675", "pdf_url": "https://arxiv.org/pdf/2603.00675v1", "arxiv_id": "2603.00675", "doi": "10.48550/arXiv.2603.00675", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.495}
{"id": "cfc930e8bc7d7b84ea9f2e118429a9bfc73724c4f778b7470e7c87392522d593", "sources": ["arxiv", "semantic_scholar"], "title": "Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization", "abstract": "Post-Training Quantization (PTQ) has emerged as an effective technique for alleviating the substantial computational and memory overheads of Vision-Language Models (VLMs) by compressing both weights and activations without retraining the full model. Existing PTQ methods primarily rely on static identification and global compensation of sensitive or outlier channels, yet they often overlook the distributional differences of these important channels across inputs, leading to unsatisfactory quantization. In this work, we observe that the distributions and occurrence frequencies of important channels vary significantly both across modalities and among tokens, even within the same modality. Accordingly, we propose \\textbf{Quant Experts (QE)}, a token-aware adaptive error compensation with mixture-of-experts for VLMs quantization. QE divides the important channels into token-independent and token-dependent groups. For the former, a shared expert is designed for most tokens to compensate for global quantization error using a low-rank adapter. For the latter, routed experts including multiple routed low-rank adapters are elaborated to compensate for local quantization error related to specific tokens. Extensive experiments demonstrate that QE consistently enhances task accuracy across various quantization settings and model scales, ranging from 2B to 70B parameters, while maintaining performance comparable to full-precision models.", "authors": ["Chenwei Jia", "Baoting Li", "Xuchong Zhang", "Mingzhuo Wei", "Bochen Lin", "Hongbin Sun"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-27", "url": "https://arxiv.org/abs/2602.24059", "pdf_url": "https://arxiv.org/pdf/2602.24059v1", "arxiv_id": "2602.24059", "doi": "10.48550/arXiv.2602.24059", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4939}
{"id": "4f1add3f314a1f834981e5549eed1560344a3745f23b1f2e050bfa89cd519a78", "sources": ["arxiv", "semantic_scholar"], "title": "KEEP: A KV-Cache-Centric Memory Management System for Efficient Embodied Planning", "abstract": "Memory-augmented Large Language Models (LLMs) have demonstrated remarkable capability for complex and long-horizon embodied planning. By keeping track of past experiences and environmental states, memory enables LLMs to maintain a global view, thereby avoiding repetitive exploration. However, existing approaches often store the memory as raw text, leading to excessively long prompts and high prefill latency. While it is possible to store and reuse the KV caches, the efficiency benefits are greatly undermined due to frequent KV cache updates. In this paper, we propose KEEP, a KV-cache-centric memory management system for efficient embodied planning. KEEP features 3 key innovations: (1) a Static-Dynamic Memory Construction algorithm that reduces KV cache recomputation by mixed-granularity memory group; (2) a Multi-hop Memory Re-computation algorithm that dynamically identifies important cross-attention among different memory groups and reconstructs memory interactions iteratively; (3) a Layer-balanced Memory Loading that eliminates unbalanced KV cache loading and cross-attention computation across different layers. Extensive experimental results have demonstrated that KEEP achieves 2.68x speedup with negligible accuracy loss compared with text-based memory methods on ALFRED dataset. Compared with the KV re-computation method CacheBlend (EuroSys'25), KEEP shows 4.13% success rate improvement and 1.90x time-to-first-token (TTFT) reduction. Our code is available on https://github.com/PKU-SEC-Lab/KEEP_Embodied_Memory.", "authors": ["Zebin Yang", "Tong Xie", "Baotong Lu", "Shaoshan Liu", "Bo Yu", "Meng Li"], "categories": ["cs.RO", "cs.AI", "cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-27", "url": "https://arxiv.org/abs/2602.23592", "pdf_url": "https://arxiv.org/pdf/2602.23592v2", "arxiv_id": "2602.23592", "doi": "10.48550/arXiv.2602.23592", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/PKU-SEC-Lab/KEEP_Embodied_Memory", "venue": "arXiv.org", "quality_score": 0.7632}
{"id": "3a7f5029f60d0fe3aee5e0f68d18a64724004bc96e2d77d3128e62a0a1125fa6", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Long-Horizon GUI Agents via Training-Free KV Cache Compression", "abstract": "Large Vision-Language Models (VLMs) have emerged as powerful engines for autonomous GUI agents, yet their deployment is severely constrained by the substantial memory footprint and latency of the Key-Value (KV) cache during long-horizon interactions. While existing cache compression methods have proven effective for LLMs, we empirically demonstrate that they suffer from suboptimal performance in GUI scenarios due to a fundamental misalignment: unlike general visual tasks where attention sparsity varies across layers, GUI attention patterns exhibit uniform high-sparsity across all transformer layers. Motivated by this insight, we propose ST-Lite, a training-free KV cache compression framework tailored for efficient GUI agents that explicitly addresses the dynamic spatio-trajectory dependencies within GUI data streams. ST-Lite introduces a novel dual-branch scoring policy incorporating Component-centric Spatial Saliency (CSS) and Trajectory-aware Semantic Gating (TSG). Specifically, CSS preserves the structural integrity of interactive UI elements by evaluating local neighborhood saliency, while TSG mitigates historical redundancy by dynamically filtering visually repetitive KV pairs within the interaction trajectory. Extensive evaluations demonstrate that with only a 10-20% cache budget, ST-Lite achieves a 2.45x decoding acceleration while maintaining comparable or even superior performance compared to full-cache baselines, offering a scalable solution for resource-constrained GUI agents.", "authors": ["Bowen Zhou", "Zhou Xu", "Wanli Li", "Jingyu Xiao", "Haoqian Wang"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-27", "url": "https://arxiv.org/abs/2603.00188", "pdf_url": "https://arxiv.org/pdf/2603.00188v1", "arxiv_id": "2603.00188", "doi": "10.48550/arXiv.2603.00188", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4939}
{"id": "333710e563850bb48541a6296b807135aac4571a4ef83f44439b230759a4791a", "sources": ["arxiv", "semantic_scholar"], "title": "ICaRus: Identical Cache Reuse for Efficient Multi Model Inference", "abstract": "Multi model inference has recently emerged as a prominent paradigm, particularly in the development of agentic AI systems. However, in such scenarios, each model must maintain its own Key-Value (KV) cache for the identical prompt, leading to substantial memory consumption. This explosive growth of KV caches forces LLM serving systems to evict previously stored caches, which in turn introduces significant recomputation overhead whenever the evicted caches are required again. Moreover, prefix caching is inherently infeasible across different models, forcing each model to recompute KV cache for the identical prompt, which leads to significant overhead. To alleviate these issues, we propose Identical Cache Reuse (ICaRus), a novel architecture that allows multiple models to share identical KV caches across all layers. ICaRus is based on the key observation that a decoder-only Transformer can be conceptually decomposed into a logical encoder, which generates KV caches, and a logical decoder, which predicts output tokens from the KV caches. ICaRus fine-tunes only the logical decoder while freezing the logical encoder, enabling multiple models to share an identical KV cache. This eliminates cache memory explosion and unexpected evictions while also allowing cross-model reuse of KV caches for new input tokens, thereby removing redundant recomputation in multi model inference achieving both efficiency and scalability. Moreover, by incorporating lightweight adapters such as LoRA, ICaRus parallelizes KV cache generation and next-token prediction during decoding. ICaRus achieves comparable accuracy to task-specific fine-tuned model across a diverse set of tasks, while allowing multiple specialized models to fully share KV caches. ICaRus achieves up to 11.1x lower P95 latency and 3.8x higher throughput in multi agent workflow with 8 different models, compared to conventional multi model system.", "authors": ["Sunghyeon Woo", "Jaeeun Kil", "Hoseung Kim", "Minsub Kim", "Joonghoon Kim", "Ahreum Seo", "Sungjae Lee", "Minjung Jo", "Jiwon Ryu", "Baeseong Park", "Se Jung Kwon", "Dongsoo Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-27", "url": "https://arxiv.org/abs/2603.13281", "pdf_url": "https://arxiv.org/pdf/2603.13281v1", "arxiv_id": "2603.13281", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3143}
{"id": "41b834de31ac4148a0fcbf58977f08fb2d19f3306a4a7bf86dd655b77e689731", "sources": ["arxiv", "semantic_scholar"], "title": "FedRot-LoRA: Mitigating Rotational Misalignment in Federated LoRA", "abstract": "Federated LoRA provides a communication-efficient mechanism for fine-tuning large language models on decentralized data. In practice, however, a discrepancy between the factor-wise averaging used to preserve low rank and the mathematically correct aggregation of local updates can cause significant aggregation error and unstable training. We argue that a major source of this problem is rotational misalignment, arising from the rotational invariance of low-rank factorizations -- semantically equivalent updates can be represented in different latent subspaces across clients since $(B_i R_i)(R_i^\\top A_i) = B_i A_i$. When such misaligned factors are averaged directly, they interfere destructively and degrade the global update. To address this issue, we propose FedRot-LoRA, a federated LoRA framework that aligns client updates via orthogonal transformations prior to aggregation. This alignment preserves the semantic update while reducing cross-client subspace mismatch, without increasing communication cost or restricting model expressivity. We provide a convergence analysis that examines the aggregation error induced by factor-wise averaging and shows how rotational alignment yields a tighter upper bound on this error. Extensive experiments on natural language understanding and generative tasks demonstrate that FedRot-LoRA consistently outperforms existing federated LoRA baselines across a range of heterogeneity levels and LoRA ranks.", "authors": ["Haoran Zhang", "Dongjun Kim", "Seohyeon Cha", "Haris Vikalo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-27", "url": "https://arxiv.org/abs/2602.23638", "pdf_url": "https://arxiv.org/pdf/2602.23638v2", "arxiv_id": "2602.23638", "doi": "10.48550/arXiv.2602.23638", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4939}
{"id": "93b370782b0fe6c8cbde7cc087d43136e3760299400ea7ed1f6f0bc0200c8a18", "sources": ["arxiv", "semantic_scholar"], "title": "InnerQ: Hardware-Aware Tuning-Free Quantization of KV Cache for Large Language Models", "abstract": "When transformer-based language models are deployed for text generation, most of the inference time is spent in the decoding stage, where output tokens are generated sequentially. Reducing the hardware cost of each decoding step is therefore critical for efficient long-context generation. A major bottleneck is the key-value (KV) cache, whose size grows with sequence length and often dominates the model's memory footprint. Prior work has proposed quantization methods to compress the KV cache while minimizing its loss of precision. We present InnerQ, a hardware-aware KV cache quantization scheme that reduces decode latency without compromising evaluation performance. InnerQ performs group-wise quantization by grouping cache matrices along their inner dimension. This grouping strategy aligns dequantization with vector-matrix multiplication and increases data reuse across GPU compute units. As a result, InnerQ reduces memory access and accelerates dequantization, achieving an average $1.3\\times$ speedup over prior KV cache quantization methods and $2.7\\times$ over the non-quantized baseline. To maintain fidelity under aggressive compression, InnerQ incorporates three techniques: (i) hybrid quantization, which chooses symmetric or asymmetric quantization for each group based on local statistics; (ii) high-precision windows for both recent tokens and attention sink tokens to mitigate outlier leakage; and (iii) per-channel normalization of the key cache, computed once during prefill and folded into the model parameters to eliminate runtime overhead. Beyond reducing latency, experiments on Llama and Mistral models show that InnerQ also improves few-shot evaluation scores relative to prior KV cache quantization methods.", "authors": ["Sayed Mohammadreza Tayaranian Hosseini", "Amir Ardakani", "Warren J. Gross"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-26", "url": "https://arxiv.org/abs/2602.23200", "pdf_url": "https://arxiv.org/pdf/2602.23200v2", "arxiv_id": "2602.23200", "doi": "10.48550/arXiv.2602.23200", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4927}
{"id": "1bdd38e73856d198e7e49b8dcd98548d7deb09f876838e097a96e17f9fce7539", "sources": ["arxiv", "semantic_scholar"], "title": "SideQuest: Model-Driven KV Cache Management for Long-Horizon Agentic Reasoning", "abstract": "Long-running agentic tasks, such as deep research, require multi-hop reasoning over information distributed across multiple webpages and documents. In such tasks, the LLM context is dominated by tokens from external retrieval, causing memory usage to grow rapidly and limiting decode performance. While several KV cache compression techniques exist for long-context inputs, we find that existing heuristics fail to support multi-step reasoning models effectively. We address this challenge with SideQuest -- a novel approach that leverages the Large Reasoning Model (LRM) itself to perform KV cache compression by reasoning about the usefulness of tokens in its context. To prevent the tokens associated with this management process from polluting the model's memory, we frame KV cache compression as an auxiliary task executed in parallel to the main reasoning task. Our evaluations, using a model trained with just 215 samples, show that SideQuest reduces peak token usage by up to 65% on agentic tasks with minimal degradation in accuracy, outperforming heuristic-based KV cache compression techniques.", "authors": ["Sanjay Kariyappa", "G. Edward Suh"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-26", "url": "https://arxiv.org/abs/2602.22603", "pdf_url": "https://arxiv.org/pdf/2602.22603v2", "arxiv_id": "2602.22603", "doi": "10.48550/arXiv.2602.22603", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4927}
{"id": "bb87aff0b9e29a3add7cc254f36ac293e4e89501783f4254ff03db86d279a694", "sources": ["arxiv", "semantic_scholar"], "title": "A Mixture-of-Experts Model for Multimodal Emotion Recognition in Conversations", "abstract": "Emotion Recognition in Conversations (ERC) presents unique challenges, requiring models to capture the temporal flow of multi-turn dialogues and to effectively integrate cues from multiple modalities. We propose Mixture of Speech-Text Experts for Recognition of Emotions (MiSTER-E), a modular Mixture-of-Experts (MoE) framework designed to decouple two core challenges in ERC: modality-specific context modeling and multimodal information fusion. MiSTER-E leverages large language models (LLMs) fine-tuned for both speech and text to provide rich utterance-level embeddings, which are then enhanced through a convolutional-recurrent context modeling layer. The system integrates predictions from three experts-speech-only, text-only, and cross-modal-using a learned gating mechanism that dynamically weighs their outputs. To further encourage consistency and alignment across modalities, we introduce a supervised contrastive loss between paired speech-text representations and a KL-divergence-based regulariza-tion across expert predictions. Importantly, MiSTER-E does not rely on speaker identity at any stage. Experiments on three benchmark datasets-IEMOCAP, MELD, and MOSI-show that our proposal achieves 70.9%, 69.5%, and 87.9% weighted F1-scores respectively, outperforming several baseline speech-text ERC systems. We also provide various ablations to highlight the contributions made in the proposed approach.", "authors": ["Soumya Dutta", "Smruthi Balaji", "Sriram Ganapathy"], "categories": ["cs.CL", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-02-26", "url": "https://arxiv.org/abs/2602.23300", "pdf_url": "https://arxiv.org/pdf/2602.23300v1", "arxiv_id": "2602.23300", "doi": "10.1016/j.csl.2026.101965", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3135}
{"id": "a1daf266bda742d9b81532f4b41843da58388050567cade5688b7835d2791492", "sources": ["arxiv"], "title": "XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression", "abstract": "Learning-based 3D visual geometry models have significantly advanced with the advent of large-scale transformers. Among these, StreamVGGT leverages frame-wise causal attention to deliver robust and efficient streaming 3D reconstruction. However, it suffers from unbounded growth in the Key-Value (KV) cache due to the massive influx of vision tokens from multi-image and long-video inputs, leading to increased memory consumption and inference latency as input frames accumulate. This ultimately limits its scalability for long-horizon applications. To address this gap, we propose XStreamVGGT, a tuning-free approach that seamlessly integrates pruning and quantization to systematically compress the KV cache, enabling extremely memory-efficient streaming inference. Specifically, redundant KVs generated from multi-frame inputs are initially pruned to conform to a fixed KV memory budget using an efficient token-importance identification mechanism that maintains full compatibility with high-performance attention kernels (e.g., FlashAttention). Additionally, leveraging the inherent distribution patterns of KV tensors, we apply dimension-adaptive KV quantization within the pruning pipeline to further minimize memory overhead while preserving numerical accuracy. Extensive evaluations show that XStreamVGGT achieves mostly negligible performance degradation while substantially reducing memory usage by 4.42$\\times$ and accelerating inference by 5.48$\\times$, enabling practical and scalable streaming 3D applications. The code is available at https://github.com/ywh187/XStreamVGGT/.", "authors": ["Zunhai Su", "Weihao Ye", "Hansen Feng", "Keyu Fan", "Jing Zhang", "Dahai Yu", "Zhengwu Liu", "Ngai Wong"], "categories": ["cs.CV"], "fields_of_study": [], "published_date": "2026-02-25", "url": "https://arxiv.org/abs/2602.21780", "pdf_url": "https://arxiv.org/pdf/2602.21780v1", "arxiv_id": "2602.21780", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ywh187/XStreamVGGT/", "venue": null, "quality_score": 0.5809}
{"id": "847243af4e3f3877a7a465d7402b380bcc4e52af8dacf5a204702c6bdf7bdf20", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Multi-Objective Tiered Storage Configuration for KV Cache in LLM Service", "abstract": "The memory-for-computation paradigm of KV caching is essential for accelerating large language model (LLM) inference service, but limited GPU high-bandwidth memory (HBM) capacity motivates offloading the KV cache to cheaper external storage tiers. While this expands capacity, it introduces the challenge of dynamically managing heterogeneous storage resources to balance cost, throughput, and latency under varying workloads. We formulate this as a multi-objective optimization problem: identifying the Pareto frontier across these metrics within the storage configuration space. Using a high-fidelity end-to-end simulator, we observe that the objective functions are non-analytic and exhibit complex variable coupling, making the Pareto frontier difficult to approximate analytically. To obtain the frontier, we introduce Kareto, a KV-cache Adaptive REsource managemenT Optimizer. Kareto leverages a diminishing-return-guided pruning method to efficiently navigate the large configuration space and approximate the Pareto frontier. Additionally, it incorporates a fine-grained adaptive tuner that uses eviction policies in tier storage and KV block access patterns for group-specific cache management, improving cache efficiency. Experiments on real-world traces show that Kareto adapts to workload and can identify configurations of better cost efficiency, covering static strategies. Compared to the fixed setup with 1024 GB DRAM, Kareto can improve throughput by up to 9.3%, or reduce latency by up to 58.3%, or lower cost by up to 20.2% under respective optimization objectives.", "authors": ["Xianzhe Zheng", "Zhengheng Wang", "Ruiyan Ma", "Rui Wang", "Xiyu Wang", "Rui Chen", "Peng Zhang", "Sicheng Pan", "Zhangheng Huang", "Chenxin Wu", "Yi Zhang", "Bo Cai", "Kan Liu", "Teng Ma", "Yin Du", "Dong Deng", "Sai Wu", "Guoyun Zhu", "Wei Zhang", "Feifei Li"], "categories": ["cs.AR", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-25", "url": "https://arxiv.org/abs/2603.08739", "pdf_url": "https://arxiv.org/pdf/2603.08739v1", "arxiv_id": "2603.08739", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3128}
{"id": "eafabe1139c966d6d214bd5c683fd581a5578106342d3b1086463af9d92674d1", "sources": ["arxiv", "semantic_scholar"], "title": "TiMi: Empower Time Series Transformers with Multimodal Mixture of Experts", "abstract": "Multimodal time series forecasting has garnered significant attention for its potential to provide more accurate predictions than traditional single-modality models by leveraging rich information inherent in other modalities. However, due to fundamental challenges in modality alignment, existing methods often struggle to effectively incorporate multimodal data into predictions, particularly textual information that has a causal influence on time series fluctuations, such as emergency reports and policy announcements. In this paper, we reflect on the role of textual information in numerical forecasting and propose Time series transformers with Multimodal Mixture-of-Experts, TiMi, to unleash the causal reasoning capabilities of LLMs. Concretely, TiMi utilizes LLMs to generate inferences on future developments, which serve as guidance for time series forecasting. To seamlessly integrate both exogenous factors and time series into predictions, we introduce a Multimodal Mixture-of-Experts (MMoE) module as a lightweight plug-in to empower Transformer-based time series models for multimodal forecasting, eliminating the need for explicit representation-level alignment. Experimentally, our proposed TiMi demonstrates consistent state-of-the-art performance on sixteen real-world multimodal forecasting benchmarks, outperforming advanced baselines while offering both strong adaptability and interpretability.", "authors": ["Jiafeng Lin", "Yuxuan Wang", "Huakun Luo", "Zhongyi Pei", "Jianmin Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-25", "url": "https://arxiv.org/abs/2602.21693", "pdf_url": "https://arxiv.org/pdf/2602.21693v1", "arxiv_id": "2602.21693", "doi": "10.48550/arXiv.2602.21693", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4916}
{"id": "619657738455ef4f5b2af3ede59c9885bfcbd07968ee8e05a72343baa5afebe5", "sources": ["arxiv", "semantic_scholar"], "title": "Sensitivity-Guided Framework for Pruned and Quantized Reservoir Computing Accelerators", "abstract": "This paper presents a compression framework for Reservoir Computing that enables systematic design-space exploration of trade-offs among quantization levels, pruning rates, model accuracy, and hardware efficiency. The proposed approach leverages a sensitivity-based pruning mechanism to identify and remove less critical quantized weights with minimal impact on model accuracy, thereby reducing computational overhead while preserving accuracy. We perform an extensive trade-off analysis to validate the effectiveness of the proposed framework and the impact of pruning and quantization on model performance and hardware parameters. For this evaluation, we employ three time-series datasets, including both classification and regression tasks. Experimental results across selected benchmarks demonstrate that our proposed approach maintains high accuracy while substantially improving computational and resource efficiency in FPGA-based implementations, with variations observed across different configurations and time series applications. For instance, for the MELBOEN dataset, an accelerator quantized to 4-bit at a 15\\% pruning rate reduces resource utilization by 1.2\\% and the Power Delay Product (PDP) by 50.8\\% compared to an unpruned model, without any noticeable degradation in accuracy.", "authors": ["Atousa Jafari", "Mahdi Taheri", "Hassan Ghasemzadeh Mohammadi", "Christian Herglotz", "Marco Platzner"], "categories": ["cs.AR", "cs.AI", "cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-24", "url": "https://arxiv.org/abs/2603.08737", "pdf_url": "https://arxiv.org/pdf/2603.08737v1", "arxiv_id": "2603.08737", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3121}
{"id": "c6782d8c86d1ce181bdd4433644e93baeaea622c64803a9e6e3a89dacc07b1c8", "sources": ["arxiv", "semantic_scholar"], "title": "Modality-Guided Mixture of Graph Experts with Entropy-Triggered Routing for Multimodal Recommendation", "abstract": "Multimodal recommendation enhances ranking by integrating user-item interactions with item content, which is particularly effective under sparse feedback and long-tail distributions. However, multimodal signals are inherently heterogeneous and can conflict in specific contexts, making effective fusion both crucial and challenging. Existing approaches often rely on shared fusion pathways, leading to entangled representations and modality imbalance. To address these issues, we propose MAGNET, a Modality-Guided Mixture of Adaptive Graph Experts Network with Progressive Entropy-Triggered Routing for Multimodal Recommendation, designed to enhance controllability, stability, and interpretability in multimodal fusion. MAGNET couples interaction-conditioned expert routing with structure-aware graph augmentation, so that both what to fuse and how to fuse are explicitly controlled and interpretable. At the representation level, a dual-view graph learning module augments the interaction graph with content-induced edges, improving coverage for sparse and long-tail items while preserving collaborative structure via parallel encoding and lightweight fusion. At the fusion level, MAGNET employs structured experts with explicit modality roles-dominant, balanced, and complementary-enabling a more interpretable and adaptive combination of behavioral, visual, and textual cues. To further stabilize sparse routing and prevent expert collapse, we introduce a two-stage entropy-weighting mechanism that monitors routing entropy. This mechanism automatically transitions training from an early coverage-oriented regime to a later specialization-oriented regime, progressively balancing expert utilization and routing confidence. Extensive experiments on public benchmarks demonstrate consistent improvements over strong baselines.", "authors": ["Ji Dai", "Quan Fang", "Dengsheng Cai"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-24", "url": "https://arxiv.org/abs/2602.20723", "pdf_url": "https://arxiv.org/pdf/2602.20723v2", "arxiv_id": "2602.20723", "doi": "10.48550/arXiv.2602.20723", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4904}
{"id": "8d4e0d766aa7178ee3264d47cca931984e5394f867867e3616b7c70b6b3895e8", "sources": ["arxiv", "semantic_scholar"], "title": "A Replicate-and-Quantize Strategy for Plug-and-Play Load Balancing of Sparse Mixture-of-Experts LLMs", "abstract": "Sparse Mixture-of-Experts (SMoE) architectures are increasingly used to scale large language models efficiently, delivering strong accuracy under fixed compute budgets. However, SMoE models often suffer from severe load imbalance across experts, where a small subset of experts receives most tokens while others are underutilized. Prior work has focused mainly on training-time solutions such as routing regularization or auxiliary losses, leaving inference-time behavior, which is critical for deployment, less explored. We present a systematic analysis of expert routing during inference and identify three findings: (i) load imbalance persists and worsens with larger batch sizes, (ii) selection frequency does not reliably reflect expert importance, and (iii) overall expert workload and importance can be estimated using a small calibration set. These insights motivate inference-time mechanisms that rebalance workloads without retraining or router modification. We propose Replicate-and-Quantize (R&Q), a training-free and near-lossless framework for dynamic workload rebalancing. In each layer, heavy-hitter experts are replicated to increase parallel capacity, while less critical experts and replicas are quantized to remain within the original memory budget. We also introduce a Load-Imbalance Score (LIS) to measure routing skew by comparing heavy-hitter load to an equal allocation baseline. Experiments across representative SMoE models and benchmarks show up to 1.4x reduction in imbalance with accuracy maintained within +/-0.6%, enabling more predictable and efficient inference.", "authors": ["Zijie Liu", "Jie Peng", "Jinhao Duan", "Zirui Liu", "Kaixiong Zhou", "Mingfu Liang", "Luke Simon", "Xi Liu", "Zhaozhuo Xu", "Tianlong Chen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-23", "url": "https://arxiv.org/abs/2602.19938", "pdf_url": "https://arxiv.org/pdf/2602.19938v1", "arxiv_id": "2602.19938", "doi": "10.48550/arXiv.2602.19938", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4893}
{"id": "2a213c36a0be521e0ef047a0696661189963c84681f9ccb430360e75a805b3ad", "sources": ["arxiv", "semantic_scholar"], "title": "Flash-VAED: Plug-and-Play VAE Decoders for Efficient Video Generation", "abstract": "Latent diffusion models have enabled high-quality video synthesis, yet their inference remains costly and time-consuming. As diffusion transformers become increasingly efficient, the latency bottleneck inevitably shifts to VAE decoders. To reduce their latency while maintaining quality, we propose a universal acceleration framework for VAE decoders that preserves full alignment with the original latent distribution. Specifically, we propose (1) an independence-aware channel pruning method to effectively mitigate severe channel redundancy, and (2) a stage-wise dominant operator optimization strategy to address the high inference cost of the widely used causal 3D convolutions in VAE decoders. Based on these innovations, we construct a Flash-VAED family. Moreover, we design a three-phase dynamic distillation framework that efficiently transfers the capabilities of the original VAE decoder to Flash-VAED. Extensive experiments on Wan and LTX-Video VAE decoders demonstrate that our method outperforms baselines in both quality and speed, achieving approximately a 6$\\times$ speedup while maintaining the reconstruction performance up to 96.9%. Notably, Flash-VAED accelerates the end-to-end generation pipeline by up to 36% with negligible quality drops on VBench-2.0.", "authors": ["Lunjie Zhu", "Yushi Huang", "Xingtong Ge", "Yufei Xue", "Zhening Liu", "Yumeng Zhang", "Zehong Lin", "Jun Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-22", "url": "https://arxiv.org/abs/2602.19161", "pdf_url": "https://arxiv.org/pdf/2602.19161v1", "arxiv_id": "2602.19161", "doi": "10.48550/arXiv.2602.19161", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Aoko955/Flash-VAED", "venue": "arXiv.org", "quality_score": 0.7544}
{"id": "52b1b3ad5eb18edf5f47250fbebd6a09e8a5ecd0bbf3edf9d5616a3544fa68e2", "sources": ["arxiv", "semantic_scholar"], "title": "Routing-Aware Explanations for Mixture of Experts Graph Models in Malware Detection", "abstract": "Mixture-of-Experts (MoE) offers flexible graph reasoning by combining multiple views of a graph through a learned router. We investigate routing-aware explanations for MoE graph models in malware detection using control flow graphs (CFGs). Our architecture builds diversity at two levels. At the node level, each layer computes multiple neighborhood statistics and fuses them with an MLP, guided by a degree reweighting factor rho and a pooling choice lambda in {mean, std, max}, producing distinct node representations that capture complementary structural cues in CFGs. At the readout level, six experts, each tied to a specific (rho, lambda) view, output graph-level logits that the router weights into a final prediction. Post-hoc explanations are generated with edge-level attributions per expert and aggregated using the router gates so the rationale reflects both what each expert highlights and how strongly it is selected. Evaluated against single-expert GNN baselines such as GCN, GIN, and GAT on the same CFG dataset, the proposed MoE achieves strong detection accuracy while yielding stable, faithful attributions under sparsity-based perturbations. The results indicate that making the router explicit and combining multi-statistic node encoding with expert-level diversity can improve the transparency of MoE decisions for malware analysis.", "authors": ["Hossein Shokouhinejad", "Roozbeh Razavi-Far", "Griffin Higgins", "Ali. A Ghorbani"], "categories": ["cs.CR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-22", "url": "https://arxiv.org/abs/2602.19025", "pdf_url": "https://arxiv.org/pdf/2602.19025v1", "arxiv_id": "2602.19025", "doi": "10.48550/arXiv.2602.19025", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4881}
{"id": "aa7b102e6dd7fe6ce6204cf14ec350faa07a1cd75ea3b62631a12ac254e86d20", "sources": ["arxiv", "semantic_scholar"], "title": "Sink-Aware Pruning for Diffusion Language Models", "abstract": "Diffusion Language Models (DLMs) incur high inference cost due to iterative denoising, motivating efficient pruning. Existing pruning heuristics largely inherited from autoregressive (AR) LLMs, typically preserve attention sink tokens because AR sinks serve as stable global anchors. We show that this assumption does not hold for DLMs: the attention-sink position exhibits substantially higher variance over the full generation trajectory (measured by how the dominant sink locations shift across timesteps), indicating that sinks are often transient and less structurally essential than in AR models. Based on this observation, we propose ${\\bf \\texttt{Sink-Aware Pruning}}$, which automatically identifies and prunes unstable sinks in DLMs (prior studies usually keep sinks for AR LLMs). Without retraining, our method achieves a better quality-efficiency trade-off and outperforms strong prior pruning baselines under matched compute. Our code is available at https://github.com/VILA-Lab/Sink-Aware-Pruning.", "authors": ["Aidar Myrzakhan", "Tianyi Li", "Bowei Guo", "Shengkun Tang", "Zhiqiang Shen"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-19", "url": "https://arxiv.org/abs/2602.17664", "pdf_url": "https://arxiv.org/pdf/2602.17664v1", "arxiv_id": "2602.17664", "doi": "10.48550/arXiv.2602.17664", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/VILA-Lab/Sink-Aware-Pruning", "venue": "arXiv.org", "quality_score": 0.7491}
{"id": "89825c64e037aa9c0e06a6cc615ab129a6ad01aa1d1ee180d5936b305e4572e0", "sources": ["arxiv", "semantic_scholar"], "title": "Phase-Aware Mixture of Experts for Agentic Reinforcement Learning", "abstract": "Reinforcement learning (RL) has equipped LLM agents with a strong ability to solve complex tasks. However, existing RL methods normally use a \\emph{single} policy network, causing \\emph{simplicity bias} where simple tasks occupy most parameters and dominate gradient updates, leaving insufficient capacity for complex tasks. A plausible remedy could be employing the Mixture-of-Experts (MoE) architecture in the policy network, as MoE allows different parameters (experts) to specialize in different tasks, preventing simple tasks from dominating all parameters. However, a key limitation of traditional MoE is its token-level routing, where the router assigns each token to specialized experts, which fragments phase-consistent patterns into scattered expert assignments and thus undermines expert specialization. In this paper, we propose \\textbf{Phase-Aware Mixture of Experts (PA-MoE)}. It first features a lightweight \\emph{phase router} that learns latent phase boundaries directly from the RL objective without pre-defining phase categories. Then, the phase router allocates temporally consistent assignments to the same expert, allowing experts to preserve phase-specific expertise. Experimental results demonstrate the effectiveness of our proposed PA-MoE.", "authors": ["Shengtian Yang", "Yu Li", "Shuo He", "Yewen Li", "Qingpeng Cai", "Peng Jiang", "Lei Feng"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-19", "url": "https://arxiv.org/abs/2602.17038", "pdf_url": "https://arxiv.org/pdf/2602.17038v3", "arxiv_id": "2602.17038", "doi": "10.48550/arXiv.2602.17038", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4847}
{"id": "bcb067e50c3f6de230c1df078b0830de05aa16da7efb59dc13305bab50b400f9", "sources": ["arxiv", "semantic_scholar"], "title": "ARKV: Adaptive and Resource-Efficient KV Cache Management under Limited Memory Budget for Long-Context Inference in LLMs", "abstract": "Large Language Models (LLMs) are increasingly deployed in scenarios demanding ultra-long context reasoning, such as agentic workflows and deep research understanding. However, long-context inference is constrained by the KV cache, a transient memory structure that grows linearly with sequence length and batch size, quickly dominating GPU memory usage. Existing memory reduction techniques, including eviction and quantization, often rely on static heuristics and suffer from degraded quality under tight budgets. In this paper, we propose ARKV, a lightweight and adaptive framework that dynamically allocates precision levels to cached tokens based on per-layer attention dynamics and token-level importance. During a short prefill phase, ARKV estimates the original quantization (OQ) ratio of each layer by computing statistical scores such as attention entropy, variance and kurtosis. During decoding, tokens are assigned to one of three states, Original (full precision), Quantization (low precision), or Eviction, according to a fast heavy-hitter scoring strategy. Our experiments on LLaMA3 and Qwen3 models across diverse long- and short-context tasks demonstrate that ARKV preserves ~97% of baseline accuracy on long-context benchmarks while reducing KV memory usage by 4x, with minimal throughput loss. On short-context tasks, ARKV matches full-precision baselines; on GSM8K math reasoning, it significantly outperforms uniform quantization. These results highlight the practical viability of ARKV for scalable LLM deployment, offering fine-grained, data-driven memory control without retraining or architectural modifications. The source code and artifacts can be found in: https://github.com/Large-scale-Sustainable-Computing-LSC/ARKV", "authors": ["Jianlong Lei", "Shashikant Ilager"], "categories": ["cs.AR", "cs.AI", "cs.DC", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-19", "url": "https://arxiv.org/abs/2603.08727", "pdf_url": "https://arxiv.org/pdf/2603.08727v1", "arxiv_id": "2603.08727", "doi": null, "citation_count": 3, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/Large-scale-Sustainable-Computing-LSC/ARKV", "venue": null, "quality_score": 0.5728}
{"id": "0cc3f2c90ec6f0e37139eea420a6776034cd71e82a5590278a8b89f3c812c940", "sources": ["arxiv", "semantic_scholar"], "title": "Grassmannian Mixture-of-Experts: Concentration-Controlled Routing on Subspace Manifolds", "abstract": "Mixture-of-Experts models rely on learned routers to assign tokens to experts, yet standard softmax gating provides no principled mechanism to control the tradeoff between sparsity and utilization. We propose Grassmannian MoE (GrMoE), a routing framework that operates on the Grassmannian manifold of subspaces, where gating weights arise from the concentration parameters of Matrix Bingham distributions. This construction yields a single, interpretable knob -- the concentration matrix $Λ$ -- that continuously controls routing entropy, replacing discrete top-$k$ selection with a smooth, geometrically principled sparsity mechanism. We further develop an amortized variational inference procedure for posterior routing distributions, enabling uncertainty-aware expert assignment that naturally resists expert collapse. We formally prove tight bounds relating the Bingham concentration spectrum to routing entropy, expected top-$k$ mass, and an exponential bound on expert collapse, establishing the first formal theory of concentration-controlled sparsity. On synthetic routing tasks, a 350M-parameter MoE language model with 8 experts, a 1.3B-parameter model with 16 experts, and a 2.7B-parameter model with 32 experts, GrMoE achieves 0\\% routing collapse across all seeds, comparable or better perplexity with 15--30\\% improved load balance, and a smooth monotonic relationship between concentration and effective sparsity that enables post-hoc sparsity tuning without retraining. Token-level analysis reveals that experts learn heterogeneous concentration values that correlate with linguistic specialization, providing interpretable routing behavior.", "authors": ["Ibne Farabi Shihab", "Sanjeda Akter", "Anuj Sharma"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-19", "url": "https://arxiv.org/abs/2602.17798", "pdf_url": "https://arxiv.org/pdf/2602.17798v1", "arxiv_id": "2602.17798", "doi": "10.48550/arXiv.2602.17798", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4847}
{"id": "26fca49392145b70da1dc794cc093d5acf46713001a940a6ea3c3b06bd1d00f1", "sources": ["arxiv", "semantic_scholar"], "title": "Fast KV Compaction via Attention Matching", "abstract": "Scaling language models to long contexts is often bottlenecked by the size of the key-value (KV) cache. In deployed settings, long contexts are typically managed through compaction in token space via summarization. However, summarization can be highly lossy, substantially harming downstream performance. Recent work on Cartridges has shown that it is possible to train highly compact KV caches in latent space that closely match full-context performance, but at the cost of slow and expensive end-to-end optimization. This work describes an approach for fast context compaction in latent space through Attention Matching, which constructs compact keys and values to reproduce attention outputs and preserve attention mass at a per-KV-head level. We show that this formulation naturally decomposes into simple subproblems, some of which admit efficient closed-form solutions. Within this framework, we develop a family of methods that significantly push the Pareto frontier of compaction time versus quality, achieving up to 50x compaction in seconds on some datasets with little quality loss.", "authors": ["Adam Zweiger", "Xinghong Fu", "Han Guo", "Yoon Kim"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-18", "url": "https://arxiv.org/abs/2602.16284", "pdf_url": "https://arxiv.org/pdf/2602.16284v2", "arxiv_id": "2602.16284", "doi": "10.48550/arXiv.2602.16284", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4835}
{"id": "5c0db29e0c4622e69425febb1a88fb7986f524ee0e1d0a2b75107c7fcc1aa275", "sources": ["arxiv", "semantic_scholar"], "title": "Agent Memory Below the Prompt: Persistent Q4 KV Cache for Multi-Agent LLM Inference on Edge Devices", "abstract": "Multi-agent LLM systems on edge devices face a memory management problem: device RAM is too small to hold every agent's KV cache simultaneously. On Apple M4 Pro with 10.2 GB of cache budget, only 3 agents fit at 8K context in FP16. A 10-agent workflow must constantly evict and reload caches. Without persistence, every eviction forces a full re-prefill through the model -- 15.7 seconds per agent at 4K context. We address this by persisting each agent's KV cache to disk in 4-bit quantized format and reloading it directly into the attention layer, eliminating redundant O(n) prefill computation via direct cache restoration. The system comprises three components: a block pool providing per-agent isolated Q4 KV caches in safetensors format, a BatchQuantizedKVCache for concurrent inference over multiple agents' quantized caches, and cross-phase context injection that accumulates attention state across conversation phases without re-computation. Evaluated on three architectures (Gemma 3 12B, dense GQA, 48 layers; DeepSeek-Coder-V2-Lite 16B, MoE MLA, 27 layers; Llama 3.1 8B, dense GQA, 32 layers), cache restoration reduces time-to-first-token by up to 136x (Gemma: 22--136x at 4K--32K; DeepSeek: 11--76x at 4K--32K; Llama: 24--111x at 4K--16K; 3--10x at 1K). Q4 quantization fits 4x more agent contexts into fixed device memory than FP16. Perplexity measured with actual Q4 KV caches shows -0.7% for Gemma, +2.8% for Llama, and +3.0% for DeepSeek. Open-source at https://github.com/yshk-mxim/agent-memory", "authors": ["Yakov Pyotr Shkolnikov"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-17", "url": "https://arxiv.org/abs/2603.04428", "pdf_url": "https://arxiv.org/pdf/2603.04428v1", "arxiv_id": "2603.04428", "doi": null, "citation_count": 3, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/yshk-mxim/agent-memory", "venue": null, "quality_score": 0.5701}
{"id": "c250a273463f13badd38b0985637fe1cfd52cd376de4ac12c1c9c898ddd6303c", "sources": ["arxiv", "semantic_scholar"], "title": "Thin Keys, Full Values: Reducing KV Cache via Low-Dimensional Attention Selection", "abstract": "Standard Transformer attention uses identical dimensionality for queries, keys, and values, yet these components serve different roles: queries and keys produce scalar attention weights (selection), while values carry rich representations (value transfer). We show that selection requires only $O(\\log N)$ dimensions to distinguish among $N$ relevant token categories (e.g., syntactic roles, semantic clusters, positional patterns) -- far fewer than value transfer needs. We introduce factored keys, which exploit this asymmetry to physically shrink the KV cache of any pretrained model without retraining from scratch -- unlike Grouped-Query Attention (GQA) and Multi-Head Latent Attention (MLA), which must be designed into the architecture before pretraining. We factorize each key projection $W_K \\approx A_{d \\times r} B_{r \\times d}$ via truncated singular value decomposition (SVD) (where $r$ is the chosen compression dimension), set $W_K' = A$ as the new key projection producing compact $r$-dimensional keys for the cache, and absorb $B^\\top$ into the query projection ($W_Q' = W_Q B^\\top$) at zero cost -- since queries are never cached. At the 7B scale, training from scratch with $r = d/4$ (where $d$ is the model dimension) matches full-attention perplexity ($9.24$ vs $9.25$ PPL after 20B tokens, mean over two seeds) while using 12% fewer parameters and training 8% faster. For existing models, SVD followed by QK fine-tuning (3 epochs, less than 1% of pretraining data) achieves 75% key cache savings at roughly 2% quality cost on both GPT-2 and Mistral-7B. The approach composes with GQA and quantization for up to $16\\times$ combined key cache compression. For a 7B model serving a 128K context, factored keys save 25 GB of KV cache per user, enabling roughly 60% more concurrent users on identical hardware.", "authors": ["Hengshuai Yao", "Xing Chen", "Ahmed Murtadha", "Guan Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-16", "url": "https://arxiv.org/abs/2603.04427", "pdf_url": "https://arxiv.org/pdf/2603.04427v4", "arxiv_id": "2603.04427", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3062}
{"id": "014591f595f53721d88489052644e7412ccb02980d3acefaacbae172550e5504", "sources": ["arxiv", "semantic_scholar"], "title": "Parameter-Efficient Fine-Tuning of LLMs with Mixture of Space Experts", "abstract": "Large Language Models (LLMs) have achieved remarkable progress, with Parameter-Efficient Fine-Tuning (PEFT) emerging as a key technique for downstream task adaptation. However, existing PEFT methods mainly operate in Euclidean space, fundamentally limiting their capacity to capture complex geometric structures inherent in language data. While alternative geometric spaces, like hyperbolic geometries for hierarchical data and spherical manifolds for circular patterns, offer theoretical advantages, forcing representations into a single manifold type ultimately limits expressiveness, even when curvature parameters are learnable. To address this, we propose Mixture of Space (MoS), a unified framework that leverages multiple geometric spaces simultaneously to learn richer, curvature-aware representations. Building on this scheme, we develop MoSLoRA, which extends Low-Rank Adaptation (LoRA) with heterogeneous geometric experts, enabling models to dynamically select or combine appropriate geometric spaces based on input context. Furthermore, to address the computational overhead of frequent manifold switching, we develop a lightweight routing mechanism. Moreover, we provide empirical insights into how curvature optimization impacts training stability and model performance. Our experiments across diverse benchmarks demonstrate that MoSLoRA consistently outperforms strong baselines, achieving up to 5.6% improvement on MATH500 and 15.9% on MAWPS.", "authors": ["Buze Zhang", "Jinkai Tao", "Zilang Zeng", "Neil He", "Ali Maatouk", "Menglin Yang", "Rex Ying"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-16", "url": "https://arxiv.org/abs/2602.14490", "pdf_url": "https://arxiv.org/pdf/2602.14490v1", "arxiv_id": "2602.14490", "doi": "10.48550/arXiv.2602.14490", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4813}
{"id": "56f91b1d295b543affa4ba5e8da0fb68c25860151d09b9874075168944469197", "sources": ["arxiv", "semantic_scholar"], "title": "Dual-Signal Adaptive KV-Cache Optimization for Long-Form Video Understanding in Vision-Language Models", "abstract": "Vision-Language Models (VLMs) face a critical memory bottleneck when processing long-form video content due to the linear growth of the Key-Value (KV) cache with sequence length. Existing solutions predominantly employ reactive eviction strategies that compute full attention matrices before discarding tokens, resulting in substantial computational waste. We propose Sali-Cache, a novel a priori optimization framework that implements dual-signal adaptive caching through proactive memory management. By integrating a temporal filter based on optical flow analysis for detecting inter-frame redundancy and a spatial filter leveraging saliency detection for identifying visually significant regions, Sali-Cache intelligently manages memory allocation before entering computationally expensive attention operations. Experimental evaluation on the LLaVA 1.6 architecture demonstrates that our method achieves a 2.20x compression ratio in effective memory usage while maintaining 100% accuracy across BLEU, ROUGE-L, and Exact Match metrics. Furthermore, under identical memory budget constraints, Sali-Cache preserves context-rich features over extended temporal durations without degrading model performance, enabling efficient processing of long-form video content on consumer-grade hardware.", "authors": ["Vishnu Sai", "Dheeraj Sai", "Srinath B", "Girish Varma", "Priyesh Shukla"], "categories": ["cs.CV", "cs.AI", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-15", "url": "https://arxiv.org/abs/2602.14236", "pdf_url": "https://arxiv.org/pdf/2602.14236v1", "arxiv_id": "2602.14236", "doi": "10.48550/arXiv.2602.14236", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4801}
{"id": "66946a8b7501d1da8bb409317595e6fd842a31f7d3797e18d0d2736d65bcbeac", "sources": ["arxiv", "semantic_scholar"], "title": "Lamer-SSL: Layer-aware Mixture of LoRA Experts for Continual Multilingual Expansion of Self-supervised Models without Forgetting", "abstract": "Despite their impressive performance, self-supervised speech models often struggle to generalize to new languages and tend to forget previously acquired knowledge during continual training. To address this, we propose Lamer-SSL, a parameter-efficient framework that integrates a Layer-Aware MixturE of LoRA Experts (Lamer) module with a replay strategy. The Lamer module enables flexible balancing between shared and language-specific representations, while layer-aware expert allocation assigns more experts to deeper layers where semantic information is richer. Meanwhile, the replay strategy retains prior knowledge using minimal data, mitigating forgetting during continual training. Experiments on automatic speech recognition (ASR) and language identification (LID) demonstrate that Lamer-SSL extends self-supervised models to new languages effectively while maintaining strong performance on previously learned languages with only 2.14% parameters being trainable.", "authors": ["Jing Xu", "Minglin Wu", "Xueyuan Chen", "Xixin Wu", "Helen Meng"], "categories": ["cs.CL", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-02-13", "url": "https://arxiv.org/abs/2602.12746", "pdf_url": "https://arxiv.org/pdf/2602.12746v1", "arxiv_id": "2602.12746", "doi": "10.48550/arXiv.2602.12746", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.4778}
{"id": "a64b4f13b2a7b3791276a2a8b84d750f33b0634fb923c17562adb35f17abe269", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Predefined Experts: Maximizing Data Usage on Vertical Federated Learning", "abstract": "Vertical Federated Learning (VFL) has emerged as a critical paradigm for collaborative model training in privacy-sensitive domains such as finance and healthcare. However, most existing VFL frameworks rely on the idealized assumption of full sample alignment across participants, a premise that rarely holds in real-world scenarios. To bridge this gap, this work introduces Split-MoPE, a novel framework that integrates Split Learning with a specialized Mixture of Predefined Experts (MoPE) architecture. Unlike standard Mixture of Experts (MoE), where routing is learned dynamically, MoPE uses predefined experts to process specific data alignments, effectively maximizing data usage during both training and inference without requiring full sample overlap. By leveraging pretrained encoders for target data domains, Split-MoPE achieves state-of-the-art performance in a single communication round, significantly reducing the communication footprint compared to multi-round end-to-end training. Furthermore, unlike existing proposals that address sample misalignment, this novel architecture provides inherent robustness against malicious or noisy participants and offers per-sample interpretability by quantifying each collaborator's contribution to each prediction. Extensive evaluations on vision (CIFAR-10/100) and tabular (Breast Cancer Wisconsin) datasets demonstrate that Split-MoPE consistently outperforms state-of-the-art systems such as LASER and Vertical SplitNN, particularly in challenging scenarios with high data missingness.", "authors": ["Jon Irureta", "Gorka Azkune", "Jon Imaz", "Aizea Lojo", "Javier Fernandez-Marques"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-13", "url": "https://arxiv.org/abs/2602.12708", "pdf_url": "https://arxiv.org/pdf/2602.12708v2", "arxiv_id": "2602.12708", "doi": "10.48550/arXiv.2602.12708", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4778}
{"id": "1a27ea726206efd0d279f03246087d06ba9fb2168932b09891ba381aaa5db7b6", "sources": ["arxiv", "semantic_scholar"], "title": "Doc-to-LoRA: Learning to Instantly Internalize Contexts", "abstract": "Long input sequences are central to in-context learning, document understanding, and multi-step reasoning of Large Language Models (LLMs). However, the quadratic attention cost of Transformers makes inference memory-intensive and slow. While context distillation (CD) can transfer information into model parameters, per-prompt distillation is impractical due to training costs and latency. To address these limitations, we propose Doc-to-LoRA (D2L), a lightweight hypernetwork that meta-learns to perform approximate CD within a single forward pass. Given an unseen prompt, D2L generates a LoRA adapter for a target LLM, enabling subsequent queries to be answered without re-consuming the original context, reducing latency and KV-cache memory consumption during inference of the target LLM. On a long-context needle-in-a-haystack task, D2L successfully learns to map contexts into adapters that store the needle information, achieving near-perfect zero-shot accuracy at sequence lengths exceeding the target LLM's native context window by more than 4x. On real-world QA datasets with limited compute, D2L outperforms standard CD while significantly reducing peak memory consumption and update latency. We envision that D2L can facilitate rapid adaptation of LLMs, opening up the possibility of frequent knowledge updates and personalized chat behavior.", "authors": ["Rujikorn Charakorn", "Edoardo Cetin", "Shinnosuke Uesaka", "Robert Tjarko Lange"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-13", "url": "https://arxiv.org/abs/2602.15902", "pdf_url": "https://arxiv.org/pdf/2602.15902v1", "arxiv_id": "2602.15902", "doi": "10.48550/arXiv.2602.15902", "citation_count": 16, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4778}
{"id": "c4612ae6d47b10c378cf414ac6e614e08ae05f7d084757921439521afbcf24b5", "sources": ["arxiv", "semantic_scholar"], "title": "Cachemir: Fully Homomorphic Encrypted Inference of Generative Large Language Model with KV Cache", "abstract": "Generative large language models (LLMs) have revolutionized multiple domains. Modern LLMs predominantly rely on an autoregressive decoding strategy, which generates output tokens sequentially and employs a key-value cache (KV cache) to avoid redundant computation. However, the widespread deployment of LLMs has raised serious privacy concerns, as users are feeding all types of data into the model, motivating the development of secure inference frameworks based on fully homomorphic encryption (FHE). A major limitation of existing FHE-based frameworks is their inability to effectively integrate the KV cache, resulting in prohibitively high latency for autoregressive decoding. In this paper, we propose Cachemir, a KV Cache Accelerated Homomorphic Encrypted LLM Inference Regime to overcome this limitation. Cachemir comprises three key technical contributions: 1) a set of novel HE packing algorithms specifically designed to leverage the computational advantages of the KV cache; 2) an interleaved replicated packing algorithm to efficiently compute the vector-matrix multiplications that result from using the KV cache in Transformer linear layers; and 3) an augmented bootstrapping placement strategy that accounts for the KV cache to minimize bootstrapping cost. We demonstrate that Cachemir achieves $48.83\\times$ and $67.16\\times$ speedup over MOAI (ICML'25) and THOR (CCS'25) respectively on CPU and consumes less than 100 seconds on GPU to generate an output token for Llama-3-8B.", "authors": ["Ye Yu", "Yifan Zhou", "Yi Chen", "Pedro Soto", "Wenjie Xiong", "Meng Li"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-12", "url": "https://arxiv.org/abs/2602.11470", "pdf_url": "https://arxiv.org/pdf/2602.11470v1", "arxiv_id": "2602.11470", "doi": "10.48550/arXiv.2602.11470", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4767}
{"id": "d65dcea243cce2e939a358fb8b0efb323a8b66a129c4bed35ada3bbf2df79d83", "sources": ["arxiv", "semantic_scholar"], "title": "GORGO: Maximizing KV-Cache Reuse While Minimizing Network Latency in Cross-Region LLM Load Balancing", "abstract": "Distributing LLM inference across geographical regions can improve Time-to-First-Token (TTFT) by regionalizing service deployments. While existing multi-region load balancers save prefill computation by prioritizing Key--Value (KV) Cache hit rate, they ignore cluster networking latency, a critical factor in routing decisions. We introduce GORGO, a method for minimizing TTFT by optimizing a total serving cost as a function of available compute, network latency, and prefix caching. Using extensive profiling on custom infrastructure, we analyze component-level latency bottlenecks and benchmark GORGO against three baselines: (1) naive least-load routing, which ignores prefix-cache overlap; (2) prefix-similarity routing, which selectively pushes requests to the replica with the highest cached-prefix overlap; and (3) a centralized HTTP proxy that runs the GORGO policy while tracking requests across all nodes. We demonstrate that GORGO reduces P99 TTFT through network-aware routing and improves average TTFT by preventing pathological cross-region forwarding. Additionally, we find that GORGO-proxy overcomes synchronization overhead in previous methods and is 2.5x faster on median TTFT, demonstrating the success of a centralized router.", "authors": ["Alessio Ricci Toniolo", "Abinaya Dinesh", "Rome Thorstenson"], "categories": ["cs.NI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-12", "url": "https://arxiv.org/abs/2602.11688", "pdf_url": "https://arxiv.org/pdf/2602.11688v1", "arxiv_id": "2602.11688", "doi": "10.48550/arXiv.2602.11688", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/atoniolo76/gotoni/tree/benchmark-load-balancing", "venue": "arXiv.org", "quality_score": 0.7367}
{"id": "20aed9382dfe0e743c409c346f4acbc869d5e9215670840d7e51b1309b860ec9", "sources": ["arxiv", "semantic_scholar"], "title": "LAER-MoE: Load-Adaptive Expert Re-layout for Efficient Mixture-of-Experts Training", "abstract": "Expert parallelism is vital for effectively training Mixture-of-Experts (MoE) models, enabling different devices to host distinct experts, with each device processing different input data. However, during expert parallel training, dynamic routing results in significant load imbalance among experts: a handful of overloaded experts hinder overall iteration, emerging as a training bottleneck. In this paper, we introduce LAER-MoE, an efficient MoE training framework. The core of LAER-MoE is a novel parallel paradigm, Fully Sharded Expert Parallel (FSEP), which fully partitions each expert parameter by the number of devices and restores partial experts at expert granularity through All-to-All communication during training. This allows for flexible re-layout of expert parameters during training to enhance load balancing. In particular, we perform fine-grained scheduling of communication operations to minimize communication overhead. Additionally, we develop a load balancing planner to formulate re-layout strategies of experts and routing schemes for tokens during training. We perform experiments on an A100 cluster, and the results indicate that our system achieves up to 1.69x acceleration compared to the current state-of-the-art training systems. Source code available at https://github.com/PKU-DAIR/Hetu-Galvatron/tree/laer-moe.", "authors": ["Xinyi Liu", "Yujie Wang", "Fangcheng Fu", "Xuefeng Xiao", "Huixia Li", "Jiashi Li", "Bin Cui"], "categories": ["cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-12", "url": "https://arxiv.org/abs/2602.11686", "pdf_url": "https://arxiv.org/pdf/2602.11686v1", "arxiv_id": "2602.11686", "doi": "10.1145/3779212.3790180", "citation_count": 4, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/PKU-DAIR/Hetu-Galvatron/tree/laer-moe", "venue": "International Conference on Architectural Support for Programming Languages and Operating Systems", "quality_score": 0.7367}
{"id": "e6418dc0dab9ff2eb91c4f1924d3082c47471e9288149c276e46eab7621f5780", "sources": ["arxiv", "semantic_scholar"], "title": "Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters", "abstract": "We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.", "authors": ["Ailin Huang", "Ang Li", "Aobo Kong", "Bin Wang", "Binxing Jiao", "Bo Dong", "Bojun Wang", "Boyu Chen", "Brian Li", "Buyun Ma", "Chang Su", "Changxin Miao", "Changyi Wan", "Chao Lou", "Chen Hu", "Chen Xu", "Chenfeng Yu", "Chengting Feng", "Chengyuan Yao", "Chunrui Han", "Dan Ma", "Dapeng Shi", "Daxin Jiang", "Dehua Ma", "Deshan Sun", "Di Qi", "Enle Liu", "Fajie Zhang", "Fanqi Wan", "Guanzhe Huang", "Gulin Yan", "Guoliang Cao", "Guopeng Li", "Han Cheng", "Hangyu Guo", "Hanshan Zhang", "Hao Nie", "Haonan Jia", "Haoran Lv", "Hebin Zhou", "Hekun Lv", "Heng Wang", "Heung-Yeung Shum", "Hongbo Huang", "Hongbo Peng", "Hongyu Zhou", "Hongyuan Wang", "Houyong Chen", "Huangxi Zhu", "Huimin Wu", "Huiyong Guo", "Jia Wang", "Jian Zhou", "Jianjian Sun", "Jiaoren Wu", "Jiaran Zhang", "Jiashu Lv", "Jiashuo Liu", "Jiayi Fu", "Jiayu Liu", "Jie Cheng", "Jie Luo", "Jie Yang", "Jie Zhou", "Jieyi Hou", "Jing Bai", "Jingcheng Hu", "Jingjing Xie", "Jingwei Wu", "Jingyang Zhang", "Jishi Zhou", "Junfeng Liu", "Junzhe Lin", "Ka Man Lo", "Kai Liang", "Kaibo Liu", "Kaijun Tan", "Kaiwen Yan", "Kaixiang Li", "Kang An", "Kangheng Lin", "Lei Yang", "Liang Lv", "Liang Zhao", "Liangyu Chen", "Lieyu Shi", "Liguo Tan", "Lin Lin", "Lina Chen", "Luck Ma", "Mengqiang Ren", "Michael Li", "Ming Li", "Mingliang Li", "Mingming Zhang", "Mingrui Chen", "Mitt Huang", "Na Wang", "Peng Liu", "Qi Han", "Qian Zhao", "Qinglin He", "Qinxin Du", "Qiuping Wu", "Quan Sun", "Rongqiu Yang", "Ruihang Miao", "Ruixin Han", "Ruosi Wan", "Ruyan Guo", "Shan Wang", "Shaoliang Pang", "Shaowen Yang", "Shengjie Fan", "Shijie Shang", "Shiliang Yang", "Shiwei Li", "Shuangshuang Tian", "Siqi Liu", "Siye Wu", "Siyu Chen", "Song Yuan", "Tiancheng Cao", "Tianchi Yue", "Tianhao Cheng", "Tianning Li", "Tingdan Luo", "Wang You", "Wei Ji", "Wei Yuan", "Wei Zhang", "Weibo Wu", "Weihao Xie", "Wen Sun", "Wenjin Deng", "Wenzhen Zheng", "Wuxun Xie", "Xiangfeng Wang", "Xiangwen Kong", "Xiangyu Liu", "Xiangyu Zhang", "Xiaobo Yang", "Xiaojia Liu", "Xiaolan Yuan", "Xiaoran Jiao", "Xiaoxiao Ren", "Xiaoyun Zhang", "Xin Li", "Xin Liu", "Xin Wu", "Xing Chen", "Xingping Yang", "Xinran Wang", "Xu Zhao", "Xuan He", "Xuanti Feng", "Xuedan Cai", "Xuqiang Zhou", "Yanbo Yu", "Yang Li", "Yang Xu", "Yanlin Lai", "Yanming Xu", "Yaoyu Wang", "Yeqing Shen", "Yibo Zhu", "Yichen Lv", "Yicheng Cao", "Yifeng Gong", "Yijing Yang", "Yikun Yang", "Yin Zhao", "Yingxiu Zhao", "Yinmin Zhang", "Yitong Zhang", "Yixuan Zhang", "Yiyang Chen", "Yongchi Zhao", "Yongshen Long", "Yongyao Wang", "Yousong Guan", "Yu Zhou", "Yuang Peng", "Yuanhao Ding", "Yuantao Fan", "Yuanwei Lu", "Yuanzhen Yang", "Yuchu Luo", "Yudi Zhao", "Yue Peng", "Yueqiang Lin", "Yufan Lu", "Yuling Zhao", "Yunzhou Ju", "Yurong Zhang", "Yusheng Li", "Yuxiang Yang", "Yuyang Chen", "Yuzhu Cai", "Zejia Weng", "Zetao Hong", "Zexi Li", "Zhe Xie", "Zheng Ge", "Zheng Gong", "Zheng Zeng", "Zhenyi Lu", "Zhewei Huang", "Zhichao Chang", "Zhiguo Huang", "Zhiheng Hu", "Zidong Yang", "Zili Wang", "Ziqi Ren", "Zixin Zhang", "Zixuan Wang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-11", "url": "https://arxiv.org/abs/2602.10604", "pdf_url": "https://arxiv.org/pdf/2602.10604v2", "arxiv_id": "2602.10604", "doi": null, "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3026}
{"id": "7fd563609643de05b2ff35c85eff8cb33f44d41cc6cc44286c74ab3826098985", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-Squeeze: Simple and Effective Post-Tuning and In-Tuning Compression of LoRA Modules", "abstract": "Despite its huge number of variants, standard Low-Rank Adaptation (LoRA) is still a dominant technique for parameter-efficient fine-tuning (PEFT). Nonetheless, it faces persistent challenges, including the pre-selection of an optimal rank and rank-specific hyper-parameters, as well as the deployment complexity of heterogeneous-rank modules and more sophisticated LoRA derivatives. In this work, we introduce LoRA-Squeeze, a simple and efficient methodology that aims to improve standard LoRA learning by changing LoRA module ranks either post-hoc or dynamically during training}. Our approach posits that it is better to first learn an expressive, higher-rank solution and then compress it, rather than learning a constrained, low-rank solution directly. The method involves fine-tuning with a deliberately high(er) source rank, reconstructing or efficiently approximating the reconstruction of the full weight update matrix, and then using Randomized Singular Value Decomposition (RSVD) to create a new, compressed LoRA module at a lower target rank. Extensive experiments across 13 text and 10 vision-language tasks show that post-hoc compression often produces lower-rank adapters that outperform those trained directly at the target rank, especially if a small number of fine-tuning steps at the target rank is allowed. Moreover, a gradual, in-tuning rank annealing variant of LoRA-Squeeze consistently achieves the best LoRA size-performance trade-off.", "authors": ["Ivan Vulić", "Adam Grycner", "Quentin de Laroussilhe", "Jonas Pfeiffer"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-11", "url": "https://arxiv.org/abs/2602.10993", "pdf_url": "https://arxiv.org/pdf/2602.10993v2", "arxiv_id": "2602.10993", "doi": "10.48550/arXiv.2602.10993", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4755}
{"id": "fa15b2ab435ec4fdbaee7a2c790c9955ba3b6f2af9ab9de5ed2375907a71a7c9", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Remote KV Cache Reuse with GPU-native Video Codec", "abstract": "Remote KV cache reuse fetches KV cache for identical contexts from remote storage, avoiding recomputation, accelerating LLM inference. While it excels in high-speed networks, its performance degrades significantly in bandwidth-limited scenarios. Recent studies address this by transmitting KV caches in compressed form, but the associated heavyweight decompression counteracts the KV reuse benefits. In this paper, we propose an efficient and widely deployable remote KV cache reuse solution that leverages GPU-native video codecs. Our system, KVCodec, enables effective KV cache coding with two techniques. The codec-friendly tensor layout compresses the KV cache in a highly compact video format, enabling fast transmission. The efficient KV fetcher orchestrates the transmission, decoding, and restoration of compressed KV caches in an efficient pipelined manner, eliminating resource contention, masking network fluctuations, and achieving minimum time-to-first-token (TTFT). We prototype KVCodec on diverse GPUs from high- to low-end. Experiments reveal that it reduces TTFT by up to 3.51 times while maintaining lossless accuracy, compared to SOTA methods.", "authors": ["Liang Mi", "Weijun Wang", "Jinghan Chen", "Ting Cao", "Haipeng Dai", "Yunxin Liu"], "categories": ["cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-10", "url": "https://arxiv.org/abs/2602.09725", "pdf_url": "https://arxiv.org/pdf/2602.09725v3", "arxiv_id": "2602.09725", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3019}
{"id": "16e331efa97df667084fd7c711121237f215b2155f9edcf770090d225733805e", "sources": ["arxiv", "semantic_scholar"], "title": "Generalizing GNNs with Tokenized Mixture of Experts", "abstract": "Deployed graph neural networks (GNNs) are frozen at deployment yet must fit clean data, generalize under distribution shifts, and remain stable to perturbations. We show that static inference induces a fundamental tradeoff: improving stability requires reducing reliance on shift-sensitive features, leaving an irreducible worst-case generalization floor. Instance-conditional routing can break this ceiling, but is fragile because shifts can mislead routing and perturbations can make routing fluctuate. We capture these effects via two decompositions separating coverage vs selection, and base sensitivity vs fluctuation amplification. Based on these insights, we propose STEM-GNN, a pretrain-then-finetune framework with a mixture-of-experts encoder for diverse computation paths, a vector-quantized token interface to stabilize encoder-to-head signals, and a Lipschitz-regularized head to bound output amplification. Across nine node, link, and graph benchmarks, STEM-GNN achieves a stronger three-way balance, improving robustness to degree/homophily shifts and to feature/edge corruptions while remaining competitive on clean graphs.", "authors": ["Xiaoguang Guo", "Zehong Wang", "Jiazheng Li", "Shawn Spitzel", "Qi Yang", "Kaize Ding", "Jundong Li", "Chuxu Zhang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.09258", "pdf_url": "https://arxiv.org/pdf/2602.09258v1", "arxiv_id": "2602.09258", "doi": "10.48550/arXiv.2602.09258", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "c34b6bc4329d87eb962b3c7a7c6fb2bb6dbcd5180f1c048a533cef1ecddaaa41", "sources": ["arxiv", "semantic_scholar"], "title": "CryptoGen: Secure Transformer Generation with Encrypted KV-Cache Reuse", "abstract": "The widespread deployment of cloud-hosted generative models raises a fundamental challenge: enabling efficient autoregressive generation while preserving the privacy of both user prompts and model parameters in untrusted environments. We address this challenge in a client-server setting where an untrusted server hosts an autoregressive Transformer and the client requires cryptographic protection for both inputs and inference. We present CryptoGen, the first system to enable scalable privacy-preserving neural generation with persistent encrypted key-value (KV) cache reuse. Discriminative-task secure inference systems incur quadratic latency and memory growth when adapted to autoregressive decoding due to the lack of native encrypted KV-cache support. In contrast, CryptoGen achieves near-linear scaling by securely reusing and updating encrypted KV caches throughout generation. CryptoGen integrates homomorphic encryption and secret sharing to support both prefilling and generation. Key techniques include a unified encrypted KV-cache framework, heterogeneous SIMD encodings for different phases, optimized cipher-cipher matrix-matrix and matrix-vector operations, and efficient noise refresh and ciphertext concatenation mechanisms. Evaluation on generative Transformer models trained on WikiText-2, PTB, and LAMBADA shows that for input lengths of 128-512 tokens, CryptoGen achieves 4.4x-7.6x lower per-token latency than state-of-the-art discriminative secure inference systems, while maintaining near-linear latency and memory scaling, with advantages increasing for longer sequences. CryptoGen is released as an open-source library.", "authors": ["Hedong Zhang", "Neusha Javidnia", "Shweta Pardeshi", "Qian Lou", "Farinaz Koushanfar"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08798", "pdf_url": "https://arxiv.org/pdf/2602.08798v2", "arxiv_id": "2602.08798", "doi": "10.48550/arXiv.2602.08798", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7314}
{"id": "6c779e752c5fdb6d8574e4ef6e2f57a37beef189cd2b13482627f3b73c5a8b57", "sources": ["arxiv", "semantic_scholar"], "title": "Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction", "abstract": "Given the quadratic complexity of attention, KV cache eviction is vital to accelerate model inference. Current KV cache eviction methods typically rely on instantaneous heuristic metrics, implicitly assuming that score magnitudes are consistent proxies for importance across all heads. However, this overlooks the heterogeneity in predictive fidelity across attention heads. While certain heads prioritize the instantaneous contribution of tokens, others are dedicated to capturing long-horizon utility. In this paper, we propose that optimal budget allocation should be governed by the marginal utility in preserving long-term semantic information. Building on this insight, we propose LU-KV, a novel framework that formulates head-level budget allocation as a global combinatorial optimization problem to maximize the long-horizon marginal contribution of reserved tokens. To solve this non-convex problem, we employ a convex-hull relaxation and a marginal-utility-based greedy solver, achieving near-optimal solutions. Furthermore, we implement a data-driven offline profiling protocol to facilitate the practical deployment of LU-KV. Evaluations on LongBench and RULER benchmarks demonstrate that LU-KV reduces KV cache size by 80% with minimal performance degradation, while also decreasing inference latency and GPU memory footprint.", "authors": ["Ziyao Tang", "Pengkun Jiao", "Xinhang Chen", "Wei Liu", "Shiyong Li", "Jingjing Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08585", "pdf_url": "https://arxiv.org/pdf/2602.08585v2", "arxiv_id": "2602.08585", "doi": "10.48550/arXiv.2602.08585", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "fb9ad87ff71f7b2053d3c1ff57fcfa8d53cc9cb2ed26bcc73c2a363ea5ac2d42", "sources": ["arxiv", "semantic_scholar"], "title": "Near-Oracle KV Selection via Pre-hoc Sparsity for Long-Context Inference", "abstract": "A core bottleneck in large language model (LLM) inference is the cost of attending over the ever-growing key-value (KV) cache. Although near-oracle top-k KV selection can preserve the quality of dense attention while sharply reducing computation and bandwidth, existing sparse methods generally rely on posterior heuristics, i.e., selectors conditioned on observed attention or proxy scores. Such conditioning introduces posterior bias: it tends to distort true token importance and miss salient tokens, thereby impairing long-range reasoning. To tackle this problem, we propose Pre-hoc Sparsity (PrHS), which selects KV entries before attention scoring and provides explicit accuracy control. Let the attention mass of discarded entries be delta (the dropped mass). Through a marginal-to-mutual-information analysis, we derive an upper bound on the mutual-information loss that depends only on the dropped mass. This relation explains failure modes of posterior heuristics and enables verifiable guarantees by controlling the dropped mass in advance. Within PrHS, we instantiate three orthogonal pre-hoc selectors along the axes of time, depth, and layer. Extensive experiments on LLaMA and Mistral families validate PrHS. Across GSM8K and CoQA, PrHS reduces retrieval overhead by over 90%, achieving 3x higher retrieval sparsity than HShare at matched or better accuracy. It incurs under 1% average degradation on LongBench, lowers attention FLOPs by about 15% versus prior sparse baselines, and yields a 9.9x speedup in attention-operator latency and 2.8x higher throughput on NVIDIA A100-80GB GPUs than the dense baseline.", "authors": ["Yifei Gao", "Lei Wang", "Rong-Cheng Tu", "Qixin Zhang", "Jun Cheng", "Dacheng Tao"], "categories": ["cs.LG", "cs.AI", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08329", "pdf_url": "https://arxiv.org/pdf/2602.08329v1", "arxiv_id": "2602.08329", "doi": "10.48550/arXiv.2602.08329", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "c34b1070147d7e19f837687dc024e9f8f53a60da27cc0443e5443f542cf0bb5b", "sources": ["arxiv", "semantic_scholar"], "title": "ManifoldKV: Training-Free KV Cache Compression via Euclidean Outlier Detection", "abstract": "Long-context inference is constrained by KV-cache memory, which grows linearly with sequence length; KV-cache compression therefore hinges on reliably selecting which past tokens to retain. Most geometry-based eviction methods score keys by cosine similarity to a global centroid, but cosine is scale-invariant and can discard magnitude cues that distinguish semantically salient tokens. We propose ManifoldKV, a training-free scorer that ranks tokens by Euclidean distance to the key centroid, capturing both angular and radial deviations. On the RULER benchmark, ManifoldKV achieves 95.7% accuracy at 4K-16K contexts with 20% compression; matching the best geometric baseline while improving robustness in two regimes where cosine scoring fails. First, on multi-key retrieval, ManifoldKV reduces directional collisions, achieving 92.4% vs KeyDiff's 77.0% (+15.4 points) on 3-key NIAH at 50% compression. Second, to address dilution and performance collapse of global centroids at 64K context, we introduce WindowedManifoldKV, which restores accuracy to 84.3% at 25% compression, a 49-point recovery over global L2 and +3.2 points over KeyDiff. The method requires only 3 lines of code and works across 4 architectures without tuning.", "authors": ["Debajyoti Datta", "Trishala Neeraj", "Bibek Paudel", "Vyom Sharma", "Subhabrata Mukherjee"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08343", "pdf_url": "https://arxiv.org/pdf/2602.08343v1", "arxiv_id": "2602.08343", "doi": "10.48550/arXiv.2602.08343", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "e447941df7ca40953846142ff0d92c01db16d9289dbdf379d79114f248ca55bf", "sources": ["arxiv", "semantic_scholar"], "title": "FlexMoRE: A Flexible Mixture of Rank-heterogeneous Experts for Efficient Federatedly-trained Large Language Models", "abstract": "Recent advances in mixture-of-experts architectures have shown that individual experts models can be trained federatedly, i.e., in isolation from other experts by using a common base model to facilitate coordination. However, we hypothesize that full-sized experts may not be necessary for all domains and that instead low-rank adapters may be sufficient. Here, we introduce FlexMoRE, a Flexible Mixture of Rank-heterogenous Experts, which may be either full-sized experts or adapters of a suitable rank. We systematically investigate the trade-off between expert rank and downstream task performance by evaluating $6$ experts with ranks $2^0$ to $2^{14}$ resulting in experiments covering 150 mixtures (96 with 2 experts, 54 with 7 experts) that are evaluated across $120$ tasks. For our experiments, we build on FlexOlmo and turn its pre-trained experts into low-rank versions. Our regression analysis from expert rank to downstream task performance reveals that the best-performing rank is substantially higher for reasoning-heavy benchmarks than for knowledge-heavy benchmarks. These findings on rank sensitivity come with direct implications for memory efficiency: Using optimal ranks, FlexMoRE yields improved downstream task performance (average score $47.18$) compared to the baseline FlexOlmo-style mixture of full-sized experts (average score $45.46$) at less than one third the parameters ($10.75$B for FlexMoRE vs. $33.27$B for FlexOlmo). All code will be made available.", "authors": ["Annemette Brok Pirchert", "Jacob Nielsen", "Mogens Henrik From", "Lukas Galke Poech", "Peter Schneider-Kamp"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08818", "pdf_url": "https://arxiv.org/pdf/2602.08818v1", "arxiv_id": "2602.08818", "doi": "10.48550/arXiv.2602.08818", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "66cf0036aa7b4d06c163b26e7eda4f5bedc1241ad6f9f94c8e27f737fc8b9017", "sources": ["arxiv", "semantic_scholar"], "title": "UniComp: A Unified Evaluation of Large Language Model Compression via Pruning, Quantization and Distillation", "abstract": "Model compression is increasingly essential for deploying large language models (LLMs), yet existing comparative studies largely focus on pruning and quantization evaluated primarily on knowledge-centric benchmarks. Thus, we introduce UniComp, a unified evaluation framework for comparing pruning, quantization, and knowledge distillation. UniComp evaluates compressed models along three dimensions: performance, reliability, and efficiency, using a diverse set of capability- and safety-oriented benchmarks together with a hardware-aware efficiency analysis. Through evaluation of six compression techniques across 40 datasets, we observe (i) a consistent knowledge bias, where factual recall is largely preserved while multi-step reasoning, multilingual, and instruction-following capabilities degrade; (ii) a decoupling between performance and reliability, indicating that retained performance does not consistently imply preserved reliability; and (iii) that task-specific calibration can yield up to 50% relative improvement of reasoning performance in pruned models.", "authors": ["Jonathan von Rad", "Yong Cao", "Andreas Geiger"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.09130", "pdf_url": "https://arxiv.org/pdf/2602.09130v5", "arxiv_id": "2602.09130", "doi": "10.48550/arXiv.2602.09130", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "60caa7ff888c963b5cb9940667dd86305a49741634dea6d384d35c6274e9c897", "sources": ["arxiv", "semantic_scholar"], "title": "DirMoE: Dirichlet-routed Mixture of Experts", "abstract": "Mixture-of-Experts (MoE) models have demonstrated exceptional performance in large-scale language models. Existing routers typically rely on non-differentiable Top-$k$+Softmax, limiting their performance and scalability. We argue that two distinct decisions, which experts to activate and how to distribute expert contributions among them, are conflated in standard Top-$k$+Softmax. We introduce Dirichlet-Routed MoE (DirMoE), a novel end-to-end differentiable routing mechanism built on a Dirichlet variational autoencoder framework. This design fundamentally disentangles the core routing problems: expert selection, modeled by a Bernoulli component, and expert contribution among chosen experts, handled by a Dirichlet component. The entire forward pass remains fully differentiable through the use of Gumbel-Sigmoid relaxation for the expert selection and implicit reparameterization for the Dirichlet distribution. Our training objective, a variational ELBO, includes a direct sparsity penalty that precisely controls the number of active experts in expectation, alongside a schedule for key hyperparameters that guides the model from an exploratory to a definitive routing state. Moreover, our DirMoE router matches or exceeds other methods while improving expert specialization.", "authors": ["Amirhossein Vahidi", "Hesam Asadollahzadeh", "Navid Akhavan Attar", "Marie Moullet", "Kevin Ly", "Xingyi Yang", "Mohammad Lotfollahi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.09001", "pdf_url": "https://arxiv.org/pdf/2602.09001v1", "arxiv_id": "2602.09001", "doi": "10.48550/arXiv.2602.09001", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "d977098d946caca03317032d4a8c37c79394480d07f46da2e4698dcfef4d32f1", "sources": ["arxiv", "semantic_scholar"], "title": "Large Language Lobotomy: Jailbreaking Mixture-of-Experts via Expert Silencing", "abstract": "The rapid adoption of Mixture-of-Experts (MoE) architectures marks a major shift in the deployment of Large Language Models (LLMs). MoE LLMs improve scaling efficiency by activating only a small subset of parameters per token, but their routing structure introduces new safety attack surfaces. We find that safety-critical behaviors in MoE LLMs (e.g., refusal) are concentrated in a small set of experts rather than being uniformly distributed. Building on this, we propose Large Language Lobotomy (L$^3$), a training-free, architecture-agnostic attack that compromises safety alignment by exploiting expert routing dynamics. L$^3$ learns routing patterns that correlate with refusal, attributes safety behavior to specific experts, and adaptively silences the most safety-relevant experts until harmful outputs are produced. We evaluate L$^3$ on eight state-of-the-art open-source MoE LLMs and show that our adaptive expert silencing increases average attack success from 7.3% to 70.4%, reaching up to 86.3%, outperforming prior training-free MoE jailbreak methods. Moreover, bypassing guardrails typically requires silencing fewer than 20% of layer-wise experts while largely preserving general language utility. These results reveal a fundamental tension between efficiency-driven MoE design and robust safety alignment and motivate distributing safety mechanisms more robustly in future MoE LLMs with architecture- and routing-aware methods.", "authors": ["Jona te Lintelo", "Lichao Wu", "Stjepan Picek"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08741", "pdf_url": "https://arxiv.org/pdf/2602.08741v1", "arxiv_id": "2602.08741", "doi": "10.48550/arXiv.2602.08741", "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7314}
{"id": "32f21f6afb68f57a35742f30f233e76580ac5e9d3ca7c32e505e15c6e01775a2", "sources": ["arxiv", "semantic_scholar"], "title": "Attn-QAT: 4-Bit Attention With Quantization-Aware Training", "abstract": "Achieving reliable 4-bit attention is a prerequisite for end-to-end FP4 computation on emerging FP4-capable GPUs, yet attention remains the main obstacle due to FP4's tiny dynamic range and attention's heavy-tailed activations. This paper presents the first systematic study of 4-bit quantization-aware training (QAT) for attention. We find that \"drop-in\" QAT, which naively combines an FP4 forward pass with a high-precision Flash Attention (FA)-style backward pass, leads to training instability. We identify two key principles for stable FP4 attention: (1) matching low-precision recomputation of attention scores in the backward pass, and (2) resolving implicit precision assumptions in FA's gradient calculation. Based on these insights, we propose Attn-QAT and implement fused Triton kernels for training as well as FP4 inference kernels. Across diffusion and language models, Attn-QAT recovers the quality drop from FP4 attention without explicit outlier-mitigation heuristics used in prior FP4 attention, and delivers up to a 1.5x speedup on an RTX 5090. Video demos can be found at https://drive.google.com/drive/folders/190F6xbBDUF2kGQYIcXBt3ehSYij5jlim?usp=sharing.", "authors": ["Peiyuan Zhang", "Matthew Noto", "Wenxuan Tan", "Chengquan Jiang", "Will Lin", "Wei Zhou", "Hao Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2603.00040", "pdf_url": "https://arxiv.org/pdf/2603.00040v2", "arxiv_id": "2603.00040", "doi": "10.48550/arXiv.2603.00040", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4732}
{"id": "59ab4754ade8f19d0870d50b7b03e8d39784c5d624a3f42f7cbc15765cfd0edd", "sources": ["arxiv", "semantic_scholar"], "title": "DeltaKV: Residual-Based KV Cache Compression via Long-Range Similarity", "abstract": "The deployment of efficient long-context LLMs in applications like autonomous agents, long-chain reasoning, and creative writing is fundamentally bottlenecked by the linear growth of KV cache memory. Existing compression and eviction methods often struggle to balance accuracy, compression ratio, and hardware efficiency. We propose DeltaKV, a residual-based KV cache compression framework motivated by two empirical findings: long-range inter-token similarity and highly shared latent components in KV representations. Instead of discarding tokens, DeltaKV encodes semantic residuals relative to retrieved historical references, preserving fidelity while substantially reducing storage. To translate compression gains into real system speedups, we further introduce Sparse-vLLM, a high-performance inference engine with decoupled memory management and kernels optimized for sparse and irregular KV layouts. Experiments show that DeltaKV reduces KV cache memory to 29\\% of the original while maintaining near-lossless accuracy on LongBench, SCBench, and AIME. When integrated with Sparse-vLLM, it achieves up to 2$\\times$ throughput improvement over vLLM in long-context scenarios, demonstrating a practical path toward scalable long-context LLM deployment. Code, model checkpoints, and datasets are available at https://github.com/CURRENTF/Sparse-vLLM.", "authors": ["Jitai Hao", "Qiang Huang", "Yaowei Wang", "Min Zhang", "Jun Yu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-08", "url": "https://arxiv.org/abs/2602.08005", "pdf_url": "https://arxiv.org/pdf/2602.08005v1", "arxiv_id": "2602.08005", "doi": "10.48550/arXiv.2602.08005", "citation_count": 4, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/CURRENTF/Sparse-vLLM", "venue": "arXiv.org", "quality_score": 0.7296}
{"id": "37171c6771373c1732548b1dd235d543e42497c01bd0a7680af9793241f8c227", "sources": ["arxiv", "semantic_scholar"], "title": "ParisKV: Fast and Drift-Robust KV-Cache Retrieval for Long-Context LLMs", "abstract": "KV-cache retrieval is essential for long-context LLM inference, yet existing methods struggle with distribution drift and high latency at scale. We introduce ParisKV, a drift-robust, GPU-native KV-cache retrieval framework based on collision-based candidate selection, followed by a quantized inner-product reranking estimator. For million-token contexts, ParisKV supports CPU-offloaded KV caches via Unified Virtual Addressing (UVA), enabling on-demand top-$k$ fetching with minimal overhead. ParisKV matches or outperforms full attention quality on long-input and long-generation benchmarks. It achieves state-of-the-art long-context decoding efficiency: it matches or exceeds full attention speed even at batch size 1 for long contexts, delivers up to 2.8$\\times$ higher throughput within full attention's runnable range, and scales to million-token contexts where full attention runs out of memory. At million-token scale, ParisKV reduces decode latency by 17$\\times$ and 44$\\times$ compared to MagicPIG and PQCache, respectively, two state-of-the-art KV-cache Top-$k$ retrieval baselines, code is available at https://github.com/amy-77/ParisKV/tree/main.", "authors": ["Yanlin Qi", "Xinhang Chen", "Huiqiang Jiang", "Qitong Wang", "Botao Peng", "Themis Palpanas"], "categories": ["cs.LG", "cs.CL", "cs.DB"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-07", "url": "https://arxiv.org/abs/2602.07721", "pdf_url": "https://arxiv.org/pdf/2602.07721v3", "arxiv_id": "2602.07721", "doi": "10.48550/arXiv.2602.07721", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/amy-77/ParisKV/tree/main", "venue": "arXiv.org", "quality_score": 0.7278}
{"id": "49f82c4226ab6af88debd9f4c6982e01476b874f4620cd6179e48e1acd632262", "sources": ["arxiv", "semantic_scholar"], "title": "Layer-wise LoRA fine-tuning: a similarity metric approach", "abstract": "Pre-training Large Language Models (LLMs) on web-scale datasets becomes fundamental for advancing general-purpose AI. In contrast, enhancing their predictive performance on downstream tasks typically involves adapting their knowledge through fine-tuning. Parameter-efficient fine-tuning techniques, such as Low-Rank Adaptation (LoRA), aim to reduce the computational cost of this process by freezing the pre-trained model and updating a smaller number of parameters. In comparison to full fine-tuning, these methods achieve over 99\\% reduction in trainable parameter count, depending on the configuration. Unfortunately, such a reduction may prove insufficient as LLMs continue to grow in scale. In this work, we address the previous problem by systematically selecting only a few layers to fine-tune using LoRA or its variants. We argue that not all layers contribute equally to the model adaptation. Leveraging this, we identify the most relevant layers to fine-tune by measuring their contribution to changes in internal representations. Our method is orthogonal to and readily compatible with existing low-rank adaptation techniques. We reduce the trainable parameters in LoRA-based techniques by up to 50\\%, while maintaining the predictive performance across different models and tasks. Specifically, on encoder-only architectures, this reduction in trainable parameters leads to a negligible predictive performance drop on the GLUE benchmark. On decoder-only architectures, we achieve a small drop or even improvements in the predictive performance on mathematical problem-solving capabilities and coding tasks. Finally, this effectiveness extends to multimodal models, for which we also observe competitive results relative to fine-tuning with LoRA modules in all layers. Code is available at: https://github.com/c2d-usp/Layer-wise-LoRA-with-CKA", "authors": ["Keith Ando Ogawa", "Bruno Lopes Yamamoto", "Lucas Lauton de Alcantara", "Lucas Pellicer", "Rosimeire Pereira Costa", "Edson Bollis", "Anna Helena Reali Costa", "Artur Jordao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-05", "url": "https://arxiv.org/abs/2602.05988", "pdf_url": "https://arxiv.org/pdf/2602.05988v1", "arxiv_id": "2602.05988", "doi": "10.48550/arXiv.2602.05988", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/c2d-usp/Layer-wise-LoRA-with-CKA", "venue": "arXiv.org", "quality_score": 0.7243}
{"id": "a7f6d943eea437ed52458866b0ed5caf86439a39a5c21a367f664ef2b1abdc61", "sources": ["arxiv", "semantic_scholar"], "title": "KV-CoRE: Benchmarking Data-Dependent Low-Rank Compressibility of KV-Caches in LLMs", "abstract": "Large language models rely on kv-caches to avoid redundant computation during autoregressive decoding, but as context length grows, reading and writing the cache can quickly saturate GPU memory bandwidth. Recent work has explored KV-cache compression, yet most approaches neglect the data-dependent nature of kv-caches and their variation across layers. We introduce KV-CoRE KV-cache Compressibility by Rank Evaluation), an SVD-based method for quantifying the data-dependent low-rank compressibility of kv-caches. KV-CoRE computes the optimal low-rank approximation under the Frobenius norm and, being gradient-free and incremental, enables efficient dataset-level, layer-wise evaluation. Using this method, we analyze multiple models and datasets spanning five English domains and sixteen languages, uncovering systematic patterns that link compressibility to model architecture, training data, and language coverage. As part of this analysis, we employ the Normalized Effective Rank as a metric of compressibility and show that it correlates strongly with performance degradation under compression. Our study establishes a principled evaluation framework and the first large-scale benchmark of kv-cache compressibility in LLMs, offering insights for dynamic, data-aware compression and data-centric model development.", "authors": ["Jian Chen", "Zhuoran Wang", "Jiayu Qin", "Ming Li", "Meng Wang", "Changyou Chen", "Yin Chen", "Qizhen Weng", "Yirui Liu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-05", "url": "https://arxiv.org/abs/2602.05929", "pdf_url": "https://arxiv.org/pdf/2602.05929v2", "arxiv_id": "2602.05929", "doi": "10.48550/arXiv.2602.05929", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4686}
{"id": "bdfaab75f8cde307e2f5f35fc9b370bfb47f48fb64da9f0c46f4f0b0bfd7608e", "sources": ["arxiv", "semantic_scholar"], "title": "FlashBlock: Attention Caching for Efficient Long-Context Block Diffusion", "abstract": "Generating long-form content, such as minute-long videos and extended texts, is increasingly important for modern generative models. Block diffusion improves inference efficiency via KV caching and block-wise causal inference and has been widely adopted in diffusion language models and video generation. However, in long-context settings, block diffusion still incurs substantial overhead from repeatedly computing attention over a growing KV cache. We identify an underexplored property of block diffusion: cross-step redundancy of attention within a block. Our analysis shows that attention outputs from tokens outside the current block remain largely stable across diffusion steps, while block-internal attention varies significantly. Based on this observation, we propose FlashBlock, a cached block-external attention mechanism that reuses stable attention output, reducing attention computation and KV cache access without modifying the diffusion process. Moreover, FlashBlock is orthogonal to sparse attention and can be combined as a complementary residual reuse strategy, substantially improving model accuracy under aggressive sparsification. Experiments on diffusion language models and video generation demonstrate up to 1.44$\\times$ higher token throughput and up to 1.6$\\times$ reduction in attention time, with negligible impact on generation quality. Project page: https://caesarhhh.github.io/FlashBlock/.", "authors": ["Zhuokun Chen", "Jianfei Cai", "Bohan Zhuang"], "categories": ["cs.CV", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-05", "url": "https://arxiv.org/abs/2602.05305", "pdf_url": "https://arxiv.org/pdf/2602.05305v2", "arxiv_id": "2602.05305", "doi": "10.48550/arXiv.2602.05305", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4686}
{"id": "e37b1309e72586ee98e4ea2aeb8f671b8267a0e7fc932c34f5ba3d4f13dde1f0", "sources": ["arxiv", "semantic_scholar"], "title": "MoSE: Mixture of Slimmable Experts for Efficient and Adaptive Language Models", "abstract": "Mixture-of-Experts (MoE) models scale large language models efficiently by sparsely activating experts, but once an expert is selected, it is executed fully. Hence, the trade-off between accuracy and computation in an MoE model typically exhibits large discontinuities. We propose Mixture of Slimmable Experts (MoSE), an MoE architecture in which each expert has a nested, slimmable structure that can be executed at variable widths. This enables conditional computation not only over which experts are activated, but also over how much of each expert is utilized. Consequently, a single pretrained MoSE model can support a more continuous spectrum of accuracy-compute trade-offs at inference time. We present a simple and stable training recipe for slimmable experts under sparse routing, combining multi-width training with standard MoE objectives. During inference, we explore strategies for runtime width determination, including a lightweight test-time training mechanism that learns how to map router confidence/probabilities to expert widths under a fixed budget. Experiments on GPT models trained on OpenWebText demonstrate that MoSE matches or improves upon standard MoE at full width and consistently shifts the Pareto frontier for accuracy vs. cost, achieving comparable performance with significantly fewer FLOPs.", "authors": ["Nurbek Tastan", "Stefanos Laskaridis", "Karthik Nandakumar", "Samuel Horvath"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-05", "url": "https://arxiv.org/abs/2602.06154", "pdf_url": "https://arxiv.org/pdf/2602.06154v1", "arxiv_id": "2602.06154", "doi": "10.48550/arXiv.2602.06154", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4686}
{"id": "cbf6ca9cd3be63f4ec6ac68fefc51c57ec487fdc9cb6914032401878afc13579", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning", "abstract": "Low-Rank Adaptation (LoRA) is the prevailing approach for efficient large language model (LLM) fine-tuning. Building on this paradigm, recent studies have proposed alternative initialization strategies, architectural modifications, and optimization adjustments, reporting substantial improvements over vanilla LoRA. However, these gains are often demonstrated under fixed or narrowly tuned hyperparameter settings, despite the known sensitivity of neural networks to training configurations. In this work, we systematically re-evaluate nine representative LoRA variants alongside vanilla LoRA through extensive hyperparameter searches over learning rate, batch size, rank, and training duration. Across tasks spanning mathematical reasoning, commonsense reasoning, code generation, and instruction following at diverse model scales, we find that different LoRA methods favor distinct learning rate ranges. Crucially, once learning rates are properly tuned, all methods achieve similar peak performance (within 1-2%), with only subtle rank-dependent behaviors. These results suggest that vanilla LoRA remains a competitive baseline and that improvements reported under a single training configuration may not reflect consistent methodological advantages. Finally, a second-order analysis attributes the differing optimal learning rate ranges to variations in the largest Hessian eigenvalue, aligning with classical learning theories.", "authors": ["Yu-Ang Lee", "Ching-Yun Ko", "Pin-Yu Chen", "Mi-Yen Yeh"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-04", "url": "https://arxiv.org/abs/2602.04998", "pdf_url": "https://arxiv.org/pdf/2602.04998v2", "arxiv_id": "2602.04998", "doi": "10.48550/arXiv.2602.04998", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/yuang-lee/lr-matters-lora", "venue": "arXiv.org", "quality_score": 0.7225}
{"id": "6eccdd70c6e1039f6e0f367c9a702c586f2730fb2a32efd944d11dce37810227", "sources": ["arxiv", "semantic_scholar"], "title": "Greedy-Gnorm: A Gradient Matrix Norm-Based Alternative to Attention Entropy for Head Pruning", "abstract": "Attention head pruning has emerged as an effective technique for transformer model compression, an increasingly important goal in the era of Green AI. However, existing pruning methods often rely on static importance scores, which fail to capture the evolving role of attention heads during iterative removal. We propose Greedy-Gradient norm (Greedy-Gnorm), a novel head pruning algorithm that dynamically recalculates head importance after each pruning step. Specifically, each head is scored by the elementwise product of the l2-norms of its Q/K/V gradient blocks, as estimated from a hold-out validation set and updated at every greedy iteration. This dynamic approach to scoring mitigates against stale rankings and better reflects gradient-informed importance as pruning progresses. Extensive experiments on BERT, ALBERT, RoBERTa, and XLM-RoBERTa demonstrate that Greedy-Gnorm consistently preserves accuracy under substantial head removal, outperforming attention entropy. By effectively reducing model size while maintaining task performance, Greedy-Gnorm offers a promising step toward more energy-efficient transformer model deployment.", "authors": ["Yuxi Guo", "Paul Sheridan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-04", "url": "https://arxiv.org/abs/2602.04491", "pdf_url": "https://arxiv.org/pdf/2602.04491v1", "arxiv_id": "2602.04491", "doi": "10.48550/arXiv.2602.04491", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4675}
{"id": "aece6c0cfe8674dbc23c441fd17884797dc83655b8233e0881aa0af0df7e2fdc", "sources": ["arxiv", "semantic_scholar"], "title": "RASA: Routing-Aware Safety Alignment for Mixture-of-Experts Models", "abstract": "Mixture-of-Experts (MoE) language models introduce unique challenges for safety alignment due to their sparse routing mechanisms, which can enable degenerate optimization behaviors under standard full-parameter fine-tuning. In our preliminary experiments, we observe that naively applying full-parameter safety fine-tuning to MoE models can reduce attack success rates through routing or expert dominance effects, rather than by directly repairing Safety-Critical Experts. To address this challenge, we propose RASA, a routing-aware expert-level alignment framework that explicitly repairs Safety-Critical Experts while preventing routing-based bypasses. RASA identifies experts disproportionately activated by successful jailbreaks, selectively fine-tunes only these experts under fixed routing, and subsequently enforces routing consistency with safety-aligned contexts. Across two representative MoE architectures and a diverse set of jailbreak attacks, RASA achieves near-perfect robustness, strong cross-attack generalization, and substantially reduced over-refusal, while preserving general capabilities on benchmarks such as MMLU, GSM8K, and TruthfulQA. Our results suggest that robust MoE safety alignment benefits from targeted expert repair rather than global parameter updates, offering a practical and architecture-preserving alternative to prior approaches.", "authors": ["Jiacheng Liang", "Yuhui Wang", "Tanqiu Jiang", "Ting Wang"], "categories": ["cs.LG", "cs.AI", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-04", "url": "https://arxiv.org/abs/2602.04448", "pdf_url": "https://arxiv.org/pdf/2602.04448v2", "arxiv_id": "2602.04448", "doi": "10.48550/arXiv.2602.04448", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4675}
{"id": "c9f4c2a325a89308bfcff78bdaf433dae2c72c81cad663963a4485d9767ed9ac", "sources": ["arxiv", "semantic_scholar"], "title": "TurboBoA: Faster and Exact Attention-aware Quantization without Backpropagation", "abstract": "The rapid growth of large language models (LLMs) has heightened the importance of post-training quantization (PTQ) for reducing memory and computation costs. Among PTQ methods, GPTQ has gained significant attention for its efficiency, enabling billion-scale LLMs to be quantized within a few GPU hours. However, GPTQ's assumption of layer-wise independence leads to severe accuracy drops in low-bit regimes. Recently, BoA improved upon GPTQ by incorporating inter-layer dependencies within attention modules, but its reliance on sequential quantization across all out-channels makes it substantially less efficient. In this paper, we propose TurboBoA, a new backpropagation-free PTQ algorithm that preserves the accuracy benefits of BoA while significantly accelerating the process. The proposed TurboBoA introduces three key innovations: (i) joint quantization of multiple out-channels with a closed-form error compensation rule, which reduces sequential bottlenecks and yields more than a three-fold speedup; (ii) a correction mechanism for errors propagated from preceding quantized layers; and (iii) adaptive grid computation with coordinate descent refinement to maintain alignment during iterative updates. Extensive experiments demonstrate that TurboBoA delivers substantial acceleration over BoA while consistently improving accuracy. When combined with outlier suppression techniques, it achieves state-of-the-art results in both weight-only and weight-activation quantization. The code will be available at https://github.com/SamsungLabs/TurboBoA.", "authors": ["Junhan Kim", "Yeo Jeong Park", "Seungwoo Son", "Chungman Lee", "Ho-young Kim", "Joonyoung Kim", "Yongkweon Jeon"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-04", "url": "https://arxiv.org/abs/2602.04929", "pdf_url": "https://arxiv.org/pdf/2602.04929v1", "arxiv_id": "2602.04929", "doi": "10.48550/arXiv.2602.04929", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SamsungLabs/TurboBoA", "venue": "arXiv.org", "quality_score": 0.7225}
{"id": "a7b4364ea8b456d5ae7bef63e9dded5e368149f75493101724d4922fcf3e5072", "sources": ["arxiv", "semantic_scholar"], "title": "Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization", "abstract": "Despite rapid progress in autoregressive video diffusion, an emerging system algorithm bottleneck limits both deployability and generation capability: KV cache memory. In autoregressive video generation models, the KV cache grows with generation history and quickly dominates GPU memory, often exceeding 30 GB, preventing deployment on widely available hardware. More critically, constrained KV cache budgets restrict the effective working memory, directly degrading long horizon consistency in identity, layout, and motion. To address this challenge, we present Quant VideoGen (QVG), a training free KV cache quantization framework for autoregressive video diffusion models. QVG leverages video spatiotemporal redundancy through Semantic Aware Smoothing, producing low magnitude, quantization friendly residuals. It further introduces Progressive Residual Quantization, a coarse to fine multi stage scheme that reduces quantization error while enabling a smooth quality memory trade off. Across LongCat Video, HY WorldPlay, and Self Forcing benchmarks, QVG establishes a new Pareto frontier between quality and memory efficiency, reducing KV cache memory by up to 7.0 times with less than 4% end to end latency overhead while consistently outperforming existing baselines in generation quality. Code is available at: https://github.com/svg-project/Quant-VideoGen", "authors": ["Haocheng Xi", "Shuo Yang", "Yilong Zhao", "Muyang Li", "Han Cai", "Xingyang Li", "Yujun Lin", "Zhuoyang Zhang", "Jintao Zhang", "Xiuyu Li", "Zhiying Xu", "Jun Wu", "Chenfeng Xu", "Ion Stoica", "Song Han", "Kurt Keutzer"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-03", "url": "https://arxiv.org/abs/2602.02958", "pdf_url": "https://arxiv.org/pdf/2602.02958v5", "arxiv_id": "2602.02958", "doi": "10.48550/arXiv.2602.02958", "citation_count": 11, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/svg-project/Quant-VideoGen", "venue": "arXiv.org", "quality_score": 0.7207}
{"id": "d52403bc879463ef032e432cc4783c0f7aa26e951b5946eec117c4c5750d2c7e", "sources": ["arxiv", "semantic_scholar"], "title": "HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing", "abstract": "This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer's token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.", "authors": ["Yizhao Gao", "Jianyu Wei", "Qihao Zhang", "Yu Cheng", "Shimao Chen", "Zhengju Tang", "Zihan Jiang", "Yifan Song", "Hailin Zhang", "Liang Zhao", "Bo Yang", "Gang Wang", "Shijie Cao", "Fuli Luo"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-03", "url": "https://arxiv.org/abs/2602.03560", "pdf_url": "https://arxiv.org/pdf/2602.03560v1", "arxiv_id": "2602.03560", "doi": "10.48550/arXiv.2602.03560", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4664}
{"id": "fdf3e0aae805d9690a058b68b6614412638c5695528f879f4bd10fae61587b4a", "sources": ["arxiv", "semantic_scholar"], "title": "ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contribution", "abstract": "Recently, large language models (LLMs) have shown remarkable reasoning abilities by producing long reasoning traces. However, as the sequence length grows, the key-value (KV) cache expands linearly, incurring significant memory and computation costs. Existing KV cache eviction methods mitigate this issue by discarding less important KV pairs, but often fail to capture complex KV dependencies, resulting in performance degradation. To better balance efficiency and performance, we introduce ForesightKV, a training-based KV cache eviction framework that learns to predict which KV pairs to evict during long-text generations. We first design the Golden Eviction algorithm, which identifies the optimal eviction KV pairs at each step using future attention scores. These traces and the scores at each step are then distilled via supervised training with a Pairwise Ranking Loss. Furthermore, we formulate cache eviction as a Markov Decision Process and apply the GRPO algorithm to mitigate the significant language modeling loss increase on low-entropy tokens. Experiments on AIME2024 and AIME2025 benchmarks of three reasoning models demonstrate that ForesightKV consistently outperforms prior methods under only half the cache budget, while benefiting synergistically from both supervised and reinforcement learning approaches. Code is available at https://github.com/RUCAIBox/ForesightKV.", "authors": ["Zican Dong", "Peiyu Liu", "Junyi Li", "Zhipeng Chen", "Han Peng", "Shuo Wang", "Wayne Xin Zhao"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-03", "url": "https://arxiv.org/abs/2602.03203", "pdf_url": "https://arxiv.org/pdf/2602.03203v2", "arxiv_id": "2602.03203", "doi": "10.48550/arXiv.2602.03203", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/RUCAIBox/ForesightKV", "venue": "arXiv.org", "quality_score": 0.7207}
{"id": "dff1e071b0fb03b53d9be7e510fff27158f3151a76f792b01fd4ef3ff84f3894", "sources": ["arxiv"], "title": "One Size Does Not Fit All: Token-Wise Adaptive Compression for KV Cache", "abstract": "Despite the remarkable progress of Large Language Models (LLMs), the escalating memory footprint of the Key-Value (KV) cache remains a critical bottleneck for efficient inference. While dimensionality reduction offers a promising compression avenue, existing approaches typically either necessitate prohibitively expensive pre-training from scratch or suffer from severe performance deterioration under high compression regimes. In this work, we propose DynaKV, a novel post-training framework for low-rank KV cache compression. To the best of our knowledge, DynaKV is the first method to dynamically allocate compression rates to individual tokens according to their semantic meaning, which allows it to achieve better fidelity at aggressive compression ratios. Extensive experiments demonstrate that our method consistently outperforms existing state-of-the-art compression techniques, achieving significant memory reduction while maintaining competitive generation quality. Furthermore, our approach is orthogonal to sequence-level pruning methods. When integrated with SnapKV, DynaKV retains only 6% of the KV cache while maintaining 94% of the baseline performance on the LongBench benchmark.", "authors": ["Liming Lu", "Kaixi Qiu", "Jiayu Zhou", "Jushi Kai", "Haoyan Zhang", "Huanyu Wang", "Jingwen Leng", "Ziwei He", "Zhouhan Lin"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": [], "published_date": "2026-02-03", "url": "https://arxiv.org/abs/2603.04411", "pdf_url": "https://arxiv.org/pdf/2603.04411v1", "arxiv_id": "2603.04411", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2968}
{"id": "67ba8bfbf171a62c17180aed3c1c6e87dbda6cd7aa612eda17b5004ded0f7924", "sources": ["arxiv", "semantic_scholar"], "title": "Online Vector Quantized Attention", "abstract": "Standard sequence mixing layers used in language models struggle to balance efficiency and performance. Self-attention performs well on long context tasks but has expensive quadratic compute and linear memory costs, while linear attention and SSMs use only linear compute and constant memory but struggle with long context processing. In this paper, we develop a sequence mixing layer that aims to find a better compromise between memory-compute costs and long-context processing, which we call online vector-quantized (OVQ) attention. OVQ-attention requires linear compute costs and constant memory, but, unlike linear attention and SSMs, it uses a sparse memory update that allows it to greatly increase the size of its memory state and, consequently, memory capacity. We develop a theoretical basis for OVQ-attention based on Gaussian mixture regression, and we test it on a variety of synthetic long context tasks and on long context language modeling. OVQ-attention shows significant improvements over linear attention baselines and the original VQ-attention, on which OVQ-attention was inspired. It demonstrates competitive, and sometimes identical, performance to strong self-attention baselines up 64k sequence length, despite using a small fraction of the memory of full self-attention.", "authors": ["Nick Alonso", "Tomas Figliolia", "Beren Millidge"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-03", "url": "https://arxiv.org/abs/2602.03922", "pdf_url": "https://arxiv.org/pdf/2602.03922v3", "arxiv_id": "2602.03922", "doi": "10.48550/arXiv.2602.03922", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4664}
{"id": "9733e6c198a3764484daf313384b870c3df6ae896e3aafde02e52af20c715fac", "sources": ["arxiv", "semantic_scholar"], "title": "Hierarchical Adaptive Eviction for KV Cache Management in Multimodal Language Models", "abstract": "The integration of visual information into Large Language Models (LLMs) has enabled Multimodal LLMs (MLLMs), but the quadratic memory and computational costs of Transformer architectures remain a bottleneck. Existing KV cache eviction strategies fail to address the heterogeneous attention distributions between visual and text tokens, leading to suboptimal efficiency or degraded performance. In this paper, we propose Hierarchical Adaptive Eviction (HAE), a KV cache eviction framework that optimizes text-visual token interaction in MLLMs by implementing Dual-Attention Pruning during pre-filling (leveraging visual token sparsity and attention variance) and a Dynamic Decoding Eviction Strategy (inspired by OS Recycle Bins) during decoding. HAE minimizes KV cache usage across layers, reduces computational overhead via index broadcasting, and theoretically ensures superior information integrity and lower error bounds compared to greedy strategies, enhancing efficiency in both comprehension and generation tasks. Empirically, HAE reduces KV-Cache memory by 41\\% with minimal accuracy loss (0.3\\% drop) in image understanding tasks and accelerates story generation inference by 1.5x while maintaining output quality on Phi3.5-Vision-Instruct model.", "authors": ["Xindian Ma", "Yidi Lu", "Peng Zhang", "Jing Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.02197", "pdf_url": "https://arxiv.org/pdf/2602.02197v1", "arxiv_id": "2602.02197", "doi": "10.48550/arXiv.2602.02197", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4652}
{"id": "6ff0eb0536e91ae9b9aa04f7a19c35bc24bdc744d2e93c18f06eacb51ca251ca", "sources": ["arxiv", "semantic_scholar"], "title": "Q Cache: Visual Attention is Valuable in Less than Half of Decode Layers for Multimodal Large Language Model", "abstract": "Multimodal large language models (MLLMs) are plagued by exorbitant inference costs attributable to the profusion of visual tokens within the vision encoder. The redundant visual tokens engenders a substantial computational load and key-value (KV) cache footprint bottleneck. Existing approaches focus on token-wise optimization, leveraging diverse intricate token pruning techniques to eliminate non-crucial visual tokens. Nevertheless, these methods often unavoidably undermine the integrity of the KV cache, resulting in failures in long-text generation tasks. To this end, we conduct an in-depth investigation towards the attention mechanism of the model from a new perspective, and discern that attention within more than half of all decode layers are semantic similar. Upon this finding, we contend that the attention in certain layers can be streamlined by inheriting the attention from their preceding layers. Consequently, we propose Lazy Attention, an efficient attention mechanism that enables cross-layer sharing of similar attention patterns. It ingeniously reduces layer-wise redundant computation in attention. In Lazy Attention, we develop a novel layer-shared cache, Q Cache, tailored for MLLMs, which facilitates the reuse of queries across adjacent layers. In particular, Q Cache is lightweight and fully compatible with existing inference frameworks, including Flash Attention and KV cache. Additionally, our method is highly flexible as it is orthogonal to existing token-wise techniques and can be deployed independently or combined with token pruning approaches. Empirical evaluations on multiple benchmarks demonstrate that our method can reduce KV cache usage by over 35% and achieve 1.5x throughput improvement, while sacrificing only approximately 1% of performance on various MLLMs. Compared with SOTA token-wise methods, our technique achieves superior accuracy preservation.", "authors": ["Jiedong Zhuang", "Lu Lu", "Ming Dai", "Rui Hu", "Jian Chen", "Qiang Liu", "Haoji Hu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.01901", "pdf_url": "https://arxiv.org/pdf/2602.01901v1", "arxiv_id": "2602.01901", "doi": "10.48550/arXiv.2602.01901", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4652}
{"id": "2a50f4ae4ed79d2d72c52291836029cb53a5895033a757de4b8da4e47e6b495f", "sources": ["arxiv", "semantic_scholar"], "title": "More Than a Quick Glance: Overcoming the Greedy Bias in KV-Cache Compression", "abstract": "While Large Language Models (LLMs) can theoretically support extensive context windows, their actual deployment is constrained by the linear growth of Key-Value (KV) cache memory. Prevailing compression strategies mitigate this through various pruning mechanisms, yet trade-off semantic recall for memory efficiency. In this work, we present LASER-KV (Layer Accumulated Selection with Exact-LSH Recall), a framework designed to test the limits of KV compression under a strict accumulative budgeting policy. We deviate from the standard fixed summary size approach by implementing a block-wise accumulation strategy governed by a protection divisor (n). This allows us to isolate the effects of compression from sliding window artifacts. Our experiments on the Babilong benchmark reveal performance degradation in previous compression methods by 15-30% on various long context tasks. LASER-KV maintains stable performance, achieving superior accuracies by a margin of upto 10% at 128k. These findings challenge the prevailing assumption that attention scores alone are a sufficient proxy for token utility.", "authors": ["Aryan Sood", "Tanvi Sharma", "Vansh Agrawal"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.02199", "pdf_url": "https://arxiv.org/pdf/2602.02199v1", "arxiv_id": "2602.02199", "doi": "10.48550/arXiv.2602.02199", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4652}
{"id": "70ed91ef416a8fd601a2835e6f0bd6378282a12e8adc6dc7da7263096a7e6304", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Concept Bottleneck Experts", "abstract": "Concept Bottleneck Models (CBMs) promote interpretability by grounding predictions in human-understandable concepts. However, existing CBMs typically constrain their task predictor to a single expression whose functional form is set a priori, limiting both predictive accuracy and adaptability to diverse user needs. We propose Mixture of Concept Bottleneck Experts (M-CBE), a framework that generalizes existing CBMs along two dimensions: the number of expressions, referred to as experts, employed by the task predictor to map concepts to the task, and the functional form each expression takes, thus exposing an underexplored region of this design space. We investigate this region by instantiating two novel models: Linear M-CBE, which learns a finite set of linear expressions, and Symbolic M-CBE, which leverages symbolic regression to discover expert functions from data subject to user-specified operator vocabularies. Empirical evaluation demonstrates that varying the number of expressions and their functional form provides a robust framework for navigating the accuracy-interpretability trade-off.", "authors": ["Francesco De Santis", "Gabriele Ciravegna", "Giovanni De Felice", "Arianna Casanova", "Francesco Giannini", "Michelangelo Diligenti", "Johannes Schneider", "Danilo Giordano", "Mateo Espinosa Zarlenga", "Pietro Barbiero"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.02886", "pdf_url": "https://arxiv.org/pdf/2602.02886v3", "arxiv_id": "2602.02886", "doi": "10.48550/arXiv.2602.02886", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4652}
{"id": "a18f04be5c73d35dd0ea8bf02e42901e2a9d7d62642e821ec7435e2e5121e909", "sources": ["arxiv", "semantic_scholar"], "title": "HQP: Sensitivity-Aware Hybrid Quantization and Pruning for Ultra-Low-Latency Edge AI Inference", "abstract": "The escalating demand for high-fidelity, real-time inference in distributed edge-cloud environments necessitates aggressive model optimization to counteract severe latency and energy constraints. This paper introduces the Hybrid Quantization and Pruning (HQP) framework, a novel, integrated methodology designed to achieve synergistic model acceleration while adhering to strict quality guarantees. We detail a sensitivity-aware structural pruning algorithm that employs a dynamic weight sensitivity metric, derived from a highly efficient approximation of the Fisher Information Matrix (FIM), to guide the iterative removal of redundant filters. This pruning is strictly conditional, enforcing an adherence to a maximum permissible accuracy drop (Delta ax) before the model proceeds to 8-bit post-training quantization. This rigorous coordination is critical, as it ensures the resultant sparse model structure is maximally robust to quantization error and hardware-specific kernel optimization. Exhaustive evaluation across heterogeneous NVIDIA Jetson edge platforms, utilizing resource-efficient architectures like MobileNetV3 and ResNet-18, demonstrates that the HQP framework achieves a peak performance gain of 3.12 times inference speedup and a 55 percent model size reduction, while rigorously containing the accuracy drop below the 1.5 percent constraint. A comprehensive comparative analysis against conventional single-objective compression techniques validates the HQP framework as a superior, hardware-agnostic solution for deploying ultra-low-latency AI in resource-limited edge infrastructures.", "authors": ["Dinesh Gopalan", "Ratul Ali"], "categories": ["cs.DC", "cs.AI", "cs.LG", "cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.06069", "pdf_url": "https://arxiv.org/pdf/2602.06069v1", "arxiv_id": "2602.06069", "doi": "10.48550/arXiv.2602.06069", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4652}
{"id": "c7f28ede15f34d58fc6d50dc331470aa22e54802e9022814f83087f31c2259b3", "sources": ["arxiv", "semantic_scholar"], "title": "Fast Autoregressive Video Diffusion and World Models with Temporal Cache Compression and Sparse Attention", "abstract": "Autoregressive video diffusion models enable streaming generation, opening the door to long-form synthesis, video world models, and interactive neural game engines. However, their core attention layers become a major bottleneck at inference time: as generation progresses, the KV cache grows, causing both increasing latency and escalating GPU memory, which in turn restricts usable temporal context and harms long-range consistency. In this work, we study redundancy in autoregressive video diffusion and identify three persistent sources: near-duplicate cached keys across frames, slowly evolving (largely semantic) queries/keys that make many attention computations redundant, and cross-attention over long prompts where only a small subset of tokens matters per frame. Building on these observations, we propose a unified, training-free attention framework for autoregressive diffusion: TempCache compresses the KV cache via temporal correspondence to bound cache growth; AnnCA accelerates cross-attention by selecting frame-relevant prompt tokens using fast approximate nearest neighbor (ANN) matching; and AnnSA sparsifies self-attention by restricting each query to semantically matched keys, also using a lightweight ANN. Together, these modules reduce attention, compute, and memory and are compatible with existing autoregressive diffusion backbones and world models. Experiments demonstrate up to x5--x10 end-to-end speedups while preserving near-identical visual quality and, crucially, maintaining stable throughput and nearly constant peak GPU memory usage over long rollouts, where prior methods progressively slow down and suffer from increasing memory usage.", "authors": ["Dvir Samuel", "Issar Tzachor", "Matan Levy", "Micahel Green", "Gal Chechik", "Rami Ben-Ari"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.01801", "pdf_url": "https://arxiv.org/pdf/2602.01801v1", "arxiv_id": "2602.01801", "doi": "10.48550/arXiv.2602.01801", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4652}
{"id": "c67be37e6c6ed2e63916a8a5a1b0823fad70ec336b1dfe2b6aec0bf6d011d56d", "sources": ["arxiv", "semantic_scholar"], "title": "SAME: Stabilized Mixture-of-Experts for Multimodal Continual Instruction Tuning", "abstract": "Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, but real-world deployment requires them to continually expand their capabilities, making Multimodal Continual Instruction Tuning (MCIT) essential. Recent methods leverage sparse expert routing to promote task specialization, but we find that the expert routing process suffers from drift as the data distribution evolves. For example, a grounding query that previously activated localization experts may instead be routed to irrelevant experts after learning OCR tasks. Meanwhile, the grounding-related experts can be overwritten by new tasks and lose their original functionality. Such failure reflects two problems: router drift, where expert selection becomes inconsistent over time, and expert drift, where shared experts are overwritten across tasks. Therefore, we propose StAbilized Mixture-of-Experts (SAME) for MCIT. To address router drift, SAME stabilizes expert selection by decomposing routing dynamics into orthogonal subspaces and updating only task-relevant directions. To mitigate expert drift, we regulate expert updates via curvature-aware scaling using historical input covariance in a rehearsal-free manner. SAME also introduces adaptive expert activation to freeze selected experts during training, reducing redundant computation and cross-task interference. We also introduce a new benchmark to evaluate MCIT with long task sequence, and extensive experiments demonstrate SAME's SOTA performance. Code is available at https://github.com/LAMDA-CL/Prism.", "authors": ["Zhen-Hao Xie", "Jun-Tao Tang", "Yu-Cheng Shi", "Han-Jia Ye", "De-Chuan Zhan", "Da-Wei Zhou"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.01990", "pdf_url": "https://arxiv.org/pdf/2602.01990v2", "arxiv_id": "2602.01990", "doi": "10.48550/arXiv.2602.01990", "citation_count": 5, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/LAMDA-CL/Prism", "venue": "arXiv.org", "quality_score": 0.719}
{"id": "d455cbb2a3a993a784e9e7cfe5478f05719d53a39bc0ec86a783d7509b803d91", "sources": ["arxiv", "semantic_scholar"], "title": "RAP: KV-Cache Compression via RoPE-Aligned Pruning", "abstract": "Long-context inference in large language models is increasingly bottlenecked by the memory and compute cost of the KV-Cache. Low-rank factorization compresses KV projections by writing $W \\approx A * B$, where A produces latent KV states and B can be absorbed into downstream weights. In modern RoPE-based LLMs, this absorption fails: RoPE forces latent KV states to be reconstructed to full dimension, reintroducing substantial memory and compute overhead. We propose RoPE-Aligned Pruning (RAP), which prunes entire RoPE-aligned column pairs to preserve RoPE's 2x2 rotation structure, restore B absorption, and eliminate reconstruction. Our evaluation on LLaMA-3-8B and Mistral-7B shows that RAP enables joint reduction of KV-Cache, attention parameters, and FLOPs by 20-30%, all at once, while maintaining strong accuracy. Notably, RAP reduces attention latency to 83% (prefill) and 77% (decode) of baseline.", "authors": ["Jihao Xin", "Tian Lyu", "David Keyes", "Hatem Ltaief", "Marco Canini"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2602.02599", "pdf_url": "https://arxiv.org/pdf/2602.02599v3", "arxiv_id": "2602.02599", "doi": "10.48550/arXiv.2602.02599", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4641}
{"id": "87b3b64477ba82042f5ee7751fe1db26e698d6cb277515c495ad644f6c3a6af6", "sources": ["arxiv", "semantic_scholar"], "title": "LRAgent: Efficient KV Cache Sharing for Multi-LoRA LLM Agents", "abstract": "Role specialization in multi-LLM agent systems is often realized via multi-LoRA, where agents share a pretrained backbone and differ only by lightweight adapters. Despite sharing base model weights, each agent independently builds and stores its own KV cache for the same long, tool-augmented trajectories, incurring substantial memory and compute overhead. Existing KV cache sharing methods largely overlook this multi-LoRA setting. We observe that, cache differences across agents are dominated by adapter outputs, while activations from the shared pretrained backbone remain highly similar. Based on this observation, we propose LRAgent, a KV cache sharing framework for multi-LoRA agents. It decomposes the cache into two components, a shared base component derived from pretrained weights and an adapter-dependent component derived from LoRA weights. LRAgent reduces memory overhead by sharing the base component across agents and storing the adapter component in its inherent low-rank form. It also reduces computational overhead by sharing the low-rank cache, enabled by a shared-A multi-LoRA architecture. This avoids redundant computations for contexts that have already been processed by other agents. To efficiently reconstruct adapter contributions at runtime, we introduce Flash-LoRA-Attention, a kernel that reorders attention computation to avoid materializing the low-rank cache to full dimension. LRAgent achieves throughput and time-to-first-token latency close to fully shared caching, while preserving accuracy near the non-shared caching baseline across agentic question-answering benchmarks.", "authors": ["Hyesung Jeon", "Hyeongju Ha", "Jae-Joon Kim"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2602.01053", "pdf_url": "https://arxiv.org/pdf/2602.01053v2", "arxiv_id": "2602.01053", "doi": "10.48550/arXiv.2602.01053", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4641}
{"id": "03ba9c87ae50fff29ecdc5c44bc17bb88737517ecf469139ef0e36b8402a6abf", "sources": ["arxiv", "semantic_scholar"], "title": "A Statistical Theory of Gated Attention through the Lens of Hierarchical Mixture of Experts", "abstract": "Self-attention has greatly contributed to the success of the widely used Transformer architecture by enabling learning from data with long-range dependencies. In an effort to improve performance, a gated attention model that leverages a gating mechanism within the multi-head self-attention has recently been proposed as a promising alternative. Gated attention has been empirically demonstrated to increase the expressiveness of low-rank mapping in standard attention and even to eliminate the attention sink phenomenon. Despite its efficacy, a clear theoretical understanding of gated attention's benefits remains lacking in the literature. To close this gap, we rigorously show that each entry in a gated attention matrix or a multi-head self-attention matrix can be written as a hierarchical mixture of experts. By recasting learning as an expert estimation problem, we demonstrate that gated attention is more sample-efficient than multi-head self-attention. In particular, while the former needs only a polynomial number of data points to estimate an expert, the latter requires exponentially many data points to achieve the same estimation error. Furthermore, our analysis also provides a theoretical justification for why gated attention yields higher performance when a gate is placed at the output of the scaled dot product attention or the value map rather than at other positions in the multi-head self-attention architecture.", "authors": ["Viet Nguyen", "Tuan Minh Pham", "Thinh Cao", "Tan Dinh", "Huy Nguyen", "Nhat Ho", "Alessandro Rinaldo"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2602.01468", "pdf_url": "https://arxiv.org/pdf/2602.01468v1", "arxiv_id": "2602.01468", "doi": "10.48550/arXiv.2602.01468", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4641}
{"id": "830c2ed0bd8288fdc9fcf1fdd6f3d98619fa492d663614c9987ffc4bd204765a", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Top-k Attention: Efficient Attention via Scalable Fast Weights", "abstract": "The vanilla self-attention mechanism in Transformers can be viewed as a two-layer fast-weight MLP, whose weights are dynamically induced by inputs and whose hidden dimension is equal to the sequence length $N$. As the context extends, the expressive capacity of such an $N$-width MLP increases, but it becomes unscalable for extremely long sequences. Recently, this fast-weight perspective has motivated the Mixture-of-Experts (MoE) attention mechanism, which partitions the sequence into rigid blocks, treats them as fast-weight experts, and sparsely routes the tokens to them. In this paper, we elevate this perspective to a unifying framework for efficient attention mechanisms, interpreting them as making fast weights scalable through either routing or compression, and organizing them into a five-dimensional taxonomy. Then, we propose Mixture-of-Top-$k$ Attention (MiTA), which employs a small set of landmark queries to gather top-$k$ attended key-value pairs as query-aware and deformable routed experts, while compressing the $N$-width MLP into a narrower shared expert. Consequently, our MiTA improves the flexibility of prior MoE attention from rigid to deformable fast-weight experts, as well as the scalability of prior top-$k$ attention from query-specific set to reusable top-$k$ set. We conduct extensive experiments on vision tasks showing the superior effectiveness and efficiency of our MiTA, and also uncovering intriguing properties such as an emergent token-pruning effect and easy generalization from standard attention. Code is available at https://github.com/QishuaiWen/MiTA.", "authors": ["Qishuai Wen", "Zhiyuan Huang", "Xianghan Meng", "Wei He", "Chun-Guang Li"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2602.01219", "pdf_url": "https://arxiv.org/pdf/2602.01219v5", "arxiv_id": "2602.01219", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/QishuaiWen/MiTA", "venue": null, "quality_score": 0.5484}
{"id": "4af2735cc6e6871363ecc4f4684437c5dde54ff3f144eda0f409b4cda50aef15", "sources": ["arxiv", "semantic_scholar"], "title": "LiME: Lightweight Mixture of Experts for Efficient Multimodal Multi-task Learning", "abstract": "MoE-PEFT methods combine Mixture of Experts with parameter-efficient fine-tuning for multi-task adaptation, but require separate adapters per expert causing trainable parameters to scale linearly with expert count and limiting applicability to adapter-based architectures. We propose LiME (Lightweight Mixture of Experts), which achieves expert specialization through lightweight modulation rather than adapter replication. Instead of separate adapters, LiME uses a single shared PEFT module and modulates its output with lightweight expert vectors, reducing expert parameters while generalizing to any PEFT method. Notably, LiME introduces zero-parameter routing by leveraging existing frozen and adapted representations eliminating learned router parameters typically required per layer. Theoretically, we prove that (i) more experts preserve more task-relevant information and (ii) modulation approximates full expert-specific PEFT with bounded error. LiME further incorporates n-gram windowed routing and adaptive expert selection (Auto Top-K) based on routing confidence. Experiments on MMT-47, a multimodal multi-task benchmark with 47 tasks spanning text, image, and video, demonstrate that LiME achieves competitive or superior performance while using up to 4x fewer trainable parameters and up to 29% faster training compared to corresponding MoE-PEFT baselines.", "authors": ["Md Kowsher", "Haris Mansoor", "Nusrat Jahan Prottasha", "Ozlem Garibay", "Victor Zhu", "Zhengping Ji", "Chen Chen"], "categories": ["cs.LG", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-01", "url": "https://arxiv.org/abs/2604.02338", "pdf_url": "https://arxiv.org/pdf/2604.02338v1", "arxiv_id": "2604.02338", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2953}
{"id": "e7c8470f1748fde46c10222d536dde97ce24852fe8169f28e4b100c1906a3637", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Minimax Estimation Rates for Contaminated Mixture of Multinomial Logistic Experts via Expert Heterogeneity", "abstract": "Contaminated mixture of experts (MoE) is motivated by transfer learning methods where a pre-trained model, acting as a frozen expert, is integrated with an adapter model, functioning as a trainable expert, in order to learn a new task. Despite recent efforts to analyze the convergence behavior of parameter estimation in this model, there are still two unresolved problems in the literature. First, the contaminated MoE model has been studied solely in regression settings, while its theoretical foundation in classification settings remains absent. Second, previous works on MoE models for classification capture pointwise convergence rates for parameter estimation without any guaranty of minimax optimality. In this work, we close these gaps by performing, for the first time, the convergence analysis of a contaminated mixture of multinomial logistic experts with homogeneous and heterogeneous structures, respectively. In each regime, we characterize uniform convergence rates for estimating parameters under challenging settings where ground-truth parameters vary with the sample size. Furthermore, we also establish corresponding minimax lower bounds to ensure that these rates are minimax optimal. Notably, our theories offer an important insight into the design of contaminated MoE, that is, expert heterogeneity yields faster parameter estimation rates and, therefore, is more sample-efficient than expert homogeneity.", "authors": ["Fanqi Yan", "Dung Le", "Trang Pham", "Huy Nguyen", "Nhat Ho"], "categories": ["math.ST", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-01-31", "url": "https://arxiv.org/abs/2602.00939", "pdf_url": "https://arxiv.org/pdf/2602.00939v1", "arxiv_id": "2602.00939", "doi": "10.48550/arXiv.2602.00939", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4629}
{"id": "efcd85330982ac8ad114c5bff59611628fbd7a34432e199e66e0ac2b79c1226a", "sources": ["arxiv", "semantic_scholar"], "title": "ProphetKV: User-Query-Driven Selective Recomputation for Efficient KV Cache Reuse in Retrieval-Augmented Generation", "abstract": "The prefill stage of long-context Retrieval-Augmented Generation (RAG) is severely bottlenecked by computational overhead. To mitigate this, recent methods assemble pre-calculated KV caches of retrieved RAG documents (by a user query) and reprocess selected tokens to recover cross-attention between these pre-calculated KV caches. However, we identify a fundamental \"crowding-out effect\" in current token selection criteria: globally salient but user-query-irrelevant tokens saturate the limited recomputation budget, displacing the tokens truly essential for answering the user query and degrading inference accuracy. We propose ProphetKV, a user-query-driven KV Cache reuse method for RAG scenarios. ProphetKV dynamically prioritizes tokens based on their semantic relevance to the user query and employs a dual-stage recomputation pipeline to fuse layer-wise attention metrics into a high-utility set. By ensuring the recomputation budget is dedicated to bridging the informational gap between retrieved context and the user query, ProphetKV achieves high-fidelity attention recovery with minimal overhead. Our extensive evaluation results show that ProphetKV retains 96%-101% of full-prefill accuracy with only a 20% recomputation ratio, while achieving accuracy improvements of 8.8%-24.9% on RULER and 18.6%-50.9% on LongBench over the state-of-the-art approaches (e.g., CacheBlend, EPIC, and KVShare).", "authors": ["Shihao Wang", "Jiahao Chen", "Yanqi Pan", "Hao Huang", "Yichen Hao", "Xiangyu Zou", "Wen Xia", "Wentao Zhang", "Chongyang Qiu", "Pengfei Wang"], "categories": ["cs.OS", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-31", "url": "https://arxiv.org/abs/2602.02579", "pdf_url": "https://arxiv.org/pdf/2602.02579v3", "arxiv_id": "2602.02579", "doi": "10.48550/arXiv.2602.02579", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4629}
{"id": "22f189e56be4b9d48a6cbdc6b44c485b6ae1c85229f52d687186bdcfc828a6d7", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Expert Sharing: Decoupling Memory from Parallelism in Mixture-of-Experts Diffusion LLMs", "abstract": "Among parallel decoding paradigms, diffusion large language models (dLLMs) have emerged as a promising candidate that balances generation quality and throughput. However, their integration with Mixture-of-Experts (MoE) architectures is constrained by an expert explosion: as the number of tokens generated in parallel increases, the number of distinct experts activated grows nearly linearly. This results in substantial memory traffic that pushes inference into a memory-bound regime, negating the efficiency gains of both MoE and parallel decoding. To address this challenge, we propose Dynamic Expert Sharing (DES), a novel technique that shifts MoE optimization from token-centric pruning and conventional expert skipping methods to sequence-level coreset selection. To maximize expert reuse, DES identifies a compact, high-utility set of experts to satisfy the requirements of an entire parallel decoding block. We introduce two innovative selection strategies: (1) Intra-Sequence Sharing (DES-Seq), which adapts optimal allocation to the sequence level, and (2) Saliency-Aware Voting (DES-Vote), a novel mechanism that allows tokens to collectively elect a coreset based on aggregated router weights. Extensive experiments on MoE dLLMs demonstrate that DES reduces unique expert activations by over 55% and latency by up to 38%, while retaining 99% of vanilla accuracy, effectively decoupling memory overhead from the degree of parallelism.", "authors": ["Hao Mark Chen", "Zhiwen Mo", "Royson Lee", "Qianzhou Wang", "Da Li", "Shell Xu Hu", "Wayne Luk", "Timothy Hospedales", "Hongxiang Fan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-31", "url": "https://arxiv.org/abs/2602.00879", "pdf_url": "https://arxiv.org/pdf/2602.00879v1", "arxiv_id": "2602.00879", "doi": "10.48550/arXiv.2602.00879", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4629}
{"id": "31f388f40f0fd052d74d756f443dea625e5dd4d1e6c18d7638dd2188b8bf754b", "sources": ["arxiv", "semantic_scholar"], "title": "MoDEx: Mixture of Depth-specific Experts for Multivariate Long-term Time Series Forecasting", "abstract": "Multivariate long-term time series forecasting (LTSF) supports critical applications such as traffic-flow management, solar-power scheduling, and electricity-transformer monitoring. The existing LTSF paradigms follow a three-stage pipeline of embedding, backbone refinement, and long-horizon prediction. However, the behaviors of individual backbone layers remain underexplored. We introduce layer sensitivity, a gradient-based metric inspired by GradCAM and effective receptive field theory, which quantifies both positive and negative contributions of each time point to a layer's latent features. Applying this metric to a three-layer MLP backbone reveals depth-specific specialization in modeling temporal dynamics in the input sequence. Motivated by these insights, we propose MoDEx, a lightweight Mixture of Depth-specific Experts, which replaces complex backbones with depth-specific MLP experts. MoDEx achieves state-of-the-art accuracy on seven real-world benchmarks, ranking first in 78 percent of cases, while using significantly fewer parameters and computational resources. It also integrates seamlessly into transformer variants, consistently boosting their performance and demonstrating robust generalizability as an efficient and high-performance LTSF framework.", "authors": ["Hyekyung Yoon", "Minhyuk Lee", "Imseung Park", "Myungjoo Kang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-31", "url": "https://arxiv.org/abs/2602.00624", "pdf_url": "https://arxiv.org/pdf/2602.00624v1", "arxiv_id": "2602.00624", "doi": "10.48550/arXiv.2602.00624", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4629}
{"id": "136ceed141070445fe1e68d932666284943a05f45234a604d6f0129a44f5d333", "sources": ["arxiv", "semantic_scholar"], "title": "SPA-Cache: Singular Proxies for Adaptive Caching in Diffusion Language Models", "abstract": "While Diffusion Language Models (DLMs) offer a flexible, arbitrary-order alternative to the autoregressive paradigm, their non-causal nature precludes standard KV caching, forcing costly hidden state recomputation at every decoding step. Existing DLM caching approaches reduce this cost by selective hidden state updates; however, they are still limited by (i) costly token-wise update identification heuristics and (ii) rigid, uniform budget allocation that fails to account for heterogeneous hidden state dynamics. To address these challenges, we present SPA-Cache that jointly optimizes update identification and budget allocation in DLM cache. First, we derive a low-dimensional singular proxy that enables the identification of update-critical tokens in a low-dimensional subspace, substantially reducing the overhead of update identification. Second, we introduce an adaptive strategy that allocates fewer updates to stable layers without degrading generation quality. Together, these contributions significantly improve the efficiency of DLMs, yielding up to an $8\\times$ throughput improvement over vanilla decoding and a $2$--$4\\times$ speedup over existing caching baselines.", "authors": ["Wenhao Sun", "Rong-Cheng Tu", "Yifu Ding", "Zhao Jin", "Jingyi Liao", "Yongcheng Jing", "Dacheng Tao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2602.02544", "pdf_url": "https://arxiv.org/pdf/2602.02544v2", "arxiv_id": "2602.02544", "doi": "10.48550/arXiv.2602.02544", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/wenhao728/spa-cache", "venue": "arXiv.org", "quality_score": 0.7136}
{"id": "c3e9faca2e36108f84ffc3df6caae157e75767e38659cd01a79ad900f029b8ed", "sources": ["arxiv", "semantic_scholar"], "title": "Decomposing and Composing: Towards Efficient Vision-Language Continual Learning via Rank-1 Expert Pool in a Single LoRA", "abstract": "Continual learning (CL) in vision-language models (VLMs) faces significant challenges in improving task adaptation and avoiding catastrophic forgetting. Existing methods usually have heavy inference burden or rely on external knowledge, while Low-Rank Adaptation (LoRA) has shown potential in reducing these issues by enabling parameter-efficient tuning. However, considering directly using LoRA to alleviate the catastrophic forgetting problem is non-trivial, we introduce a novel framework that restructures a single LoRA module as a decomposable Rank-1 Expert Pool. Our method learns to dynamically compose a sparse, task-specific update by selecting from this expert pool, guided by the semantics of the [CLS] token. In addition, we propose an Activation-Guided Orthogonal (AGO) loss that orthogonalizes critical parts of LoRA weights across tasks. This sparse composition and orthogonalization enable fewer parameter updates, resulting in domain-aware learning while minimizing inter-task interference and maintaining downstream task performance. Extensive experiments across multiple settings demonstrate state-of-the-art results in all metrics, surpassing zero-shot upper bounds in generalization. Notably, it reduces trainable parameters by 96.7% compared to the baseline method, eliminating reliance on external datasets or task-ID discriminators. The merged LoRAs retain less weights and incur no inference latency, making our method computationally lightweight.", "authors": ["Zhan Fa", "Yue Duan", "Jian Zhang", "Lei Qi", "Wanqi Yang", "Yinghuan Shi"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2601.22828", "pdf_url": "https://arxiv.org/pdf/2601.22828v1", "arxiv_id": "2601.22828", "doi": "10.48550/arXiv.2601.22828", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4618}
{"id": "6365c4e91177bc171ce629e5691e83764dfa209e510d5811d5cac725473d6a95", "sources": ["arxiv", "semantic_scholar"], "title": "Competitive Non-Clairvoyant KV-Cache Scheduling for LLM Inference", "abstract": "Large Language Model (LLM) inference presents a unique scheduling challenge due to the Key-Value (KV) cache, where a job's memory footprint grows linearly with the number of decoded tokens. This growth couples scheduling decisions with feasibility: a scheduler must minimize latency under a hard memory budget, yet the response lengths of requests are inherently unknown. While recent works have explored this problem either assuming clairvoyance -- exact knowledge of response lengths -- or relying on machine-learned predictions, obtaining robust performance guarantees without any prior knowledge of job sizes remains a theoretically fundamental and practically important open problem. In this work, we propose the Geometric Slicing Algorithm (GSA), the non-clairvoyant policy to achieve the first constant competitive ratio for this problem in the offline batch setting. GSA manages uncertainty through a geometric phase structure that periodically restarts jobs to bound memory exposure, combined with a staggered pipeline mechanism that enables high concurrency by smoothing aggregate memory consumption. We prove that GSA achieves a competitive ratio of at most 61.92 for general instances, improving to 32 in the large-memory regime. Our algorithmic framework also yields a clairvoyant counterpart, the Geometric Batching Algorithm (GBA), which achieves an approximation ratio of 10.67 for general instances and 6.75 in the large-memory regime -- significantly improving upon the best previously known bound of over 9000. Numerical experiments on real request traces demonstrate that our algorithms perform robustly while preserving these worst-case guarantees.", "authors": ["Yiding Feng", "Zonghan Yang", "Yuhao Zhang"], "categories": ["cs.DS"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2601.22996", "pdf_url": "https://arxiv.org/pdf/2601.22996v1", "arxiv_id": "2601.22996", "doi": "10.48550/arXiv.2601.22996", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4618}
{"id": "d3877ef10b4b010558ced84e5b0ef55918fecc8edd1c5e0f67379b5a5b1d22dc", "sources": ["arxiv", "semantic_scholar"], "title": "A Unified Study of LoRA Variants: Taxonomy, Review, Codebase, and Empirical Evaluation", "abstract": "Low-Rank Adaptation (LoRA) is a fundamental parameter-efficient fine-tuning method that balances efficiency and performance in large-scale neural networks. However, the proliferation of LoRA variants has led to fragmentation in methodology, theory, code, and evaluation. To this end, this work presents the first unified study of LoRA variants, offering a systematic taxonomy, unified theoretical review, structured codebase, and standardized empirical assessment. First, we categorize LoRA variants along four principal axes: rank, optimization dynamics, initialization, and integration with Mixture-of-Experts. Then, we review their relationships and evolution within a common theoretical framework focused on low-rank update dynamics. Further, we introduce LoRAFactory, a modular codebase that implements variants through a unified interface, supporting plug-and-play experimentation and fine-grained analysis. Last, using this codebase, we conduct a large-scale evaluation across natural language generation, natural language understanding, and image classification tasks, systematically exploring key hyperparameters. Our results uncover several findings, notably: LoRA and its variants exhibit pronounced sensitivity to the choices of learning rate compared to other hyperparameters; moreover, with proper hyperparameter configurations, LoRA consistently matches or surpasses the performance of most of its variants.", "authors": ["Haonan He", "Jingqi Ye", "Minglei Li", "Zhengbo Wang", "Tao Chen", "Lei Bai", "Peng Ye"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-30", "url": "https://arxiv.org/abs/2601.22708", "pdf_url": "https://arxiv.org/pdf/2601.22708v1", "arxiv_id": "2601.22708", "doi": "10.48550/arXiv.2601.22708", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4618}
{"id": "f30ad4a9d62c14e5d71500615657a2af07e8727a65c3e093c44620bb0f549573", "sources": ["arxiv", "semantic_scholar"], "title": "Don't be so Stief! Learning KV Cache low-rank approximation over the Stiefel manifold", "abstract": "Key-value (KV) caching enables fast autoregressive decoding but at long contexts becomes a dominant bottleneck in High Bandwidth Memory (HBM) capacity and bandwidth. A common mitigation is to compress cached keys and values by projecting per-head matrices to a lower rank, storing only the projections in the HBM. However, existing post-training approaches typically fit these projections using SVD-style proxy objectives, which may poorly reflect end-to-end reconstruction after softmax, value mixing, and subsequent decoder-layer transformations. For these reasons, we introduce StiefAttention, a post-training KV-cache compression method that learns orthonormal projection bases by directly minimizing decoder-layer output reconstruction error. StiefAttention additionally constructs layer-wise error-rank profiles over candidate ranks, enabling sequential rank allocation under a user-specified KV cache budget. Notably, on Llama3-8B under the same conditions, StiefAttention outperforms EigenAttention by $4.2$ points on C4 perplexity and $8.9$ points on 0-shot MMLU accuracy at iso-compression, yielding lower relative error and higher cosine similarity with respect to the original decoder-layer outputs.", "authors": ["Luca Benfenati", "Matteo Risso", "Andrea Vannozzi", "Ahmet Caner Yüzügüler", "Lukas Cavigelli", "Enrico Macii", "Daniele Jahier Pagliari", "Alessio Burrello"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.21686", "pdf_url": "https://arxiv.org/pdf/2601.21686v2", "arxiv_id": "2601.21686", "doi": "10.48550/arXiv.2601.21686", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4606}
{"id": "1630d3f87d09a4e2807053fe336a9031620a61ac4a9c902fade28c41e348f5fb", "sources": ["arxiv", "semantic_scholar"], "title": "Past- and Future-Informed KV Cache Policy with Salience Estimation in Autoregressive Video Diffusion", "abstract": "Video generation is pivotal to digital media creation, and recent advances in autoregressive video generation have markedly enhanced the efficiency of real-time video synthesis. However, existing approaches generally rely on heuristic KV Cache policies, which ignore differences in token importance in long-term video generation. This leads to the loss of critical spatiotemporal information and the accumulation of redundant, invalid cache, thereby degrading video generation quality and efficiency. To address this limitation, we first observe that token contributions to video generation are highly time-heterogeneous and accordingly propose a novel Past- and Future-Informed KV Cache Policy (PaFu-KV). Specifically, PaFu-KV introduces a lightweight Salience Estimation Head distilled from a bidirectional teacher to estimate salience scores, allowing the KV cache to retain informative tokens while discarding less relevant ones. This policy yields a better quality-efficiency trade-off by shrinking KV cache capacity and reducing memory footprint at inference time. Extensive experiments on benchmarks demonstrate that our method preserves high-fidelity video generation quality while enables accelerated inference, thereby enabling more efficient long-horizon video generation. Our code will be released upon paper acceptance.", "authors": ["Hanmo Chen", "Chenghao Xu", "Xu Yang", "Xuan Chen", "Cheng Deng"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.21896", "pdf_url": "https://arxiv.org/pdf/2601.21896v3", "arxiv_id": "2601.21896", "doi": "10.48550/arXiv.2601.21896", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4606}
{"id": "c5b808d6c0d7cddfdc3ed1353be05bd2127804cf68f627a03589520b9c4420ff", "sources": ["arxiv", "semantic_scholar"], "title": "MoHETS: Long-term Time Series Forecasting with Mixture-of-Heterogeneous-Experts", "abstract": "Real-world multivariate time series can exhibit intricate multi-scale structures, including global trends, local periodicities, and non-stationary regimes, which makes long-horizon forecasting challenging. Although sparse Mixture-of-Experts (MoE) approaches improve scalability and specialization, they typically rely on homogeneous MLP experts that poorly capture the diverse temporal dynamics of time series data. We address these limitations with MoHETS, an encoder-only Transformer that integrates sparse Mixture-of-Heterogeneous-Experts (MoHE) layers. MoHE routes temporal patches to a small subset of expert networks, combining a shared depthwise-convolution expert for sequence-level continuity with routed Fourier-based experts for patch-level periodic structures. MoHETS further improves robustness to non-stationary dynamics by incorporating exogenous information via cross-attention over covariate patch embeddings. Finally, we replace parameter-heavy linear projection heads with a lightweight convolutional patch decoder, improving parameter efficiency, reducing training instability, and allowing a single model to generalize across arbitrary forecast horizons. We validate across seven multivariate benchmarks and multiple horizons, with MoHETS consistently achieving state-of-the-art performance, reducing the average MSE by $12\\%$ compared to strong recent baselines, demonstrating effective heterogeneous specialization for long-term forecasting.", "authors": ["Evandro S. Ortigossa", "Guy Lutsker", "Eran Segal"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.21866", "pdf_url": "https://arxiv.org/pdf/2601.21866v2", "arxiv_id": "2601.21866", "doi": "10.48550/arXiv.2601.21866", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4606}
{"id": "c1d0b836d8d8d564ddf0b6100f5437da025da428da0feec901d0f711ceac2139", "sources": ["arxiv", "semantic_scholar"], "title": "Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves", "abstract": "Depth-recurrence facilitates latent reasoning by sharing parameters across depths. However, prior work lacks combined FLOP-, parameter-, and memory-matched baselines, underutilizes depth-recurrence due to partially fixed layer stacks, and ignores the bottleneck of constant hidden-sizes that restricts many-step latent reasoning. To address this, we introduce a modular framework of depth-recurrent attention mixtures (Dreamer), combining sequence attention, depth attention, and sparse expert attention. It alleviates the hidden-size bottleneck through attention along depth, decouples scaling dimensions, and allows depth-recurrent models to scale efficiently and effectively. Across language reasoning benchmarks, our models require 2 to 8x fewer training tokens for the same accuracy as FLOP-, parameter-, and memory-matched SOTA, and outperform ca. 2x larger SOTA models with the same training tokens. We further present insights into knowledge usage across depths, e.g., showing 2 to 11x larger expert selection diversity than SOTA MoEs.", "authors": ["Jonas Knupp", "Jan Hendrik Metzen", "Jeremias Bohn", "Georg Groh", "Kristian Kersting"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.21582", "pdf_url": "https://arxiv.org/pdf/2601.21582v1", "arxiv_id": "2601.21582", "doi": "10.48550/arXiv.2601.21582", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4606}
{"id": "b8c38cde7506e5cce0cff63874f44389785f7f072e2c85dfecf8b3c2d80a5943", "sources": ["arxiv", "semantic_scholar"], "title": "Abstracting Robot Manipulation Skills via Mixture-of-Experts Diffusion Policies", "abstract": "Diffusion-based policies have recently shown strong results in robot manipulation, but their extension to multi-task scenarios is hindered by the high cost of scaling model size and demonstrations. We introduce Skill Mixture-of-Experts Policy (SMP), a diffusion-based mixture-of-experts policy that learns a compact orthogonal skill basis and uses sticky routing to compose actions from a small, task-relevant subset of experts at each step. A variational training objective supports this design, and adaptive expert activation at inference yields fast sampling without oversized backbones. We validate SMP in simulation and on a real dual-arm platform with multi-task learning and transfer learning tasks, where SMP achieves higher success rates and markedly lower inference cost than large diffusion baselines. These results indicate a practical path toward scalable, transferable multi-task manipulation: learn reusable skills once, activate only what is needed, and adapt quickly when tasks change.", "authors": ["Ce Hao", "Xuanran Zhai", "Yaohua Liu", "Harold Soh"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.21251", "pdf_url": "https://arxiv.org/pdf/2601.21251v1", "arxiv_id": "2601.21251", "doi": "10.48550/arXiv.2601.21251", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4606}
{"id": "a9d5209a953bdde15f46ebfb3e87d2ef8cbeb89d48990a11559a582cb0bf2a98", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Speedup -- Utilizing KV Cache for Sampling and Reasoning", "abstract": "KV caches, typically used only to speed up autoregressive decoding, encode contextual information that can be reused for downstream tasks at no extra cost. We propose treating the KV cache as a lightweight representation, eliminating the need to recompute or store full hidden states. Despite being weaker than dedicated embeddings, KV-derived representations are shown to be sufficient for two key applications: \\textbf{(i) Chain-of-Embedding}, where they achieve competitive or superior performance on Llama-3.1-8B-Instruct and Qwen2-7B-Instruct; and \\textbf{(ii) Fast/Slow Thinking Switching}, where they enable adaptive reasoning on Qwen3-8B and DeepSeek-R1-Distil-Qwen-14B, reducing token generation by up to $5.7\\times$ with minimal accuracy loss. Our findings establish KV caches as a free, effective substrate for sampling and reasoning, opening new directions for representation reuse in LLM inference. Code: https://github.com/cmd2001/ICLR2026_KV-Embedding.", "authors": ["Zeyu Xing", "Xing Li", "Hui-Ling Zhen", "Mingxuan Yuan", "Sinno Jialin Pan"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-28", "url": "https://arxiv.org/abs/2601.20326", "pdf_url": "https://arxiv.org/pdf/2601.20326v1", "arxiv_id": "2601.20326", "doi": "10.48550/arXiv.2601.20326", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/cmd2001/ICLR2026_KV-Embedding", "venue": "arXiv.org", "quality_score": 0.7101}
{"id": "8ee86a3cddd917287b6aa5b850a3d889d1355e9b5a29750a825f17c07c71dfd9", "sources": ["arxiv", "semantic_scholar"], "title": "Window-Diffusion: Accelerating Diffusion Language Model Inference with Windowed Token Pruning and Caching", "abstract": "Diffusion language models (DLMs) generate text through iterative denoising, but inference requires full-sequence attention at every iteration, resulting in substantial redundant computation on masked tokens. Block-wise diffusion can reduce this cost, yet it typically relies on retraining and constrained update orders, limiting its direct applicability to pretrained DLMs. Our token-level analysis reveals pronounced structural locality in DLM inference. Decoding is driven by a small set of prefix-localized active tokens; the influence of distant undecoded context diminishes rapidly, and decoded tokens exhibit stage-wise temporal stability, enabling reuse of intermediate representations except for a brief post-decode transient. Motivated by these observations, we propose \\textbf{\\placeholder}\\footnote{The source code is available at https://github.com/vhicrgit/Window-Diffusion.}, a window-based token pruning and caching method for inference. We maintain a local computation window that slides rightward as denoising progresses, and partition undecoded tokens into: (i) \\textit{active tokens} that are computed online, (ii) \\textit{buffer tokens} whose KV states are cached and periodically refreshed, and (iii) \\textit{far-field tokens} that are pruned outside the window. Computation is restricted to active and buffer tokens within the window, while far-field tokens are omitted at each stage. Experiments on LLaDA and Dream show that, under matched compute budgets, our method achieves up to $99\\times$ inference speedup while largely preserving generation performance.", "authors": ["Fengrui Zuo", "Zhiwei Ke", "Yiming Liu", "Wenqi Lou", "Chao Wang", "Xuehai Zhou"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-28", "url": "https://arxiv.org/abs/2601.20332", "pdf_url": "https://arxiv.org/pdf/2601.20332v2", "arxiv_id": "2601.20332", "doi": "10.48550/arXiv.2601.20332", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/vhicrgit/Window-Diffusion.}", "venue": "arXiv.org", "quality_score": 0.7101}
{"id": "5f52586dde4bdb31aed80bdf01645e0e303906ca413234aab3947f1459c924e0", "sources": ["arxiv", "semantic_scholar"], "title": "CompSRT: Quantization and Pruning for Image Super Resolution Transformers", "abstract": "Model compression has become an important tool for making image super resolution models more efficient. However, the gap between the best compressed models and the full precision model still remains large and a need for deeper understanding of compression theory on more performant models remains. Prior research on quantization of LLMs has shown that Hadamard transformations lead to weights and activations with reduced outliers, which leads to improved performance. We argue that while the Hadamard transform does reduce the effect of outliers, an empirical analysis on how the transform functions remains needed. By studying the distributions of weights and activations of SwinIR-light, we show with statistical analysis that lower errors is caused by the Hadamard transforms ability to reduce the ranges, and increase the proportion of values around $0$. Based on these findings, we introduce CompSRT, a more performant way to compress the image super resolution transformer network SwinIR-light. We perform Hadamard-based quantization, and we also perform scalar decomposition to introduce two additional trainable parameters. Our quantization performance statistically significantly surpasses the SOTA in metrics with gains as large as 1.53 dB, and visibly improves visual quality by reducing blurriness at all bitwidths. At $3$-$4$ bits, to show our method is compatible with pruning for increased compression, we also prune $40\\%$ of weights and show that we can achieve $6.67$-$15\\%$ reduction in bits per parameter with comparable performance to SOTA.", "authors": ["Dorsa Zeinali", "Hailing Wang", "Yitian Zhang", "Yun Fu"], "categories": ["eess.IV"], "fields_of_study": ["Engineering"], "published_date": "2026-01-28", "url": "https://arxiv.org/abs/2601.21069", "pdf_url": "https://arxiv.org/pdf/2601.21069v2", "arxiv_id": "2601.21069", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2924}
{"id": "2e10b5567754f8b2b0b6e44fa7187c97fd2c378a1a3e4216cc10677dd4d168d3", "sources": ["arxiv", "semantic_scholar"], "title": "Switchcodec: Adaptive residual-expert sparse quantization for high-fidelity neural audio coding", "abstract": "Recent neural audio compression models often rely on residual vector quantization for high-fidelity coding, but using a fixed number of per-frame codebooks is suboptimal for the wide variability of audio content-especially for signals that are either very simple or highly complex. To address this limitation, we propose SwitchCodec, a neural audio codec based on Residual Experts Vector Quantization (REVQ). REVQ combines a shared quantizer with dynamically routed expert quantizers that are activated according to the input audio, decoupling bitrate from codebook capacity and improving compression efficiency. This design ensures full training and utilization of each quantizer. In addition, a variable-bitrate mechanism adjusts the number of active expert quantizers at inference, enabling multi-bitrate operation without retraining. Experiments demonstrate that SwitchCodec surpasses existing baselines on both objective metrics and subjective listening tests.", "authors": ["Xiangbo Wang", "Wenbin Jiang", "Jin Wang", "Yubo You", "Sheng Fang", "Fei Wen"], "categories": ["cs.SD", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-28", "url": "https://arxiv.org/abs/2601.20362", "pdf_url": "https://arxiv.org/pdf/2601.20362v2", "arxiv_id": "2601.20362", "doi": "10.48550/arXiv.2601.20362", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.4595}
{"id": "d1eb94374b0c0945c3e731c7c2dbe2dcd6e3b6f10c63682ad7632696c8b1fc4b", "sources": ["arxiv", "semantic_scholar"], "title": "ChunkWise LoRA: Adaptive Sequence Partitioning for Memory-Efficient Low-Rank Adaptation and Accelerated LLM Inference", "abstract": "Recent advances in low-rank adaptation (LoRA) have enabled efficient fine-tuning of large language models (LLMs) with minimal additional parameters. However, existing LoRA methods apply static rank configurations uniformly across all input tokens, ignoring variation in token complexity and computational requirements. In this work, we propose ChunkWise LoRA, a dynamic and adaptive approach that partitions sequences into variable-length chunks based on token complexity and assigns each chunk a tailored low-rank configuration. Our system introduces a runtime scheduler that estimates token difficulty, performs adaptive chunking, and selects per-chunk LoRA rank and scaling using a rank-ladder mechanism. To preserve output consistency, we further introduce a boundary-safe composition module and integrate policy-driven KV-cache strategies. Experiments on benchmark datasets such as Wikitext-103 and SQuAD demonstrate that ChunkWise LoRA achieves up to 34\\% lower latency and 38% memory reduction compared to baseline LoRA, while maintaining or improving task performance metrics like BLEU, EM, and perplexity. The proposed framework remains fully compatible with existing transformer architectures and inference frameworks, providing a practical solution for real-world deployment of parameter-efficient LLMs.", "authors": ["Ketan Thakkar", "Maitreyi Chatterjee", "Ramasubramanian Balasubramanian", "Achyuthan Jootoo", "Rajendra Ugrani"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-28", "url": "https://arxiv.org/abs/2601.21109", "pdf_url": "https://arxiv.org/pdf/2601.21109v1", "arxiv_id": "2601.21109", "doi": "10.1109/ISED67359.2025.11405311", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Symposium on Embedded Computing and System Design", "quality_score": 0.4595}
{"id": "720e24afeb3c1ee437b343a388545d41fc2fe86c49d3efd4b69b860a5e379be3", "sources": ["arxiv", "semantic_scholar"], "title": "Hyperparameter Transfer with Mixture-of-Expert Layers", "abstract": "Mixture-of-Experts (MoE) layers have emerged as an important tool in scaling up modern neural networks by decoupling total trainable parameters from activated parameters in the forward pass for each token. However, sparse MoEs add complexity to training due to (i) new trainable parameters (router weights) that, like all other parameter groups, require hyperparameter (HP) tuning; (ii) new architecture scale dimensions (number of and size of experts) that must be chosen and potentially taken large. To make HP selection cheap and reliable, we propose a new parameterization for transformer models with MoE layers when scaling model width, depth, number of experts, and expert (hidden) size. Our parameterization is justified by a novel dynamical mean-field theory (DMFT) analysis. When varying different model dimensions trained at a fixed token budget, we find empirically that our parameterization enables reliable HP transfer across models from 51M to over 2B total parameters. We further take HPs identified from sweeping small models on a short token horizon to train larger models on longer horizons and report performant model behaviors.", "authors": ["Tianze Jiang", "Blake Bordelon", "Cengiz Pehlevan", "Boris Hanin"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-28", "url": "https://arxiv.org/abs/2601.20205", "pdf_url": "https://arxiv.org/pdf/2601.20205v3", "arxiv_id": "2601.20205", "doi": "10.48550/arXiv.2601.20205", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4595}
{"id": "b3699373930cc6de0cae049724b0c390cf6bd2f2b5b61dffe06c4198453f54ac", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring Fine-Tuning for In-Context Retrieval and Efficient KV-Caching in Long-Context Language Models", "abstract": "With context windows of millions of tokens, Long-Context Language Models (LCLMs) can encode entire document collections, offering a strong alternative to conventional retrieval-augmented generation (RAG). However, it remains unclear whether fine-tuning strategies can improve long-context performance and translate to greater robustness under KV-cache compression techniques. In this work, we investigate which training strategies most effectively enhance LCLMs' ability to identify and use relevant information, as well as enhancing their robustness under KV-cache compression. Our experiments show substantial in-domain improvements, achieving gains of up to +20 points over the base model. However, out-of-domain generalization remains task dependent with large variance -- LCLMs excels on finance questions (+9 points), while RAG shows stronger performance on multiple-choice questions (+6 points) over the baseline models. Finally, we show that our fine-tuning approaches bring moderate improvements in robustness under KV-cache compression, with gains varying across tasks.", "authors": ["Francesco Maria Molfese", "Momchil Hardalov", "Rexhina Blloshmi", "Bill Byrne", "Adrià de Gispert"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-26", "url": "https://arxiv.org/abs/2601.18527", "pdf_url": "https://arxiv.org/pdf/2601.18527v1", "arxiv_id": "2601.18527", "doi": "10.48550/arXiv.2601.18527", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference of the European Chapter of the Association for Computational Linguistics", "quality_score": 0.4572}
{"id": "7459dc3a1bb7a09dc713b5e40235308624e17a8e3236406478452888121a7edf", "sources": ["arxiv", "semantic_scholar"], "title": "Randomization Boosts KV Caching, Learning Balances Query Load: A Joint Perspective", "abstract": "KV caching is a fundamental technique for accelerating Large Language Model (LLM) inference by reusing key-value (KV) pairs from previous queries, but its effectiveness under limited memory is highly sensitive to the eviction policy. The default Least Recently Used (LRU) eviction algorithm struggles with dynamic online query arrivals, especially in multi-LLM serving scenarios, where balancing query load across workers and maximizing cache hit rate of each worker are inherently conflicting objectives. We give the first unified mathematical model that captures the core trade-offs between KV cache eviction and query routing. Our analysis reveals the theoretical limitations of existing methods and leads to principled algorithms that integrate provably competitive randomized KV cache eviction with learning-based methods to adaptively route queries with evolving patterns, thus balancing query load and cache hit rate. Our theoretical results are validated by extensive experiments across 4 benchmarks and 3 prefix-sharing settings, demonstrating improvements of up to 6.92$\\times$ in cache hit rate, 11.96$\\times$ reduction in latency, 14.06$\\times$ reduction in time-to-first-token (TTFT), and 77.4% increase in throughput over the state-of-the-art methods. Our code is available at https://github.com/fzwark/KVRouting.", "authors": ["Fangzhou Wu", "Sandeep Silwal", " Qiuyi", " Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-26", "url": "https://arxiv.org/abs/2601.18999", "pdf_url": "https://arxiv.org/pdf/2601.18999v1", "arxiv_id": "2601.18999", "doi": "10.48550/arXiv.2601.18999", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/fzwark/KVRouting", "venue": "arXiv.org", "quality_score": 0.7066}
{"id": "779454640ef85adfd6089ba2b7e2352f9cfebf42d4473e58d691729b7d996ed9", "sources": ["arxiv", "semantic_scholar"], "title": "$\\infty$-MoE: Generalizing Mixture of Experts to Infinite Experts", "abstract": "The Mixture of Experts (MoE) selects a few feed-forward networks (FFNs) per token, achieving an effective trade-off between computational cost and performance. In conventional MoE, each expert is treated as entirely independent, and experts are combined in a discrete space. As a result, when the number of experts increases, it becomes difficult to train each expert effectively. To stabilize training while increasing the number of experts, we propose $\\infty$-MoE that selects a portion of the parameters of large FFNs based on continuous values sampled for each token. By considering experts in a continuous space, this approach allows for an infinite number of experts while maintaining computational efficiency. Experiments show that a GPT-2 Small-based $\\infty$-MoE model, with 129M active and 186M total parameters, achieves comparable performance to a dense GPT-2 Medium with 350M parameters. Adjusting the number of sampled experts at inference time allows for a flexible trade-off between accuracy and speed, with an improvement of up to 2.5\\% in accuracy over conventional MoE.", "authors": ["Shota Takashiro", "Takeshi Kojima", "Shohei Taniguchi", "Yusuke Iwasawa", "Yutaka Matsuo"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-25", "url": "https://arxiv.org/abs/2601.17680", "pdf_url": "https://arxiv.org/pdf/2601.17680v1", "arxiv_id": "2601.17680", "doi": "10.48550/arXiv.2601.17680", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.456}
{"id": "59502184008164e50c76bcaa88d6d30522944297bfeef3561b843abb7c1ecf31", "sources": ["arxiv", "semantic_scholar"], "title": "Systematic Characterization of Minimal Deep Learning Architectures: A Unified Analysis of Convergence, Pruning, and Quantization", "abstract": "Deep learning networks excel at classification, yet identifying minimal architectures that reliably solve a task remains challenging. We present a computational methodology for systematically exploring and analyzing the relationships among convergence, pruning, and quantization. The workflow first performs a structured design sweep across a large set of architectures, then evaluates convergence behavior, pruning sensitivity, and quantization robustness on representative models. Focusing on well-known image classification of increasing complexity, and across Deep Neural Networks, Convolutional Neural Networks, and Vision Transformers, our initial results show that, despite architectural diversity, performance is largely invariant and learning dynamics consistently exhibit three regimes: unstable, learning, and overfitting. We further characterize the minimal learnable parameters required for stable learning, uncover distinct convergence and pruning phases, and quantify the effect of reduced numeric precision on trainable parameters. Aligning with intuition, the results confirm that deeper architectures are more resilient to pruning than shallower ones, with parameter redundancy as high as 60%, and quantization impacts models with fewer learnable parameters more severely and has a larger effect on harder image datasets. These findings provide actionable guidance for selecting compact, stable models under pruning and low-precision constraints in image classification.", "authors": ["Ziwei Zheng", "Huizhi Liang", "Vaclav Snasel", "Vito Latora", "Panos Pardalos", "Giuseppe Nicosia", "Varun Ojha"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-25", "url": "https://arxiv.org/abs/2601.17987", "pdf_url": "https://arxiv.org/pdf/2601.17987v1", "arxiv_id": "2601.17987", "doi": "10.1109/CAI68641.2026.11536272", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Algebraic Informatics", "quality_score": 0.456}
{"id": "5c59984e14ccdde51a9339f89d69e5d4ab2a0cab58504c8b31294f096417a565", "sources": ["arxiv", "semantic_scholar"], "title": "Domain-Expert-Guided Hybrid Mixture-of-Experts for Medical AI: Integrating Data-Driven Learning with Clinical Priors", "abstract": "Mixture-of-Experts (MoE) models increase representational capacity with modest computational cost, but their effectiveness in specialized domains such as medicine is limited by small datasets. In contrast, clinical practice offers rich expert knowledge, such as physician gaze patterns and diagnostic heuristics, that models cannot reliably learn from limited data. Combining data-driven experts, which capture novel patterns, with domain-expert-guided experts, which encode accumulated clinical insights, provides complementary strengths for robust and clinically meaningful learning. To this end, we propose Domain-Knowledge-Guided Hybrid MoE (DKGH-MoE), a plug-and-play and interpretable module that unifies data-driven learning with domain expertise. DKGH-MoE integrates a data-driven MoE to extract novel features from raw imaging data, and a domain-expert-guided MoE incorporates clinical priors, specifically clinician eye-gaze cues, to emphasize regions of high diagnostic relevance. By integrating domain expert insights with data-driven features, DKGH-MoE improves both performance and interpretability.", "authors": ["Jinchen Gu", "Nan Zhao", "Lei Qiu", "Lu Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-25", "url": "https://arxiv.org/abs/2601.17977", "pdf_url": "https://arxiv.org/pdf/2601.17977v1", "arxiv_id": "2601.17977", "doi": "10.1109/ISBI61048.2026.11515744", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Symposium on Biomedical Imaging", "quality_score": 0.456}
{"id": "388ef9690bb443c6cac313ba44813aad4dd1684fafbb9c15d15846575b1ed745", "sources": ["arxiv", "semantic_scholar"], "title": "Split-on-Share: Mixture of Sparse Experts for Task-Agnostic Continual Learning", "abstract": "Continual learning in Large Language Models (LLMs) is hindered by the plasticity-stability dilemma, where acquiring new capabilities often leads to catastrophic forgetting of previous knowledge. Existing methods typically treat parameters uniformly, failing to distinguish between specific task knowledge and shared capabilities. We introduce Mixture of Sparse Experts for Task-Agnostic Continual Learning, referred to as SETA, a framework that resolves the plasticity-stability conflict by decomposing the model into modular subspaces. Unlike standard updates, where tasks compete for the same parameters, SETA separates knowledge into unique experts, designed to isolate task-specific patterns, and shared experts, responsible for capturing common features. This structure is maintained through elastic weight anchoring, which protects critical shared knowledge and enables a unified gating network to automatically retrieve the correct expert combination for each task during inference. Extensive experiments across diverse domain-specific and general benchmarks demonstrate that SETA consistently outperforms state-of-the-art parameter-efficient fine-tuning-based continual learning methods.", "authors": ["Fatema Siddika", "Md Anwar Hossen", "Tanwi Mallick", "Ali Jannesari"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-24", "url": "https://arxiv.org/abs/2601.17616", "pdf_url": "https://arxiv.org/pdf/2601.17616v2", "arxiv_id": "2601.17616", "doi": "10.48550/arXiv.2601.17616", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4549}
{"id": "47b89e88e5464d0e31bd5d295bfd4c0e4ca7874e17df4b0148ffd55df591b416", "sources": ["arxiv", "semantic_scholar"], "title": "MambaNet: Mamba-assisted Channel Estimation Neural Network With Attention Mechanism", "abstract": "This paper proposes a Mamba-assisted neural network framework incorporating self-attention mechanism to achieve improved channel estimation with low complexity for orthogonal frequency-division multiplexing (OFDM) waveforms, particularly for configurations with a large number of subcarriers. With the integration of customized Mamba architecture, the proposed framework handles large-scale subcarrier channel estimation efficiently while capturing long-distance dependencies among these subcarriers effectively. Unlike conventional Mamba structure, this paper implements a bidirectional selective scan to improve channel estimation performance, because channel gains at different subcarriers are non-causal. Moreover, the proposed framework exhibits relatively lower space complexity than transformer-based neural networks. Simulation results tested on the 3GPP TS 36.101 channel demonstrate that compared to other baseline neural network solutions, the proposed method achieves improved channel estimation performance with a reduced number of tunable parameters.", "authors": ["Dianxin Luan", "Chengsi Liang", "Jie Huang", "Zheng Lin", "Kaitao Meng", "John Thompson", "Cheng-Xiang Wang"], "categories": ["cs.LG", "cs.AI", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2026-01-23", "url": "https://arxiv.org/abs/2601.17108", "pdf_url": "https://arxiv.org/pdf/2601.17108v1", "arxiv_id": "2601.17108", "doi": "10.48550/arXiv.2601.17108", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4537}
{"id": "b8895b93b5edba9d0016971351d1ade615f440483ae99c564fbb3a4210a30404", "sources": ["arxiv", "semantic_scholar"], "title": "Least-Loaded Expert Parallelism: Load Balancing An Imbalanced Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) models are typically pre-trained with explicit load-balancing constraints to ensure statistically balanced expert routing. Despite this, we observe that even well-trained MoE models exhibit significantly imbalanced routing. This behavior is arguably natural-and even desirable - as imbalanced routing allows models to concentrate domain-specific knowledge within a subset of experts. Expert parallelism (EP) is designed to scale MoE models by distributing experts across multiple devices, but with a less-discussed assumption of balanced routing. Under extreme imbalance, EP can funnel a disproportionate number of tokens to a small number of experts, leading to compute- and memory-bound failures on overloaded devices during post-training or inference, where explicit load balancing is often inapplicable. We propose Least-Loaded Expert Parallelism (LLEP), a novel EP algorithm that dynamically reroutes excess tokens and associated expert parameters from overloaded devices to underutilized ones. This ensures that all devices complete their workloads within the minimum collective latency while respecting memory constraints. Across different model scales, LLEP achieves up to 5x speedup and 4x reduction in peak memory usage compared to standard EP. This enables faster and higher-throughput post-training and inference, with ~1.9x faster for gpt-oss-120b. We support our method with extensive theoretical analysis and comprehensive empirical evaluations, including ablation studies. These results illuminate key trade-offs and enable a principled framework for hardware-specific hyper-parameter tuning to achieve optimal performance.", "authors": ["Xuan-Phi Nguyen", "Shrey Pandit", "Austin Xu", "Caiming Xiong", "Shafiq Joty"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-23", "url": "https://arxiv.org/abs/2601.17111", "pdf_url": "https://arxiv.org/pdf/2601.17111v1", "arxiv_id": "2601.17111", "doi": "10.48550/arXiv.2601.17111", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4537}
{"id": "eea63b78ed99767bacbfaf73608754767432ecea0e36b2b12e246528619d6e1d", "sources": ["arxiv", "semantic_scholar"], "title": "Clustering-Guided Spatial-Spectral Mamba for Hyperspectral Image Classification", "abstract": "Although Mamba models greatly improve Hyperspectral Image (HSI) classification, they have critical challenges in terms defining efficient and adaptive token sequences for improve performance. This paper therefore presents CSSMamba (Clustering-guided Spatial-Spectral Mamba) framework to better address the challenges, with the following contributions. First, to achieve efficient and adaptive token sequences for improved Mamba performance, we integrate the clustering mechanism into a spatial Mamba architecture, leading to a cluster-guided spatial Mamba module (CSpaMamba) that reduces the Mamba sequence length and improves Mamba feature learning capability. Second, to improve the learning of both spatial and spectral information, we integrate the CSpaMamba module with a spectral mamba module (SpeMamba), leading to a complete clustering-guided spatial-spectral Mamba framework. Third, to further improve feature learning capability, we introduce an Attention-Driven Token Selection mechanism to optimize Mamba token sequencing. Last, to seamlessly integrate clustering into the Mamba model in a coherent manner, we design a Learnable Clustering Module that learns the cluster memberships in an adaptive manner. Experiments on the Pavia University, Indian Pines, and Liao-Ning 01 datasets demonstrate that CSSMamba achieves higher accuracy and better boundary preservation compared to state-of-the-art CNN, Transformer, and Mamba-based methods.", "authors": ["Zack Dewis", "Yimin Zhu", "Zhengsen Xu", "Mabel Heffring", "Saeid Taleghanidoozdoozan", "Quinn Ledingham", "Lincoln Linlin Xu"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-22", "url": "https://arxiv.org/abs/2601.16098", "pdf_url": "https://arxiv.org/pdf/2601.16098v1", "arxiv_id": "2601.16098", "doi": "10.48550/arXiv.2601.16098", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4526}
{"id": "c68300df7d856da59b5d9c50d52ce3e0d5d84025c170ed9902ac83e80f3a3a80", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Experts Models in Vision: Routing, Optimization, and Generalization", "abstract": "Mixture-of-Experts (MoE) architectures enable conditional computation by routing inputs to multiple expert subnetworks and are often motivated as a mechanism for scaling large language models. In this project, we instead study MoE behavior in an image classification setting, focusing on predictive performance, expert utilization, and generalization. We compare dense, SoftMoE, and SparseMoE classifier heads on the CIFAR10 dataset under comparable model capacity. Both MoE variants achieve slightly higher validation accuracy than the dense baseline while maintaining balanced expert utilization through regularization, avoiding expert collapse. To analyze generalization, we compute Hessian-based sharpness metrics at convergence, including the largest eigenvalue and trace of the loss Hessian, evaluated on both training and test data. We find that SoftMoE exhibits higher sharpness by these metrics, while Dense and SparseMoE lie in a similar curvature regime, despite all models achieving comparable generalization performance. Complementary loss surface perturbation analyses reveal qualitative differences in non-local behavior under finite parameter perturbations between dense and MoE models, which help contextualize curvature-based measurements without directly explaining validation accuracy. We further evaluate empirical inference efficiency and show that naively implemented conditional routing does not yield inference speedups on modern hardware at this scale, highlighting the gap between theoretical and realized efficiency in sparse MoE models.", "authors": ["Adam Rokah", "Daniel Veress", "Caleb Caulk", "Sourav Sharan"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-21", "url": "https://arxiv.org/abs/2601.15021", "pdf_url": "https://arxiv.org/pdf/2601.15021v1", "arxiv_id": "2601.15021", "doi": "10.48550/arXiv.2601.15021", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/moe-project-uu/mixture-of-experts-project", "venue": "arXiv.org", "quality_score": 0.6977}
{"id": "ec2b9cb9f2183fe88bcd53af6a574f8e3f50eddec1162cec0e0cb2246a6f711b", "sources": ["arxiv", "semantic_scholar"], "title": "HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding", "abstract": "Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated significant improvement in offline video understanding. However, extending these capabilities to streaming video inputs, remains challenging, as existing models struggle to simultaneously maintain stable understanding performance, real-time responses, and low GPU memory overhead. To address this challenge, we propose HERMES, a novel training-free architecture for real-time and accurate understanding of video streams. Based on a mechanistic attention investigation, we conceptualize KV cache as a hierarchical memory framework that encapsulates video information across multiple granularities. During inference, HERMES reuses a compact KV cache, enabling efficient streaming understanding under resource constraints. Notably, HERMES requires no auxiliary computations upon the arrival of user queries, thereby guaranteeing real-time responses for continuous video stream interactions, which achieves 10$\\times$ faster TTFT compared to prior SOTA. Even when reducing video tokens by up to 68% compared with uniform sampling, HERMES achieves superior or comparable accuracy across all benchmarks, with up to 11.4% gains on streaming datasets.", "authors": ["Haowei Zhang", "Shudong Yang", "Jinlan Fu", "See-Kiong Ng", "Xipeng Qiu"], "categories": ["cs.CV", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-21", "url": "https://arxiv.org/abs/2601.14724", "pdf_url": "https://arxiv.org/pdf/2601.14724v4", "arxiv_id": "2601.14724", "doi": "10.48550/arXiv.2601.14724", "citation_count": 12, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4515}
{"id": "8c459d10d9d858873362f8df999211d52990a7d734a719c5f318783d650d6b5f", "sources": ["arxiv", "semantic_scholar"], "title": "Robustness of Mixtures of Experts to Feature Noise", "abstract": "Despite their practical success, it remains unclear why Mixture of Experts (MoE) models can outperform dense networks beyond sheer parameter scaling. We study an iso-parameter regime where inputs exhibit latent modular structure but are corrupted by feature noise, a proxy for noisy internal activations. We show that sparse expert activation acts as a noise filter: compared to a dense estimator, MoEs achieve lower generalization error under feature noise, improved robustness to perturbations, and faster convergence speed. Empirical results on synthetic data and real-world language tasks corroborate the theoretical insights, demonstrating consistent robustness and efficiency gains from sparse modular computation.", "authors": ["Dong Sun", "Rahul Nittala", "Rebekka Burkholz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-21", "url": "https://arxiv.org/abs/2601.14792", "pdf_url": "https://arxiv.org/pdf/2601.14792v2", "arxiv_id": "2601.14792", "doi": "10.48550/arXiv.2601.14792", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4515}
{"id": "75d6ac27d16bd822f1ae754df3a958b2c875d9e3d8a8705dee6459f1cfeb3a85", "sources": ["arxiv", "semantic_scholar"], "title": "ContiguousKV: Accelerating LLM Prefill with Granularity-Aligned KV Cache Management", "abstract": "Efficiently serving Large Language Models (LLMs) with persistent Prefix Key-Value (KV) Cache is critical for applications like conversational search and multi-turn dialogue. Serving a request requires loading the pre-computed prefix KV cache and generating the first token, defined as the Re-Prefill Phase. Offloading this shared prefix cache to secondary storage is essential for memory scalability. Re-Prefill with offloading suffers from severe I/O bottlenecks in two aspects. First, semantic-aware KV cache pruning algorithms select important tokens in fine granularity, while systems manage I/O in coarse, fixed-size blocks, causing severe read amplification. Second, the sequential dependency between identifying important tokens and loading KV cache creates idle I/O and compute bubbles, under-utilizing system resources. This paper proposes \\textit{ContiguousKV}, a high-performance prefix KV cache offloading system that bridges algorithmic semantics with I/O efficiency to accelerate the Re-Prefill phase. We first introduce \\textit{ContiguousChunk}, a unified data management granularity that aligns KV cache pruning with I/O operations. All the mechanisms critical for I/O performance are performed at the granularity of ContiguousChunk, thereby eliminating read amplification. By exploiting the high similarity in important ContiguousChunk indices across layers, we propose intra- and inter-period asynchronous prefetching to break the sequential dependency between I/O and compute, effectively eliminating idle bubbles. Finally, we propose attention-guided cache management to retain semantically critical prefix data in memory. Evaluations on Qwen2.5 series models show that ContiguousKV achieves a 3.85x speedup in the Re-Prefill phase over the state-of-the-art offloading system IMPRESS, while maintaining high output quality.", "authors": ["Jing Zou", "Shangyu Wu", "Hancong Duan", "Qiao Li", "Chun Jason Xue"], "categories": ["cs.OS", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-20", "url": "https://arxiv.org/abs/2601.13631", "pdf_url": "https://arxiv.org/pdf/2601.13631v1", "arxiv_id": "2601.13631", "doi": "10.48550/arXiv.2601.13631", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4503}
{"id": "1e9b467e183ce8590e7f6265dbe578374760c88ef85f3ecb097a50ae6685088f", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding Multilingualism in Mixture-of-Experts LLMs: Routing Mechanism, Expert Specialization, and Layerwise Steering", "abstract": "Mixture-of-Experts (MoE) architectures have shown strong multilingual capabilities, yet the internal mechanisms underlying performance gains and cross-language differences remain insufficiently understood. In this work, we conduct a systematic analysis of MoE models, examining routing behavior and expert specialization across languages and network depth. Our analysis reveals that multilingual processing in MoE models is highly structured: routing aligns with linguistic families, expert utilization follows a clear layerwise pattern, and high-resource languages rely on shared experts while low-resource languages depend more on language-exclusive experts despite weaker performance. Layerwise interventions further show that early and late MoE layers support language-specific processing, whereas middle layers serve as language-agnostic capacity hubs. Building on these insights, we propose a routing-guided steering method that adaptively guides routing behavior in middle layers toward shared experts associated with dominant languages at inference time, leading to consistent multilingual performance improvements, particularly for linguistically related language pairs. Our code is available at https://github.com/conctsai/Multilingualism-in-Mixture-of-Experts-LLMs.", "authors": ["Yuxin Chen", "Zhengzhou Cai", "Xiangtian Ji", "Weixiang Zhao", "An Zhang", "Xiang Wang", "Tat-Seng Chua"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-20", "url": "https://arxiv.org/abs/2601.14050", "pdf_url": "https://arxiv.org/pdf/2601.14050v1", "arxiv_id": "2601.14050", "doi": "10.48550/arXiv.2601.14050", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/conctsai/Multilingualism-in-Mixture-of-Experts-LLMs", "venue": "arXiv.org", "quality_score": 0.6959}
{"id": "3d55190638a5459cf0a60fe3f4d132f7f7e23d280ce11162ef58c5df620b03dd", "sources": ["arxiv", "semantic_scholar"], "title": "HeteroCache: A Dynamic Retrieval Approach to Heterogeneous KV Cache Compression for Long-Context LLM Inference", "abstract": "The linear memory growth of the KV cache poses a significant bottleneck for LLM inference in long-context tasks. Existing static compression methods often fail to preserve globally important information. Although recent dynamic retrieval approaches attempt to address this issue, they typically suffer from coarse-grained caching strategies and incur high I/O overhead. To overcome these limitations, we propose HeteroCache, a training-free dynamic compression framework. Our method is built on two key insights: attention heads exhibit diverse temporal heterogeneity, and there is significant spatial redundancy among heads within the same layer. Guided by these insights, HeteroCache categorizes heads based on stability and similarity, applying a fine-grained weighting strategy that allocates larger cache budgets to heads with rapidly shifting attention to capture context changes. Furthermore, it features a hierarchical storage mechanism where representative heads monitor attention drift to trigger asynchronous, on-demand context retrieval, thereby hiding I/O latency. Experiments demonstrate that HeteroCache achieves state-of-the-art performance on long-context benchmarks and accelerates decoding by up to $3\\times$ compared to the original model with a 224K context. Our code is available at https://github.com/ponytaill/HeteroCache.", "authors": ["Zhiyuan Shi", "Qibo Qiu", "Feng Xue", "Zhonglin Jiang", "Li Yu", "Jian Jiang", "Xiaofei He", "Wenxiao Wang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-20", "url": "https://arxiv.org/abs/2601.13684", "pdf_url": "https://arxiv.org/pdf/2601.13684v2", "arxiv_id": "2601.13684", "doi": "10.48550/arXiv.2601.13684", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ponytaill/HeteroCache", "venue": "arXiv.org", "quality_score": 0.6959}
{"id": "70744c993cb0b5ba2c708aaf8fa3a0c729b8095ca53270e277409150805769fc", "sources": ["arxiv", "semantic_scholar"], "title": "ButterflyMoE: Sub-Linear Ternary Experts via Structured Butterfly Orbits", "abstract": "Linear memory scaling stores $N$ independent expert weight matrices requiring $\\mathcal{O}(N \\cdot d^2)$ memory, which exceeds edge devices memory budget. Current compression methods like quantization, pruning and low-rank factorization reduce constant factors but leave the scaling bottleneck unresolved. We introduce ButterflyMoE, a method that treats experts not as independent weight matrices but as geometric reorientations of a unified shared quantized substrate. Diversity among experts arises from viewing different angles of shared capacity, not from redundant storage. By applying learned rotations to a shared ternary prototype, each expert yields $\\mathcal{O}(d^2 + N \\cdot d \\log d)$ memory,sub-linear in the number of experts. The key insight: training these rotations with quantization reduces activation outliers and stabilizes extreme low bit training, where static methods collapse. Across language modeling benchmarks, ButterflyMoE achieves 150$\\times$ memory reduction at 256 experts with negligible accuracy loss. ButterflyMoE allows multiple experts to fit on edge-constrained devices showing that geometric parameterization breaks linear scaling.", "authors": ["Aryan Karmore"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-20", "url": "https://arxiv.org/abs/2601.13563", "pdf_url": "https://arxiv.org/pdf/2601.13563v4", "arxiv_id": "2601.13563", "doi": "10.48550/arXiv.2601.13563", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4503}
{"id": "6a97a24e4c9fd18821d7cd4f4a68c7ba1cd98dc572ee3a5d5a35528f5b6cf709", "sources": ["arxiv", "semantic_scholar"], "title": "Topology-Aware Multiscale Mixture of Experts for Efficient Molecular Property Prediction", "abstract": "Many molecular properties depend on 3D geometry, where non-covalent interactions, stereochemical effects, and medium- to long-range forces are determined by spatial distances and angles that cannot be uniquely captured by a 2D bond graph. Yet most 3D molecular graph neural networks still rely on globally fixed neighborhood heuristics, typically defined by distance cutoffs and maximum neighbor limits, to define local message-passing neighborhoods, leading to rigid, data-agnostic interaction budgets. We propose Multiscale Interaction Mixture of Experts (MI-MoE) to adapt interaction modeling across geometric regimes. Our contributions are threefold: (1) we introduce a distance-cutoff expert ensemble that explicitly captures short-, mid-, and long-range interactions without committing to a single cutoff; (2) we design a topological gating encoder that routes inputs to experts using filtration-based descriptors, including persistent homology features, summarizing how connectivity evolves across radii; and (3) we show that MI-MoE is a plug-in module that consistently improves multiple strong 3D molecular backbones across diverse molecular and polymer property prediction benchmark datasets, covering both regression and classification tasks. These results highlight topology-aware multiscale routing as an effective principle for 3D molecular graph learning.", "authors": ["Long D. Nguyen", "Kelin Xia", "Binh P. Nguyen"], "categories": ["cs.LG", "cs.AI", "q-bio.QM"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2026-01-19", "url": "https://arxiv.org/abs/2601.12637", "pdf_url": "https://arxiv.org/pdf/2601.12637v1", "arxiv_id": "2601.12637", "doi": "10.48550/arXiv.2601.12637", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4492}
{"id": "ff37156af496e86c77bd14a45cbcd519113d277b644f9aac5157b8287ad90c78", "sources": ["arxiv", "semantic_scholar"], "title": "Actionable Advice from Reviews via Mixture of LoRA Experts: A Two-LLM Pipeline for Issue Extraction and Business Recommendations", "abstract": "Customer reviews contain detailed, domain specific signals about service failures and user expectations, but converting this unstructured feedback into actionable business decisions remains difficult. We study review-to-action generation: producing concrete, implementable recommendations grounded in review text. We propose a modular two-LLM framework in which an Issue model extracts salient issues and assigns coarse themes, and an Advice model generates targeted operational fixes conditioned on the extracted issue representation. To enable specialization without expensive full fine-tuning, we adapt the Advice model using a mixture of LoRA experts strategy: multiple low-rank adapters are trained and a lightweight gating mechanism performs token-level expert mixing at inference, combining complementary expertise across issue types. We construct synthetic review-issue-advice triples from Yelp reviews (airlines and restaurants) to supervise training, and evaluate recommendations using an eight dimension operational rubric spanning actionability, specificity, feasibility, expected impact, novelty, non-redundancy, bias, and clarity. Across both domains, our approach consistently outperforms prompting-only and single-adapter baselines, yielding higher actionability and specificity while retaining favorable efficiency-quality trade-offs.", "authors": ["Kartikey Singh Bhandari", "Manav Ganesh", "Yashwant Viswanathan", "Archit Agrawal", "Dhruv Kumar", "Pratik Narang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-18", "url": "https://arxiv.org/abs/2601.12338", "pdf_url": "https://arxiv.org/pdf/2601.12338v1", "arxiv_id": "2601.12338", "doi": "10.48550/arXiv.2601.12338", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.448}
{"id": "572d07523ca3dd01391793fc2360a668d16b00f47f1a23554c34b0d01f48ac1f", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Diverse Skills for Behavior Models with Mixture of Experts", "abstract": "Imitation learning has demonstrated strong performance in robotic manipulation by learning from large-scale human demonstrations. While existing models excel at single-task learning, it is observed in practical applications that their performance degrades in the multi-task setting, where interference across tasks leads to an averaging effect. To address this issue, we propose to learn diverse skills for behavior models with Mixture of Experts, referred to as Di-BM. Di-BM associates each expert with a distinct observation distribution, enabling experts to specialize in sub-regions of the observation space. Specifically, we employ energy-based models to represent expert-specific observation distributions and jointly train them alongside the corresponding action models. Our approach is plug-and-play and can be seamlessly integrated into standard imitation learning methods. Extensive experiments on multiple real-world robotic manipulation tasks demonstrate that Di-BM significantly outperforms state-of-the-art baselines. Moreover, fine-tuning the pretrained Di-BM on novel tasks exhibits superior data efficiency and the reusable of expert-learned knowledge. Code is available at https://github.com/robotnav-bot/Di-BM.", "authors": ["Wangtian Shen", "Jinming Ma", "Mingliang Zhou", "Ziyang Meng"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-18", "url": "https://arxiv.org/abs/2601.12397", "pdf_url": "https://arxiv.org/pdf/2601.12397v1", "arxiv_id": "2601.12397", "doi": "10.48550/arXiv.2601.12397", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/robotnav-bot/Di-BM", "venue": "arXiv.org", "quality_score": 0.6924}
{"id": "4a633722ab848010d8eddea89ed516e0ae257ab21164a41f334c5ccd38fb2c20", "sources": ["arxiv", "semantic_scholar"], "title": "EMoE: Eigenbasis-Guided Routing for Mixture-of-Experts", "abstract": "The relentless scaling of deep learning models has led to unsustainable computational demands, positioning Mixture-of-Experts (MoE) architectures as a promising path towards greater efficiency. However, MoE models are plagued by two fundamental challenges: 1) a load imbalance problem known as the``rich get richer\" phenomenon, where a few experts are over-utilized, and 2) an expert homogeneity problem, where experts learn redundant representations, negating their purpose. Current solutions typically employ an auxiliary load-balancing loss that, while mitigating imbalance, often exacerbates homogeneity by enforcing uniform routing at the expense of specialization. To resolve this, we introduce the Eigen-Mixture-of-Experts (EMoE), a novel architecture that leverages a routing mechanism based on a learned orthonormal eigenbasis. EMoE projects input tokens onto this shared eigenbasis and routes them based on their alignment with the principal components of the feature space. This principled, geometric partitioning of data intrinsically promotes both balanced expert utilization and the development of diverse, specialized experts, all without the need for a conflicting auxiliary loss function. Our code is publicly available at https://github.com/Belis0811/EMoE.", "authors": ["Anzhe Cheng", "Shukai Duan", "Shixuan Li", "Chenzhong Yin", "Mingxi Cheng", "Shahin Nazarian", "Paul Thompson", "Paul Bogdan"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-17", "url": "https://arxiv.org/abs/2601.12137", "pdf_url": "https://arxiv.org/pdf/2601.12137v1", "arxiv_id": "2601.12137", "doi": "10.48550/arXiv.2601.12137", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Belis0811/EMoE", "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.6906}
{"id": "1f6ee9eeb79bac95fe2c8633686439b3fe4c4f13ff649bee7d5fa3de9bc75cf6", "sources": ["arxiv", "semantic_scholar"], "title": "MoST: Mixing Speech and Text with Modality-Aware Mixture of Experts", "abstract": "We present MoST (Mixture of Speech and Text), a novel multimodal large language model that seamlessly integrates speech and text processing through our proposed Modality-Aware Mixture of Experts (MAMoE) architecture. While current multimodal models typically process diverse modality representations with identical parameters, disregarding their inherent representational differences, we introduce specialized routing pathways that direct tokens to modality-appropriate experts based on input type. MAMoE simultaneously enhances modality-specific learning and cross-modal understanding through two complementary components: modality-specific expert groups that capture domain-specific patterns and shared experts that facilitate information transfer between modalities. Building on this architecture, we develop an efficient transformation pipeline that adapts the pretrained MoE language model through strategic post-training on ASR and TTS datasets, followed by fine-tuning with a carefully curated speech-text instruction dataset. A key feature of this pipeline is that it relies exclusively on fully accessible, open-source datasets to achieve strong performance and data efficiency. Comprehensive evaluations across ASR, TTS, audio language modeling, and spoken question answering benchmarks show that MoST consistently outperforms existing models of comparable parameter counts. Our ablation studies confirm that the modality-specific routing mechanism and shared experts design significantly contribute to performance gains across all tested domains. To our knowledge, MoST represents the first fully open-source speech-text LLM built on a Mixture of Experts architecture. \\footnote{We release MoST model, training code, inference code, and training data at https://github.com/NUS-HPC-AI-Lab/MoST", "authors": ["Yuxuan Lou", "Kai Yang", "Yang You"], "categories": ["cs.CL", "cs.AI", "cs.LG", "cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-15", "url": "https://arxiv.org/abs/2601.10272", "pdf_url": "https://arxiv.org/pdf/2601.10272v1", "arxiv_id": "2601.10272", "doi": "10.48550/arXiv.2601.10272", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/NUS-HPC-AI-Lab/MoST", "venue": "arXiv.org", "quality_score": 0.6871}
{"id": "461b0c4022e647115e3a9acb07091ca072385b778a513e5a94053144545ab327", "sources": ["arxiv", "semantic_scholar"], "title": "LOOKAT: Lookup-Optimized Key-Attention for Memory-Efficient Transformers", "abstract": "Compressing the KV cache is a required step to deploy large language models on edge devices. Current quantization methods compress storage but fail to reduce bandwidth as attention calculation requires dequantizing keys from INT4/INT8 to FP16 before use. We observe that attention scoring is mathematically equivalent to the inner product similarity search and we can apply some compression techniques from vector databases to compress KV-cache better. We propose LOOKAT, which applies product quantization and asymmetric distance computation, to transformer architecture by decomposing key vectors into subspaces, learning codebooks and computing attention tables via lookup tables. This transforms attention from memory-bound to compute-bound. LOOKAT achieves 64 $\\times$ compression at 95.7\\% output fidelity and 32 $\\times$ compression at 95.0\\% fidelity when tested on GPT-2. LOOKAT requires no architecture changes or training while maintaining rank correlation $ρ> 0.95$. Theoretical analysis confirms that rank correlation degrades as $O(d_k/mK)$, with guarantees validated across sequence lengths up to 1024 tokens.", "authors": ["Aryan Karmore"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-15", "url": "https://arxiv.org/abs/2601.10155", "pdf_url": "https://arxiv.org/pdf/2601.10155v1", "arxiv_id": "2601.10155", "doi": "10.48550/arXiv.2601.10155", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4446}
{"id": "fe13e99e36545b2df993d7bb01ea5b52cbdcceb71c6bb4c3b916acb1bb30027f", "sources": ["arxiv", "semantic_scholar"], "title": "Horseshoe Mixtures-of-Experts (HS-MoE)", "abstract": "Horseshoe mixtures-of-experts (HS-MoE) models provide a Bayesian framework for sparse expert selection in mixture-of-experts architectures. We combine the horseshoe prior's adaptive global-local shrinkage with input-dependent gating, yielding data-adaptive sparsity in expert usage. Our primary methodological contribution is a particle learning algorithm for sequential inference, in which the filter is propagated forward in time while tracking only sufficient statistics. We also discuss how HS-MoE relates to modern mixture-of-experts layers in large language models, which are deployed under extreme sparsity constraints (e.g., activating a small number of experts per token out of a large pool).", "authors": ["Nick Polson", "Vadim Sokolov"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-01-14", "url": "https://arxiv.org/abs/2601.09043", "pdf_url": "https://arxiv.org/pdf/2601.09043v1", "arxiv_id": "2601.09043", "doi": "10.48550/arXiv.2601.09043", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4434}
{"id": "9638d74686276314e06363a454a114ae50b9a2f8600cdf8898d6b3749f4b1b6d", "sources": ["arxiv", "semantic_scholar"], "title": "On the Limits of Learned Importance Scoring for KV Cache Compression", "abstract": "We investigate learned KV cache compression through Speculative Importance Prediction (SIP), a 1.7M parameter non-query-aware scorer that predicts token importance from KV representations alone. Despite architectural sophistication (multi-horizon lookahead, cross-attention), SIP does not outperform simple baselines, including random selection, across 5 seeds, 4 retention levels, and 3 tasks. Key findings: (1) position-based heuristics (keep first 4 + last N tokens) match or exceed learned approaches; (2) prefill attention provides equivalent signal to complex learned scorers; (3) marginal information in KV representations beyond position and prefill attention appears limited for importance prediction. We hypothesize that circular dependence between future queries and generation trajectories contributes to this difficulty.", "authors": ["Brady Steele"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-13", "url": "https://arxiv.org/abs/2601.14279", "pdf_url": "https://arxiv.org/pdf/2601.14279v1", "arxiv_id": "2601.14279", "doi": "10.48550/arXiv.2601.14279", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4423}
{"id": "309e623ed4fae58053636be06004c77eea28addddf38823d9f06140138092b0a", "sources": ["arxiv", "semantic_scholar"], "title": "TableCache: Primary Foreign Key Guided KV Cache Precomputation for Low Latency Text-to-SQL", "abstract": "In Text-to-SQL tasks, existing LLM-based methods often include extensive database schemas in prompts, leading to long context lengths and increased prefilling latency. While user queries typically focus on recurrent table sets-offering an opportunity for KV cache sharing across queries-current inference engines, such as SGLang and vLLM, generate redundant prefix cache copies when processing user queries with varying table orders. To address this inefficiency, we propose precomputing table representations as KV caches offline and querying the required ones online. A key aspect of our approach is the computation of table caches while preserving primary foreign key relationships between tables. Additionally, we construct a Table Trie structure to facilitate efficient KV cache lookups during inference. To enhance cache performance, we introduce a cache management system with a query reranking strategy to improve cache hit rates and a computation loading pipeline for parallelizing model inference and cache loading. Experimental results show that our proposed TableCache achieves up to a 3.62x speedup in Time to First Token (TTFT) with negligible performance degradation.", "authors": ["Jinbo Su", "Yuxuan Hu", "Cuiping Li", "Hong Chen", "Jia Li", "Lintao Ma", "Jing Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-13", "url": "https://arxiv.org/abs/2601.08743", "pdf_url": "https://arxiv.org/pdf/2601.08743v1", "arxiv_id": "2601.08743", "doi": "10.48550/arXiv.2601.08743", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4423}
{"id": "0f2b022fd563bcb9eb39d7c6ef95b68a073fc26064b4b25bd3f643aaa125db7e", "sources": ["arxiv", "semantic_scholar"], "title": "When KV Cache Reuse Fails in Multi-Agent Systems: Cross-Candidate Interaction is Crucial for LLM Judges", "abstract": "Multi-agent LLM systems routinely generate multiple candidate responses that are aggregated by an LLM judge. To reduce the dominant prefill cost in such pipelines, recent work advocates KV cache reuse across partially shared contexts and reports substantial speedups for generation agents. In this work, we show that these efficiency gains do not transfer uniformly to judge-centric inference. Across GSM8K, MMLU, and HumanEval, we find that reuse strategies that are effective for execution agents can severely perturb judge behavior: end-task accuracy may appear stable, yet the judge's selection becomes highly inconsistent with dense prefill. We quantify this risk using Judge Consistency Rate (JCR) and provide diagnostics showing that reuse systematically weakens cross-candidate attention, especially for later candidate blocks. Our ablation further demonstrates that explicit cross-candidate interaction is crucial for preserving dense-prefill decisions. Overall, our results identify a previously overlooked failure mode of KV cache reuse and highlight judge-centric inference as a distinct regime that demands dedicated, risk-aware system design.", "authors": ["Sichu Liang", "Zhenglin Wang", "Jiajia Chu", "Pengfei Xia", "Hui Zang", "Deyu Zhou"], "categories": ["cs.MA", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-13", "url": "https://arxiv.org/abs/2601.08343", "pdf_url": "https://arxiv.org/pdf/2601.08343v1", "arxiv_id": "2601.08343", "doi": "10.48550/arXiv.2601.08343", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4423}
{"id": "7679f13beec95e066bcd457f343856e5a5fa9bc71fcb812d885f8afe5955e4e4", "sources": ["arxiv", "semantic_scholar"], "title": "KVzap: Fast, Adaptive, and Faithful KV Cache Pruning", "abstract": "Growing context lengths in transformer-based language models have made the key-value (KV) cache a critical inference bottleneck. While many KV cache pruning methods have been proposed, they have not yet been adopted in major inference engines due to speed--accuracy trade-offs. We introduce KVzap, a fast, input-adaptive approximation of KVzip that works in both prefilling and decoding. On Qwen3-8B, Llama-3.1-8B-Instruct, and Qwen3-32B across long-context and reasoning tasks, KVzap achieves $2$--$4\\times$ KV cache compression with negligible accuracy loss and achieves state-of-the-art performance on the KVpress leaderboard. Code and models are available at https://github.com/NVIDIA/kvpress.", "authors": ["Simon Jegou", "Maximilian Jeblick"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-12", "url": "https://arxiv.org/abs/2601.07891", "pdf_url": "https://arxiv.org/pdf/2601.07891v2", "arxiv_id": "2601.07891", "doi": "10.48550/arXiv.2601.07891", "citation_count": 8, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/NVIDIA/kvpress", "venue": "arXiv.org", "quality_score": 0.6818}
{"id": "7278e81ef9098d766f6d3acb7d2db59e9d11c200a2689c31cedf8960958d2031", "sources": ["arxiv", "semantic_scholar"], "title": "Gated Sparse Attention: Combining Computational Efficiency with Training Stability for Long-Context Language Models", "abstract": "The computational burden of attention in long-context language models has motivated two largely independent lines of work: sparse attention mechanisms that reduce complexity by attending to selected tokens, and gated attention variants that improve training sta-bility while mitigating the attention sink phenomenon. We observe that these approaches address complementary weaknesses and propose Gated Sparse Attention (GSA), an architecture that realizes the benefits of both. GSA incorporates a gated lightning indexer with sigmoid activations that produce bounded, interpretable selection scores, an adaptive sparsity controller that modulates the number of attended tokens based on local uncertainty, and dual gating at the value and output stages. We establish theoretical foundations for the approach, including complexity analysis, expressiveness results, and convergence guarantees. In experiments with 1.7B parameter models trained on 400B tokens, GSA matches the efficiency of sparse-only baselines (12-16x speedup at 128K context) while achieving the quality gains associated with gated attention: perplexity improves from 6.03 to 5.70, RULER scores at 128K context nearly double, and attention to the first token, a proxy for attention sinks, drops from 47% to under 4%. Training stability improves markedly, with loss spikes reduced by 98%.", "authors": ["Alfred Shen", "Aaron Shen"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-12", "url": "https://arxiv.org/abs/2601.15305", "pdf_url": "https://arxiv.org/pdf/2601.15305v1", "arxiv_id": "2601.15305", "doi": "10.48550/arXiv.2601.15305", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4411}
{"id": "2335fbd3edc9bf03c67ea4a649574dd28723e6a546600eedc48599d7efc34084", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Layer Selection for Layer-Wise Token Pruning in LLM Inference", "abstract": "Due to the prevalence of large language models (LLMs), key-value (KV) cache reduction for LLM inference has received remarkable attention. Among numerous works that have been proposed in recent years, layer-wise token pruning approaches, which select a subset of tokens at particular layers to retain in KV cache and prune others, are one of the most popular schemes. They primarily adopt a set of pre-defined layers, at which tokens are selected. Such design is inflexible in the sense that the accuracy significantly varies across tasks and deteriorates in harder tasks such as KV retrieval. In this paper, we propose ASL, a training-free method that adaptively chooses the selection layer for KV cache reduction, exploiting the variance of token ranks ordered by attention score. The proposed method balances the performance across different tasks while meeting the user-specified KV budget requirement. ASL operates during the prefilling stage and can be jointly used with existing KV cache reduction methods such as SnapKV to optimize the decoding stage. By evaluations on the InfiniteBench, RULER, and NIAH benchmarks, we show that ASL, equipped with one-shot token selection, adaptively trades inference speed for accuracy, outperforming state-of-the-art layer-wise token pruning methods in difficult tasks.", "authors": ["Rei Taniguchi", "Yuyang Dong", "Makoto Onizuka", "Chuan Xiao"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-12", "url": "https://arxiv.org/abs/2601.07667", "pdf_url": "https://arxiv.org/pdf/2601.07667v2", "arxiv_id": "2601.07667", "doi": "10.48550/arXiv.2601.07667", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/TANIGUCHIREI/ASL", "venue": "arXiv.org", "quality_score": 0.6818}
{"id": "bf5c949f533fc8087422d933d85141d95c8329d7f287e32defbb7d2f9ce44aab", "sources": ["arxiv", "semantic_scholar"], "title": "CRAFT: Fine-Grained Cost-Aware Expert Replication For Efficient Mixture-of-Experts Serving", "abstract": "Mixture-of-Experts (MoE) has recently emerged as the mainstream architecture for efficiently scaling large language models while maintaining near-constant computational cost. Expert parallelism distributes parameters by partitioning experts across devices, but this introduces token-level load imbalance during inference. Expert replication is a widely adopted load-balancing technique in serving frameworks that alleviates load imbalance in large-scale deployments by replicating experts with high loads. In this work, we demonstrate that existing replication schemes often over-replicate, with many replicas providing marginal improvement. Replicas consume substantial GPU memory, which may lead to resource contention and throughput degradation. We present CRAFT, an efficient expert replication framework that maximizes load balance under a given memory budget by performing fine-grained, per-layer replication based on the estimated replication benefit. CRAFT can be seamlessly integrated into existing serving frameworks without any additional training or model changes. Our evaluation shows that CRAFT increases end-to-end serving throughput by $1.14\\times$ on average (up to $1.2\\times$) over existing replication techniques in large-scale deployments with models ranging from hundreds of billions to a trillion parameters.", "authors": ["Adrian Zhao", "Zhenkun Cai", "Zhenyu Song", "Lingfan Yu", "Haozheng Fan", "Jun Wu", "Yida Wang", "Nandita Vijaykumar"], "categories": ["cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-12", "url": "https://arxiv.org/abs/2603.28768", "pdf_url": "https://arxiv.org/pdf/2603.28768v2", "arxiv_id": "2603.28768", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Proceedings of the Ninth Conference on Machine Learning and Systems (MLSys 2026)", "quality_score": 0.4411}
{"id": "227b992a7adae049cfc6ee4ab94f38db76e1176c2a19c0eb73cebe93c0c6a499", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Specialized Generalists: A Multi-Task MoE-LoRA Framework for Domain-Specific LLM Adaptation", "abstract": "The rapid evolution of Large Language Models (LLMs) has shifted focus from general-purpose capabilities to domain-specific expertise. However, adapting LLMs to specialized fields such as medicine presents two challenge: (1) the \"Stability-Plasticity Dilemma\", where the model must acquire complex clinical knowledge without suffering from catastrophic forgetting of general world knowledge; and (2) \"Task Interference\", where disparate sub-tasks, such as medical diagnosis, report summarization, and drug-drug interaction prediction, compete for limited low-rank parameter space. In this paper, we propose Med-MoE-LoRA, a novel framework that integrates Mixture-of-Experts (MoE) with Low-Rank Adaptation (LoRA) to enable efficient multi-task domain adaptation, especially for medical scenarios. Drawing inspiration from recent advances, our framework employs an asymmetric expert distribution where deeper layers are equipped with a higher density of LoRA experts to capture complex semantic abstractions. We further introduce a \"Knowledge-Preservation Plugin\", inspired by LoRA MoE, to isolate and protect general-purpose reasoning. By utilizing soft merging with adaptive routing and rank-wise decoupling, Med-MoE-LoRA achieves superior performance in medical benchmarks while reducing interference. Experimental results demonstrate that our approach consistently outperforms standard LoRA and conventional MoE architectures across multiple clinical NLP tasks while retaining the model's general cognitive capabilities.", "authors": ["Yuxin Yang", "Aoxiong Zeng", "Xiangquan Yang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-12", "url": "https://arxiv.org/abs/2601.07935", "pdf_url": "https://arxiv.org/pdf/2601.07935v1", "arxiv_id": "2601.07935", "doi": "10.48550/arXiv.2601.07935", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4411}
{"id": "2ede4dec107d0bf7f40531254415bc3119f9e631585fcd9105914f577db10f31", "sources": ["arxiv", "semantic_scholar"], "title": "Garbage Attention in Large Language Models: BOS Sink Heads and Sink-aware Pruning", "abstract": "Large Language Models (LLMs) are known to contain significant redundancy, yet a systematic explanation for why certain components, particularly in higher layers, are more redundant has remained elusive. In this work, we identify the BOS sink phenomenon as a key mechanism driving this layer-wise sensitivity. We show that attention heads with high BOS sink scores are strongly associated with functional redundancy: such heads, especially in deeper layers, contribute little to predictive performance and effectively serve as \\emph{dumping grounds} for superfluous attention weights. This provides a concrete functional explanation for the structural redundancy reported in prior studies. Leveraging this insight, we introduce a simple pruning strategy that removes high-BOS sink heads. Experiments on Gemma-3, Llama-3.1, and Qwen3 demonstrate that this approach identifies redundant transformer components more reliably than weight- or activation-based criteria, while preserving performance close to dense baselines even under aggressive pruning. Moreover, we find that the behavior of sink heads remains stable across different sequence lengths. Overall, our results suggest that structural properties of attention offer a more intuitive and robust basis for model compression than magnitude-based methods.", "authors": ["Jaewon Sok", "Jewon Yeom", "Seonghyeon Park", "Jeongjae Park", "Taesup Kim"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-11", "url": "https://arxiv.org/abs/2601.06787", "pdf_url": "https://arxiv.org/pdf/2601.06787v1", "arxiv_id": "2601.06787", "doi": "10.48550/arXiv.2601.06787", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.44}
{"id": "4df1244db8a0122524ad699eb57bc3ed7ace11d10ad3e529ee38d18abfbbfb74", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-DisCo:Low Economy Cost Training Mixture-of-Experts Models", "abstract": "Training large-scale Mixture-of-Experts (MoE) models typically requires high-memory, high-bandwidth GPUs (e.g., A100), and their high cost has become a major barrier to large-model training. In contrast, affordable hardware is low-cost but constrained by memory capacity and bandwidth, making it unsuitable for direct LLM training. To address this, we propose MoE-DisCo (Mixture-of-Experts with Disentangled Clustering and Coordination), a staged training framework. MoE-DisCo decomposes the MoE model into multiple dense submodels, each consisting of a shared backbone and a single expert, and partitions the training data into subsets using unsupervised clustering. Each submodel is trained independently and in parallel on its assigned data subset using low-cost devices, without any inter-device communication. Subsequently, all experts are integrated into a complete MoE model and fine-tuned globally for a short period on high-memory, high-bandwidth GPUs. Experiments show that our method matches or even surpasses full-parameter training in performance across multiple downstream tasks, loss function, and perplexity (PPL), while reducing training cost by 47.6 percent to 69.5 percent on Qwen1.5-MoE-2.7B and Llama-MoE-3.5B across different datasets.", "authors": ["Xin Ye", "Daning Cheng", "Boyang Zhang", "Yunquan Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-11", "url": "https://arxiv.org/abs/2601.06857", "pdf_url": "https://arxiv.org/pdf/2601.06857v1", "arxiv_id": "2601.06857", "doi": "10.48550/arXiv.2601.06857", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.44}
{"id": "11fda8ee4e298601be74a27e34af4486a70bada98da0b30a912e5de3e911023e", "sources": ["arxiv", "semantic_scholar"], "title": "GRASP LoRA: GRPO Guided Adapter Sparsity Policy for Cross Lingual Transfer", "abstract": "Parameter efficient fine tuning is a way to adapt LLMs to new languages when compute or data are limited, yet adapter pipelines usually choose a global prune ratio by grid search. This practice is computationally expensive and development set intensive, since it repeats training, freezes sparsity, and misses fractional optima. We introduce GRASP LoRA (GRPO Guided Adapter Sparsity Policy), which treats global sparsity as a learnable control variable. A GRPO controller interleaves with training, periodically probing candidate prune ratios on a small micro development set and updating a single global prune ratio online from its reward signal. It operates on merged source and target LoRA adapters on a frozen backbone and replaces grid search with one controller run that learns a prune ratio, followed by a single final merge and prune fine tuning run with pruning fixed to that ratio. On cross lingual transfer from English into Arabic and Chinese, including XL-Sum summarization and MLQA extractive question answering with Llama 3 8B, GRASP LoRA improves semantic faithfulness, content coverage, and answer quality over strong target only and merge and prune baselines. It reduces end to end runtime by multiple times relative to grid search, lowers reliance on large development sets, and makes adapter reuse practical for low resource deployment.", "authors": ["Besher Hassan", "Xiuying Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-10", "url": "https://arxiv.org/abs/2601.06702", "pdf_url": "https://arxiv.org/pdf/2601.06702v1", "arxiv_id": "2601.06702", "doi": "10.48550/arXiv.2601.06702", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4389}
{"id": "0d75bba4777a3dc8c4749fb4b874c039990d3d0fd1f10a47d814e0b3f2614bd3", "sources": ["arxiv", "semantic_scholar"], "title": "DR-LoRA: Dynamic Rank LoRA for Fine-Tuning Mixture-of-Experts Models", "abstract": "Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning methods, such as LoRA, are widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches typically assign identical LoRA ranks to all expert modules, ignoring the heterogeneous specialization of pretrained experts. This uniform allocation leads to a resource mismatch: task-relevant experts are under-provisioned, while less relevant ones receive redundant parameters. To address this, we propose DR-LoRA, a Dynamic Rank LoRA framework for fine-tuning pretrained MoE models. Specifically, DR-LoRA initializes all expert LoRA modules with a small active rank and uses an expert saliency score, which combines routing frequency and gradient-based rank importance, to identify which experts would benefit most from additional capacity. It then periodically expands the active ranks of the task-critical expert LoRA, progressively constructing a heterogeneous rank distribution tailored to the target task. Experiments on three MoE models across six tasks show that DR-LoRA consistently outperforms LoRA and other strong baselines, demonstrating that task-adaptive heterogeneous rank allocation is an effective strategy to improve active capacity utilization in MoE fine-tuning.", "authors": ["Guanzhi Deng", "Bo Li", "Ronghao Chen", "Xiujin Liu", "Zhuo Han", "Huacan Wang", "Lijie Wen", "Linqi Song"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-08", "url": "https://arxiv.org/abs/2601.04823", "pdf_url": "https://arxiv.org/pdf/2601.04823v5", "arxiv_id": "2601.04823", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2778}
{"id": "24b48fce1400d4ff75434ce96c84936a8cc4a73eeb3c9f463d5b63120880d4a6", "sources": ["arxiv", "semantic_scholar"], "title": "GPU-Accelerated INT8 Quantization for KV Cache Compression in Large Language Models", "abstract": "The key-value (KV) cache in large language models presents a significant memory bottleneck during inference, growing linearly with sequence length and often exceeding the memory footprint of model weights themselves. We implement and evaluate GPU-accelerated INT8 quantization for KV cache compression, achieving 4$\\times$ memory reduction with minimal accuracy degradation. We develop four CUDA kernel variants -- naive, tiled, coarsened, and vectorized -- and benchmark them across realistic workload sizes up to 1 billion elements. Our vectorized kernel achieves up to 1,694$\\times$ speedup over CPU baselines while maintaining reconstruction error below 0.004 and attention score error below 0.1 even for 8K-dimensional heads. These results demonstrate that INT8 quantization provides a practical approach for reducing memory pressure in LLM inference with negligible computational overhead (6--58ms) and minimal impact on downstream model behavior", "authors": ["Maanas Taneja", "Purab Shingvi"], "categories": ["cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-08", "url": "https://arxiv.org/abs/2601.04719", "pdf_url": "https://arxiv.org/pdf/2601.04719v1", "arxiv_id": "2601.04719", "doi": "10.48550/arXiv.2601.04719", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4366}
{"id": "d941441e02bd676da27ea312f2996241f10597eb8dacbf6fc83c261d26dbe990", "sources": ["arxiv", "semantic_scholar"], "title": "PackCache: A Training-Free Acceleration Method for Unified Autoregressive Video Generation via Compact KV-Cache", "abstract": "A unified autoregressive model is a Transformer-based framework that addresses diverse multimodal tasks (e.g., text, image, video) as a single sequence modeling problem under a shared token space. Such models rely on the KV-cache mechanism to reduce attention computation from O(T^2) to O(T); however, KV-cache size grows linearly with the number of generated tokens, and it rapidly becomes the dominant bottleneck limiting inference efficiency and generative length. Unified autoregressive video generation inherits this limitation. Our analysis reveals that KV-cache tokens exhibit distinct spatiotemporal properties: (i) text and conditioning-image tokens act as persistent semantic anchors that consistently receive high attention, and (ii) attention to previous frames naturally decays with temporal distance. Leveraging these observations, we introduce PackCache, a training-free KV-cache management method that dynamically compacts the KV cache through three coordinated mechanisms: condition anchoring that preserves semantic references, cross-frame decay modeling that allocates cache budget according to temporal distance, and spatially preserving position embedding that maintains coherent 3D structure under cache removal. In terms of efficiency, PackCache accelerates end-to-end generation by 1.7-2.2x on 48-frame long sequences, showcasing its strong potential for enabling longer-sequence video generation. Notably, the final four frames - the portion most impacted by the progressively expanding KV-cache and thus the most expensive segment of the clip - PackCache delivers a 2.6x and 3.7x acceleration on A40 and H200, respectively, for 48-frame videos.", "authors": ["Kunyang Li", "Mubarak Shah", "Yuzhang Shang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-07", "url": "https://arxiv.org/abs/2601.04359", "pdf_url": "https://arxiv.org/pdf/2601.04359v1", "arxiv_id": "2601.04359", "doi": "10.48550/arXiv.2601.04359", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4354}
{"id": "5862837f2413b6b23e203d7b1e27d33df72c9a3aa77a388f95e8e54fcd4e8e53", "sources": ["arxiv", "semantic_scholar"], "title": "Joint Encoding of KV-Cache Blocks for Scalable LLM Serving", "abstract": "Modern large language models (LLMs) drive interactive AI systems but are bottlenecked by the memory-heavy growth of key-value (KV) caches, which limits real-time throughput under concurrent loads. Existing KV-cache compression methods rely on rigid heuristics, disrupt tensor layouts, or require specialized compute, hindering scalability and deployment. We propose joint encoding of KV-cache blocks, which fuses similar blocks across requests and input chunks into shared representations while preserving standard cache structure. This alleviates the KV-cache memory bottleneck, supporting high-concurrency serving without specialized hardware. Theoretically, we analyze the rate-distortion tradeoff of fused cache blocks under a Poisson process model. Empirically, our method achieves up to 4.38 $\\times$ KV-cache compression with negligible accuracy loss across diverse LLMs and benchmarks, outperforming recent structured and adaptive compression baselines. In real LLM serving, joint encoding improves the token throughput by $\\sim$40\\% on a single-machine vLLM benchmark, demonstrating substantial gains in inference throughput. Code is available at https://github.com/sef1/kv_fast_fusion kv_joint_encoding.", "authors": ["Joseph Kampeas", "Emir Haleva"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-06", "url": "https://arxiv.org/abs/2601.03067", "pdf_url": "https://arxiv.org/pdf/2601.03067v1", "arxiv_id": "2601.03067", "doi": "10.48550/arXiv.2601.03067", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/sef1/kv_fast_fusion", "venue": "arXiv.org", "quality_score": 0.6711}
{"id": "35c4f2959d11dbef4a63e53095ad74e88bd0656e7f7c94aff7cd4a3183690a2e", "sources": ["arxiv", "semantic_scholar"], "title": "The Illusion of Specialization: Unveiling the Domain-Invariant \"Standing Committee\" in Mixture-of-Experts Models", "abstract": "Mixture of Experts models are widely assumed to achieve domain specialization through sparse routing. In this work, we question this assumption by introducing COMMITTEEAUDIT, a post hoc framework that analyzes routing behavior at the level of expert groups rather than individual experts. Across three representative models and the MMLU benchmark, we uncover a domain-invariant Standing Committee. This is a compact coalition of routed experts that consistently captures the majority of routing mass across domains, layers, and routing budgets, even when architectures already include shared experts. Qualitative analysis further shows that Standing Committees anchor reasoning structure and syntax, while peripheral experts handle domain-specific knowledge. These findings reveal a strong structural bias toward centralized computation, suggesting that specialization in Mixture of Experts models is far less pervasive than commonly believed. This inherent bias also indicates that current training objectives, such as load-balancing losses that enforce uniform expert utilization, may be working against the model's natural optimization path, thereby limiting training efficiency and performance.", "authors": ["Yan Wang", "Yitao Xu", "Nanhan Shen", "Jinyan Su", "Jimin Huang", "Zining Zhu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-06", "url": "https://arxiv.org/abs/2601.03425", "pdf_url": "https://arxiv.org/pdf/2601.03425v2", "arxiv_id": "2601.03425", "doi": "10.48550/arXiv.2601.03425", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4343}
{"id": "a722e3112d9fae4c862b16de66f53b283e2fce7244bef960ef18de12303bbeee", "sources": ["arxiv", "semantic_scholar"], "title": "Spectral-Efficient LoRa with Low Complexity Detection", "abstract": "In this paper, we propose a spectral-efficient LoRa (SE-LoRa) modulation scheme with a low complexity successive interference cancellation (SIC)-based detector. The proposed communication scheme significantly improves the spectral efficiency of LoRa modulation, while achieving an acceptable error performance compared to conventional LoRa modulation, especially in higher spreading factor (SF) settings. We derive the joint maximum likelihood (ML) detection rule for the SE-LoRa transmission scheme that turns out to be of high computational complexity. To overcome this issue, and by exploiting the frequency-domain characteristics of the dechirped SE-LoRa signal, we propose a low complexity SIC-based detector with a computation complexity at the order of conventional LoRa detection. By computer simulations, we show that the proposed SE-LoRa with low complexity SIC-based detector can improve the spectral efficiency of LoRa modulation up to $445.45\\%$, $1011.11\\%$, and $1071.88\\%$ for SF values of $7$, $9$, and $11$, respectively, while maintaining the error performance within less than $3$ dB of conventional LoRa at symbol error rate (SER) of $10^{-3}$ in Rician channel conditions.", "authors": ["Alireza Maleki", "Ebrahim Bedeer", "Robert Barton"], "categories": ["eess.SP"], "fields_of_study": ["Engineering"], "published_date": "2026-01-06", "url": "https://arxiv.org/abs/2601.03148", "pdf_url": "https://arxiv.org/pdf/2601.03148v1", "arxiv_id": "2601.03148", "doi": "10.1109/jiot.2026.3685102", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Internet of Things Journal", "quality_score": 0.4343}
{"id": "0793cfd42f30de28a24dc2d064667b87e7c9c4743adc988cda85de9f8339359b", "sources": ["arxiv", "semantic_scholar"], "title": "Crystal-KV: Efficient KV Cache Management for Chain-of-Thought LLMs via Answer-First Principle", "abstract": "Chain-of-Thought (CoT) reasoning in large language models (LLMs) significantly improves accuracy on complex tasks, yet incurs excessive memory overhead due to the long think-stage sequences stored in the Key-Value (KV) cache. Unlike traditional generation tasks where all tokens are uniformly important, CoT emphasizes the final answer, rendering conventional KV compression strategies ineffective. In this paper, we present Crystal-KV, an efficient KV cache management framework tailored for CoT reasoning. Our key insight is the answer-first principle. By mapping answer preferences into think-stage attention map, we distinguish between SlipKV, which mainly maintains the reasoning flow but may occasionally introduce misleading context, and CrystalKV, which truly contributes to the correctness of the final answer. Next, we propose an attention-based Least Recently Frequently Used algorithm. It precisely identifies when a SlipKV entry's utility expires and evicts it, retaining CrystalKV without disrupting reasoning flow. Finally, we introduce an adaptive cache budget allocation algorithm. Based on the dynamic proportion of CrystalKV, it estimates the importance of each layer/head and adjusts the KV cache budget during inference, amplifying critical components to improve budget utilization. Results show that Crystal-KV achieves state-of-the-art KV cache compression, significantly improves throughput, and enables faster response time, while maintaining, or even improving, answer accuracy for CoT reasoning.", "authors": ["Zihan Wang", "Cheng Tang", "Lei Gong", "Cheng Li", "Chao Wang", "teng wang", "Wenqi Lou", "Xuehai Zhou"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-05", "url": "https://arxiv.org/abs/2601.16986", "pdf_url": "https://arxiv.org/pdf/2601.16986v1", "arxiv_id": "2601.16986", "doi": "10.48550/arXiv.2601.16986", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4331}
{"id": "d46349153a3709ab81f615bc8c6471466fdbf2a9fdd4931a6860ab4049b92ff2", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-Drop: Temporal LoRA Decoding for Efficient LLM Inference", "abstract": "Autoregressive large language models (LLMs) are bottlenecked by sequential decoding, where each new token typically requires executing all transformer layers. Existing dynamic-depth and layer-skipping methods reduce this cost, but often rely on auxiliary routing mechanisms or incur accuracy degradation when bypassed layers are left uncompensated. We present \\textbf{LoRA-Drop}, a plug-and-play inference framework that accelerates decoding by applying a \\emph{temporal compute schedule} to a fixed subset of intermediate layers: on most decoding steps, selected layers reuse the previous-token hidden state and apply a low-rank LoRA correction, while periodic \\emph{refresh} steps execute the full model to prevent drift. LoRA-Drop requires no routing network, is compatible with standard KV caching, and can reduce KV-cache footprint by skipping KV updates in droppable layers during LoRA steps and refreshing periodically. Across \\textbf{LLaMA2-7B}, \\textbf{LLaMA3-8B}, \\textbf{Qwen2.5-7B}, and \\textbf{Qwen2.5-14B}, LoRA-Drop achieves up to \\textbf{2.6$\\times$ faster decoding} and \\textbf{45--55\\% KV-cache reduction} while staying within \\textbf{0.5 percentage points (pp)} of baseline accuracy. Evaluations on reasoning (GSM8K, MATH, BBH), code generation (HumanEval, MBPP), and long-context/multilingual benchmarks (LongBench, XNLI, XCOPA) identify a consistent \\emph{safe zone} of scheduling configurations that preserves quality while delivering substantial efficiency gains, providing a simple path toward adaptive-capacity inference in LLMs. Codes are available at https://github.com/hosseinbv/LoRA-Drop.git.", "authors": ["Hossein Rajabzadeh", "Maryam Dialameh", "Chul B. Park", "Il-Min Kim", "Hyock Ju Kwon"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-05", "url": "https://arxiv.org/abs/2601.02569", "pdf_url": "https://arxiv.org/pdf/2601.02569v1", "arxiv_id": "2601.02569", "doi": "10.48550/arXiv.2601.02569", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/hosseinbv/LoRA-Drop.git", "venue": "arXiv.org", "quality_score": 0.6694}
{"id": "4f97bbb9cf53ca2374dbe1a6ba2889c93f0d7ce4dcc74752ad96c83d6c62afb1", "sources": ["arxiv", "semantic_scholar"], "title": "Varying-Coefficient Mixture of Experts Model", "abstract": "Mixture-of-Experts (MoE) is a flexible framework that combines multiple specialized submodels (``experts''), by assigning covariate-dependent weights (``gating functions'') to each expert, and have been commonly used for analyzing heterogeneous data. Existing statistical MoE formulations typically assume constant coefficients, for covariate effects within the expert or gating models, which can be inadequate for longitudinal, spatial, or other dynamic settings where covariate influences and latent subpopulation structure evolve across a known dimension. We propose a Varying-Coefficient Mixture of Experts (VCMoE) model that allows all coefficient effects in both the gating functions and expert models to vary along an indexing variable. We establish identifiability and consistency of the proposed model, and develop an estimation procedure, label-consistent EM algorithm, for both fully functional and hybrid specifications, along with the corresponding asymptotic distributions of the resulting estimators. For inference, simultaneous confidence bands are constructed using both asymptotic theory for the maximum discrepancy between the estimated functional coefficients and their true counterparts, and with bootstrap methods. In addition, a generalized likelihood ratio test is developed to examine whether a coefficient function is genuinely varying across the index variable. Simulation studies demonstrate good finite-sample performance, with acceptable bias and satisfactory coverage rates. We illustrate the proposed VCMoE model using a dataset of single nucleus gene expression in embryonic mice to characterize the temporal dynamics of the associations between the expression levels of genes Satb2 and Bcl11b across two latent cell subpopulations of neurons, yielding results that are consistent with prior findings.", "authors": ["Qicheng Zhao", "Celia M. T. Greenwood", "Qihuang Zhang"], "categories": ["stat.ME", "stat.ML"], "fields_of_study": ["Mathematics"], "published_date": "2026-01-05", "url": "https://arxiv.org/abs/2601.01699", "pdf_url": "https://arxiv.org/pdf/2601.01699v1", "arxiv_id": "2601.01699", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2756}
{"id": "0b29470a5ad9b1dc80b3cba48a59ac99bdc339dda8decc99c69186b3354e36da", "sources": ["arxiv", "semantic_scholar"], "title": "OrbitFlow: SLO-Aware Long-Context LLM Serving with Fine-Grained KV Cache Reconfiguration", "abstract": "Serving long-context LLMs is challenging because request lengths and batch composition vary during token generation, causing the memory footprint to fluctuate significantly at runtime. Offloading KV caches to host memory limits effective memory usage, but existing static and predetermined offloading strategies cannot adapt to the rapidly shifting memory demands of long-context serving. This often leads to excessive CPU-to-GPU KV transfers that translate into latency spikes and frequent SLO violations. To address these challenges, we introduce OrbitFlow, a fine-grained and adaptive KV cache management system that meets latency SLOs in long-context LLM serving. OrbitFlow employs a lightweight ILP solver to decide which layers' KV caches to retain on the GPU for each request, within memory capacity constraints. It continuously refines KV placements based on runtime feedback when the active plan becomes suboptimal during token generation. Under heavy load, OrbitFlow invokes a fallback mechanism to temporarily defer in-flight requests with large memory footprints, preserving overall SLO attainment. Our experiments demonstrate that OrbitFlow improves SLO attainment for TPOT and TBT by up to 66% and 48%, respectively, while reducing the 95th percentile latency by 38% and achieving up to 3.3x higher throughput compared to existing offloading methods.", "authors": ["Xinyue Ma", "Heelim Hong", "Taegeon Um", "Jongseop Lee", "Seoyeong Choy", "Woo-Yeon Lee", "Myeongjae Jeon"], "categories": ["cs.AI", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-05", "url": "https://arxiv.org/abs/2601.10729", "pdf_url": "https://arxiv.org/pdf/2601.10729v2", "arxiv_id": "2601.10729", "doi": "10.48550/arXiv.2601.10729", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Proceedings of the VLDB Endowment", "quality_score": 0.4331}
{"id": "e741de7adc9cf1ffd1ce278f7d31e6d9107befa4ac8dd31a0352a5162858de1f", "sources": ["arxiv", "semantic_scholar"], "title": "Routing by Analogy: kNN-Augmented Expert Assignment for Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) architectures scale large language models efficiently by employing a parametric ``router'' to dispatch tokens to a sparse subset of experts. Typically, this router is trained once and then frozen, rendering routing decisions brittle under distribution shifts. We address this limitation by introducing kNN-MoE, a retrieval-augmented routing framework that reuses locally optimal expert assignments from a memory of similar past cases. This memory is constructed offline by directly optimizing token-wise routing logits to maximize the likelihood on a reference set. Crucially, we use the average similarity of retrieved neighbors as a confidence-driven mixing coefficient, thus allowing the method to fall back to the frozen router when no relevant cases are found. Experiments show that kNN-MoE outperforms the zero-shot baseline and is competitive with computationally intensive supervised fine-tuning.", "authors": ["Boxuan Lyu", "Soichiro Murakami", "Hidetaka Kamigaito", "Peinan Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-05", "url": "https://arxiv.org/abs/2601.02144", "pdf_url": "https://arxiv.org/pdf/2601.02144v2", "arxiv_id": "2601.02144", "doi": "10.48550/arXiv.2601.02144", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4331}
{"id": "2277b39c6bcb8b417c07d4bc6e587225b78a5d02808b651fdd7246385dbf63de", "sources": ["arxiv", "semantic_scholar"], "title": "XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression", "abstract": "Learning-based 3D visual geometry models have benefited substantially from large-scale transformers. Among these, StreamVGGT leverages frame-wise causal attention for strong streaming reconstruction, but suffers from unbounded KV cache growth, leading to escalating memory consumption and inference latency as input frames accumulate. We propose XStreamVGGT, a tuning-free approach that systematically compresses the KV cache through joint pruning and quantization, enabling extremely memory-efficient streaming inference. Specifically, redundant KVs originating from multi-view inputs are pruned through efficient token importance identification, enabling a fixed memory budget. Leveraging the unique distribution of KV tensors, we incorporate KV quantization to further reduce memory consumption. Extensive evaluations show that XStreamVGGT achieves mostly negligible performance degradation while substantially reducing memory usage by 4.42$\\times$ and accelerating inference by 5.48$\\times$, enabling scalable and practical streaming 3D applications. The code is available at https://github.com/ywh187/XStreamVGGT/.", "authors": ["Zunhai Su", "Weihao Ye", "Hansen Feng", "Keyu Fan", "Jing Zhang", "Dahai Yu", "Zhengwu Liu", "Ngai Wong"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-03", "url": "https://arxiv.org/abs/2601.01204", "pdf_url": "https://arxiv.org/pdf/2601.01204v1", "arxiv_id": "2601.01204", "doi": "10.48550/arXiv.2601.01204", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ywh187/XStreamVGGT/", "venue": "Journal of the Society for Information Display", "quality_score": 0.6658}
{"id": "1230e5d4bb15257effde3798385f7823755c1c82a72fa9012a509716055fcb86", "sources": ["arxiv", "semantic_scholar"], "title": "HFedMoE: Resource-aware Heterogeneous Federated Learning with Mixture-of-Experts", "abstract": "While federated learning (FL) enables fine-tuning of large language models (LLMs) without compromising data privacy, the substantial size of an LLM renders on-device training impractical for resource-constrained clients, such as mobile devices. Thus, Mixture-of-Experts (MoE) models have emerged as a computation-efficient solution, which activates only a sparse subset of experts during model training to reduce computing burden without sacrificing performance. Though integrating MoE into FL fine-tuning holds significant potential, it still encounters three key challenges: i) selecting appropriate experts for clients remains challenging due to the lack of a reliable metric to measure each expert's impact on local fine-tuning performance, ii) the heterogeneous computing resources across clients severely hinder MoE-based LLM fine-tuning, as dynamic expert activations across diverse input samples can overwhelm resource-constrained devices, and iii) client-specific expert subsets and routing preference undermine global aggregation, where misaligned expert updates and inconsistent gating networks in troduce destructive interference. To address these challenges, we propose HFedMoE, a heterogeneous MoE-based FL fine-tuning framework that customizes a subset of experts to each client for computation-efficient LLM fine-tuning. Specifically, HFedMoE identifies the expert importance based on its contributions to fine-tuning performance, and then adaptively selects a subset of experts from an information bottleneck perspective to align with each client' s computing budget. A sparsity-aware model aggregation strategy is also designed to aggregate the actively fine-tuned experts and gating parameters with importance weighted contributions. Extensive experiments demonstrate that HFedMoE outperforms state-of-the-art benchmarks in training accuracy and convergence speed.", "authors": ["Zihan Fang", "Zheng Lin", "Senkang Hu", "Yanan Ma", "Yihang Tao", "Yiqin Deng", "Xianhao Chen", "Yuguang Fang"], "categories": ["cs.LG", "cs.AI", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-02", "url": "https://arxiv.org/abs/2601.00583", "pdf_url": "https://arxiv.org/pdf/2601.00583v1", "arxiv_id": "2601.00583", "doi": "10.48550/arXiv.2601.00583", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4297}
{"id": "bd2b0b4255ac211f5f250e4aaed8820e84eeddca7ee722c756a7a080bf42288c", "sources": ["arxiv", "semantic_scholar"], "title": "PackKV: Reducing KV Cache Memory Footprint through LLM-Aware Lossy Compression", "abstract": "Transformer-based large language models (LLMs) have demonstrated remarkable potential across a wide range of practical applications. However, long-context inference remains a significant challenge due to the substantial memory requirements of the key-value (KV) cache, which can scale to several gigabytes as sequence length and batch size increase. In this paper, we present \\textbf{PackKV}, a generic and efficient KV cache management framework optimized for long-context generation. %, which synergistically supports both latency-critical and throughput-critical inference scenarios. PackKV introduces novel lossy compression techniques specifically tailored to the characteristics of KV cache data, featuring a careful co-design of compression algorithms and system architecture. Our approach is compatible with the dynamically growing nature of the KV cache while preserving high computational efficiency. Experimental results show that, under the same and minimum accuracy drop as state-of-the-art quantization methods, PackKV achieves, on average, \\textbf{153.2}\\% higher memory reduction rate for the K cache and \\textbf{179.6}\\% for the V cache. Furthermore, PackKV delivers extremely high execution throughput, effectively eliminating decompression overhead and accelerating the matrix-vector multiplication operation. Specifically, PackKV achieves an average throughput improvement of \\textbf{75.7}\\% for K and \\textbf{171.7}\\% for V across A100 and RTX Pro 6000 GPUs, compared to cuBLAS matrix-vector multiplication kernels, while demanding less GPU memory bandwidth. Code available on https://github.com/BoJiang03/PackKV", "authors": ["Bo Jiang", "Taolue Yang", "Youyuan Liu", "Xubin He", "Sheng Di", "Sian Jin"], "categories": ["cs.DC", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-30", "url": "https://arxiv.org/abs/2512.24449", "pdf_url": "https://arxiv.org/pdf/2512.24449v2", "arxiv_id": "2512.24449", "doi": "10.48550/arXiv.2512.24449", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/BoJiang03/PackKV", "venue": "arXiv.org", "quality_score": 0.6587}
{"id": "b85c5d2f0f8fc1f534bca81e19f5cad67683765ce4041839b42eff178454285d", "sources": ["arxiv", "semantic_scholar"], "title": "Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary Loss", "abstract": "Mixture-of-Experts (MoE) models lack explicit constraints to ensure the router's decisions align well with the experts' capabilities, which ultimately limits model performance. To address this, we propose expert-router coupling (ERC) loss, a lightweight auxiliary loss that tightly couples the router's decisions with expert capabilities. Our approach treats each expert's router embedding as a proxy token for the tokens assigned to that expert, and feeds perturbed router embeddings through the experts to obtain intermediate activations. The ERC loss enforces two constraints on these activations: (1) Each expert must exhibit higher activation for its own proxy token than for the proxy tokens of any other expert. (2) Each proxy token must elicit stronger activation from its corresponding expert than from any other expert. These constraints jointly ensure that each router embedding faithfully represents its corresponding expert's capability, while each expert specializes in processing the tokens actually routed to it. The ERC loss is computationally efficient, operating only on $n^2$ activations, where $n$ is the number of experts. This represents a fixed cost independent of batch size, unlike prior coupling methods that scale with the number of tokens (often millions per batch). Through pre-training MoE-LLMs ranging from 3B to 15B parameters and extensive analysis on trillions of tokens, we demonstrate the effectiveness of the ERC loss. Moreover, the ERC loss offers flexible control and quantitative tracking of expert specialization levels during training, providing valuable insights into MoEs.", "authors": ["Ang Lv", "Jin Ma", "Yiyuan Ma", "Siyuan Qiao"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-29", "url": "https://arxiv.org/abs/2512.23447", "pdf_url": "https://arxiv.org/pdf/2512.23447v2", "arxiv_id": "2512.23447", "doi": "10.48550/arXiv.2512.23447", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4251}
{"id": "164a2755673998011dab2f78c04966400a8cf2ff8c39459c1db176b51f8d4491", "sources": ["arxiv", "semantic_scholar"], "title": "Fragile Knowledge, Robust Instruction-Following: The Width Pruning Dichotomy in Llama-3.2", "abstract": "Structured width pruning of GLU-MLP layers, guided by the Maximum Absolute Weight (MAW) criterion, reveals a systematic dichotomy in how reducing the expansion ratio affects different model capabilities. While performance on tasks relying on parametric knowledge (e.g., MMLU, GSM8K) and perplexity metrics degrades predictably, instruction-following capabilities improve substantially (+46% to +75% in IFEval for Llama-3.2-1B and 3B models), and multi-step reasoning remains robust (MUSR). This pattern challenges the prevailing assumption that pruning induces uniform degradation. We evaluated seven expansion ratio configurations using comprehensive benchmarks assessing factual knowledge, mathematical reasoning, language comprehension, instruction-following, and truthfulness. Our analysis identifies the expansion ratio as a critical architectural parameter that selectively modulates cognitive capabilities, rather than merely serving as a compression metric. We provide the first systematic characterization of this selective preservation phenomenon. Notably, we document a robust inverse correlation (r = -0.864, p = 0.012 in Llama-3B) between factual knowledge capacity (MMLU) and truthfulness metrics (TruthfulQA-MC2): as knowledge degrades, the model's ability to discriminate misconceptions improves consistently. This connects two previously distinct research areas, demonstrating that MAW-guided width pruning acts as a selective filter, reducing parametric knowledge while preserving or enhancing behavioral alignment. Additionally, we quantify context-dependent efficiency trade-offs: pruned configurations achieve up to 23% reduction in energy consumption (J/token) but incur penalties in single-request latency, whereas batch processing workloads benefit uniformly.", "authors": ["Pere Martra"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-27", "url": "https://arxiv.org/abs/2512.22671", "pdf_url": "https://arxiv.org/pdf/2512.22671v2", "arxiv_id": "2512.22671", "doi": "10.48550/arXiv.2512.22671", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/peremartra/llama-glu-expansion-pruning", "venue": "arXiv.org", "quality_score": 0.6534}
{"id": "681f6e27cce4dd91a5bfb650b514be1fd47721bc456bbaef958220b6fc457ec6", "sources": ["arxiv", "semantic_scholar"], "title": "InstructMoLE: Instruction-Guided Mixture of Low-rank Experts for Multi-Conditional Image Generation", "abstract": "Parameter-Efficient Fine-Tuning of Diffusion Transformers (DiTs) for diverse, multi-conditional tasks often suffers from task interference when using monolithic adapters like LoRA. The Mixture of Low-rank Experts (MoLE) architecture offers a modular solution, but its potential is usually limited by routing policies that operate at a token level. Such local routing can conflict with the global nature of user instructions, leading to artifacts like spatial fragmentation and semantic drift in complex image generation tasks. To address these limitations, we introduce InstructMoLE, a novel framework that employs an Instruction-Guided Mixture of Low-Rank Experts. Instead of per-token routing, InstructMoLE utilizes a global routing signal, Instruction-Guided Routing (IGR), derived from the user's comprehensive instruction. This ensures that a single, coherently chosen expert council is applied uniformly across all input tokens, preserving the global semantics and structural integrity of the generation process. To complement this, we introduce an output-space orthogonality loss, which promotes expert functional diversity and mitigates representational collapse. Extensive experiments demonstrate that InstructMoLE significantly outperforms existing LoRA adapters and MoLE variants across challenging multi-conditional generation benchmarks. Our work presents a robust and generalizable framework for instruction-driven fine-tuning of generative models, enabling superior compositional control and fidelity to user intent.", "authors": ["Jinqi Xiao", "Qing Yan", "Liming Jiang", "Zichuan Liu", "Hao Kang", "Shen Sang", "Tiancheng Zhi", "Jing Liu", "Cheng Yang", "Xin Lu", "Bo Yuan"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-25", "url": "https://arxiv.org/abs/2512.21788", "pdf_url": "https://arxiv.org/pdf/2512.21788v3", "arxiv_id": "2512.21788", "doi": "10.48550/arXiv.2512.21788", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4205}
{"id": "455933274fbc47b4a61e0f08ef3bdd410c6ccb6b6e3ef38351100f8e0ebe7c78", "sources": ["arxiv", "semantic_scholar"], "title": "Spatiotemporal-Untrammelled Mixture of Experts for Multi-Person Motion Prediction", "abstract": "Comprehensively and flexibly capturing the complex spatio-temporal dependencies of human motion is critical for multi-person motion prediction. Existing methods grapple with two primary limitations: i) Inflexible spatiotemporal representation due to reliance on positional encodings for capturing spatiotemporal information. ii) High computational costs stemming from the quadratic time complexity of conventional attention mechanisms. To overcome these limitations, we propose the Spatiotemporal-Untrammelled Mixture of Experts (ST-MoE), which flexibly explores complex spatio-temporal dependencies in human motion and significantly reduces computational cost. To adaptively mine complex spatio-temporal patterns from human motion, our model incorporates four distinct types of spatiotemporal experts, each specializing in capturing different spatial or temporal dependencies. To reduce the potential computational overhead while integrating multiple experts, we introduce bidirectional spatiotemporal Mamba as experts, each sharing bidirectional temporal and spatial Mamba in distinct combinations to achieve model efficiency and parameter economy. Extensive experiments on four multi-person benchmark datasets demonstrate that our approach not only outperforms state-of-art in accuracy but also reduces model parameter by 41.38% and achieves a 3.6x speedup in training. The code is available at https://github.com/alanyz106/ST-MoE.", "authors": ["Zheng Yin", "Chengjian Li", "Xiangbo Shu", "Meiqi Cao", "Rui Yan", "Jinhui Tang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-25", "url": "https://arxiv.org/abs/2512.21707", "pdf_url": "https://arxiv.org/pdf/2512.21707v1", "arxiv_id": "2512.21707", "doi": "10.48550/arXiv.2512.21707", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/alanyz106/ST-MoE", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.6499}
{"id": "dfcd97d4cfd46aa071450c96d6d51cc6fab58b715790856086319ab493078c03", "sources": ["arxiv", "semantic_scholar"], "title": "Hybrid Quantum-Classical Mixture of Experts: Unlocking Topological Advantage via Interference-Based Routing", "abstract": "The Mixture-of-Experts (MoE) architecture has emerged as a powerful paradigm for scaling deep learning models, yet it is fundamentally limited by challenges such as expert imbalance and the computational complexity of classical routing mechanisms. This paper investigates the potential of Quantum Machine Learning (QML) to address these limitations through a novel Hybrid Quantum-Classical Mixture of Experts (QMoE) architecture. Specifically, we conduct an ablation study using a Quantum Gating Network (Router) combined with classical experts to isolate the source of quantum advantage. Our central finding validates the Interference Hypothesis: by leveraging quantum feature maps (Angle Embedding) and wave interference, the Quantum Router acts as a high-dimensional kernel method, enabling the modeling of complex, non-linear decision boundaries with superior parameter efficiency compared to its classical counterparts. Experimental results on non-linearly separable data, such as the Two Moons dataset, demonstrate that the Quantum Router achieves a significant topological advantage, effectively \"untangling\" data distributions that linear classical routers fail to separate efficiently. Furthermore, we analyze the architecture's robustness against simulated quantum noise, confirming its feasibility for near-term intermediate-scale quantum (NISQ) hardware. We discuss practical applications in federated learning, privacy-preserving machine learning, and adaptive systems that could benefit from this quantum-enhanced routing paradigm.", "authors": ["Reda Heddad", "Lamiae Bouanane"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-25", "url": "https://arxiv.org/abs/2512.22296", "pdf_url": "https://arxiv.org/pdf/2512.22296v1", "arxiv_id": "2512.22296", "doi": "10.48550/arXiv.2512.22296", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/RH2004/QMOE", "venue": "arXiv.org", "quality_score": 0.6499}
{"id": "c6654c2b0914954b53a1d886d831eccbf8d75b06db012c8dc0e60601a4064f2d", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient MoE Inference with Fine-Grained Scheduling of Disaggregated Expert Parallelism", "abstract": "The mixture-of-experts (MoE) architecture scales model size with sublinear computational increase but suffers from memory-intensive inference due to KV caches and sparse expert activation. Recent disaggregated expert parallelism (DEP) distributes attention and experts to dedicated GPU groups but lacks support for shared experts and efficient task scheduling, limiting performance. We propose FinDEP, a fine-grained task scheduling algorithm for DEP that maximizes task overlap to improve MoE inference throughput. FinDEP introduces three innovations: 1) partitioning computation/communication into smaller tasks for fine-grained pipelining, 2) formulating a scheduling optimization supporting variable granularity and ordering, and 3) developing an efficient solver for this large search space. Experiments on four GPU systems with DeepSeek-V2 and Qwen3-MoE show FinDEP improves throughput by up to 1.61x over prior methods, achieving up to 1.24x speedup on a 32-GPU system.", "authors": ["Xinglin Pan", "Shaohuai Shi", "Wenxiang Lin", "Yuxin Wang", "Zhenheng Tang", "Wei Wang", "Xiaowen Chu"], "categories": ["cs.DC", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-25", "url": "https://arxiv.org/abs/2512.21487", "pdf_url": "https://arxiv.org/pdf/2512.21487v1", "arxiv_id": "2512.21487", "doi": "10.48550/arXiv.2512.21487", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4205}
{"id": "1bdd68c96cedd0aa33b448b92e8b679c9a5e39a391336f20bcd5245831d37250", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Experts with Gradient Conflict-Driven Subspace Topology Pruning for Emergent Modularity", "abstract": "Mixture-of-Experts (MoE) architectures achieve parameter efficiency through conditional computation, yet contemporary designs suffer from two fundamental limitations: structural parameter isolation that causes catastrophic forgetting, and instruction-overfitting that degrades performance in instruction-free scenarios. We propose CDSP-MoE (Conflict-Driven Subspace Pruning MoE), a framework that addresses these issues through a paradigm shift from isolated expert containers to dynamic expert instantiation within a shared physical subspace. Grounded in the Universal Weight Subspace Hypothesis, CDSP-MoE maintains a super-complete parameter backbone where logical experts are carved out via learnable topology masks. Unlike prior work that uses gradient conflict for token reassignment or optimization surgery, we leverage it as a structural supervisory signal: a Lagged Gradient Game penalizes interfering connections in the shared manifold, enabling the topology to spontaneously prune conflicting pathways and evolve interpretable modular structures. Experimental results demonstrate that CDSP-MoE achieves robust content-driven routing without human-defined task labels, maintaining semantic specialization even under strict blind inference protocols where explicit instructions are absent. Code is available at: https://github.com/konodiodaaaaa1/Conflict-Driven-Subspace-Pruning-Mixture-of-Experts", "authors": ["Yuxing Gan", "Ziyu Lei"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-23", "url": "https://arxiv.org/abs/2512.20291", "pdf_url": "https://arxiv.org/pdf/2512.20291v4", "arxiv_id": "2512.20291", "doi": "10.48550/arXiv.2512.20291", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/konodiodaaaaa1/Conflict-Driven-Subspace-Pruning-Mixture-of-Experts", "venue": "arXiv.org", "quality_score": 0.6464}
{"id": "3023f80e009287d851b250be96d9c16ae2ba218b8fc62400bdd96d54591ab238", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-DiffuSeq: Enhancing Long-Document Diffusion Models with Sparse Attention and Mixture of Experts", "abstract": "We propose \\textbf{MoE-DiffuSeq}, a diffusion-based framework for efficient long-form text generation that integrates sparse attention with a Mixture-of-Experts (MoE) architecture. Existing sequence diffusion models suffer from prohibitive computational and memory costs when scaling to long documents, largely due to dense attention and slow iterative reconstruction. MoE-DiffuSeq addresses these limitations by combining expert routing with a tailored sparse attention mechanism, substantially reducing attention complexity while preserving global coherence and textual fidelity. In addition, we introduce a \\emph{soft absorbing state} within the diffusion process that reshapes attention dynamics during denoising, enabling faster sequence reconstruction and more precise token refinement. This design accelerates both training and sampling without sacrificing generation quality. Extensive experiments on long-document benchmarks demonstrate that MoE-DiffuSeq consistently outperforms prior diffusion-based and sparse-attention baselines in training efficiency, inference speed, and generation quality. Our approach is particularly effective for long-context applications such as scientific document generation, large-scale code synthesis, and extended dialogue modeling, establishing a scalable and expressive solution for diffusion-based long-form text generation.", "authors": ["Alexandros Christoforos", "Chadbourne Davis"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-23", "url": "https://arxiv.org/abs/2512.20604", "pdf_url": "https://arxiv.org/pdf/2512.20604v2", "arxiv_id": "2512.20604", "doi": "10.48550/arXiv.2512.20604", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4182}
{"id": "406885e16802a43a66f72768b4da27a27d8975e21c773fdc2a613ee1450277ba", "sources": ["arxiv", "semantic_scholar"], "title": "HEART-VIT: Hessian-Guided Efficient Dynamic Attention and Token Pruning in Vision Transformer", "abstract": "Vision Transformers (ViTs) deliver state-of-the-art accuracy but their quadratic attention cost and redundant computations severely hinder deployment on latency and resource-constrained platforms. Existing pruning approaches treat either tokens or heads in isolation, relying on heuristics or first-order signals, which often sacrifice accuracy or fail to generalize across inputs. We introduce HEART-ViT, a Hessian-guided efficient dynamic attention and token pruning framework for vision transformers, which to the best of our knowledge is the first unified, second-order, input-adaptive framework for ViT optimization. HEART-ViT estimates curvature-weighted sensitivities of both tokens and attention heads using efficient Hessian-vector products, enabling principled pruning decisions under explicit loss budgets.This dual-view sensitivity reveals an important structural insight: token pruning dominates computational savings, while head pruning provides fine-grained redundancy removal, and their combination achieves a superior trade-off. On ImageNet-100 and ImageNet-1K with ViT-B/16 and DeiT-B/16, HEART-ViT achieves up to 49.4 percent FLOPs reduction, 36 percent lower latency, and 46 percent higher throughput, while consistently matching or even surpassing baseline accuracy after fine-tuning, for example 4.7 percent recovery at 40 percent token pruning. Beyond theoretical benchmarks, we deploy HEART-ViT on different edge devices such as AGX Orin, demonstrating that our reductions in FLOPs and latency translate directly into real-world gains in inference speed and energy efficiency. HEART-ViT bridges the gap between theory and practice, delivering the first unified, curvature-driven pruning framework that is both accuracy-preserving and edge-efficient.", "authors": ["Mohammad Helal Uddin", "Liam Seymour", "Sabur Baidya"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-23", "url": "https://arxiv.org/abs/2512.20120", "pdf_url": "https://arxiv.org/pdf/2512.20120v1", "arxiv_id": "2512.20120", "doi": "10.48550/arXiv.2512.20120", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4182}
{"id": "21e8a7db09b2d8217753f0b1b75ed44aab1d7cb056fa97ae8c9f0fd58fed991b", "sources": ["arxiv", "semantic_scholar"], "title": "Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning", "abstract": "We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.", "authors": [" NVIDIA", " :", "Aaron Blakeman", "Aaron Grattafiori", "Aarti Basant", "Abhibha Gupta", "Abhinav Khattar", "Adi Renduchintala", "Aditya Vavre", "Akanksha Shukla", "Akhiad Bercovich", "Aleksander Ficek", "Aleksandr Shaposhnikov", "Alex Kondratenko", "Alexander Bukharin", "Alexandre Milesi", "Ali Taghibakhshi", "Alisa Liu", "Amelia Barton", "Ameya Sunil Mahabaleshwarkar", "Amir Klein", "Amit Zuker", "Amnon Geifman", "Amy Shen", "Anahita Bhiwandiwalla", "Andrew Tao", "Ann Guan", "Anubhav Mandarwal", "Arham Mehta", "Ashwath Aithal", "Ashwin Poojary", "Asif Ahamed", "Asma Kuriparambil Thekkumpate", "Ayush Dattagupta", "Banghua Zhu", "Bardiya Sadeghi", "Barnaby Simkin", "Ben Lanir", "Benedikt Schifferer", "Besmira Nushi", "Bilal Kartal", "Bita Darvish Rouhani", "Boris Ginsburg", "Brandon Norick", "Brandon Soubasis", "Branislav Kisacanin", "Brian Yu", "Bryan Catanzaro", "Carlo del Mundo", "Chantal Hwang", "Charles Wang", "Cheng-Ping Hsieh", "Chenghao Zhang", "Chenhan Yu", "Chetan Mungekar", "Chintan Patel", "Chris Alexiuk", "Christopher Parisien", "Collin Neale", "Damon Mosk-Aoyama", "Dan Su", "Dane Corneil", "Daniel Afrimi", "Daniel Rohrer", "Daniel Serebrenik", "Daria Gitman", "Daria Levy", "Darko Stosic", "David Mosallanezhad", "Deepak Narayanan", "Dhruv Nathawani", "Dima Rekesh", "Dina Yared", "Divyanshu Kakwani", "Dong Ahn", "Duncan Riach", "Dusan Stosic", "Edgar Minasyan", "Edward Lin", "Eileen Long", "Eileen Peters Long", "Elena Lantz", "Ellie Evans", "Elliott Ning", "Eric Chung", "Eric Harper", "Eric Tramel", "Erick Galinkin", "Erik Pounds", "Evan Briones", "Evelina Bakhturina", "Faisal Ladhak", "Fay Wang", "Fei Jia", "Felipe Soares", "Feng Chen", "Ferenc Galko", "Frankie Siino", "Gal Hubara Agam", "Ganesh Ajjanagadde", "Gantavya Bhatt", "Gargi Prasad", "George Armstrong", "Gerald Shen", "Gorkem Batmaz", "Grigor Nalbandyan", "Haifeng Qian", "Harsh Sharma", "Hayley Ross", "Helen Ngo", "Herman Sahota", "Hexin Wang", "Himanshu Soni", "Hiren Upadhyay", "Huizi Mao", "Huy C Nguyen", "Huy Q Nguyen", "Iain Cunningham", "Ido Shahaf", "Igor Gitman", "Ilya Loshchilov", "Ivan Moshkov", "Izzy Putterman", "Jan Kautz", "Jane Polak Scowcroft", "Jared Casper", "Jatin Mitra", "Jeffrey Glick", "Jenny Chen", "Jesse Oliver", "Jian Zhang", "Jiaqi Zeng", "Jie Lou", "Jimmy Zhang", "Jining Huang", "Joey Conway", "Joey Guman", "John Kamalu", "Johnny Greco", "Jonathan Cohen", "Joseph Jennings", "Joyjit Daw", "Julien Veron Vialard", "Junkeun Yi", "Jupinder Parmar", "Kai Xu", "Kan Zhu", "Kari Briski", "Katherine Cheung", "Katherine Luna", "Keshav Santhanam", "Kevin Shih", "Kezhi Kong", "Khushi Bhardwaj", "Krishna C. Puvvada", "Krzysztof Pawelec", "Kumar Anik", "Lawrence McAfee", "Laya Sleiman", "Leon Derczynski", "Li Ding", "Lucas Liebenwein", "Luis Vega", "Maanu Grover", "Maarten Van Segbroeck", "Maer Rodrigues de Melo", "Makesh Narsimhan Sreedhar", "Manoj Kilaru", "Maor Ashkenazi", "Marc Romeijn", "Mark Cai", "Markus Kliegl", "Maryam Moosaei", "Matvei Novikov", "Mehrzad Samadi", "Melissa Corpuz", "Mengru Wang", "Meredith Price", "Michael Boone", "Michael Evans", "Miguel Martinez", "Mike Chrzanowski", "Mohammad Shoeybi", "Mostofa Patwary", "Nabin Mulepati", "Natalie Hereth", "Nave Assaf", "Negar Habibi", "Neta Zmora", "Netanel Haber", "Nicola Sessions", "Nidhi Bhatia", "Nikhil Jukar", "Nikki Pope", "Nikolai Ludwig", "Nima Tajbakhsh", "Nirmal Juluru", "Oleksii Hrinchuk", "Oleksii Kuchaiev", "Olivier Delalleau", "Oluwatobi Olabiyi", "Omer Ullman Argov", "Ouye Xie", "Parth Chadha", "Pasha Shamis", "Pavlo Molchanov", "Pawel Morkisz", "Peter Dykas", "Peter Jin", "Pinky Xu", "Piotr Januszewski", "Pranav Prashant Thombre", "Prasoon Varshney", "Pritam Gundecha", "Qing Miao", "Rabeeh Karimi Mahabadi", "Ran El-Yaniv", "Ran Zilberstein", "Rasoul Shafipour", "Rich Harang", "Rick Izzo", "Rima Shahbazyan", "Rishabh Garg", "Ritika Borkar", "Ritu Gala", "Riyad Islam", "Roger Waleffe", "Rohit Watve", "Roi Koren", "Ruoxi Zhang", "Russell J. Hewett", "Ryan Prenger", "Ryan Timbrook", "Sadegh Mahdavi", "Sahil Modi", "Samuel Kriman", "Sanjay Kariyappa", "Sanjeev Satheesh", "Saori Kaji", "Satish Pasumarthi", "Sean Narentharen", "Sean Narenthiran", "Seonmyeong Bak", "Sergey Kashirsky", "Seth Poulos", "Shahar Mor", "Shanmugam Ramasamy", "Shantanu Acharya", "Shaona Ghosh", "Sharath Turuvekere Sreenivas", "Shelby Thomas", "Shiqing Fan", "Shreya Gopal", "Shrimai Prabhumoye", "Shubham Pachori", "Shubham Toshniwal", "Shuoyang Ding", "Siddharth Singh", "Simeng Sun", "Smita Ithape", "Somshubra Majumdar", "Soumye Singhal", "Stefania Alborghetti", "Stephen Ge", "Sugam Dipak Devare", "Sumeet Kumar Barua", "Suseella Panguluri", "Suyog Gupta", "Sweta Priyadarshi", "Syeda Nahida Akter", "Tan Bui", "Teodor-Dumitru Ene", "Terry Kong", "Thanh Do", "Tijmen Blankevoort", "Tom Balough", "Tomer Asida", "Tomer Bar Natan", "Tugrul Konuk", "Twinkle Vashishth", "Udi Karpas", "Ushnish De", "Vahid Noorozi", "Vahid Noroozi", "Venkat Srinivasan", "Venmugil Elango", "Vijay Korthikanti", "Vitaly Kurin", "Vitaly Lavrukhin", "Wanli Jiang", "Wasi Uddin Ahmad", "Wei Du", "Wei Ping", "Wenfei Zhou", "Will Jennings", "William Zhang", "Wojciech Prazuch", "Xiaowei Ren", "Yashaswi Karnati", "Yejin Choi", "Yev Meyer", "Yi-Fu Wu", "Yian Zhang", "Ying Lin", "Yonatan Geifman", "Yonggan Fu", "Yoshi Subara", "Yoshi Suhara", "Yubo Gao", "Zach Moshe", "Zhen Dong", "Zihan Liu", "Zijia Chen", "Zijie Yan"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-23", "url": "https://arxiv.org/abs/2512.20848", "pdf_url": "https://arxiv.org/pdf/2512.20848v1", "arxiv_id": "2512.20848", "doi": "10.48550/arXiv.2512.20848", "citation_count": 44, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4515}
{"id": "437ff87c96562df2eda809fde85534f71252a945888678145de51839e3ac3826", "sources": ["arxiv", "semantic_scholar"], "title": "MixKVQ: Query-Aware Mixed-Precision KV Cache Quantization for Long-Context Reasoning", "abstract": "Long Chain-of-Thought (CoT) reasoning has significantly advanced the capabilities of Large Language Models (LLMs), but this progress is accompanied by substantial memory and latency overhead from the extensive Key-Value (KV) cache. Although KV cache quantization is a promising compression technique, existing low-bit quantization methods often exhibit severe performance degradation on complex reasoning tasks. Fixed-precision quantization struggles to handle outlier channels in the key cache, while current mixed-precision strategies fail to accurately identify components requiring high-precision representation. We find that an effective low-bit KV cache quantization strategy must consider two factors: a key channel's intrinsic quantization difficulty and its relevance to the query. Based on this insight, we propose MixKVQ, a novel plug-and-play method that introduces a lightweight, query-aware algorithm to identify and preserve critical key channels that need higher precision, while applying per-token quantization for value cache. Experiments on complex reasoning datasets demonstrate that our approach significantly outperforms existing low-bit methods, achieving performance comparable to a full-precision baseline at a substantially reduced memory footprint.", "authors": ["Tao Zhang", "Ziqian Zeng", "Hao Peng", "Huiping Zhuang", "Cen Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-22", "url": "https://arxiv.org/abs/2512.19206", "pdf_url": "https://arxiv.org/pdf/2512.19206v1", "arxiv_id": "2512.19206", "doi": "10.48550/arXiv.2512.19206", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4171}
{"id": "440e47857aa842016431c22b65e74a2e71de14e51f00f332f6eb68c2c71c2d28", "sources": ["arxiv", "semantic_scholar"], "title": "Selective LoRA for Visual Tokens and Attention Heads", "abstract": "Low-rank adaptation (LoRA) is widely used for parameter-efficient fine-tuning, but its standard all-token, all-head design ignores the heterogeneous structure of vision language model (VLM) inputs. We introduce \\emph{Image-LoRA}, a vision-oriented PEFT recipe that views LoRA as a token-level residual update and applies this update only to visual tokens. Image-LoRA further restricts adaptation to the value path of a compact subset of attention heads, selected using a one-pass influence estimate from a rank-1 visual-token-only probe. This token-, head-, and value-selective design reduces trainable parameters and adapter-only training FLOPs while leaving the pure-text forward pass of the frozen backbone unchanged when no visual tokens are present. Across visual localization benchmarks with controlled text:image token ratios, Image-LoRA matches or closely approaches standard LoRA, while showing especially favorable trade-offs in image-token-heavy regimes. We further validate its generality on TextVQA and VideoQA, verify pure-text preservation on GSM8K, and show on ViLP that a stronger information bottleneck can yield gains over standard LoRA.", "authors": ["Tiange Luo", "Lajanugen Logeswaran", "Jaekyeom Kim", "Justin Johnson", "Honglak Lee"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-22", "url": "https://arxiv.org/abs/2512.19219", "pdf_url": "https://arxiv.org/pdf/2512.19219v2", "arxiv_id": "2512.19219", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2654}
{"id": "addc616e33cb4da7b01fa2ed6a7f23a02bc2a99d034f878fe7601e7f5da63a83", "sources": ["arxiv", "semantic_scholar"], "title": "SAP: Syntactic Attention Pruning for Transformer-based Language Models", "abstract": "This paper introduces Syntactic Attention Pruning (SAP), a novel method for effectively pruning attention heads in Transformer models. Unlike conventional approaches that rely solely on mathematical analysis of model weights and activations, SAP incorporates both the syntactic structure and attention patterns of sentences to guide the pruning process. By leveraging these linguistic features, SAP not only achieves performance comparable to state-of-the-art methods but also enhances the interpretability of model behavior. To further improve robustness, we propose Candidate Filtering (CF), a mechanism that prioritizes heads based on their contribution to model performance, mitigating degradation during pruning. Experimental results indicate that SAP effectively preserves critical heads of a high density of strong attention values, outperforming existing head pruning strategies in retrain-free settings. These findings position SAP as a promising foundation for a new direction in model compression research, offering high flexibility for pruning across all transformer-based language models.", "authors": ["Tzu-Yun Lee", "Ding-Yong Hong", "Jan-Jan Wu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-22", "url": "https://arxiv.org/abs/2512.19125", "pdf_url": "https://arxiv.org/pdf/2512.19125v1", "arxiv_id": "2512.19125", "doi": "10.48550/arXiv.2512.19125", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4171}
{"id": "16d417c1ee13fdeba3ccf3a93950d9983f92e5e8334805d7bca977060dbd6585", "sources": ["arxiv", "semantic_scholar"], "title": "How Many Experts Are Enough? Towards Optimal Semantic Specialization for Mixture-of-Experts", "abstract": "Finding the optimal configuration of Sparse Mixture-ofExperts (SMoE) that maximizes semantic differentiation among experts is essential for exploiting the full potential of MoE architectures. However, existing SMoE frameworks either heavily rely on hyperparameter tuning or overlook the importance of diversifying semantic roles across experts when adapting the expert pool size. We propose Mixture-of-Experts for Adaptive Semantic Specialization (MASS), a semanticaware MoE framework for adaptive expert expansion and dynamic routing. MASS introduces two key advancements: (i) a gradient-based semantic drift detector that prompts targeted expert expansion when the existing expert pool lacks capacity to capture the full semantic diversity of the data, and (ii) an integration of adaptive routing strategy that dynamically adjusts expert usage based on token-level routing confidence mass. We first demonstrate that MASS reliably converges to the point of optimal balance between cost-performance trade-off with notably improved sematic specialization in a highly controlled synthetic setup. Further empirical results on real-world datasets across language and vision domains show that MASS consistently outperforms a range of strong MoE baselines, demonstrating its domain robustness and enhanced expert specialization.", "authors": ["Sumin Park", "Noseong Park"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-21", "url": "https://arxiv.org/abs/2512.19765", "pdf_url": "https://arxiv.org/pdf/2512.19765v1", "arxiv_id": "2512.19765", "doi": "10.48550/arXiv.2512.19765", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4159}
{"id": "05f5473d33354834f290b8e6827089498eac6c53586f490a7280caaa52710564", "sources": ["arxiv", "semantic_scholar"], "title": "Tempo as the Stable Cue: Hierarchical Mixture of Tempo and Beat Experts for Music to 3D Dance Generation", "abstract": "Music to 3D dance generation aims to synthesize realistic and rhythmically synchronized human dance from music. While existing methods often rely on additional genre labels to further improve dance generation, such labels are typically noisy, coarse, unavailable, or insufficient to capture the diversity of real-world music, which can result in rhythm misalignment or stylistic drift. In contrast, we observe that tempo, a core property reflecting musical rhythm and pace, remains relatively consistent across datasets and genres, typically ranging from 60 to 200 BPM. Based on this finding, we propose TempoMoE, a hierarchical tempo-aware Mixture-of-Experts module that enhances the diffusion model and its rhythm perception. TempoMoE organizes motion experts into tempo-structured groups for different tempo ranges, with multi-scale beat experts capturing fine- and long-range rhythmic dynamics. A Hierarchical Rhythm-Adaptive Routing dynamically selects and fuses experts from music features, enabling flexible, rhythm-aligned generation without manual genre labels. Extensive experiments demonstrate that TempoMoE achieves state-of-the-art results in dance quality and rhythm alignment.", "authors": ["Guangtao Lyu", "Chenghao Xu", "Qi Liu", "Jiexi Yan", "Muli Yang", "Fen Fang", "Cheng Deng"], "categories": ["cs.CV", "cs.MM", "cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-21", "url": "https://arxiv.org/abs/2512.18804", "pdf_url": "https://arxiv.org/pdf/2512.18804v1", "arxiv_id": "2512.18804", "doi": "10.48550/arXiv.2512.18804", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4159}
{"id": "7ddffe35f92b69a66cea873119849100e8a0a6fbedc9385bf1c570b2a1bc9993", "sources": ["arxiv", "semantic_scholar"], "title": "MoE Pathfinder: Trajectory-driven Expert Pruning", "abstract": "Mixture-of-experts (MoE) architectures used in large language models (LLMs) achieve state-of-the-art performance across diverse tasks yet face practical challenges such as deployment complexity and low activation efficiency. Expert pruning has thus emerged as a promising solution to reduce computational overhead and simplify the deployment of MoE models. However, existing expert pruning approaches conventionally rely on local importance metrics and often apply uniform layer-wise pruning, leveraging only partial evaluation signals and overlooking the heterogeneous contributions of experts across layers. To address these limitations, we propose an expert pruning approach based on the trajectory of activated experts across layers, which treats MoE as a weighted computation graph and casts expert selection as a global optimal path planning problem. Within this framework, we integrate complementary importance signals from reconstruction error, routing probabilities, and activation strength at the trajectory level, which naturally yields non-uniform expert retention across layers. Experiments show that our approach achieves superior pruning performance on nearly all tasks compared with most existing approaches.", "authors": ["Xican Yang", "Yuanhe Tian", "Yan Song"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-20", "url": "https://arxiv.org/abs/2512.18425", "pdf_url": "https://arxiv.org/pdf/2512.18425v1", "arxiv_id": "2512.18425", "doi": "10.48550/arXiv.2512.18425", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4148}
{"id": "2aef0e3579e747ff1dfbc5df2e168762f056324342662625dbcb179f5e23b211", "sources": ["arxiv", "semantic_scholar"], "title": "TraCT: Disaggregated LLM Serving with CXL Shared Memory KV Cache at Rack-Scale", "abstract": "Disaggregated LLM serving improves resource efficiency by separating the compute-intensive prefill phase from the latency-critical decode phase. However, this architecture introduces a fundamental bottleneck: key/value (KV) tensors generated during prefill must be transferred to decode workers, and existing systems rely on RDMA-based network paths for this exchange. As model sizes and context lengths increase, KV transfer dominates both time-to-first-token (TTFT) and peak throughput, and remains highly sensitive to network contention even when prefix reuse is high. This paper presents TraCT, a rack-scale LLM serving system that uses CXL shared memory as both a KV-transfer substrate and a rack-wide prefix-aware KV cache. TraCT enables GPUs to write and read KV blocks directly through CXL load/store and DMA operations, eliminating the NIC hop that constrains existing disaggregated pipelines. However, to realize this design, multiple new challenges such as synchronization, consistency, and data management on non-coherent CXL memory need to be addressed. TraCT proposes various software solutions such as the two-tier inter-node synchronization mechanism to address these challenges. We implement TraCT on the Dynamo LLM inference framework and show that, across static and synthetic workloads, TraCT reduces average TTFT by up to 9.8x, lowers P99 latency by up to 6.2x, and improves peak throughput by up to 1.6x compared to RDMA and DRAM-based caching baselines.", "authors": ["Dongha Yoon", "Younghoon Min", "Hoshik Kim", "Sam H. Noh", "Jongryool Kim"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-20", "url": "https://arxiv.org/abs/2512.18194", "pdf_url": "https://arxiv.org/pdf/2512.18194v1", "arxiv_id": "2512.18194", "doi": "10.48550/arXiv.2512.18194", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4148}
{"id": "09cdadce27619dd6ee8ccadbe388198877ef454c248dcfb03fc4b5c390151476", "sources": ["arxiv", "semantic_scholar"], "title": "Secret mixtures of experts inside your LLM", "abstract": "Despite being one of the earliest neural network layers, the Multilayer Perceptron (MLP) is arguably one of the least understood parts of the transformer architecture due to its dense computation and lack of easy visualization. This paper seeks to understand the MLP layers in dense LLM models by hypothesizing that these layers secretly approximately perform a sparse computation -- namely, that they can be well approximated by sparsely-activating Mixture of Experts (MoE) layers. Our hypothesis is based on a novel theoretical connection between MoE models and Sparse Autoencoder (SAE) structure in activation space. We empirically validate the hypothesis on pretrained LLMs, and demonstrate that the activation distribution matters -- these results do not hold for Gaussian data, but rather rely crucially on structure in the distribution of neural network activations. Our results shine light on a general principle at play in MLP layers inside LLMs, and give an explanation for the effectiveness of modern MoE-based transformers. Additionally, our experimental explorations suggest new directions for more efficient MoE architecture design based on low-rank routers.", "authors": ["Enric Boix-Adsera"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-12-20", "url": "https://arxiv.org/abs/2512.18452", "pdf_url": "https://arxiv.org/pdf/2512.18452v1", "arxiv_id": "2512.18452", "doi": "10.48550/arXiv.2512.18452", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4148}
{"id": "4a86d20da85bcea8fdcc0fba16f943a992da3f8eff93cad554e0cf25086d077d", "sources": ["arxiv", "semantic_scholar"], "title": "KV Admission: Learning What to Write for Efficient Long-Context Inference", "abstract": "Long-context LLM inference is bottlenecked by the quadratic attention complexity and linear KV cache growth. Prior approaches mitigate this via post-hoc selection or eviction but overlook the root inefficiency: indiscriminate writing to memory. In this paper, we formalize KV cache management as a causal system of three primitives: KV Admission, Selection, and Eviction. We instantiate KV Admission via Write-Gated KV (WG-KV), a lightweight mechanism that learns to predict token utility before cache entry. By filtering out low-utility states early to maintain a compact global cache alongside a sliding local cache, WG-KV reduces memory usage by 46-68% and delivers 3.03-3.70x prefill and 1.85-2.56x decode speedups on Llama and Qwen models, while maintaining compatibility with FlashAttention and Paged-KV systems. These results demonstrate that learning what to write is a principled and practical recipe for efficient long-context inference. Code is available at https://github.com/EMCLab-Sinica/WG-KV.", "authors": ["Yen-Chieh Huang", "Pi-Cheng Hsiu", "Rui Fang", "Ming-Syan Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-19", "url": "https://arxiv.org/abs/2512.17452", "pdf_url": "https://arxiv.org/pdf/2512.17452v3", "arxiv_id": "2512.17452", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/EMCLab-Sinica/WG-KV", "venue": null, "quality_score": 0.4889}
{"id": "5f981d05e721e3e486ab4da801dafab6d7bf60877cbce7cc10c1f99b16f9471e", "sources": ["arxiv", "semantic_scholar"], "title": "Bandwidth-Efficient Adaptive Mixture-of-Experts via Low-Rank Compensation", "abstract": "Mixture-of-Experts (MoE) models scale capacity via sparse activation but stress memory and bandwidth. Offloading alleviates GPU memory by fetching experts on demand, yet token-level routing causes irregular transfers that make inference I/O-bound. Static uniform quantization reduces traffic but degrades accuracy under aggressive compression by ignoring expert heterogeneity. We present Bandwidth-Efficient Adaptive Mixture-of-Experts via Low-Rank Compensation, which performs router-guided precision restoration using precomputed low-rank compensators. At inference time, our method transfers compact low-rank factors with Top-n (n<k) experts per token and applies compensation to them, keeping others low-bit. Integrated with offloading on GPU and GPU-NDP systems, our method delivers a superior bandwidth-accuracy trade-off and improved throughput.", "authors": ["Zhenyu Liu", "Yunzhen Liu", "Zehao Fan", "Garrett Gagnon", "Yayue Hou", "Nan Wu", "Yangwook Kang", "Liu Liu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-18", "url": "https://arxiv.org/abs/2512.17073", "pdf_url": "https://arxiv.org/pdf/2512.17073v1", "arxiv_id": "2512.17073", "doi": "10.48550/arXiv.2512.17073", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4125}
{"id": "b6e0babc7aa5d1ee70e5fc0c78c58f9ce6212cfdca75d7835064a9a6b44f1599", "sources": ["arxiv", "semantic_scholar"], "title": "Economical Jet Taggers -- Equivariant, Slim, and Quantized", "abstract": "Modern machine learning is transforming jet tagging at the LHC, but the leading transformer architectures are large, not particularly fast, and training-intensive. We present a slim version of the L-GATr tagger, reduce the number of parameters of jet-tagging transformers, and quantize them. We compare different quantization methods for standard and Lorentz-equivariant transformers and estimate their gains in resource efficiency. We find an order-of-magnitude reduction in energy cost for an moderate performance decrease, down to 1000-parameter taggers. This might be a step towards trigger-level jet tagging with small and quantized versions of the leading equivariant transformer architectures.", "authors": ["Antoine Petitjean", "Tilman Plehn", "Jonas Spinner", "Ullrich Köthe"], "categories": ["hep-ph", "hep-ex"], "fields_of_study": ["Physics"], "published_date": "2025-12-18", "url": "https://arxiv.org/abs/2512.17011", "pdf_url": "https://arxiv.org/pdf/2512.17011v2", "arxiv_id": "2512.17011", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2625}
{"id": "45df00274edacfa1dcf91ac1f5c52e11370f70776d790df055bf606cfd5d765a", "sources": ["arxiv", "semantic_scholar"], "title": "CTkvr: KV Cache Retrieval for Long-Context LLMs via Centroid then Token Indexing", "abstract": "Large language models (LLMs) are increasingly applied in long-context scenarios such as multi-turn conversations. However, long contexts pose significant challenges for inference efficiency, including high memory overhead from Key-Value (KV) cache and increased latency due to excessive memory accesses. Recent methods for dynamic KV selection struggle with trade-offs: block-level indexing degrades accuracy by retrieving irrelevant KV entries, while token-level indexing incurs high latency from inefficient retrieval mechanisms. In this paper, we propose CTKVR, a novel centroid-then-token KV retrieval scheme that addresses these limitations. CTKVR leverages a key observation: query vectors adjacent in position exhibit high similarity after Rotary Position Embedding (RoPE) and share most of their top-k KV cache entries. Based on this insight, CTKVR employs a two-stage retrieval strategy: lightweight centroids are precomputed during prefilling for centroid-grained indexing, followed by token-level refinement for precise KV retrieval. This approach balances retrieval efficiency and accuracy. To further enhance performance, we implement an optimized system for indexing construction and search using CPU-GPU co-execution. Experimentally, CTKVR achieves superior performance across multiple benchmarks with less than 1% accuracy degradation. Meanwhile, CTKVR delivers 3 times and 4 times throughput speedups on Llama-3-8B and Yi-9B at 96K context length across diverse GPU hardware.", "authors": ["Kuan Lu", "Shuhang Lin", "Sai Wu", "Yichen Yao", "Junhan Yang", "Huan Li", "Wei Chu", "Xu Yinghui", "Yuan Qi", "Gang Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-17", "url": "https://arxiv.org/abs/2512.15550", "pdf_url": "https://arxiv.org/pdf/2512.15550v1", "arxiv_id": "2512.15550", "doi": "10.48550/arXiv.2512.15550", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4114}
{"id": "7d7881fd1eca95c4c482b2bb026e7aa6e88db6d778839b543359fea168cd5ab5", "sources": ["arxiv", "semantic_scholar"], "title": "EVICPRESS: Joint KV-Cache Compression and Eviction for Efficient LLM Serving", "abstract": "Reusing KV cache is essential for high efficiency of Large Language Model (LLM) inference systems. With more LLM users, the KV cache footprint can easily exceed GPU memory capacity, so prior work has proposed to either evict KV cache to lower-tier storage devices, or compress KV cache so that more KV cache can be fit in the fast memory. However, prior work misses an important opportunity: jointly optimizing the eviction and compression decisions across all KV caches to minimize average generation latency without hurting quality. We propose EVICPRESS, a KV-cache management system that applies lossy compression and adaptive eviction to KV cache across multiple storage tiers. Specifically, for each KV cache of a context, EVICPRESS considers the effect of compression and eviction of the KV cache on the average generation quality and delay across all contexts as a whole. To achieve this, EVICPRESS proposes a unified utility function that quantifies the effect of quality and delay of the lossy compression or eviction. To this end, EVICPRESS's profiling module periodically updates the utility function scores on all possible eviction-compression configurations for all contexts and places KV caches using a fast heuristic to rearrange KV caches on all storage tiers, with the goal of maximizing the utility function scores on each storage tier. Compared to the baselines that evict KV cache or compress KV cache, EVICPRESS achieves higher KV-cache hit rates on fast devices, i.e., lower delay, while preserving high generation quality by applying conservative compression to contexts that are sensitive to compression errors. Evaluation on 12 datasets and 5 models demonstrates that EVICPRESS achieves up to 2.19x faster time-to-first-token (TTFT) at equivalent generation quality.", "authors": ["Shaoting Feng", "Yuhan Liu", "Hanchen Li", "Xiaokun Chen", "Samuel Shen", "Kuntai Du", "Zhuohan Gu", "Rui Zhang", "Yuyang Huang", "Yihua Cheng", "Jiayi Yao", "Qizheng Zhang", "Ganesh Ananthanarayanan", "Junchen Jiang"], "categories": ["cs.OS", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-16", "url": "https://arxiv.org/abs/2512.14946", "pdf_url": "https://arxiv.org/pdf/2512.14946v1", "arxiv_id": "2512.14946", "doi": "10.48550/arXiv.2512.14946", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4102}
{"id": "dcb6f96a5a550202b49502a97d43002743a4621f1de2a542e88e1f553f302c31", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Attention Schemes (MoAS): Learning to Route Between MHA, GQA, and MQA", "abstract": "The choice of attention mechanism in Transformer models involves a critical trade-off between modeling quality and inference efficiency. Multi-Head Attention (MHA) offers the best quality but suffers from large Key-Value (KV) cache memory requirements during inference. Multi-Query Attention (MQA) and Grouped-Query Attention (GQA) reduce memory usage but often at the cost of model performance. In this work, we propose Mixture of Attention Schemes (MoAS), a novel architecture that dynamically selects the optimal attention scheme (MHA, GQA, or MQA) for each token via a learned router. We demonstrate that dynamic routing performs better than static averaging of schemes and achieves performance competitive with the MHA baseline while offering potential for conditional compute efficiency. Experimental results on WikiText-2 show that dynamic routing (val loss 2.3074) outperforms a static mixture (2.3093), validating the effectiveness of the proposed method. Our code is available at https://github.com/Esmail-ibraheem/Mixture-of-Attention-Schemes-MoAS.", "authors": ["Esmail Gumaan"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-16", "url": "https://arxiv.org/abs/2512.20650", "pdf_url": "https://arxiv.org/pdf/2512.20650v1", "arxiv_id": "2512.20650", "doi": "10.48550/arXiv.2512.20650", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Esmail-ibraheem/Mixture-of-Attention-Schemes-MoAS", "venue": "arXiv.org", "quality_score": 0.634}
{"id": "3e9593febf6b879410c1cac2fc07ef20511567171021ed24fae2fb4d2ff93d86", "sources": ["arxiv", "semantic_scholar"], "title": "Kinetic-Mamba: Mamba-Assisted Predictions of Stiff Chemical Kinetics", "abstract": "Accurate chemical kinetics modeling is essential for combustion simulations, as it governs the evolution of complex reaction pathways and thermochemical states. In this work, we introduce Kinetic-Mamba, a Mamba-based neural operator framework that integrates the expressive power of neural operators with the efficient temporal modeling capabilities of Mamba architectures. The framework comprises three complementary models: (i) a standalone Mamba model that predicts the time evolution of thermochemical state variables from given initial conditions; (ii) a constrained Mamba model that enforces mass conservation while learning the state dynamics; and (iii) a regime-informed architecture employing two standalone Mamba models to capture dynamics across temperature-dependent regimes. We additionally develop a latent Kinetic-Mamba variant that evolves dynamics in a reduced latent space and reconstructs the full state on the physical manifold. The accuracy and robustness of Kinetic-Mamba was evaluated using both time-decomposition and recursive-prediction strategies. We further assess the extrapolation capabilities of the model on varied out-of-distribution datasets. Computational experiments on Syngas and GRI-Mech 3.0 reaction mechanisms demonstrate that our framework achieves high fidelity in predicting complex kinetic behavior using only the initial conditions of the state variables.", "authors": ["Additi Pandey", "Liang Wei", "Hessam Babaee", "George Em Karniadakis"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-16", "url": "https://arxiv.org/abs/2512.14471", "pdf_url": "https://arxiv.org/pdf/2512.14471v2", "arxiv_id": "2512.14471", "doi": "10.48550/arXiv.2512.14471", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4102}
{"id": "3c01cfba72f29b4b35903cb2a0a49c04903aaa98d824c34d8846d248b196f7f3", "sources": ["arxiv", "semantic_scholar"], "title": "Janus: Disaggregating Attention and Experts for Scalable MoE Inference", "abstract": "Serving large Mixture-of-Experts (MoE) models is challenging because of their large memory footprints, heterogeneous resource demands, and highly dynamic inference workloads. Most existing MoE inference systems deploy the entire model as a monolithic unit, forcing attention and MoE layers to share the same resource configuration despite their different scaling behaviors and resource bottlenecks. Such coarse-grained provisioning leads to resource inefficiency and suboptimal performance. We present JANUS, a scalable and resource-efficient MoE inference system built around three key principles. First, JANUS disaggregates attention and MoE layers onto separate GPU worker pools, enabling independent resource provisioning for the two layer types, and uses an adaptive two-phase communication mechanism for low-latency data exchange. Second, because MoE-layer execution is often memory-bound and highly sensitive to activated-expert imbalance, JANUS introduces a lightweight, microsecond-scale activation scheduler that balances per-layer activated experts across MoE instances to reduce inference latency. Third, JANUS employs a fine-grained, SLO-aware resource scaling scheme that jointly selects attention resources, MoE resources, and expert placement to minimize GPU cost under token-level SLOs. Evaluation shows that JANUS improves per-GPU throughput by up to 4.7x over state-of-the-art MoE inference baselines while satisfying token-level latency SLOs.", "authors": ["Zhexiang Zhang", "Ye Wang", "Yumiao Zhao", "Jiayu Xiao", "Qianjing Yang", "Xiangyu Wang", "Jingzhe Jiang", "Qizhen Weng", "Ruichuan Chen", "Shaohuai Shi", "Adel N. Toosi", "Yin Chen", "Minchen Yu"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-15", "url": "https://arxiv.org/abs/2512.13525", "pdf_url": "https://arxiv.org/pdf/2512.13525v3", "arxiv_id": "2512.13525", "doi": "10.48550/arXiv.2512.13525", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4091}
{"id": "93d69adb75e04b3db87759a100904e69abdbdf56737f3bc7987273582e4aeba0", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Recursive Transformers with Mixture of LoRAs", "abstract": "Parameter sharing in recursive transformers reduces model size but collapses layer-wise expressivity. We propose Mixture of LoRAs (MoL), a lightweight conditional-computation mechanism that inserts Low-Rank Adaptation (LoRA) experts inside a shared feed-forward network (FFN). MoL enables token-conditional weight-space modulation of the shared FFN without untying backbone parameters, unlike prior approaches that add fixed or externally attached adapters. We pretrain a modernised recursive architecture, ModernALBERT, integrating rotary embeddings, GeGLU, FlashAttention, and a distillation-based initialisation. Across GLUE, SQuAD-v2, and BEIR, ModernALBERT (50M--120M) achieves state-of-the-art performance among compact models and surpasses larger fully parameterised baselines. We also propose an expert-merging procedure that compresses MoL into a single adapter at inference while preserving accuracy, enabling efficient deployment. Our results show that conditional weight-space modulation effectively restores the expressivity lost under aggressive parameter sharing in recursive transformers.", "authors": ["Mohammadmahdi Nouriborji", "Morteza Rohanian", "Omid Rohanian"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-14", "url": "https://arxiv.org/abs/2512.12880", "pdf_url": "https://arxiv.org/pdf/2512.12880v2", "arxiv_id": "2512.12880", "doi": "10.48550/arXiv.2512.12880", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4079}
{"id": "e27959e599e15b06fcaefc1dfc36a78a217e43aba7a28705552faec6c8a8cf34", "sources": ["arxiv", "semantic_scholar"], "title": "V-Rex: Real-Time Streaming Video LLM Acceleration via Dynamic KV Cache Retrieval", "abstract": "Streaming video large language models (LLMs) are increasingly used for real-time multimodal tasks such as video captioning, question answering, conversational agents, and augmented reality. However, these models face fundamental memory and computational challenges because their key-value (KV) caches grow substantially with continuous streaming video input. This process requires an iterative prefill stage, which is a unique feature of streaming video LLMs. Due to its iterative prefill stage, it suffers from significant limitations, including extensive computation, substantial data transfer, and degradation in accuracy. Crucially, this issue is exacerbated for edge deployment, which is the primary target for these models. In this work, we propose V-Rex, the first software-hardware co-designed accelerator that comprehensively addresses both algorithmic and hardware bottlenecks in streaming video LLM inference. At its core, V-Rex introduces ReSV, a training-free dynamic KV cache retrieval algorithm. ReSV exploits temporal and spatial similarity-based token clustering to reduce excessive KV cache memory across video frames. To fully realize these algorithmic benefits, V-Rex offers a compact, low-latency hardware accelerator with a dynamic KV cache retrieval engine (DRE), featuring bit-level and early-exit based computing units. V-Rex achieves unprecedented real-time of 3.9-8.3 FPS and energy-efficient streaming video LLM inference on edge deployment with negligible accuracy loss. While DRE only accounts for 2.2% power and 2.0% area, the system delivers 1.9-19.7x speedup and 3.1-18.5x energy efficiency improvements over AGX Orin GPU. This work is the first to comprehensively tackle KV cache retrieval across algorithms and hardware, enabling real-time streaming video LLM inference on resource-constrained edge devices.", "authors": ["Donghyuk Kim", "Sejeong Yang", "Wonjin Shin", "Joo-Young Kim"], "categories": ["eess.IV", "cs.AI", "cs.AR", "cs.CV", "cs.MM"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-12-13", "url": "https://arxiv.org/abs/2512.12284", "pdf_url": "https://arxiv.org/pdf/2512.12284v3", "arxiv_id": "2512.12284", "doi": "10.1109/HPCA68181.2026.11408603", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Symposium on High-Performance Computer Architecture", "quality_score": 0.4068}
{"id": "be6a52b9e4560a830c51d70d7cdb023691c45c2e59b639555f921a44d1c70050", "sources": ["arxiv", "semantic_scholar"], "title": "Hold Onto That Thought: Assessing KV Cache Compression On Reasoning", "abstract": "Large language models (LLMs) have demonstrated remarkable performance on long-context tasks, but are often bottlenecked by memory constraints. Namely, the KV cache, which is used to significantly speed up attention computations, grows linearly with context length. A suite of compression algorithms has been introduced to alleviate cache growth by evicting unimportant tokens. However, several popular strategies are targeted towards the prefill phase, i.e., processing long prompt context, and their performance is rarely assessed on reasoning tasks requiring long decoding. In particular, short but complex prompts, such as those in benchmarks like GSM8K and MATH500, often benefit from multi-step reasoning and self-reflection, resulting in thinking sequences thousands of tokens long. In this work, we benchmark the performance of several popular compression strategies on long-reasoning tasks. For the non-reasoning Llama-3.1-8B-Instruct, we determine that no singular strategy fits all, and that performance is heavily influenced by dataset type. However, we discover that H2O and our decoding-enabled variant of SnapKV are dominant strategies for reasoning models, indicating the utility of heavy-hitter tracking for reasoning traces. We also find that eviction strategies at low budgets can produce longer reasoning traces, revealing a tradeoff between cache size and inference costs.", "authors": ["Minghui Liu", "Aadi Palnitkar", "Tahseen Rabbani", "Hyunwoo Jae", "Kyle Rui Sang", "Dixi Yao", "Shayan Shabihi", "Fuheng Zhao", "Tian Li", "Ce Zhang", "Furong Huang", "Kunpeng Zhang"], "categories": ["cs.CL", "cs.AI", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-12", "url": "https://arxiv.org/abs/2512.12008", "pdf_url": "https://arxiv.org/pdf/2512.12008v1", "arxiv_id": "2512.12008", "doi": "10.48550/arXiv.2512.12008", "citation_count": 3, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4056}
{"id": "670a39eec13d5f66aa9e877ad2f1d84fbcb03bf2c49eea0cf0014f0004c268c8", "sources": ["arxiv", "semantic_scholar"], "title": "CXL-SpecKV: A Disaggregated FPGA Speculative KV-Cache for Datacenter LLM Serving", "abstract": "Large Language Models (LLMs) have revolutionized natural language processing tasks, but their deployment in datacenter environments faces significant challenges due to the massive memory requirements of key-value (KV) caches. During the autoregressive decoding process, KV caches consume substantial GPU memory, limiting batch sizes and overall system throughput. To address these challenges, we propose \\textbf{CXL-SpecKV}, a novel disaggregated KV-cache architecture that leverages Compute Express Link (CXL) interconnects and FPGA accelerators to enable efficient speculative execution and memory disaggregation. Our approach introduces three key innovations: (i) a CXL-based memory disaggregation framework that offloads KV-caches to remote FPGA memory with low latency, (ii) a speculative KV-cache prefetching mechanism that predicts and preloads future tokens' cache entries, and (iii) an FPGA-accelerated KV-cache compression and decompression engine that reduces memory bandwidth requirements by up to 4$\\times$. When evaluated on state-of-the-art LLM models, CXL-SpecKV achieves up to 3.2$\\times$ higher throughput compared to GPU-only baselines, while reducing memory costs by 2.8$\\times$ and maintaining accuracy. Our system demonstrates that intelligent memory disaggregation combined with speculative execution can effectively address the memory wall challenge in large-scale LLM serving. Our code implementation has been open-sourced at https://github.com/FastLM/CXL-SpecKV.", "authors": ["Dong Liu", "Yanxuan Yu"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-11", "url": "https://arxiv.org/abs/2512.11920", "pdf_url": "https://arxiv.org/pdf/2512.11920v1", "arxiv_id": "2512.11920", "doi": "10.1145/3748173.3779188", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/FastLM/CXL-SpecKV", "venue": "Symposium on Field Programmable Gate Arrays", "quality_score": 0.6251}
{"id": "a800f70992152c4f7782434e7a0fe32dc56432f978506e66e693a1462769ad30", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Lookup Key-Value Experts", "abstract": "Recent research has developed several LLM architectures suitable for inference on end-user devices, such as the Mixture of Lookup Experts (MoLE)~\\parencite{jie_mixture_2025}. A key feature of MoLE is that each token id is associated with a dedicated group of experts. For a given input, only the experts corresponding to the input token id will be activated. Since the communication overhead of loading this small number of activated experts into RAM during inference is negligible, expert parameters can be offloaded to storage, making MoLE suitable for resource-constrained devices. However, MoLE's context-independent expert selection mechanism, based solely on input ids, may limit model performance. To address this, we propose the \\textbf{M}ixture \\textbf{o}f \\textbf{L}ookup \\textbf{K}ey-\\textbf{V}alue Experts (\\textbf{MoLKV}) model. In MoLKV, each expert is structured as a key-value pair. For a given input, the input-derived query interacts with the cached key-value experts from the current sequence, generating a context-aware expert output. This context-aware mechanism alleviates the limitation of MoLE, and experimental results demonstrate that MoLKV achieves significantly lower validation loss in small-scale evaluations.", "authors": ["Zongcheng Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-10", "url": "https://arxiv.org/abs/2512.09723", "pdf_url": "https://arxiv.org/pdf/2512.09723v1", "arxiv_id": "2512.09723", "doi": "10.48550/arXiv.2512.09723", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4033}
{"id": "d73b9e6e65f1a828446f9fe3042864e79b70479ff35349333c8f5f255f054b50", "sources": ["arxiv", "semantic_scholar"], "title": "Block Sparse Flash Attention", "abstract": "Modern large language models increasingly require long contexts for reasoning and multi-document tasks, but attention's quadratic complexity creates a severe computational bottleneck. We present Block-Sparse FlashAttention (BSFA), a drop-in replacement that accelerates long-context inference while preserving model quality. Unlike methods that predict importance before computing scores, BSFA computes exact query-key similarities to select the top-k most important value blocks for each query. By comparing per-block maximum scores against calibrated thresholds, we skip approximately 50% of the computation and memory transfers for pruned blocks. Our training-free approach requires only a one-time threshold calibration on a small dataset to learn the per-layer and per-head attention score distributions. We provide a CUDA kernel implementation that can be used as a drop-in replacement for FlashAttention. On Llama-3.1-8B, BSFA achieves up to 1.10x speedup on real-world reasoning benchmarks and up to 1.24x for needle-in-a-haystack retrieval tasks while maintaining above 99% baseline accuracy, with certain configurations even improving accuracy by focusing on the most relevant content, substantially outperforming existing sparse attention methods. The implementation is available at https://github.com/Danielohayon/Block-Sparse-Flash-Attention", "authors": ["Daniel Ohayon", "Itay Lamprecht", "Itay Hubara", "Israel Cohen", "Daniel Soudry", "Noam Elata"], "categories": ["cs.LG", "cs.CL", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-07", "url": "https://arxiv.org/abs/2512.07011", "pdf_url": "https://arxiv.org/pdf/2512.07011v1", "arxiv_id": "2512.07011", "doi": "10.48550/arXiv.2512.07011", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Danielohayon/Block-Sparse-Flash-Attention", "venue": "arXiv.org", "quality_score": 0.618}
{"id": "865cd2d9190ddc6c27cb92ad5a01b875249ea7d48d11ca527e34b51bd9190c1f", "sources": ["arxiv", "semantic_scholar"], "title": "KV-CAR: KV Cache Compression using Autoencoders and KV Reuse in Large Language Models", "abstract": "As Large Language Models (LLMs) scale in size and context length, the memory requirements of the key value (KV) cache have emerged as a major bottleneck during autoregressive decoding. The KV cache grows with sequence length and embedding dimension, often exceeding the memory footprint of the model itself and limiting achievable batch sizes and context windows. To address this challenge, we present KV CAR, a unified and architecture agnostic framework that significantly reduces KV cache storage while maintaining model fidelity. KV CAR combines two complementary techniques. First, a lightweight autoencoder learns compact representations of key and value tensors along the embedding dimension, compressing them before they are stored in the KV cache and restoring them upon retrieval. Second, a similarity driven reuse mechanism identifies opportunities to reuse KV tensors of specific attention heads across adjacent layers. Together, these methods reduce the dimensional and structural redundancy in KV tensors without requiring changes to the transformer architecture. Evaluations on GPT 2 and TinyLLaMA models across Wikitext, C4, PIQA, and Winogrande datasets demonstrate that KV CAR achieves up to 47.85 percent KV cache memory reduction with minimal impact on perplexity and zero shot accuracy. System level measurements on an NVIDIA A40 GPU show that the reduced KV footprint directly translates into longer sequence lengths and larger batch sizes during inference. These results highlight the effectiveness of KV CAR in enabling memory efficient LLM inference.", "authors": ["Sourjya Roy", "Shrihari Sridharan", "Surya Selvam", "Anand Raghunathan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-07", "url": "https://arxiv.org/abs/2512.06727", "pdf_url": "https://arxiv.org/pdf/2512.06727v1", "arxiv_id": "2512.06727", "doi": "10.48550/arXiv.2512.06727", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3999}
{"id": "9bac629d49540568baa0dda7a57cb4df81ed17c5b877ac0c83d1f7fe040bd03d", "sources": ["arxiv", "semantic_scholar"], "title": "KQ-SVD: Compressing the KV Cache with Provable Guarantees on Attention Fidelity", "abstract": "The Key-Value (KV) cache is central to the efficiency of transformer-based large language models (LLMs), storing previously computed vectors to accelerate inference. Yet, as sequence length and batch size grow, the cache becomes a major memory bottleneck. Prior compression methods typically apply low-rank decomposition to keys alone or attempt to jointly embed queries and keys, but both approaches neglect that attention fundamentally depends on their inner products. In this work, we prove that such strategies are suboptimal for approximating the attention matrix. We introduce KQ-SVD, a simple and computationally efficient method that directly performs an optimal low-rank decomposition of the attention matrix via a closed-form solution. By targeting the true source of redundancy, KQ-SVD preserves attention outputs with higher fidelity under compression. Extensive evaluations on LLaMA and Mistral models demonstrate that our approach consistently delivers superior projection quality.", "authors": ["Damien Lesens", "Beheshteh T. Rakhshan", "Guillaume Rabusseau"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-05", "url": "https://arxiv.org/abs/2512.05916", "pdf_url": "https://arxiv.org/pdf/2512.05916v1", "arxiv_id": "2512.05916", "doi": "10.48550/arXiv.2512.05916", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3976}
{"id": "1b489da4de5b04f1ad1ef0aed2c557c8df74915e62a3916b8d7ce6808d23830e", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache Recycling to Expand Usable Context Capacity in Low Parameter LLMs", "abstract": "Whether attention key value (KV) states computed for one prompt for a small LLM can be reused to accelerate inference on a new similar prompt, giving an increase to the space to its context memory using an approach called token recycling. Using a standard Hugging Face setup with DialoGPT-medium (a 345M parameter GPT-2 style decoder trained on 147M Reddit exchanges, 2005 to 2017) as the testbed, we build a cache of past activations and get entries by sentence embeddings, then reuse cached past key values when the cached prompt is an exact prefix of the new input. We compare recycled vs. baseline runs on latency and output fidelity, and log reuse depth in tokens. Reproducibility requires no model modifications, cached KVs are serialized to the CPU, reloaded, and supplied to the generate function to continue decoding from the cached prefix. In tests, we observe consistent speedups when prefix overlap exists, with no material degradation in output semantics, and when overlap is absent, behavior matches baseline.", "authors": ["Prashant Pandey"], "categories": ["cs.LG", "cs.AI", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-04", "url": "https://arxiv.org/abs/2512.11851", "pdf_url": "https://arxiv.org/pdf/2512.11851v1", "arxiv_id": "2512.11851", "doi": "10.51244/ijrsi.2025.12120067", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International journal of research and scientific innovation", "quality_score": 0.3965}
{"id": "c702175cc431f621b0771743e1578962c0552daa393d62f5b909c78a85d38d15", "sources": ["arxiv", "semantic_scholar"], "title": "Context-Aware Mixture-of-Experts Inference on CXL-Enabled GPU-NDP Systems", "abstract": "Mixture-of-Experts (MoE) models scale large language models through conditional computation, but inference becomes memory-bound once expert weights exceed the capacity of GPU memory. In this case, weights must be offloaded to external memory, and fetching them incurs costly and repeated transfers. We address this by adopting CXL-attached near-data processing (CXL-NDP) as the offloading tier to execute cold experts in place, converting expensive parameter movement into cheaper activation movement. Unlike prior GPU-NDP systems that are largely context-agnostic and reactive, we develop a context-aware MoE system that uses prefill-stage activation statistics to guide decoding-stage expert placement, dynamically pins hot experts in GPU-side HBM, and maps the remainder to CXL-NDP. To meet NDP's limited compute throughput, we introduce context-aware mixed-precision quantization that allocates per-expert bitwidths (1-4 bit) based on prefill stage. The resulting MoE inference system overlaps GPU and NDP execution while minimizing cross-device movement. The evaluation on the GPU-NDP system shows that our approach achieves up to an 8.7-fold decoding throughput improvement over the state-of-the-art method, while incurring only a 0.13% average accuracy drop.", "authors": ["Zehao Fan", "Zhenyu Liu", "Yunzhen Liu", "Yayue Hou", "Hadjer Benmeziane", "Kaoutar El Maghraoui", "Liu Liu"], "categories": ["cs.LG", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-04", "url": "https://arxiv.org/abs/2512.04476", "pdf_url": "https://arxiv.org/pdf/2512.04476v1", "arxiv_id": "2512.04476", "doi": "10.48550/arXiv.2512.04476", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3965}
{"id": "616e1ad132ef87b06a0cc08a7a9f7e9993089aa1893d769a28fbdeeb43adca4b", "sources": ["arxiv", "semantic_scholar"], "title": "Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs", "abstract": "Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), conversational long-memory benchmarks (LongMemEval), and long-context understanding (LongBenchV2 and SCBench), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.", "authors": ["Ngoc Bui", "Shubham Sharma", "Simran Lamba", "Saumitra Mishra", "Rex Ying"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-03", "url": "https://arxiv.org/abs/2512.03324", "pdf_url": "https://arxiv.org/pdf/2512.03324v2", "arxiv_id": "2512.03324", "doi": "10.48550/arXiv.2512.03324", "citation_count": 9, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3953}
{"id": "7f575f80281b23ba49b54aa04bd23edc6c553c7025672c52f1648c329a5202b6", "sources": ["arxiv", "semantic_scholar"], "title": "Reconstructing KV Caches with Cross-layer Fusion For Enhanced Transformers", "abstract": "Transformer decoders have achieved strong results across tasks, but the memory required for the KV cache becomes prohibitive at long sequence lengths. Although Cross-layer KV Cache sharing (e.g., YOCO, CLA) offers a path to mitigate KV Cache bottleneck, it typically underperforms within-layer methods like GQA. To understand the root cause, we investigate the information flow of keys and values of the top-layers. Our preliminary reveals a clear distribution: values are predominantly derived from the bottom layer, while keys draw more information from both bottom and middle layers. Building upon this, we propose FusedKV, whose top-layer KV caches are a learnable fusion of the most informative ones from the bottom and middle layers. This fusion operates directly on post-RoPE keys, preserving relative positional information without the computational cost of re-applying rotary embeddings. To further improve efficiency, we propose FusedKV-Lite, an cross-layer sharing approach, where top-layer KV caches are directly derived from the bottom-layer values and the middle-layer keys. Compared to FusedKV, FusedKV-Lite reduces I/O overhead at the cost of a slight increase in perplexity. In experiments on LLMs ranging from 332M to 4B parameters, our proposed method reduce 50\\% cache memory while achieving lower validation perplexity than the standard Transformer decoder, establishing it as a memory-efficient, high-performance architectural alternative.", "authors": ["Hongzhan Lin", "Zhiqi Bai", "Xinmiao Zhang", "Sen Yang", "Xiang Li", "Siran Yang", "Yunlong Xu", "Jiaheng Liu", "Yongchi Zhao", "Jiamang Wang", "Yuchi Xu", "Wenbo Su", "Bo Zheng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-03", "url": "https://arxiv.org/abs/2512.03870", "pdf_url": "https://arxiv.org/pdf/2512.03870v3", "arxiv_id": "2512.03870", "doi": "10.48550/arXiv.2512.03870", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3953}
{"id": "25ce0156ade95cdbee88160332d0c99b62eb606fbfc18644cf0b9df950ec1110", "sources": ["arxiv", "semantic_scholar"], "title": "Data-Free Pruning of Self-Attention Layers in LLMs", "abstract": "Many self-attention sublayers in large language models (LLMs) can be removed with little to no loss. We attribute this to the Attention Suppression Hypothesis: during pre-training, some deep attention layers learn to mute their own contribution, leaving the residual stream and the MLP to carry the representation. We propose Gate-Norm, a one-shot, weight-only criterion that ranks attention sublayers by query--key coupling and removes the least coupled ones, requiring no calibration data, no forward passes, no fine-tuning, and no specialized kernels. On 40-layer, 13B-parameter LLaMA models, Gate-Norm prunes the model in under a second. Pruning $8$--$16$ attention sublayers yields up to $1.30\\times$ higher inference throughput while keeping average zero-shot accuracy within $2\\%$ of the unpruned baseline across BoolQ, RTE, HellaSwag, WinoGrande, ARC-Easy/Challenge, and OpenBookQA. Across these settings, Gate-Norm matches data-driven pruning methods in accuracy while being $\\sim 1000\\times$ faster to score layers, enabling practical, data-free compression of LLMs.", "authors": ["Dhananjay Saikumar", "Blesson Varghese"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-03", "url": "https://arxiv.org/abs/2512.20636", "pdf_url": "https://arxiv.org/pdf/2512.20636v1", "arxiv_id": "2512.20636", "doi": "10.48550/arXiv.2512.20636", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3953}
{"id": "c76a1db3ffbbd19a281e0fe22416183410a4708e2ec83f087509884554f73b98", "sources": ["arxiv", "semantic_scholar"], "title": "KVNAND: Efficient On-Device Large Language Model Inference Using DRAM-Free In-Flash Computing", "abstract": "Deploying large language models (LLMs) on edge devices enables personalized agents with strong privacy and low cost. However, with tens to hundreds of billions of parameters, single-batch autoregressive inference suffers from extremely low arithmetic intensity, creating severe weight-loading and bandwidth pressures on resource-constrained platforms. Recent in-flash computing (IFC) solutions alleviate this bottleneck by co-locating weight-related linear computations in the decode phase with flash, yet still rely on DRAM for the key-value (KV) cache. As context length grows, the KV cache can exceed model weights in size, imposing prohibitive DRAM cost and capacity requirements. Attempts to offload KV cache to flash suffer from severe performance penalties. We propose KVNAND, the first DRAM-free, IFC-based architecture that stores both model weights and KV cache entirely in compute-enabled 3D NAND flash. KVNAND addresses the fundamental performance challenges of flash under intensive KV cache access by leveraging IFC for all memory-bound operations to reduce data transfer overhead, introducing head-group parallelism to boost throughput, and employing page-level KV cache mapping to align token access patterns with flash organization. In addition, we propose a design space exploration framework that evaluates discrete and compact KVNAND variants to balance weight and KV placement, automatically identifying the optimal design trade-off. These techniques mitigate latency, energy, and reliability concerns, turning flash into a practical medium for long-context KV storage. Evaluations on MHA 7B and GQA 70B LLMs show that KVNAND achieves 1.98\\(\\times\\)/1.94\\(\\times\\)/2.05\\(\\times\\) geomean speedup at 128/1K/10K-token contexts compared to DRAM-equipped IFC designs and addresses out-of-memory failures at 100K context length.", "authors": ["Lishuo Deng", "Shaojie Xu", "Jinwu Chen", "Changwei Yan", "Jiajie Wang", "Zhe Jiang", "Weiwei Shan"], "categories": ["cs.AR", "cs.AI", "cs.ET"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-03", "url": "https://arxiv.org/abs/2512.03608", "pdf_url": "https://arxiv.org/pdf/2512.03608v1", "arxiv_id": "2512.03608", "doi": "10.48550/arXiv.2512.03608", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3953}
{"id": "ada476253d6386742ad9f5c9b72ec583db800790f93062e436c33b8bac7d6ea1", "sources": ["arxiv", "semantic_scholar"], "title": "In-Context Sync-LoRA for Portrait Video Editing", "abstract": "Editing portrait videos is a challenging task that requires flexible yet precise control over a wide range of modifications, such as appearance changes, expression edits, or the addition of objects. The key difficulty lies in preserving the subject's original temporal behavior, demanding that every edited frame remains precisely synchronized with the corresponding source frame. We present Sync-LoRA, a method for editing portrait videos that achieves high-quality visual modifications while maintaining frame-accurate synchronization and identity consistency. Our approach uses an image-to-video diffusion model, where the edit is defined by modifying the first frame and then propagated to the entire sequence. To enable accurate synchronization, we train an in-context LoRA using paired videos that depict identical motion trajectories but differ in appearance. These pairs are automatically generated and curated through a synchronization-based filtering process that selects only the most temporally aligned examples for training. This training setup teaches the model to combine motion cues from the source video with the visual changes introduced in the edited first frame. Trained on a compact, highly curated set of synchronized human portraits, Sync-LoRA generalizes to unseen identities and diverse edits (e.g., modifying appearance, adding objects, or changing backgrounds), robustly handling variations in pose and expression. Our results demonstrate high visual fidelity and strong temporal coherence, achieving a robust balance between edit fidelity and precise motion preservation.", "authors": ["Sagi Polaczek", "Or Patashnik", "Ali Mahdavi-Amiri", "Daniel Cohen-Or"], "categories": ["cs.CV", "cs.AI", "cs.GR"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-02", "url": "https://arxiv.org/abs/2512.03013", "pdf_url": "https://arxiv.org/pdf/2512.03013v1", "arxiv_id": "2512.03013", "doi": "10.48550/arXiv.2512.03013", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3942}
{"id": "a4de734316c895992b1f99158cb2ac257ed582e3083d649a9dd39f5fdae3011c", "sources": ["arxiv", "semantic_scholar"], "title": "SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts", "abstract": "The emergence of large vision-language models (VLMs) has significantly enhanced the efficiency and flexibility of geospatial interpretation. However, general-purpose VLMs remain suboptimal for remote sensing (RS) tasks. Existing geospatial VLMs typically adopt a unified modeling strategy and struggle to differentiate between task types and interpretation granularities, limiting their ability to balance local detail perception and global contextual understanding. In this paper, we present SkyMoE, a Mixture-of-Experts (MoE) vision-language model tailored for multimodal, multi-task RS interpretation. SkyMoE employs an adaptive router that generates task- and granularity-aware routing instructions, enabling specialized large language model experts to handle diverse sub-tasks. To further promote expert decoupling and granularity sensitivity, we introduce a context-disentangled augmentation strategy that creates contrastive pairs between local and global features, guiding experts toward level-specific representation learning. We also construct MGRS-Bench, a comprehensive benchmark covering multiple RS interpretation tasks and granularity levels, to evaluate generalization in complex scenarios. Extensive experiments on 21 public datasets demonstrate that SkyMoE achieves state-of-the-art performance across tasks, validating its adaptability, scalability, and superior multi-granularity understanding in remote sensing.", "authors": ["Jiaqi Liu", "Ronghao Fu", "Lang Sun", "Haoran Liu", "Xiao Yang", "Weipeng Zhang", "Xu Na", "Zhuoran Duan", "Bo Yang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-02", "url": "https://arxiv.org/abs/2512.02517", "pdf_url": "https://arxiv.org/pdf/2512.02517v1", "arxiv_id": "2512.02517", "doi": "10.48550/arXiv.2512.02517", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2508}
{"id": "87c8a4eb69f8ba59d9d6256800a7871efbda94c00f458dcc84abbfb7354bef4c", "sources": ["arxiv", "semantic_scholar"], "title": "KVReviver: Reversible KV Cache Compression with Sketch-Based Token Reconstruction", "abstract": "As the context length of current large language models (LLMs) rapidly increases, the memory demand for the Key-Value (KV) cache is becoming a bottleneck for LLM deployment and batch processing. Traditional KV cache compression methods typically involve permanently evicting or irreversibly merging \"less important\" tokens with low attention scores. This approach results in the unrecoverable loss of token information, which we call Contextual Amnesia, significantly degrading the model's information retrieval capability. To address this issue, we propose KVReviver, a reversible KV cache compression method based on the sketch algorithm. This method allows reconstructing compressed tokens from an additional data structure, thus enabling full-scale computation within limited memory. Experiments showed that in 2k-length contexts, it requires only 10% of KV Cache budget while maintaining identical end-to-end inference accuracy. For 32k-length contexts, it achieves equivalent or comparable accuracy ~2% accuracy loss) using merely 25% of KV Cache budget.", "authors": ["Aomufei Yuan", "Zhiming Wang", "Ruijie Miao", "Dayu Wang", "Yuxuan Tian", "Zihan Wang", "Yebo Peng", "Yuhan Wu", "Bairen Yi", "Xin Liu", "Tong Yang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-01", "url": "https://arxiv.org/abs/2512.17917", "pdf_url": "https://arxiv.org/pdf/2512.17917v1", "arxiv_id": "2512.17917", "doi": "10.48550/arXiv.2512.17917", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.393}
{"id": "d12ede5727bc61eb16c177b636a14a003bb86d6053e5a242a33053b52da44e1a", "sources": ["arxiv", "semantic_scholar"], "title": "KV Pareto: Systems-Level Optimization of KV Cache and Model Compression for Long Context Inference", "abstract": "Long-context Large Language Models (LLMs) face significant memory bottlenecks during inference due to the linear growth of key-value (KV) cache with sequence length. While individual optimization techniques like KV cache quantization, chunked prefill, and model weight quantization have shown promise, their joint effects and optimal configurations for edge deployment remain underexplored. We introduce KV Pareto, a systems-level framework that systematically maps the trade-off frontier between total memory consumption and task accuracy across these three complementary optimization techniques. Our framework evaluates multiple LLM architectures (Qwen, Llama, Mistral) with varying KV quantization schemes (int2/4/8, mixed-precision), granularities (per-token, per-tensor, per-block), and 4-bit weight quantization via AWQ. Our framework identifies model-specific Pareto-optimal configurations that achieve 68-78% total memory reduction with minimal (1-3%) accuracy degradation on long-context tasks. We additionally verify the selected frontiers on additional benchmarks of Needle-in-a-Haystack, GSM8k and MMLU as well as extended context lengths of up to 128k to demonstrate the practical need of joint optimization for efficient LLM inference.", "authors": ["Sai Gokhale", "Devleena Das", "Rajeev Patwari", "Ashish Sirasao", "Elliott Delaye"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-01", "url": "https://arxiv.org/abs/2512.01953", "pdf_url": "https://arxiv.org/pdf/2512.01953v1", "arxiv_id": "2512.01953", "doi": "10.48550/arXiv.2512.01953", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2501}
{"id": "92c97177476f097e6bc4cf426f71ac314313d76ff2d248d5eec3732552fa4934", "sources": ["arxiv", "semantic_scholar"], "title": "Elastic Mixture of Rank-Wise Experts for Knowledge Reuse in Federated Fine-Tuning", "abstract": "Federated fine-tuning offers a promising solution for adapting Large Language Models (LLMs) to downstream tasks while safeguarding data privacy. However, its high computational and communication demands hinder its deployment on resource-constrained devices. In this paper, we propose SmartFed, a resource-efficient federated fine-tuning framework. SmartFed intelligently reuses knowledge embedded in existing LoRA modules, eliminating the need for expensive training from scratch when adapting LLMs to new tasks. To effectively exploit this knowledge and ensure scalability, we introduce the Mixture of Rank-Wise Experts (MoRE). MoRE decomposes LoRA modules into fine-grained rank-level experts. These experts are selectively activated and combined based on input semantics and resource budgets. Moreover, to optimize resource utilization, we present the Elastic Expert Quota Allocation (EEQA). EEQA adaptively allocates expert capacity across parameter matrices based on their contribution to model performance, focusing computing resources on the critical experts. Extensive evaluations across multiple benchmarks demonstrate that SmartFed significantly outperforms existing methods in model performance and training efficiency.", "authors": ["Yebo Wu", "Jingguang Li", "Zhijiang Guo", "Li Li"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-30", "url": "https://arxiv.org/abs/2512.00902", "pdf_url": "https://arxiv.org/pdf/2512.00902v1", "arxiv_id": "2512.00902", "doi": "10.48550/arXiv.2512.00902", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3919}
{"id": "edf6f0145bce25258de3070033e6fd327b20d3d555478ae9ee95fc0c0b4e88ba", "sources": ["arxiv", "semantic_scholar"], "title": "G-KV: Decoding-Time KV Cache Eviction with Global Attention", "abstract": "Recent reasoning large language models (LLMs) excel in complex tasks but encounter significant computational and memory challenges due to long sequence lengths. KV cache compression has emerged as an effective approach to greatly enhance the efficiency of reasoning. However, existing methods often focus on prompt compression or token eviction with local attention score, overlooking the long-term importance of tokens. We propose G-KV, a KV cache eviction method that employs a global scoring mechanism, combining local and historical attention scores to more accurately assess token importance. Additionally, we introduce post-training techniques, including reinforcement learning and distillation, to optimize models for compressed KV cache settings. The code of this paper is available on: https://github.com/microsoft/G-KV.", "authors": ["Mengqi Liao", "Lu Wang", "Chaoyun Zhang", "Zekai Shen", "Xiaowei Mao", "Si Qin", "Qingwei Lin", "Saravan Rajmohan", "Dongmei Zhang", "Huaiyu Wan"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-29", "url": "https://arxiv.org/abs/2512.00504", "pdf_url": "https://arxiv.org/pdf/2512.00504v1", "arxiv_id": "2512.00504", "doi": "10.48550/arXiv.2512.00504", "citation_count": 5, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/microsoft/G-KV", "venue": "arXiv.org", "quality_score": 0.6039}
{"id": "77c12673812ec955a3b29862e1d306fea057df7e7f2da406361816f0bdb0b9d9", "sources": ["arxiv", "semantic_scholar"], "title": "Ada-MoGE: Adaptive Mixture of Gaussian Expert Model for Time Series Forecasting", "abstract": "Multivariate time series forecasts are widely used, such as industrial, transportation and financial forecasts. However, the dominant frequencies in time series may shift with the evolving spectral distribution of the data. Traditional Mixture of Experts (MoE) models, which employ a fixed number of experts, struggle to adapt to these changes, resulting in frequency coverage imbalance issue. Specifically, too few experts can lead to the overlooking of critical information, while too many can introduce noise. To this end, we propose Ada-MoGE, an adaptive Gaussian Mixture of Experts model. Ada-MoGE integrates spectral intensity and frequency response to adaptively determine the number of experts, ensuring alignment with the input data's frequency distribution. This approach prevents both information loss due to an insufficient number of experts and noise contamination from an excess of experts. Additionally, to prevent noise introduction from direct band truncation, we employ Gaussian band-pass filtering to smoothly decompose the frequency domain features, further optimizing the feature representation. The experimental results show that our model achieves state-of-the-art performance on six public benchmarks with only 0.2 million parameters.", "authors": ["Zhenliang Ni", "Xiaowen Ma", "Zhenkai Wu", "Shuai Xiao", "Han Shu", "Xinghao Chen"], "categories": ["cs.LG", "cs.AI", "cs.CE"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-29", "url": "https://arxiv.org/abs/2512.02061", "pdf_url": "https://arxiv.org/pdf/2512.02061v1", "arxiv_id": "2512.02061", "doi": "10.48550/arXiv.2512.02061", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3907}
{"id": "ec1484a0961f4621076cfa6af62e9aade24b02a0c26d34d567b9ccc31f7a9d6a", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Dataset Quantization: A New Direction for Dataset Pruning", "abstract": "This paper addresses the challenges of storage and communication costs for large-scale datasets in resource-constrained edge devices by proposing a novel dataset quantization approach to reduce intra-sample redundancy. Unlike traditional dataset pruning and distillation methods that focus on inter-sample redundancy, the proposed method compresses each image by reducing redundant or less informative content within samples while preserving essential features. It first applies linear symmetric quantization to obtain an initial quantization range and scale for each sample. Then, an adaptive quantization allocation algorithm is introduced to distribute different quantization ratios for samples with varying precision requirements, maintaining a constant total compression ratio. The main contributions include: (1) being the first to use limited bits to represent datasets for storage reduction; (2) introducing a dataset-level quantization algorithm with adaptive ratio allocation; and (3) validating the method's effectiveness through extensive experiments on CIFAR-10, CIFAR-100, and ImageNet-1K. Results show that the method maintains model training performance while achieving significant dataset compression, outperforming traditional quantization and dataset pruning baselines under the same compression ratios.", "authors": ["Chenyue Yu", "Jianyu Yu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-28", "url": "https://arxiv.org/abs/2512.05987", "pdf_url": "https://arxiv.org/pdf/2512.05987v1", "arxiv_id": "2512.05987", "doi": "10.48550/arXiv.2512.05987", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3896}
{"id": "62ef60c20d4777733669831a88f552cae96a62992805009f60bab7ddcf595a97", "sources": ["arxiv", "semantic_scholar"], "title": "MoE3D: Mixture of Experts meets Multi-Modal 3D Understanding", "abstract": "Multi-modal 3D understanding is a fundamental task in computer vision. Previous multi-modal fusion methods typically employ a single, dense fusion network, struggling to handle the significant heterogeneity and complexity across modalities, leading to suboptimal performance. In this paper, we propose MoE3D, which integrates Mixture of Experts (MoE) into the multi-modal learning framework. The core is that we deploy a set of specialized \"expert\" networks, each adept at processing a specific modality or a mode of cross-modal interaction. Specifically, the MoE-based transformer is designed to better utilize the complementary information hidden in the visual features. Information aggregation module is put forward to further enhance the fusion performance. Top-1 gating is employed to make one expert process features with expert groups, ensuring high efficiency. We further propose a progressive pre-training strategy to better leverage the semantic and 2D prior, thus equipping the network with good initialization. Our MoE3D achieves competitive performance across four prevalent 3D understanding tasks. Notably, our MoE3D surpasses the top-performing counterpart by 6.1 mIoU on Multi3DRefer.", "authors": ["Yu Li", "Yuenan Hou", "Yingmei Wei", "Xinge Zhu", "Yuexin Ma", "Wenqi Shao", "Yanming Guo"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-27", "url": "https://arxiv.org/abs/2511.22103", "pdf_url": "https://arxiv.org/pdf/2511.22103v1", "arxiv_id": "2511.22103", "doi": "10.48550/arXiv.2511.22103", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3884}
{"id": "3846027f41cda380d882c78e391f31c2a3959ef2f46061386c543e13a90776b6", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Multi-Adapter LLM Serving via Cross-Model KV-Cache Reuse with Activated LoRA", "abstract": "Modern large language model (LLM) systems increasingly rely on multi-turn pipelines that are composed of multiple task-specific adapters, yet existing serving frameworks remain inefficient, incurring substantial recomputation overhead when switching between adapters. We present the first LLM serving engine that supports cross-model prefix cache reuse between base and adapted models via Activated LoRA (aLoRA), enabling efficient and fine-grained adapter switching during inference. Our design extends the vLLM framework by introducing base-aligned block hashing and activation-aware masking within the model execution path, permitting cache reuse across models while preserving compatibility with existing serving engine optimizations. Integrated into a production-grade inference stack, this approach supports dynamic adapter activation without excessive key-value tensor recomputation. Evaluation across representative multi-turn, multi-adapter pipelines demonstrates up to 58x end-to-end latency reduction and over 100x time-to-first-token improvement relative to standard LoRA baselines, with benefits that scale with model size and sequence length and manifest across all stages of the request lifecycle. This work bridges parameter-efficient model adaptation with high-performance serving, providing the first complete realization of cross-model KV-cache reuse in modern LLM inference engines.", "authors": ["Allison Li", "Kristjan Greenewald", "Thomas Parnell", "Navid Azizan"], "categories": ["cs.DC", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-26", "url": "https://arxiv.org/abs/2512.17910", "pdf_url": "https://arxiv.org/pdf/2512.17910v1", "arxiv_id": "2512.17910", "doi": "10.48550/arXiv.2512.17910", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3873}
{"id": "044e3293136beaf00f0ee3f0e4e23762c927ebca036ec77266e9fa1b05b12b54", "sources": ["arxiv", "semantic_scholar"], "title": "MLPMoE: Zero-Shot Architectural Metamorphosis of Dense LLM MLPs into Static Mixture-of-Experts", "abstract": "Large Language Models (LLMs) are predominantly deployed as dense transformers, where every parameter in every feed-forward block is activated for every token. While architecturally simple, this is computationally inefficient, since inference costs scale linearly with parameter count. Recent upcycling methods such as MoEfication, CMoE, ToMoE, and MoORE reveal that much of the useful computation lives in sparse, semi-modular substructures inside dense feed-forward networks, but these approaches typically rely on clustering, activation profiling, singular value decomposition, or custom routing that requires calibration data. This paper introduces MLPMoE (MLP Mixture-of-Experts), a training-free, deterministic transformation that restructures the dense MLP in transformer blocks into a static, high-cardinality mixture of experts. The transformation uses simple tensor slicing and summation, reinterpreting the algebra of tensor parallelism as a topological conversion rather than a distributed training pattern. We further introduce Fractal Fade (differential branch sparsity) and Compensated Pruning (variance-preserving branch reduction) as lightweight mechanisms for structured sparsity. On Qwen2.5-0.5B-Instruct and DeepSeek-R1-Distill-Llama-8B, the zero-shot MLPMoE transform changes a proxy perplexity metric by less than 0.05 percent while keeping the parameter count effectively constant. On the 8B model, differential sparsity removes about 20 percent of MLP parameters while keeping perplexity within about 2 percent of the dense baseline. The method operates entirely post hoc on existing checkpoints and does not require gradients, calibration sets, or router training. Code is available at https://gist.github.com/iwallarm/fc2ef1eddf226ca7814f9e5e2ae9bad1", "authors": ["Ivan Novikov"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-26", "url": "https://arxiv.org/abs/2511.21089", "pdf_url": "https://arxiv.org/pdf/2511.21089v1", "arxiv_id": "2511.21089", "doi": "10.48550/arXiv.2511.21089", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5985}
{"id": "3ba25882c63e479da29994695948734279fb7fd583ff1e5f1f5cf595cb38eb79", "sources": ["arxiv", "semantic_scholar"], "title": "Mosaic Pruning: A Hierarchical Framework for Generalizable Pruning of Mixture-of-Experts Models", "abstract": "Sparse Mixture-of-Experts (SMoE) architectures have enabled a new frontier in scaling Large Language Models (LLMs), offering superior performance by activating only a fraction of their total parameters during inference. However, their practical deployment is severely hampered by substantial static memory overhead, as all experts must be loaded into memory. Existing post-training pruning methods, while reducing model size, often derive their pruning criteria from a single, general-purpose corpus. This leads to a critical limitation: a catastrophic performance degradation when the pruned model is applied to other domains, necessitating a costly re-pruning for each new domain. To address this generalization gap, we introduce Mosaic Pruning (MoP). The core idea of MoP is to construct a functionally comprehensive set of experts through a structured ``cluster-then-select\" process. This process leverages a similarity metric that captures expert performance across different task domains to functionally cluster the experts, and subsequently selects the most representative expert from each cluster based on our proposed Activation Variability Score. Unlike methods that optimize for a single corpus, our proposed Mosaic Pruning ensures that the pruned model retains a functionally complementary set of experts, much like the tiles of a mosaic that together form a complete picture of the original model's capabilities, enabling it to handle diverse downstream tasks.Extensive experiments on various MoE models demonstrate the superiority of our approach. MoP significantly outperforms prior work, achieving a 7.24\\% gain on general tasks and 8.92\\% on specialized tasks like math reasoning and code generation.", "authors": ["Wentao Hu", "Mingkuan Zhao", "Shuangyong Song", "Xiaoyan Zhu", "Xin Lai", "Jiayin Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-25", "url": "https://arxiv.org/abs/2511.19822", "pdf_url": "https://arxiv.org/pdf/2511.19822v1", "arxiv_id": "2511.19822", "doi": "10.48550/arXiv.2511.19822", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3861}
{"id": "9656a2ac2314a3be3fd2a33492b3ddf8d481528511bc1440bcff5bcc0895db22", "sources": ["arxiv", "semantic_scholar"], "title": "MapReduce LoRA: Advancing the Pareto Front in Multi-Preference Optimization for Generative Models", "abstract": "Reinforcement learning from human feedback (RLHF) with reward models has advanced alignment of generative models to human aesthetic and perceptual preferences. However, jointly optimizing multiple rewards often incurs an alignment tax, improving one dimension while degrading others. To address this, we introduce two complementary methods: MapReduce LoRA and Reward-aware Token Embedding (RaTE). MapReduce LoRA trains preference-specific LoRA experts in parallel and iteratively merges them to refine a shared base model; RaTE learns reward-specific token embeddings that compose at inference for flexible preference control. Experiments on Text-to-Image generation (Stable Diffusion 3.5 Medium and FLUX.1-dev) show improvements of 36.1%, 4.6%, and 55.7%, and 32.7%, 4.3%, and 67.1% on GenEval, PickScore, and OCR, respectively. On Text-to-Video generation (HunyuanVideo), visual and motion quality improve by 48.1% and 90.0%, respectively. On the language task, Helpful Assistant, with Llama-2 7B, helpful and harmless improve by 43.4% and 136.7%, respectively. Our framework sets a new state-of-the-art multi-preference alignment recipe across modalities.", "authors": ["Chieh-Yun Chen", "Zhonghao Wang", "Qi Chen", "Zhifan Ye", "Min Shi", "Yue Zhao", "Yinan Zhao", "Hui Qu", "Wei-An Lin", "Yiru Shen", "Ajinkya Kale", "Irfan Essa", "Humphrey Shi"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-25", "url": "https://arxiv.org/abs/2511.20629", "pdf_url": "https://arxiv.org/pdf/2511.20629v5", "arxiv_id": "2511.20629", "doi": "10.48550/arXiv.2511.20629", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SHI-Labs/MapReduce-LoRA", "venue": "arXiv.org", "quality_score": 0.5968}
{"id": "146460477d14673b3378b9633ef796f1ed7f6f64d148e8dd624a54c8c82849cd", "sources": ["arxiv", "semantic_scholar"], "title": "Context-Aware Token Pruning and Discriminative Selective Attention for Transformer Tracking", "abstract": "One-stream Transformer-based trackers have demonstrated remarkable performance by concatenating template and search region tokens, thereby enabling joint attention across all tokens. However, enabling an excessive proportion of background search tokens to attend to the target template tokens weakens the tracker's discriminative capability. Several token pruning methods have been proposed to mitigate background interference; however, they often remove tokens near the target, leading to the loss of essential contextual information and degraded tracking performance. Moreover, the presence of distractors within the search tokens further reduces the tracker's ability to accurately identify the target. To address these limitations, we propose CPDATrack, a novel tracking framework designed to suppress interference from background and distractor tokens while enhancing computational efficiency. First, a learnable module is integrated between two designated encoder layers to estimate the probability of each search token being associated with the target. Based on these estimates, less-informative background tokens are pruned from the search region while preserving the contextual cues surrounding the target. To further suppress background interference, a discriminative selective attention mechanism is employed that fully blocks search-to-template attention in the early layers. In the subsequent encoder layers, high-probability target tokens are selectively extracted from a localized region to attend to the template tokens, thereby reducing the influence of background and distractor tokens. The proposed CPDATrack achieves state-of-the-art performance across multiple benchmarks, particularly on GOT-10k, where it attains an average overlap of 75.1 percent.", "authors": ["Janani Kugarajeevan", "Thanikasalam Kokul", "Amirthalingam Ramanan", "Subha Fernando"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-25", "url": "https://arxiv.org/abs/2511.19928", "pdf_url": "https://arxiv.org/pdf/2511.19928v1", "arxiv_id": "2511.19928", "doi": "10.48550/arXiv.2511.19928", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3861}
{"id": "06778b39a75c95d651a15b147afb0f19019b55b330dd32640ecfa5bc509beaea", "sources": ["arxiv", "semantic_scholar"], "title": "EfficientXpert: Efficient Domain Adaptation for Large Language Models via Propagation-Aware Pruning", "abstract": "Large language models (LLMs) are increasingly adapted into domain-specific variants for applications in law, healthcare, and finance. Their scale, however, limits deployment in resource-constrained settings, and existing compression approaches often either degrade after domain adaptation or require substantial additional computation. We introduce EfficientXpert, a lightweight framework for domain pruning that integrates ForeSight Mask, a propagation-aware criterion for selecting weights to prune without backpropagation, and Partial Brain Surgeon, an efficient closed-form update for low-rank adapters under a fixed sparsity pattern. With fine-tuning cost comparable to standard LoRA, EfficientXpert converts a general pretrained model into a sparse, domain-adapted expert in a single pruning step. Across health and legal benchmarks, EfficientXpert reaches up to 98 percent of dense performance at 40 percent sparsity, improving over prior pruning baselines while matching LoRA training time and staying within 1 percent of LoRA peak GPU memory in our experiments.", "authors": ["Songlin Zhao", "Michael Pitts", "Zhuwei Qin"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-25", "url": "https://arxiv.org/abs/2511.19935", "pdf_url": "https://arxiv.org/pdf/2511.19935v2", "arxiv_id": "2511.19935", "doi": "10.48550/arXiv.2511.19935", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3861}
{"id": "a30c2209fe9c5a4325ad57318e3df63e618df2b3e85bbb41266e96be6aa5b1ad", "sources": ["arxiv", "semantic_scholar"], "title": "SWAN: Sparse Winnowed Attention for Reduced Inference Memory via Decompression-Free KV-Cache Compression", "abstract": "Large Language Models (LLMs) face a significant bottleneck during autoregressive inference due to the massive memory footprint of the Key-Value (KV) cache. Existing compression techniques like token eviction, quantization, or other low-rank methods often risk information loss, have fixed limits, or introduce significant computational overhead from explicit decompression steps. In this work, we introduce SWAN, a novel, fine-tuning-free framework that eliminates this overhead. Our method uses an offline orthogonal matrix to rotate and prune the KV-cache, which is then used directly in the attention computation without any reconstruction. Our extensive experiments demonstrate that SWAN, augmented with a small dense buffer, offers a robust trade-off, maintaining performance close to the uncompressed baseline even at aggressive 50-60% memory savings per-token on KV-cache. A key advantage is its runtime-tunable compression level, allowing operators to dynamically adjust the memory footprint, a flexibility absent in methods requiring fixed offline configurations. This combination of a decompression-free design, high performance under compression, and adaptability makes SWAN a practical and efficient solution for serving LLMs with long contexts.", "authors": ["Santhosh G S", "Saurav Prakash", "Balaraman Ravindran"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-24", "url": "https://arxiv.org/abs/2511.18936", "pdf_url": "https://arxiv.org/pdf/2511.18936v1", "arxiv_id": "2511.18936", "doi": "10.48550/arXiv.2511.18936", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.385}
{"id": "48e08624aba51f515e02ad8281b925d51b942454e48a188d6c417f59617affea", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Mixture of Experts Against Severe Distribution Shifts", "abstract": "The challenge of building neural networks that can continuously learn and adapt to evolving data streams is central to the fields of continual learning (CL) and reinforcement learning (RL). This lifelong learning problem is often framed in terms of the plasticity-stability dilemma, focusing on issues like loss of plasticity and catastrophic forgetting. Unlike neural networks, biological brains maintain plasticity through capacity growth, inspiring researchers to explore similar approaches in artificial networks, such as adding capacity dynamically. Prior solutions often lack parameter efficiency or depend on explicit task indices, but Mixture-of-Experts (MoE) architectures offer a promising alternative by specializing experts for distinct distributions. This paper aims to evaluate a DynamicMoE approach for continual and reinforcement learning environments and benchmark its effectiveness against existing network expansion methods.", "authors": ["Donghu Kim"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-24", "url": "https://arxiv.org/abs/2511.18987", "pdf_url": "https://arxiv.org/pdf/2511.18987v1", "arxiv_id": "2511.18987", "doi": "10.48550/arXiv.2511.18987", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.385}
{"id": "a380f99a141b43336633d8eb1d41af5be60024b26d14433a63f29f1d9cf8a37b", "sources": ["arxiv", "semantic_scholar"], "title": "OrdMoE: Preference Alignment via Hierarchical Expert Group Ranking in Multimodal Mixture-of-Experts LLMs", "abstract": "Preference learning has recently emerged as a pivotal strategy for post-training alignment of Multimodal Large Language Models (MLLMs). However, existing approaches predominantly rely on external human-annotated preference data, which is costly and labor-intensive to collect. In this work, we propose OrdMoE, a novel preference alignment framework that bypasses the reliance on external human preferences entirely by leveraging intrinsic signals within Mixture-of-Experts (MoE) architectures. Specifically, we observe that the router's expert selection scores implicitly encode a quality-aware ranking of responses (i.e. higher-scoring experts consistently generate higher-quality outputs). Building on this insight, OrdMoE constructs an internal preference hierarchy by grouping experts into ranked tiers based on their per-token routing scores and activating each tier separately to produce a sequence of responses with increasing quality. This yields a zero-cost, self-supervised preference ordering over generated responses, which can be directly optimized using standard preference learning objectives. Extensive experiments across multiple multimodal benchmarks demnstrate that OrdMoE significantly enhances both alignment and overall performance of multimodal Mixture-of-Experts LLMs, achieving competitive results without requiring any human-annotated preference data.", "authors": ["Yuting Gao", "Weihao Chen", "Lan Wang", "Ruihan Xu", "Qingpei Guo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-24", "url": "https://arxiv.org/abs/2511.19023", "pdf_url": "https://arxiv.org/pdf/2511.19023v1", "arxiv_id": "2511.19023", "doi": "10.48550/arXiv.2511.19023", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.385}
{"id": "c79dd9a5989a5064692ef6e24ab3ad56433ee2ea497141403a2ad98da5bfe049", "sources": ["arxiv", "semantic_scholar"], "title": "Kitty: Accurate and Efficient 2-bit KV Cache Quantization with Dynamic Channel-wise Precision Boost", "abstract": "The KV cache is a dominant memory bottleneck for LLM inference. While 4-bit KV quantization preserves accuracy, 2-bit often degrades it, especially on long-context reasoning. We close this gap via an algorithm-system co-design for mixed-precision KV caching: Kitty. On the algorithm side, extensive experiments show that Dynamic Channel-wise Precision Boost -- which ranks Key-cache channels by sensitivity and keeps only a small fraction at higher precision -- maintains near-zero loss in accuracy drop while approaching 2-bit memory. The main challenge is handling dynamic 4-bit channel boosts while keeping the page layout coalesced and the dequantization uniform, with no scattered reads or hard-coded masks. Kitty addresses these issues by decompose each mixed-precision Key page into two tensors with unified 2-bit precision. Based on this, Kitty provides a page-centric KV layout, Triton-compatible page dequantization kernels, and a lightweight runtime pipeline that preserves coalescing and avoids divergence. Across seven tasks and two model families (Qwen3, LLaMA3), Kitty cuts KV memory by nearly 8x with negligible accuracy loss, enabling up to 8x larger batches and 2.1x-4.1x higher throughput under the same memory budget. We release the full implementation of Kitty at https://github.com/Summer-Summer/Kitty.", "authors": ["Haojun Xia", "Xiaoxia Wu", "Jisen Li", "Robert Wu", "Junxiong Wang", "Jue Wang", "Chenxi Li", "Aman Singhal", "Alay Dilipbhai Shah", "Alpay Ariyak", "Donglin Zhuang", "Zhongzhu Zhou", "Ben Athiwaratkun", "Zhen Zheng", "Shuaiwen Leon Song"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-23", "url": "https://arxiv.org/abs/2511.18643", "pdf_url": "https://arxiv.org/pdf/2511.18643v1", "arxiv_id": "2511.18643", "doi": "10.48550/arXiv.2511.18643", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Summer-Summer/Kitty", "venue": "arXiv.org", "quality_score": 0.5932}
{"id": "6605765fb75777e450f514bbaaac1e8e2ac9d11837ae11c7aff7fec525455407", "sources": ["arxiv", "semantic_scholar"], "title": "AnyExperts: On-Demand Expert Allocation for Multimodal Language Models with Mixture of Expert", "abstract": "Multimodal Mixture-of-Experts (MoE) models offer a promising path toward scalable and efficient large vision-language systems. However, existing approaches rely on rigid routing strategies (typically activating a fixed number of experts per token) ignoring the inherent heterogeneity in semantic importance across modalities. This leads to suboptimal compute allocation, where redundant tokens consume as many resources as critical ones. To address this, we propose AnyExperts, a novel on-demand, budget-aware dynamic routing framework that allocates a variable total number of expert slots per token based on its semantic importance. Crucially, to prevent uncontrolled compute growth, the total slots per token are constrained within a fixed range, and each slot is filled by either a real expert or a virtual expert, with the virtual share capped at a small maximum (e.g., 20%). The model then adaptively balances the real-to-virtual ratio per token, assigning more real experts to semantically rich regions and relying more on virtual experts for redundant content. Evaluated across diverse tasks in visual understanding, audio understanding, and NLP understanding, AnyExperts improves performance under the same compute budget. Notably, on general image/video tasks, it achieves comparable accuracy with 40% fewer real expert activations; on text-dense tasks (OCR and NLP), it maintains performance while reducing real expert usage by 10%. These results demonstrate that fine-grained, importance-driven expert allocation significantly enhances both the efficiency and effectiveness of multimodal MoE models.", "authors": ["Yuting Gao", "Wang Lan", "Hengyuan Zhao", "Linjiang Huang", "Si Liu", "Qingpei Guo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-23", "url": "https://arxiv.org/abs/2511.18314", "pdf_url": "https://arxiv.org/pdf/2511.18314v1", "arxiv_id": "2511.18314", "doi": "10.48550/arXiv.2511.18314", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3839}
{"id": "836ac94acd8c963c4312b0b1003319a9166f7640bbbee56a0024b9b590ddd6da", "sources": ["arxiv", "semantic_scholar"], "title": "FastMMoE: Accelerating Multimodal Large Language Models through Dynamic Expert Activation and Routing-Aware Token Pruning", "abstract": "Multimodal large language models (MLLMs) have achieved impressive performance, but high-resolution visual inputs result in long sequences of visual tokens and substantial inference latency. Reducing redundant visual tokens is critical to ease computational/memory burdens while preserving performance, enabling MLLM deployment in resource-constrained or latency-sensitive scenarios. Current visual token pruning methods mainly rely on attention-based redundancy analysis and are tailored to dense architectures. We propose Fast Multimodal Mixture-of-Experts (FastMMoE), a training-free acceleration framework for mixture-of-experts (MoE) based MLLMs, developed from a routing analysis perspective. FastMMoE combines two complementary strategies: (i) expert activation reduction for visual tokens to minimize unnecessary expert computation; and (ii) routing-aware token pruning that leverages similarity in routing probability distributions to identify and remove highly redundant visual tokens. Experiments on large-scale MoE-MLLMs such as DeepSeek-VL2 and InternVL3.5 demonstrate that FastMMoE can reduce FLOPs by up to 55.0% while retaining approximately 95.5% of the original performance, consistently outperforming dense-model pruning baselines including FastV and SparseVLM across multiple retention rates.", "authors": ["Guoyang Xia", "Yifeng Ding", "Fengfa Li", "Lei Ren", "Wei Chen", "Fangxiang Feng", "Xiaojie Wang"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-22", "url": "https://arxiv.org/abs/2511.17885", "pdf_url": "https://arxiv.org/pdf/2511.17885v2", "arxiv_id": "2511.17885", "doi": "10.48550/arXiv.2511.17885", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3827}
{"id": "6a258fcdaaaa9aebceb18605f4754bffb6c061c0b583ff44e9ddead2028437fd", "sources": ["arxiv", "semantic_scholar"], "title": "Exploiting the Experts: Unauthorized Compression in MoE-LLMs", "abstract": "Mixture-of-Experts (MoE) architectures are increasingly adopted in large language models (LLMs) for their scalability and efficiency. However, their modular structure introduces a unique vulnerability: adversaries can attempt to compress or repurpose models by pruning experts and cheaply fine-tuning the remainder, effectively bypassing licensing and security constraints. In this paper, we systematically study the prunability of MoE-LLMs under task-specific usage. We first develop an expert attribution framework that identifies the subset of experts most responsible for a given task, then evaluate the performance trade-offs of pruning and re-aligning these experts using active learning-driven fine-tuning. Our findings reveal a critical knowledge loss--recovery trade-off: while certain experts can be isolated to retain task accuracy, significant degradation occurs without targeted re-alignment. Based on this analysis, we propose defense strategies that aim to make MoE models harder to compress and fine-tune without authorization, including entangled expert training and selective fine-tuning protocols that resist unauthorized adaptation. By positioning expert pruning as both a threat vector and a defense target, this work highlights the dual-use nature of MoE modularity and provides the first systematic evaluation framework for secure specialization of MoE-LLMs.", "authors": ["Pinaki Prasad Guha Neogi", "Ahmad Mohammadshirazi", "Dheeraj Kulshrestha", "Rajiv Ramnath"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-22", "url": "https://arxiv.org/abs/2511.19480", "pdf_url": "https://arxiv.org/pdf/2511.19480v1", "arxiv_id": "2511.19480", "doi": "10.48550/arXiv.2511.19480", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3827}
{"id": "94b763374f79cd14a21766b9fa5c622e912417bbd8794fa21fc77b4506aa37f8", "sources": ["arxiv", "semantic_scholar"], "title": "MCMoE: Completing Missing Modalities with Mixture of Experts for Incomplete Multimodal Action Quality Assessment", "abstract": "Multimodal Action Quality Assessment (AQA) has recently emerged as a promising paradigm. By leveraging complementary information across shared contextual cues, it enhances the discriminative evaluation of subtle intra-class variations in highly similar action sequences. However, partial modalities are frequently unavailable at the inference stage in reality. The absence of any modality often renders existing multimodal models inoperable. Furthermore, it triggers catastrophic performance degradation due to interruptions in cross-modal interactions. To address this issue, we propose a novel Missing Completion Framework with Mixture of Experts (MCMoE) that unifies unimodal and joint representation learning in single-stage training. Specifically, we propose an adaptive gated modality generator that dynamically fuses available information to reconstruct missing modalities. We then design modality experts to learn unimodal knowledge and dynamically mix the knowledge of all experts to extract cross-modal joint representations. With a mixture of experts, missing modalities are further refined and complemented. Finally, in the training phase, we mine the complete multimodal features and unimodal expert knowledge to guide modality generation and generation-based joint representation extraction. Extensive experiments demonstrate that our MCMoE achieves state-of-the-art results in both complete and incomplete multimodal learning on three public AQA benchmarks. Code is available at https://github.com/XuHuangbiao/MCMoE.", "authors": ["Huangbiao Xu", "Huanqi Wu", "Xiao Ke", "Junyi Wu", "Rui Xu", "Jinglin Xu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-21", "url": "https://arxiv.org/abs/2511.17397", "pdf_url": "https://arxiv.org/pdf/2511.17397v2", "arxiv_id": "2511.17397", "doi": "10.48550/arXiv.2511.17397", "citation_count": 5, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/XuHuangbiao/MCMoE", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.5897}
{"id": "999fdbacd54528530eda8ff5b8eef9bfadaee9f0c06d98ffd94784adde4a0a8e", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Mixture-of-Experts for Multi-Channel Imaging: Are All Channel Interactions Required?", "abstract": "Vision Transformers ($\\text{ViTs}$) have become the backbone of vision foundation models, yet their optimization for multi-channel domains - such as cell painting or satellite imagery - remains underexplored. A key challenge in these domains is capturing interactions between channels, as each channel carries different information. While existing works have shown efficacy by treating each channel independently during tokenization, this approach naturally introduces a major computational bottleneck in the attention block - channel-wise comparisons leads to a quadratic growth in attention, resulting in excessive $\\text{FLOPs}$ and high training cost. In this work, we shift focus from efficacy to the overlooked efficiency challenge in cross-channel attention and ask: \"Is it necessary to model all channel interactions?\". Inspired by the philosophy of Sparse Mixture-of-Experts ($\\text{MoE}$), we propose MoE-ViT, a Mixture-of-Experts architecture for multi-channel images in $\\text{ViTs}$, which treats each channel as an expert and employs a lightweight router to select only the most relevant experts per patch for attention. Proof-of-concept experiments on real-world datasets - JUMP-CP and So2Sat - demonstrate that $\\text{MoE-ViT}$ achieves substantial efficiency gains without sacrificing, and in some cases enhancing, performance, making it a practical and attractive backbone for multi-channel imaging.", "authors": ["Sukwon Yun", "Heming Yao", "Burkhard Hoeckendorf", "David Richmond", "Aviv Regev", "Russell Littman"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-21", "url": "https://arxiv.org/abs/2511.17400", "pdf_url": "https://arxiv.org/pdf/2511.17400v1", "arxiv_id": "2511.17400", "doi": "10.48550/arXiv.2511.17400", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3816}
{"id": "e95b508c687327f7060fcdf66cbe004812856424ec426d2388101c70ff3ca55f", "sources": ["arxiv", "semantic_scholar"], "title": "Glass Surface Detection: Leveraging Reflection Dynamics in Flash/No-flash Imagery", "abstract": "Glass surfaces are ubiquitous in daily life, typically appearing colorless, transparent, and lacking distinctive features. These characteristics make glass surface detection a challenging computer vision task. Existing glass surface detection methods always rely on boundary cues (e.g., window and door frames) or reflection cues to locate glass surfaces, but they fail to fully exploit the intrinsic properties of the glass itself for accurate localization. We observed that in most real-world scenes, the illumination intensity in front of the glass surface differs from that behind it, which results in variations in the reflections visible on the glass surface. Specifically, when standing on the brighter side of the glass and applying a flash towards the darker side, existing reflections on the glass surface tend to disappear. Conversely, while standing on the darker side and applying a flash towards the brighter side, distinct reflections will appear on the glass surface. Based on this phenomenon, we propose NFGlassNet, a novel method for glass surface detection that leverages the reflection dynamics present in flash/no-flash imagery. Specifically, we propose a Reflection Contrast Mining Module (RCMM) for extracting reflections, and a Reflection Guided Attention Module (RGAM) for fusing features from reflection and glass surface for accurate glass surface detection. For learning our network, we also construct a dataset consisting of 3.3K no-flash and flash image pairs captured from various scenes with corresponding ground truth annotations. Extensive experiments demonstrate that our method outperforms the state-of-the-art methods. Our code, model, and dataset will be available upon acceptance of the manuscript.", "authors": ["Tao Yan", "Hao Huang", "Yiwei Lu", "Zeyu Wang", "Ke Xu", "Yinghui Wang", "Xiaojun Chang", "Rynson W. H. Lau"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-21", "url": "https://arxiv.org/abs/2511.16887", "pdf_url": "https://arxiv.org/pdf/2511.16887v2", "arxiv_id": "2511.16887", "doi": "10.48550/arXiv.2511.16887", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3816}
{"id": "0d5b9f936e8f8a935bf46e3237388f74511c36de5e949cc09e569d4fb254f6e8", "sources": ["arxiv", "semantic_scholar"], "title": "Revisiting Multimodal KV Cache Compression: A Frequency-Domain-Guided Outlier-KV-Aware Approach", "abstract": "Multimodal large language models suffer from substantial inference overhead since multimodal KV Cache grows proportionally with the visual input length. Existing multimodal KV Cache compression methods mostly rely on attention score to reduce cache size, which makes them are incompatible with established efficient attention kernels (e.g., FlashAttention) and ignores the contribution of value vectors to the attention output. In this work, we revisit multimodal KV Cache compression from the perspective of the KV matrices' distribution. First, we observe that frequency-domain energy of multimodal KV matrices is predominantly concentrated in low-frequency and extract this principal energy via a low-pass filter. Further, we find that removing KV pairs that deviate substantially from this principal energy leads to a pronounced performance drop, which we define as Outlier KVs. Considering Outlier KVs are more likely to encode features critical for inference, we propose FlashCache, a frequency-domain-guided, Outlier-KV-aware KV Cache compression framework. First, we introduce an Outlier KV Recognition Module that models the principal component of multimodal KV matrices in the frequency domain and preferentially retains KV pairs that significantly deviate from it. Furthermore, Dynamic Budget Allocation Module is designed to adaptively determine the per-layer KV Cache size to retain more Outlier KVs. Experiments on multiple MLLMs and benchmarks demonstrate that FlashCache outperforms state-of-the-art multimoal KV compression methods, achieving up to 1.69 times faster decoding with 80% lower KV memory usage while maintaining task performance.", "authors": ["Yaoxin Yang", "Peng Ye", "Xudong Tan", "Chongjun Tu", "Maosen Zhao", "Jia Hao", "Tao Chen"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-20", "url": "https://arxiv.org/abs/2511.16786", "pdf_url": "https://arxiv.org/pdf/2511.16786v2", "arxiv_id": "2511.16786", "doi": "10.48550/arXiv.2511.16786", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3804}
{"id": "f4445fe1edbfd3bc79c4e542e04187282da073984032043c5f7199162259fcf8", "sources": ["arxiv", "semantic_scholar"], "title": "AMS-KV: Adaptive KV Caching in Multi-Scale Visual Autoregressive Transformers", "abstract": "Visual autoregressive modeling (VAR) via next-scale prediction has emerged as a scalable image generation paradigm. While Key and Value (KV) caching in large language models (LLMs) has been extensively studied, next-scale prediction presents unique challenges, and KV caching design for next-scale based VAR transformers remains largely unexplored. A major bottleneck is the excessive KV memory growth with the increasing number of scales-severely limiting scalability. Our systematic investigation reveals that: (1) Attending to tokens from local scales significantly contributes to generation quality (2) Allocating a small amount of memory for the coarsest scales, termed as condensed scales, stabilizes multi-scale image generation (3) Strong KV similarity across finer scales is predominantly observed in cache-efficient layers, whereas cache-demanding layers exhibit weaker inter-scale similarity. Based on the observations, we introduce AMS-KV, a scale-adaptive KV caching policy for next-scale prediction in VAR models. AMS-KV prioritizes storing KVs from condensed and local scales, preserving the most relevant tokens to maintain generation quality. It further optimizes KV cache utilization and computational efficiency identifying cache-demanding layers through inter-scale similarity analysis. Compared to the vanilla next-scale prediction-based VAR models, AMS-KV reduces KV cache usage by up to 84.83% and self-attention latency by 60.48%. Moreover, when the baseline VAR-d30 model encounters out-of-memory failures at a batch size of 128, AMS-KV enables stable scaling to a batch size of 256 with improved throughput.", "authors": ["Boxun Xu", "Yu Wang", "Zihu Wang", "Peng Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-20", "url": "https://arxiv.org/abs/2511.16047", "pdf_url": "https://arxiv.org/pdf/2511.16047v1", "arxiv_id": "2511.16047", "doi": "10.48550/arXiv.2511.16047", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3804}
{"id": "aa5753d37c114363b00f462e91a86f420d5a8c43cd5a751b52d7f6b732d62e54", "sources": ["arxiv", "semantic_scholar"], "title": "On 10x Better Scalability: KV Stores Scale Up KV Cache", "abstract": "Large language models (LLMs) rely on Key-Value (KV) cache to reduce time-to-first-token (TTFT) latency, but existing disk-based KV cache systems using file-per-object layouts suffer from severe scalability bottlenecks due to file system metadata overhead, I/O inefficiency, and poor spatial locality. This paper presents SGLANG-LSM, a database-inspired system that leverages Log-Structured Merge-tree (LSM-tree) architectures for scalable KV cache management. SGLANG-LSM implements a layered system design with three coordinated components: (1) a prefix-preserving storage engine that maintains token sequence locality while efficiently storing large KV cache tensors through key-value separation, (2) an adaptive controller that dynamically optimizes LSM-tree configurations based on shifting workload characteristics, and (3) runtime services including batch operations and automatic resource management for production deployment. Evaluation on large-scale dynamic workloads demonstrates that SGLANG-LSM significantly improves cache hits by up to 143% and reduces TTFT by up to 24% compared to state-of-the-art systems, representing the first systematic application of database storage architectures to large-scale LLM cache management.", "authors": ["Weiping Yu", "Ye Jiarui", "He Mengke", "Junfeng Liu", "Siqiang Luo"], "categories": ["cs.DB"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-20", "url": "https://arxiv.org/abs/2511.16138", "pdf_url": "https://arxiv.org/pdf/2511.16138v2", "arxiv_id": "2511.16138", "doi": "10.48550/arXiv.2511.16138", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3804}
{"id": "099ecc68acddd44a69c80f723c4998e52e73c574ec28e8410fb24567aedf51dd", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Expert Quantization for Scalable Mixture-of-Experts Inference", "abstract": "Mixture-of-Experts (MoE) has become a practical architecture for scaling LLM capacity while keeping per-token compute modest, but deploying MoE models on a single, memory-limited GPU remains difficult because expert weights dominate the HBM footprint. Existing expert offloading and prefetching systems reduce the resident set, yet they often pay expert-loading costs on the critical path when activation becomes dense. Post-training quantization (PTQ) lowers the footprint without transfers, but prevailing pipelines fix expert bit-widths offline and assume routing remains stable, even though MoE expert utilization is heavy-tailed and the hot set can shift across workloads. We present DynaExq, a runtime-aware mixed-precision serving system that treats single-GPU MoE inference under a hard HBM envelope as an online, budget-constrained precision allocation problem. The key insight is to keep the experts that dominate runtime traffic resident at higher precision, while maintaining a low-precision fallback for the remaining experts, so the system can reduce transfer volume and avoid the waiting latency that limits offloading and prefetching under dense activation. DynaExq estimates long-horizon expert hotness from router traces, selects a per-layer high-precision resident set via a budget-feasible top-$n$ rule, and applies promotions and demotions asynchronously through stable expert handles so the forward pass always executes on a fully materialized expert version. Across Qwen3-MoE-30B/80B and six benchmarks, DynaExq improves accuracy over static PTQ on Qwen3-80B (73.09% to 77.57%) under comparable device-memory budgets and achieves up to 2.73x higher throughput than offloading/prefetch baselines at batch size 32.", "authors": ["Kexin Chu", "Dawei Xiang", "Zixu Shen", "Yiwei Yang", "Zecheng Liu", "Wei Zhang"], "categories": ["cs.PF", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-19", "url": "https://arxiv.org/abs/2511.15015", "pdf_url": "https://arxiv.org/pdf/2511.15015v3", "arxiv_id": "2511.15015", "doi": "10.48550/arXiv.2511.15015", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3793}
{"id": "320006d041411aa514800ba7fb4a0074161d41dd417a46dc723e28a959295691", "sources": ["arxiv", "semantic_scholar"], "title": "DiffuMamba: High-Throughput Diffusion LMs with Mamba Backbone", "abstract": "Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) generation, yet their reliance on Transformer backbones limits inference efficiency due to quadratic attention or KV-cache overhead. We introduce DiffuMamba, a masked diffusion language model built on a bidirectional Mamba backbone that combines the diffusion objective with linear-time sequence modeling, and DiffuMamba-H, a hybrid variant with interleaved attention. Across scales up to 1.3B parameters, our models match Transformer-based diffusion in downstream performance while achieving up to 8.2x and 4.3x higher inference throughput, respectively, on long sequences. We further present a systematic analysis of inference efficiency across modern DLM variants combining asymptotic complexity with empirical measurements. Notably, cache-efficient block diffusion with Mamba mixers emerges as the only strategy that scales linearly with sequence length and achieves the strongest performance across all baselines, suggesting a promising direction for future diffusion-based generation systems.", "authors": ["Vaibhav Singh", "Oleksiy Ostapenko", "Pierre-André Noël", "Eugene Belilovsky", "Torsten Scholak"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-19", "url": "https://arxiv.org/abs/2511.15927", "pdf_url": "https://arxiv.org/pdf/2511.15927v3", "arxiv_id": "2511.15927", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2414}
{"id": "8f7d7a3eb9ea2ac0ab4a889de6a845e9bafcca0b6a87419fa848d760a972caed", "sources": ["arxiv", "semantic_scholar"], "title": "discretize_distributions: Efficient Quantization of Gaussian Mixtures with Guarantees in Wasserstein Distance", "abstract": "We present discretize_distributions, a Python package that efficiently constructs discrete approximations of Gaussian mixture distributions and provides guarantees on the approximation error in Wasserstein distance. The package implements state-of-the-art quantization methods for Gaussian mixture models and extends them to improve scalability. It further integrates complementary quantization strategies such as sigma-point methods and provides a modular interface that supports custom schemes and integration into control and verification pipelines for cyber-physical systems. We benchmark the package on various examples, including high-dimensional, large, and degenerate Gaussian mixtures, and demonstrate that discretize_distributions produces accurate approximations at low computational cost.", "authors": ["Steven Adams", "Elize Alwash", "Luca Laurenti"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-19", "url": "https://arxiv.org/abs/2511.15854", "pdf_url": "https://arxiv.org/pdf/2511.15854v1", "arxiv_id": "2511.15854", "doi": "10.48550/arXiv.2511.15854", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3793}
{"id": "b016cbe4c7f4b8c46dc6908507c39743e1ea5bce46cd4eef5c2be4f54c174360", "sources": ["arxiv", "semantic_scholar"], "title": "MoDES: Accelerating Mixture-of-Experts Multimodal Large Language Models via Dynamic Expert Skipping", "abstract": "Mixture-of-Experts (MoE) Multimodal large language models (MLLMs) excel at vision-language tasks, but they suffer from high computational inefficiency. To reduce inference overhead, expert skipping methods have been proposed to deactivate redundant experts based on the current input tokens. However, we find that applying these methods-originally designed for unimodal large language models (LLMs)-to MLLMs results in considerable performance degradation. This is primarily because such methods fail to account for the heterogeneous contributions of experts across MoE layers and modality-specific behaviors of tokens within these layers. Motivated by these findings, we propose MoDES, the first training-free framework that adaptively skips experts to enable efficient and accurate MoE MLLM inference. It incorporates a globally-modulated local gating (GMLG) mechanism that integrates global layer-wise importance into local routing probabilities to accurately estimate per-token expert importance. A dual-modality thresholding (DMT) method is then applied, which processes tokens from each modality separately, to derive the skipping schedule. To set the optimal thresholds, we introduce a frontier search algorithm that exploits monotonicity properties, cutting convergence time from several days to a few hours. Extensive experiments for 3 model series across 13 benchmarks demonstrate that MoDES far outperforms previous approaches. For instance, when skipping 88% experts for Qwen3-VL-MoE-30B-A3B-Instruct, the performance boost is up to 10.67% (97.33% vs. 86.66%). Furthermore, MoDES significantly enhances inference speed, improving the prefilling time by 2.16$\\times$ and the decoding time by 1.26$\\times$. Our code is available at https://github.com/ModelTC/MoDES.", "authors": ["Yushi Huang", "Zining Wang", "Zhihang Yuan", "Yifu Ding", "Ruihao Gong", "Jinyang Guo", "Xianglong Liu", "Jun Zhang"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-19", "url": "https://arxiv.org/abs/2511.15690", "pdf_url": "https://arxiv.org/pdf/2511.15690v2", "arxiv_id": "2511.15690", "doi": "10.48550/arXiv.2511.15690", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ModelTC/MoDES", "venue": "arXiv.org", "quality_score": 0.5861}
{"id": "16420223fb875466df3924b1d6c115ca1527bb818e901a088e7fc754977eefaf", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-SpeQ: Speculative Quantized Decoding with Proactive Expert Prefetching and Offloading for Mixture-of-Experts", "abstract": "The immense memory requirements of state-of-the-art Mixture-of-Experts (MoE) models present a significant challenge for inference, often exceeding the capacity of a single accelerator. While offloading experts to host memory is a common solution, it introduces a severe I/O bottleneck over the PCIe bus, as the data-dependent nature of expert selection places these synchronous transfers directly on the critical path of execution, crippling performance. This paper argues that the I/O bottleneck can be overcome by trading a small amount of cheap, on-device computation to hide the immense cost of data movement. We present MoE-SpeQ, a new inference system built on a novel co-design of speculative execution and expert offloading. MoE-SpeQ employs a small, on-device draft model to predict the sequence of required experts for future tokens. This foresight enables a runtime orchestrator to prefetch these experts from host memory, effectively overlapping the expensive I/O with useful computation and hiding the latency from the critical path. To maximize performance, an adaptive governor, guided by an Amortization Roofline Model, dynamically tunes the speculation strategy to the underlying hardware. Our evaluation on memory-constrained devices shows that for the Phi-MoE model, MoE-SpeQ achieves at most 2.34x speedup over the state-of-the-art offloading framework. Our work establishes a new, principled approach for managing data-dependent memory access in resource-limited environments, making MoE inference more accessible on commodity hardware.", "authors": ["Wenfeng Wang", "Jiacheng Liu", "Xiaofeng Hou", "Xinfeng Xia", "Peng Tang", "Mingxuan Zhang", "Chao Li", "Minyi Guo"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-18", "url": "https://arxiv.org/abs/2511.14102", "pdf_url": "https://arxiv.org/pdf/2511.14102v1", "arxiv_id": "2511.14102", "doi": "10.48550/arXiv.2511.14102", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3781}
{"id": "25ed0a77af424617aa242c7c923b6e5c9ef0aef8aa6e58ca282073d0048701e3", "sources": ["arxiv", "semantic_scholar"], "title": "MoMoE: A Mixture of Expert Agent Model for Financial Sentiment Analysis", "abstract": "We present a novel approach called Mixture of Mixture of Expert (MoMoE) that combines the strengths of Mixture-of-Experts (MoE) architectures with collaborative multi-agent frameworks. By modifying the LLaMA 3.1 8B architecture to incorporate MoE layers in each agent of a layered collaborative structure, we create an ensemble of specialized expert agents that iteratively refine their outputs. Each agent leverages an MoE layer in its final attention block, enabling efficient task decomposition while maintaining computational feasibility. This hybrid approach creates specialized pathways through both the model architecture and the agent collaboration layers. Experimental results demonstrate significant improvements across multiple language understanding and generation benchmarks, highlighting the synergistic benefits of combining expert routing at both the neural and agent levels.", "authors": ["Peng Shu", "Junhao Chen", "Zhengliang Liu", "Hanqi Jiang", "Yi Pan", "Khanh Nhu Nguyen", "Zihao Wu", "Huaqin Zhao", "Yiwei Li", "Enze Shi", "ShaoChen Xu"], "categories": ["cs.CE"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-17", "url": "https://arxiv.org/abs/2511.13983", "pdf_url": "https://arxiv.org/pdf/2511.13983v1", "arxiv_id": "2511.13983", "doi": "10.48550/arXiv.2511.13983", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.377}
{"id": "fa3499c61e180b0cfbf393d18ccbda6cafb3678af006963e24d05b888da91ffc", "sources": ["arxiv", "semantic_scholar"], "title": "YOLO Meets Mixture-of-Experts: Adaptive Expert Routing for Robust Object Detection", "abstract": "This paper presents a novel Mixture-of-Experts framework for object detection, incorporating adaptive routing among multiple YOLOv9-T experts to enable dynamic feature specialization and achieve higher mean Average Precision (mAP) and Average Recall (AR) compared to a single YOLOv9-T model.", "authors": ["Ori Meiraz", "Sharon Shalev", "Avishai Weizman"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-17", "url": "https://arxiv.org/abs/2511.13344", "pdf_url": "https://arxiv.org/pdf/2511.13344v4", "arxiv_id": "2511.13344", "doi": "10.48550/arXiv.2511.13344", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.377}
{"id": "40fb5ba5f89936b0625d3c9d354d55aa5d000d817ee1c2b5dbd94f29781134e7", "sources": ["arxiv", "semantic_scholar"], "title": "Whose Narrative is it Anyway? A KV Cache Manipulation Attack", "abstract": "The Key Value(KV) cache is an important component for efficient inference in autoregressive Large Language Models (LLMs), but its role as a representation of the model's internal state makes it a potential target for integrity attacks. This paper introduces \"History Swapping,\" a novel block-level attack that manipulates the KV cache to steer model generation without altering the user-facing prompt. The attack involves overwriting a contiguous segment of the active generation's cache with a precomputed cache from a different topic. We empirically evaluate this method across 324 configurations on the Qwen 3 family of models, analyzing the impact of timing, magnitude, and layer depth of the cache overwrite. Our findings reveal that only full-layer overwrites can successfully hijack the conversation's topic, leading to three distinct behaviors: immediate and persistent topic shift, partial recovery, or a delayed hijack. Furthermore, we observe that high-level structural plans are encoded early in the generation process and local discourse structure is maintained by the final layers of the model. This work demonstrates that the KV cache is a significant vector for security analysis, as it encodes not just context but also topic trajectory and structural planning, making it a powerful interface for manipulating model behavior.", "authors": ["Mukkesh Ganesh", "Kaushik Iyer", "Arun Baalaaji Sankar Ananthan"], "categories": ["cs.CR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-16", "url": "https://arxiv.org/abs/2511.12752", "pdf_url": "https://arxiv.org/pdf/2511.12752v1", "arxiv_id": "2511.12752", "doi": "10.48550/arXiv.2511.12752", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3758}
{"id": "8742ba2ac0e8734eec0def96b5366684229cea06e9f947a440c8c089c8f02765", "sources": ["arxiv", "semantic_scholar"], "title": "AMR-MoEGA: Antimicrobial Resistance Prediction using Mixture of Experts and Genetic Algorithms", "abstract": "Antimicrobial resistance (AMR) poses a mounting global health crisis, requiring rapid and reliable prediction frameworks that capture its complex evolutionary dynamics. Traditional antimicrobial susceptibility testing (AST), while accurate, remains laborious and time-consuming, limiting its clinical scalability. Existing computational approaches, primarily reliant on single nucleotide polymorphism (SNP)-based analysis, fail to account for evolutionary drivers such as horizontal gene transfer (HGT) and genome-level interactions. This study introduces a novel Evolutionary Mixture of Experts (Evo-MoE) framework that integrates genomic sequence analysis, machine learning, and evolutionary algorithms to model and predict AMR evolution. A Mixture of Experts model, trained on labeled genomic data for multiple antibiotics, serves as the predictive core, estimating the likelihood of resistance for each genome. This model is embedded as a fitness function within a Genetic Algorithm designed to simulate AMR development across generations. Each genome, encoded as an individual in the population, undergoes mutation, crossover, and selection guided by predicted resistance probabilities. The resulting evolutionary trajectories reveal dynamic pathways of resistance acquisition, offering mechanistic insights into genomic evolution under selective antibiotic pressure. Sensitivity analysis of mutation rates and selection pressures demonstrates the model's robustness and biological plausibility. Validation against curated AMR databases and literature evidence further substantiates the framework's predictive fidelity. This integrative approach bridges genomic prediction and evolutionary simulation, offering a powerful tool for understanding and anticipating AMR dynamics, and potentially guiding rational antibiotic design and policy interventions.", "authors": ["Anshul Bagaria"], "categories": ["q-bio.PE"], "fields_of_study": ["Biology"], "published_date": "2025-11-15", "url": "https://arxiv.org/abs/2511.12223", "pdf_url": "https://arxiv.org/pdf/2511.12223v1", "arxiv_id": "2511.12223", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2384}
{"id": "772620ba4f44a797ed18b61c785f8f01c1b1024767819a37fbd2f928fa4b0ecf", "sources": ["arxiv", "semantic_scholar"], "title": "KVSwap: Disk-aware KV Cache Offloading for Long-Context On-device Inference", "abstract": "Language models (LMs) underpin emerging mobile and embedded AI applications like meeting and video summarization and document analysis, which often require processing multiple long-context inputs. Running an LM locally on-device improves privacy, enables offline use, and reduces cost, but long-context inference quickly hits a \\emph{memory capacity wall} as the key-value (KV) cache grows linearly with context length and batch size. Existing KV-cache offloading schemes are designed to transfer cache data from GPU memory to CPU memory; however, they are not suitable for embedded and mobile systems, where the CPU and GPU (or NPU) typically share a unified memory and the non-volatile secondary storage (disk) offers limited I/O bandwidth. We present KVSwap, a software framework tailored for local devices that achieves high memory efficiency while effectively leveraging disk storage. KVSwap stores the full cache on disk, uses highly compact in-memory metadata to predict which entries to preload, overlaps computation with hardware-aware disk access, and orchestrates read patterns to match storage device characteristics. Our evaluation shows that across representative LMs and storage types, KVSwap delivers higher throughput under tight memory budgets while maintaining generation quality over existing KV cache offloading schemes.", "authors": ["Huawei Zhang", "Chunwei Xia", "Zheng Wang"], "categories": ["cs.DC", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.11907", "pdf_url": "https://arxiv.org/pdf/2511.11907v2", "arxiv_id": "2511.11907", "doi": "10.48550/arXiv.2511.11907", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2377}
{"id": "0aa714316104206161519bb1a0c1d2a1736c992e6ca4615122a80a32f6f66422", "sources": ["arxiv", "semantic_scholar"], "title": "AccKV: Towards Efficient Audio-Video LLMs Inference via Adaptive-Focusing and Cross-Calibration KV Cache Optimization", "abstract": "Recent advancements in Audio-Video Large Language Models (AV-LLMs) have enhanced their capabilities in tasks like audio-visual question answering and multimodal dialog systems. Video and audio introduce an extended temporal dimension, resulting in a larger key-value (KV) cache compared to static image embedding. A naive optimization strategy is to selectively focus on and retain KV caches of audio or video based on task. However, in the experiment, we observed that the attention of AV-LLMs to various modalities in the high layers is not strictly dependent on the task. In higher layers, the attention of AV-LLMs shifts more towards the video modality. In addition, we also found that directly integrating temporal KV of audio and spatial-temporal KV of video may lead to information confusion and significant performance degradation of AV-LLMs. If audio and video are processed indiscriminately, it may also lead to excessive compression or reservation of a certain modality, thereby disrupting the alignment between modalities. To address these challenges, we propose AccKV, an Adaptive-Focusing and Cross-Calibration KV cache optimization framework designed specifically for efficient AV-LLMs inference. Our method is based on layer adaptive focusing technology, selectively focusing on key modalities according to the characteristics of different layers, and enhances the recognition of heavy hitter tokens through attention redistribution. In addition, we propose a Cross-Calibration technique that first integrates inefficient KV caches within the audio and video modalities, and then aligns low-priority modalities with high-priority modalities to selectively evict KV cache of low-priority modalities. The experimental results show that AccKV can significantly improve the computational efficiency of AV-LLMs while maintaining accuracy.", "authors": ["Zhonghua Jiang", "Kui Chen", "Kunxi Li", "Keting Yin", "Yiyun Zhou", "Zhaode Wang", "Chengfei Lv", "Shengyu Zhang"], "categories": ["cs.MM", "cs.CV", "cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.11106", "pdf_url": "https://arxiv.org/pdf/2511.11106v1", "arxiv_id": "2511.11106", "doi": "10.48550/arXiv.2511.11106", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2377}
{"id": "b8d49c42d3f205519e8ad1f5f8d182e710e675fa5bc842a87ab7113ef1ed390a", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Efficient Mixture-of-Experts for Remote Sensing Modality-Missing Classification", "abstract": "Multimodal remote sensing classification often suffers from missing modalities caused by sensor failures and environmental interference, leading to severe performance degradation. In this work, we rethink missing-modality learning from a conditional computation perspective and investigate whether Mixture-of-Experts (MoE) models can inherently adapt to diverse modality-missing scenarios. We first conduct a systematic study of representative MoE paradigms under various missing-modality settings, revealing both their potential and limitations. Building on these insights, we propose a Missing-aware Mixture-of-LoRAs (MaMOL), a parameter-efficient MoE framework that unifies multiple modality-missing cases within a single model. MaMOL introduces a dual-routing mechanism to decouple modality-invariant shared experts and modality-aware dynamic experts, enabling automatic expert activation conditioned on available modalities. Extensive experiments on multiple remote sensing benchmarks demonstrate that MaMOL significantly improves robustness and generalization under diverse missing-modality scenarios with minimal computational overhead. Transfer experiments on natural image datasets further validate its scalability and cross-domain applicability.", "authors": ["Qinghao Gao", "Jiahui Qu", "Wenqian Dong"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.11460", "pdf_url": "https://arxiv.org/pdf/2511.11460v3", "arxiv_id": "2511.11460", "doi": "10.48550/arXiv.2511.11460", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3735}
{"id": "be8513770f14f02cfcde3f88e75dd97f5c78c8ae0dbbdd2400b880b88aa3963a", "sources": ["arxiv", "semantic_scholar"], "title": "DoReMi: Bridging 3D Domains via Topology-Aware Domain-Representation Mixture of Experts", "abstract": "Constructing a unified 3D scene understanding model has long been hindered by the significant topological discrepancies across different sensor modalities. While applying the Mixture-of-Experts (MoE) architecture is an effective approach to achieving universal understanding, we observe that existing 3D MoE networks often suffer from semantics-driven routing bias. This makes it challenging to address cross-domain data characterized by \"semantic consistency yet topological heterogeneity.\" To overcome this challenge, we propose DoReMi (Topology-Aware Domain-Representation Mixture of Experts). Specifically, we introduce a self-supervised pre-training branch based on multi attributes, such as topological and texture variations, to anchor cross-domain structural priors. Building upon this, we design a domain-aware expert branch comprising two core mechanisms: Domain Spatial-Guided Routing (DSR), which achieves an acute perception of local topological variations by extracting spatial contexts, and Entropy-controlled Dynamic Allocation (EDA), which dynamically adjusts the number of activated experts by quantifying routing uncertainty to ensure training stability. Through the synergy of these dual branches, DoReMi achieves a deep integration of universal feature extraction and highly adaptive expert allocation. Extensive experiments across various tasks, encompassing both indoor and outdoor scenes, validate the superiority of DoReMi. It achieves 80.1% mIoU on the ScanNet validation set and 77.2% mIoU on S3DIS, comprehensively outperforming existing state-of-the-art methods. The code will be released soon.", "authors": ["Mingwei Xing", "Xinliang Wang", "Yifeng Shi"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.11232", "pdf_url": "https://arxiv.org/pdf/2511.11232v2", "arxiv_id": "2511.11232", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2377}
{"id": "2d6a47ed1eecc3b71b199053f41327f611b9a275cd5acf95f6a15a9c2d781cdc", "sources": ["arxiv", "semantic_scholar"], "title": "NP-LoRA: Null Space Projection for Subject-Style LoRA Fusion", "abstract": "Low-Rank Adaptation (LoRA) fusion enables the composition of subject and style representations for controllable generation without retraining. However, existing approaches primarily operate through weight-level merging, without explicitly modeling how independently trained LoRAs interact in the shared parameter space. We adopt a geometric perspective on LoRA fusion, interpreting content and style LoRAs as occupying overlapping, non-orthogonal low-rank subspaces, where such overlap can lead to conflicting parameter updates that affect generation quality. This observation motivates us to reformulate LoRA fusion not merely as parameter combination, but as a problem of controlling how updates from overlapping subspaces are combined. Based on this insight, we propose Null Space Projection LoRA (NP-LoRA), a training-free framework that employs projection as a fusion operator to explicitly modulate cross-LoRA interactions. Specifically, NP-LoRA uses principal directions of the style LoRA to define a projection subspace and projects the content LoRA onto the complementary subspace (i.e., the null space of the style LoRA), suppressing interference along dominant style directions while preserving complementary information. To avoid the overly aggressive suppression of hard projection, we further formulate soft projection as a regularized optimization problem that balances content preservation against style-subspace suppression. This objective admits a closed-form solution, yielding a projection operator controlled by a single parameter that continuously interpolates between linear merging and hard projection. Extensive experiments across multiple pretrained LoRA pairs show that NP-LoRA achieves more balanced content-style composition compared to strong baselines, without requiring retraining.", "authors": ["Chuheng Chen", "Xiaofei Zhou", "Geyuan Zhang", "Yong Huang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.11051", "pdf_url": "https://arxiv.org/pdf/2511.11051v3", "arxiv_id": "2511.11051", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2377}
{"id": "53dde8638b756c7c38420f25f963bd355e61aedade31bc5c4fc56c689cd03880", "sources": ["arxiv", "semantic_scholar"], "title": "Optimizing Mixture of Block Attention", "abstract": "Mixture of Block Attention (MoBA) (Lu et al., 2025) is a promising building block for efficiently processing long contexts in LLMs by enabling queries to sparsely attend to a small subset of key-value blocks, drastically reducing computational cost. However, the design principles governing MoBA's performance are poorly understood, and it lacks an efficient GPU implementation, hindering its practical adoption. In this paper, we first develop a statistical model to analyze MoBA's underlying mechanics. Our model reveals that performance critically depends on the router's ability to accurately distinguish relevant from irrelevant blocks based on query-key affinities. We derive a signal-to-noise ratio that formally connects architectural parameters to this retrieval accuracy. Guided by our analysis, we identify two key pathways for improvement: using smaller block sizes and applying a short convolution on keys to cluster relevant signals, which enhances routing accuracy. While theoretically better, small block sizes are inefficient on GPUs. To bridge this gap, we introduce FlashMoBA, a hardware-aware CUDA kernel that enables efficient MoBA execution even with the small block sizes our theory recommends. We validate our insights by training LLMs from scratch, showing that our improved MoBA models match the performance of dense attention baselines. FlashMoBA achieves up to 14.7x speedup over FlashAttention-2 for small blocks, making our theoretically-grounded improvements practical. Code is available at: https://github.com/mit-han-lab/flash-moba.", "authors": ["Guangxuan Xiao", "Junxian Guo", "Kasra Mazaheri", "Song Han"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.11571", "pdf_url": "https://arxiv.org/pdf/2511.11571v2", "arxiv_id": "2511.11571", "doi": "10.48550/arXiv.2511.11571", "citation_count": 5, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/mit-han-lab/flash-moba", "venue": "arXiv.org", "quality_score": 0.5773}
{"id": "11cfa4713ba423cd55cc98969a4b1cf09a5818786e9ca69f98dbab39ff7c419e", "sources": ["arxiv", "semantic_scholar"], "title": "Uncertainty Makes It Stable: Curiosity-Driven Quantized Mixture-of-Experts", "abstract": "Deploying deep neural networks on resource-constrained devices faces two critical challenges: maintaining accuracy under aggressive quantization while ensuring predictable inference latency. We present a curiosity-driven quantized Mixture-of-Experts framework that addresses both through Bayesian epistemic uncertainty-based routing across heterogeneous experts (BitNet ternary, 1-16 bit BitLinear, post-training quantization). Evaluated on audio classification benchmarks (ESC-50, Quinn, UrbanSound8K), our 4-bit quantization maintains 99.9 percent of full-precision F1 (0.858 vs 0.859) with 4x compression and 31 percent energy savings versus 8-bit, while both achieve statistical parity with full precision (p > 0.05). Crucially, curiosity-driven routing simultaneously improves accuracy and stability: on Quinn, F1 increases from 0.802 to 0.809 while cross-fold variance drops by 85 percent (p < 0.001, Levene's test), with reductions of 50 to 94 percent across datasets. The routing is self-organizing, with the high-precision 8-bit expert automatically receiving the most uncertain samples (20 percent lower confidence, p < 0.001), while lightweight experts handle easier inputs. Datasets with already low baseline variance show no artificial stability gain, confirming the mechanism targets genuine epistemic uncertainty rather than overfitting routing decisions. At 1.2M parameters, the framework provides interpretable, precision-aware routing suitable for safety-sensitive edge deployments where both accuracy and predictability are critical.", "authors": ["Sebastián Andrés Cajas Ordóñez", "Luis Fernando Torres Torres", "Mackenzie J. Meni", "Carlos Andrés Duran Paredes", "Eric Arazo", "Cristian Bosch", "Ricardo Simon Carbajo", "Yuan Lai", "Leo Anthony Celi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-13", "url": "https://arxiv.org/abs/2511.11743", "pdf_url": "https://arxiv.org/pdf/2511.11743v3", "arxiv_id": "2511.11743", "doi": "10.48550/arXiv.2511.11743", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3724}
{"id": "cae793b6fa8c2b5569adbad28950a72a442cfabd532d0c26e7a5f3c8b4f44f0d", "sources": ["arxiv", "semantic_scholar"], "title": "$A^3$: Attention-Aware Accurate KV Cache Fusion for Fast Large Language Model Serving", "abstract": "Large language models (LLMs) have demonstrated strong capabilities in processing long contexts, enabling them to tackle tasks involving long textual inputs such as multi-turn conversations, legal documents, or retrieved documents in Retrieval-Augmented Generation (RAG) systems. However, despite their ability to handle long sequences, the resulting decoding latency and memory overhead remain substantial, posing challenges for real-world deployment. Recent advances in KV Cache reuse have shown potential to mitigate these costs, but still suffer from notable performance degradation. To address this issue, we conduct an in-depth investigation of recomputation-based reuse methods and observe that the recomputed tokens often fail to align with the context segments most relevant to the question. This misalignment hinders proper updates to the critical contextual representations. Therefore, we propose the $\\textbf{A}$ttention-$\\textbf{A}$ware $\\textbf{A}$ccurate KV Cache Fusion algorithm ($A^3$), which precomputes and selectively fuses the KV Cache of text chunks based on their relevance to the question, achieving accurate integration with minimal computational overhead. Extensive experiments on various benchmarks and LLMs demonstrate that $A^3$ achieves the best task performance compared to four baselines while reducing the time-to-first-token (TTFT) by 2$\\times$.", "authors": ["Yuechi Zhou", "Yi Su", "Jianxin Zhang", "Juntao Li", "Qingrong Xia", "Zhefeng Wang", "Xinyu Duan", "Baoxing Huai"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-13", "url": "https://arxiv.org/abs/2511.17560", "pdf_url": "https://arxiv.org/pdf/2511.17560v1", "arxiv_id": "2511.17560", "doi": "10.48550/arXiv.2511.17560", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3724}
{"id": "b60454f5bd4d66b85ccd1d94fa414627df375c4deeeb7ed52e1589a944f8b3f5", "sources": ["arxiv", "semantic_scholar"], "title": "ExpertAD: Enhancing Autonomous Driving Systems with Mixture of Experts", "abstract": "Recent advancements in end-to-end autonomous driving systems (ADSs) underscore their potential for perception and planning capabilities. However, challenges remain. Complex driving scenarios contain rich semantic information, yet ambiguous or noisy semantics can compromise decision reliability, while interference between multiple driving tasks may hinder optimal planning. Furthermore, prolonged inference latency slows decision-making, increasing the risk of unsafe driving behaviors. To address these challenges, we propose ExpertAD, a novel framework that enhances the performance of ADS with Mixture of Experts (MoE) architecture. We introduce a Perception Adapter (PA) to amplify task-critical features, ensuring contextually relevant scene understanding, and a Mixture of Sparse Experts (MoSE) to minimize task interference during prediction, allowing for effective and efficient planning. Our experiments show that ExpertAD reduces average collision rates by up to 20% and inference latency by 25% compared to prior methods. We further evaluate its multi-skill planning capabilities in rare scenarios (e.g., accidents, yielding to emergency vehicles) and demonstrate strong generalization to unseen urban environments. Additionally, we present a case study that illustrates its decision-making process in complex driving scenarios.", "authors": ["Haowen Jiang", "Xinyu Huang", "You Lu", "Dingji Wang", "Yuheng Cao", "Chaofeng Sha", "Bihuan Chen", "Keyu Chen", "Xin Peng"], "categories": ["cs.RO", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-13", "url": "https://arxiv.org/abs/2511.11740", "pdf_url": "https://arxiv.org/pdf/2511.11740v1", "arxiv_id": "2511.11740", "doi": "10.48550/arXiv.2511.11740", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.237}
{"id": "ad26914fe8d89446d641be6ee047b53803a0fa653b9999fdab22c2361ccaec83", "sources": ["arxiv", "semantic_scholar"], "title": "BuddyMoE: Exploiting Expert Redundancy to Accelerate Memory-Constrained Mixture-of-Experts Inference", "abstract": "Mixture-of-Experts (MoE) architectures scale language models by activating only a subset of specialized expert networks for each input token, thereby reducing the number of floating-point operations. However, the growing size of modern MoE models causes their full parameter sets to exceed GPU memory capacity; for example, Mixtral-8x7B has 45 billion parameters and requires 87 GB of memory even though only 14 billion parameters are used per token. Existing systems alleviate this limitation by offloading inactive experts to CPU memory, but transferring experts across the PCIe interconnect incurs significant latency (about 10 ms). Prefetching heuristics aim to hide this latency by predicting which experts are needed, but prefetch failures introduce significant stalls and amplify inference latency. In the event of a prefetch failure, prior work offers two primary solutions: either fetch the expert on demand, which incurs a long stall due to the PCIe bottleneck, or drop the expert from the computation, which significantly degrades model accuracy. The critical challenge, therefore, is to maintain both high inference speed and model accuracy when prefetching fails.", "authors": ["Yun Wang", "Lingyun Yang", "Senhao Yu", "Yixiao Wang", "Ruixing Li", "Zhixiang Wei", "James Yen", "Zhengwei Qi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-13", "url": "https://arxiv.org/abs/2511.10054", "pdf_url": "https://arxiv.org/pdf/2511.10054v1", "arxiv_id": "2511.10054", "doi": "10.48550/arXiv.2511.10054", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3724}
{"id": "b823686561f844afc54da41cefdb7fe3a0b36c83a8cec98df5623a441d7ea766", "sources": ["arxiv", "semantic_scholar"], "title": "Bayesian Mixture of Experts For Large Language Models", "abstract": "We present Bayesian Mixture of Experts (Bayesian-MoE), a post-hoc uncertainty estimation framework for fine-tuned large language models (LLMs) based on Mixture-of-Experts architectures. Our method applies a structured Laplace approximation to the second linear layer of each expert, enabling calibrated uncertainty estimation without modifying the original training procedure or introducing new parameters. Unlike prior approaches, which apply Bayesian inference to added adapter modules, Bayesian-MoE directly targets the expert pathways already present in MoE models, leveraging their modular design for tractable block-wise posterior estimation. We use Kronecker-factored low-rank approximations to model curvature and derive scalable estimates of predictive uncertainty and marginal likelihood. Experiments on common-sense reasoning benchmarks with Qwen1.5-MoE and DeepSeek-MoE demonstrate that Bayesian-MoE improves both expected calibration error (ECE) and negative log-likelihood (NLL) over baselines, confirming its effectiveness for reliable downstream decision-making.", "authors": ["Maryam Dialameh", "Hossein Rajabzadeh", "Weiwei Zhang", "Walid Ahmed", "Hyock Ju Kwon"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-12", "url": "https://arxiv.org/abs/2511.08968", "pdf_url": "https://arxiv.org/pdf/2511.08968v1", "arxiv_id": "2511.08968", "doi": "10.48550/arXiv.2511.08968", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3713}
{"id": "14ff08dceb360fea2d99af8aa5c1b61fe6e27848acc6d4a7db0b0b4b6546a8e3", "sources": ["arxiv", "semantic_scholar"], "title": "Selective Sinkhorn Routing for Improved Sparse Mixture of Experts", "abstract": "Sparse Mixture-of-Experts (SMoE) models are scalable and computationally efficient, enabling large increases in model capacity with limited inference overhead. Existing SMoE methods often depend on auxiliary objectives, such as load-balancing loss and z-loss, or additional trainable components such as noisy gating. While these techniques encourage expert diversity, they can introduce objective misalignment, increase model complexity, or incur substantial training overhead, especially in Sinkhorn-based routing methods. In this paper, we revisit the token-to-expert assignment as an optimal transport problem. We add constraints to ensure balanced expert utilization. We show that even minimal optimal transport-based routing improves SMoE performance without requiring auxiliary balancing losses. Unlike prior approaches, our method derives gating scores directly from the transport map, leading to more balanced and effective token-to-expert assignments. Building on this insight, we introduce Selective Sinkhorn Routing (SSR), a lightweight routing mechanism that replaces complex auxiliary losses with efficient Sinkhorn-based routing while preserving flexible expert selection. Experiments on language modeling and image classification show that SSR improves training efficiency, accuracy, and robustness to input corruption.", "authors": ["Duc Anh Nguyen", "Huu Binh Ta", "Nhuan Le Duc", "Tan Minh Nguyen", "Toan Tran"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-12", "url": "https://arxiv.org/abs/2511.08972", "pdf_url": "https://arxiv.org/pdf/2511.08972v2", "arxiv_id": "2511.08972", "doi": "10.48550/arXiv.2511.08972", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3713}
{"id": "ba3fea36a4c6116bb6c5e3fb2b8fdf63cc78838b54e7f8f2808760cb0d655088", "sources": ["arxiv", "semantic_scholar"], "title": "Sharp Eyes and Memory for VideoLLMs: Information-Aware Visual Token Pruning for Efficient and Reliable VideoLLM Reasoning", "abstract": "Current Video Large Language Models (VideoLLMs) suffer from quadratic computational complexity and key-value cache scaling, due to their reliance on processing excessive redundant visual tokens. To address this problem, we propose SharpV, a minimalist and efficient method for adaptive pruning of visual tokens and KV cache. Different from most uniform compression approaches, SharpV dynamically adjusts pruning ratios based on spatial-temporal information. Remarkably, this adaptive mechanism occasionally achieves performance gains over dense models, offering a novel paradigm for adaptive pruning. During the KV cache pruning stage, based on observations of visual information degradation, SharpV prunes degraded visual features via a self-calibration manner, guided by similarity to original visual features. In this way, SharpV achieves hierarchical cache pruning from the perspective of information bottleneck, offering a new insight into VideoLLMs' information flow. Experiments on multiple public benchmarks demonstrate the superiority of SharpV. Moreover, to the best of our knowledge, SharpV is notably the first two-stage pruning framework that operates without requiring access to exposed attention scores, ensuring full compatibility with hardware acceleration techniques like Flash Attention.", "authors": ["Jialong Qin", "Xin Zou", "Di Lu", "Yibo Yan", "Xuming Hu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-11", "url": "https://arxiv.org/abs/2511.08003", "pdf_url": "https://arxiv.org/pdf/2511.08003v2", "arxiv_id": "2511.08003", "doi": "10.48550/arXiv.2511.08003", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2355}
{"id": "a780d26c30c8e55e67ca791c47bbb32548e73c3c991371fbdf625f8ec16e6667", "sources": ["arxiv", "semantic_scholar"], "title": "Let the Experts Speak: Improving Survival Prediction & Calibration via Mixture-of-Experts Heads", "abstract": "Deep mixture-of-experts models have attracted a lot of attention for survival analysis problems, particularly for their ability to cluster similar patients together. In practice, grouping often comes at the expense of key metrics such as calibration error and predictive accuracy. This is due to the restrictive inductive bias that mixture-of-experts imposes, that predictions for individual patients must look like predictions for the group they're assigned to. Might we be able to discover patient group structure, where it exists, while improving calibration and predictive accuracy? In this work, we introduce several discrete-time deep mixture-of-experts (MoE)-based architectures for survival analysis problems, one of which achieves all desiderata: clustering, calibration, and predictive accuracy. We show that a key differentiator between this array of MoEs is how expressive their experts are. We find that more expressive experts that tailor predictions per patient outperform experts that rely on fixed group prototypes.", "authors": ["Todd Morrill", "Aahlad Puli", "Murad Megjhani", "Soojin Park", "Richard Zemel"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-11", "url": "https://arxiv.org/abs/2511.09567", "pdf_url": "https://arxiv.org/pdf/2511.09567v2", "arxiv_id": "2511.09567", "doi": "10.48550/arXiv.2511.09567", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3701}
{"id": "e73516a72f7a7d4e9e074a4fcb544575905ca3039874cdd7a70d77cdc7332c6f", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning as Regularization: Sensitivity-Aware One-Shot Pruning in ASR", "abstract": "We challenge the conventional view of neural network pruning as solely a compression technique, demonstrating that one-shot magnitude pruning serves as a powerful implicit regularizer for ASR. Using Whisper-small, we combine gradient- and Fisher-based sensitivity diagnostics with targeted, component-wise pruning. This reveals architectural asymmetries: decoder FFNs are pruning-fragile, whereas decoder self-attention and the last encoder layers contain redundancy that, when removed, improves generalization. Without fine-tuning, pruning 50% of decoder self-attention reduces WER by 2.38% absolute (20.44% relative) on LibriSpeech test-other; pruning the last four encoder layers at 50% instead yields a 1.72% absolute (14.8% relative) improvement. Gains persisted on Common Voice and TED-LIUM datasets. Beyond regularization benefits, our sensitivity-aware approach enables more aggressive one-shot compression. At 40% sparsity, where established global pruning approaches catastrophically fail, our method preserves near-baseline accuracy. This positions pruning as a first-class architectural design tool: knowing where to prune is as important as how much to prune.", "authors": ["Julian Irigoyen", "Arthur Söhler", "Andreas Søeborg Kirkedal"], "categories": ["eess.AS", "cs.CL", "cs.SD"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-11-11", "url": "https://arxiv.org/abs/2511.08092", "pdf_url": "https://arxiv.org/pdf/2511.08092v1", "arxiv_id": "2511.08092", "doi": "10.48550/arXiv.2511.08092", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3701}
{"id": "8f0e78a5f958959eab017d6cd22ef69c2daa54a4016538187d6562367a1c8c42", "sources": ["arxiv", "semantic_scholar"], "title": "Pre-Attention Expert Prediction and Prefetching for Mixture-of-Experts Large Language Models", "abstract": "Mixture-of-Experts (MoE) Large Language Models (LLMs) efficiently scale-up the model while keeping relatively low inference cost. As MoE models only activate part of the experts, related work has proposed expert prediction and caching methods to prefetch the experts for faster inference. However, existing approaches utilize the activations from the previous layer for prediction, incurring low accuracy and leave the first layer unoptimized. Applying complex layers or even training standalone networks for better prediction introduces high computation overhead. In this paper, we propose pre-attention expert prediction to achieve accurate and lightweight expert prefetching. The key insight is that some functions in LLMs are ranking-preserving, indicating that matching the ranking of selected experts using simple linear functions is possible. Therefore, we utilize the activations before the attention block in the same layer with 2 linear functions and ranking-aware loss to achieve accurate prediction, which also supports prefetching in the first layer. Our lightweight, pre-attention expert routers achieve 93.03% accuracy on DeepSeek V2 Lite, 94.69% on Qwen3-30B, and 97.62% on Phi-mini-MoE, showing about 15% improvement on absolute accuracy over the state-of-the-art methods.", "authors": ["Shien Zhu", "Samuel Bohl", "Robin Oester", "Gustavo Alonso"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-10", "url": "https://arxiv.org/abs/2511.10676", "pdf_url": "https://arxiv.org/pdf/2511.10676v1", "arxiv_id": "2511.10676", "doi": "10.48550/arXiv.2511.10676", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.369}
{"id": "348cd4c74e2a7121c7236c452e1901776dec46885139d6e95e6322a9cbb7606c", "sources": ["arxiv", "semantic_scholar"], "title": "StreamKV: Streaming Video Question-Answering with Segment-based KV Cache Retrieval and Compression", "abstract": "Video Large Language Models (Video-LLMs) have demonstrated significant potential in the areas of video captioning, search, and summarization. However, current Video-LLMs still face challenges with long real-world videos. Recent methods have introduced a retrieval mechanism that retrieves query-relevant KV caches for question answering, enhancing the efficiency and accuracy of long real-world videos. However, the compression and retrieval of KV caches are still not fully explored. In this paper, we propose \\textbf{StreamKV}, a training-free framework that seamlessly equips Video-LLMs with advanced KV cache retrieval and compression. Compared to previous methods that used uniform partitioning, StreamKV dynamically partitions video streams into semantic segments, which better preserves semantic information. For KV cache retrieval, StreamKV calculates a summary vector for each segment to retain segment-level information essential for retrieval. For KV cache compression, StreamKV introduces a guidance prompt designed to capture the key semantic elements within each segment, ensuring only the most informative KV caches are retained for answering questions. Moreover, StreamKV unifies KV cache retrieval and compression within a single module, performing both in a layer-adaptive manner, thereby further improving the effectiveness of streaming video question answering. Extensive experiments on public StreamingVQA benchmarks demonstrate that StreamKV significantly outperforms existing Online Video-LLMs, achieving superior accuracy while substantially improving both memory efficiency and computational latency. The code has been released at https://github.com/sou1p0wer/StreamKV.", "authors": ["Yilong Chen", "Xiang Bai", "Zhibin Wang", "Chengyu Bai", "Yuhan Dai", "Ming Lu", "Shanghang Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-10", "url": "https://arxiv.org/abs/2511.07278", "pdf_url": "https://arxiv.org/pdf/2511.07278v1", "arxiv_id": "2511.07278", "doi": "10.48550/arXiv.2511.07278", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/sou1p0wer/StreamKV", "venue": "arXiv.org", "quality_score": 0.5702}
{"id": "89f15a7a41fdb02ee639bed67e8fc0c92292caf33e83370f3c1afbfd19f90ceb", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA on the Go: Instance-level Dynamic LoRA Selection and Merging", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient approach for fine-tuning large language models. However, conventional LoRA adapters are typically trained for a single task, limiting their applicability in real-world settings where inputs may span diverse and unpredictable domains. At inference time, existing approaches combine multiple LoRAs for improving performance on diverse tasks, while usually requiring labeled data or additional task-specific training, which is expensive at scale. In this work, we introduce LoRA on the Go (LoGo), a training-free framework that dynamically selects and merges adapters at the instance level without any additional requirements. LoGo leverages signals extracted from a single forward pass through LoRA adapters, to identify the most relevant adapters and determine their contributions on-the-fly. Across 5 NLP benchmarks, 27 datasets, and 3 model families, LoGo outperforms training-based baselines on some tasks upto a margin of 3.6% while remaining competitive on other tasks and maintaining inference throughput, highlighting its effectiveness and practicality.", "authors": ["Seungeon Lee", "Soumi Das", "Manish Gupta", "Krishna P. Gummadi"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-10", "url": "https://arxiv.org/abs/2511.07129", "pdf_url": "https://arxiv.org/pdf/2511.07129v3", "arxiv_id": "2511.07129", "doi": "10.48550/arXiv.2511.07129", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.369}
{"id": "e46d70cb8e3e5e283f3c009db7c8adb208f0ab92f28ac73083be8858eac942f7", "sources": ["arxiv", "semantic_scholar"], "title": "Spatially-Aware Mixture of Experts with Log-Logistic Survival Modeling for Whole-Slide Images", "abstract": "Accurate survival prediction from histopathology whole-slide images (WSIs) remains challenging due to their gigapixel resolution, strong spatial heterogeneity, and complex survival distributions. We introduce a comprehensive computational pathology framework that addresses these limitations through four complementary innovations: (1) Quantile-Gated Patch Selection for dynamically identifying prognostically relevant regions, (2) Graph-Guided Clustering to group patches by spatial and morphological similarity, (3) Hierarchical Context Attention to model both local tissue interactions and global slide-level context, and (4) an Expert-Driven Mixture of Log-Logistics module that flexibly models complex survival distributions. Across large TCGA cohorts, our method achieves state-of-the-art performance, yielding time-dependent concordance indices of 0.644 on LUAD, 0.751 on KIRC, and 0.752 on BRCA, consistently outperforming both histology-only and multimodal baselines. The framework further provides improved calibration and interpretability, advancing the use of WSIs for personalized cancer prognosis.", "authors": ["Ardhendu Sekhar", "Vasu Soni", "Keshav Aske", "Shivam Madnoorkar", "Pranav Jeevan", "Amit Sethi"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-09", "url": "https://arxiv.org/abs/2511.06266", "pdf_url": "https://arxiv.org/pdf/2511.06266v3", "arxiv_id": "2511.06266", "doi": "10.1109/ISBI61048.2026.11515974", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Symposium on Biomedical Imaging", "quality_score": 0.3678}
{"id": "2522864b4b384872f113d3c206a63a771c029847851078f6ff61cc5774967416", "sources": ["arxiv", "semantic_scholar"], "title": "Lethe: Layer- and Time-Adaptive KV Cache Pruning for Reasoning-Intensive LLM Serving", "abstract": "Generative reasoning with large language models (LLMs) often involves long decoding sequences, leading to substantial memory and latency overheads from accumulating key-value (KV) caches. While existing KV compression methods primarily focus on reducing prefill memory from long input sequences, they fall short in addressing the dynamic and layer-sensitive nature of long-form generation, which is central to reasoning tasks. We propose Lethe, a dynamic KV cache management framework that introduces adaptivity along both the spatial and temporal dimensions of decoding. Along the spatial dimension, Lethe performs layerwise sparsity-aware allocation, assigning token pruning budgets to each transformer layer based on estimated attention redundancy. Along the temporal dimension, Lethe conducts multi-round token pruning during generation, driven by a Recency-Aware Selective Retention} (RASR) mechanism. RASR extends traditional recency-based heuristics by also considering token relevance derived from evolving attention patterns, enabling informed decisions about which tokens to retain or evict. Empirical results demonstrate that Lethe achieves a favorable balance between efficiency and generation quality across diverse models and tasks, increases throughput by up to 2.56x.", "authors": ["Hui Zeng", "Daming Zhao", "Pengfei Yang", "WenXuan Hou", "Tianyang Zheng", "Hui Li", "Weiye Ji", "Jidong Zhai"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-08", "url": "https://arxiv.org/abs/2511.06029", "pdf_url": "https://arxiv.org/pdf/2511.06029v3", "arxiv_id": "2511.06029", "doi": "10.48550/arXiv.2511.06029", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3667}
{"id": "f7cb19e63e312ec7674a3a372038b4460bde6d609d97db92aabf90687fa4b703", "sources": ["arxiv", "semantic_scholar"], "title": "MoSKA: Mixture of Shared KV Attention for Efficient Long-Sequence LLM Inference", "abstract": "The escalating context length in Large Language Models (LLMs) creates a severe performance bottleneck around the Key-Value (KV) cache, whose memory-bound nature leads to significant GPU under-utilization. This paper introduces Mixture of Shared KV Attention (MoSKA), an architecture that addresses this challenge by exploiting the heterogeneity of context data. It differentiates between per-request unique and massively reused shared sequences. The core of MoSKA is a novel Shared KV Attention mechanism that transforms the attention on shared data from a series of memory-bound GEMV operations into a single, compute-bound GEMM by batching concurrent requests. This is supported by an MoE-inspired sparse attention strategy that prunes the search space and a tailored Disaggregated Infrastructure that specializes hardware for unique and shared data. This comprehensive approach demonstrates a throughput increase of up to 538.7x over baselines in workloads with high context sharing, offering a clear architectural path toward scalable LLM inference.", "authors": ["Myunghyun Rhee", "Sookyung Choi", "Euiseok Kim", "Joonseop Sim", "Youngpyo Joo", "Hoshik Kim"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-08", "url": "https://arxiv.org/abs/2511.06010", "pdf_url": "https://arxiv.org/pdf/2511.06010v1", "arxiv_id": "2511.06010", "doi": "10.1109/LCA.2025.3627539", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE computer architecture letters", "quality_score": 0.3667}
{"id": "5929d16ed5ad097a043d67703e814dff9c3f3a9d30d7d60bca43007cb8e783a4", "sources": ["arxiv", "semantic_scholar"], "title": "In-depth Analysis on Caching and Pre-fetching in Mixture of Experts Offloading", "abstract": "In today's landscape, Mixture of Experts (MoE) is a crucial architecture that has been used by many of the most advanced models. One of the major challenges of MoE models is that they usually require much more memory than their dense counterparts due to their unique architecture, and hence are harder to deploy in environments with limited GPU memory, such as edge devices. MoE offloading is a promising technique proposed to overcome this challenge, especially if it is enhanced with caching and pre-fetching, but prior work stopped at suboptimal caching algorithm and offered limited insights. In this work, we study MoE offloading in depth and make the following contributions: 1. We analyze the expert activation and LRU caching behavior in detail and provide traces. 2. We propose LFU caching optimization based on our analysis and obtain strong improvements from LRU. 3. We implement and experiment speculative expert pre-fetching, providing detailed trace showing its huge potential . 4. In addition, our study extensively covers the behavior of the MoE architecture itself, offering information on the characteristic of the gating network and experts. This can inspire future work on the interpretation of MoE models and the development of pruning techniques for MoE architecture with minimal performance loss.", "authors": ["Shuning Lin", "Yifan He", "Yitong Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-08", "url": "https://arxiv.org/abs/2511.05814", "pdf_url": "https://arxiv.org/pdf/2511.05814v1", "arxiv_id": "2511.05814", "doi": "10.48550/arXiv.2511.05814", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3667}
{"id": "14e03451f17122b86e8adabfe0cbd0a9c92c71b1c1645ec372d6baebd688854d", "sources": ["arxiv", "semantic_scholar"], "title": "Q$^2$: Quantization-Aware Gradient Balancing and Attention Alignment for Low-Bit Quantization", "abstract": "Quantization-aware training (QAT) has achieved remarkable success in low-bit ($\\leq$4-bit) quantization for classification networks. However, when applied to more complex visual tasks such as object detection and image segmentation, performance still suffers significant degradation. A key cause of this limitation has been largely overlooked in the literature. In this work, we revisit this phenomenon from a new perspective and identify a major failure factor: gradient imbalance at feature fusion stages, induced by accumulated quantization errors. This imbalance biases the optimization trajectory and impedes convergence under low-bit quantization. Based on this diagnosis, we propose Q$^2$, a two-pronged framework comprising: (1) Quantization-aware Gradient Balancing Fusion (Q-GBFusion), a closed-loop mechanism that dynamically rebalances gradient contributions during feature fusion; and (2) Quantization-aware Attention Distribution Alignment (Q-ADA), a parameter-free supervision strategy that reconstructs the supervision distribution using semantic relevance and quantization sensitivity, yielding more stable and reliable supervision to stabilize training and accelerate convergence. Extensive experiments show that our method, as a plug-and-play and general strategy, can be integrated into various state-of-the-art QAT pipelines, achieving an average +2.5\\% mAP gain on object detection and a +3.7\\% mDICE improvement on image segmentation. Notably, it is applied only during training and introduces no inference-time overhead, making it highly practical for real-world deployment.", "authors": ["Zhaoyang Wang", "Dong Wang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-08", "url": "https://arxiv.org/abs/2511.05898", "pdf_url": "https://arxiv.org/pdf/2511.05898v2", "arxiv_id": "2511.05898", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2333}
{"id": "824cfa0d94a3338b22cf76defdd02c7eb5ab40f6e3c767058309964100be8d75", "sources": ["arxiv", "semantic_scholar"], "title": "GMoPE:A Prompt-Expert Mixture Framework for Graph Foundation Models", "abstract": "Graph Neural Networks (GNNs) have demonstrated impressive performance on task-specific benchmarks, yet their ability to generalize across diverse domains and tasks remains limited. Existing approaches often struggle with negative transfer, scalability issues, and high adaptation costs. To address these challenges, we propose GMoPE (Graph Mixture of Prompt-Experts), a novel framework that seamlessly integrates the Mixture-of-Experts (MoE) architecture with prompt-based learning for graphs. GMoPE leverages expert-specific prompt vectors and structure-aware MoE routing to enable each expert to specialize in distinct subdomains and dynamically contribute to predictions. To promote diversity and prevent expert collapse, we introduce a soft orthogonality constraint across prompt vectors, encouraging expert specialization and facilitating a more balanced expert utilization. Additionally, we adopt a prompt-only fine-tuning strategy that significantly reduces spatiotemporal complexity during transfer. We validate GMoPE through extensive experiments under various pretraining strategies and multiple downstream tasks. Results show that GMoPE consistently outperforms state-of-the-art baselines and achieves performance comparable to full parameter fine-tuning-while requiring only a fraction of the adaptation overhead. Our work provides a principled and scalable framework for advancing generalizable and efficient graph foundation models.", "authors": ["Zhibin Wang", "Zhixing Zhang", "Shuqi Wang", "Xuanting Xie", "Zhao Kang"], "categories": ["cs.LG", "cs.AI", "cs.SI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-05", "url": "https://arxiv.org/abs/2511.03251", "pdf_url": "https://arxiv.org/pdf/2511.03251v1", "arxiv_id": "2511.03251", "doi": "10.48550/arXiv.2511.03251", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3632}
{"id": "750ce5e0b468c1f5aa44985bfc1d78a7acbd22bc1f05fc62e5319e5616545282", "sources": ["arxiv", "semantic_scholar"], "title": "IG-Pruning: Input-Guided Block Pruning for Large Language Models", "abstract": "With the growing computational demands of large language models (LLMs), efficient inference has become increasingly critical for practical deployment. Depth pruning has emerged as a promising approach for reducing the computational costs of large language models by removing transformer layers. However, existing methods typically rely on fixed block masks, which can lead to suboptimal performance across different tasks and inputs. In this paper, we propose IG-Pruning, a novel input-aware block-wise pruning method that dynamically selects layer masks at inference time. Our approach consists of two stages: (1) Discovering diverse mask candidates through semantic clustering and L0 optimization, and (2) Implementing efficient dynamic pruning without the need for extensive training. Experimental results demonstrate that our method consistently outperforms state-of-the-art static depth pruning methods, making it particularly suitable for resource-constrained deployment scenarios.", "authors": ["Kangyu Qiao", "Shaolei Zhang", "Yang Feng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-04", "url": "https://arxiv.org/abs/2511.02213", "pdf_url": "https://arxiv.org/pdf/2511.02213v1", "arxiv_id": "2511.02213", "doi": "10.18653/v1/2025.emnlp-main.537", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ictnlp/IG-Pruning", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.5596}
{"id": "4ebdb60cfd6770c46b9a521364985b51930e17c48743c42a19f72a3a1acf0c8f", "sources": ["arxiv", "semantic_scholar"], "title": "Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live", "abstract": "KV cache management is essential for efficient LLM inference. To maximize utilization, existing inference engines evict finished requests' KV cache if new requests are waiting. This policy breaks for agentic workloads, which interleave LLM calls with tools, introducing pauses that prevent effective KV reuse across turns. Since many tool calls have much shorter durations than human response multi-turn chatbot, it would be promising to retain the KV cache in during these tools. However, many challenges remain. First, we need to consider both the potential cost of recomputation or reloading (if offloading enabled) as well as the increasing queueing delays after eviction from GPU. Second, due to the internal variance of tool call durations, the method needs to remain robust under limited predictability of tool call durations. We present Continuum, a serving system to optimize job completion time for multi-turn agent workloads by introducing time-to-live mechanism for KV cache retention. For requests that generate tool calls, Continuum selectively pins the KV cache in GPU memory with a time-to-live value determined by the reload cost and potential queueing delay induced by eviction. When the TTL expires, the KV cache can be automatically evicted to free up GPU memory, providing robust performance under edge cases. When combined with program-level first-come-first-serve, Continuum preserves multi-turn continuity, and reduces delay for agentic workflows. Evaluations on real-world agents (SWE-Bench, BFCL, OpenHand) with Llama-3.1 8B/70B, Gemma-3 12B, and GLM-4.5 355B shows that Continuum improves the average job completion times by over 8x while improving throughput.", "authors": ["Hanchen Li", "Runyuan He", "Qiuyang Mang", "Qizheng Zhang", "Huanzhi Mao", "Xiaokun Chen", "Hangrui Zhou", "Alvin Cheung", "Joseph Gonzalez", "Ion Stoica"], "categories": ["cs.OS", "cs.AI", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-04", "url": "https://arxiv.org/abs/2511.02230", "pdf_url": "https://arxiv.org/pdf/2511.02230v6", "arxiv_id": "2511.02230", "doi": "10.48550/arXiv.2511.02230", "citation_count": 31, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3891}
{"id": "356ab277face0402a89ac531552d2465d8c0a5251b934a1c2cf4f4394d6169fa", "sources": ["arxiv", "semantic_scholar"], "title": "Using Span Queries to Optimize for Cache and Attention Locality", "abstract": "Clients are evolving beyond chat completion, and now include a variety of innovative inference-time scaling and deep reasoning techniques. At the same time, inference servers remain heavily optimized for chat completion. Prior work has shown that large improvements to KV cache hit rate are possible if inference servers evolve towards these non-chat use cases. However, they offer solutions that are also optimized for a single use case, RAG. In this paper, we introduce the span query to generalize the interface to the inference server. We demonstrate that chat, RAG, inference-time scaling, and agentic workloads can all be expressed as span queries. We show how the critical distinction that had been assumed by prior work lies in whether the order of the inputs matter -- do they commute? In chat, they do not. In RAG, they often do. This paper introduces span queries, which are expression trees of inference calls, linked together with commutativity constraints. We describe span query syntax and semantics. We show how they can be automatically optimized to improve KV cache locality. We show how a small change to vLLM (affecting only 492 lines) can enable high-performance execution of span queries. Using this stack, we demonstrate that span queries can achieve 10-20x reductions in TTFT for two distinct non-chat use cases. Finally, we show that span queries can also be optimized to improve attention locality, so as to avoid the so-called lost-in-the-middle problem. We demonstrate that an attention-optimized span query on a 2b parameter model vastly outperforms the accuracy of a stock inference server using an 8b model.", "authors": ["Paul Castro", "Nick Mitchell", "Nathan Ordonez", "Thomas Parnell", "Mudhakar Srivatsa", "Antoni Viros i Martin"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-04", "url": "https://arxiv.org/abs/2511.02749", "pdf_url": "https://arxiv.org/pdf/2511.02749v1", "arxiv_id": "2511.02749", "doi": "10.48550/arXiv.2511.02749", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3621}
{"id": "fc920cc9395448c7d861efb05326791660a762568567b3c0ed229f16fbb227ba", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache Transform Coding for Compact Storage in LLM Inference", "abstract": "Serving large language models (LLMs) at scale necessitates efficient key-value (KV) cache management. KV caches can be reused across conversation turns via shared-prefix prompts that are common in iterative code editing and chat. However, stale caches consume scarce GPU memory, require offloading, or force recomputation. We present KVTC, a lightweight transform coder that compresses KV caches for compact on-GPU and off-GPU storage. Drawing on classical media compression, KVTC combines PCA-based feature decorrelation, adaptive quantization, and entropy coding. It requires only a brief initial calibration and leaves model parameters unchanged. By exploiting redundancies in KV caches, KVTC achieves up to 20$\\times$ compression while maintaining reasoning and long-context accuracy, and 40$\\times$ or higher for specific use cases. We test KVTC with Llama 3, Mistral NeMo, and R1-Qwen 2.5 models across benchmarks including AIME25, GSM8K, LiveCodeBench, LongBench, MATH-500, MMLU, Qasper and RULER. It consistently outperforms inference-time baselines such as token eviction, quantization, and SVD-based methods, while achieving higher compression ratios. These results support KVTC as a practical building block for memory-efficient LLM serving with reusable KV caches.", "authors": ["Konrad Staniszewski", "Adrian Łańcucki"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-03", "url": "https://arxiv.org/abs/2511.01815", "pdf_url": "https://arxiv.org/pdf/2511.01815v2", "arxiv_id": "2511.01815", "doi": "10.48550/arXiv.2511.01815", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3609}
{"id": "04fb65c141d981f69671f96f47d9f767667c35ae872dcc68574515535ae0d069", "sources": ["arxiv", "semantic_scholar"], "title": "CryptoMoE: Privacy-Preserving and Scalable Mixture of Experts Inference via Balanced Expert Routing", "abstract": "Private large language model (LLM) inference based on cryptographic primitives offers a promising path towards privacy-preserving deep learning. However, existing frameworks only support dense LLMs like LLaMA-1 and struggle to scale to mixture-of-experts (MoE) architectures. The key challenge comes from securely evaluating the dynamic routing mechanism in MoE layers, which may reveal sensitive input information if not fully protected. In this paper, we propose CryptoMoE, the first framework that enables private, efficient, and accurate inference for MoE-based models. CryptoMoE balances expert loads to protect expert routing information and proposes novel protocols for secure expert dispatch and combine. CryptoMoE also develops a confidence-aware token selection strategy and a batch matrix multiplication protocol to improve accuracy and efficiency further. Extensive experiments on DeepSeekMoE-16.4B, OLMoE-6.9B, and QWenMoE-14.3B show that CryptoMoE achieves $2.8\\sim3.5\\times$ end-to-end latency reduction and $2.9\\sim4.3\\times$ communication reduction over a dense baseline with minimum accuracy loss. We also adapt CipherPrune (ICLR'25) for MoE inference and demonstrate CryptoMoE can reduce the communication by up to $4.3 \\times$. Code is available at: https://github.com/PKU-SEC-Lab/CryptoMoE.", "authors": ["Yifan Zhou", "Tianshi Xu", "Jue Hong", "Ye Wu", "Meng Li"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-03", "url": "https://arxiv.org/abs/2511.01197", "pdf_url": "https://arxiv.org/pdf/2511.01197v3", "arxiv_id": "2511.01197", "doi": "10.48550/arXiv.2511.01197", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/PKU-SEC-Lab/CryptoMoE", "venue": "arXiv.org", "quality_score": 0.5578}
{"id": "2243822438aa43ef74e9522989694eec73fb4ab7b6c6d510b8c1f31d43030ca1", "sources": ["arxiv", "semantic_scholar"], "title": "FlexiCache: Leveraging Temporal Stability of Attention Heads for Efficient KV Cache Management", "abstract": "Large Language Model (LLM) serving is increasingly constrained by the growing size of the key-value (KV) cache, which scales with both context length and generation length. Prior work shows that attention is dominated by a small subset of critical tokens, yet existing systems struggle to exploit this efficiently without degrading accuracy, especially in long generation. We make a key observation: the temporal stability of these critical tokens varies significantly across KV heads: some heads consistently focus on the same tokens, while others shift frequently. Building on this insight, we introduce FlexiCache, a hierarchical KV-cache management system that leverages the temporal stability of KV heads to reduce GPU memory usage and computation overhead, while preserving model accuracy. FlexiCache classifies KV heads as stable or unstable: it retains all KV-cache pages from unstable heads in GPU memory, whereas for stable heads, it keeps only the top-K pages on the GPU and offloads the rest to host memory. By exploiting temporal stability, FlexiCache performs periodic reranking for stable heads to fetch newly promoted top pages. Implemented atop vLLM, FlexiCache reduces GPU memory footprint for long-context requests by up to 70%, improves offline serving throughput by 1.38-1.55x, and lowers online token latency by 1.6-2.1x, all while maintaining accuracy in long-context, long-generation scenarios.", "authors": ["Nazmul Takbir", "Hamidreza Alikhani", "Nikil Dutt", "Sangeetha Abdu Jyothi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-02", "url": "https://arxiv.org/abs/2511.00868", "pdf_url": "https://arxiv.org/pdf/2511.00868v2", "arxiv_id": "2511.00868", "doi": "10.48550/arXiv.2511.00868", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3598}
{"id": "8dac8885b5ce0484721c128ce8e2d372677b061f5cbf7cc0f14fdf6ff30ef1eb", "sources": ["arxiv", "semantic_scholar"], "title": "Scalable Processing-Near-Memory for 1M-Token LLM Inference: CXL-Enabled KV-Cache Management Beyond GPU Limits", "abstract": "The expansion of context windows in large language models (LLMs) to multi-million tokens introduces severe memory and compute bottlenecks, particularly in managing the growing Key-Value (KV) cache. While Compute Express Link (CXL) enables non-eviction frameworks that offload the full KV-cache to scalable external memory, these frameworks still suffer from costly data transfers when recalling non-resident KV tokens to limited GPU memory as context lengths increase. This work proposes scalable Processing-Near-Memory (PNM) for 1M-Token LLM Inference, a CXL-enabled KV-cache management system that coordinates memory and computation beyond GPU limits. Our design offloads token page selection to a PNM accelerator within CXL memory, eliminating costly recalls and enabling larger GPU batch sizes. We further introduce a hybrid parallelization strategy and a steady-token selection mechanism to enhance compute efficiency and scalability. Implemented atop a state-of-the-art CXL-PNM system, our solution delivers consistent performance gains for LLMs with up to 405B parameters and 1M-token contexts. Our PNM-only offloading scheme (PNM-KV) and GPU-PNM hybrid with steady-token execution (PnG-KV) achieve up to 21.9x throughput improvement, up to 60x lower energy per token, and up to 7.3x better total cost efficiency than the baseline, demonstrating that CXL-enabled multi-PNM architectures can serve as a scalable backbone for future long-context LLM inference.", "authors": ["Dowon Kim", "MinJae Lee", "Janghyeon Kim", "HyuckSung Kwon", "Hyeonggyu Jeong", "Sang-Soo Park", "Minyong Yoon", "Si-Dong Roh", "Yongsuk Kwon", "Jinin So", "Jungwook Choi"], "categories": ["cs.AR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-31", "url": "https://arxiv.org/abs/2511.00321", "pdf_url": "https://arxiv.org/pdf/2511.00321v1", "arxiv_id": "2511.00321", "doi": "10.1109/PACT65351.2025.00013", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Parallel Architectures and Compilation Techniques", "quality_score": 0.3575}
{"id": "e7868729b0f272755379ce9710bfc9487ce9c0b48b799bbf3b8fb78b70503599", "sources": ["arxiv", "semantic_scholar"], "title": "Fine-Tuning Open Video Generators for Cinematic Scene Synthesis: A Small-Data Pipeline with LoRA and Wan2.1 I2V", "abstract": "We present a practical pipeline for fine-tuning open-source video diffusion transformers to synthesize cinematic scenes for television and film production from small datasets. The proposed two-stage process decouples visual style learning from motion generation. In the first stage, Low-Rank Adaptation (LoRA) modules are integrated into the cross-attention layers of the Wan2.1 I2V-14B model to adapt its visual representations using a compact dataset of short clips from Ay Yapim's historical television film El Turco. This enables efficient domain transfer within hours on a single GPU. In the second stage, the fine-tuned model produces stylistically consistent keyframes that preserve costume, lighting, and color grading, which are then temporally expanded into coherent 720p sequences through the model's video decoder. We further apply lightweight parallelization and sequence partitioning strategies to accelerate inference without quality degradation. Quantitative and qualitative evaluations using FVD, CLIP-SIM, and LPIPS metrics, supported by a small expert user study, demonstrate measurable improvements in cinematic fidelity and temporal stability over the base model. The complete training and inference pipeline is released to support reproducibility and adaptation across cinematic domains.", "authors": ["Meftun Akarsu", "Kerem Catay", "Sedat Bin Vedat", "Enes Kutay Yarkan", "Ilke Senturk", "Arda Sar", "Dafne Eksioglu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-31", "url": "https://arxiv.org/abs/2510.27364", "pdf_url": "https://arxiv.org/pdf/2510.27364v1", "arxiv_id": "2510.27364", "doi": "10.5281/zenodo.17370356", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5525}
{"id": "cfb7cd10eeda13022c24c714692623e1ff887590142403929abd960e263fcab8", "sources": ["arxiv", "semantic_scholar"], "title": "MossNet: Mixture of State-Space Experts is a Multi-Head Attention", "abstract": "Large language models (LLMs) have significantly advanced generative applications in natural language processing (NLP). Recent trends in model architectures revolve around efficient variants of transformers or state-space/gated-recurrent models (SSMs, GRMs). However, prevailing SSM/GRM-based methods often emulate only a single attention head, potentially limiting their expressiveness. In this work, we propose MossNet, a novel mixture-of-state-space-experts architecture that emulates a linear multi-head attention (MHA). MossNet leverages a mixture-of-experts (MoE) implementation not only in channel-mixing multi-layered perceptron (MLP) blocks but also in the time-mixing SSM kernels to realize multiple \"attention heads.\" Extensive experiments on language modeling and downstream evaluations show that MossNet outperforms both transformer- and SSM-based architectures of similar model size and data budgets. Larger variants of MossNet, trained on trillions of tokens, further confirm its scalability and superior performance. In addition, real-device profiling on a Samsung Galaxy S24 Ultra and an Nvidia A100 GPU demonstrate favorable runtime speed and resource usage compared to similarly sized baselines. Our results suggest that MossNet is a compelling new direction for efficient, high-performing recurrent LLM architectures.", "authors": ["Shikhar Tuli", "James Seale Smith", "Haris Jeelani", "Chi-Heng Lin", "Abhishek Patel", "Vasili Ramanishka", "Yen-Chang Hsu", "Hongxia Jin"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-30", "url": "https://arxiv.org/abs/2510.26182", "pdf_url": "https://arxiv.org/pdf/2510.26182v1", "arxiv_id": "2510.26182", "doi": "10.48550/arXiv.2510.26182", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2268}
{"id": "001bcb94d3a4abef1f7fd949a8cf09a5d42c4b379805290a788ee0ed0f0ecf22", "sources": ["arxiv", "semantic_scholar"], "title": "LoRAQuant: Mixed-Precision Quantization of LoRA to Ultra-Low Bits", "abstract": "Low-Rank Adaptation (LoRA) has become a popular technique for parameter-efficient fine-tuning of large language models (LLMs). In many real-world scenarios, multiple adapters are loaded simultaneously to enable LLM customization for personalized user experiences or to support a diverse range of tasks. Although each adapter is lightweight in isolation, their aggregate cost becomes substantial at scale. To address this, we propose LoRAQuant, a mixed-precision post-training quantization method tailored to LoRA. Specifically, LoRAQuant reparameterizes each adapter by singular value decomposition (SVD) to concentrate the most important information into specific rows and columns. This makes it possible to quantize the important components to higher precision, while quantizing the rest to ultra-low bitwidth. We conduct comprehensive experiments with LLaMA 2-7B, LLaMA 2-13B, and Mistral 7B models on mathematical reasoning, coding, and summarization tasks. Results show that our LoRAQuant uses significantly lower bits than other quantization methods, but achieves comparable or even higher performance.", "authors": ["Amir Reza Mirzaei", "Yuqiao Wen", "Yanshuai Cao", "Lili Mou"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-30", "url": "https://arxiv.org/abs/2510.26690", "pdf_url": "https://arxiv.org/pdf/2510.26690v2", "arxiv_id": "2510.26690", "doi": "10.48550/arXiv.2510.26690", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3564}
{"id": "7d86f4be384c78d5d99f734ed6130f543e6f49cda362f6e3d9e48a9ba3f5197a", "sources": ["arxiv", "semantic_scholar"], "title": "MoME: Mixture of Visual Language Medical Experts for Medical Imaging Segmentation", "abstract": "In this study, we propose MoME, a Mixture of Visual Language Medical Experts, for Medical Image Segmentation. MoME adapts the successful Mixture of Experts (MoE) paradigm, widely used in Large Language Models (LLMs), for medical vision-language tasks. The architecture enables dynamic expert selection by effectively utilizing multi-scale visual features tailored to the intricacies of medical imagery, enriched with textual embeddings. This work explores a novel integration of vision-language models for this domain. Utilizing an assembly of 10 datasets, encompassing 3,410 CT scans, MoME demonstrates strong performance on a comprehensive medical imaging segmentation benchmark. Our approach explores the integration of foundation models for medical imaging, benefiting from the established efficacy of MoE in boosting model performance by incorporating textual information. Demonstrating competitive precision across multiple datasets, MoME explores a novel architecture for achieving robust results in medical image analysis.", "authors": ["Arghavan Rezvani", "Xiangyi Yan", "Anthony T. Wu", "Kun Han", "Pooya Khosravi", "Xiaohui Xie"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-30", "url": "https://arxiv.org/abs/2510.26996", "pdf_url": "https://arxiv.org/pdf/2510.26996v1", "arxiv_id": "2510.26996", "doi": "10.48550/arXiv.2510.26996", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3564}
{"id": "db1093227b9cdaed2907c9a518af2bf0ad0c4a6bb7e8384f25ed66ee50e7fc0d", "sources": ["arxiv", "semantic_scholar"], "title": "PureKV: Plug-and-Play KV Cache Optimization with Spatial-Temporal Sparse Attention for Vision-Language Large Models", "abstract": "Vision-Language Large Models (VLLMs) face significant efficiency challenges when processing high-resolution inputs. The quadratic complexity in attention and autoregressive generation, as well as the constantly growing key value (KV) cache size, severely hinder the prefilling and decoding stages. Recent efforts have attempted to compress KV cache by identifying and pruning KV cache of less important tokens, but these methods typically rely on attention scores to estimate token importance, making them incompatible with efficient attention mechanisms such as FlashAttention and Sparse Attention, which do not explicitly compute attention matrices. Moreover, existing methods overlook how sparse attention, while accelerating the prefilling stage, alters the information structure of the KV cache, thereby compromising the effectiveness of downstream KV cache compression strategies. To address this issue, we propose PureKV, a plug-and-play framework for joint optimization of sparse attention and KV cache compression. We first introduce a KV cache compression strategy that is fully compatible with efficient attention accelerators. Our method utilizes lower layer attention scores to estimate the importance of high layers' KV cache, enabling active pruning without compromising accuracy. In addition, we have designed a Spatial-Temporal Sparse Attention (ST-SpAttn) module specifically tailored for video KV cache compression algorithms. This module combines spatial and temporal attention sparsity to improve the compression efficiency of KV cache optimization algorithms by purifying spatial noise and temporal redundancy in KV cache. At the same time, ST-SpAttn also accelerated the prefilling stage of VLLMs. Extensive experiments on VLLMs (VideoLLaMA2, Qwen2.5-VL) have shown that PureKV achieves 5.0 times KV cache compression and 3.16 times prefill acceleration, with negligible quality degradation.", "authors": ["Zhonghua Jiang", "Kunxi Li", "Yiyun Zhou", "Sihao Liu", "Zhaode Wang", "Chengfei lv", "Shengyu Zhang"], "categories": ["cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-29", "url": "https://arxiv.org/abs/2510.25600", "pdf_url": "https://arxiv.org/pdf/2510.25600v2", "arxiv_id": "2510.25600", "doi": "10.48550/arXiv.2510.25600", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3552}
{"id": "474acae81d6b8689ba3b4d6d41e3cb3dc8ba7ad054a9c1bbf25defe3fc80db46", "sources": ["arxiv", "semantic_scholar"], "title": "FlowMM: Cross-Modal Information Flow Guided KV Cache Merging for Efficient Multimodal Context Inference", "abstract": "Traditional KV cache eviction strategies, which discard less critical KV-pairs based on attention scores, often degrade generation quality, causing context loss or hallucinations. Recent efforts shift toward KV merging, merging eviction tokens with retention tokens based on similarity. However, in multimodal scenarios, distributional biases across modality tokens and attentional biases in cross-modal interactions limit its effectiveness. This work introduces FlowMM, an adaptive framework for cross-modal information flow-guided multimodal KV cache merging. FlowMM leverages cross-modal information flow to dynamically apply layer-specific merging strategies, capturing modality-specific patterns while preserving contextual integrity. Furthermore, we introduce a sensitivity-adaptive token matching mechanism that jointly evaluates token similarity and task-critical sensitivity, merging low-risk tokens while safeguarding high-sensitivity ones. Extensive experiments across diverse leading MLLMs show that FlowMM reduces KV cache memory by 80% to 95% and decoding latency by 1.3-1.8x, while maintaining competitive task performance.", "authors": ["Kunxi Li", "Yufan Xiong", "Zhonghua Jiang", "Yiyun Zhou", "Zhaode Wang", "Chengfei Lv", "Shengyu Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-29", "url": "https://arxiv.org/abs/2511.05534", "pdf_url": "https://arxiv.org/pdf/2511.05534v2", "arxiv_id": "2511.05534", "doi": "10.48550/arXiv.2511.05534", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3552}
{"id": "29a2898925e291890e73ae890218e9fe153c97914db78f315a3f47d6fa1dee9a", "sources": ["arxiv", "semantic_scholar"], "title": "H3M-SSMoEs: Hypergraph-based Multimodal Learning with LLM Reasoning and Style-Structured Mixture of Experts", "abstract": "Stock movement prediction remains fundamentally challenging due to complex temporal dependencies, heterogeneous modalities, and dynamically evolving inter-stock relationships. Existing approaches often fail to unify structural, semantic, and regime-adaptive modeling within a scalable framework. This work introduces H3M-SSMoEs, a novel Hypergraph-based MultiModal architecture with LLM reasoning and Style-Structured Mixture of Experts, integrating three key innovations: (1) a Multi-Context Multimodal Hypergraph that hierarchically captures fine-grained spatiotemporal dynamics via a Local Context Hypergraph (LCH) and persistent inter-stock dependencies through a Global Context Hypergraph (GCH), employing shared cross-modal hyperedges and Jensen-Shannon Divergence weighting mechanism for adaptive relational learning and cross-modal alignment; (2) a LLM-enhanced reasoning module, which leverages a frozen large language model with lightweight adapters to semantically fuse and align quantitative and textual modalities, enriching representations with domain-specific financial knowledge; and (3) a Style-Structured Mixture of Experts (SSMoEs) that combines shared market experts and industry-specialized experts, each parameterized by learnable style vectors enabling regime-aware specialization under sparse activation. Extensive experiments on three major stock markets demonstrate that H3M-SSMoEs surpasses state-of-the-art methods in both superior predictive accuracy and investment performance, while exhibiting effective risk control. Datasets, source code, and model weights are available at our GitHub repository: https://github.com/PeilinTime/H3M-SSMoEs.", "authors": ["Peilin Tan", "Liang Xie", "Churan Zhi", "Dian Tu", "Chuanqi Shi"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-29", "url": "https://arxiv.org/abs/2510.25091", "pdf_url": "https://arxiv.org/pdf/2510.25091v1", "arxiv_id": "2510.25091", "doi": "10.48550/arXiv.2510.25091", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/PeilinTime/H3M-SSMoEs", "venue": "arXiv.org", "quality_score": 0.549}
{"id": "a7cd7d26cda3ab2c0cc62973a5a54bbe7c4aaadcee277cad0c5b1a8adca076cf", "sources": ["arxiv", "semantic_scholar"], "title": "SALS: Sparse Attention in Latent Space for KV cache Compression", "abstract": "Large Language Models capable of handling extended contexts are in high demand, yet their inference remains challenging due to substantial Key-Value cache size and high memory bandwidth requirements. Previous research has demonstrated that KV cache exhibits low-rank characteristics within the hidden dimension, suggesting the potential for effective compression. However, due to the widely adopted Rotary Position Embedding mechanism in modern LLMs, naive low-rank compression suffers severe accuracy degradation or creates a new speed bottleneck, as the low-rank cache must first be reconstructed in order to apply RoPE. In this paper, we introduce two key insights: first, the application of RoPE to the key vectors increases their variance, which in turn results in a higher rank; second, after the key vectors are transformed into the latent space, they largely maintain their representation across most layers. Based on these insights, we propose the Sparse Attention in Latent Space framework. SALS projects the KV cache into a compact latent space via low-rank projection, and performs sparse token selection using RoPE-free query-key interactions in this space. By reconstructing only a small subset of important tokens, it avoids the overhead of full KV cache reconstruction. We comprehensively evaluate SALS on various tasks using two large-scale models: LLaMA2-7b-chat and Mistral-7b, and additionally verify its scalability on the RULER-128k benchmark with LLaMA3.1-8B-Instruct. Experimental results demonstrate that SALS achieves SOTA performance by maintaining competitive accuracy. Under different settings, SALS achieves 6.4-fold KV cache compression and 5.7-fold speed-up in the attention operator compared to FlashAttention2 on the 4K sequence. For the end-to-end throughput performance, we achieves 1.4-fold and 4.5-fold improvement compared to GPT-fast on 4k and 32K sequences, respectively.", "authors": ["Junlin Mu", "Hantao Huang", "Jihang Zhang", "Minghui Yu", "Tao Wang", "Yidong Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-28", "url": "https://arxiv.org/abs/2510.24273", "pdf_url": "https://arxiv.org/pdf/2510.24273v1", "arxiv_id": "2510.24273", "doi": "10.48550/arXiv.2510.24273", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3541}
{"id": "d886229ff828aba6da94fbc16683349e2593f143fd36af63409fa0f551bdbe8e", "sources": ["arxiv", "semantic_scholar"], "title": "FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic", "abstract": "Low-bit floating-point (FP) formats, such as FP8, provide significant acceleration and memory savings in model training thanks to native hardware support on modern GPUs and NPUs. However, we analyze that FP8 quantization offers speedup primarily for large-dimensional matrix multiplications, while inherent quantization overheads diminish speedup when applied to low-rank adaptation (LoRA), which uses small-dimensional matrices for efficient fine-tuning of large language models (LLMs). To address this limitation, we propose FALQON, a novel framework that eliminates the quantization overhead from separate LoRA computational paths by directly merging LoRA adapters into an FP8-quantized backbone during fine-tuning. Furthermore, we reformulate the forward and backward computations for merged adapters to significantly reduce quantization overhead, and introduce a row-wise proxy update mechanism that efficiently integrates substantial updates into the quantized backbone. Experimental evaluations demonstrate that FALQON achieves approximately a 3$\\times$ training speedup over existing quantized LoRA methods with a similar level of accuracy, providing a practical solution for efficient large-scale model fine-tuning. Moreover, FALQON's end-to-end FP8 workflow removes the need for post-training quantization, facilitating efficient deployment. Code is available at https://github.com/iamkanghyunchoi/falqon.", "authors": ["Kanghyun Choi", "Hyeyoon Lee", "SunJong Park", "Dain Kwon", "Jinho Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-28", "url": "https://arxiv.org/abs/2510.24061", "pdf_url": "https://arxiv.org/pdf/2510.24061v1", "arxiv_id": "2510.24061", "doi": "10.48550/arXiv.2510.24061", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/iamkanghyunchoi/falqon", "venue": "arXiv.org", "quality_score": 0.5472}
{"id": "e3fb5dd9d5a2e0c3b0393b0497b55f4cea3e831dc3a61ff014cac7208cd8461d", "sources": ["arxiv", "semantic_scholar"], "title": "Block-Diagonal LoRA for Eliminating Communication Overhead in Tensor Parallel LoRA Serving", "abstract": "When serving a single base LLM with several different LoRA adapters simultaneously, the adapters cannot simply be merged with the base model's weights as the adapter swapping would create overhead and requests using different adapters could not be batched. Rather, the LoRA computations have to be separated from the base LLM computations, and in a multi-device setup the LoRA adapters can be sharded in a way that is well aligned with the base model's tensor parallel execution, as proposed in S-LoRA. However, the S-LoRA sharding strategy encounters some communication overhead, which may be small in theory, but can be large in practice. In this paper, we propose to constrain certain LoRA factors to be block-diagonal, which allows for an alternative way of sharding LoRA adapters that does not require any additional communication for the LoRA computations. We demonstrate in extensive experiments that our block-diagonal LoRA approach is similarly parameter efficient as standard LoRA (i.e., for a similar number of parameters it achieves similar downstream performance) and that it leads to significant end-to-end speed-up over S-LoRA. For example, when serving on eight A100 GPUs, we observe up to 1.79x (1.23x) end-to-end speed-up with 0.87x (1.74x) the number of adapter parameters for Llama-3.1-70B, and up to 1.63x (1.3x) end-to-end speed-up with 0.86x (1.73x) the number of adapter parameters for Llama-3.1-8B.", "authors": ["Xinyu Wang", "Jonas M. Kübler", "Kailash Budhathoki", "Yida Wang", "Matthäus Kleindessner"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-27", "url": "https://arxiv.org/abs/2510.23346", "pdf_url": "https://arxiv.org/pdf/2510.23346v2", "arxiv_id": "2510.23346", "doi": "10.48550/arXiv.2510.23346", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3529}
{"id": "4a0cb0bb579d09fe5c6d776fd4286498a3324aa19fdaffc29f862714ce231712", "sources": ["arxiv", "semantic_scholar"], "title": "Jarvis: Towards Personalized AI Assistant via Personal KV-Cache Retrieval", "abstract": "The rapid development of Vision-language models (VLMs) enables open-ended perception and reasoning. Recent works have started to investigate how to adapt general-purpose VLMs into personalized assistants. Even commercial models such as ChatGPT now support model personalization by incorporating user-specific information. However, existing methods either learn a set of concept tokens or train a VLM to utilize user-specific information. However, both pipelines struggle to generate accurate answers as personalized assistants. We introduce Jarvis, an innovative framework for a personalized AI assistant through personal KV-Cache retrieval, which stores user-specific information in the KV-Caches of both textual and visual tokens. The textual tokens are created by summarizing user information into metadata, while the visual tokens are produced by extracting distinct image patches from the user's images. When answering a question, Jarvis first retrieves related KV-Caches from personal storage and uses them to ensure accuracy in responses. We also introduce a fine-grained benchmark built with the same distinct image patch mining pipeline, emphasizing accurate question answering based on fine-grained user-specific information. Jarvis is capable of providing more accurate responses, particularly when they depend on specific local details. Jarvis achieves state-of-the-art results in both visual question answering and text-only tasks across multiple datasets, indicating a practical path toward personalized AI assistants. The code and dataset will be released.", "authors": ["Binxiao Xu", "Junyu Feng", "Shaolin Lu", "Yulin Luo", "Shilin Yan", "Hao Liang", "Ming Lu", "Wentao Zhang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-26", "url": "https://arxiv.org/abs/2510.22765", "pdf_url": "https://arxiv.org/pdf/2510.22765v2", "arxiv_id": "2510.22765", "doi": "10.48550/arXiv.2510.22765", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3518}
{"id": "e3a9c820040e6819ff238e5ca22e8f11a801156d2f2c24ff4b2b9e81e8cf7654", "sources": ["arxiv", "semantic_scholar"], "title": "Sparsity and Superposition in Mixture of Experts", "abstract": "Mixture of Experts (MoE) models have become central to scaling large language models, yet their mechanistic differences from dense networks remain poorly understood. Previous work has explored how dense models use \\textit{superposition} to represent more features than dimensions, and how superposition is a function of feature sparsity and feature importance. MoE models cannot be explained mechanistically through the same lens. We find that neither feature sparsity nor feature importance cause discontinuous phase changes, and that network sparsity (the ratio of active to total experts) better characterizes MoEs. We develop new metrics for measuring superposition across experts. Our findings demonstrate that models with greater network sparsity exhibit greater \\emph{monosemanticity}. We propose a new definition of expert specialization based on monosemantic feature representation rather than load balancing, showing that experts naturally organize around coherent feature combinations when initialized appropriately. These results suggest that network sparsity in MoEs may enable more interpretable models without sacrificing performance, challenging the common assumption that interpretability and capability are fundamentally at odds.", "authors": ["Marmik Chaudhari", "Jeremi Nuer", "Rome Thorstenson"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-26", "url": "https://arxiv.org/abs/2510.23671", "pdf_url": "https://arxiv.org/pdf/2510.23671v2", "arxiv_id": "2510.23671", "doi": "10.48550/arXiv.2510.23671", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3518}
{"id": "e73fa4630b5a128a51c06db3f1860a8d991ff5ab795cd8f7e289229cc4715b97", "sources": ["arxiv", "semantic_scholar"], "title": "SABlock: Semantic-Aware KV Cache Eviction with Adaptive Compression Block Size", "abstract": "The growing memory footprint of the Key-Value (KV) cache poses a severe scalability bottleneck for long-context Large Language Model (LLM) inference. While KV cache eviction has emerged as an effective solution by discarding less critical tokens, existing token-, block-, and sentence-level compression methods struggle to balance semantic coherence and memory efficiency. To this end, we introduce SABlock, a \\underline{s}emantic-aware KV cache eviction framework with \\underline{a}daptive \\underline{block} sizes. Specifically, SABlock first performs semantic segmentation to align compression boundaries with linguistic structures, then applies segment-guided token scoring to refine token importance estimation. Finally, for each segment, a budget-driven search strategy adaptively determines the optimal block size that preserves semantic integrity while improving compression efficiency under a given cache budget. Extensive experiments on long-context benchmarks demonstrate that SABlock consistently outperforms state-of-the-art baselines under the same memory budgets. For instance, on Needle-in-a-Haystack (NIAH), SABlock achieves 99.9% retrieval accuracy with only 96 KV entries, nearly matching the performance of the full-cache baseline that retains up to 8K entries. Under a fixed cache budget of 1,024, SABlock further reduces peak memory usage by 46.28% and achieves up to 9.5x faster decoding on a 128K context length.", "authors": ["Jinhan Chen", "Jianchun Liu", "Hongli Xu", "Xianjun Gao", "Shilong Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-26", "url": "https://arxiv.org/abs/2510.22556", "pdf_url": "https://arxiv.org/pdf/2510.22556v1", "arxiv_id": "2510.22556", "doi": "10.48550/arXiv.2510.22556", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3518}
{"id": "22a9b5817999e71fff757d59912134cc45decad04c1b92d71dc6403f0a685f30", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Low Rank Attention for Long-Context Inference in Large Language Models", "abstract": "As the length of input text increases, the key-value (KV) cache in LLMs imposes prohibitive GPU memory costs and limits long-context inference on resource constrained devices. Existing approaches, such as KV quantization and pruning, reduce memory usage but suffer from numerical precision loss or suboptimal retention of key-value pairs. In this work, Low Rank Query and Key attention (LRQK) is introduced, a two-stage framework that jointly decomposes full-precision query and key matrices into compact rank-\\(r\\) factors during the prefill stage, and then employs these low-dimensional projections to compute proxy attention scores in \\(\\mathcal{O}(lr)\\) time at each decode step. By selecting only the top-\\(k\\) tokens and a small fixed set of recent tokens, LRQK employs a mixed GPU-CPU cache with a hit-and-miss mechanism where only missing full-precision KV pairs are transferred, thereby preserving exact attention outputs while reducing CPU-GPU data movement. Extensive experiments on the RULER and LongBench benchmarks with LLaMA-3-8B and Qwen2.5-7B demonstrate that LRQK matches or surpasses leading sparse-attention methods in long context settings, while delivering significant memory savings with minimal accuracy loss. Our code is available at https://github.com/tenghuilee/LRQK.", "authors": ["Tenghui Li", "Guoxu Zhou", "Xuyang Zhao", "Yuning Qiu", "Qibin Zhao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-25", "url": "https://arxiv.org/abs/2510.23649", "pdf_url": "https://arxiv.org/pdf/2510.23649v3", "arxiv_id": "2510.23649", "doi": "10.48550/arXiv.2510.23649", "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/tenghuilee/LRQK", "venue": "arXiv.org", "quality_score": 0.5419}
{"id": "038ee918fc357b068b811716d6fccf12fb06fe91780ee4e03de92b49c482c69e", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning and Quantization Impact on Graph Neural Networks", "abstract": "Graph neural networks (GNNs) are known to operate with high accuracy on learning from graph-structured data, but they suffer from high computational and resource costs. Neural network compression methods are used to reduce the model size while maintaining reasonable accuracy. Two of the common neural network compression techniques include pruning and quantization. In this research, we empirically examine the effects of three pruning methods and three quantization methods on different GNN models, including graph classification tasks, node classification tasks, and link prediction. We conducted all experiments on three graph datasets, including Cora, Proteins, and BBBP. Our findings demonstrate that unstructured fine-grained and global pruning can significantly reduce the model's size(50\\%) while maintaining or even improving precision after fine-tuning the pruned model. The evaluation of different quantization methods on GNN shows diverse impacts on accuracy, inference time, and model size across different datasets.", "authors": ["Khatoon Khedri", "Reza Rawassizadeh", "Qifu Wen", "Mehdi Hosseinzadeh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-24", "url": "https://arxiv.org/abs/2510.22058", "pdf_url": "https://arxiv.org/pdf/2510.22058v1", "arxiv_id": "2510.22058", "doi": "10.48550/arXiv.2510.22058", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "086412055d33fa9098139d661e3b633d7ca39986687e797bb6daf869a779ec4c", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Task Vehicle Routing Solver via Mixture of Specialized Experts under State-Decomposable MDP", "abstract": "Existing neural methods for multi-task vehicle routing problems (VRPs) typically learn unified solvers to handle multiple constraints simultaneously. However, they often underutilize the compositional structure of VRP variants, each derivable from a common set of basis VRP variants. This critical oversight causes unified solvers to miss out the potential benefits of basis solvers, each specialized for a basis VRP variant. To overcome this limitation, we propose a framework that enables unified solvers to perceive the shared-component nature across VRP variants by proactively reusing basis solvers, while mitigating the exponential growth of trained neural solvers. Specifically, we introduce a State-Decomposable MDP (SDMDP) that reformulates VRPs by expressing the state space as the Cartesian product of basis state spaces associated with basis VRP variants. More crucially, this formulation inherently yields the optimal basis policy for each basis VRP variant. Furthermore, a Latent Space-based SDMDP extension is developed by incorporating both the optimal basis policies and a learnable mixture function to enable the policy reuse in the latent space. Under mild assumptions, this extension provably recovers the optimal unified policy of SDMDP through the mixture function that computes the state embedding as a mapping from the basis state embeddings generated by optimal basis policies. For practical implementation, we introduce the Mixture-of-Specialized-Experts Solver (MoSES), which realizes basis policies through specialized Low-Rank Adaptation (LoRA) experts, and implements the mixture function via an adaptive gating mechanism. Extensive experiments conducted across VRP variants showcase the superiority of MoSES over prior methods.", "authors": ["Yuxin Pan", "Zhiguang Cao", "Chengyang Gu", "Liu Liu", "Peilin Zhao", "Yize Chen", "Fangzhen Lin"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-24", "url": "https://arxiv.org/abs/2510.21453", "pdf_url": "https://arxiv.org/pdf/2510.21453v1", "arxiv_id": "2510.21453", "doi": "10.48550/arXiv.2510.21453", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "8f44d8334ccdc0b6c3015b182133f7ba8691e692898c98b596e87af23dc60aea", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Graph Mixture of Residual Experts: Unsupervised Learning on Diverse Graphs with Heterogeneous Specialization", "abstract": "Graph Neural Networks (GNNs) face a fundamental adaptability challenge: their fixed message-passing architectures struggle with the immense diversity of real-world graphs, where optimal computational strategies vary by local structure and task. While Mixture-of-Experts (MoE) offers a promising pathway to adaptability, existing graph MoE methods remain constrained by their reliance on supervised signals and instability when training heterogeneous experts. We introduce ADaMoRE (Adaptive Mixture of Residual Experts), a principled framework that enables robust, fully unsupervised training of heterogeneous MoE on graphs. ADaMoRE employs a backbone-residual expert architecture where foundational encoders provide stability while specialized residual experts capture diverse computational patterns. A structurally-aware gating network performs fine-grained node routing. The entire architecture is trained end-to-end using a unified unsupervised objective, which integrates a primary reconstruction task with an information-theoretic diversity regularizer to explicitly enforce functional specialization among the experts. Theoretical analysis confirms our design improves data efficiency and training stability. Extensive evaluation across 16 benchmarks validates ADaMoRE's state-of-the-art performance in unsupervised node classification and few-shot learning, alongside superior generalization, training efficiency, and faster convergence on diverse graphs and tasks.", "authors": ["Yunlong Chu", "Minglai Shao", "Zengyi Wo", "Bing Hao", "Yuhang Liu", "Ruijie Wang", "Jianxin Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-24", "url": "https://arxiv.org/abs/2510.21207", "pdf_url": "https://arxiv.org/pdf/2510.21207v1", "arxiv_id": "2510.21207", "doi": "10.48550/arXiv.2510.21207", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "ede646babbd934fd5ac95ae033857f21db9552f51d2af1ee9bde941edd0ee1fe", "sources": ["arxiv", "semantic_scholar"], "title": "Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models", "abstract": "Recent large vision-language models (LVLMs) demonstrate remarkable capabilities in processing extended multi-modal sequences, yet the resulting key-value (KV) cache expansion creates a critical memory bottleneck that fundamentally limits deployment scalability. While existing KV cache compression methods focus on retaining high-importance KV pairs to minimize storage, they often overlook the modality-specific semantic redundancy patterns that emerge distinctively in multi-modal KV caches. In this work, we first analyze how, beyond simple importance, the KV cache in LVLMs exhibits varying levels of redundancy across attention heads. We show that relying solely on importance can only cover a subset of the full KV cache information distribution, leading to potential loss of semantic coverage. To address this, we propose MixKV, a novel method that mixes importance with diversity for optimized KV cache compression in LVLMs. MixKV adapts to head-wise semantic redundancy, selectively balancing diversity and importance when compressing KV pairs. Extensive experiments demonstrate that MixKV consistently enhances existing methods across multiple LVLMs. Under extreme compression (budget=64), MixKV improves baseline methods by an average of 5.1% across five multi-modal understanding benchmarks and achieves remarkable gains of 8.0% and 9.0% for SnapKV and AdaKV on GUI grounding tasks, all while maintaining comparable inference efficiency. Furthermore, MixKV extends seamlessly to LLMs with comparable performance gains. Our code is available at https://github.com/xuyang-liu16/MixKV.", "authors": ["Xuyang Liu", "Xiyan Gui", "Yuchao Zhang", "Linfeng Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-23", "url": "https://arxiv.org/abs/2510.20707", "pdf_url": "https://arxiv.org/pdf/2510.20707v2", "arxiv_id": "2510.20707", "doi": "10.48550/arXiv.2510.20707", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/xuyang-liu16/MixKV", "venue": "arXiv.org", "quality_score": 0.5383}
{"id": "58293fed3069df40a7b1d36a7110d7a724aa1d07c93e2255ed2575f370244616", "sources": ["arxiv", "semantic_scholar"], "title": "Stateful KV Cache Management for LLMs: Balancing Space, Time, Accuracy, and Positional Fidelity", "abstract": "The Key-Value (KV) cache is integral to efficient autoregressive inference in large language models (LLMs), yet its unbounded growth in stateful multi-turn scenarios presents major challenges. This paper examines the interplay between KV cache management strategies, the architectural context limits of models like meta-llama/Meta-Llama-3-8b-instruct, and the often-overlooked integrity of positional encodings. Through empirical analysis using a stateful benchmarking framework, we show that LLM generation quality degrades sharply when the accumulated KV cache approaches or exceeds the model's trained context window (e.g., 8192 tokens for Llama 3), a failure mode distinct from GPU memory exhaustion. Common eviction strategies, even high-retention ones (e.g., 99% via AttentionTop), can worsen performance if they disrupt positional coherence. Because LLMs rely on consistent positional signals (e.g., RoPE), compacting a cache by removing non-contiguous tokens can scramble these signals and lead to degenerative outputs. We further show that simple strategies preserving contiguous context blocks (e.g., keeping an initial \"gist\") can yield more coherent generations than complex or positionally disruptive ones. We advocate for eviction techniques that respect architectural limits, preserve positional structure, and view \"cache health\" holistically beyond mere size.", "authors": ["Pratik Poudel"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-23", "url": "https://arxiv.org/abs/2511.04686", "pdf_url": "https://arxiv.org/pdf/2511.04686v1", "arxiv_id": "2511.04686", "doi": "10.48550/arXiv.2511.04686", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3483}
{"id": "84266fddf36f9d3b108145c14c582af9ed8da3d2320ef05c0799c59930b2641d", "sources": ["arxiv", "semantic_scholar"], "title": "PruneHal: Reducing Hallucinations in Multi-modal Large Language Models through Adaptive KV Cache Pruning", "abstract": "While multi-modal large language models (MLLMs) have made significant progress in recent years, the issue of hallucinations remains a major challenge. To mitigate this phenomenon, existing solutions either introduce additional data for further training or incorporate external or internal information during inference. However, these approaches inevitably introduce extra computational costs. In this paper, we observe that hallucinations in MLLMs are strongly associated with insufficient attention allocated to visual tokens. In particular, the presence of redundant visual tokens disperses the model's attention, preventing it from focusing on the most informative ones. As a result, critical visual cues are often under-attended, which in turn exacerbates the occurrence of hallucinations. Building on this observation, we propose \\textbf{PruneHal}, a training-free, simple yet effective method that leverages adaptive KV cache pruning to enhance the model's focus on critical visual information, thereby mitigating hallucinations. To the best of our knowledge, we are the first to apply token pruning for hallucination mitigation in MLLMs. Notably, our method don't require additional training and incurs nearly no extra inference cost. Moreover, PruneHal is model-agnostic and can be seamlessly integrated with different decoding strategies, including those specifically designed for hallucination mitigation. We evaluate PruneHal on several widely used hallucination evaluation benchmarks using four mainstream MLLMs, achieving robust and outstanding results that highlight the effectiveness and superiority of our method. Our code will be publicly available.", "authors": ["Fengyuan Sun", "Hui Chen", "Xinhao Xu", "Dandan Zheng", "Jingdong Chen", "Jun Zhou", "Jungong Han", "Guiguang Ding"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-22", "url": "https://arxiv.org/abs/2510.19183", "pdf_url": "https://arxiv.org/pdf/2510.19183v1", "arxiv_id": "2510.19183", "doi": "10.48550/arXiv.2510.19183", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3472}
{"id": "931a2a5dd445188983c1e9800df854cd09a854647a660f8c0a878f7d23b3e11c", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-GS: Mixture of Experts for Dynamic Gaussian Splatting", "abstract": "Recent advances in dynamic scene reconstruction have significantly benefited from 3D Gaussian Splatting, yet existing methods show inconsistent performance across diverse scenes, indicating no single approach effectively handles all dynamic challenges. To overcome these limitations, we propose Mixture of Experts for Dynamic Gaussian Splatting (MoE-GS), a unified framework integrating multiple specialized experts via a novel Volume-aware Pixel Router. Unlike sparsity-oriented MoE architectures in large language models, MoE-GS is designed to improve dynamic novel view synthesis quality by combining heterogeneous deformation priors, rather than to reduce training or inference-time FLOPs. Our router adaptively blends expert outputs by projecting volumetric Gaussian-level weights into pixel space through differentiable weight splatting, ensuring spatially and temporally coherent results. Although MoE-GS improves rendering quality, the increased model capacity and reduced FPS are inherent to the MoE architecture. To mitigate this, we explore two complementary directions: (1) single-pass multi-expert rendering and gate-aware Gaussian pruning, which improve efficiency within the MoE framework, and (2) a distillation strategy that transfers MoE performance to individual experts, enabling lightweight deployment without architectural changes. To the best of our knowledge, MoE-GS is the first approach incorporating Mixture-of-Experts techniques into dynamic Gaussian splatting. Extensive experiments on the N3V and Technicolor datasets demonstrate that MoE-GS consistently outperforms state-of-the-art methods with improved efficiency. Video demonstrations are available at cvsp-lab.github.io/MoE-GS.", "authors": ["In-Hwan Jin", "Hyeongju Mun", "Joonsoo Kim", "Kugjin Yun", "Kyeongbo Kong"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-22", "url": "https://arxiv.org/abs/2510.19210", "pdf_url": "https://arxiv.org/pdf/2510.19210v2", "arxiv_id": "2510.19210", "doi": "10.48550/arXiv.2510.19210", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3472}
{"id": "d39c380c75d5b392ed4e0fb9e7519ac7be4840d5b6a964e3a3e26627d7e00933", "sources": ["arxiv", "semantic_scholar"], "title": "TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications", "abstract": "Large Language Models (LLMs) are increasingly deployed in complex multi-agent applications that rely on external function calls. This workload creates severe performance challenges for the KV Cache: spatial contention leads to the eviction of critical agents' caches and temporal underutilization leaves the cache of agents stalled on long-running function calls idling in GPU memory. We present TokenCake, a KV-Cache-centric serving framework that bridges this gap by co-optimizing scheduling and memory management through an agent-aware design. TokenCake's Temporal Scheduler employs an event-driven, opportunistic policy to proactively offload idle KV Caches during function calls and uses predictive uploading to hide data transfer latency. TokenCake's Spatial Scheduler uses dynamic memory partitioning, guided by a hybrid priority metric combining graph structure and runtime state, to reserve GPU memory for critical-path agents. Our evaluation on representative multi-agent benchmarks shows that TokenCake reduces end-to-end latency by over 47.06% and improves effective GPU memory utilization by up to 16.9% compared to vLLM.", "authors": ["Zhuohang Bian", "Feiyang Wu", "Zhuoran Li", "Teng Ma", "Youwei Zhuo"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-21", "url": "https://arxiv.org/abs/2510.18586", "pdf_url": "https://arxiv.org/pdf/2510.18586v3", "arxiv_id": "2510.18586", "doi": "10.48550/arXiv.2510.18586", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.346}
{"id": "55eb100a6d6958df14726ab9c8382661eb1c42253e60fb7b47551eca89306fc6", "sources": ["arxiv", "semantic_scholar"], "title": "Ensembling Pruned Attention Heads For Uncertainty-Aware Efficient Transformers", "abstract": "Uncertainty quantification (UQ) is essential for deploying deep neural networks in safety-critical settings. Although methods like Deep Ensembles achieve strong UQ performance, their high computational and memory costs hinder scalability to large models. We introduce Hydra Ensembles, an efficient transformer-based ensemble that prunes attention heads to create diverse members and merges them via a new multi-head attention with grouped fully-connected layers. This yields a compact model with inference speed close to a single network, matching or surpassing Deep Ensembles in UQ performance without retraining from scratch. We also provide an in-depth analysis of pruning, showing that naive approaches can harm calibration, whereas Hydra Ensembles preserves robust uncertainty. Experiments on image and text classification tasks, with various architectures, show consistent gains over Deep Ensembles. Remarkably, in zero-shot classification on ImageNet-1k, our approach surpasses state of the art methods, even without requiring additional training.", "authors": ["Firas Gabetni", "Giuseppe Curci", "Andrea Pilzer", "Subhankar Roy", "Elisa Ricci", "Gianni Franchi"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-21", "url": "https://arxiv.org/abs/2510.18358", "pdf_url": "https://arxiv.org/pdf/2510.18358v2", "arxiv_id": "2510.18358", "doi": "10.48550/arXiv.2510.18358", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.346}
{"id": "af317171f31a701a31a7feb1cb798a10e005882cdb769546c0298617d98a88ff", "sources": ["arxiv", "semantic_scholar"], "title": "ReXMoE: Reusing Experts with Minimal Overhead in Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) architectures have emerged as a promising approach to scale Large Language Models (LLMs). MoE boosts the efficiency by activating a subset of experts per token. Recent works show that fine-grained experts substantially enriches the combinatorial flexibility of active experts and enhances model expressiveness. However, such a design is fundamentally limited by the layer-local routing mechanism: each layer is restricted to its own expert pool. This requires a careful trade-off between expert dimensionality and routing diversity given fixed parameter budgets. We describe ReXMoE, a novel MoE architecture that improves routing beyond the existing layer-local approaches by allowing routers to reuse experts across adjacent layers. ReXMoE decouples expert dimensionality from per-layer budgets, enabling richer expert combinations without sacrificing individual expert capacity or inflating overall parameters. To this end, we propose a new progressive scaling routing (PSR) strategy to gradually increase the candidate expert pool during training. As a result, ReXMoE improves both language modeling and downstream task performance. Extensive experiments on models ranging from 0.5B to 7B parameters across different architectures demonstrate that ReXMoE consistently improves performance under fixed architectural dimensions, confirming ReXMoE as new design paradigm for parameter-efficient and scalable MoE-based LLMs.", "authors": ["Zheyue Tan", "Zhiyuan Li", "Tao Yuan", "Dong Zhou", "Weilin Liu", "Yueqing Zhuang", "Yadong Li", "Guowei Niu", "Cheng Qin", "Zhuyu Yao", "Congyi Liu", "Haiyang Xu", "Boxun Li", "Guohao Dai", "Bo Zhao", "Yu Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-20", "url": "https://arxiv.org/abs/2510.17483", "pdf_url": "https://arxiv.org/pdf/2510.17483v1", "arxiv_id": "2510.17483", "doi": "10.48550/arXiv.2510.17483", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3449}
{"id": "12b74bca580902e2a4b23c50795a0263535c18acd1fafb02e37b3229ebfc22e5", "sources": ["arxiv", "semantic_scholar"], "title": "MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation", "abstract": "Active Surveillance (AS) is a treatment option for managing low and intermediate-risk prostate cancer (PCa), aiming to avoid overtreatment while monitoring disease progression through serial MRI and clinical follow-up. Accurate prostate segmentation is an important preliminary step for automating this process, enabling automated detection and diagnosis of PCa. However, existing deep-learning segmentation models are often trained on single-time-point and expertly annotated datasets, making them unsuitable for longitudinal AS analysis, where multiple time points and a scarcity of expert labels hinder their effective fine-tuning. To address these challenges, we propose MambaX-Net, a novel semi-supervised, dual-scan 3D segmentation architecture that computes the segmentation for time point t by leveraging the MRI and the corresponding segmentation mask from the previous time point. We introduce two new components: (i) a Mamba-enhanced Cross-Attention Module, which integrates the Mamba block into cross attention to efficiently capture temporal evolution and long-range spatial dependencies, and (ii) a Shape Extractor Module that encodes the previous segmentation mask into a latent anatomical representation for refined zone delination. Moreover, we introduce a semi-supervised self-training strategy that leverages pseudo-labels generated from a pre-trained nnU-Net, enabling effective learning without expert annotations. MambaX-Net was evaluated on a longitudinal AS dataset, and results showed that it significantly outperforms state-of-the-art U-Net and Transformer-based models, achieving superior prostate zone segmentation even when trained on limited and noisy data.", "authors": ["Yovin Yahathugoda", "Davide Prezzi", "Piyalitt Ittichaiwong", "Vicky Goh", "Sebastien Ourselin", "Michela Antonelli"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-20", "url": "https://arxiv.org/abs/2510.17529", "pdf_url": "https://arxiv.org/pdf/2510.17529v2", "arxiv_id": "2510.17529", "doi": "10.48550/arXiv.2510.17529", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3449}
{"id": "a462a993ba6195ac3387c023cdbdb1a72ab352d1c45f4c496651ba16a7abf89d", "sources": ["arxiv", "semantic_scholar"], "title": "Learned Inertial Odometry for Cycling Based on Mixture of Experts Algorithm", "abstract": "With the rapid growth of bike sharing and the increasing diversity of cycling applications, accurate bicycle localization has become essential. traditional GNSS-based methods suffer from multipath effects, while existing inertial navigation approaches rely on precise modeling and show limited robustness. Tight Learned Inertial Odometry (TLIO) achieves low position drift by combining raw IMU data with predicted displacements by neural networks, but its high computational cost restricts deployment on mobile devices. To overcome this, we extend TLIO to bicycle localization and introduce an improved Mixture-of Experts (MoE) model that reduces both training and inference costs. Experiments show that, compared to the state-of-the-art LLIO framework, our method achieves comparable accuracy while reducing parameters by 64.7% and computational cost by 81.8%.", "authors": ["Hao Qiao", "Yan Wang", "Shuo Yang", "Xiaoyao Yu", "Jian kuang", "Xiaoji Niu"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-20", "url": "https://arxiv.org/abs/2510.17604", "pdf_url": "https://arxiv.org/pdf/2510.17604v1", "arxiv_id": "2510.17604", "doi": "10.48550/arXiv.2510.17604", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3449}
{"id": "b7a0950ed6e315642fe010f1a5099a99552cd3d4f94bd846b3d093276279a44a", "sources": ["arxiv", "semantic_scholar"], "title": "L-MoE: End-to-End Training of a Lightweight Mixture of Low-Rank Adaptation Experts", "abstract": "The Mixture of Experts (MoE) architecture enables the scaling of Large Language Models (LLMs) to trillions of parameters by activating a sparse subset of weights for each input, maintaining constant computational cost during inference. Concurrently, Low-Rank Adaptation (LoRA) has emerged as a dominant technique for parameter-efficiently fine-tuning LLMs on specialized tasks. In this work, we unify these two paradigms into a novel, end-to-end trainable framework named L-MoE: a Lightweight Mixture of LoRA Experts. L-MoE redefines MoE experts not as dense feed-forward networks, but as a collection of task-specialized, low-rank adapters. A lightweight gating network, trained jointly with the experts, learns to dynamically compose these LoRA adapters by computing a weighted average of their parameters for each input token. This composition is fully differentiable, allowing gradients from a standard auto-regressive language modeling objective to flow back through the entire architecture, simultaneously refining both the expert adapters and the routing strategy. This approach creates a highly parameter-efficient MoE model that is modular by design, allows for dynamic skill composition, and is trainable from end-to-end. We present the formal mathematical framework for L-MoE, detailing the differentiable routing mechanism and the joint optimization objective, thereby providing a new path toward building more efficient, scalable, and specialized language models.", "authors": ["Shihao Ji", "Zihui Song"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-19", "url": "https://arxiv.org/abs/2510.17898", "pdf_url": "https://arxiv.org/pdf/2510.17898v2", "arxiv_id": "2510.17898", "doi": "10.48550/arXiv.2510.17898", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3438}
{"id": "820924cf422d00a74a255f1f0d7a35a36e63dad2132c3468ad997d5d14ba0b6f", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Model Representation and Reducing KV Cache via Skip Connections with First Value Heads", "abstract": "Transformer models have driven breakthroughs across various language tasks by their strong capability to learn rich contextual representations. Scaling them to improve representation, however, often demands substantial memory and compute costs, such as the Key-Value (KV) cache used during auto-regressive decoding. Skip connections offer a promising way to improve representation without bloating resource usage, yet most prior works either improve expressivity while leaving KV costs unchanged, or reduce memory at the cost of weaker representation. In this work, we propose SkipV1Former, a Transformer variant that uses skip connections from the first layer's Value heads to strengthen model representation and reduce KV cache. Specifically, from the second block onward, each layer reuses half of its Value heads from the very first layer, while computing the other half as usual-cutting Value projections and V cache by nearly 50 \\%. Theoretically, we show that routing uncompressed first-layer Values into deeper layers restores information lost to compression and accelerates the model's implicit mesa-optimization-a key pattern of Transformer in auto-regressive tasks. Empirically, across different model scales, SkipV1Former delivers consistent reductions of approximately 25 \\% in KV cache while improving perplexity relative to standard Multi-Head Attention (MHA) Transformers and some advanced variants. Moreover, we propose a recipe for uptraining existing MHA Transformer checkpoints to SkipV1Former with only 10-15\\% additional compute. Finally, SkipV1Former can seamlessly combine advanced methods like Group-Query Attention and Multi-Latent Attention to achieve further KV cache savings and performance improvement. When combined with YOCO, it cuts KV cache size by nearly 50 \\% while still improving performance.", "authors": ["Zhoutong Wu", "Yuan Zhang", "Yiming Dong", "Chenheng Zhang", "Cong Fang", "Kun Yuan", "Zhouchen Lin"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-19", "url": "https://arxiv.org/abs/2510.16807", "pdf_url": "https://arxiv.org/pdf/2510.16807v2", "arxiv_id": "2510.16807", "doi": "10.48550/arXiv.2510.16807", "citation_count": 4, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Zhoutong-Wu/SkipV1Former}", "venue": "arXiv.org", "quality_score": 0.5312}
{"id": "97f54b8fd675c8d89ce02fbffdc05a22449fd42c692ce0d4b8933ffa211c0324", "sources": ["arxiv", "semantic_scholar"], "title": "Online Mixture of Experts: No-Regret Learning for Optimal Collective Decision-Making", "abstract": "We explore the use of expert-guided bandit learning, which we refer to as online mixture-of-experts (OMoE). In this setting, given a context, a candidate committee of experts must determine how to aggregate their outputs to achieve optimal results in terms of aggregate accuracy. We propose two algorithms to address this problem. The first algorithm combines aggregate voting with UCB-driven successive elimination, efficiently pruning suboptimal exploration actions. The second algorithm employs an online weighted-majority-voting mechanism, leveraging the respective voting power of each expert proportional to their predictive power. We derive theoretical guarantees for the regret properties in the bandit setting under ideal circumstances, and empirical results are provided accordingly. As a modern study on applications, these methods are applied to the online fine-tuning of a set of expert large language models (LLMs), where after each response, the generative LLM dynamically reweighs its set of experts and/or selects the optimal committee of experts to generate the most accurate response. Our results introduce new methodologies and no-regret guarantees for combining multiple experts to improve on the performance of the an aggregate model overall.", "authors": ["Larkin Liu", "Jalal Etesami"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-19", "url": "https://arxiv.org/abs/2510.21788", "pdf_url": "https://arxiv.org/pdf/2510.21788v2", "arxiv_id": "2510.21788", "doi": "10.48550/arXiv.2510.21788", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3438}
{"id": "a8db6cdee50c9d137ee2b7c740aa7d11ecb3275603618d9b08b64a5acdf744b6", "sources": ["arxiv", "semantic_scholar"], "title": "Schrödinger Bridge Mamba for One-Step Speech Enhancement", "abstract": "We present Schrödinger Bridge Mamba (SBM), a novel model for efficient speech enhancement by integrating the Schrödinger Bridge (SB) training paradigm and the Mamba architecture. Experiments of joint denoising and dereverberation tasks demonstrate SBM outperforms strong generative and discriminative methods on multiple metrics with only one step of inference while achieving a competitive real-time factor for streaming feasibility. Ablation studies reveal that the SB paradigm consistently yields improved performance across diverse architectures over conventional mapping. Furthermore, Mamba exhibits a stronger performance under the SB paradigm compared to Multi-Head Self-Attention (MHSA) and Long Short-Term Memory (LSTM) backbones. These findings highlight the synergy between the Mamba architecture and the SB trajectory-based training, providing a high-quality solution for real-world speech enhancement. Demo page: https://sbmse.github.io", "authors": ["Jing Yang", "Sirui Wang", "Chao Wu", "Lei Guo", "Fan Fan"], "categories": ["cs.SD", "cs.AI", "cs.LG", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-10-19", "url": "https://arxiv.org/abs/2510.16834", "pdf_url": "https://arxiv.org/pdf/2510.16834v2", "arxiv_id": "2510.16834", "doi": "10.48550/arXiv.2510.16834", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3438}
{"id": "e5da4ed5302283dbc91c0c7f4886d9d62b16c257ec299bab8bf7af43993aeda1", "sources": ["arxiv", "semantic_scholar"], "title": "Input Domain Aware MoE: Decoupling Routing Decisions from Task Optimization in Mixture of Experts", "abstract": "Sparse Mixture of Experts (sMoE) has become a pivotal approach for scaling large vision-language models, offering substantial capacity while maintaining computational efficiency through dynamic, sparse activation of experts. However, existing routing mechanisms, typically based on similarity scoring, struggle to effectively capture the underlying input structure. This limitation leads to a trade-off between expert specialization and balanced computation, hindering both scalability and performance. We propose Input Domain Aware MoE, a novel routing framework that leverages a probabilistic mixture model to better partition the input space. By modeling routing probabilities as a mixture of distributions, our method enables experts to develop clear specialization boundaries while achieving balanced utilization. Unlike conventional approaches, our routing mechanism is trained independently of task-specific objectives, allowing for stable optimization and decisive expert assignments. Empirical results on vision-language tasks demonstrate that our method consistently outperforms existing sMoE approaches, achieving higher task performance and improved expert utilization balance.", "authors": ["Yongxiang Hua", "Haoyu Cao", "Zhou Tao", "Bocheng Li", "Zihao Wu", "Chaohu Liu", "Linli Xu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-18", "url": "https://arxiv.org/abs/2510.16448", "pdf_url": "https://arxiv.org/pdf/2510.16448v1", "arxiv_id": "2510.16448", "doi": "10.1145/3746027.3755754", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ACM Multimedia", "quality_score": 0.3426}
{"id": "afd1624e4774072deb075cc400fdb6950ff9a25dc9fc83129f3197937f0c5677", "sources": ["arxiv", "semantic_scholar"], "title": "Modeling Expert Interactions in Sparse Mixture of Experts via Graph Structures", "abstract": "Sparse Mixture of Experts (SMoE) has emerged as a promising solution to achieving unparalleled scalability in deep learning by decoupling model parameter count from computational cost. By activating only a small subset of parameters per sample, SMoE enables significant growth in model capacity while maintaining efficiency. However, SMoE struggles to adapt to distributional shifts, leading to reduced robustness under data contamination. In this work, we introduce SymphonySMoE, a novel family of SMoE that introduces a social graph to model interactions among experts. This graph-based structure enhances the token routing process, addressing the robustness challenges that are inherent in conventional SMoE designs. SymphonySMoE is lightweight, modular, and integrates seamlessly with existing SMoE-based models such as the XMoE and the Generalist Language Model. We provide both theoretical analysis and empirical evidence demonstrating SymphonySMoE's advantages over baseline SMoE. Extensive experiments on language modeling and visual instruction tuning validate our method's effectiveness. We further highlight the scalability of SymphonySMoE to models with 4.2 and 7.4 billion parameters, showcasing its applicability in fine-tuning tasks for large-scale systems.", "authors": ["Minh-Khoi Nguyen-Nhat", "Rachel S. Y. Teo", "Laziz Abdullaev", "Maurice Mok", "Viet-Hoang Tran", "Tan Minh Nguyen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-18", "url": "https://arxiv.org/abs/2510.16411", "pdf_url": "https://arxiv.org/pdf/2510.16411v1", "arxiv_id": "2510.16411", "doi": "10.48550/arXiv.2510.16411", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3426}
{"id": "cde562fbb4c913e83504097dee86bfe2f258961deefa653e83036aab4f1f1071", "sources": ["arxiv", "semantic_scholar"], "title": "Expert Merging in Sparse Mixture of Experts with Nash Bargaining", "abstract": "Existing expert merging strategies for Sparse Mixture of Experts (SMoE) typically rely on input-dependent or input-independent averaging of expert parameters, but often lack a principled weighting mechanism. In this work, we reinterpret expert merging through the lens of game theory, revealing cooperative and competitive dynamics among experts. Based on this perspective, we introduce Nash Merging of Experts (NAMEx), a novel framework that incorporates Nash Bargaining into the merging process, enabling more balanced and efficient collaboration among experts. Additionally, we incorporate complex momentum into NAMEx to accelerate expert propagation with theoretical guarantees for convergence. Extensive experiments across language modelling, text classification, image classification, and zero-shot robustness under data corruption show that NAMEx consistently outperforms competing methods while integrating seamlessly with popular MoE architectures. Finally, we demonstrate NAMEx's scalability by applying it to large-scale systems, including Qwen1.5-MoE (14B) and DeepSeek-MoE (16B), where it proves effective in both zero-shot and fine-tuning settings. The code is publicly available at: https://github.com/anh147/NAMEx.", "authors": ["Dung V. Nguyen", "Anh T. Nguyen", "Minh H. Nguyen", "Luc Q. Nguyen", "Shiqi Jiang", "Ethan Fetaya", "Linh Duy Tran", "Gal Chechik", "Tan M. Nguyen"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-17", "url": "https://arxiv.org/abs/2510.16138", "pdf_url": "https://arxiv.org/pdf/2510.16138v2", "arxiv_id": "2510.16138", "doi": "10.48550/arXiv.2510.16138", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/anh147/NAMEx", "venue": "arXiv.org", "quality_score": 0.5277}
{"id": "26d940015776b288659bae5856f4c2eb0b0a886a71f22fb14d8bac431351f3e3", "sources": ["arxiv", "semantic_scholar"], "title": "MTmixAtt: Integrating Mixture-of-Experts with Multi-Mix Attention for Large-Scale Recommendation", "abstract": "Industrial recommender systems critically depend on high-quality ranking models. However, traditional pipelines still rely on manual feature engineering and scenario-specific architectures, which hinder cross-scenario transfer and large-scale deployment. To address these challenges, we propose \\textbf{MTmixAtt}, a unified Mixture-of-Experts (MoE) architecture with Multi-Mix Attention, designed for large-scale recommendation tasks. MTmixAtt integrates two key components. The \\textbf{AutoToken} module automatically clusters heterogeneous features into semantically coherent tokens, removing the need for human-defined feature groups. The \\textbf{MTmixAttBlock} module enables efficient token interaction via a learnable mixing matrix, shared dense experts, and scenario-aware sparse experts, capturing both global patterns and scenario-specific behaviors within a single framework. Extensive experiments on the industrial TRec dataset from Meituan demonstrate that MTmixAtt consistently outperforms state-of-the-art baselines including Transformer-based models, WuKong, HiFormer, MLP-Mixer, and RankMixer. At comparable parameter scales, MTmixAtt achieves superior CTR and CTCVR metrics; scaling to MTmixAtt-1B yields further monotonic gains. Large-scale online A/B tests validate the real-world impact: in the \\textit{Homepage} scenario, MTmixAtt increases Payment PV by \\textbf{+3.62\\%} and Actual Payment GTV by \\textbf{+2.54\\%}. Overall, MTmixAtt provides a unified and scalable solution for modeling arbitrary heterogeneous features across scenarios, significantly improving both user experience and commercial outcomes.", "authors": ["Xianyang Qi", "Yuan Tian", "Zhaoyu Hu", "Zhirui Kuai", "Chang Liu", "Hongxiang Lin", "Lei Wang"], "categories": ["cs.IR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-17", "url": "https://arxiv.org/abs/2510.15286", "pdf_url": "https://arxiv.org/pdf/2510.15286v1", "arxiv_id": "2510.15286", "doi": "10.48550/arXiv.2510.15286", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3415}
{"id": "b44311c043fe5cf51f6371f88b67fd6ec4dcbfa0c1c8f071fa4bcdad84f8d5d2", "sources": ["arxiv", "semantic_scholar"], "title": "Backdoor or Manipulation? Graph Mixture of Experts Can Defend Against Various Graph Adversarial Attacks", "abstract": "Extensive research has highlighted the vulnerability of graph neural networks (GNNs) to adversarial attacks, including manipulation, node injection, and the recently emerging threat of backdoor attacks. However, existing defenses typically focus on a single type of attack, lacking a unified approach to simultaneously defend against multiple threats. In this work, we leverage the flexibility of the Mixture of Experts (MoE) architecture to design a scalable and unified framework for defending against backdoor, edge manipulation, and node injection attacks. Specifically, we propose an MI-based logic diversity loss to encourage individual experts to focus on distinct neighborhood structures in their decision processes, thus ensuring a sufficient subset of experts remains unaffected under perturbations in local structures. Moreover, we introduce a robustness-aware router that identifies perturbation patterns and adaptively routes perturbed nodes to corresponding robust experts. Extensive experiments conducted under various adversarial settings demonstrate that our method consistently achieves superior robustness against multiple graph adversarial attacks.", "authors": ["Yuyuan Feng", "Bin Ma", "Enyan Dai"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-17", "url": "https://arxiv.org/abs/2510.15333", "pdf_url": "https://arxiv.org/pdf/2510.15333v1", "arxiv_id": "2510.15333", "doi": "10.48550/arXiv.2510.15333", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3415}
{"id": "d26aea775ddf018982b4f67a5babf8c613ff862dbda79bd7a7c02568f16dbd06", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Is All You Need for KV Cache in Diffusion LLMs", "abstract": "This work studies how to adaptively recompute key-value (KV) caches for diffusion large language models (DLMs) to maximize prediction accuracy while minimizing decoding latency. Prior methods' decoders recompute QKV for all tokens at every denoising step and layer, despite KV states changing little across most steps, especially in shallow layers, leading to substantial redundancy. We make three observations: (1) distant ${\\bf MASK}$ tokens primarily act as a length-bias and can be cached block-wise beyond the active prediction window; (2) KV dynamics increase with depth, suggesting that selective refresh starting from deeper layers is sufficient; and (3) the most-attended token exhibits the smallest KV drift, providing a conservative lower bound on cache change for other tokens. Building on these, we propose ${\\bf Elastic-Cache}$, a training-free, architecture-agnostic strategy that jointly decides ${when}$ to refresh (via an attention-aware drift test on the most-attended token) and ${where}$ to refresh (via a depth-aware schedule that recomputes from a chosen layer onward while reusing shallow-layer caches and off-window MASK caches). Unlike fixed-period schemes, Elastic-Cache performs adaptive, layer-aware cache updates for diffusion LLMs, reducing redundant computation and accelerating decoding with negligible loss in generation quality. Experiments on LLaDA-Instruct, LLaDA-1.5, and LLaDA-V across mathematical reasoning and code generation tasks demonstrate consistent speedups: $8.7\\times$ on GSM8K (256 tokens), and $45.1\\times$ on longer sequences, while consistently maintaining higher accuracy than the baseline. Our method achieves significantly higher throughput ($6.8\\times$ on GSM8K) than existing confidence-based approaches while preserving generation quality, enabling practical deployment of diffusion LLMs.", "authors": ["Quan Nguyen-Tri", "Mukul Ranjan", "Zhiqiang Shen"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-16", "url": "https://arxiv.org/abs/2510.14973", "pdf_url": "https://arxiv.org/pdf/2510.14973v2", "arxiv_id": "2510.14973", "doi": "10.48550/arXiv.2510.14973", "citation_count": 17, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/VILA-Lab/Elastic-Cache", "venue": "arXiv.org", "quality_score": 0.5259}
{"id": "5023fda90496996a62c04842016e1abff9a56b74578db0ab7c1f89692758cbed", "sources": ["arxiv", "semantic_scholar"], "title": "Kelle: Co-design KV Caching and eDRAM for Efficient LLM Serving in Edge Computing", "abstract": "Running Large Language Models (LLMs) on edge devices is crucial for reducing latency, improving real-time processing, and enhancing privacy. By performing inference directly on the device, data does not need to be sent to the cloud, ensuring faster responses and reducing reliance on network connectivity. However, implementing LLMs on edge devices presents challenges, particularly with managing key-value (KV) caches, which plays a pivotal role in LLM serving. As the input text lengthens, the size of the KV cache increases linearly with the sequence length, leading to a significant memory footprint and data access costs. On the other hand, edge devices have limited memory and computational power, making it hard to store and efficiently access the large caches needed for LLM inference. To mitigate the substantial overhead caused by KV cache, we propose using embedded DRAM (eDRAM) as the primary storage for LLM serving in edge device, which offers higher storage density compared to SRAM. However, to ensure data integrity, eDRAM needs periodic refresh operations, which are power-intensive. To reduce eDRAM costs and improve overall system performance, we propose~\\textit{Kelle}, a software-hardware co-design solution optimized for deploying LLMs on eDRAM-based edge systems. Combined with our fine-grained memory eviction, recomputation, and refresh control algorithms, the \\textit{Kelle} accelerator delivers a $3.9\\times$ speedup and $4.5\\times$ energy savings compared to existing baseline solutions.", "authors": ["Tianhua Xia", "Sai Qian Zhang"], "categories": ["cs.AR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-16", "url": "https://arxiv.org/abs/2510.16040", "pdf_url": "https://arxiv.org/pdf/2510.16040v1", "arxiv_id": "2510.16040", "doi": "10.1145/3725843.3756071", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Micro", "quality_score": 0.3403}
{"id": "5cb2f6b4936330350a7c5c3cd0a60035136ed7efe3231e7e314c04d124da7f82", "sources": ["arxiv", "semantic_scholar"], "title": "REAP the Experts: Why Pruning Prevails for One-Shot MoE compression", "abstract": "Sparsely-activated Mixture-of-Experts (SMoE) models offer efficient pre-training and low latency but their large parameter counts create significant memory overhead, motivating research into expert compression. Contrary to recent findings favouring expert merging on discriminative benchmarks, we find that expert pruning is a superior strategy for generative tasks. We demonstrate that existing merging techniques introduce an irreducible error due to the loss of fine-grained routing control over experts. Leveraging this insight, we propose Router-weighted Expert Activation Pruning (REAP), a novel pruning criterion that considers both router gate-values and expert activation norms to minimize the reconstruction error bound. Across a diverse set of SMoE models ranging from 20B to 1T parameters, REAP consistently outperforms merging and other pruning methods on generative benchmarks, especially at 50% compression. Notably, our method achieves near-lossless compression on code generation tasks with Qwen3-Coder-480B and Kimi-K2, even after pruning 50% of experts.", "authors": ["Mike Lasby", "Ivan Lazarevich", "Nish Sinnadurai", "Sean Lie", "Yani Ioannou", "Vithursan Thangarasa"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-15", "url": "https://arxiv.org/abs/2510.13999", "pdf_url": "https://arxiv.org/pdf/2510.13999v3", "arxiv_id": "2510.13999", "doi": "10.48550/arXiv.2510.13999", "citation_count": 17, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3891}
{"id": "efeca5f7269d0fafa1386f8d0a2ed4323cffcf9273af93423eecd4edf6e919df", "sources": ["arxiv", "semantic_scholar"], "title": "BanaServe: Unified KV Cache and Dynamic Module Migration for Balancing Disaggregated LLM Serving in AI Infrastructure", "abstract": "Large language models (LLMs) are increasingly deployed in AI infrastructure, driving the need for high throughput, resource efficient serving systems. Disaggregated LLM serving, which separates prompt prefill from auto-regressive decode, has emerged as a promising architecture by isolating their heterogeneous compute and memory demands. However, current disaggregated systems face three key limitations: (i) static resource allocation cannot adapt to highly dynamic workloads, causing over-provisioning that wastes resources or under-provisioning that violates service level objectives (SLOs); (ii) inherent load imbalance between prefill and decode stages, where prefill is compute-bound and decode is memory-bound, causes under-utilization in one tier while the other becomes a bottleneck; and (iii) prefix cache aware routing skews load distribution, as high cache hit rate prefill nodes attract disproportionately more requests, further degrading balance and efficiency. To address these issues, we present BanaServe, a dynamic orchestration framework that continuously rebalances computational and memory resources across prefill and decode instances while eliminating hotspots induced by cache. BanaServe introduces layer level weight migration, attention level Key Value Cache (KV Cache) migration, and Global KV Cache Store sharing with layer wise overlapped transmission, enabling both coarse grained (layer level) and fine grained (attention level) load redistribution with minimal latency overhead. These mechanisms allow routers to perform purely load aware scheduling, unconstrained by cache placement. Compared to vLLM, BanaServe achieves 1.2x-3.9x higher throughput with 3.9%-78.4% lower total processing time, and outperforms DistServe by 1.1x-2.8x in throughput with 1.4%-70.1% latency reduction.", "authors": ["Yiyuan He", "Minxian Xu", "Jingfeng Wu", "Jianmin Hu", "Chong Ma", "Min Shen", "Le Chen", "Chengzhong Xu", "Lin Qu", "Kejiang Ye"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-15", "url": "https://arxiv.org/abs/2510.13223", "pdf_url": "https://arxiv.org/pdf/2510.13223v1", "arxiv_id": "2510.13223", "doi": "10.1002/spe.70054", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Software: Practice and Experience 2026", "quality_score": 0.3392}
{"id": "54201d5ec0b891344ba947cc04ce3302209e9e40b6ddfc20d7df0ef49a70d3a8", "sources": ["arxiv", "semantic_scholar"], "title": "GatePro: Parameter-Free Expert Selection Optimization for Mixture-of-Experts Models", "abstract": "Modern large language models leverage Mixture-of-Experts (MoE) architectures for efficient scaling, but face a critical challenge: functionally similar experts are often selected simultaneously, creating redundant computation and limiting effective model capacity. Existing auxiliary balance loss methods improve token distribution but fail to address the underlying expert diversity problem. We introduce GatePro, a novel parameter-free method that directly promotes expert selection diversity. GatePro identifies the most similar expert pairs and introduces localized competition mechanisms, preventing redundant expert co-activation while maintaining natural expert specialization. Our comprehensive evaluation demonstrates GatePro's effectiveness across model scales and benchmarks. Analysis demonstrates GatePro's ability to achieve enhanced expert diversity, where experts develop more distinct and complementary capabilities, avoiding functional redundancy. This approach can be deployed hot-swappable during any training phase without additional learnable parameters, offering a practical solution for improving MoE effectiveness.", "authors": ["Chen Zheng", "Yuhang Cai", "Deyi Liu", "Jin Ma", "Yiyuan Ma", "Yuan Yang", "Jing Liu", "Yutao Zeng", "Xun Zhou", "Siyuan Qiao"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-15", "url": "https://arxiv.org/abs/2510.13079", "pdf_url": "https://arxiv.org/pdf/2510.13079v1", "arxiv_id": "2510.13079", "doi": "10.48550/arXiv.2510.13079", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3392}
{"id": "c72d3ddfe8a61de0449ab6bea395f7716355f88d72566c412238b836ab89b609", "sources": ["arxiv", "semantic_scholar"], "title": "Who Speaks for the Trigger? Dynamic Expert Routing in Backdoored Mixture-of-Experts Transformers", "abstract": "Large language models (LLMs) with Mixture-of-Experts (MoE) architectures achieve impressive performance and efficiency by dynamically routing inputs to specialized subnetworks, known as experts. However, this sparse routing mechanism inherently exhibits task preferences due to expert specialization, introducing a new and underexplored vulnerability to backdoor attacks. In this work, we investigate the feasibility and effectiveness of injecting backdoors into MoE-based LLMs by exploiting their inherent expert routing preferences. We thus propose BadSwitch, a novel backdoor framework that integrates task-coupled dynamic trigger optimization with a sensitivity-guided Top-S expert tracing mechanism. Our approach jointly optimizes trigger embeddings during pretraining while identifying S most sensitive experts, subsequently constraining the Top-K gating mechanism to these targeted experts. Unlike traditional backdoor attacks that rely on superficial data poisoning or model editing, BadSwitch primarily embeds malicious triggers into expert routing paths with strong task affinity, enabling precise and stealthy model manipulation. Through comprehensive evaluations across three prominent MoE architectures (Switch Transformer, QwenMoE, and DeepSeekMoE), we demonstrate that BadSwitch can efficiently hijack pre-trained models with up to 100% success rate (ASR) while maintaining the highest clean accuracy (ACC) among all baselines. Furthermore, BadSwitch exhibits strong resilience against both text-level and model-level defense mechanisms, achieving 94.07% ASR and 87.18% ACC on the AGNews dataset. Our analysis of expert activation patterns reveals fundamental insights into MoE vulnerabilities. We anticipate this work will expose security risks in MoE systems and contribute to advancing AI safety.", "authors": ["Xin Zhao", "Xiaojun Chen", "Bingshan Liu", "Haoyu Gao", "Zhendong Zhao", "Yilong Chen"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-15", "url": "https://arxiv.org/abs/2510.13462", "pdf_url": "https://arxiv.org/pdf/2510.13462v1", "arxiv_id": "2510.13462", "doi": "10.48550/arXiv.2510.13462", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3392}
{"id": "dc267a9d5e013280fe9a65b49b1d3d0b177002756c3a81c86298d6b9e5f775fb", "sources": ["arxiv", "semantic_scholar"], "title": "MoBiLE: Efficient Mixture-of-Experts Inference on Consumer GPU with Mixture of Big Little Experts", "abstract": "Mixture-of-Experts (MoE) models have recently demonstrated exceptional performance across a diverse range of applications. The principle of sparse activation in MoE models facilitates an offloading strategy, wherein active experts are maintained in GPU HBM, while inactive experts are stored in CPU DRAM. The efficacy of this approach, however, is fundamentally constrained by the limited bandwidth of the CPU-GPU interconnect. To mitigate this bottleneck, existing approaches have employed prefetching to accelerate MoE inference. These methods attempt to predict and prefetch the required experts using specially trained modules. Nevertheless, such techniques are often encumbered by significant training overhead and have shown diminished effectiveness on recent MoE models with fine-grained expert segmentation. In this paper, we propose MoBiLE, a plug-and-play offloading-based MoE inference framework with \\textit{mixture of big-little experts}. It reduces the number of experts for unimportant tokens to half for acceleration while maintaining full experts for important tokens to guarantee model quality. Further, a dedicated fallback and prefetching mechanism is designed for switching between little and big experts to improve memory efficiency. We evaluate MoBiLE on four typical modern MoE architectures and challenging generative tasks. Our results show that MoBiLE achieves a speedup of 1.60x to 1.72x compared to the baseline on a consumer GPU system, with negligible degradation in accuracy.", "authors": ["Yushu Zhao", "Yubin Qin", "Yang Wang", "Xiaolong Yang", "Huiming Han", "Shaojun Wei", "Yang Hu", "Shouyi Yin"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-14", "url": "https://arxiv.org/abs/2510.12357", "pdf_url": "https://arxiv.org/pdf/2510.12357v1", "arxiv_id": "2510.12357", "doi": "10.1109/ASP-DAC66049.2026.11420472", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Asia and South Pacific Design Automation Conference", "quality_score": 0.338}
{"id": "eda9635eac189526881811363d3080c3866d76ee844afe339bee914a8a6126ff", "sources": ["arxiv", "semantic_scholar"], "title": "KVCOMM: Online Cross-context KV-cache Communication for Efficient LLM-based Multi-agent Systems", "abstract": "Multi-agent large language model (LLM) systems are increasingly adopted for complex language processing tasks that require communication and coordination among agents. However, these systems often suffer substantial overhead from repeated reprocessing of overlapping contexts across agents. In typical pipelines, once an agent receives a message from its predecessor, the full context-including prior turns-must be reprocessed from scratch, leading to inefficient processing. While key-value (KV) caching is an effective solution for avoiding redundant computation in single-agent settings where prefixes remain unchanged, it cannot be directly reused in multi-agent scenarios due to diverging prefixes introduced by agent-specific context extensions. We identify that the core challenge lies in the offset variance of KV-caches across agents. To address this, we propose KVCOMM, a training-free framework that enables efficient prefilling in multi-agent inference by reusing KV-caches and aligning cache offsets of overlapping contexts under diverse prefix contexts. KVCOMM estimates and adjusts KV-caches for shared content by referencing a pool of cached examples-termed anchors-that store observed cache deviations under varying prefixes. The anchor pool is maintained and updated online, allowing dynamic adaptation to distinct user requests and context structures. KVCOMM achieves over 70% reuse rate across diverse multi-agent workloads, including retrieval-augmented generation, math reasoning, and collaborative coding tasks, all without quality degradation. Particularly, when each fully-connected agent receives 1K input tokens with 512 prefix tokens and 512 output tokens under a five-agent setting, KVCOMM achieves up to 7.8x speedup compared to the standard prefill pipeline, reducing TTFT from ~430 ms to ~55 ms.", "authors": ["Hancheng Ye", "Zhengqi Gao", "Mingyuan Ma", "Qinsi Wang", "Yuzhe Fu", "Ming-Yu Chung", "Yueqian Lin", "Zhijian Liu", "Jianyi Zhang", "Danyang Zhuo", "Yiran Chen"], "categories": ["cs.MA", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-14", "url": "https://arxiv.org/abs/2510.12872", "pdf_url": "https://arxiv.org/pdf/2510.12872v2", "arxiv_id": "2510.12872", "doi": "10.48550/arXiv.2510.12872", "citation_count": 17, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/FastMAS/KVCOMM}", "venue": "arXiv.org", "quality_score": 0.5224}
{"id": "46801520395ffd6b39a2f771cc83f6a6d3f4032ba8dcac0b4ba37b7af9e43541", "sources": ["arxiv", "semantic_scholar"], "title": "Dendrograms of Mixing Measures for Softmax-Gated Gaussian Mixture of Experts: Consistency Without Model Sweeps", "abstract": "We develop a unified statistical framework for softmax-gated Gaussian mixture of experts (SGMoE) that addresses three long-standing obstacles in parameter estimation and model selection: (i) non-identifiability of gating parameters up to common translations, (ii) intrinsic gate-expert interactions that induce coupled differential relations in the likelihood, and (iii) the tight numerator-denominator coupling in the softmax-induced conditional density. Our approach introduces Voronoi-type loss functions aligned with the gate-partition geometry and establishes finite-sample convergence rates for the maximum likelihood estimator (MLE). In over-specified models, we reveal a link between the MLE's convergence rate and the solvability of an associated system of polynomial equations characterizing near-nonidentifiable directions. For model selection, we adapt dendrograms of mixing measures to SGMoE, yielding a consistent, sweep-free selector of the number of experts that attains pointwise-optimal parameter rates under overfitting while avoiding multi-size training. Simulations on synthetic data corroborate the theory, accurately recovering the expert count and achieving the predicted rates for parameter estimation while closely approximating the regression function. Under model misspecification (e.g., $ε$-contamination), the dendrogram selection criterion is robust, recovering the true number of mixture components, while the Akaike information criterion, the Bayesian information criterion, and the integrated completed likelihood tend to overselect as sample size grows. On a maize proteomics dataset of drought-responsive traits, our dendrogram-guided SGMoE selects two experts, exposes a clear mixing-measure hierarchy, stabilizes the likelihood early, and yields interpretable genotype-phenotype maps, outperforming standard criteria without multi-size training.", "authors": ["Do Tien Hai", "Trung Nguyen Mai", "TrungTin Nguyen", "Nhat Ho", "Binh T. Nguyen", "Christopher Drovandi"], "categories": ["stat.ML", "cs.LG", "math.ST", "stat.CO", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-14", "url": "https://arxiv.org/abs/2510.12744", "pdf_url": "https://arxiv.org/pdf/2510.12744v2", "arxiv_id": "2510.12744", "doi": "10.48550/arXiv.2510.12744", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.338}
{"id": "956d803e0963c45579e9d400b6df172f46c9804bceb0a5d23afbdcdce2a03eca", "sources": ["arxiv", "semantic_scholar"], "title": "XQuant: Achieving Ultra-Low Bit KV Cache Quantization with Cross-Layer Compression", "abstract": "Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks. However, their extensive memory requirements, particularly due to KV cache growth during long-text understanding and generation, present significant challenges for deployment in resource-constrained environments. Quantization has emerged as a promising solution to reduce memory consumption while preserving historical information. We propose XQuant, a training-free and plug-and-play framework that achieves ultra-low equivalent bit-width KV cache quantization. XQuant introduces two key innovations: a computationally negligible data-free calibration method and cross-layer KV cache compression, enabling quantization to sub-1.4 bits. Extensive experiments on TruthfulQA and LongBench demonstrate that XQuant outperforms state-of-the-art methods (e.g., KIVI-2bit and AsymKV-1.5bit) by achieving lower bit-width while maintaining superior performance, establishing a better trade-off between memory efficiency and model accuracy.", "authors": ["Haoqi Yang", "Yao Yao", "Zuchao Li", "Baoyuan Qi", "Guoming Liu", "Hai Zhao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-13", "url": "https://arxiv.org/abs/2510.11236", "pdf_url": "https://arxiv.org/pdf/2510.11236v1", "arxiv_id": "2510.11236", "doi": "10.18653/v1/2025.emnlp-main.494", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3369}
{"id": "914ec280ff9dbb3558c419c79e7768ff8d5f158b49e911353325643514a416ed", "sources": ["arxiv", "semantic_scholar"], "title": "MC#: Mixture Compressor for Mixture-of-Experts Large Models", "abstract": "Mixture-of-Experts (MoE) effectively scales large language models (LLMs) and vision-language models (VLMs) by increasing capacity through sparse activation. However, preloading all experts into memory and activating multiple experts per input introduces significant computational and memory overhead, making the expert module a major contributor to model size and inference cost. To address this, we propose MC# (Mixture-Compressor-sharp), a framework that combines static quantization and dynamic expert pruning by leveraging the significance of experts and tokens for aggressive compression of MoE-LLMs/VLMs. To reduce storage and loading costs, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which optimizes bit allocation via linear programming, balancing expert importance and quantization error for a Pareto-optimal trade-off between size and performance. To reduce runtime computation, Online Top-any Pruning (OTP) uses Gumbel-Softmax sampling to dynamically select a subset of experts per token, enabling fine-grained control over activation. By combining PMQ's static bit-width optimization with OTP's dynamic routing, MC# achieves extreme compression with minimal accuracy loss. On DeepSeek-VL2, MC# achieves a 6.2 times weight reduction at 2.57 average bits with only a 1.7% accuracy drop across five multimodal benchmarks. Additionally, OTP reduces expert activation over 20% with less than 1% performance degradation, demonstrating strong potential for efficient MoE-based model deployment.", "authors": ["Wei Huang", "Yue Liao", "Yukang Chen", "Jianhui Liu", "Haoru Tan", "Si Liu", "Shiming Zhang", "Shuicheng Yan", "Xiaojuan Qi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Medicine", "Computer Science"], "published_date": "2025-10-13", "url": "https://arxiv.org/abs/2510.10962", "pdf_url": "https://arxiv.org/pdf/2510.10962v1", "arxiv_id": "2510.10962", "doi": "10.1109/TPAMI.2026.3664873", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.3369}
{"id": "2d1ded7f0b0adeab4d8c18600cbce26dabb57e92ea66ac41bcddf53c9dae3e1a", "sources": ["arxiv", "semantic_scholar"], "title": "LouisKV: Efficient KV Cache Retrieval for Long Input-Output Sequences", "abstract": "While Key-Value (KV) cache succeeds in reducing redundant computations in auto-regressive models, it introduces significant memory overhead, limiting its practical deployment in long-sequence scenarios. Existing KV retrieval methods mitigate this by dynamically retaining only a subset of KV entries on the GPU. However, they still suffer from notable efficiency and accuracy bottlenecks due to per-token retrieval and coarse-grained page-level KV management, especially in long-output reasoning scenarios. With the emergence of large reasoning models, efficiently handling such scenarios has become increasingly important. To address this issue, we present two key observations: (1) critical KVs exhibit strong temporal locality during decoding, and (2) these KVs exhibit distinct distribution patterns across the input prompt and generated output. Building on these observations, we propose LouisKV, an efficient KV cache retrieval framework designed for various long-sequence scenarios. Specifically, LouisKV introduces a semantic-aware retrieval strategy leveraging temporal locality to trigger retrieval only at semantic boundaries, drastically reducing computation and data transfer overhead. LouisKV also designs a decoupled, fine-grained management scheme that tailors differentiated strategies for input and output sequences to create retrieval units that better match the model's attention patterns, enabling precise identification of critical KVs. Furthermore, to boost efficiency, LouisKV incorporates several kernel-level optimizations, including custom Triton and CUDA kernels to accelerate the KV clustering and retrieval. Evaluations show that LouisKV achieves up to 4.7$\\times$ speedup over state-of-the-art KV retrieval methods while maintaining near-lossless accuracy across diverse long-sequence tasks, including long-input short-output, short-input long-output, and long-input long-output scenarios.", "authors": ["Wenbo Wu", "Qingyi Si", "Xiurui Pan", "Ye Wang", "Jie Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-13", "url": "https://arxiv.org/abs/2510.11292", "pdf_url": "https://arxiv.org/pdf/2510.11292v1", "arxiv_id": "2510.11292", "doi": "10.48550/arXiv.2510.11292", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3369}
{"id": "8de2619e9871f1487e62a134083f6e88b2256a081e14a9b6f5f08236f693086f", "sources": ["arxiv", "semantic_scholar"], "title": "CacheClip: Accelerating RAG with Effective KV Cache Reuse", "abstract": "Retrieval-Augmented Generation (RAG) systems suffer from severe time-to-first-token (TTFT) bottlenecks due to long input sequences. Existing KV cache reuse methods face a fundamental trade-off: prefix caching requires identical prefixes that rarely occur in RAG scenarios, while direct precomputation sacrifices quality due to missing inter-chunk attention and repeated attention sinks. Recent methods like APE and CacheBlend partially address these issues but remain inadequate for robust RAG applications. This paper presents CacheClip, a novel framework that achieves both fast TTFT and high generation quality. Our key insight is that small auxiliary LLMs exhibit similar last-layer attention distributions to primary LLMs (the target model for generation), enabling efficient identification of tokens critical for restoring inter-chunk attention, thereby significantly improving response quality on cross-chunk reasoning tasks. CacheClip integrates four techniques: (1) auxiliary-model-guided token selection for selective KV cache recomputation, (2) shared prefixes to eliminate redundant attention sinks, (3) a sliding-window grouping strategy to maintain local coherence during partial KV cache updates, and (4) a CPU-GPU hybrid design that offloads auxiliary model inference to idle CPU resources, avoiding additional GPU overhead. The recomputation ratio is adjustable, allowing users to flexibly balance efficiency and quality for different deployment requirements. Experiments show CacheClip retains up to 85.2% and 91.1% of full-attention performance on NIAH and LongBench, outperforming CacheBlend and APE by 16.1 and 12.8 points on NIAH, and by 4.5 and 4.2 points on LongBench (with recomp% = 20%). Meanwhile, CacheClip accelerates LLM inference by up to 3.33$\\times$ in prefill time (with recomp% = 20%), providing a practical solution to the efficiency-quality trade-off in RAG systems.", "authors": ["Bin Yang", "Qiuyu Leng", "Jun Zeng", "Zhenhua Wu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-11", "url": "https://arxiv.org/abs/2510.10129", "pdf_url": "https://arxiv.org/pdf/2510.10129v2", "arxiv_id": "2510.10129", "doi": "10.48550/arXiv.2510.10129", "citation_count": 6, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3346}
{"id": "23a23d5783bca9a05d971a96c3d9d595107437e779d05f230db1cdd559b6a58b", "sources": ["arxiv", "semantic_scholar"], "title": "TC-LoRA: Temporally Modulated Conditional LoRA for Adaptive Diffusion Control", "abstract": "Current controllable diffusion models typically rely on fixed architectures that modify intermediate activations to inject guidance conditioned on a new modality. This approach uses a static conditioning strategy for a dynamic, multi-stage denoising process, limiting the model's ability to adapt its response as the generation evolves from coarse structure to fine detail. We introduce TC-LoRA (Temporally Modulated Conditional LoRA), a new paradigm that enables dynamic, context-aware control by conditioning the model's weights directly. Our framework uses a hypernetwork to generate LoRA adapters on-the-fly, tailoring weight modifications for the frozen backbone at each diffusion step based on time and the user's condition. This mechanism enables the model to learn and execute an explicit, adaptive strategy for applying conditional guidance throughout the entire generation process. Through experiments on various data domains, we demonstrate that this dynamic, parametric control significantly enhances generative fidelity and adherence to spatial conditions compared to static, activation-based methods. TC-LoRA establishes an alternative approach in which the model's conditioning strategy is modified through a deeper functional adaptation of its weights, allowing control to align with the dynamic demands of the task and generative stage.", "authors": ["Minkyoung Cho", "Ruben Ohana", "Christian Jacobsen", "Adityan Jothi", "Min-Hung Chen", "Z. Morley Mao", "Ethem Can"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-10", "url": "https://arxiv.org/abs/2510.09561", "pdf_url": "https://arxiv.org/pdf/2510.09561v2", "arxiv_id": "2510.09561", "doi": "10.48550/arXiv.2510.09561", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3334}
{"id": "1c1b97dad7827c583d2c1c45f62e11621fcfd0c90001a3ee9e6fe41172234494", "sources": ["arxiv", "semantic_scholar"], "title": "OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference", "abstract": "Large language models (LLMs) with extended context windows enable powerful applications but impose significant memory overhead, as caching all key-value (KV) states scales linearly with sequence length and batch size. Existing cache eviction methods address this by exploiting attention sparsity, yet they typically rank tokens heuristically using accumulated attention weights without considering their true impact on attention outputs. We propose Optimal Brain Cache (OBCache), a principled framework that formulates cache eviction as a layer-wise structured pruning problem. Building upon the Optimal Brain Damage (OBD) theory, OBCache quantifies token saliency by measuring the perturbation in attention outputs induced by pruning tokens, with closed-form scores derived for isolated keys, isolated values, and joint key-value pairs. Our scores account not only for attention weights but also for information from value states and attention outputs, thereby enhancing existing eviction strategies with output-aware signals. Experiments on LLaMA and Qwen models demonstrate that replacing the heuristic scores in existing works, which estimate token saliency across different query positions, with OBCache's output-aware scores consistently improves long-context accuracy. Code is available at https://github.com/DreamSoul-AI/OBCache.", "authors": ["Yuzhe Gu", "Xiyu Liang", "Jiaojiao Zhao", "Enmao Diao"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-09", "url": "https://arxiv.org/abs/2510.07651", "pdf_url": "https://arxiv.org/pdf/2510.07651v2", "arxiv_id": "2510.07651", "doi": "10.48550/arXiv.2510.07651", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/DreamSoul-AI/OBCache", "venue": "arXiv.org", "quality_score": 0.5135}
{"id": "2cde25cf684e9932b80ec008719629a8064251e862da07d56e582f3056723ba1", "sources": ["arxiv", "semantic_scholar"], "title": "Fewer Weights, More Problems: A Practical Attack on LLM Pruning", "abstract": "Model pruning, i.e., removing a subset of model weights, has become a prominent approach to reducing the memory footprint of large language models (LLMs) during inference. Notably, popular inference engines, such as vLLM, enable users to conveniently prune downloaded models before they are deployed. While the utility and efficiency of pruning methods have improved significantly, the security implications of pruning remain underexplored. In this work, for the first time, we show that modern LLM pruning methods can be maliciously exploited. In particular, an adversary can construct a model that appears benign yet, once pruned, exhibits malicious behaviors. Our method is based on the idea that the adversary can compute a proxy metric that estimates how likely each parameter is to be pruned. With this information, the adversary can first inject a malicious behavior into those parameters that are unlikely to be pruned. Then, they can repair the model by using parameters that are likely to be pruned, effectively canceling out the injected behavior in the unpruned model. We demonstrate the severity of our attack through extensive evaluation on five models; after any of the pruning in vLLM are applied (Magnitude, Wanda, and SparseGPT), it consistently exhibits strong malicious behaviors in a diverse set of attack scenarios (success rates of up to $95.7\\%$ for jailbreak, $98.7\\%$ for benign instruction refusal, and $99.5\\%$ for targeted content injection). Our results reveal a critical deployment-time security gap and underscore the urgent need for stronger security awareness in model compression.", "authors": ["Kazuki Egashira", "Robin Staab", "Thibaud Gloaguen", "Mark Vero", "Martin Vechev"], "categories": ["cs.LG", "cs.AI", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-09", "url": "https://arxiv.org/abs/2510.07985", "pdf_url": "https://arxiv.org/pdf/2510.07985v3", "arxiv_id": "2510.07985", "doi": "10.48550/arXiv.2510.07985", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/eth-sri/llm-pruning-attack", "venue": "arXiv.org", "quality_score": 0.5135}
{"id": "b0925594045eb289df1e7f77db68394e5f947a81ed08c883bd072c26930ccc42", "sources": ["arxiv", "semantic_scholar"], "title": "Which Heads Matter for Reasoning? RL-Guided KV Cache Compression", "abstract": "Reasoning large language models exhibit complex reasoning behaviors via extended chain-of-thought generation that are highly fragile to information loss during decoding, creating critical challenges for KV cache compression. Existing token-dropping methods directly disrupt reasoning chains by removing intermediate steps, while head-reallocation methods, designed for retrieval tasks, fail to preserve the heads essential for generative reasoning. However, no existing method can identify which attention heads genuinely maintain reasoning consistency and control generation termination. To address this, we propose RLKV, which uses reinforcement learning as a probe to discover which heads contribute to reasoning quality by directly optimizing their cache usage against actual generation outcomes. This discovery naturally leads to an efficient compression strategy: we allocate full KV cache to reasoning-critical heads while aggressively compressing others with constant-size KV cache. Experiments reveal that a fraction of heads proves essential for reasoning, enabling 20--60% cache reduction with near-lossless performance across diverse tasks and models, and up to 2.06x end-to-end speedup at 60% reduction.", "authors": ["Wenjie Du", "Li Jiang", "Keda Tao", "Xue Liu", "Huan Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-09", "url": "https://arxiv.org/abs/2510.08525", "pdf_url": "https://arxiv.org/pdf/2510.08525v3", "arxiv_id": "2510.08525", "doi": "10.48550/arXiv.2510.08525", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3323}
{"id": "68cf3b382f3f8a7491f53d2d045ebc3c037a948e5936431566c7b716a493c5dc", "sources": ["arxiv", "semantic_scholar"], "title": "LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference", "abstract": "KV cache has traditionally been stored in GPU memory to accelerate the decoding phase of large language model (LLM) inference. However, it is increasingly necessary to move KV caches outside GPU devices, to enable cache reuse across different queries and inference engines. Our real-world usage statistics confirm this trend: over time, the total KV cache stored by users has grown rapidly, far exceeding the capacity of GPU memory. Despite this need, there lacks an efficient solution for offloading and transferring KV caches. We present LMCACHE, the first and so far the most efficient open-source KV caching solution, which extracts and stores KV caches generated by modern LLM engines (vLLM and SGLang) out of the GPU memory and shares them across engines and queries. LMCACHE supports both cache offloading (prefix reuse across queries) and prefill-decode (PD) disaggregation (cross-engine/GPU cache transfer). LMCACHE's high performance and wide adoption stem from the following contributions: (1) highly optimized KV cache data movement powered by batched data movement operations, compute and I/O pipelining; (2) a modular KV cache connector component, decoupling LMCACHE from the rapid evolution of inference engines; (3) a first-class control API for flexible cache orchestration across GPU, CPU, storage, and network layers. Our evaluation shows that combining LMCACHE with vLLM achieves up to 15x improvement in throughput across workloads such as multi-round question answering and document analysis. Large-scale adoption of LMCACHE in enterprise settings provides us valuable insights, for example, fetching KV cache from remote storage has unsurprisingly benefits to prefill delay, and that context truncation, which is a widely applied technique in industry, can greatly reduce prefix cache hit ratio by half. The source code of LMCACHE is at: https://github.com/LMCache/LMCache.", "authors": ["Yuhan Liu", "Yihua Cheng", "Jiayi Yao", "Yuwei An", "Xiaokun Chen", "Shaoting Feng", "Yuyang Huang", "Samuel Shen", "Rui Zhang", "Kuntai Du", "Junchen Jiang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-08", "url": "https://arxiv.org/abs/2510.09665", "pdf_url": "https://arxiv.org/pdf/2510.09665v2", "arxiv_id": "2510.09665", "doi": "10.48550/arXiv.2510.09665", "citation_count": 76, "influential_citation_count": 11, "has_code": true, "code_url": "https://github.com/LMCache/LMCache", "venue": "arXiv.org", "quality_score": 0.5396}
{"id": "2a530942eb5597836afd7e677be65dbd213603f7d792e0e4d9d0dc7dfba89836", "sources": ["arxiv", "semantic_scholar"], "title": "Guided by the Experts: Provable Feature Learning Dynamic of Soft-Routed Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) architectures have emerged as a cornerstone of modern AI systems. In particular, MoEs route inputs dynamically to specialized experts whose outputs are aggregated through weighted summation. Despite their widespread application, theoretical understanding of MoE training dynamics remains limited to either separate expert-router optimization or only top-1 routing scenarios with carefully constructed datasets. This paper advances MoE theory by providing convergence guarantees for joint training of soft-routed MoE models with non-linear routers and experts in a student-teacher framework. We prove that, with moderate over-parameterization, the student network undergoes a feature learning phase, where the router's learning process is ``guided'' by the experts, that recovers the teacher's parameters. Moreover, we show that a post-training pruning can effectively eliminate redundant neurons, followed by a provably convergent fine-tuning process that reaches global optimality. To our knowledge, our analysis is the first to bring novel insights in understanding the optimization landscape of the MoE architecture.", "authors": ["Fangshuo Liao", "Anastasios Kyrillidis"], "categories": ["cs.LG", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-08", "url": "https://arxiv.org/abs/2510.07205", "pdf_url": "https://arxiv.org/pdf/2510.07205v1", "arxiv_id": "2510.07205", "doi": "10.48550/arXiv.2510.07205", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3311}
{"id": "b963b1b75158e2f9ab39c3618752e66e6277c9242204cfcdb71776d8e128a749", "sources": ["arxiv", "semantic_scholar"], "title": "Red-Bandit: Test-Time Adaptation for LLM Red-Teaming via Bandit-Guided LoRA Experts", "abstract": "Automated red-teaming has emerged as a scalable approach for auditing Large Language Models (LLMs) prior to deployment, yet existing approaches lack mechanisms to efficiently adapt to model-specific vulnerabilities at inference. We introduce Red-Bandit, a red-teaming framework that adapts online to identify and exploit model failure modes under distinct attack styles (e.g., manipulation, slang). Red-Bandit post-trains a set of parameter-efficient LoRA experts, each specialized for a particular attack style, using reinforcement learning that rewards the generation of unsafe prompts via a rule-based safety model. At inference, a multi-armed bandit policy dynamically selects among these attack-style experts based on the target model's response safety, balancing exploration and exploitation. Red-Bandit achieves state-of-the-art results on AdvBench under sufficient exploration (ASR@10), while producing more human-readable prompts (lower perplexity). Moreover, Red-Bandit's bandit policy serves as a diagnostic tool for uncovering model-specific vulnerabilities by indicating which attack styles most effectively elicit unsafe behaviors.", "authors": ["Christos Ziakas", "Nicholas Loo", "Nishita Jain", "Alessandra Russo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-08", "url": "https://arxiv.org/abs/2510.07239", "pdf_url": "https://arxiv.org/pdf/2510.07239v2", "arxiv_id": "2510.07239", "doi": "10.48550/arXiv.2510.07239", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3311}
{"id": "350d5a6403db90ec0548d2478b19b0a8195f88a44bcbdff36e4dfbdba4a2b78f", "sources": ["arxiv", "semantic_scholar"], "title": "Incorporating Expert Knowledge into Bayesian Causal Discovery of Mixtures of Directed Acyclic Graphs", "abstract": "Bayesian causal discovery benefits from prior information elicited from domain experts, and in heterogeneous domains any prior knowledge would be badly needed. However, so far prior elicitation approaches have assumed a single causal graph and hence are not suited to heterogeneous domains. We propose a causal elicitation strategy for heterogeneous settings, based on Bayesian experimental design (BED) principles, and a variational mixture structure learning (VaMSL) method -- extending the earlier differentiable Bayesian structure learning (DiBS) method -- to iteratively infer mixtures of causal Bayesian networks (CBNs). We construct an informative graph prior incorporating elicited expert feedback in the inference of mixtures of CBNs. Our proposed method successfully produces a set of alternative causal models (mixture components or clusters), and achieves an improved structure learning performance on heterogeneous synthetic data when informed by a simulated expert. Finally, we demonstrate that our approach is capable of capturing complex distributions in a breast cancer database.", "authors": ["Zachris Björkman", "Jorge Loría", "Sophie Wharrie", "Samuel Kaski"], "categories": ["cs.LG", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-08", "url": "https://arxiv.org/abs/2510.06735", "pdf_url": "https://arxiv.org/pdf/2510.06735v2", "arxiv_id": "2510.06735", "doi": "10.48550/arXiv.2510.06735", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3311}
{"id": "6cccce419062bb72c1caeb5495a4f19d3e52b10d3158817c9565490260db3550", "sources": ["arxiv", "semantic_scholar"], "title": "VecInfer: Efficient LLM Inference with Low-Bit KV Cache via Outlier-Suppressed Vector Quantization", "abstract": "The Key-Value (KV) cache introduces substantial memory overhead during large language model (LLM) inference. Although existing vector quantization (VQ) methods reduce KV cache usage and provide flexible representational capacity across bit-widths, they suffer severe performance degradation at ultra-low bit-widths due to key cache outliers that hinder effective codebook utilization. To address this challenge, we propose VecInfer, a novel VQ method for aggressive KV cache compression while enabling efficient inference. By applying smooth and Hadamard transformations, VecInfer suppresses outliers in the key cache, enabling the codebook to comprehensively cover the original data distribution and thereby reducing quantization difficulty. To facilitate efficient deployment, we design an optimized CUDA kernel that fuses computation with dequantization to minimize memory access overhead. Extensive evaluations demonstrate that VecInfer consistently outperforms existing quantization baselines across both long-context understanding and mathematical reasoning tasks. With only 2-bit quantization, VecInfer achieves performance comparable to full precision, while delivering up to $\\mathbf{2.7\\times}$ speedup in large-batch self-attention computation and $\\mathbf{8.3\\times}$ reduction in single-batch end-to-end latency on Llama-3.1-8B with a 196k sequence length.", "authors": ["Dingyu Yao", "Chenxu Yang", "Zhengyang Tong", "Zheng Lin", "Wei Liu", "Jian Luan", "Weiping Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-07", "url": "https://arxiv.org/abs/2510.06175", "pdf_url": "https://arxiv.org/pdf/2510.06175v1", "arxiv_id": "2510.06175", "doi": "10.48550/arXiv.2510.06175", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.33}
{"id": "e41b4339a83e0da574a8282f9f186fc0b30560298ecaac8f5cb7d353c5e6a163", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Neuron Experts", "abstract": "In this work, we first explore whether the parameters activated by the MoE layer remain highly sparse at inference. We perform a sparsification study on several representative MoE models. For each expert, we rank parameters by the magnitude of their activations from the gate projection and progressively prune the activated subset. Pruning up to 60% of parameters within that subset causes only negligible task-performance degradation; substantial drops occur only after more than 90% are removed. We further decompose experts into neuron-granular MoE and visualize their activation values, finding that most neuron activations are near zero. This observation motivates us to select only high-activation neuron experts during pretraining. Based on this insight, we propose Mixture of Neuron Experts (MoNE). MoNE achieves neuron-granular expert selection by only applying a simple top-k selection within each expert, incurs negligible latency, and requires no additional routing parameters or inter-expert communication. Extensive experiments demonstrate that MoNE matches traditional MoE performance while activating only 50% of the MoE-layer parameters, and it consistently outperforms traditional MoE when compared at equal numbers of activated parameters. These results suggest that MoNE is a practical approach to improving parameter utilization and inference efficiency in MoE-like models.", "authors": ["Runxi Cheng", "Yuchen Guan", "Yucheng Ding", "Qingguo Hu", "Yongxian Wei", "Chun Yuan", "Yelong Shen", "Weizhu Chen", "Yeyun Gong"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-07", "url": "https://arxiv.org/abs/2510.05781", "pdf_url": "https://arxiv.org/pdf/2510.05781v1", "arxiv_id": "2510.05781", "doi": "10.48550/arXiv.2510.05781", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.33}
{"id": "3c166058b9ec9928c77f9fc2beebe57fc8eba3ea334c403d9840f3ca6db2afc9", "sources": ["arxiv", "semantic_scholar"], "title": "H1B-KV: Hybrid One-Bit Caches for Memory-Efficient Large Language Model Inference", "abstract": "Autoregressive decoding in large language models (LLMs) requires caching a growing list of past key-value (KV) pairs, making long-context inference a memory-bound problem. While recent methods have explored quantizing the cache, evicting tokens, or using binary sketches for keys (e.g., Loki), these approaches often provide an incomplete solution by leaving one component (like values) uncompressed or by discarding context information. This paper introduces the Hybrid One-Bit KV Cache (H1B-KV), a comprehensive compression scheme that radically reduces memory usage without sacrificing context. H1B-KV represents each key vector using a 1-bit binary sketch, enabling hardware-friendly bitwise attention, and further compresses value vectors using 4-bit quantization. This holistic, hybrid approach allows a 7-billion parameter LLM to handle an 8k-token context with under 60 MB of cache memory - a 70x reduction. We demonstrate that after a lightweight finetuning, H1B-KV matches full-precision performance not only on perplexity benchmarks but also on complex downstream tasks like mathematical reasoning (GSM8K), multi-task understanding (MMLU), and code generation (HumanEval). Our results show H1B-KV significantly outperforms leading quantization (KIVI), token eviction (SparseLLM), and key-only sketching (Loki) methods in quality-per-byte, establishing it as a robust solution for deploying LLMs in memory-constrained environments.", "authors": ["Harshil Vejendla"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-07", "url": "https://arxiv.org/abs/2510.05529", "pdf_url": "https://arxiv.org/pdf/2510.05529v1", "arxiv_id": "2510.05529", "doi": "10.1109/URTC68753.2025.11533071", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.21}
{"id": "725e8bec7a0d70b8b0fbbc6d73f14df6f62cf798a6fb9c5c015d79088f5f375d", "sources": ["arxiv", "semantic_scholar"], "title": "Rasterized Steered Mixture of Experts for Efficient 2D Image Regression", "abstract": "The Steered Mixture of Experts regression framework has demonstrated strong performance in image reconstruction, compression, denoising, and super-resolution. However, its high computational cost limits practical applications. This work introduces a rasterization-based optimization strategy that combines the efficiency of rasterized Gaussian kernel rendering with the edge-aware gating mechanism of the Steered Mixture of Experts. The proposed method is designed to accelerate two-dimensional image regression while maintaining the model's inherent sparsity and reconstruction quality. By replacing global iterative optimization with a rasterized formulation, the method achieves significantly faster parameter updates and more memory-efficient model representations. In addition, the proposed framework supports applications such as native super-resolution and image denoising, which are not directly achievable with standard rasterized Gaussian kernel approaches. The combination of fast rasterized optimization with the edge-aware structure of the Steered Mixture of Experts provides a new balance between computational efficiency and reconstruction fidelity for two-dimensional image processing tasks.", "authors": ["Yi-Hsin Li", "Mårten Sjöström", "Sebastian Knorr", "Thomas Sikora"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-07", "url": "https://arxiv.org/abs/2510.05814", "pdf_url": "https://arxiv.org/pdf/2510.05814v2", "arxiv_id": "2510.05814", "doi": "10.48550/arXiv.2510.05814", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.33}
{"id": "95149839ac167f62f6ecdf1193bfb058ac1b77e79e386e657ed9ea80b0d53047", "sources": ["arxiv", "semantic_scholar"], "title": "KVLinC : KV Cache Quantization with Hadamard Rotation and Linear Correction", "abstract": "Quantizing the key-value (KV) cache is a promising strategy for improving the inference efficiency of large language models (LLMs). However, aggressive quantization to very low precision (e.g., 2 bits) introduces significant errors in the stored key and value tensors, which propagate through the dot-product attention mechanism and ultimately degrade generation quality. To address this, we propose KVLinC, a framework to mitigate attention errors introduced by KV cache quantization in the extreme low-precision regime. KVLinC combines a Hadamard rotation, which reduces quantization error in values, with lightweight linear correction adapters that explicitly compensate for errors introduced by quantized keys. Across extensive evaluations on the LLaMA, Qwen2.5, and Qwen3 model families, KVLinC consistently matches or surpasses strong baselines while achieving higher KV-cache compression. Furthermore, we implement a custom attention kernel that results in upto 2.55x faster inference compared to Flash Attention baseline, enabling efficient long-context LLM inference.", "authors": ["Utkarsh Saxena", "Kaushik Roy"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-06", "url": "https://arxiv.org/abs/2510.05373", "pdf_url": "https://arxiv.org/pdf/2510.05373v1", "arxiv_id": "2510.05373", "doi": "10.48550/arXiv.2510.05373", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3289}
{"id": "da5d45d4581f54114191ad30d6a1e9a62496bcd227b2f708bac9b9897f2d6689", "sources": ["arxiv", "semantic_scholar"], "title": "Recover-LoRA: Data-Free Accuracy Recovery of Degraded Language Models via Low-Rank Adaptation", "abstract": "Inference optimizations such as quantization, pruning, format and datatype conversion, model export, and serialization can lead to functional degradations in language model task performance. While most efforts on performance recovery for deployment focus on robust quantization techniques, we focus on recovering model accuracies from any sources that degrade model weights, such as improper model serialization. In this work, we propose Recover-LoRA, a lightweight and dataset agnostic method to recover accuracy in degraded models. Recover-LoRA uses synthetic data and logit distillation to learn LoRA adapters on selective layers that facilitate aligning the degraded model to its full precision model. We investigate the utility of Recover-LoRA across a diverse set of small language models (SLMs), including models with varying attention architectures, multi-head attention (MHA) and group-query attention (GQA), as well as several evaluation datasets. Our results show that Recover-LoRA recovers model accuracies by 5-17% on MHA and GQA SLMs.", "authors": ["Devleena Das", "Rajeev Patwari", "Ashish Sirasao"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-06", "url": "https://arxiv.org/abs/2510.08600", "pdf_url": "https://arxiv.org/pdf/2510.08600v1", "arxiv_id": "2510.08600", "doi": "10.18653/v1/2025.emnlp-industry.164", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3289}
{"id": "24fba133d6f3a011721a7a48c79fab4915917754baa8fd3e7cabf4cf2be938d6", "sources": ["arxiv", "semantic_scholar"], "title": "Speak, Edit, Repeat: High-Fidelity Voice Editing and Zero-Shot TTS with Cross-Attentive Mamba", "abstract": "We introduce MAVE (Mamba with Cross-Attention for Voice Editing and Synthesis), a novel autoregressive architecture for text-conditioned voice editing and high-fidelity text-to-speech (TTS) synthesis, built on a cross-attentive Mamba backbone. MAVE achieves state-of-the-art performance in speech editing and very competitive results in zero-shot TTS, while not being explicitly trained on the latter task, outperforming leading autoregressive and diffusion models on diverse, real-world audio. By integrating Mamba for efficient audio sequence modeling with cross-attention for precise text-acoustic alignment, MAVE enables context-aware voice editing with exceptional naturalness and speaker consistency. In pairwise human evaluations on a random 40-sample subset of the RealEdit benchmark (400 judgments), 57.2% of listeners rated MAVE - edited speech as perceptually equal to the original, while 24.8% prefered the original and 18.0% MAVE - demonstrating that in the majority of cases edits are indistinguishable from the source. MAVE compares favorably with VoiceCraft and FluentSpeech both on pairwise comparisons and standalone mean opinion score (MOS) evaluations. For zero-shot TTS, MAVE exceeds VoiceCraft in both speaker similarity and naturalness, without requiring multiple inference runs or post-processing. Remarkably, these quality gains come with a significantly lower memory cost and approximately the same latency: MAVE requires ~6x less memory than VoiceCraft during inference on utterances from the RealEdit database (mean duration: 6.21s, A100, FP16, batch size 1). Our results demonstrate that MAVE establishes a new standard for flexible, high-fidelity voice editing and synthesis through the synergistic integration of structured state-space modeling and cross-modal attention.", "authors": ["Baher Mohammad", "Magauiya Zhussip", "Stamatios Lefkimmiatis"], "categories": ["cs.SD", "cs.AI", "cs.CL", "cs.LG", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-10-06", "url": "https://arxiv.org/abs/2510.04738", "pdf_url": "https://arxiv.org/pdf/2510.04738v1", "arxiv_id": "2510.04738", "doi": "10.48550/arXiv.2510.04738", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3289}
{"id": "a7d127629ac9ceaef5112ee129a0a425bf236aaafe5988f12d01945a8876f4ad", "sources": ["arxiv", "semantic_scholar"], "title": "Multilingual Routing in Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) architectures have become the key to scaling modern LLMs, yet little is understood about how their sparse routing dynamics respond to multilingual data. In this work, we analyze expert routing patterns using parallel multilingual datasets and present highly interpretable layer-wise phenomena. We find that MoE models route tokens in language-specific ways in the early and late decoder layers but exhibit significant cross-lingual routing alignment in middle layers, mirroring parameter-sharing trends observed in dense LLMs. In particular, we reveal a clear, strong correlation between a model's performance in a given language and how similarly its tokens are routed to English in these layers. Extending beyond correlation, we explore inference-time interventions that induce higher cross-lingual routing alignment. We introduce a method that steers the router by promoting middle-layer task experts frequently activated in English, and it successfully increases multilingual performance. These 1-2% gains are remarkably consistent across two evaluation tasks, three models, and 15+ languages, especially given that these simple interventions override routers of extensively trained, state-of-the-art LLMs. In comparison, interventions outside of the middle layers or targeting multilingual-specialized experts only yield performance degradation. Altogether, we present numerous findings that explain how MoEs process non-English text and demonstrate that generalization is limited by the model's ability to leverage language-universal experts in all languages.", "authors": ["Lucas Bandarkar", "Chenyuan Yang", "Mohsen Fayyaz", "Junlin Hu", "Nanyun Peng"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-06", "url": "https://arxiv.org/abs/2510.04694", "pdf_url": "https://arxiv.org/pdf/2510.04694v2", "arxiv_id": "2510.04694", "doi": "10.48550/arXiv.2510.04694", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3289}
{"id": "77f621bfddeb8d0224b17dfebd3e303b2dc7eb29bd2736e9063a1fe14371e308", "sources": ["arxiv", "semantic_scholar"], "title": "Compressed Convolutional Attention: Efficient Attention in a Compressed Latent Space", "abstract": "Multi-headed Attention's (MHA) quadratic compute and linearly growing KV-cache make long-context transformers expensive to train and serve. Prior works such as Grouped Query Attention (GQA) and Multi-Latent Attention (MLA) shrink the cache, speeding decode, but leave compute, which determines prefill and training speed, largely unchanged. We introduce Compressed Convolutional Attention (CCA), a novel attention method which down-projects queries, keys, and values and performs the entire attention operation inside the shared latent space. This simple design dramatically cuts parameters, KV-cache, and FLOPs all at once by the desired compression factor. Because CCA is orthogonal to head-sharing, we combine the two to form Compressed Convolutional Grouped Query Attention (CCGQA), which further tightens the compute-bandwidth Pareto frontier so that users can tune compression toward either FLOP or memory limits without sacrificing quality. Experiments show that CCGQA consistently outperforms both GQA and MLA at equal KV-cache compression on dense and MoE models. Additionally, we show that CCGQA outperforms all other attention methods on MoE models with half the KV-cache of GQA and MLA, achieving an 8x KV-cache compression with no drop in performance compared to standard MHA. CCA and CCGQA also dramatically reduce the FLOP cost of attention which leads to substantially faster training and prefill than existing methods. On H100 GPUs, our fused CCA/CCGQA kernel reduces prefill latency by about 1.7x at a sequence length of 16k relative to MHA, and accelerates backward by about 1.3x.", "authors": ["Tomas Figliolia", "Nicholas Alonso", "Rishi Iyer", "Quentin Anthony", "Beren Millidge"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-06", "url": "https://arxiv.org/abs/2510.04476", "pdf_url": "https://arxiv.org/pdf/2510.04476v2", "arxiv_id": "2510.04476", "doi": "10.48550/arXiv.2510.04476", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3289}
{"id": "5d6313a08f744b2108343a21ba2b681eff886496b79b496460fd08b74ea70a19", "sources": ["arxiv", "semantic_scholar"], "title": "MoME: Estimating Psychological Traits from Gait with Multi-Stage Mixture of Movement Experts", "abstract": "Gait encodes rich biometric and behavioural information, yet leveraging the manner of walking to infer psychological traits remains a challenging and underexplored problem. We introduce a hierarchical Multi-Stage Mixture of Movement Experts (MoME) architecture for multi-task prediction of psychological attributes from gait sequences represented as 2D poses. MoME processes the walking cycle in four stages of movement complexity, employing lightweight expert models to extract spatio-temporal features and task-specific gating modules to adaptively weight experts across traits and stages. Evaluated on the PsyMo benchmark covering 17 psychological traits, our method outperforms state-of-the-art gait analysis models, achieving a 37.47% weighted F1 score at the run level and 44.6% at the subject level. Our experiments show that integrating auxiliary tasks such as identity recognition, gender prediction, and BMI estimation further improves psychological trait estimation. Our findings demonstrate the viability of multi-task gait-based learning for psychological trait estimation and provide a foundation for future research on movement-informed psychological inference.", "authors": ["Andy Cǎtrunǎ", "Adrian Cosma", "Emilian Rǎdoi"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-06", "url": "https://arxiv.org/abs/2510.04654", "pdf_url": "https://arxiv.org/pdf/2510.04654v2", "arxiv_id": "2510.04654", "doi": "10.48550/arXiv.2510.04654", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3289}
{"id": "ef5fe025e1a22694685d245d8e0c641c7576e620546fcdff02d5a31eb712a962", "sources": ["arxiv", "semantic_scholar"], "title": "PatternKV: Flattening KV Representation Expands Quantization Headroom", "abstract": "KV cache in autoregressive LLMs eliminates redundant recomputation but has emerged as the dominant memory and bandwidth bottleneck during inference, notably with long contexts and test-time scaling. KV quantization is a key lever for reducing cache cost, but accuracy drops sharply as the native KV distribution lacks flatness and thus maintains a wide quantization range. Prior work focuses on isolating outliers, which caps their error but fails to flatten the overall distribution, leaving performance fragile under low-bit settings. In this work, we show that the K cache maintains a stable, context-evolving structure, while the V cache carries latent semantic regularities, with both contributing to the organization of vectors into shared patterns. Building on these insights, we propose PatternKV, a pattern-aligned residual quantization scheme. It mines representative pattern vectors online, aligns each KV vector to its nearest pattern, and quantizes only the residual. This reshaping of the KV distribution flattens the quantization target and narrows its range, thereby improving the fidelity of low-bit KV quantization. Across long-context and test-time scaling settings on multiple backbones, PatternKV delivers consistent 2-bit gains, with a 0.08% average 4-bit drop relative to FP16, improves test-time scaling accuracy by 10% on average, and raises throughput by 1.5x while supporting 1.25x larger batches.", "authors": ["Ji Zhang", "Yiwei Li", "Shaoxiong Feng", "Peiwen Yuan", "Xinglin Wang", "Jiayi Shi", "Yueqi Zhang", "Chuyi Tan", "Boyuan Pan", "Yao Hu", "Kan Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-05", "url": "https://arxiv.org/abs/2510.05176", "pdf_url": "https://arxiv.org/pdf/2510.05176v2", "arxiv_id": "2510.05176", "doi": "10.48550/arXiv.2510.05176", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3277}
{"id": "1f1a1c446f079671533f93e5b51eb9458158e209e78ebcdff3f44a975720033b", "sources": ["arxiv", "semantic_scholar"], "title": "Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention", "abstract": "The pursuit of computational efficiency has driven the adoption of low-precision formats for training transformer models. However, this progress is often hindered by notorious training instabilities. This paper provides the first mechanistic explanation for a long-standing and unresolved failure case where training with flash attention in low-precision settings leads to catastrophic loss explosion. Our in-depth analysis reveals that the failure is not a random artifact but caused by two intertwined phenomena: the emergence of similar low-rank representations within the attention mechanism and the compounding effect of biased rounding errors inherent in low-precision arithmetic. We demonstrate how these factors create a vicious cycle of error accumulation that corrupts weight updates, ultimately derailing the training dynamics. To validate our findings, we introduce a minimal modification to the flash attention that mitigates the bias in rounding errors. This simple change stabilizes the training process, confirming our analysis and offering a practical solution to this persistent problem. Code is available at https://github.com/ucker/why-low-precision-training-fails.", "authors": ["Haiquan Qiu", "Quanming Yao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-05", "url": "https://arxiv.org/abs/2510.04212", "pdf_url": "https://arxiv.org/pdf/2510.04212v3", "arxiv_id": "2510.04212", "doi": "10.48550/arXiv.2510.04212", "citation_count": 5, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/ucker/why-low-precision-training-fails", "venue": "arXiv.org", "quality_score": 0.5065}
{"id": "fa52cd9d5d93612f0a8e9a8db0731eaa0ff6e107b0a7398cf832e378d22dfee3", "sources": ["arxiv", "semantic_scholar"], "title": "MoME: Mixture of Matryoshka Experts for Audio-Visual Speech Recognition", "abstract": "Large language models (LLMs) have recently shown strong potential in audio-visual speech recognition (AVSR), but their high computational demands and sensitivity to token granularity limit their practicality in resource-constrained settings. Token compression methods can reduce inference cost, but they require fixing a compression rate in advance and produce a single fixed-length output, offering no flexibility to balance information density and efficiency at inference time. Matryoshka representation learning (MRL) addresses this by enabling a single model to operate across multiple token granularities, allowing compression rates to be adjusted dynamically. However, current MRL-based methods treat each scale independently during training, limiting cross-scale generalization, robustness at high compression, and interpretability. To overcome these limitations, we propose MoME (Mixture of Matryoshka Experts), a novel framework that integrates sparse Mixture-of-Experts (MoE) into MRL-based LLMs for AVSR. MoME augments a frozen LLM with top-k routed and shared experts, allowing dynamic capacity allocation across scales and modalities. A shared router promotes consistent expert activation across granularities, enabling compressed sequences to benefit from representations learned at lower compression. Experiments on LRS2 and LRS3 demonstrate that MoME achieves state-of-the-art performance across AVSR, ASR, and VSR tasks, while requiring significantly fewer parameters and maintaining robustness under noise. MoME unifies the adaptability of MRL with the efficiency of MoE, offering a scalable and interpretable solution for resource-aware speech recognition.", "authors": ["Umberto Cappellazzo", "Minsu Kim", "Pingchuan Ma", "Honglie Chen", "Xubo Liu", "Stavros Petridis", "Maja Pantic"], "categories": ["eess.AS", "cs.CV", "cs.SD"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-10-05", "url": "https://arxiv.org/abs/2510.04136", "pdf_url": "https://arxiv.org/pdf/2510.04136v1", "arxiv_id": "2510.04136", "doi": "10.48550/arXiv.2510.04136", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3277}
{"id": "8f94575634c0bc3902d8235431628b7c2b91df284df04767d897825e3ddded42", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Many Zero-Compute Experts: A High-Rate Quantization Theory Perspective", "abstract": "This paper uses classical high-rate quantization theory to provide new insights into mixture-of-experts (MoE) models for regression tasks. Our MoE is defined by a segmentation of the input space to regions, each with a single-parameter expert that acts as a constant predictor with zero-compute at inference. Motivated by high-rate quantization theory assumptions, we assume that the number of experts is sufficiently large to make their input-space regions very small. This lets us to study the approximation error of our MoE model class: (i) for one-dimensional inputs, we formulate the test error and its minimizing segmentation and experts; (ii) for multidimensional inputs, we formulate an upper bound for the test error and study its minimization. Moreover, we consider the learning of the expert parameters from a training dataset, given an input-space segmentation, and formulate their statistical learning properties. This leads us to theoretically and empirically show how the tradeoff between approximation and estimation errors in MoE learning depends on the number of experts.", "authors": ["Yehuda Dar"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-03", "url": "https://arxiv.org/abs/2510.03151", "pdf_url": "https://arxiv.org/pdf/2510.03151v1", "arxiv_id": "2510.03151", "doi": "10.48550/arXiv.2510.03151", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3254}
{"id": "fd695f1d8d0cce2fc74916021503295a5c6378c6ac27a65ab953aa9235f953e0", "sources": ["arxiv", "semantic_scholar"], "title": "Cache-to-Cache: Direct Semantic Communication Between Large Language Models", "abstract": "Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains that are not attainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 6.4-14.2% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.1-5.4%, while delivering an average 2.5x speedup in latency. Our code is available at https://github.com/thu-nics/C2C.", "authors": ["Tianyu Fu", "Zihan Min", "Hanling Zhang", "Jichao Yan", "Guohao Dai", "Wanli Ouyang", "Yu Wang"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-03", "url": "https://arxiv.org/abs/2510.03215", "pdf_url": "https://arxiv.org/pdf/2510.03215v2", "arxiv_id": "2510.03215", "doi": "10.48550/arXiv.2510.03215", "citation_count": 26, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/thu-nics/C2C", "venue": "arXiv.org", "quality_score": 0.5029}
{"id": "29a9c005481b6e5b482230e8770b09202ebe5bed5144c9a9e61cba6bd60fb576", "sources": ["arxiv", "semantic_scholar"], "title": "KaVa: Latent Reasoning via Compressed KV-Cache Distillation", "abstract": "Large Language Models (LLMs) excel at multi-step reasoning problems with explicit chain-of-thought (CoT), but verbose traces incur significant computational costs and memory overhead, and often carry redundant, stylistic artifacts. Latent reasoning has emerged as an efficient alternative that internalizes the thought process, but it suffers from a critical lack of supervision, limiting its effectiveness on complex, natural-language reasoning traces. In this work we propose KaVa, the first framework that bridges this gap by distilling knowledge directly from a compressed KV-cache of the teacher into a latent-reasoning student via self-distillation, leveraging the representational flexibility of continuous latent tokens to align stepwise KV trajectories. We show that the abstract, unstructured knowledge within compressed KV-cache, which lacks direct token correspondence, can serve as a rich supervisory signal for a latent reasoning student. Empirically, the approach consistently outperforms strong latent baselines, exhibits markedly smaller degradation from equation-only to natural-language traces, and scales to larger backbones while preserving efficiency. These results establish compressed KV-cache distillation as a scalable supervision signal for latent reasoning, combining the accuracy of CoT-trained teachers with the efficiency and deployability of latent inference.", "authors": ["Anna Kuzina", "Maciej Pioro", "Paul N. Whatmough", "Babak Ehteshami Bejnordi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-02", "url": "https://arxiv.org/abs/2510.02312", "pdf_url": "https://arxiv.org/pdf/2510.02312v2", "arxiv_id": "2510.02312", "doi": "10.48550/arXiv.2510.02312", "citation_count": 5, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3243}
{"id": "d49553c66191fd2fefc99119f1fa685c22d6ef31570689a09433d2df0eca96d9", "sources": ["arxiv", "semantic_scholar"], "title": "ElasticMoE: An Efficient Auto Scaling Method for Mixture-of-Experts Models", "abstract": "Mixture-of-Experts (MoE) models promise efficient scaling of large language models (LLMs) by activating only a small subset of experts per token, but their parallelized inference pipelines make elastic serving challenging. Existing strategies fall short: horizontal scaling provisions entire replicas of the current configuration, often tens to hundreds of accelerators, leading to coarse granularity, long provisioning delays, and costly overprovisioning. Vertical scaling offers finer adjustments but typically requires instance restarts, incurring downtime. These limitations make current approaches ill-suited for the bursty, short-lived traffic patterns common in cloud deployments. We present ElasticMoE, an elastic scaling framework for MoE LLMs that achieves fine-grained, low-latency, and zero-downtime scaling. ElasticMoE decouples inference execution from memory operations, enabling scaling steps to proceed concurrently with serving. An HBM Management Module (HMM) reuses weights and KV caches via zero-copy remapping, while high-bandwidth peer-to-peer transfers bring newly added accelerators online without interrupting service. A virtual memory based expert redistribution mechanism migrates MoE experts without costly buffer reallocations, reducing peak memory usage during expert parallelism reconfiguration. Our evaluation on Ascend NPUs with three popular MoE LLMs shows that ElasticMoE achieves up to 9x lower scale-up latency, up to 2x better throughput during scaling, and significantly improves SLO attainment compared to baselines. By enabling fine-grained, concurrent scaling with minimal disruption, ElasticMoE advances the practicality of deploying massive MoE LLMs in dynamic cloud environments.", "authors": ["Gursimran Singh", "Timothy Yu", "Haley Li", "Cheng Chen", "Hanieh Sadri", "Qintao Zhang", "Yu Zhang", "Ying Xiong", "Yong Zhang", "Zhenan Fan"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-02", "url": "https://arxiv.org/abs/2510.02613", "pdf_url": "https://arxiv.org/pdf/2510.02613v1", "arxiv_id": "2510.02613", "doi": "10.48550/arXiv.2510.02613", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3243}
{"id": "2cc7f64f0137adff7e2f0c4f8f68dbadd3551bf960bb1511dbcd09e08a211b1c", "sources": ["arxiv", "semantic_scholar"], "title": "Expected Attention: KV Cache Compression by Estimating Attention from Future Queries Distribution", "abstract": "Memory consumption of the Key-Value (KV) cache represents a major bottleneck for efficient large language model inference. While attention-score-based KV cache pruning shows promise, it faces critical practical limitations: attention scores from future tokens are unavailable during compression, and modern implementations like Flash Attention do not materialize the full attention matrix, making past scores inaccessible. To overcome these challenges, we introduce $\\textbf{Expected Attention, a training-free compression method}$ that estimates KV pairs importance by predicting how future queries will attend to them. Our approach leverages the distributional properties of LLM activations to compute expected attention scores in closed form for each KV pair. These scores enable principled ranking and pruning of KV pairs with minimal impact on the residual stream, achieving effective compression without performance degradation. Importantly, our method operates seamlessly across both prefilling and decoding phases, consistently outperforming state-of-the-art baselines in both scenarios. Finally, $\\textbf{we release KVPress, a comprehensive library to enable researchers to implement and benchmark KV cache compression methods, already including more than 20 techniques}$.", "authors": ["Alessio Devoto", "Maximilian Jeblick", "Simon Jégou"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-01", "url": "https://arxiv.org/abs/2510.00636", "pdf_url": "https://arxiv.org/pdf/2510.00636v1", "arxiv_id": "2510.00636", "doi": "10.48550/arXiv.2510.00636", "citation_count": 42, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4515}
{"id": "7306b0a046a460995257046c7e55f4a79e71d73011ed82a760636fd96379c4de", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Shared Experts with LoRA-Based Mixture of Experts for Multi-Task Learning", "abstract": "Mixture-of-Experts (MoE) has emerged as a powerful framework for multi-task learning (MTL). However, existing MoE-MTL methods often rely on single-task pretrained backbones and suffer from redundant adaptation and inefficient knowledge sharing during the transition from single-task to multi-task learning (STL to MTL). To address these limitations, we propose adaptive shared experts (ASE) within a low-rank adaptation (LoRA) based MoE, where shared experts are assigned router-computed gating weights jointly normalized with sparse experts. This design facilitates STL to MTL transition, enhances expert specialization, and cooperation. Furthermore, we incorporate fine-grained experts by increasing the number of LoRA experts while proportionally reducing their rank, enabling more effective knowledge sharing under a comparable parameter budget. Extensive experiments on the PASCAL-Context benchmark, under unified training settings, demonstrate that ASE consistently improves performance across diverse configurations and validates the effectiveness of fine-grained designs for MTL.", "authors": ["Minghao Yang", "Ren Togo", "Guang Li", "Takahiro Ogawa", "Miki Haseyama"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-01", "url": "https://arxiv.org/abs/2510.00570", "pdf_url": "https://arxiv.org/pdf/2510.00570v1", "arxiv_id": "2510.00570", "doi": "10.48550/arXiv.2510.00570", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.3231}
{"id": "9b2784eed88fdd9f06872e9abda7244b6d88e09374aec46bcf0f5d453ff8605d", "sources": ["arxiv", "semantic_scholar"], "title": "GUI-KV: Efficient GUI Agents via KV Cache with Spatio-Temporal Awareness", "abstract": "Graphical user interface (GUI) agents built on vision-language models have emerged as a promising approach to automate human-computer workflows. However, they also face the inefficiency challenge as they process long sequences of high-resolution screenshots and solving long-horizon tasks, making inference slow, costly and memory-bound. While key-value (KV) caching can mitigate this, storing the full cache is prohibitive for image-heavy contexts. Existing cache-compression methods are sub-optimal as they do not account for the spatial and temporal redundancy of GUIs. In this work, we first analyze attention patterns in GUI agent workloads and find that, unlike in natural images, attention sparsity is uniformly high across all transformer layers. This insight motivates a simple uniform budget allocation strategy, which we show empirically outperforms more complex layer-varying schemes. Building on this, we introduce GUI-KV, a plug-and-play KV cache compression method for GUI agents that requires no retraining. GUI-KV combines two novel techniques: (i) spatial saliency guidance, which augments attention scores with the L2 norm of hidden states to better preserve semantically important visual tokens, and (ii) temporal redundancy scoring, which projects previous frames' keys onto the current frame's key subspace to preferentially prune redundant history. Across standard GUI agent benchmarks and models, GUI-KV outperforms competitive KV compression baselines, closely matching full-cache accuracy at modest budgets. Notably, in a 5-screenshot setting on the AgentNetBench benchmark, GUI-KV reduces decoding FLOPs by 38.9% while increasing step accuracy by 4.1% over the full-cache baseline. These results demonstrate that exploiting GUI-specific redundancies enables efficient and reliable agent performance.", "authors": ["Kung-Hsiang Huang", "Haoyi Qiu", "Yutong Dai", "Caiming Xiong", "Chien-Sheng Wu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-01", "url": "https://arxiv.org/abs/2510.00536", "pdf_url": "https://arxiv.org/pdf/2510.00536v1", "arxiv_id": "2510.00536", "doi": "10.48550/arXiv.2510.00536", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3231}
{"id": "2ca47670badc8501fb5f6092359cd95a1f59e0922fd2289021f1b4736107b428", "sources": ["arxiv", "semantic_scholar"], "title": "ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning Models", "abstract": "The long-output context generation of large reasoning models enables extended chain of thought (CoT) but also drives rapid growth of the key-value (KV) cache, quickly overwhelming GPU memory. To address this challenge, we propose ThinKV, a thought-adaptive KV cache compression framework. ThinKV is based on the observation that attention sparsity reveals distinct thought types with varying importance within the CoT. It applies a hybrid quantization-eviction strategy, assigning token precision by thought importance and progressively evicting tokens from less critical thoughts as reasoning trajectories evolve. Furthermore, to implement ThinKV, we design a kernel that extends PagedAttention to enable efficient reuse of evicted tokens' memory slots, eliminating compaction overheads. Extensive experiments on DeepSeek-R1-Distill, GPT-OSS, and NVIDIA AceReason across mathematics and coding benchmarks show that ThinKV achieves near-lossless accuracy with less than 5% of the original KV cache, while improving performance with up to 5.8x higher inference throughput over state-of-the-art baselines.", "authors": ["Akshat Ramachandran", "Marina Neseem", "Charbel Sakr", "Rangharajan Venkatesan", "Brucek Khailany", "Tushar Krishna"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-01", "url": "https://arxiv.org/abs/2510.01290", "pdf_url": "https://arxiv.org/pdf/2510.01290v2", "arxiv_id": "2510.01290", "doi": "10.48550/arXiv.2510.01290", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3231}
{"id": "7d77908d9c55a5b0fa5f76092bb7cdc3b18503eade304bf19415670ebbcde5ad", "sources": ["arxiv", "semantic_scholar"], "title": "Can Mamba Learn In Context with Outliers? A Theoretical Generalization Analysis", "abstract": "The Mamba model has gained significant attention for its computational advantages over Transformer-based models, while achieving comparable performance across a wide range of language tasks. Like Transformers, Mamba exhibits in-context learning (ICL) capabilities, i.e., making predictions for new tasks based on a prompt containing input-label pairs and a query, without requiring fine-tuning. Despite its empirical success, the theoretical understanding of Mamba remains limited, largely due to the nonlinearity introduced by its gating mechanism. To the best of our knowledge, this paper presents the first theoretical analysis of the training dynamics of a one-layer Mamba model, which consists of a linear attention component followed by a nonlinear gating layer, and its ICL generalization on unseen binary classification tasks, even when the prompt includes additive outliers. Our analysis shows that Mamba leverages the linear attention layer to select informative context examples and uses the nonlinear gating layer to suppress the influence of outliers. By establishing and comparing to the analysis of linear Transformers under the same setting, we show that although Mamba may require more training iterations to converge, it maintains accurate predictions even when the proportion of outliers exceeds the threshold that a linear Transformer can tolerate. These theoretical findings are supported by empirical experiments.", "authors": ["Hongkang Li", "Songtao Lu", "Xiaodong Cui", "Pin-Yu Chen", "Meng Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-01", "url": "https://arxiv.org/abs/2510.00399", "pdf_url": "https://arxiv.org/pdf/2510.00399v1", "arxiv_id": "2510.00399", "doi": "10.48550/arXiv.2510.00399", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3231}
{"id": "ef63e89fc84f1d521c062ef328712cd9007508578cb91933ffadc9b63da846aa", "sources": ["arxiv", "semantic_scholar"], "title": "LD-MoLE: Learnable Dynamic Routing for Mixture of LoRA Experts", "abstract": "Recent studies have shown that combining parameter-efficient fine-tuning (PEFT) with mixture-of-experts (MoE) is an effective strategy for adapting large language models (LLMs) to the downstream tasks. However, most existing approaches rely on conventional TopK routing, which requires careful hyperparameter tuning and assigns a fixed number of experts to each token. In this work, we propose LD-MoLE, a Learnable Dynamic routing mechanism for Mixture of LoRA Experts that enables adaptive, token-dependent, and layer-wise expert allocation. Our method replaces the non-differentiable TopK selection with a differentiable routing function and a closed-form solution. Moreover, our design allows the model to adaptively determine the number of experts to activate for each token at different layers. In addition, we introduce an analytical sparsity control objective to regularize the number of activated experts. Extensive experiments on the Qwen3-1.7B and Llama-3.2-3B models show that LD-MoLE achieves the highest average scores compared to state-of-the-art baselines, across a diverse set of benchmarks. Our method not only achieves superior performance, but also demonstrates the ability to learn token-dependent and layer-wise expert allocation.", "authors": ["Yuan Zhuang", "Yi Shen", "Yuexin Bian", "Qing Su", "Shihao Ji", "Yuanyuan Shi", "Fei Miao"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2509.25684", "pdf_url": "https://arxiv.org/pdf/2509.25684v2", "arxiv_id": "2509.25684", "doi": "10.48550/arXiv.2509.25684", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.322}
{"id": "b97036dd5c0e9e17e9e41772089d05cef3fe926b405556bf80c35556a6fae01f", "sources": ["arxiv", "semantic_scholar"], "title": "MOLM: Mixture of LoRA Markers", "abstract": "Generative models can generate photorealistic images at scale. This raises urgent concerns about the ability to detect synthetically generated images and attribute these images to specific sources. While watermarking has emerged as a possible solution, existing methods remain fragile to realistic distortions, susceptible to adaptive removal, and expensive to update when the underlying watermarking key changes. We propose a general watermarking framework that formulates the encoding problem as key-dependent perturbation of the parameters of a generative model. Within this framework, we introduce Mixture of LoRA Markers (MOLM), a routing-based instantiation in which binary keys activate lightweight LoRA adapters inside residual and attention blocks. This design avoids key-specific re-training and achieves the desired properties such as imperceptibility, fidelity, verifiability, and robustness. Experiments on Stable Diffusion and FLUX show that MOLM preserves image quality while achieving robust key recovery against distortions, compression and regeneration, averaging attacks, and black-box adversarial attacks on the extractor.", "authors": ["Samar Fares", "Nurbek Tastan", "Noor Hussein", "Karthik Nandakumar"], "categories": ["cs.CV", "cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2510.00293", "pdf_url": "https://arxiv.org/pdf/2510.00293v2", "arxiv_id": "2510.00293", "doi": "10.48550/arXiv.2510.00293", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.322}
{"id": "cdd45d17cb21c8df347da7aefdc35aed00f4bbba3a4739bc6d4ac22b9333a84a", "sources": ["arxiv", "semantic_scholar"], "title": "The Pitfalls of KV Cache Compression", "abstract": "KV cache compression promises increased throughput and efficiency with negligible loss in performance. While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multi-instruction prompting have been insufficiently studied. In this paper, we identify several pitfalls that practitioners should be aware of when deploying KV cache compressed LLMs. We evaluate five KV cache compression methods (StreamingLLM, SnapKV, TOVA, H2O, and K-Norm) on Llama3.1 8B and Qwen2.5 14B under multi-instruction prompting with IFEval. Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM. As a practical example, we highlight system prompt leakage as a case study, empirically demonstrating the impact of compression on leakage and general instruction-following. We identify several factors that contribute to system prompt leakage: compression method, instruction order, and KV eviction bias. We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks.", "authors": ["Alex Chen", "Renato Geh", "Aditya Grover", "Guy Van den Broeck", "Daniel Israel"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2510.00231", "pdf_url": "https://arxiv.org/pdf/2510.00231v2", "arxiv_id": "2510.00231", "doi": "10.48550/arXiv.2510.00231", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.322}
{"id": "75481c89196c634a7098b558dd2db919cdf5acbae19ab3b0823925b79be2fcf4", "sources": ["arxiv", "semantic_scholar"], "title": "The silence of the weights: a structural pruning strategy for attention-based audio signal architectures with second order metrics", "abstract": "Transformer-based models have become the state of the art across multiple domains, from natural language processing to machine listening, thanks to the attention mechanisms. However, the attention layers require a large number of parameters and high-end hardware for both training and inference. We propose a novel channel-pruning technique explicitly targeted at the attention mechanism, decoupling the pruning of each head and the four layers in the attention block: query, key, value, and output projection matrices, employing a second-order metric to score the network's parameters. We compare our technique against head-pruning strategies and magnitude-driven scoring metrics, investigating the effects of pruning on Audio Spectrogram Transformer (AST) and Whisper. Our results show that even after pruning 50\\% of the parameters in the attention block, performance is largely preserved.", "authors": ["Andrea Diecidue", "Carlo Alberto Barbano", "Piero Fraternali", "Mathieu Fontaine", "Enzo Tartaglione"], "categories": ["cs.SD", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2509.26207", "pdf_url": "https://arxiv.org/pdf/2509.26207v2", "arxiv_id": "2509.26207", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2049}
{"id": "f33955c65579c67742c7a535942e194ff3b71dc1f1138dfb5c2a23b0c7cb13d3", "sources": ["arxiv", "semantic_scholar"], "title": "Training Matryoshka Mixture-of-Experts for Elastic Inference-Time Expert Utilization", "abstract": "Mixture-of-Experts (MoE) has emerged as a promising paradigm for efficiently scaling large language models without a proportional increase in computational cost. However, the standard training strategy of Top-K router prevents MoE models from realizing their full potential for elastic inference. When the number of activated experts is altered at inference time, these models exhibit precipitous performance degradation. In this work, we introduce Matryoshka MoE (M-MoE), a training framework that instills a coarse-to-fine structure directly into the expert ensemble. By systematically varying the number of activated experts during training, M-MoE compels the model to learn a meaningful ranking: top-ranked experts collaborate to provide essential, coarse-grained capabilities, while subsequent experts add progressively finer-grained detail. We explore this principle at multiple granularities, identifying a layer-wise randomization strategy as the most effective. Our experiments demonstrate that a single M-MoE model achieves remarkable elasticity, with its performance at various expert counts closely matching that of an entire suite of specialist models, but at only a fraction of the total training cost. This flexibility not only unlocks elastic inference but also enables optimizing performance by allocating different computational budgets to different model layers. Our work paves the way for more practical and adaptable deployments of large-scale MoE models.", "authors": ["Yaoxiang Wang", "Qingguo Hu", "Yucheng Ding", "Ruizhe Wang", "Yeyun Gong", "Jian Jiao", "Yelong Shen", "Peng Cheng", "Jinsong Su"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2509.26520", "pdf_url": "https://arxiv.org/pdf/2509.26520v1", "arxiv_id": "2509.26520", "doi": "10.48550/arXiv.2509.26520", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.322}
{"id": "61191a6e3bfa5a700f3074a9320719fab3e34b15f9492380ee0f3921c0ce2601", "sources": ["arxiv", "semantic_scholar"], "title": "PrunedLoRA: Robust Gradient-Based structured pruning for Low-rank Adaptation in Fine-tuning", "abstract": "Low-rank adaptation (LoRA) has become a widely used paradigm for parameter-efficient fine-tuning of large language models, yet its representational capacity often lags behind full fine-tuning. Within the context of LoRA, a key open question is how to obtain expressive low-rank adapters from over-parameterized spaces. We propose \\textit{PrunedLoRA}, a new framework that leverages structured pruning to obtain highly representative low-rank adapters from an over-parameterized initialization. Unlike prior approaches that impose a fixed low-rank budget, PrunedLoRA dynamically prunes less important components during fine-tuning and prevents their reactivation, enabling flexible and adaptive rank allocation. For structured pruning, by minimizing the pruning error for overall loss, we provide fine-grained pruning and recovery updates in a gradient-based pruning strategy with grounded interpretation. We provide the first theoretical analysis of the robustness of structured pruning and provably show that under the impact of weight perturbation, gradient-based pruning is more robust than activation-based pruning with respect to overall loss. Empirically, PrunedLoRA consistently outperforms LoRA and its variants across supervised fine-tuning tasks in mathematical reasoning, code generation, and natural language understanding, and it also demonstrates advantages over existing structured pruning methods across diverse sparsity levels.", "authors": ["Xin Yu", "Cong Xie", "Ziyu Zhao", "Tiantian Fan", "Lingzhou Xue", "Zhi Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2510.00192", "pdf_url": "https://arxiv.org/pdf/2510.00192v2", "arxiv_id": "2510.00192", "doi": "10.48550/arXiv.2510.00192", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.322}
{"id": "190d4cbd0318ad77662df7b8bf792df9db6114899783f00da39c5b48da0002de", "sources": ["arxiv", "semantic_scholar"], "title": "One-Prompt Strikes Back: Sparse Mixture of Experts for Prompt-based Continual Learning", "abstract": "Prompt-based methods have recently gained prominence in Continual Learning (CL) due to their strong performance and memory efficiency. A prevalent strategy in this paradigm assigns a dedicated subset of prompts to each task, which, while effective, incurs substantial computational overhead and causes memory requirements to scale linearly with the number of tasks. Conversely, approaches employing a single shared prompt across tasks offer greater efficiency but often suffer from degraded performance due to knowledge interference. To reconcile this trade-off, we propose SMoPE, a novel framework that integrates the benefits of both task-specific and shared prompt strategies. Inspired by recent findings on the relationship between Prefix Tuning and Mixture of Experts (MoE), SMoPE organizes a shared prompt into multiple \"prompt experts\" within a sparse MoE architecture. For each input, only a select subset of relevant experts is activated, effectively mitigating interference. To facilitate expert selection, we introduce a prompt-attention score aggregation mechanism that computes a unified proxy score for each expert, enabling dynamic and sparse activation. Additionally, we propose an adaptive noise mechanism to encourage balanced expert utilization while preserving knowledge from prior tasks. To further enhance expert specialization, we design a prototype-based loss function that leverages prefix keys as implicit memory representations. Extensive experiments across multiple CL benchmarks demonstrate that SMoPE consistently outperforms task-specific prompt methods and achieves performance competitive with state-of-the-art approaches, all while significantly reducing parameter counts and computational costs.", "authors": ["Minh Le", "Bao-Ngoc Dao", "Huy Nguyen", "Quyen Tran", "Anh Nguyen", "Nhat Ho"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-29", "url": "https://arxiv.org/abs/2509.24483", "pdf_url": "https://arxiv.org/pdf/2509.24483v3", "arxiv_id": "2509.24483", "doi": "10.48550/arXiv.2509.24483", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3208}
{"id": "8deb5a6f0f24ff37abffa7cf8cb7f64eb7d938cfa5049b48065142d50bf85970", "sources": ["arxiv", "semantic_scholar"], "title": "From Score Distributions to Balance: Plug-and-Play Mixture-of-Experts Routing", "abstract": "Mixture-of-Experts (MoE) models can scale parameter capacity by routing each token to a subset of experts through a learned gate function. While conditional routing reduces training costs, it shifts the burden on inference memory: expert parameters and activations consume memory, limiting the number of experts per device. As tokens are routed, some experts become overloaded while others are underutilized. Because experts are mapped to GPUs, this imbalance translates directly into degraded system performance in terms of latency, throughput, and cost. We present LASER, a plug-and-play, inference-time routing algorithm that balances load while preserving accuracy. LASER adapts to the shape of the gate's score distribution. When scores provide a clear preference, it routes to the strongest experts; when scores are more uniform, it broadens the set of viable experts and routes to the least-loaded among them. Because LASER relies only on gate scores from a trained model, it integrates directly into existing MoE inference pipelines without retraining or finetuning. We evaluate LASER on Mixtral-8x7B and DeepSeek-MoE-16b-chat across four datasets (ARC-Easy, ARC-Challenge, MMLU, and GSM8K). LASER improves load balancing, translating into lower latency and higher throughput, while keeping the accuracy changes negligible.", "authors": ["Rana Shahout", "Colin Cai", "Yilun Du", "Minlan Yu", "Michael Mitzenmacher"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-29", "url": "https://arxiv.org/abs/2510.03293", "pdf_url": "https://arxiv.org/pdf/2510.03293v1", "arxiv_id": "2510.03293", "doi": "10.48550/arXiv.2510.03293", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3208}
{"id": "1a7358276cd7cce4b6146d004f458ad6f15d7b9a545939f4b98f128136932966", "sources": ["arxiv", "semantic_scholar"], "title": "QuantSparse: Comprehensively Compressing Video Diffusion Transformer with Model Quantization and Attention Sparsification", "abstract": "Diffusion transformers exhibit remarkable video generation capability, yet their prohibitive computational and memory costs hinder practical deployment. Model quantization and attention sparsification are two promising directions for compression, but each alone suffers severe performance degradation under aggressive compression. Combining them promises compounded efficiency gains, but naive integration is ineffective. The sparsity-induced information loss exacerbates quantization noise, leading to amplified attention shifts. To address this, we propose \\textbf{QuantSparse}, a unified framework that integrates model quantization with attention sparsification. Specifically, we introduce \\textit{Multi-Scale Salient Attention Distillation}, which leverages both global structural guidance and local salient supervision to mitigate quantization-induced bias. In addition, we develop \\textit{Second-Order Sparse Attention Reparameterization}, which exploits the temporal stability of second-order residuals to efficiently recover information lost under sparsity. Experiments on HunyuanVideo-13B demonstrate that QuantSparse achieves 20.88 PSNR, substantially outperforming the state-of-the-art quantization baseline Q-VDiT (16.85 PSNR), while simultaneously delivering a \\textbf{3.68$\\times$} reduction in storage and \\textbf{1.88$\\times$} acceleration in end-to-end inference. Our code will be released in https://github.com/wlfeng0509/QuantSparse.", "authors": ["Weilun Feng", "Chuanguang Yang", "Haotong Qin", "Mingqiang Wu", "Yuqi Li", "Xiangqi Li", "Zhulin An", "Libo Huang", "Yulun Zhang", "Michele Magno", "Yongjun Xu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-28", "url": "https://arxiv.org/abs/2509.23681", "pdf_url": "https://arxiv.org/pdf/2509.23681v4", "arxiv_id": "2509.23681", "doi": "10.48550/arXiv.2509.23681", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/wlfeng0509/QuantSparse", "venue": "arXiv.org", "quality_score": 0.4941}
{"id": "2b37cee9eed4c2be861141ad559f2b8314b0129de6488e667b696da551b15a9d", "sources": ["arxiv", "semantic_scholar"], "title": "d$^2$Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching", "abstract": "Diffusion-based large language models (dLLMs), despite their promising performance, still suffer from inferior inference efficiency. This is because dLLMs rely on bidirectional attention and cannot directly benefit from the standard key-value (KV) cache as autoregressive models (ARMs) do. To tackle this issue, we introduce \\textit{Dual aDaptive Cache} (d$^2$Cache), which is a training-free approximate KV cache framework for accelerating dLLM inference. d$^2$Cache features a two-stage fine-grained selection strategy to identify tokens and adaptively update their KV states at each decoding step, while caching the KV states of the remaining tokens for reuse. Furthermore, d$^2$Cache naturally offers a more reliable decoding alternative, which can enable quasi left-to-right generation and mitigate premature overconfidence in tokens at the end of the sequence. Extensive experimental results on two representative dLLMs (\\ie, LLaDA and Dream) demonstrate that d$^2$Cache not only achieves substantial inference speedups, but also yields consistent improvements in generation quality. The code is available at https://github.com/Kamichanw/d2Cache.", "authors": ["Yuchu Jiang", "Yue Cai", "Xiangzhong Luo", "Jiale Fu", "Jiarui Wang", "Chonghan Liu", "Xu Yang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-27", "url": "https://arxiv.org/abs/2509.23094", "pdf_url": "https://arxiv.org/pdf/2509.23094v2", "arxiv_id": "2509.23094", "doi": "10.48550/arXiv.2509.23094", "citation_count": 19, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Kamichanw/d2Cache", "venue": "arXiv.org", "quality_score": 0.4923}
{"id": "f65fa38a26c2131b771b5c8ae6930e95561c341a89782fd8aab4894e6e699e9c", "sources": ["arxiv", "semantic_scholar"], "title": "HEAPr: Hessian-based Efficient Atomic Expert Pruning in Output Space", "abstract": "Mixture-of-Experts (MoE) architectures in large language models (LLMs) deliver exceptional performance and reduced inference costs compared to dense LLMs. However, their large parameter counts result in prohibitive memory requirements, limiting practical deployment. While existing pruning methods primarily focus on expert-level pruning, this coarse granularity often leads to substantial accuracy degradation. In this work, we introduce HEAPr, a novel pruning algorithm that decomposes experts into smaller, indivisible atomic experts, enabling more precise and flexible atomic expert pruning. To measure the importance of each atomic expert, we leverage second-order information based on principles similar to the Optimal Brain Surgeon theory. To address the computational and storage challenges posed by second-order information, HEAPr exploits the inherent properties of atomic experts to transform the second-order information from expert parameters into that of atomic expert parameters, and further simplifies it to the second-order information of atomic expert outputs. This approach reduces the space complexity from $O(d^4)$, where $d$ is the model's dimensionality, to $O(d^2)$. HEAPr requires only two forward passes and one backward pass on a small calibration set to compute the importance of atomic experts. Extensive experiments on MoE models, including DeepSeek MoE and Qwen MoE family, demonstrate that HEAPr outperforms existing expert-level pruning methods across a wide range of pruning ratios and benchmarks. Specifically, HEAPr achieves nearly lossless compression at pruning ratios of 20% ~ 25% in most models, while also reducing FLOPs nearly by 20%. The code can be found at [https://github.com/LLIKKE/HEAPr](https://github.com/LLIKKE/HEAPr).", "authors": ["Ke Li", "Zheng Yang", "Zhongbin Zhou", "Feng Xue", "Zhonglin Jiang", "Wenxiao Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.22299", "pdf_url": "https://arxiv.org/pdf/2509.22299v3", "arxiv_id": "2509.22299", "doi": "10.48550/arXiv.2509.22299", "citation_count": 1, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/LLIKKE/HEAPr", "venue": "arXiv.org", "quality_score": 0.4905}
{"id": "7bfe590415727f390e1af8c4c16da7663c8818002440d4b4e45cb407cc316506", "sources": ["arxiv", "semantic_scholar"], "title": "Elastic MoE: Unlocking the Inference-Time Scalability of Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) models typically fix the number of activated experts $k$ at both training and inference. However, real-world deployments often face heterogeneous hardware, fluctuating workloads, and diverse quality-latency requirements, while training separate models for each scenario is costly. Considering that MoE models already operate with sparse activation, adjusting the number of activated experts offers a natural path to serving diverse budgets with a single model. Yet, we find that activating more experts $k'$ ($> k$) at inference does not yield the expected gains. Instead, performance degrades rapidly after only a slight increase, a phenomenon we term the \\textit{inference-time scaling wall}. Further investigation reveals that this degradation stems from a lack of learned collaboration among experts. To address this, we introduce \\textbf{Elastic Mixture-of-Experts (EMoE)}, a novel training framework that enables MoE models to elastically vary the number of activated experts at inference. By simultaneously training experts to collaborate in diverse combinations and encouraging the router to make high-quality selections, EMoE ensures robust performance across inference budgets. Extensive experiments across four MoE architectures (7B--21B) and nine benchmarks show that EMoE significantly expands the effective scaling range to 2-3$\\times$ the training-time $k$, while also achieving higher peak performance.", "authors": ["Naibin Gu", "Zhenyu Zhang", "Yuchen Feng", "Yilong Chen", "Peng Fu", "Zheng Lin", "Shuohuan Wang", "Yu Sun", "Hua Wu", "Weiping Wang", "Haifeng Wang"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.21892", "pdf_url": "https://arxiv.org/pdf/2509.21892v2", "arxiv_id": "2509.21892", "doi": "10.48550/arXiv.2509.21892", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3174}
{"id": "bdd0d701020c0d7ead0f0118c47d1168829419e0a3f9c969cc6d32764c4c5ed4", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Experts Search: Enhancing Reasoning in Mixture-of-Experts LLMs at Test Time", "abstract": "Test-Time Scaling (TTS) enhances the reasoning ability of large language models (LLMs) by allocating additional computation during inference. However, existing approaches primarily rely on output-level sampling while overlooking the role of model architecture. In mainstream Mixture-of-Experts (MoE) LLMs, we observe that varying the number of activated experts yields complementary solution sets with stable accuracy, revealing a new and underexplored source of diversity. Motivated by this observation, we propose Dynamic Experts Search (DES), a TTS strategy that elevates expert activation into a controllable dimension of the search space. DES integrates two key components: (1) Dynamic MoE, which enables direct control of expert counts during inference to generate diverse reasoning trajectories without additional cost; and (2) Expert Configuration Inheritance, which preserves consistent expert counts within a reasoning path while varying them across runs, thereby balancing stability and diversity throughout the search. Extensive experiments across MoE architectures, verifiers and reasoning benchmarks (i.e., math, code and knowledge) demonstrate that DES reliably outperforms TTS baselines, enhancing accuracy and stability without additional cost. These results highlight DES as a practical and scalable form of architecture-aware TTS, illustrating how structural flexibility in modern LLMs can advance reasoning.", "authors": ["Yixuan Han", "Fan Ma", "Ruijie Quan", "Yi Yang"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.22572", "pdf_url": "https://arxiv.org/pdf/2509.22572v1", "arxiv_id": "2509.22572", "doi": "10.48550/arXiv.2509.22572", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3174}
{"id": "35e30bccc6936cfd06cd5cc58d27f06c05a9248f24abc8a2f68af6ed07fc6ff2", "sources": ["arxiv", "semantic_scholar"], "title": "OjaKV: Context-Aware Online Low-Rank KV Cache Compression", "abstract": "The expanding long-context capabilities of large language models are constrained by a significant memory bottleneck: the key-value (KV) cache required for autoregressive generation. This bottleneck is substantial; for instance, a Llama-3.1-8B model processing a 32K-token prompt at a batch size of 4 requires approximately 16GB for its KV cache, a size exceeding the model's weights. While KV-cache compression via low-rank projection is a promising direction, existing methods rely on a static, offline-learned subspace that performs poorly under data distribution shifts. To overcome these limitations, we introduce OjaKV, a novel framework that integrates a strategic hybrid storage policy with online subspace adaptation. First, OjaKV recognizes that not all tokens are equally important for compression; it preserves the crucial first and most recent tokens in full-rank, maintaining high-fidelity anchors for attention. Second, for the vast majority of intermediate tokens, it applies low-rank compression by incrementally adapting the projection basis using Oja's algorithm for online principal component analysis. This adaptation involves a comprehensive update during prompt prefilling and lightweight periodic updates during decoding, ensuring the subspace remains aligned with the evolving context. Crucially, our framework is fully compatible with modern attention modules like FlashAttention. Experiments demonstrate that OjaKV maintains or even improves zero-shot accuracy at high compression ratios. In particular, OjaKV achieves its strongest gains on very long-context benchmarks that require complex reasoning, highlighting the importance of online subspace adaptation in dynamically tracking context shifts. These results establish our hybrid framework as a practical, plug-and-play solution for memory-efficient long-context inference without requiring model fine-tuning.", "authors": ["Yuxuan Zhu", "David H. Yang", "Mohammad Mohammadi Amiri", "Keerthiram Murugesan", "Tejaswini Pedapati", "Pin-Yu Chen"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-25", "url": "https://arxiv.org/abs/2509.21623", "pdf_url": "https://arxiv.org/pdf/2509.21623v2", "arxiv_id": "2509.21623", "doi": "10.48550/arXiv.2509.21623", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3162}
{"id": "59e988fbb18ddb31554e7eca25745d361d45c4d1b553a6fb4d6dd6fb5a91cb00", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Thoughts: Learning to Aggregate What Experts Think, Not Just What They Say", "abstract": "Open-source Large Language Models (LLMs) increasingly specialize by domain (e.g., math, code, general reasoning), motivating systems that leverage complementary strengths across models. Prior multi-LLM approaches either (i) route a query to one or a few experts and generate independently, (ii) aggregate outputs from each model via costly multi-turn exchanges, or (iii) fuse weights into a single model-typically requiring architectural homogeneity. We introduce Mixture of Thoughts (MoT), a simple method for latent-level collaboration among heterogeneous experts under a global routing scheme. For each query, a lightweight router selects top-$K$ experts and designates a primary expert; uniformly placed interaction layers project hidden states into a shared latent space where the primary expert performs cross-attention over its active (selected) peers. Pre-trained experts remain frozen; only the router and the lightweight interaction layers are trained with a novel joint training objective that improves both the expert selection and inter-expert collaboration. Across five in-distribution (ID) and three out-of-distribution (OOD) benchmarks, MoT surpasses the current routing and aggregation-based state-of-the-art, Avengers, by $+0.38\\%$ and $+2.92\\%$, respectively. Further, MoT significantly outperforms the best-performing single model. It achieves this with single-pass inference, runtime comparable to routing baselines, and none of the overheads of iterative aggregation. MoT offers a simple latent-space mechanism for combining heterogeneous LLMs, a practical step toward broader multi-LLM collaboration. Our code is publicly available at https://github.com/jacobfa/mot.", "authors": ["Jacob Fein-Ashley", "Dhruv Parikh", "Rajgopal Kannan", "Viktor Prasanna"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-25", "url": "https://arxiv.org/abs/2509.21164", "pdf_url": "https://arxiv.org/pdf/2509.21164v1", "arxiv_id": "2509.21164", "doi": "10.48550/arXiv.2509.21164", "citation_count": 8, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jacobfa/mot", "venue": "arXiv.org", "quality_score": 0.4887}
{"id": "70dd8b7d5a7dba6c990b386da17f6845f59b4bf28ac54da7800d8755d5ada458", "sources": ["arxiv", "semantic_scholar"], "title": "A Reformulation of UVN-Flash for Multicomponent Two-Phase Systems with Application to CO2-rich Mixture Transport in Pipelines", "abstract": "Pipeline transport of dense-phase CO2-rich mixtures is a crucial component in carbon capture and storage (CCS). Accurate modeling requires coupling of fluid dynamics and thermodynamics, especially during transient events such as depressurization. In this work, we present a unified framework for two-phase multicomponent transport in pipelines that integrates both aspects. Specifically, we employ the homogeneous equilibrium model (HEM) for modeling the transport of two-phase CO2-rich mixture, with thermodynamic closure provided by a Helmholtz energy-based equation of state. Phase equilibrium calculations are performed using UVN-flash, supplemented with a stability analysis procedure to detect phase separation and generate initial guesses for the phase-equilibrium calculations. Specifically, we introduce a novel tailored UVN-flash routine that aligns with the fluid dynamics formulation. This is achieved by introducing an alternative and better-scaled set of variables for the phase-equilibrium calculations. The proposed framework is applied to the depressurization of tanks and pipelines containing CO2-rich mixtures, demonstrating its effectiveness for CCS-relevant applications.", "authors": ["Pardeep Kumar", "Patricio I. Rosen Esquivel"], "categories": ["physics.comp-ph", "math.NA", "physics.flu-dyn"], "fields_of_study": ["Computer Science", "Physics", "Mathematics"], "published_date": "2025-09-25", "url": "https://arxiv.org/abs/2509.20965", "pdf_url": "https://arxiv.org/pdf/2509.20965v1", "arxiv_id": "2509.20965", "doi": "10.48550/arXiv.2509.20965", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3162}
{"id": "085f81014b5b1234d148d1fca33b857a28b0bb776696b9c27adc1afc90b77fe7", "sources": ["arxiv", "semantic_scholar"], "title": "SHMoAReg: Spark Deformable Image Registration via Spatial Heterogeneous Mixture of Experts and Attention Heads", "abstract": "Encoder-Decoder architectures are widely used in deep learning-based Deformable Image Registration (DIR), where the encoder extracts multi-scale features and the decoder predicts deformation fields by recovering spatial locations. However, current methods lack specialized extraction of features (that are useful for registration) and predict deformation jointly and homogeneously in all three directions. In this paper, we propose a novel expert-guided DIR network with Mixture of Experts (MoE) mechanism applied in both encoder and decoder, named SHMoAReg. Specifically, we incorporate Mixture of Attention heads (MoA) into encoder layers, while Spatial Heterogeneous Mixture of Experts (SHMoE) into the decoder layers. The MoA enhances the specialization of feature extraction by dynamically selecting the optimal combination of attention heads for each image token. Meanwhile, the SHMoE predicts deformation fields heterogeneously in three directions for each voxel using experts with varying kernel sizes. Extensive experiments conducted on two publicly available datasets show consistent improvements over various methods, with a notable increase from 60.58% to 65.58% in Dice score for the abdominal CT dataset. Furthermore, SHMoAReg enhances model interpretability by differentiating experts' utilities across/within different resolution layers. To the best of our knowledge, we are the first to introduce MoE mechanism into DIR tasks. The code will be released soon.", "authors": ["Yuxi Zheng", "Jianhui Feng", "Tianran Li", "Marius Staring", "Yuchuan Qiao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-24", "url": "https://arxiv.org/abs/2509.20073", "pdf_url": "https://arxiv.org/pdf/2509.20073v1", "arxiv_id": "2509.20073", "doi": "10.1109/BIBM66473.2025.11356999", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Bioinformatics and Biomedicine", "quality_score": 0.3151}
{"id": "98e1eaf6fb53e80d3592c2e71c4234dd674820a64932df3adea5a588fea0de74", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Reasoning Chains through Depth-Specialized Mixture-of-Experts in Transformer Architectures", "abstract": "Contemporary transformer architectures apply identical processing depth to all inputs, creating inefficiencies and limiting reasoning quality. Simple factual queries are subjected to the same multilayered computation as complex logical problems, wasting resources while constraining deep inference. To overcome this, we came up with a concept of Dynamic Reasoning Chains through Depth Specialised Mixture of Experts (DS-MoE), a modular framework that extends the Mixture of Experts paradigm from width-based to depth specialised computation. DS-MoE introduces expert modules optimised for distinct reasoning depths, shallow pattern recognition, compositional reasoning, logical inference, memory integration, and meta-cognitive supervision. A learned routing network dynamically assembles custom reasoning chains, activating only the necessary experts to match input complexity. The dataset on which we trained and evaluated DS-MoE is on The Pile, an 800GB corpus covering diverse domains such as scientific papers, legal texts, programming code, and web content, enabling systematic assessment across reasoning depths. Experimental results demonstrate that DS-MoE achieves up to 16 per cent computational savings and 35 per cent faster inference compared to uniform-depth transformers, while delivering 2.8 per cent higher accuracy on complex multi-step reasoning benchmarks. Furthermore, routing decisions yield interpretable reasoning chains, enhancing transparency and scalability. These findings establish DS-MoE as a significant advancement in adaptive neural architectures, demonstrating that depth-specialised modular processing can simultaneously improve efficiency, reasoning quality, and interpretability in large-scale language models.", "authors": ["Sampurna Roy", "Ayan Sar", "Anurag Kaushish", "Kanav Gupta", "Tanupriya Choudhury", "Abhijit Kumar"], "categories": ["cs.CL", "cs.AI", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-24", "url": "https://arxiv.org/abs/2509.20577", "pdf_url": "https://arxiv.org/pdf/2509.20577v1", "arxiv_id": "2509.20577", "doi": "10.1109/BigData66926.2025.11401520", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "BigData Congress [Services Society]", "quality_score": 0.3151}
{"id": "f33389032cb9bbdf15e5d037727dcb410c140647fafa44f8cd6199e2908d58f5", "sources": ["arxiv", "semantic_scholar"], "title": "Symphony-MoE: Harmonizing Disparate Pre-trained Models into a Coherent Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) models enable scalable performance by activating large parameter sets sparsely, minimizing computational overhead. To mitigate the prohibitive cost of training MoEs from scratch, recent work employs upcycling, reusing a single pre-trained dense model by replicating its feed-forward network (FFN) layers into experts. However, this limits expert diversity, as all experts originate from a single pre-trained dense model. This paper addresses this limitation by constructing powerful MoE models using experts sourced from multiple identically-architected but disparate pre-trained models (e.g., Qwen2.5-Coder and Qwen2). A key challenge lies in the fact that these source models occupy disparate, dissonant regions of the parameter space, making direct upcycling prone to severe performance degradation. To overcome this, we propose Symphony-MoE, a novel two-stage framework designed to harmonize these models into a single, coherent expert mixture. First, we establish this harmony in a training-free manner: we construct a shared backbone via a layer-aware fusion strategy and, crucially, alleviate parameter misalignment among experts using activation-based functional alignment. Subsequently, a stage of post-training coordinates the entire architecture. Experiments demonstrate that our method successfully integrates experts from heterogeneous sources, achieving an MoE model that significantly surpasses baselines in multi-domain tasks and out-of-distribution generalization.", "authors": ["Qi Wang", "Hanyang Peng", "Yue Yu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-23", "url": "https://arxiv.org/abs/2509.18542", "pdf_url": "https://arxiv.org/pdf/2509.18542v2", "arxiv_id": "2509.18542", "doi": "10.48550/arXiv.2509.18542", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.314}
{"id": "e3b59aa4ab8718cbabd3b80ea71ac3dc5a1a72129b8e07d395f0c0c94c83fb95", "sources": ["arxiv", "semantic_scholar"], "title": "DevFD: Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces", "abstract": "The rise of realistic digital face generation and manipulation poses significant social risks. The primary challenge lies in the rapid and diverse evolution of generation techniques, which often outstrip the detection capabilities of existing models. To defend against the ever-evolving new types of forgery, we need to enable our model to quickly adapt to new domains with limited computation and data while avoiding forgetting previously learned forgery types. In this work, we posit that genuine facial samples are abundant and relatively stable in acquisition methods, while forgery faces continuously evolve with the iteration of manipulation techniques. Given the practical infeasibility of exhaustively collecting all forgery variants, we frame face forgery detection as a continual learning problem and allow the model to develop as new forgery types emerge. Specifically, we employ a Developmental Mixture of Experts (MoE) architecture that uses LoRA models as its individual experts. These experts are organized into two groups: a Real-LoRA to learn and refine knowledge of real faces, and multiple Fake-LoRAs to capture incremental information from different forgery types. To prevent catastrophic forgetting, we ensure that the learning direction of Fake-LoRAs is orthogonal to the established subspace. Moreover, we integrate orthogonal gradients into the orthogonal loss of Fake-LoRAs, preventing gradient interference throughout the training process of each task. Experimental results under both the datasets and manipulation types incremental protocols demonstrate the effectiveness of our method.", "authors": ["Tianshuo Zhang", "Li Gao", "Siran Peng", "Xiangyu Zhu", "Zhen Lei"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-23", "url": "https://arxiv.org/abs/2509.19230", "pdf_url": "https://arxiv.org/pdf/2509.19230v3", "arxiv_id": "2509.19230", "doi": "10.48550/arXiv.2509.19230", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.314}
{"id": "3539aa9fc355ef284dc7cf32817cd9be9bb5250524aad67fa587404a854dfeca", "sources": ["arxiv", "semantic_scholar"], "title": "Do Sparse Subnetworks Exhibit Cognitively Aligned Attention? Effects of Pruning on Saliency Map Fidelity, Sparsity, and Concept Coherence", "abstract": "Prior works have shown that neural networks can be heavily pruned while preserving performance, but the impact of pruning on model interpretability remains unclear. In this work, we investigate how magnitude-based pruning followed by fine-tuning affects both low-level saliency maps and high-level concept representations. Using a ResNet-18 trained on ImageNette, we compare post-hoc explanations from Vanilla Gradients (VG) and Integrated Gradients (IG) across pruning levels, evaluating sparsity and faithfulness. We further apply CRAFT-based concept extraction to track changes in semantic coherence of learned concepts. Our results show that light-to-moderate pruning improves saliency-map focus and faithfulness while retaining distinct, semantically meaningful concepts. In contrast, aggressive pruning merges heterogeneous features, reducing saliency map sparsity and concept coherence despite maintaining accuracy. These findings suggest that while pruning can shape internal representations toward more human-aligned attention patterns, excessive pruning undermines interpretability.", "authors": ["Sanish Suwal", "Dipkamal Bhusal", "Michael Clifford", "Nidhi Rastogi"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-23", "url": "https://arxiv.org/abs/2509.21387", "pdf_url": "https://arxiv.org/pdf/2509.21387v3", "arxiv_id": "2509.21387", "doi": "10.48550/arXiv.2509.21387", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.314}
{"id": "31c9fa116c405953eb562ef8c595cd27dbe3d87b965548d27d00d06400f0f1de", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba Modulation: On the Length Generalization of Mamba", "abstract": "The quadratic complexity of the attention mechanism in Transformer models has motivated the development of alternative architectures with sub-quadratic scaling, such as state-space models. Among these, Mamba has emerged as a leading architecture, achieving state-of-the-art results across a range of language modeling tasks. However, Mamba's performance significantly deteriorates when applied to contexts longer than those seen during pre-training, revealing a sharp sensitivity to context length extension. Through detailed analysis, we attribute this limitation to the out-of-distribution behaviour of its state-space dynamics, particularly within the parameterization of the state transition matrix $\\mathbf{A}$. Unlike recent works which attribute this sensitivity to the vanished accumulation of discretization time steps, $\\exp(-\\sum_{t=1}^NΔ_t)$, we establish a connection between state convergence behavior as the input length approaches infinity and the spectrum of the transition matrix $\\mathbf{A}$, offering a well-founded explanation of its role in length extension. Next, to overcome this challenge, we propose an approach that applies spectrum scaling to pre-trained Mamba models to enable robust long-context generalization by selectively modulating the spectrum of $\\mathbf{A}$ matrices in each layer. We show that this can significantly improve performance in settings where simply modulating $Δ_t$ fails, validating our insights and providing avenues for better length generalization of state-space models with structured transition matrices.", "authors": ["Peng Lu", "Jerry Huang", "Qiuhao Zeng", "Xinyu Wang", "Boxing Chen", "Philippe Langlais", "Yufei Cui"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-09-23", "url": "https://arxiv.org/abs/2509.19633", "pdf_url": "https://arxiv.org/pdf/2509.19633v3", "arxiv_id": "2509.19633", "doi": "10.48550/arXiv.2509.19633", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.314}
{"id": "6d53c8f73f8ccf5aef337727db3ca935d431a5ea8c7a4e1170e1fe781466487f", "sources": ["arxiv", "semantic_scholar"], "title": "On-the-Fly Adaptation to Quantization: Configuration-Aware LoRA for Efficient Fine-Tuning of Quantized LLMs", "abstract": "As increasingly large pre-trained models are released, deploying them on edge devices for privacy-preserving applications requires effective compression. Recent works combine quantization with the fine-tuning of high-precision LoRA adapters, which can substantially reduce model size while mitigating the accuracy loss from quantization. However, edge devices have inherently heterogeneous capabilities, while performing configuration-wise fine-tuning for every quantization setting is computationally prohibitive. In this paper, we propose CoA-LoRA, a method that dynamically adjusts the LoRA adapter to arbitrary quantization configurations (i.e., the per-layer bit-width choices of a pre-trained model) without requiring repeated fine-tuning. This is accomplished via a configuration-aware model that maps each configuration to its low-rank adjustments. The effectiveness of this model critically depends on the training configuration set, a collection of configurations chosen to cover different total bit-width budgets. However, constructing a high-quality configuration set is non-trivial. We therefore design a Pareto-based configuration search that iteratively optimizes the training configuration set, yielding more precise low-rank adjustments. Our experiments demonstrate that, unlike the state-of-the-art methods that require fine-tuning a separate LoRA adapter for each configuration, CoA-LoRA incurs no additional time cost while achieving comparable or even superior performance to those methods.", "authors": ["Rongguang Ye", "Ming Tang", "Edith C. H. Ngai"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-22", "url": "https://arxiv.org/abs/2509.25214", "pdf_url": "https://arxiv.org/pdf/2509.25214v3", "arxiv_id": "2509.25214", "doi": "10.48550/arXiv.2509.25214", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3128}
{"id": "22dd2446885448dcdd15552639cd8eb9e523cfd087ed9f0bc930500ab9b127aa", "sources": ["arxiv", "semantic_scholar"], "title": "Attention-based Mixture of Experts for Robust Speech Deepfake Detection", "abstract": "AI-generated speech is becoming increasingly used in everyday life, powering virtual assistants, accessibility tools, and other applications. However, it is also being exploited for malicious purposes such as impersonation, misinformation, and biometric spoofing. As speech deepfakes become nearly indistinguishable from real human speech, the need for robust detection methods and effective countermeasures has become critically urgent. In this paper, we present the ISPL's submission to the SAFE challenge at IH&MMSec 2025, where our system ranked first across all tasks. Our solution introduces a novel approach to audio deepfake detection based on a Mixture of Experts architecture. The proposed system leverages multiple state-of-the-art detectors, combining their outputs through an attention-based gating network that dynamically weights each expert based on the input speech signal. In this design, each expert develops a specialized understanding of the shared training data by learning to capture different complementary aspects of the same input through inductive biases. Experimental results indicate that our method outperforms existing approaches across multiple datasets. We further evaluate and analyze the performance of our system in the SAFE challenge.", "authors": ["Viola Negroni", "Davide Salvi", "Alessandro Ilic Mezza", "Paolo Bestagini", "Stefano Tubaro"], "categories": ["cs.SD"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-22", "url": "https://arxiv.org/abs/2509.17585", "pdf_url": "https://arxiv.org/pdf/2509.17585v1", "arxiv_id": "2509.17585", "doi": "10.1109/WIFS66636.2025.00017", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Workshop on Information Forensics and Security", "quality_score": 0.3128}
{"id": "b5890047b93b2e07e88d000473b1f840bd0e617a6059eee3d17418fbb07b6030", "sources": ["arxiv", "semantic_scholar"], "title": "EpiCache: Episodic KV Cache Management for Long-Term Conversation on Resource-Constrained Environments", "abstract": "Modern large language models (LLMs) extend context lengths to millions of tokens, enabling coherent, personalized responses grounded in long conversational history. However, the Key-Value (KV) cache grows linearly with the extended dialogue history, causing the model's memory footprint to quickly exceed device limits. While recent KV cache compression methods attempt to reduce memory usage, most apply cache eviction after processing the entire context, incurring unbounded peak memory usage. Additionally, query-dependent eviction narrows the cache semantics to a single query, leading to failure cases in multi-turn conversations. In this paper, we introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and performs episode-specific KV cache eviction. Across three LongConvQA benchmarks (LongMemEval, Realtalk, and LoCoMo), EpiCache improves accuracy by up to 30%, achieves near full-cache accuracy under 4-6x compression, and reduces latency and peak memory by up to 2.4x and 3.7x, respectively.", "authors": ["Minsoo Kim", "Arnav Kundu", "Han-Byul Kim", "Richa Dixit", "Minsik Cho"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-22", "url": "https://arxiv.org/abs/2509.17396", "pdf_url": "https://arxiv.org/pdf/2509.17396v4", "arxiv_id": "2509.17396", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1991}
{"id": "45ed0cbfc69fc6e307b7099068bf0ee8db104c9cfcf5d75c9b508baf470158bc", "sources": ["arxiv", "semantic_scholar"], "title": "Achilles' Heel of Mamba: Essential difficulties of the Mamba architecture demonstrated by synthetic data", "abstract": "State Space Models (SSMs) have emerged as promising alternatives to attention mechanisms, with the Mamba architecture demonstrating impressive performance and linear complexity for processing long sequences. However, the fundamental differences between Mamba and Transformer architectures remain incompletely understood. In this work, we use carefully designed synthetic tasks to reveal Mamba's inherent limitations. Through experiments, we identify that Mamba's nonlinear convolution introduces an asymmetry bias that significantly impairs its ability to recognize symmetrical patterns and relationships. Using composite function and inverse sequence matching tasks, we demonstrate that Mamba strongly favors compositional solutions over symmetrical ones and struggles with tasks requiring the matching of reversed sequences. We show these limitations stem not from the SSM module itself but from the nonlinear convolution preceding it, which fuses token information asymmetrically. These insights provide a new understanding of Mamba's constraints and suggest concrete architectural improvements for future sequence models.", "authors": ["Tianyi Chen", "Pengxiao Lin", "Zhiwei Wang", "Zhi-Qin John Xu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-22", "url": "https://arxiv.org/abs/2509.17514", "pdf_url": "https://arxiv.org/pdf/2509.17514v2", "arxiv_id": "2509.17514", "doi": "10.48550/arXiv.2509.17514", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3128}
{"id": "993ebde026886a1b93e350dbf0ce66a975df811f5f959b4acf3049494d71e7c3", "sources": ["arxiv", "semantic_scholar"], "title": "PTQTP: Post-Training Quantization to Trit-Planes for Large Language Models", "abstract": "Post-training quantization (PTQ) of large language models (LLMs) to extremely low bit-widths remains challenging due to the fundamental trade-off between computational efficiency and representational capacity. While existing ultra-low-bit methods rely on binary approximations or quantization-aware training(QAT), they often suffer from either limited representational capacity or huge training resource overhead. We introduce PTQ to Trit-Planes (PTQTP), a structured PTQ framework that decomposes weight matrices into dual ternary {-1, 0, 1} trit-planes. This approach achieves multiplication-free additive inference by decoupling weights into discrete topology (trit-planes) and continuous magnitude (scales), effectively enabling high-fidelity sparse approximation. PTQTP provides: (1) a theoretically grounded progressive approximation algorithm ensuring global weight consistency; (2) model-agnostic deployment without architectural modifications; and (3) uniform ternary operations that eliminate mixed-precision overhead. Comprehensive experiments on LLaMA3.x and Qwen3 (0.6B-70B) demonstrate that PTQTP significantly outperforms sub-4bit PTQ methods on both language reasoning tasks and mathematical reasoning as well as coding. PTQTP rivals the 1.58-bit QAT performance while requiring only single-hour quantization compared to 10-14 GPU days for training-based methods, and the end-to-end inference speed achieves 4.63$\\times$ faster than the FP16 baseline model, establishing a new and practical solution for efficient LLM deployment in resource-constrained environments. Code will available at https://github.com/HeXiao-55/PTQTP.", "authors": ["He Xiao", "Runming Yang", "Qingyao Yang", "Wendong Xu", "Zhen Li", "Yupeng Su", "Zhengwu Liu", "Hongxia Yang", "Ngai Wong"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-21", "url": "https://arxiv.org/abs/2509.16989", "pdf_url": "https://arxiv.org/pdf/2509.16989v3", "arxiv_id": "2509.16989", "doi": "10.48550/arXiv.2509.16989", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/HeXiao-55/PTQTP", "venue": "arXiv.org", "quality_score": 0.4817}
{"id": "c48b05bcfb0d885d3754c43cd2e9d5211d820e84923d00802bdf200203b0f7ab", "sources": ["arxiv", "semantic_scholar"], "title": "ShadowServe: Interference-Free KV Cache Fetching for Distributed Prefix Caching", "abstract": "Distributed prefix caching accelerates long-context LLM serving by reusing KV cache entries for common context prefixes. However, KV cache fetches can become a bottleneck when network bandwidth is limited. Compression mitigates the bandwidth issue, but can degrade overall performance when decompression interferes with model computation. We present ShadowServe, the first SmartNIC-accelerated, interference-free prefix caching system for LLM serving. ShadowServe separates a control plane on the host and a data plane fully offloaded to the SmartNIC, which eliminates interference to both host GPU and CPU. To overcome the SmartNIC's limited compute and memory resources, we design a chunked pipeline that parallelizes data plane operations across the SmartNIC's compute resources, and a minimal-copy memory management scheme that reduces memory pressure on the SmartNIC. Compared to state-of-the-art solutions, ShadowServe achieves up to 2.2x lower loaded time-per-output-token (TPOT), and reduces time-to-first-token (TTFT) by up to 1.38x in low-bandwidth scenarios (<= 20 Gbps), translating to up to 1.35x higher throughput.", "authors": ["Xingyu Xiang", "Raj Joshi", "Yuhan Liu", "Jiayi Yao", "Chenxingyu Zhao", "Junchen Jiang", "Yang Zhou", "Eddie Kohler", "Minlan Yu"], "categories": ["cs.DC", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-21", "url": "https://arxiv.org/abs/2509.16857", "pdf_url": "https://arxiv.org/pdf/2509.16857v1", "arxiv_id": "2509.16857", "doi": "10.48550/arXiv.2509.16857", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3117}
{"id": "91c5ca0225e34b643fbb70e37e764d85bfffac1941d08bac44dae0c593b5d191", "sources": ["arxiv", "semantic_scholar"], "title": "KV-Efficient VLA: A Method to Speed up Vision Language Models with RNN-Gated Chunked KV Cache", "abstract": "Vision-Language-Action (VLA) models offer a unified framework for robotic perception and control, but their ability to scale to real-world, long-horizon tasks is limited by the high computational cost of attention and the large memory required for storing key-value (KV) pairs during inference, particularly when retaining historical image tokens as context. Recent methods have focused on scaling backbone architectures to improve generalization, with less emphasis on addressing inference inefficiencies essential for real-time use. In this work, we present KV-Efficient VLA, a model-agnostic memory compression approach designed to address these limitations by introducing a lightweight mechanism to selectively retain high-utility context. Our method partitions the KV cache into fixed-size chunks and employs a recurrent gating module to summarize and filter the historical context according to learned utility scores. This design aims to preserve recent fine-grained detail while aggressively pruning stale, low-relevance memory. Based on experiments, our approach can yield an average of 24.6% FLOPs savings, 1.34x inference speedup, and 1.87x reduction in KV memory. Our method integrates seamlessly into recent VLA stacks, enabling scalable inference without modifying downstream control logic.", "authors": ["Wanshun Xu", "Long Zhuang", "Lianlei Shan"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-20", "url": "https://arxiv.org/abs/2509.21354", "pdf_url": "https://arxiv.org/pdf/2509.21354v2", "arxiv_id": "2509.21354", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1976}
{"id": "afc4e42ba3ead7ae090ee7ce0d42d6c5e0e095538be00daeb3ffc1e0af507efd", "sources": ["arxiv", "semantic_scholar"], "title": "MoPE: A Mixture of Password Experts for Improving Password Guessing", "abstract": "Textual passwords remain a predominant authentication mechanism in web security. To evaluate their strength, existing research has proposed several data-driven models across various scenarios. However, these models generally treat passwords uniformly, neglecting the structural differences among passwords. This typically results in biased training that favors frequent password structural patterns. To mitigate the biased training, we argue that passwords, as a type of complex short textual data, should be processed in a structure-aware manner by identifying their structural patterns and routing them to specialized models accordingly. In this paper, we propose MoPE, a Mixture of Password Experts framework, specifically designed to leverage the structural patterns in passwords to improveguessing performance. Motivated by the observation that passwords with similar structural patterns (e.g., fixed-length numeric strings) tend to cluster in high-density regions within the latent space, our MoPE introduces: (1) a novel structure-based method for generating specialized expert models; (2) a lightweight gate method to select appropriate expert models to output reliable guesses, better aligned with the high computational frequency of password guessing tasks. Our evaluation shows that MoPE significantly outperforms existing state-of-the-art baselines in both offline and online guessing scenarios, achieving up to 38.80% and 9.27% improvement in cracking rate, respectively, showcasing that MoPE can effectively exploit the capabilities of data-driven models for password guessing. Additionally, we implement a real-time Password Strength Meter (PSM) based on offline MoPE, assisting users in choosing stronger passwords more precisely with millisecond-level response latency.", "authors": ["Mingjian Duan", "Ming Xu", "Shenghao Zhang", "Jiaheng Zhang", "Weili Han"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-20", "url": "https://arxiv.org/abs/2509.16558", "pdf_url": "https://arxiv.org/pdf/2509.16558v1", "arxiv_id": "2509.16558", "doi": "10.48550/arXiv.2509.16558", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3105}
{"id": "edd15fa2815010bdd94caf6bb8e4223e6800b0efd600c30ac835a0f0e4e97188", "sources": ["arxiv", "semantic_scholar"], "title": "DiEP: Adaptive Mixture-of-Experts Compression through Differentiable Expert Pruning", "abstract": "Despite the significant breakthrough of Mixture-of-Experts (MoE), the increasing scale of these MoE models presents huge memory and storage challenges. Existing MoE pruning methods, which involve reducing parameter size with a uniform sparsity across all layers, often lead to suboptimal outcomes and performance degradation due to varying expert redundancy in different MoE layers. To address this, we propose a non-uniform pruning strategy, dubbed \\textbf{Di}fferentiable \\textbf{E}xpert \\textbf{P}runing (\\textbf{DiEP}), which adaptively adjusts pruning rates at the layer level while jointly learning inter-layer importance, effectively capturing the varying redundancy across different MoE layers. By transforming the global discrete search space into a continuous one, our method handles exponentially growing non-uniform expert combinations, enabling adaptive gradient-based pruning. Extensive experiments on five advanced MoE models demonstrate the efficacy of our method across various NLP tasks. Notably, \\textbf{DiEP} retains around 92\\% of original performance on Mixtral 8$\\times$7B with only half the experts, outperforming other pruning methods by up to 7.1\\% on the challenging MMLU dataset.", "authors": ["Sikai Bai", "Haoxi Li", "Jie Zhang", "Zicong Hong", "Song Guo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-19", "url": "https://arxiv.org/abs/2509.16105", "pdf_url": "https://arxiv.org/pdf/2509.16105v1", "arxiv_id": "2509.16105", "doi": "10.48550/arXiv.2509.16105", "citation_count": 10, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3094}
{"id": "aa66404c7953d86b31a00188d01c5434d1d1433cfcc7e3a275c9bb8131e934cd", "sources": ["arxiv", "semantic_scholar"], "title": "TrueMoE: Dual-Routing Mixture of Discriminative Experts for Synthetic Image Detection", "abstract": "The rapid progress of generative models has made synthetic image detection an increasingly critical task. Most existing approaches attempt to construct a single, universal discriminative space to separate real from fake content. However, such unified spaces tend to be complex and brittle, often struggling to generalize to unseen generative patterns. In this work, we propose TrueMoE, a novel dual-routing Mixture-of-Discriminative-Experts framework that reformulates the detection task as a collaborative inference across multiple specialized and lightweight discriminative subspaces. At the core of TrueMoE is a Discriminative Expert Array (DEA) organized along complementary axes of manifold structure and perceptual granularity, enabling diverse forgery cues to be captured across subspaces. A dual-routing mechanism, comprising a granularity-aware sparse router and a manifold-aware dense router, adaptively assigns input images to the most relevant experts. Extensive experiments across a wide spectrum of generative models demonstrate that TrueMoE achieves superior generalization and robustness.", "authors": ["Laixin Zhang", "Shuaibo Li", "Wei Ma", "Hongbin Zha"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-19", "url": "https://arxiv.org/abs/2509.15741", "pdf_url": "https://arxiv.org/pdf/2509.15741v1", "arxiv_id": "2509.15741", "doi": "10.48550/arXiv.2509.15741", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3094}
{"id": "8d8e6775462662175f407e3794ec3d86eea40a89cc217455e696cc9b757b0be5", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive LoRA Experts Allocation and Selection for Federated Fine-Tuning", "abstract": "Large Language Models (LLMs) have demonstrated impressive capabilities across various tasks, but fine-tuning them for domain-specific applications often requires substantial domain-specific data that may be distributed across multiple organizations. Federated Learning (FL) offers a privacy-preserving solution, but faces challenges with computational constraints when applied to LLMs. Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient fine-tuning approach, though a single LoRA module often struggles with heterogeneous data across diverse domains. This paper addresses two critical challenges in federated LoRA fine-tuning: 1. determining the optimal number and allocation of LoRA experts across heterogeneous clients, and 2. enabling clients to selectively utilize these experts based on their specific data characteristics. We propose FedLEASE (Federated adaptive LoRA Expert Allocation and SElection), a novel framework that adaptively clusters clients based on representation similarity to allocate and train domain-specific LoRA experts. It also introduces an adaptive top-$M$ Mixture-of-Experts mechanism that allows each client to select the optimal number of utilized experts. Our extensive experiments on diverse benchmark datasets demonstrate that FedLEASE significantly outperforms existing federated fine-tuning approaches in heterogeneous client settings while maintaining communication efficiency.", "authors": ["Lei Wang", "Jieming Bian", "Letian Zhang", "Jie Xu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-18", "url": "https://arxiv.org/abs/2509.15087", "pdf_url": "https://arxiv.org/pdf/2509.15087v2", "arxiv_id": "2509.15087", "doi": "10.48550/arXiv.2509.15087", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3082}
{"id": "205a89c53a3901ce042329e1c941988461bfc3c47dd768d38f346b23686acfaf", "sources": ["arxiv", "semantic_scholar"], "title": "Super-Linear: A Lightweight Pretrained Mixture of Linear Experts for Time Series Forecasting", "abstract": "Time series forecasting (TSF) is critical in domains like energy, finance, healthcare, and logistics, requiring models that generalize across diverse datasets. Large pre-trained models such as Chronos and Time-MoE show strong zero-shot (ZS) performance but suffer from high computational costs. In this work, we introduce Super-Linear, a lightweight and scalable mixture-of-experts (MoE) model for general forecasting. It replaces deep architectures with simple frequency-specialized linear experts, trained on resampled data across multiple frequency regimes. A lightweight spectral gating mechanism dynamically selects relevant experts, enabling efficient, accurate forecasting. Despite its simplicity, Super-Linear demonstrates strong performance across benchmarks, while substantially improving efficiency, robustness to sampling rates, and interpretability. The implementation of Super-Linear is available at: \\href{https://github.com/azencot-group/SuperLinear}{https://github.com/azencot-group/SuperLinear}.", "authors": ["Liran Nochumsohn", "Raz Marshanski", "Hedi Zisling", "Omri Azencot"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-18", "url": "https://arxiv.org/abs/2509.15105", "pdf_url": "https://arxiv.org/pdf/2509.15105v3", "arxiv_id": "2509.15105", "doi": "10.48550/arXiv.2509.15105", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/azencot-group/SuperLinear}{https://github.com/azencot-group/SuperLinear}", "venue": "arXiv.org", "quality_score": 0.4764}
{"id": "5cb949dc7352f8c65a4d8272e8219f631b32bf4a0dd4f37ae48759610dc79cad", "sources": ["arxiv", "semantic_scholar"], "title": "FURINA: Free from Unmergeable Router via LINear Aggregation of mixed experts", "abstract": "The Mixture of Experts (MoE) paradigm has been successfully integrated into Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning (PEFT), delivering performance gains with minimal parameter overhead. However, a key limitation of existing MoE-LoRA methods is their reliance on a discrete router, which prevents the integration of the MoE components into the backbone model. To overcome this, we propose FURINA, a novel Free from Unmergeable Router framework based on the LINear Aggregation of experts. FURINA eliminates the router by introducing a Self-Routing mechanism. This is achieved through three core innovations: (1) decoupled learning of the direction and magnitude for LoRA adapters, (2) a shared learnable magnitude vector for consistent activation scaling, and (3) expert selection loss that encourages divergent expert activation. The proposed mechanism leverages the angular similarity between the input and each adapter's directional component to activate experts, which are then scaled by the shared magnitude vector. This design allows the output norm to naturally reflect the importance of each expert, thereby enabling dynamic, router-free routing. The expert selection loss further sharpens this behavior by encouraging sparsity and aligning it with standard MoE activation patterns. We also introduce a shared expert within the MoE-LoRA block that provides stable, foundational knowledge. To the best of our knowledge, FURINA is the first router-free, MoE-enhanced LoRA method that can be fully merged into the backbone model, introducing zero additional inference-time cost or complexity. Extensive experiments demonstrate that FURINA not only significantly outperforms standard LoRA but also matches or surpasses the performance of existing MoE-LoRA methods, while eliminating the extra inference-time overhead of MoE.", "authors": ["Jiayi Han", "Liang Du", "Yinda Chen", "Xiao Kang", "Weiyang Ding", "Donghong Han"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-18", "url": "https://arxiv.org/abs/2509.14900", "pdf_url": "https://arxiv.org/pdf/2509.14900v2", "arxiv_id": "2509.14900", "doi": "10.48550/arXiv.2509.14900", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3082}
{"id": "f03d3abd63dea9997c1fd6203dd2b900383ef3dfcde06bd11ef0361bba3c213a", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Low-Rank Adapter Experts in Generalizable Audio Deepfake Detection", "abstract": "Foundation models such as Wav2Vec2 excel at representation learning in speech tasks, including audio deepfake detection. However, after being fine-tuned on a fixed set of bonafide and spoofed audio clips, they often fail to generalize to novel deepfake methods not represented in training. To address this, we propose a mixture-of-LoRA-experts approach that integrates multiple low-rank adapters (LoRA) into the model's attention layers. A routing mechanism selectively activates specialized experts, enhancing adaptability to evolving deepfake attacks. Experimental results show that our method outperforms standard fine-tuning in both in-domain and out-of-domain scenarios, reducing equal error rates relative to baseline models. Notably, our best MoE-LoRA model lowers the average out-of-domain EER from 8.55\\% to 6.08\\%, demonstrating its effectiveness in achieving generalizable audio deepfake detection.", "authors": ["Janne Laakkonen", "Ivan Kukanov", "Ville Hautamäki"], "categories": ["eess.AS", "cs.LG", "cs.SD"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2025-09-17", "url": "https://arxiv.org/abs/2509.13878", "pdf_url": "https://arxiv.org/pdf/2509.13878v1", "arxiv_id": "2509.13878", "doi": "10.1109/APSIPAASC65261.2025.11249190", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Asia-Pacific Signal and Information Processing Association Annual Summit and Conference", "quality_score": 0.3071}
{"id": "f963620882b6e849a3452e70cc0fd3e4167dfd95cafc1948c9453974ec48e394", "sources": ["arxiv", "semantic_scholar"], "title": "SparseDoctor: Towards Efficient Chat Doctor with Mixture of Experts Enhanced Large Language Models", "abstract": "Large language models (LLMs) have achieved great success in medical question answering and clinical decision-making, promoting the efficiency and popularization of the personalized virtual doctor in society. However, the traditional fine-tuning strategies on LLM require the updates of billions of parameters, substantially increasing the training cost, including the training time and utility cost. To enhance the efficiency and effectiveness of the current medical LLMs and explore the boundary of the representation capability of the LLMs on the medical domain, apart from the traditional fine-tuning strategies from the data perspective (i.e., supervised fine-tuning or reinforcement learning from human feedback), we instead craft a novel sparse medical LLM named SparseDoctor armed with contrastive learning enhanced LoRA-MoE (low rank adaptation-mixture of experts) architecture. To this end, the crafted automatic routing mechanism can scientifically allocate the computational resources among different LoRA experts supervised by the contrastive learning. Additionally, we also introduce a novel expert memory queue mechanism to further boost the efficiency of the overall framework and prevent the memory overflow during training. We conduct comprehensive evaluations on three typical medical benchmarks: CMB, CMExam, and CMMLU-Med. Experimental results demonstrate that the proposed LLM can consistently outperform the strong baselines such as the HuatuoGPT series.", "authors": ["Jianbin Zhang", "Yulin Zhu", "Wai Lun Lo", "Richard Tai-Chiu Hsung", "Harris Sik-Ho Tsang", "Kai Zhou"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-15", "url": "https://arxiv.org/abs/2509.14269", "pdf_url": "https://arxiv.org/pdf/2509.14269v2", "arxiv_id": "2509.14269", "doi": "10.48550/arXiv.2509.14269", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3048}
{"id": "f6dec2ecae50adeaaa96fc5b2ac304ddf357cde9ff152b7b05ee0ab2e76330f7", "sources": ["arxiv", "semantic_scholar"], "title": "LoRALib: A Standardized Benchmark for Evaluating LoRA-MoE Methods", "abstract": "As a parameter efficient fine-tuning (PEFT) method, low-rank adaptation (LoRA) can save significant costs in storage and computing, but its strong adaptability to a single task is often accompanied by insufficient cross-task generalization capabilities. To improve this, existing work combines LoRA with mixture-of-experts (MoE) to enhance the model's adaptability through expert modules and routing mechanisms. However, existing LoRA-MoE methods lack unified standards in models, datasets, hyperparameters, and evaluation methods, making it difficult to conduct fair comparisons between different methods. To this end, we proposed a unified benchmark named LoRALib. Specifically, we standardized datasets from $40$ downstream tasks into a unified format, fine-tuned them using the same hyperparameters and obtained $680$ LoRA modules across $17$ model architectures. Based on this LoRA library, we conduct large-scale experiments on $3$ representative LoRA-MoE methods and different LoRA selection mechanisms using the open-sourced testing tool OpenCompass. Extensive experiments show that LoRAMoE performs best, and that prioritizing LoRAs relevant to the target task can further improve the performance of MoE. We hope these findings will inspire future work. Our datasets and LoRA library are available at https://huggingface.co/datasets/YaoLuzjut/LoRAOcean_dataset and https://huggingface.co/YaoLuzjut/models.", "authors": ["Shaoheng Wang", "Yao Lu", "Yuqi Li", "Yaxin Gao", "Jiaqi Nie", "Shanqing Yu", "Yingli Tian", "Qi Xuan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-14", "url": "https://arxiv.org/abs/2509.18137", "pdf_url": "https://arxiv.org/pdf/2509.18137v1", "arxiv_id": "2509.18137", "doi": "10.48550/arXiv.2509.18137", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4693}
{"id": "c056e1882c24dfc7c752eee9a543daf11759d21935edfad17be1ade74b1080d7", "sources": ["arxiv", "semantic_scholar"], "title": "Knowledge-Guided Adaptive Mixture of Experts for Precipitation Prediction", "abstract": "Accurate precipitation forecasting is indispensable in agriculture, disaster management, and sustainable strategies. However, predicting rainfall has been challenging due to the complexity of climate systems and the heterogeneous nature of multi-source observational data, including radar, satellite imagery, and surface-level measurements. The multi-source data vary in spatial and temporal resolution, and they carry domain-specific features, making it challenging for effective integration in conventional deep learning models. Previous research has explored various machine learning techniques for weather prediction; however, most struggle with the integration of data with heterogeneous modalities. To address these limitations, we propose an Adaptive Mixture of Experts (MoE) model tailored for precipitation rate prediction. Each expert within the model specializes in a specific modality or spatio-temporal pattern. We also incorporated a dynamic router that learns to assign inputs to the most relevant experts. Our results show that this modular design enhances predictive accuracy and interpretability. In addition to the modeling framework, we introduced an interactive web-based visualization tool that enables users to intuitively explore historical weather patterns over time and space. The tool was designed to support decision-making for stakeholders in climate-sensitive sectors. We evaluated our approach using a curated multimodal climate dataset capturing real-world conditions during Hurricane Ian in 2022. The benchmark results show that the Adaptive MoE significantly outperformed all the baselines.", "authors": ["Chen Jiang", "Kofi Osei", "Sai Deepthi Yeddula", "Dongji Feng", "Wei-Shinn Ku"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-14", "url": "https://arxiv.org/abs/2509.11459", "pdf_url": "https://arxiv.org/pdf/2509.11459v1", "arxiv_id": "2509.11459", "doi": "10.48550/arXiv.2509.11459", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3036}
{"id": "3bb69ede300a762a177514c76e07bcaa00ebf7394450247e812371e06254fcff", "sources": ["arxiv", "semantic_scholar"], "title": "Judge Q: Trainable Queries for Optimized Information Retention in KV Cache Eviction", "abstract": "Large language models (LLMs) utilize key-value (KV) cache to store historical information during sequence processing. The size of KV cache grows linearly as the length of the sequence extends, which seriously affects memory usage and decoding efficiency. Current methods for KV cache eviction typically utilize the last window from the pre-filling phase as queries to compute the KV importance scores for eviction. Although this scheme is simple to implement, it tends to overly focus on local information, potentially leading to the neglect or omission of crucial global information. To mitigate this issue, we propose Judge Q, a novel training method which incorporates a soft token list. This method only tunes the model's embedding layer at a low training cost. By concatenating the soft token list at the end of the input sequence, we train these tokens' attention map to the original input sequence to align with that of the actual decoded tokens. In this way, the queries corresponding to the soft tokens can effectively capture global information and better evaluate the importance of the keys and values within the KV cache, thus maintaining decoding quality when KV cache is evicted. Under the same eviction budget, our method exhibits less performance degradation compared to existing eviction approaches. We validate our approach through experiments conducted on models such as Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3, using benchmarks including LongBench, RULER, and Needle-in-a-Haystack. Results indicate an improvement of approximately 1 point on the LongBench and over 3 points on RULER. This proposed methodology can be seamlessly integrated into existing open-source models with minimal training overhead, thereby enhancing performance in KV cache eviction scenarios.", "authors": ["Yijun Liu", "Yixuan Wang", "Yuzhuang Xu", "Shiyu Ji", "Yang Xu", "Qingfu Zhu", "Wanxiang Che"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-13", "url": "https://arxiv.org/abs/2509.10798", "pdf_url": "https://arxiv.org/pdf/2509.10798v2", "arxiv_id": "2509.10798", "doi": "10.48550/arXiv.2509.10798", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4675}
{"id": "2b29dd056f7ecbd8ccd59426a901abfabb5ba9d056c9c5c1f7daed1d3894cc90", "sources": ["arxiv", "semantic_scholar"], "title": "Dropping Experts, Recombining Neurons: Retraining-Free Pruning for Sparse Mixture-of-Experts LLMs", "abstract": "Sparse Mixture-of-Experts (SMoE) architectures are widely used in large language models (LLMs) due to their computational efficiency. However, though only a few experts are activated for each token, SMoE still requires loading all expert parameters, leading to high memory usage and challenges in deployment. Previous work has tried to reduce the overhead by pruning and merging experts, but primarily focused on expert-level operations, leaving neuron-level structure underexplored. We propose DERN (Dropping Experts, Recombining Neurons), a task-agnostic and retraining-free framework for expert pruning and reconstruction. We observe that experts are often misaligned and contain semantic conflicts at the neuron level, which poses challenges for direct merging. To solve this, DERN works in three steps: it first prunes redundant experts using router statistics; then it decomposes them into neuron-level expert segments, assigning each segment to its most compatible retained expert; and finally, it merges segments within each retained expert to build a compact representation. Experiments on Mixtral, Qwen, and DeepSeek SMoE models show that DERN improves performance by more than 5% on commonsense reasoning and MMLU benchmarks under 50% expert sparsity, without extra training. It also greatly reduces the number of experts and memory usage, making SMoE LLMs easier to deploy in practice.", "authors": ["Yixiao Zhou", "Ziyu Zhao", "Dongzhou Cheng", "zhiliang wu", "Jie Gui", "Yi Yang", "Fei Wu", "Yu Cheng", "Hehe Fan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-12", "url": "https://arxiv.org/abs/2509.10377", "pdf_url": "https://arxiv.org/pdf/2509.10377v1", "arxiv_id": "2509.10377", "doi": "10.48550/arXiv.2509.10377", "citation_count": 24, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3495}
{"id": "b50471c90aea8c3e103f00ae4afdb2e692c6d590239d269431ea6dbad65d5a1e", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring Expert Specialization through Unsupervised Training in Sparse Mixture of Experts", "abstract": "Understanding the internal organization of neural networks remains a fundamental challenge in deep learning interpretability. We address this challenge by exploring a novel Sparse Mixture of Experts Variational Autoencoder (SMoE-VAE) architecture. We test our model on the QuickDraw dataset, comparing unsupervised expert routing against a supervised baseline guided by ground-truth labels. Surprisingly, we find that unsupervised routing consistently achieves superior reconstruction performance. The experts learn to identify meaningful sub-categorical structures that often transcend human-defined class boundaries. Through t-SNE visualizations and reconstruction analysis, we investigate how MoE models uncover fundamental data structures that are more aligned with the model's objective than predefined labels. Furthermore, our study on the impact of dataset size provides insights into the trade-offs between data quantity and expert specialization, offering guidance for designing efficient MoE architectures.", "authors": ["Strahinja Nikolic", "Ilker Oguz", "Demetri Psaltis"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-12", "url": "https://arxiv.org/abs/2509.10025", "pdf_url": "https://arxiv.org/pdf/2509.10025v1", "arxiv_id": "2509.10025", "doi": "10.48550/arXiv.2509.10025", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3014}
{"id": "84b7c5e9888c2f6d634846e62f6290c25dae0b86563957484ac56fc42ac628cf", "sources": ["arxiv", "semantic_scholar"], "title": "MoLEx: Mixture of LoRA Experts in Speech Self-Supervised Models for Audio Deepfake Detection", "abstract": "While self-supervised learning (SSL)-based models have boosted audio deepfake detection accuracy, fully finetuning them is computationally expensive. To address this, we propose a parameter-efficient framework that combines Low-Rank Adaptation with a Mixture-of-Experts router, called Mixture of LoRA Experts (MoLEx). It preserves pre-trained knowledge of SSL models while efficiently finetuning only selected experts, reducing training costs while maintaining robust performance. The observed utility of experts during inference shows the router reactivates the same experts for similar attacks but switches to other experts for novel spoofs, confirming MoLEx's domain-aware adaptability. MoLEx additionally offers flexibility for domain adaptation by allowing extra experts to be trained without modifying the entire model. We mainly evaluate our approach on the ASVSpoof 5 dataset and achieve the state-of-the-art (SOTA) equal error rate (EER) of 5.56% on the evaluation set without augmentation.", "authors": ["Zihan Pan", "Sailor Hardik Bhupendra", "Jinyang Wu"], "categories": ["cs.SD", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-11", "url": "https://arxiv.org/abs/2509.09175", "pdf_url": "https://arxiv.org/pdf/2509.09175v1", "arxiv_id": "2509.09175", "doi": "10.1109/ASRU65441.2025.11434670", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Automatic Speech Recognition & Understanding", "quality_score": 0.3002}
{"id": "9ef35730f34a1a429cafba2403bccb570ad18cd6b226ca49bfce3b4668b29cce", "sources": ["arxiv", "semantic_scholar"], "title": "SQAP-VLA: A Synergistic Quantization-Aware Pruning Framework for High-Performance Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models exhibit unprecedented capabilities for embodied intelligence. However, their extensive computational and memory costs hinder their practical deployment. Existing VLA compression and acceleration approaches conduct quantization or token pruning in an ad-hoc manner but fail to enable both for a holistic efficiency improvement due to an observed incompatibility. This work introduces SQAP-VLA, the first structured, training-free VLA inference acceleration framework that simultaneously enables state-of-the-art quantization and token pruning. We overcome the incompatibility by co-designing the quantization and token pruning pipeline, where we propose new quantization-aware token pruning criteria that work on an aggressively quantized model while improving the quantizer design to enhance pruning effectiveness. When applied to standard VLA models, SQAP-VLA yields significant gains in computational efficiency and inference speed while successfully preserving core model performance, achieving a $\\times$1.93 speedup and up to a 4.5\\% average success rate enhancement compared to the original model.", "authors": ["Hengyu Fang", "Yijiang Liu", "Yuan Du", "Li Du", "Huanrui Yang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-11", "url": "https://arxiv.org/abs/2509.09090", "pdf_url": "https://arxiv.org/pdf/2509.09090v1", "arxiv_id": "2509.09090", "doi": "10.48550/arXiv.2509.09090", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3002}
{"id": "2209297f4378d3defb9d164d8e709a3f7b7c14f5eaa396ad7f2b9753c74cef5f", "sources": ["arxiv", "semantic_scholar"], "title": "LAVa: Layer-wise KV Cache Eviction with Dynamic Budget Allocation", "abstract": "KV Cache is commonly used to accelerate LLM inference with long contexts, yet its high memory demand drives the need for cache compression. Existing compression methods, however, are largely heuristic and lack dynamic budget allocation. To address this limitation, we introduce a unified framework for cache compression by minimizing information loss in Transformer residual streams. Building on it, we analyze the layer attention output loss and derive a new metric to compare cache entries across heads, enabling layer-wise compression with dynamic head budgets. Additionally, by contrasting cross-layer information, we also achieve dynamic layer budgets. LAVa is the first unified strategy for cache eviction and dynamic budget allocation that, unlike prior methods, does not rely on training or the combination of multiple strategies. Experiments with benchmarks (LongBench, Needle-In-A-Haystack, Ruler, and InfiniteBench) demonstrate its superiority. Moreover, our experiments reveal a new insight: dynamic layer budgets are crucial for generation tasks (e.g., code completion), while dynamic head budgets play a key role in extraction tasks (e.g., extractive QA). As a fully dynamic compression method, LAVa consistently maintains top performance across task types. Our code is available at https://github.com/MGDDestiny/Lava.", "authors": ["Yiqun Shen", "Song Yuan", "Zhengze Zhang", "Xiaoliang Wang", "Daxin Jiang", "Nguyen Cam-Tu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-11", "url": "https://arxiv.org/abs/2509.09754", "pdf_url": "https://arxiv.org/pdf/2509.09754v1", "arxiv_id": "2509.09754", "doi": "10.48550/arXiv.2509.09754", "citation_count": 10, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/MGDDestiny/Lava", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.464}
{"id": "501f7942bc67b498407f58228fd2b1673f3221d5778de0f10e496fc0df7464b0", "sources": ["arxiv", "semantic_scholar"], "title": "ButterflyQuant: Ultra-low-bit LLM Quantization through Learnable Orthogonal Butterfly Transforms", "abstract": "Large language models require massive memory footprints, severely limiting deployment on consumer hardware. Quantization reduces memory through lower numerical precision, but extreme 2-bit quantization suffers from catastrophic performance loss due to outliers in activations. Rotation-based methods such as QuIP and QuaRot apply orthogonal transforms to eliminate outliers before quantization, using computational invariance: $\\mathbf{y} = \\mathbf{Wx} = (\\mathbf{WQ}^T)(\\mathbf{Qx})$ for orthogonal $\\mathbf{Q}$. However, these methods use fixed transforms--Hadamard matrices achieving optimal worst-case coherence $μ= 1/\\sqrt{n}$--that cannot adapt to specific weight distributions. We identify that different transformer layers exhibit distinct outlier patterns, motivating layer-adaptive rotations rather than one-size-fits-all approaches. In this work, we propose ButterflyQuant, which replaces Hadamard rotations with learnable butterfly transforms parameterized by continuous Givens rotation angles. Unlike Hadamard's discrete $\\{+1, -1\\}$ entries that are non-differentiable and thus prohibit gradient-based learning, butterfly transforms' continuous parameterization enables smooth optimization while guaranteeing orthogonality by construction. This orthogonal constraint ensures theoretical guarantees in outlier suppression while achieving $O(n \\log n)$ computational complexity with only $\\frac{n \\log n}{2}$ learnable parameters. We further introduce a uniformity regularization on post-transformation activations to promote smoother distributions amenable to quantization. Learning requires only 128 calibration samples and converges in minutes on a single GPU.", "authors": ["Bingxin Xu", "Zhen Dong", "Oussama Elachqar", "Yuzhang Shang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-11", "url": "https://arxiv.org/abs/2509.09679", "pdf_url": "https://arxiv.org/pdf/2509.09679v3", "arxiv_id": "2509.09679", "doi": "10.48550/arXiv.2509.09679", "citation_count": 7, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3002}
{"id": "3e1acd3047a02e1bbc0c60c332725eb78478aac83c712897c309e6ae9863b2ce", "sources": ["arxiv", "semantic_scholar"], "title": "MoSE: Unveiling Structural Patterns in Graphs via Mixture of Subgraph Experts", "abstract": "While graph neural networks (GNNs) have achieved great success in learning from graph-structured data, their reliance on local, pairwise message passing restricts their ability to capture complex, high-order subgraph patterns. leading to insufficient structural expressiveness. Recent efforts have attempted to enhance structural expressiveness by integrating random walk kernels into GNNs. However, these methods are inherently designed for graph-level tasks, which limits their applicability to other downstream tasks such as node classification. Moreover, their fixed kernel configurations hinder the model's flexibility in capturing diverse subgraph structures. To address these limitations, this paper proposes a novel Mixture of Subgraph Experts (MoSE) framework for flexible and expressive subgraph-based representation learning across diverse graph tasks. Specifically, MoSE extracts informative subgraphs via anonymous walks and dynamically routes them to specialized experts based on structural semantics, enabling the model to capture diverse subgraph patterns with improved flexibility and interpretability. We further provide a theoretical analysis of MoSE's expressivity within the Subgraph Weisfeiler-Lehman (SWL) Test, proving that it is more powerful than SWL. Extensive experiments, together with visualizations of learned subgraph experts, demonstrate that MoSE not only outperforms competitive baselines but also provides interpretable insights into structural patterns learned by the model.", "authors": ["Junda Ye", "Zhongbao Zhang", "Li Sun", "Siqiang Luo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-11", "url": "https://arxiv.org/abs/2509.09337", "pdf_url": "https://arxiv.org/pdf/2509.09337v1", "arxiv_id": "2509.09337", "doi": "10.48550/arXiv.2509.09337", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3002}
{"id": "3d33c3ade4ef771d62f8305eb0008f7859bcfa3d5b3502021c7f2569fde17334", "sources": ["arxiv", "semantic_scholar"], "title": "EvolKV: Evolutionary KV Cache Compression for LLM Inference", "abstract": "Existing key-value (KV) cache compression methods typically rely on heuristics, such as uniform cache allocation across layers or static eviction policies, however, they ignore the critical interplays among layer-specific feature patterns and task performance, which can lead to degraded generalization. In this paper, we propose EvolKV, an adaptive framework for layer-wise, task-driven KV cache compression that jointly optimizes the memory efficiency and task performance. By reformulating cache allocation as a multi-objective optimization problem, EvolKV leverages evolutionary search to dynamically configure layer budgets while directly maximizing downstream performance. Extensive experiments on 11 tasks demonstrate that our approach outperforms all baseline methods across a wide range of KV cache budgets on long-context tasks and surpasses heuristic baselines by up to 7 percentage points on GSM8K. Notably, EvolKV achieves superior performance over the full KV cache setting on code completion while utilizing only 1.5% of the original budget, suggesting the untapped potential in learned compression strategies for KV cache budget allocation.", "authors": ["Bohan Yu", "Yekun Chai"], "categories": ["cs.LG", "cs.CL", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-10", "url": "https://arxiv.org/abs/2509.08315", "pdf_url": "https://arxiv.org/pdf/2509.08315v1", "arxiv_id": "2509.08315", "doi": "10.48550/arXiv.2509.08315", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2991}
{"id": "0fb3848b56b1ee155e57a6066b6b69d3791b24fd727702b5c0177f4504f5e062", "sources": ["arxiv", "semantic_scholar"], "title": "MoWE : A Mixture of Weather Experts", "abstract": "Data-driven weather models have recently achieved state-of-the-art performance, yet progress has plateaued in recent years. This paper introduces a Mixture of Experts (MoWE) approach as a novel paradigm to overcome these limitations, not by creating a new forecaster, but by optimally combining the outputs of existing models. The MoWE model is trained with significantly lower computational resources than the individual experts. Our model employs a Vision Transformer-based gating network that dynamically learns to weight the contributions of multiple \"expert\" models at each grid point, conditioned on forecast lead time. This approach creates a synthesized deterministic forecast that is more accurate than any individual component in terms of Root Mean Squared Error (RMSE). Our results demonstrate the effectiveness of this method, achieving up to a 10% lower RMSE than the best-performing AI weather model on a 2-day forecast horizon, significantly outperforming individual experts as well as a simple average across experts. This work presents a computationally efficient and scalable strategy to push the state of the art in data-driven weather prediction by making the most out of leading high-quality forecast models.", "authors": ["Dibyajyoti Chakraborty", "Romit Maulik", "Peter Harrington", "Dallas Foster", "Mohammad Amin Nabian", "Sanjay Choudhry"], "categories": ["cs.LG", "cs.AI", "physics.ao-ph", "physics.geo-ph"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2025-09-10", "url": "https://arxiv.org/abs/2509.09052", "pdf_url": "https://arxiv.org/pdf/2509.09052v1", "arxiv_id": "2509.09052", "doi": "10.48550/arXiv.2509.09052", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2991}
{"id": "b620fef51927a14eb641d5ca90c188ee4bd0fb8cb3b31c12cce515e0ce72fa1f", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerating Mixture-of-Expert Inference with Adaptive Expert Split Mechanism", "abstract": "Mixture-of-Experts (MoE) has emerged as a promising architecture for modern large language models (LLMs). However, massive parameters impose heavy GPU memory (i.e., VRAM) demands, hindering the widespread adoption of MoE LLMs. Offloading the expert parameters to CPU RAM offers an effective way to alleviate the VRAM requirements for MoE inference. Existing approaches typically cache a small subset of experts in VRAM and dynamically prefetch experts from RAM during inference, leading to significant degradation in inference speed due to the poor cache hit rate and substantial expert loading latency. In this work, we propose MoEpic, an efficient MoE inference system with a novel expert split mechanism. Specifically, each expert is vertically divided into two segments: top and bottom. MoEpic caches the top segment of hot experts, so that more experts will be stored under the limited VRAM budget, thereby improving the cache hit rate. During each layer's inference, MoEpic predicts and prefetches the activated experts for the next layer. Since the top segments of cached experts are exempt from fetching, the loading time is reduced, which allows efficient transfer-computation overlap. Nevertheless, the performance of MoEpic critically depends on the cache configuration (i.e., each layer's VRAM budget and expert split ratio). To this end, we propose a divide-and-conquer algorithm based on fixed-point iteration for adaptive cache configuration. Extensive experiments on popular MoE LLMs demonstrate that MoEpic can save about half of the GPU cost, while lowering the inference latency by about 37.51%-65.73% compared to the baselines.", "authors": ["Jiaming Yan", "Jianchun Liu", "Hongli Xu", "Liusheng Huang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-10", "url": "https://arxiv.org/abs/2509.08342", "pdf_url": "https://arxiv.org/pdf/2509.08342v1", "arxiv_id": "2509.08342", "doi": "10.48550/arXiv.2509.08342", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2991}
{"id": "da81a940ac07e25609284eec31bbd5870cff2b80a71abcfa48630f3506f64e4c", "sources": ["arxiv", "semantic_scholar"], "title": "Customizing the Inductive Biases of Softmax Attention using Structured Matrices", "abstract": "The core component of attention is the scoring function, which transforms the inputs into low-dimensional queries and keys and takes the dot product of each pair. While the low-dimensional projection improves efficiency, it causes information loss for certain tasks that have intrinsically high-dimensional inputs. Additionally, attention uses the same scoring function for all input pairs, without imposing a distance-dependent compute bias for neighboring tokens in the sequence. In this work, we address these shortcomings by proposing new scoring functions based on computationally efficient structured matrices with high ranks, including Block Tensor-Train (BTT) and contiguous Multi-Level Low Rank (MLR) matrices. On in-context regression tasks with high-dimensional inputs, our proposed scoring functions outperform standard attention for any fixed compute budget. On language modeling, a task that exhibits locality patterns, our MLR-based attention method achieves improved scaling laws compared to both standard attention and variants of sliding window attention. Additionally, we show that both BTT and MLR fall under a broader family of efficient structured matrices capable of encoding either full-rank or distance-dependent compute biases, thereby addressing significant shortcomings of standard attention. Finally, we show that MLR attention has promising results for long-range time-series forecasting.", "authors": ["Yilun Kuang", "Noah Amsel", "Sanae Lotfi", "Shikai Qiu", "Andres Potapczynski", "Andrew Gordon Wilson"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-09", "url": "https://arxiv.org/abs/2509.07963", "pdf_url": "https://arxiv.org/pdf/2509.07963v2", "arxiv_id": "2509.07963", "doi": "10.48550/arXiv.2509.07963", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/YilunKuang/structured-attention", "venue": "International Conference on Machine Learning", "quality_score": 0.4604}
{"id": "482d72a14540cca6420a3fe918bba00f9f26cec0b5377413ccd1e73c4e782146", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Adaptive Shared Experts with Grouped Multi-Head Attention Mixture of Experts", "abstract": "Transformer models based on the Mixture of Experts (MoE) architecture have made significant progress in long-sequence modeling, but existing models still have shortcomings in computational efficiency and the ability to capture long-range dependencies, especially in terms of the dynamic adaptability of expert resource allocation. In this paper, we propose a Dynamic Adaptive Shared Expert and Grouped Multi-Head Attention Hybrid Model (DASG-MoE) to enhance long-sequence modeling capabilities by integrating three modules. First, we employ the Grouped Multi-Head Attention (GMHA) mechanism to effectively reduce the computational complexity of long sequences. By parallel processing through sequence grouping, local sliding window attention, and feature aggregation, we address long-range dependency issues and the model's lack of generalization for local information. Second, we design a Dual-Scale Shared Expert Structure (DSSE), where shallow experts use lightweight computations to quickly respond to low-dimensional features, while deep experts process high-dimensional complex semantics through pre-training transfer and post-training optimization, achieving a dynamic balance between efficiency and accuracy. Third, we propose a hierarchical Adaptive Dynamic Routing (ADR) mechanism that dynamically selects expert levels based on feature complexity and task requirements, and optimizes resource allocation through a local expert activation strategy. Experiments on multiple long-sequence benchmark datasets demonstrate that our DASG-MoE model outperforms state-of-the-art models.", "authors": ["Cheng Li", "Jiexiong Liu", "Yixuan Chen", "Jie ji"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-05", "url": "https://arxiv.org/abs/2509.10530", "pdf_url": "https://arxiv.org/pdf/2509.10530v1", "arxiv_id": "2509.10530", "doi": "10.48550/arXiv.2509.10530", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2933}
{"id": "c1a08b67e5ee9748564ff053864f9c1e5a358203d76033b753c0d1dbb5d8f110", "sources": ["arxiv", "semantic_scholar"], "title": "KVCompose: Efficient Structured KV Cache Compression with Composite Tokens", "abstract": "Large language models (LLMs) rely on key-value (KV) caches for efficient autoregressive decoding; however, cache size grows linearly with context length and model depth, becoming a major bottleneck in long-context inference. Prior KV cache compression methods either enforce rigid heuristics, disrupt tensor layouts with per-attention-head variability, or require specialized compute kernels. We propose a simple, yet effective, KV cache compression framework based on attention-guided, layer-adaptive composite tokens. Our method aggregates attention scores to estimate token importance, selects head-specific tokens independently, and aligns them into composite tokens that respect the uniform cache structure required by existing inference engines. A global allocation mechanism further adapts retention budgets across layers, assigning more capacity to layers with informative tokens. This approach achieves significant memory reduction while preserving accuracy, consistently outperforming prior structured and semi-structured methods. Crucially, our approach remains fully compatible with standard inference pipelines, offering a practical and scalable solution for efficient long-context LLM deployment.", "authors": ["Dmitry Akulov", "Mohamed Sana", "Antonio De Domenico", "Tareq Si Salem", "Nicola Piovesan", "Fadhel Ayed"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-05", "url": "https://arxiv.org/abs/2509.05165", "pdf_url": "https://arxiv.org/pdf/2509.05165v2", "arxiv_id": "2509.05165", "doi": "10.48550/arXiv.2509.05165", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2933}
{"id": "32827409f2ffae9010d3b25e615f7bf888691801016463500e48b5d5575916e7", "sources": ["arxiv", "semantic_scholar"], "title": "Extracting Uncertainty Estimates from Mixtures of Experts for Semantic Segmentation", "abstract": "Estimating accurate and well-calibrated predictive uncertainty is important for enhancing the reliability of computer vision models, especially in safety-critical applications like traffic scene perception. While ensemble methods are commonly used to quantify uncertainty by combining multiple models, a mixture of experts (MoE) offers an efficient alternative by leveraging a gating network to dynamically weight expert predictions based on the input. Building on the promising use of MoEs for semantic segmentation in our previous works, we show that well-calibrated predictive uncertainty estimates can be extracted from MoEs without architectural modifications. We investigate three methods to extract predictive uncertainty estimates: predictive entropy, mutual information, and expert variance. We evaluate these methods for an MoE with two experts trained on a semantical split of the A2D2 dataset. Our results show that MoEs yield more reliable uncertainty estimates than ensembles in terms of conditional correctness metrics under out-of-distribution (OOD) data. Additionally, we evaluate routing uncertainty computed via gate entropy and find that simple gating mechanisms lead to better calibration of routing uncertainty estimates than more complex classwise gates. Finally, our experiments on the Cityscapes dataset suggest that increasing the number of experts can further enhance uncertainty calibration. Our code is available at https://github.com/KASTEL-MobilityLab/mixtures-of-experts/.", "authors": ["Svetlana Pavlitska", "Beyza Keskin", "Alwin Faßbender", "Christian Hubschneider", "J. Marius Zöllner"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-05", "url": "https://arxiv.org/abs/2509.04816", "pdf_url": "https://arxiv.org/pdf/2509.04816v1", "arxiv_id": "2509.04816", "doi": "10.1109/ICCVW69036.2025.00038", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/KASTEL-MobilityLab/mixtures-of-experts/", "venue": null, "quality_score": 0.3467}
{"id": "d434ddcf0f2b17817c292fc6dd788ff22adc2eddc90bff528d85a3aca7dad68a", "sources": ["arxiv", "semantic_scholar"], "title": "Semantic-guided LoRA Parameters Generation", "abstract": "Low-Rank Adaptation (LoRA) has demonstrated strong generalization capabilities across a variety of tasks for efficiently fine-tuning AI models, especially on resource-constrained edges. However, in real-world applications, edge users often exhibit task-specific preferences that are difficult to handle with a unified model trained under a closed-world assumption, and the challenge may further increase when there are significant domain shifts between training and deployment. Meanwhile, retraining/fine-tuning models for each user is also impractical due to its cost-intensive nature and privacy concerns over raw data utilization from edges. To address these challenges, we propose Semantic-guided LoRA Parameter Generation (SG-LoRA), the first of its kind framework to efficiently produce user-specific LoRA parameters without any additional training on user tasks or access to user-specific data. Concretely, SG-LoRA uses task descriptions as the semantic bridge, measuring their proximity to a set of known expert tasks in a shared embedding space. Based on this semantic guidance, it models the target task's LoRA parameter distribution to generate high-performing parameters for novel tasks. SG-LoRA enables the real-time construction of LoRA models aligned with individual intents by distilling knowledge from prominent LoRA experts and, meanwhile, offering a privacy-preserving solution for personalized model adaptation in a novel zero-shot open-world setting proposed in this work. Extensive experiments on multiple challenging tasks confirm the superior performance and remarkable adaptability of SG-LoRA. Code is available at https://github.com/keepgoingjkg/SG-LoRA.", "authors": ["Miaoge Li", "Yang Chen", "Zhijie Rao", "Can Jiang", "Jingcai Guo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-05", "url": "https://arxiv.org/abs/2509.10535", "pdf_url": "https://arxiv.org/pdf/2509.10535v1", "arxiv_id": "2509.10535", "doi": "10.48550/arXiv.2509.10535", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/keepgoingjkg/SG-LoRA", "venue": "arXiv.org", "quality_score": 0.4533}
{"id": "8cb1af491c54d2df9e237d053d443c229858b8f6ffa135cc268226ecd6eb6a3c", "sources": ["arxiv", "semantic_scholar"], "title": "Robust Experts: the Effect of Adversarial Training on CNNs with Sparse Mixture-of-Experts Layers", "abstract": "Robustifying convolutional neural networks (CNNs) against adversarial attacks remains challenging and often requires resource-intensive countermeasures. We explore the use of sparse mixture-of-experts (MoE) layers to improve robustness by replacing selected residual blocks or convolutional layers, thereby increasing model capacity without additional inference cost. On ResNet architectures trained on CIFAR-100, we find that inserting a single MoE layer in the deeper stages leads to consistent improvements in robustness under PGD and AutoPGD attacks when combined with adversarial training. Furthermore, we discover that when switch loss is used for balancing, it causes routing to collapse onto a small set of overused experts, thereby concentrating adversarial training on these paths and inadvertently making them more robust. As a result, some individual experts outperform the gated MoE model in robustness, suggesting that robust subpaths emerge through specialization. Our code is available at https://github.com/KASTEL-MobilityLab/robust-sparse-moes.", "authors": ["Svetlana Pavlitska", "Haixi Fan", "Konstantin Ditschuneit", "J. Marius Zöllner"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-05", "url": "https://arxiv.org/abs/2509.05086", "pdf_url": "https://arxiv.org/pdf/2509.05086v1", "arxiv_id": "2509.05086", "doi": "10.1109/ICCVW69036.2025.00032", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/KASTEL-MobilityLab/robust-sparse-moes", "venue": null, "quality_score": 0.3467}
{"id": "9e3b6b5c7f86350a90b7a361d089adc531c85d723044c2840d36446e7121e2da", "sources": ["arxiv", "semantic_scholar"], "title": "PagedEviction: Structured Block-wise KV Cache Pruning for Efficient Large Language Model Inference", "abstract": "KV caching significantly improves the efficiency of Large Language Model (LLM) inference by storing attention states from previously processed tokens, enabling faster generation of subsequent tokens. However, as sequence length increases, the KV cache quickly becomes a major memory bottleneck. To address this, we propose PagedEviction, a novel fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM's PagedAttention. Unlike existing approaches that rely on attention-based token importance or evict tokens across different vLLM pages, PagedEviction introduces an efficient block-wise eviction algorithm tailored for paged memory layouts. Our method integrates seamlessly with PagedAttention without requiring any modifications to its CUDA attention kernels. We evaluate PagedEviction across Llama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, and Llama-3.2-3B-Instruct models on the LongBench benchmark suite, demonstrating improved memory usage with better accuracy than baselines on long context tasks.", "authors": ["Krishna Teja Chitty-Venkata", "Jie Ye", "Xian-He Sun", "Anthony Kougkas", "Murali Emani", "Venkatram Vishwanath", "Bogdan Nicolae"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-04", "url": "https://arxiv.org/abs/2509.04377", "pdf_url": "https://arxiv.org/pdf/2509.04377v1", "arxiv_id": "2509.04377", "doi": "10.48550/arXiv.2509.04377", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference of the European Chapter of the Association for Computational Linguistics", "quality_score": 0.2922}
{"id": "434d271267a7c06e0ba1fdd376861ff140a99be7277d3b22f8a1756831257ede", "sources": ["arxiv", "semantic_scholar"], "title": "Integrating Pruning with Quantization for Efficient Deep Neural Networks Compression", "abstract": "Deep Neural Networks (DNNs) have achieved significant advances in a wide range of applications. However, their deployment on resource-constrained devices remains a challenge due to the large number of layers and parameters, which result in considerable computational and memory demands. To address this issue, pruning and quantization are two widely used compression techniques, commonly applied individually in most studies to reduce model size and enhance processing speed. Nevertheless, combining these two techniques can yield even greater compression benefits. Effectively integrating pruning and quantization to harness their complementary advantages poses a challenging task, primarily due to their potential impact on model accuracy and the complexity of jointly optimizing both processes. In this paper, we propose two approaches that integrate similarity-based filter pruning with Adaptive Power-of-Two (APoT) quantization to achieve higher compression efficiency while preserving model accuracy. In the first approach, pruning and quantization are applied simultaneously during training. In the second approach, pruning is performed first to remove less important parameters, followed by quantization of the pruned model using low-bit representations. Experimental results demonstrate that our proposed approaches achieve effective model compression with minimal accuracy degradation, making them well-suited for deployment on devices with limited computational resources.", "authors": ["Sara Makenali", "Babak Rokh", "Ali Azarpeyvand"], "categories": ["cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-04", "url": "https://arxiv.org/abs/2509.04244", "pdf_url": "https://arxiv.org/pdf/2509.04244v1", "arxiv_id": "2509.04244", "doi": "10.48550/arXiv.2509.04244", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2922}
{"id": "9cc0baffb9b71d66e5ef619517a423209bd23b09f38caca349aee3d0e5632e2b", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive KV-Cache Compression without Manually Setting Budget", "abstract": "Large language models (LLMs) inference relies heavily on KV-caches to accelerate autoregressive decoding, but the resulting memory footprint grows rapidly with sequence length, posing significant efficiency challenges. Current KV-cache compression methods suffer from a Procrustes' bed problem: they force diverse workloads into fixed compression ratios, leading to suboptimal resource allocation and inference performance. To this end, we present GVote, an adaptive KV-cache compression scheme that eliminates manual budget specification while achieving superior accuracy-efficiency trade-offs. GVote operates on the principle that the important keys are the aggregation of keys required by future queries. The method predicts future query attention demands by Monte-Carlo style sampling potential queries and aggregating selected keys to determine the optimal cache budget without manual specification. Experimental evaluation demonstrates GVote's effectiveness across multiple benchmarks, including GSM8K, RULER and Longbench. Compared to baselines, GVote exhibits 2$\\times$ memory reduction while the accuracy maintains higher or comparable.", "authors": ["Chenxia Tang", "Jianchun Liu", "Hongli Xu", "Liusheng Huang"], "categories": ["cs.DB", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-03", "url": "https://arxiv.org/abs/2509.03136", "pdf_url": "https://arxiv.org/pdf/2509.03136v1", "arxiv_id": "2509.03136", "doi": "10.48550/arXiv.2509.03136", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.291}
{"id": "2746d45bd43354c84d81501fa0fe35c7237b7fc8d78aa30221d02490e2509950", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Clustered-Experts: Advancing Expert Specialization and Generalization in Instruction Tuning", "abstract": "A sparse Mixture-of-Experts (MoE) architecture has emerged as a highly scalable solution by conditionally activating sub-modules without a proportional increase in computational costs. However, improving expert specialization to enhance performance and generalization remains a challenge for MoE, especially in instruction tuning scenarios characterized by significant input heterogeneity. In this work, we propose the Mixture-of-Clustered-Experts (MoCE) to address this limitation through a dual-stage routing mechanism. The first stage in the mechanism performs expert group routing based on sequence-level features, while the second stage activates the top-$k$ experts within the group at the token level. This approach enables the effective partitioning of heterogeneous inputs based on their knowledge requirements, encouraging expert group specialization while maintaining the advantages of token-level routing. We evaluate MoCE across a comprehensive set of benchmarks, demonstrating its consistent superiority over strong baselines and its enhanced generalization capabilities. Detailed analysis further highlights the robustness and effectiveness of MoCE.", "authors": ["Sugyeong Eo", "Jungjun Lee", "Chanjun Park", "Heuiseok Lim"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-03", "url": "https://arxiv.org/abs/2509.10513", "pdf_url": "https://arxiv.org/pdf/2509.10513v1", "arxiv_id": "2509.10513", "doi": "10.48550/arXiv.2509.10513", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.291}
{"id": "c3cb171ec0fec3a732c185391b2c7fbec78b0afd56dbc77c8866ee8aaa13e6b4", "sources": ["arxiv", "semantic_scholar"], "title": "MoPEQ: Mixture of Mixed Precision Quantized Experts", "abstract": "Large Language and Vision Models using a Mixture-of-Experts (MoE) architecture pose significant challenges for deployment due to their computational and memory demands. Mixed Precision Quantization assigns different precisions to different layers of an LLM/VLM based on layer sensitivity and importance within the model. In this work, we propose a Post Training Quantization algorithm, MoPEQ, that assigns optimal bit width to each expert. Our method balances accuracy and model size by analyzing each expert's sensitivity using Hessian trace approximation instead of relying on the activation frequency of the expert. This per-expert granularity approach clusters similar experts to maintain model performance while reducing memory requirements. The experimental results on VLMEvalKit benchmark datasets using State-of-the-art VLMs Deepseek-VL2 -tiny, -small, -base, and MolmoE models demonstrate that our mixed precision quantized MoEs achieve competitive accuracy with substantial improvements in memory footprint compared to uniform-precision baseline methods. We perform a comprehensive study to analyze the impact of expert activation frequency and sensitivity using Hessian trace approximation at both layer-wise and model-wide expert precision allocation of 2, 3, and 4 bits to provide a thorough understanding of mixed precision quantization of VLM-MoEs.", "authors": ["Krishna Teja Chitty-Venkata", "Jie Ye", "Murali Emani"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-02", "url": "https://arxiv.org/abs/2509.02512", "pdf_url": "https://arxiv.org/pdf/2509.02512v1", "arxiv_id": "2509.02512", "doi": "10.1109/ICCVW69036.2025.00424", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1845}
{"id": "35910223297b091446d2473b3db2cebf94ea0f4ee6e7d3d238a948c723470c7b", "sources": ["arxiv", "semantic_scholar"], "title": "Performance Evaluation of LoRa for IoT Applications in Non-Terrestrial Networks via ns-3", "abstract": "The integration of Internet of Things (IoT) and Non-Terrestrial Networks (NTNs) has emerged as a key paradigm to provide connectivity for sensors and actuators via satellite gateways in remote areas where terrestrial infrastructure is limited or unavailable. Among other Low-Power Wide-Area Network (LPWAN) technologies for IoT, Long Range (LoRa) holds great potential given its long range, energy efficiency, and flexibility. In this paper, we explore the feasibility and performance of LoRa to support large-scale IoT connectivity through Low Earth Orbit (LEO) satellite gateways. To do so, we developed a new ns3-LoRa-NTN simulation module, which integrates and extends the ns3-LoRa and ns3-NTN modules, to enable full-stack end-to-end simulation of satellite communication in LoRa networks. Our results, given in terms of average data rate and Packet Reception Ratio (PRR), confirm that LoRa can effectively support direct communication from the ground to LEO satellites, but network optimization is required to mitigate collision probability when end nodes use the same Spreading Factors (SFs) over long distances.", "authors": ["Alessandro Traspadini", "Michele Zorzi", "Marco Giordani"], "categories": ["cs.NI", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-02", "url": "https://arxiv.org/abs/2509.02811", "pdf_url": "https://arxiv.org/pdf/2509.02811v1", "arxiv_id": "2509.02811", "doi": "10.1109/GLOBECOM59602.2025.11432761", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Global Communications Conference", "quality_score": 0.2899}
{"id": "5c9179b3d5330344223584866b2e8371e7e3d01d00dd9164c3dfec7a00c66c72", "sources": ["arxiv", "semantic_scholar"], "title": "LExI: Layer-Adaptive Active Experts for Efficient MoE Model Inference", "abstract": "Mixture-of-Experts (MoE) models scale efficiently by activating only a subset of experts per token, offering a computationally sparse alternative to dense architectures. While prior post-training optimizations, such as inter- and intra-expert pruning, reduce memory usage they provide limited gains in inference-time compute efficiency. Moreover, existing MoE architectures typically activate a fixed number of experts uniformly across all layers, resulting in redundant computation and suboptimal performance. In this work, we first demonstrate that MoE pruning strategies improve only the memory footprint but do not significantly improve inference performance on GPU using optimized frameworks such as vLLM. To address this, we introduce LExI, a data-free optimization technique that determines the optimal number of active experts per layer in a pretrained MoE model. LExI leverages only the model weights to estimate the relative importance of each layer and adaptively assigns the number of active experts accordingly per layer. Experiments on state-of-the-art language and vision MoE benchmarks demonstrate that LExI significantly outperforms traditional MoE pruning approaches in terms of inference efficiency with negligible accuracy loss. For example, using LExI, Qwen1.5-MoE achieves the same throughput on Nvidia H100 GPU with 10% better accuracy than traditional expert pruning.", "authors": ["Krishna Teja Chitty-Venkata", "Sandeep Madireddy", "Murali Emani", "Venkatram Vishwanath"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-02", "url": "https://arxiv.org/abs/2509.02753", "pdf_url": "https://arxiv.org/pdf/2509.02753v1", "arxiv_id": "2509.02753", "doi": "10.48550/arXiv.2509.02753", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2899}
{"id": "86a755cd04e000d5bdc906a28f4d6d56aac424c0888512ec2dfeeef31768e6cf", "sources": ["arxiv", "semantic_scholar"], "title": "Cache Management for Mixture-of-Experts LLMs -- extended version", "abstract": "Large language models (LLMs) have demonstrated remarkable capabilities across a variety of tasks. One of the main challenges towards the successful deployment of LLMs is memory management, since they typically involve billions of parameters. To this end, architectures based on Mixture-of-Experts have been proposed, which aim to reduce the size of the parameters that are activated when producing a token. This raises the equally critical issue of efficiently managing the limited cache of the system, in that frequently used experts should be stored in the fast cache rather than in the slower secondary memory. In this work, we introduce and study a new paging problem that models expert management optimization. Our formulation captures both the layered architecture of LLMs and the requirement that experts are cached efficiently. We first present lower bounds on the competitive ratio of both deterministic and randomized algorithms, which show that under mild assumptions, LRU-like policies have good theoretical competitive performance. We then propose a layer-based extension of LRU that is tailored to the problem at hand. Extensive simulations on both synthetic datasets and actual traces of MoE usage show that our algorithm outperforms policies for the classic paging problem, such as the standard LRU.", "authors": ["Spyros Angelopoulos", "Loris Marchal", "Adrien Obrecht", "Bertrand Simon"], "categories": ["cs.LG", "cs.DS"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-02", "url": "https://arxiv.org/abs/2509.02408", "pdf_url": "https://arxiv.org/pdf/2509.02408v1", "arxiv_id": "2509.02408", "doi": "10.1007/978-3-031-99872-0_2", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "European Conference on Parallel Processing", "quality_score": 0.2899}
{"id": "524026dc4f30a365e54cc6bfeacbc6015364ed392a4b5b634ef98ce7fa0485ba", "sources": ["arxiv", "semantic_scholar"], "title": "EndoGMDE: Generalizable Monocular Depth Estimation with Mixture of Low-Rank Experts for Diverse Endoscopic Scenes", "abstract": "Self-supervised monocular depth estimation is a significant task for low-cost and efficient 3D scene perception and measurement in endoscopy. However, the variety of illumination conditions and scene features is still the primary challenges for depth estimation in endoscopic scenes. In this work, a novel self-supervised framework is proposed for monocular depth estimation in diverse endoscopy. Firstly, considering the diverse features in endoscopic scenes with different tissues, a novel block-wise mixture of dynamic low-rank experts is proposed to efficiently finetune the foundation model for endoscopic depth estimation. In the proposed module, based on the input feature, different experts with a small amount of trainable parameters are adaptively selected for weighted inference, from low-rank experts which are allocated based on the generalization of each block. Moreover, a novel self-supervised training framework is proposed to jointly cope with brightness inconsistency and reflectance interference. The proposed method outperforms state-of-the-art works on SCARED dataset and SimCol dataset. Furthermore, the proposed network also achieves the best generalization based on zero-shot depth estimation on C3VD, Hamlyn and SERV-CT dataset. The outstanding performance of our model is further demonstrated with 3D reconstruction and ego-motion estimation. The proposed method could contribute to accurate endoscopy for minimally invasive measurement and surgery. The evaluation codes will be released upon acceptance, while the demo videos can be found on: https://endo-gmde.netlify.app/.", "authors": ["Liangjing Shao", "Chenkang Du", "Benshuang Chen", "Xueli Liu", "Xinrong Chen"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-01", "url": "https://arxiv.org/abs/2509.01206", "pdf_url": "https://arxiv.org/pdf/2509.01206v3", "arxiv_id": "2509.01206", "doi": "10.48550/arXiv.2509.01206", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2888}
{"id": "e8728912ae9f33154fa02a0db2c8a5a86045749406af3a84d07835656f96fd03", "sources": ["arxiv", "semantic_scholar"], "title": "MEPT: Mixture of Expert Prompt Tuning as a Manifold Mapper", "abstract": "Considering deep neural networks as manifold mappers, the pretrain-then-fine-tune paradigm can be interpreted as a two-stage process: pretrain establishes a broad knowledge base, and fine-tune adjusts the model parameters to activate specific neural pathways to align with the target manifold. Although prior fine-tuning approaches demonstrate success, their rigid parameter space limits their ability to dynamically activate appropriate neural pathways, rendering them ill-equipped to adapt flexibly to the diverse and evolving data distributions. In light of this view, we propose a novel approach, Mixture of Expert Prompt Tuning (MEPT), as an effective and efficient manifold-mapping framework. MEPT leverages the Mixture of Experts architecture by integrating multiple prompt experts to adaptively learn diverse and non-stationary data distributions. Empirical evaluations demonstrate that MEPT outperforms several state-of-the-art parameter efficient baselines on SuperGLUE, achieving notable improvements in mean accuracy (e.g., 1.94%) while significantly reducing activated prompts by 79.25%. The effectiveness of MEPT is further supported by theoretical insights from manifold learning and validated through neural activation pathway visualization results. Our code is avaliable at https://runjia.tech/emnlp_mept/.", "authors": ["Runjia Zeng", "Guangyan Sun", "Qifan Wang", "Tong Geng", "Sohail Dianat", "Xiaotian Han", "Raghuveer Rao", "Xueling Zhang", "Cheng Han", "Lifu Huang", "Dongfang Liu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-31", "url": "https://arxiv.org/abs/2509.00996", "pdf_url": "https://arxiv.org/pdf/2509.00996v2", "arxiv_id": "2509.00996", "doi": "10.48550/arXiv.2509.00996", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2876}
{"id": "8dba47b3b4419f5735e217f52b60708aafa7875643c097f92ea7f64f486c8828", "sources": ["arxiv", "semantic_scholar"], "title": "Router Upcycling: Leveraging Mixture-of-Routers in Mixture-of-Experts Upcycling", "abstract": "The Mixture-of-Experts (MoE) models have gained significant attention in deep learning due to their dynamic resource allocation and superior performance across diverse tasks. However, efficiently training these models remains challenging. The MoE upcycling technique has been proposed to reuse and improve existing model components, thereby minimizing training overhead. Despite this, simple routers, such as linear routers, often struggle with complex routing tasks within MoE upcycling. In response, we propose a novel routing technique called Router Upcycling to enhance the performance of MoE upcycling models. Our approach initializes multiple routers from the attention heads of preceding attention layers during upcycling. These routers collaboratively assign tokens to specialized experts in an attention-like manner. Each token is processed into diverse queries and aligned with the experts' features (serving as keys). Experimental results demonstrate that our method achieves state-of-the-art (SOTA) performance, outperforming other upcycling baselines.", "authors": ["Junfeng Ran", "Guangxiang Zhao", "Yuhan Wu", "Dawei Zhu", "Longyun Wu", "Yikai Zhao", "Tong Yang", "Lin Sun", "Xiangzheng Zhang", "Sujian Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-31", "url": "https://arxiv.org/abs/2509.00679", "pdf_url": "https://arxiv.org/pdf/2509.00679v1", "arxiv_id": "2509.00679", "doi": "10.48550/arXiv.2509.00679", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "European Conference on Artificial Intelligence", "quality_score": 0.2876}
{"id": "ac73d2917721adc49607d2bbeea8a5c4caea488762d801161a683f483efadcd7", "sources": ["arxiv", "semantic_scholar"], "title": "GraphKV: Breaking the Static Selection Paradigm with Graph-Based KV Cache Eviction", "abstract": "Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance. Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference. To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression. In GraphKV, tokens are modeled as nodes with importance scores, and edges represent their similarity relationships. Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens. GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-and-play manner. Codes will be released on Github.", "authors": ["Xuelin Li", "Xiangqi Jin", "Linfeng Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-30", "url": "https://arxiv.org/abs/2509.00388", "pdf_url": "https://arxiv.org/pdf/2509.00388v1", "arxiv_id": "2509.00388", "doi": "10.48550/arXiv.2509.00388", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2865}
{"id": "9439ebae2a0e38e6082044f089793a7d6a6040937d660e7df55c2ca44a8d21e8", "sources": ["arxiv", "semantic_scholar"], "title": "KVComp: A High-Performance, LLM-Aware, Lossy Compression Framework for KV Cache", "abstract": "Transformer-based large language models (LLMs) demonstrate impressive potential in various practical applications. However, long context inference poses a significant challenge due to the enormous memory requirements of the key-value (KV) cache, which can scale to multiple gigabytes as sequence length and batch size increase. In this paper, we present KVComp, a generic and efficient KV cache management framework optimized for long-text generation that synergistically works with both latency-critical and throughput-critical inference systems. KVComp employs novel lossy compression techniques specifically designed for KV cache data characteristics, featuring careful co-design of compression algorithms and system architecture. Our approach maintains compatibility with the growing nature of KV cache while preserving high computational efficiency. Experimental results show that KVComp achieves on average 47\\% and up to 83\\% higher memory reduction rate compared to existing methods with little/no model accuracy degradation. Furthermore, KVComp achieves extremely high execution throughput, effectively reducing decompression overhead and, in some cases, even accelerating the matrix-vector multiplication operation and outperform cuBLAS-based attention kernels with less data movement.", "authors": ["Bo Jiang", "Taolue Yang", "Youyuan Liu", "Chengming Zhang", "Xubin He", "Sian Jin"], "categories": ["cs.DC", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-30", "url": "https://arxiv.org/abs/2509.00579", "pdf_url": "https://arxiv.org/pdf/2509.00579v1", "arxiv_id": "2509.00579", "doi": "10.48550/arXiv.2509.00579", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2865}
{"id": "79c82a20f86a9d097206c786e5a3db6cd589c856a8148e7b30b1133705050664", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Global and Local Experts with Diffusion Transformer for Controllable Face Generation", "abstract": "Controllable face generation poses critical challenges in generative modeling due to the intricate balance required between semantic controllability and photorealism. While existing approaches struggle with disentangling semantic controls from generation pipelines, we revisit the architectural potential of Diffusion Transformers (DiTs) through the lens of expert specialization. This paper introduces Face-MoGLE, a novel framework featuring: (1) Semantic-decoupled latent modeling through mask-conditioned space factorization, enabling precise attribute manipulation; (2) A mixture of global and local experts that captures holistic structure and region-level semantics for fine-grained controllability; (3) A dynamic gating network producing time-dependent coefficients that evolve with diffusion steps and spatial locations. Face-MoGLE provides a powerful and flexible solution for high-quality, controllable face generation, with strong potential in generative modeling and security applications. Extensive experiments demonstrate its effectiveness in multimodal and monomodal face generation settings and its robust zero-shot generalization capability. Project page is available at https://github.com/XavierJiezou/Face-MoGLE.", "authors": ["Xuechao Zou", "Shun Zhang", "Xing Fu", "Yue Li", "Kai Li", "Yushe Cao", "Congyan Lang", "Pin Tao", "Junliang Xing"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-30", "url": "https://arxiv.org/abs/2509.00428", "pdf_url": "https://arxiv.org/pdf/2509.00428v1", "arxiv_id": "2509.00428", "doi": "10.48550/arXiv.2509.00428", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/XavierJiezou/Face-MoGLE", "venue": "arXiv.org", "quality_score": 0.4427}
{"id": "15d6a2f927d4c2ebd145bd467769306cfdc6d0ccd385a856fcf02f73aae45288", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerating Mixture-of-Experts Inference by Hiding Offloading Latency with Speculative Decoding", "abstract": "Recent advancements in Mixture of Experts (MoE) models have significantly increased their parameter scale as well as model performance. Extensive offloading techniques have been proposed to address the GPU memory limitations of MoE inference. However, due to the I/O bottleneck and sparse computation of MoE models, existing offloading techniques still suffer from low hardware utilization. To fully utilize the hardware resources, we propose SpecMoEOff, which employs the speculative decoding technique to enlarge the workload of each expert. SpecMoEOff orchestrates the GPU and CPU by both theoretical and empirical roofline analysis. In addition, we develop a dedicated CPU chunked attention verification kernel to fit the speculative decoding in offloading scenarios as well as minimizing the additional overhead led by draft models. SpecMoEOff further integrates an optimizer to automatically tune the hyperparameters of speculative decoding under given hardware and workload. Experimental results show that SpecMoEOff achieves up to 2.5x decode throughput improvement over the state-of-the-art MoE offloading techniques.", "authors": ["Zhibin Wang", "Zhonghui Zhang", "Yuhang Zhou", "Zibo Wang", "Mo Zhou", "Peng Jiang", "Weilin Cai", "Chengying Huan", "Rong Gu", "Sheng Zhong", "Chen Tian"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-29", "url": "https://arxiv.org/abs/2508.21706", "pdf_url": "https://arxiv.org/pdf/2508.21706v2", "arxiv_id": "2508.21706", "doi": "10.48550/arXiv.2508.21706", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2853}
{"id": "6229c08cc79d8a23d2cc6781be888375871334e73337be5c042971bb591fda1f", "sources": ["arxiv", "semantic_scholar"], "title": "From TLinFormer to TConstFormer: The Leap to Constant-Time Transformer Attention: Achieving O(1) Computation and O(1) KV Cache during Autoregressive Inference", "abstract": "Although the Transformer has become the cornerstone of modern AI, its autoregressive inference suffers from a linearly growing KV Cache and a computational complexity of O(N^2 d), severely hindering its ability to process ultra-long sequences. To overcome this limitation, this paper introduces the TConstFormer architecture, building upon our previous work, TLinFormer. TConstFormer employs an innovative periodic state update mechanism to achieve a truly constant-size O(1) KV Cache. The computational complexity of this mechanism is also O(1) in an amortized sense: it performs purely constant-time computations for $k-1$ consecutive steps (e.g., $k=256$) and executes a single linear-time global information synchronization only on the $k$-th step. Theoretical calculations and experimental results demonstrate that TConstFormer exhibits an overwhelming advantage over baseline models in terms of speed, memory efficiency, and overall performance on long-text inference tasks. This breakthrough paves the way for efficient and robust streaming language model applications.", "authors": ["Zhongpan Tang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-29", "url": "https://arxiv.org/abs/2509.00202", "pdf_url": "https://arxiv.org/pdf/2509.00202v1", "arxiv_id": "2509.00202", "doi": "10.48550/arXiv.2509.00202", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2853}
{"id": "2bf68e3f3728a591967d3b99de57a2198e82171e4af2a530d8516d56ff915306", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-Health: A Mixture of Experts Framework for Robust Multimodal Healthcare Prediction", "abstract": "Healthcare systems generate diverse multimodal data, including Electronic Health Records (EHR), clinical notes, and medical images. Effectively leveraging this data for clinical prediction is challenging, particularly as real-world samples often present with varied or incomplete modalities. Existing approaches typically require complete modality data or rely on manual selection strategies, limiting their applicability in real-world clinical settings where data availability varies across patients and institutions. To address these limitations, we propose MoE-Health, a novel Mixture of Experts framework designed for robust multimodal fusion in healthcare prediction. MoE-Health architecture is specifically developed to handle samples with differing modalities and improve performance on critical clinical tasks. By leveraging specialized expert networks and a dynamic gating mechanism, our approach dynamically selects and combines relevant experts based on available data modalities, enabling flexible adaptation to varying data availability scenarios. We evaluate MoE-Health on the MIMIC-IV dataset across three critical clinical prediction tasks: in-hospital mortality prediction, long length of stay, and hospital readmission prediction. Experimental results demonstrate that MoE-Health achieves superior performance compared to existing multimodal fusion methods while maintaining robustness across different modality availability patterns. The framework effectively integrates multimodal information, offering improved predictive performance and robustness in handling heterogeneous and incomplete healthcare data, making it particularly suitable for deployment in diverse healthcare environments with heterogeneous data availability.", "authors": ["Xiaoyang Wang", "Christopher C. Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-29", "url": "https://arxiv.org/abs/2508.21793", "pdf_url": "https://arxiv.org/pdf/2508.21793v1", "arxiv_id": "2508.21793", "doi": "10.1145/3765612.3767207", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "ACM International Conference on Bioinformatics, Computational Biology and Biomedicine", "quality_score": 0.2853}
{"id": "c75add2215b366c55a2f4a51ea6488050ac0a7fc2db2f4a925f58e03f76fb7fe", "sources": ["arxiv", "semantic_scholar"], "title": "AdaptCache: KV Cache Native Storage Hierarchy for Low-Delay and High-Quality Language Model Serving", "abstract": "Large language model (LLM) applications often reuse previously processed context, such as chat history and documents, which introduces significant redundant computation. Existing LLM serving systems address such redundant computation by storing the KV caches of processed context and loading the corresponding KV cache when a new request reuses the context. Further, as these LLM applications scale, the total size of KV caches becomes excessively large and requires both DRAM and SSD for full storage. However, prior work that stores KV caches in DRAM and SSD suffers from high loading delays, as most KV cache hits come from SSD, which is slow to load. To increase the KV cache hit rate on DRAM, we identify lossy KV cache compression as a promising approach. We design a lossy compression system that decides the compression algorithm, compression rate and device placement for each KV cache entry to maximise DRAM hits and minimise loading delay without significantly degrading generation quality. Compared to various static compression baselines across three tasks, our system AdaptCache achieves 1.43--2.4 x delay savings at the same quality and 6--55% quality improvements at the same delay.", "authors": ["Shaoting Feng", "Hanchen Li", "Kuntai Du", "Zhuohan Gu", "Yuhan Liu", "Jiayi Yao", "Siddhant Ray", "Samuel Shen", "Yihua Cheng", "Ganesh Ananthanarayanan", "Junchen Jiang"], "categories": ["cs.OS", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-28", "url": "https://arxiv.org/abs/2509.00105", "pdf_url": "https://arxiv.org/pdf/2509.00105v2", "arxiv_id": "2509.00105", "doi": "10.48550/arXiv.2509.00105", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2842}
{"id": "7744760690fc648b1a23cad74cdb2d7a3433ff0bf1ba92350a81260d548abf17", "sources": ["arxiv", "semantic_scholar"], "title": "A Mixture of Experts Gating Network for Enhanced Surrogate Modeling in External Aerodynamics", "abstract": "The computational cost associated with high-fidelity CFD simulations remains a significant bottleneck in the automotive design and optimization cycle. While ML-based surrogate models have emerged as a promising alternative to accelerate aerodynamic predictions, the field is characterized by a diverse and rapidly evolving landscape of specialized neural network architectures, with no single model demonstrating universal superiority. This paper introduces a novel meta-learning framework that leverages this architectural diversity as a strength. We propose a Mixture of Experts (MoE) model that employs a dedicated gating network to dynamically and optimally combine the predictions from three heterogeneous, state-of-the-art surrogate models: DoMINO, a decomposable multi-scale neural operator; X-MeshGraphNet, a scalable multi-scale graph neural network; and FigConvNet, a factorized implicit global convolution network. The gating network learns a spatially-variant weighting strategy, assigning credibility to each expert based on its localized performance in predicting surface pressure and wall shear stress fields. To prevent model collapse and encourage balanced expert contributions, we integrate an entropy regularization term into the training loss function. The entire system is trained and validated on the DrivAerML dataset, a large-scale, public benchmark of high-fidelity CFD simulations for automotive aerodynamics. Quantitative results demonstrate that the MoE model achieves a significant reduction in L-2 prediction error, outperforming not only the ensemble average but also the most accurate individual expert model across all evaluated physical quantities. This work establishes the MoE framework as a powerful and effective strategy for creating more robust and accurate composite surrogate models by synergistically combining the complementary strengths of specialized architectures.", "authors": ["Mohammad Amin Nabian", "Sanjay Choudhry"], "categories": ["cs.LG", "cs.AI", "math.NA", "physics.flu-dyn"], "fields_of_study": ["Computer Science", "Mathematics", "Physics"], "published_date": "2025-08-28", "url": "https://arxiv.org/abs/2508.21249", "pdf_url": "https://arxiv.org/pdf/2508.21249v1", "arxiv_id": "2508.21249", "doi": "10.48550/arXiv.2508.21249", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2842}
{"id": "b045dc2d4b001dad0deedd910366e026cf9a37ea8a1c2c995b80148b13cef5b1", "sources": ["arxiv", "semantic_scholar"], "title": "Spotlight Attention: Towards Efficient LLM Generation via Non-linear Hashing-based KV Cache Retrieval", "abstract": "Reducing the key-value (KV) cache burden in Large Language Models (LLMs) significantly accelerates inference. Dynamically selecting critical KV caches during decoding helps maintain performance. Existing methods use random linear hashing to identify important tokens, but this approach is inefficient due to the orthogonal distribution of queries and keys within two narrow cones in LLMs. We introduce Spotlight Attention, a novel method that employs non-linear hashing functions to optimize the embedding distribution of queries and keys, enhancing coding efficiency and robustness. We also developed a lightweight, stable training framework using a Bradley-Terry ranking-based loss, enabling optimization of the non-linear hashing module on GPUs with 16GB memory in 8 hours. Experimental results show that Spotlight Attention drastically improves retrieval precision while shortening the length of the hash code at least 5$\\times$ compared to traditional linear hashing. Finally, we exploit the computational advantages of bitwise operations by implementing specialized CUDA kernels, achieving hashing retrieval for 512K tokens in under 100$μ$s on a single A100 GPU, with end-to-end throughput up to 3$\\times$ higher than vanilla decoding.", "authors": ["Wenhao Li", "Yuxin Zhang", "Gen Luo", "Haiyuan Wan", "Ziyang Gong", "Fei Chao", "Rongrong Ji"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-27", "url": "https://arxiv.org/abs/2508.19740", "pdf_url": "https://arxiv.org/pdf/2508.19740v4", "arxiv_id": "2508.19740", "doi": "10.48550/arXiv.2508.19740", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.283}
{"id": "9223e94241cbcf0f8c9deaee9e22188c193c1fabc0be8c23fe54562ad4202982", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding and Leveraging the Expert Specialization of Context Faithfulness in Mixture-of-Experts LLMs", "abstract": "Context faithfulness is essential for reliable reasoning in context-dependent scenarios. However, large language models often struggle to ground their outputs in the provided context, resulting in irrelevant responses. Inspired by the emergent expert specialization observed in mixture-of-experts architectures, this work investigates whether certain experts exhibit specialization in context utilization, offering a potential pathway toward targeted optimization for improved context faithfulness. To explore this, we propose Router Lens, a method that accurately identifies context-faithful experts. Our analysis reveals that these experts progressively amplify attention to relevant contextual information, thereby enhancing context grounding. Building on this insight, we introduce Context-faithful Expert Fine-Tuning (CEFT), a lightweight optimization approach that selectively fine-tunes context-faithful experts. Experiments across a wide range of benchmarks and models demonstrate that CEFT matches or surpasses the performance of full fine-tuning while being significantly more efficient.", "authors": ["Jun Bai", "Minghao Tong", "Yang Liu", "Zixia Jia", "Zilong Zheng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-27", "url": "https://arxiv.org/abs/2508.19594", "pdf_url": "https://arxiv.org/pdf/2508.19594v3", "arxiv_id": "2508.19594", "doi": "10.48550/arXiv.2508.19594", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.283}
{"id": "80791913f9e62b8e6a9523dabe33c0995b920b83fbea3027396ab3469045e844", "sources": ["arxiv", "semantic_scholar"], "title": "MODE: Mixture of Document Experts for RAG", "abstract": "Retrieval-Augmented Generation (RAG) often relies on large vector databases and cross-encoders tuned for large-scale corpora, which can be excessive for small, domain-specific collections. We present MODE (Mixture of Document Experts), a lightweight alternative that replaces fine-grained nearest-neighbor search with cluster-and-route retrieval. Documents are embedded, grouped into semantically coherent clusters, and represented by cached centroids. At query time, we route to the top centroid(s) and retrieve context only within those clusters, eliminating external vector-database infrastructure and reranking while keeping latency low. On HotpotQA and SQuAD corpora with 100-500 chunks, MODE matches or exceeds a dense-retrieval baseline in answer quality while reducing end-to-end retrieval time. Ablations show that cluster granularity and multi-cluster routing control the recall/precision trade-off, and that tighter clusters improve downstream accuracy. MODE offers a practical recipe for small and medium corpora where simplicity, speed, and topical focus matter.", "authors": ["Rahul Anand"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-27", "url": "https://arxiv.org/abs/2509.00100", "pdf_url": "https://arxiv.org/pdf/2509.00100v1", "arxiv_id": "2509.00100", "doi": "10.48550/arXiv.2509.00100", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.283}
{"id": "9bfc8def0d53d482a3c05ac788c337213d79988b62878062b297a04f39b36418", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Debiasing for Token Pruning in Vision Language Models", "abstract": "Vision-language models (VLMs) typically encode substantially more visual tokens than text tokens, resulting in significant token redundancy. Pruning uninformative visual tokens is therefore crucial for improving computational efficiency, and language-to-vision attention has become a widely used importance criterion for this purpose. However, we find that attention in VLMs is systematically biased. It disproportionately favors tokens appearing later in the sequence, manifesting as over-attention to lower image regions, and assigns inflated scores to semantically empty padding tokens. These behaviors stem from intrinsic recency bias and attention sink effects inherited from large language models (LLMs), and they distort attention-based pruning by preserving irrelevant visual content. To derive a pruning criterion better aligned with semantic relevance, we introduce two lightweight yet effective debiasing techniques that restore the reliability of attention. The first compensates for positional distortions by removing recency-induced attention trends, producing a content-aware and position-agnostic importance measure. The second suppresses attention sink effects by eliminating spurious attention on padding tokens. Our method is model-agnostic, pruning-method-agnostic, and task-agnostic, enabling plug-and-play integration with existing VLM pruning models. Despite its simplicity, our approach consistently delivers strong performance gains. We evaluate our method on ten vision-language benchmarks spanning both image-based and video-based tasks, in comparison with seven state-of-the-art visual token pruning methods and across two representative VLM architectures. Our method achieves substantial performance gains, demonstrating strong effectiveness and generalizability. Our code is available at https://github.com/intcomp/attention-bias.", "authors": ["Kai Zhao", "Wubang Yuan", "Yuchen Lin", "Liting Ruan", "Xiaofeng Lu", "Deng-Ping Fan", "Ming-Ming Cheng", "Dan Zeng"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-25", "url": "https://arxiv.org/abs/2508.17807", "pdf_url": "https://arxiv.org/pdf/2508.17807v2", "arxiv_id": "2508.17807", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/intcomp/attention-bias", "venue": null, "quality_score": 0.3318}
{"id": "27005a78fad2a9b7bfd063c9b499f62140d43dfec831ae0b40fae92dfa7af688", "sources": ["arxiv", "semantic_scholar"], "title": "ISALux: Illumination and Segmentation Aware Transformer Employing Mixture of Experts for Low Light Image Enhancement", "abstract": "We introduce ISALux, a novel transformer-based approach for Low-Light Image Enhancement (LLIE) that seamlessly integrates illumination and semantic priors. Our architecture includes an original self-attention block, Hybrid Illumination and Semantics-Aware Multi-Headed Self- Attention (HISA-MSA), which integrates illumination and semantic segmentation maps for en- hanced feature extraction. ISALux employs two self-attention modules to independently process illumination and semantic features, selectively enriching each other to regulate luminance and high- light structural variations in real-world scenarios. A Mixture of Experts (MoE)-based Feed-Forward Network (FFN) enhances contextual learning, with a gating mechanism conditionally activating the top K experts for specialized processing. To address overfitting in LLIE methods caused by distinct light patterns in benchmarking datasets, we enhance the HISA-MSA module with low-rank matrix adaptations (LoRA). Extensive qualitative and quantitative evaluations across multiple specialized datasets demonstrate that ISALux is competitive with state-of-the-art (SOTA) methods. Addition- ally, an ablation study highlights the contribution of each component in the proposed model. Code will be released upon publication.", "authors": ["Raul Balmez", "Alexandru Brateanu", "Ciprian Orhei", "Codruta Ancuti", "Cosmin Ancuti"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-25", "url": "https://arxiv.org/abs/2508.17885", "pdf_url": "https://arxiv.org/pdf/2508.17885v1", "arxiv_id": "2508.17885", "doi": "10.48550/arXiv.2508.17885", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2807}
{"id": "9775c8f168cf645d487b8da8fe033c7879a80e9d4ce6055226c2548d9797d6a2", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Quantized Planetary Crater Detection System for Autonomous Space Exploration", "abstract": "Autonomous planetary exploration demands real-time, high-fidelity environmental perception. Standard deep learning models require massive computational resources. Conversely, space-qualified onboard computers operate under strict power, thermal, and memory limits. This disparity creates a severe engineering bottleneck, preventing the deployment of highly capable perception architectures on extraterrestrial exploration platforms. In this foundational concept paper, we propose the theoretical architecture for the Adaptive Quantized Planetary Crater Detection System (AQ-PCDSys) to resolve this bottleneck. We present a mathematical blueprint integrating an INT8 Quantized Neural Network (QNN) designed specifically for Quantization Aware Training (QAT). To address sensor fragility, we mathematically formalize an Adaptive Multi-Sensor Fusion (AMF) module. By deriving the exact integer requantization multiplier required for spatial attention gating, this module actively selects and fuses Optical Imagery (OI) and Digital Elevation Models (DEMs) at the feature level, ensuring reliable perception during extreme cross-illuminations and optical hardware dropouts. Furthermore, the architecture introduces anchor-free, center-to-edge regression heads, protected by a localized FP16 coordinate conversion, to accurately frame asymmetrical lunar craters without catastrophic integer truncation. Rather than presenting physical hardware telemetry, this manuscript establishes the theoretical bounds, structural logic, and mathematical justifications for the architecture. We outline a rigorous Hardware-in-the-Loop (HITL) evaluation protocol to define the exact testing criteria required for future empirical validation, paving the way for next-generation space-mission software design.", "authors": ["Aditri Paul", "Archan Paul"], "categories": ["cs.LG", "cs.AI", "cs.CV", "cs.ET", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-08-25", "url": "https://arxiv.org/abs/2508.18025", "pdf_url": "https://arxiv.org/pdf/2508.18025v4", "arxiv_id": "2508.18025", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1786}
{"id": "db724565b6206a474fe9b243d00aa03111359e97c715b0f3df37e7d427c1c35a", "sources": ["arxiv", "semantic_scholar"], "title": "Routing Distilled Knowledge via Mixture of LoRA Experts for Large Language Model based Bundle Generation", "abstract": "Large Language Models (LLMs) have shown potential in automatic bundle generation but suffer from prohibitive computational costs. Although knowledge distillation offers a pathway to more efficient student models, our preliminary study reveals that naively integrating diverse types of distilled knowledge from teacher LLMs into student LLMs leads to knowledge conflict, negatively impacting the performance of bundle generation. To address this, we propose RouteDK, a framework for routing distilled knowledge through a mixture of LoRA expert architecture. Specifically, we first distill knowledge from the teacher LLM for bundle generation in two complementary types: high-level knowledge (generalizable rules) and fine-grained knowledge (session-specific reasoning). We then train knowledge-specific LoRA experts for each type of knowledge together with a base LoRA expert. For effective integration, we propose a dynamic fusion module, featuring an input-aware router, where the router balances expert contributions by dynamically determining optimal weights based on input, thereby effectively mitigating knowledge conflicts. To further improve inference reliability, we design an inference-time enhancement module to reduce variance and mitigate suboptimal reasoning. Experiments on three public datasets show that our RouteDK achieves accuracy comparable to or even better than the teacher LLM, while maintaining strong computational efficiency. In addition, it outperforms state-of-the-art approaches for bundle generation.", "authors": ["Kaidong Feng", "Zhu Sun", "Hui Fang", "Jie Yang", "Wenyuan Liu", "Yew-Soon Ong"], "categories": ["cs.CL", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-24", "url": "https://arxiv.org/abs/2508.17250", "pdf_url": "https://arxiv.org/pdf/2508.17250v1", "arxiv_id": "2508.17250", "doi": "10.48550/arXiv.2508.17250", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2796}
{"id": "a7be5241cfe765939d030a842075d3d40770121f1f808883fc77bcb8bfe4664a", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-Inference-Bench: Performance Evaluation of Mixture of Expert Large Language and Vision Models", "abstract": "Mixture of Experts (MoE) models have enabled the scaling of Large Language Models (LLMs) and Vision Language Models (VLMs) by achieving massive parameter counts while maintaining computational efficiency. However, MoEs introduce several inference-time challenges, including load imbalance across experts and the additional routing computational overhead. To address these challenges and fully harness the benefits of MoE, a systematic evaluation of hardware acceleration techniques is essential. We present MoE-Inference-Bench, a comprehensive study to evaluate MoE performance across diverse scenarios. We analyze the impact of batch size, sequence length, and critical MoE hyperparameters such as FFN dimensions and number of experts on throughput. We evaluate several optimization techniques on Nvidia H100 GPUs, including pruning, Fused MoE operations, speculative decoding, quantization, and various parallelization strategies. Our evaluation includes MoEs from the Mixtral, DeepSeek, OLMoE and Qwen families. The results reveal performance differences across configurations and provide insights for the efficient deployment of MoEs.", "authors": ["Krishna Teja Chitty-Venkata", "Sylvia Howland", "Golara Azar", "Daria Soboleva", "Natalia Vassilieva", "Siddhisanket Raskar", "Murali Emani", "Venkatram Vishwanath"], "categories": ["cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-24", "url": "https://arxiv.org/abs/2508.17467", "pdf_url": "https://arxiv.org/pdf/2508.17467v1", "arxiv_id": "2508.17467", "doi": "10.1145/3731599.3767706", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2258}
{"id": "eb6ecb82e75c02ca36f201b03fa6e1fd2bad7af2d4ef1f3e9781bdd30aaf62b4", "sources": ["arxiv", "semantic_scholar"], "title": "CommonKV: Compressing KV Cache with Cross-layer Parameter Sharing", "abstract": "Large Language Models (LLMs) confront significant memory challenges due to the escalating KV cache with increasing sequence length. As a crucial technique, existing cross-layer KV cache sharing methods either necessitate modified model architectures with subsequent pre-training or incur significant performance degradation at high compression rates. To mitigate these challenges, we propose CommonKV, a training-free method for cross-layer KV cache compression through adjacent parameters sharing. Inspired by the high similarity observed in cross-layer hidden states, we utilize Singular Value Decomposition (SVD) to achieve weight sharing across adjacent parameters, resulting in a more easily mergeable latent KV cache. Furthermore, we also introduce an adaptive budget allocation strategy. It dynamically assigns compression budgets based on cosine similarity, ensuring that dissimilar caches are not over-compressed. Experiments across multiple backbone models and benchmarks including LongBench and Ruler demonstrate that the proposed method consistently outperforms existing low-rank and cross-layer approaches at various compression ratios. Moreover, we find that the benefits of CommonKV are orthogonal to other quantization and eviction methods. By integrating these approaches, we can ultimately achieve a 98\\% compression ratio without significant performance loss.", "authors": ["Yixuan Wang", "Haoyu Qiao", "Lujun Li", "Qingfu Zhu", "Wanxiang Che"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-22", "url": "https://arxiv.org/abs/2508.16134", "pdf_url": "https://arxiv.org/pdf/2508.16134v1", "arxiv_id": "2508.16134", "doi": "10.48550/arXiv.2508.16134", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2773}
{"id": "6c0dec220b933f81c418bba3429e7cf9402f6f3aed989bb9804c32af63be7416", "sources": ["arxiv", "semantic_scholar"], "title": "Anchor-MoE: A Mean-Anchored Mixture of Experts For Probabilistic Regression", "abstract": "Regression under uncertainty is fundamental across science and engineering. We present an Anchored Mixture of Experts (Anchor-MoE), a model that handles both probabilistic and point regression. For simplicity, we use a tuned gradient-boosting model to furnish the anchor mean; however, any off-the-shelf point regressor can serve as the anchor. The anchor prediction is projected into a latent space, where a learnable metric-window kernel scores locality and a soft router dispatches each sample to a small set of mixture-density-network experts; the experts produce a heteroscedastic correction and predictive variance. We train by minimizing negative log-likelihood, and on a disjoint calibration split fit a post-hoc linear map on predicted means to improve point accuracy. On the theory side, assuming a Hölder smooth regression function of order~$α$ and fixed Lipschitz partition-of-unity weights with bounded overlap, we show that Anchor-MoE attains the minimax-optimal $L^2$ risk rate $O\\!\\big(N^{-2α/(2α+d)}\\big)$. In addition, the CRPS test generalization gap scales as $\\widetilde{O}\\!\\Big(\\sqrt{(\\log(Mh)+P+K)/N}\\Big)$; it is logarithmic in $Mh$ and scales as the square root in $P$ and $K$. Under bounded-overlap routing, $K$ can be replaced by $k$, and any dependence on a latent dimension is absorbed into $P$. Under uniformly bounded means and variances, an analogous $\\widetilde{O}\\!\\big(\\sqrt{(\\log(Mh)+P+K)/N}\\big)$ scaling holds for the test NLL up to constants. Empirically, across standard UCI regressions, Anchor-MoE consistently matches or surpasses the strong NGBoost baseline in RMSE and NLL; on several datasets it achieves new state-of-the-art probabilistic regression results on our benchmark suite. Code is available at https://github.com/BaozhuoSU/Probabilistic_Regression.", "authors": ["Baozhuo Su", "Zhengxian Qu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-22", "url": "https://arxiv.org/abs/2508.16802", "pdf_url": "https://arxiv.org/pdf/2508.16802v1", "arxiv_id": "2508.16802", "doi": "10.48550/arXiv.2508.16802", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/BaozhuoSU/Probabilistic_Regression", "venue": "arXiv.org", "quality_score": 0.4285}
{"id": "56449fa655389ce8deeea04116aed3da5600c11a2fffbdda0cf1cf8d2cd58acb", "sources": ["arxiv", "semantic_scholar"], "title": "SparK: Query-Aware Unstructured Sparsity with Recoverable KV Cache Channel Pruning", "abstract": "Long-context inference in large language models (LLMs) is increasingly constrained by the KV cache bottleneck: memory usage grows linearly with sequence length, while attention computation scales quadratically. Existing approaches address this issue by compressing the KV cache along the temporal axis through strategies such as token eviction or merging to reduce memory and computational overhead. However, these methods often neglect fine-grained importance variations across feature dimensions (i.e., the channel axis), thereby limiting their ability to effectively balance efficiency and model accuracy. In reality, we observe that channel saliency varies dramatically across both queries and positions: certain feature channels carry near-zero information for a given query, while others spike in relevance. To address this oversight, we propose SPARK, a training-free plug-and-play method that applies unstructured sparsity by pruning KV at the channel level, while dynamically restoring the pruned entries during attention score computation. Notably, our approach is orthogonal to existing KV compression and quantization techniques, making it compatible for integration with them to achieve further acceleration. By reducing channel-level redundancy, SPARK enables processing of longer sequences within the same memory budget. For sequences of equal length, SPARK not only preserves or improves model accuracy but also reduces KV cache storage by over 30% compared to eviction-based methods. Furthermore, even with an aggressive pruning ratio of 80%, SPARK maintains performance with less degradation than 5% compared to the baseline eviction method, demonstrating its robustness and effectiveness. Our code will be available at https://github.com/Xnhyacinth/SparK.", "authors": ["Huanxuan Liao", "Yixing Xu", "Shizhu He", "Guanchen Li", "Xuanwu Yin", "Dong Li", "Emad Barsoum", "Jun Zhao", "Kang Liu"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-21", "url": "https://arxiv.org/abs/2508.15212", "pdf_url": "https://arxiv.org/pdf/2508.15212v3", "arxiv_id": "2508.15212", "doi": "10.48550/arXiv.2508.15212", "citation_count": 4, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Xnhyacinth/SparK", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4268}
{"id": "bdf177d9768c1a063a849cff3b79b2df42cfa9a6f53fdd0f755956709ec5e661", "sources": ["arxiv", "semantic_scholar"], "title": "StreamMem: Query-Agnostic KV Cache Memory for Streaming Video Understanding", "abstract": "Multimodal large language models (MLLMs) have made significant progress in visual-language reasoning, but their ability to efficiently handle long videos remains limited. Despite recent advances in long-context MLLMs, storing and attending to the key-value (KV) cache for long visual contexts incurs substantial memory and computational overhead. Existing visual compression methods require either encoding the entire visual context before compression or having access to the questions in advance, which is impractical for long video understanding and multi-turn conversational settings. In this work, we propose StreamMem, a query-agnostic KV cache memory mechanism for streaming video understanding. Specifically, StreamMem encodes new video frames in a streaming manner, compressing the KV cache using attention scores between visual tokens and generic query tokens, while maintaining a fixed-size KV memory to enable efficient question answering (QA) in memory-constrained, long-video scenarios. Evaluation on three long video understanding and two streaming video question answering benchmarks shows that StreamMem achieves state-of-the-art performance in query-agnostic KV cache compression and is competitive with query-aware compression approaches.", "authors": ["Yanlai Yang", "Zhuokai Zhao", "Satya Narayan Shukla", "Aashu Singh", "Shlok Kumar Mishra", "Lizhu Zhang", "Mengye Ren"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-21", "url": "https://arxiv.org/abs/2508.15717", "pdf_url": "https://arxiv.org/pdf/2508.15717v1", "arxiv_id": "2508.15717", "doi": "10.48550/arXiv.2508.15717", "citation_count": 37, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4515}
{"id": "d80c2cf45029459643dead88fbf69b6dd7f75ad5d656b72d431e835bd8acea9d", "sources": ["arxiv", "semantic_scholar"], "title": "ASDFormer: A Transformer with Mixtures of Pooling-Classifier Experts for Robust Autism Diagnosis and Biomarker Discovery", "abstract": "Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition marked by disruptions in brain connectivity. Functional MRI (fMRI) offers a non-invasive window into large-scale neural dynamics by measuring blood-oxygen-level-dependent (BOLD) signals across the brain. These signals can be modeled as interactions among Regions of Interest (ROIs), which are grouped into functional communities based on their underlying roles in brain function. Emerging evidence suggests that connectivity patterns within and between these communities are particularly sensitive to ASD-related alterations. Effectively capturing these patterns and identifying interactions that deviate from typical development is essential for improving ASD diagnosis and enabling biomarker discovery. In this work, we introduce ASDFormer, a Transformer-based architecture that incorporates a Mixture of Pooling-Classifier Experts (MoE) to capture neural signatures associated with ASD. By integrating multiple specialized expert branches with attention mechanisms, ASDFormer adaptively emphasizes different brain regions and connectivity patterns relevant to autism. This enables both improved classification performance and more interpretable identification of disorder-related biomarkers. Applied to the ABIDE dataset, ASDFormer achieves state-of-the-art diagnostic accuracy and reveals robust insights into functional connectivity disruptions linked to ASD, highlighting its potential as a tool for biomarker discovery.", "authors": ["Mohammad Izadi", "Mehran Safayani"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-19", "url": "https://arxiv.org/abs/2508.14005", "pdf_url": "https://arxiv.org/pdf/2508.14005v1", "arxiv_id": "2508.14005", "doi": "10.48550/arXiv.2508.14005", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2739}
{"id": "6fd5e8a8a126c4205de32dc3a43644fb28a1bd7d0d3451f465a4d80ec96ed567", "sources": ["arxiv", "semantic_scholar"], "title": "Maximum Score Routing For Mixture-of-Experts", "abstract": "Routing networks in sparsely activated mixture-of-experts (MoE) dynamically allocate input tokens to top-k experts through differentiable sparse transformations, enabling scalable model capacity while preserving computational efficiency. Traditional MoE networks impose an expert capacity constraint to ensure GPU-friendly computation. However, this leads to token dropping when capacity is saturated and results in low hardware efficiency due to padding in underutilized experts. Removing the capacity constraint, in turn, compromises load balancing and computational efficiency. To address these issues, we propose Maximum Score Routing ($\\mathbf{MaxScore}$), a novel MoE routing paradigm that models routing as a minimum-cost maximum-flow problem and integrates a SoftTopk operator. MaxScore resolves the fundamental limitations of iterative rerouting and optimal transport formulations, achieving lower training losses and higher evaluation scores at equivalent FLOPs compared to both constrained and unconstrained baselines. Implementation details and experimental configurations can be obtained from $\\href{https://github.com/dongbw18/MaxScore.git}{MaxScore}$.", "authors": ["Bowen Dong", "Yilong Fan", "Yutao Sun", "Zhenyu Li", "Tengyu Pan", "Xun Zhou", "Jianyong Wang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-18", "url": "https://arxiv.org/abs/2508.12801", "pdf_url": "https://arxiv.org/pdf/2508.12801v1", "arxiv_id": "2508.12801", "doi": "10.18653/v1/2025.findings-acl.653", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/dongbw18/MaxScore.git}{MaxScore}$", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4215}
{"id": "b24cfff988aec0dcfc52f83397b6b6e23e87d6d40819f7cb813209fb45923abf", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerating LLM Inference via Dynamic KV Cache Placement in Heterogeneous Memory System", "abstract": "Large Language Model (LLM) inference is increasingly constrained by memory bandwidth, with frequent access to the key-value (KV) cache dominating data movement. While attention sparsity reduces some memory traffic, the relevance of past tokens varies over time, requiring the full KV cache to remain accessible and sustaining pressure on both bandwidth and capacity. With advances in interconnects such as NVLink and LPDDR5X, modern AI hardware now integrates high-bandwidth memory (HBM) with high-speed off-package DRAM, making heterogeneous memory systems a practical solution. This work investigates dynamic KV cache placement across such systems to maximize aggregated bandwidth utilization under capacity constraints. Rather than proposing a specific scheduling policy, we formulate the placement problem mathematically and derive a theoretical upper bound, revealing substantial headroom for runtime optimization. To our knowledge, this is the first formal treatment of dynamic KV cache scheduling in heterogeneous memory systems for LLM inference.", "authors": ["Yunhua Fang", "Rui Xie", "Asad Ul Haq", "Linsen Ma", "Kaoutar El Maghraoui", "Naigang Wang", "Meng Wang", "Liu Liu", "Tong Zhang"], "categories": ["cs.AR", "cs.AI", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-17", "url": "https://arxiv.org/abs/2508.13231", "pdf_url": "https://arxiv.org/pdf/2508.13231v2", "arxiv_id": "2508.13231", "doi": "10.1109/LCA.2025.3622724", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE computer architecture letters", "quality_score": 0.2716}
{"id": "1e0b060baa11c88100b826231ad9da03940af6159918b15c218c8404888ea020", "sources": ["arxiv", "semantic_scholar"], "title": "STM3: Mixture of Multiscale Mamba for Long-Term Spatio-Temporal Time-Series Prediction", "abstract": "Recently, spatio-temporal time-series prediction has developed rapidly, yet existing deep learning methods struggle with learning complex long-term spatio-temporal dependencies efficiently. The long-term spatio-temporal dependency learning brings two new challenges: 1) The long-term temporal sequence naturally includes multiscale information, which is hard to extract efficiently; 2) The multiscale temporal information from different nodes is highly correlated and hard to model. To address these challenges, we propose Spatio-Temporal Mixture of Multiscale Mamba (STM3). STM3 integrates a Multiscale Mamba architecture within a novel Disentangled Mixture-of-Experts (DMoE) framework to capture diverse multiscale information efficiently, while utilizing an adaptive graph causal network to model complex spatial dependencies. To ensure robust representation learning, we introduce a stable routing strategy and a causal contrastive learning strategy, which work in tandem with hierarchical information aggregation to guarantee scale distinguishability. We theoretically prove that STM3 achieves superior routing smoothness and guarantees pattern disentanglement for each expert. Extensive experiments on 10 real-world benchmarks across domains demonstrate STM3's superior performance, achieving state-of-the-art results in long-term spatio-temporal time-series prediction. Notably, on the PEMSD8 dataset, it achieves significant improvements, surpassing the second-best model by 7.1% in MAE, 8.5% in RMSE, and 15.9% in MAPE. Code is available at https://github.com/IfReasonable/STM3_KDD26.", "authors": ["Haolong Chen", "Liang Zhang", "Zhengyuan Xin", "Guangxu Zhu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-17", "url": "https://arxiv.org/abs/2508.12247", "pdf_url": "https://arxiv.org/pdf/2508.12247v3", "arxiv_id": "2508.12247", "doi": "10.48550/arXiv.2508.12247", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/IfReasonable/STM3_KDD26", "venue": "arXiv.org", "quality_score": 0.4197}
{"id": "2ece3ce47203461f55daf9c14f8e731276a8d37c96dba0eb9c09b04523e390cd", "sources": ["arxiv", "semantic_scholar"], "title": "SurfaceLogicKV: Surface and Logic Attention Behaviors are All You Need for Robust KV Cache Compression", "abstract": "The increasing input sequence length in Large Language Models (LLMs) puts significant pressure on key-value (KV) cache storage, making efficient inference challenging. Explicitly distinguishing attention behavior into our self-defined surface memorization and logic construction reveals essential roles in long-context reasoning. We observe that an individual attention head can display various behaviors, with nearly 98.5% effectively ignoring completely irrelevant information. The remaining 1.5% behaves as logic construction, and 0.5% behaves as surface memorization. Based on layer- and head-wise integration, we propose a novel two-stage SurfaceLogicKV method to utilize these attention behaviors for KV Cache compression. As a result, it achieves improved compressing robustness while maintaining competitive performance across various tasks and long sequences compared to baselines or even FullKV in some specific situations", "authors": ["Mengjie Li", "William J. Song"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-14", "url": "https://arxiv.org/abs/2508.15806", "pdf_url": "https://arxiv.org/pdf/2508.15806v1", "arxiv_id": "2508.15806", "doi": "10.48550/arXiv.2508.15806", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2681}
{"id": "1e6976bb85536388f65ec74508ca8755ba312b045a3600463e94f910bd0de0b5", "sources": ["arxiv", "semantic_scholar"], "title": "XQuant: Breaking the Memory Wall for LLM Inference with KV Cache Rematerialization", "abstract": "Although LLM inference has emerged as a critical workload for many downstream applications, efficiently inferring LLMs is challenging due to the substantial memory footprint and bandwidth requirements. In parallel, compute capabilities have steadily outpaced both memory capacity and bandwidth over the last few decades, a trend that remains evident in modern GPU hardware and exacerbates the challenge of LLM inference. As such, new algorithms are emerging that trade increased computation for reduced memory operations. To that end, we present XQuant, which takes advantage of this trend, enabling an order-of-magnitude reduction in memory consumption through low-bit quantization with substantial accuracy benefits relative to state-of-the-art KV cache quantization methods. We accomplish this by quantizing and caching the layer input activations X, instead of using standard KV caching, and then rematerializing the Keys and Values on-the-fly during inference. This results in an immediate 2$\\times$ memory savings compared to KV caching. By applying XQuant, we achieve up to $\\sim 7.7\\times$ memory savings with $<0.1$ perplexity degradation compared to the FP16 baseline. Furthermore, our approach leverages the fact that X values are similar across layers. Building on this observation, we introduce XQuant-CL, which exploits the cross-layer similarity in the X embeddings for extreme compression. Across different models, XQuant-CL attains up to 10$\\times$ memory savings relative to the FP16 baseline with only 0.01 perplexity degradation, and 12.5$\\times$ memory savings with only $0.1$ perplexity degradation. XQuant exploits the rapidly increasing compute capabilities of hardware platforms to eliminate the memory bottleneck, while surpassing state-of-the-art KV cache quantization methods and achieving near-FP16 accuracy across a wide range of models.", "authors": ["Aditya Tomar", "Coleman Hooper", "Minjae Lee", "Haocheng Xi", "Rishabh Tiwari", "Wonjun Kang", "Luca Manolache", "Michael W. Mahoney", "Kurt Keutzer", "Amir Gholami"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-14", "url": "https://arxiv.org/abs/2508.10395", "pdf_url": "https://arxiv.org/pdf/2508.10395v1", "arxiv_id": "2508.10395", "doi": "10.48550/arXiv.2508.10395", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2681}
{"id": "8346bfd82ef12572bb435d18ed01bb26f8148294dc58eeb1dff887c9a5b5c14d", "sources": ["arxiv", "semantic_scholar"], "title": "Quantization vs Pruning: Insights from the Strong Lottery Ticket Hypothesis", "abstract": "Quantization is an essential technique for making neural networks more efficient, yet our theoretical understanding of it remains limited. Previous works demonstrated that extremely low-precision networks, such as binary networks, can be constructed by pruning large, randomly-initialized networks, and showed that the ratio between the size of the original and the pruned networks is at most polylogarithmic. The specific pruning method they employed inspired a line of theoretical work known as the Strong Lottery Ticket Hypothesis (SLTH), which leverages insights from the Random Subset Sum Problem. However, these results primarily address the continuous setting and cannot be applied to extend SLTH results to the quantized setting. In this work, we build on foundational results by Borgs et al. on the Number Partitioning Problem to derive new theoretical results for the Random Subset Sum Problem in a quantized setting. Using these results, we then extend the SLTH framework to finite-precision networks. While prior work on SLTH showed that pruning allows approximation of a certain class of neural networks, we demonstrate that, in the quantized setting, the analogous class of target discrete neural networks can be represented exactly, and we prove optimal bounds on the necessary overparameterization of the initial network as a function of the precision of the target network.", "authors": ["Aakash Kumar", "Emanuele Natale"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-14", "url": "https://arxiv.org/abs/2508.11020", "pdf_url": "https://arxiv.org/pdf/2508.11020v1", "arxiv_id": "2508.11020", "doi": "10.48550/arXiv.2508.11020", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2681}
{"id": "f09612ed424658a8575eca1ae9e0a3b36f16e7293b01a2de196c5ebfc47116c6", "sources": ["arxiv", "semantic_scholar"], "title": "Federated Learning Over LoRa Networks: Simulator Design and Performance Evaluation", "abstract": "Federated learning (FL) over long-range (LoRa) low-power wide area networks faces unique challenges due to limited bandwidth, interference, and strict duty-cycle constraints. We develop a Python-based simulator that integrates and extends the Flower and LoRaSim frameworks to evaluate centralized FL over LoRa networks. The simulator employs a detailed link-level model for FL update transfer over LoRa channels, capturing LoRa's receiver sensitivity, interference characteristics, block-fading effects, and constraints on the maximum transmission unit. It supports update sparsification, quantization, compression, forward frame-erasure correction (FEC), and duty cycling. Numerical results illustrate the impact of transmission parameters (spreading factor, FEC rate) and interference on FL performance. Demonstrating the critical role of FEC in enabling FL over LoRa networks, we perform an in-depth evaluation of the impact of FEC on FL convergence and device airtime, providing insights for communication protocol design for FL over LoRa networks.", "authors": ["Anshika Singh", "Siddhartha S. Borkotoky"], "categories": ["cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-14", "url": "https://arxiv.org/abs/2508.10574", "pdf_url": "https://arxiv.org/pdf/2508.10574v1", "arxiv_id": "2508.10574", "doi": "10.48550/arXiv.2508.10574", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2681}
{"id": "8af38fd7ee21f775e8df3b1b45d1eb44dc601f0ee3984f9d85b588f4315d11b8", "sources": ["arxiv", "semantic_scholar"], "title": "Shadow in the Cache: Unveiling and Mitigating Privacy Risks of KV-cache in LLM Inference", "abstract": "The Key-Value (KV) cache, which stores intermediate attention computations (Key and Value pairs) to avoid redundant calculations, is a fundamental mechanism for accelerating Large Language Model (LLM) inference. However, this efficiency optimization introduces significant yet underexplored privacy risks. This paper provides the first comprehensive analysis of these vulnerabilities, demonstrating that an attacker can reconstruct sensitive user inputs directly from the KV-cache. We design and implement three distinct attack vectors: a direct Inversion Attack, a more broadly applicable and potent Collision Attack, and a semantic-based Injection Attack. These methods demonstrate the practicality and severity of KV-cache privacy leakage issues. To mitigate this, we propose KV-Cloak, a novel, lightweight, and efficient defense mechanism. KV-Cloak uses a reversible matrix-based obfuscation scheme, combined with operator fusion, to secure the KV-cache. Our extensive experiments show that KV-Cloak effectively thwarts all proposed attacks, reducing reconstruction quality to random noise. Crucially, it achieves this robust security with virtually no degradation in model accuracy and minimal performance overhead, offering a practical solution for trustworthy LLM deployment.", "authors": ["Zhifan Luo", "Shuo Shao", "Su Zhang", "Lijing Zhou", "Yuke Hu", "Chenxu Zhao", "Zhihao Liu", "Zhan Qin"], "categories": ["cs.CR", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-13", "url": "https://arxiv.org/abs/2508.09442", "pdf_url": "https://arxiv.org/pdf/2508.09442v4", "arxiv_id": "2508.09442", "doi": "10.48550/arXiv.2508.09442", "citation_count": 19, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SiO-2/kvcloak", "venue": "Published in the Proceedings of the 33rd Network and Distributed System Security Symposium (NDSS 2026)", "quality_score": 0.4126}
{"id": "a53393b02c06a95e03f886eb7649d570e276105dfbd5647150c72f824bb5811c", "sources": ["arxiv", "semantic_scholar"], "title": "MoIIE: Mixture of Intra- and Inter-Modality Experts for Large Vision Language Models", "abstract": "Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across multi-modal tasks by scaling model size and training data. However, these dense LVLMs incur significant computational costs and motivate the exploration of sparse Mixture of Experts (MoE) architectures. While MoE improve parameter efficiency, effectively applying MoE to simultaneously model modality-specific features and cross-modal associations in LVLMs remains challenging. In this work, we propose to incorporate Mixture of Intra- and Inter-Modality Experts (MoIIE) to LVLMs. For each token, expert routing is guided by its modality, directing tokens to their respective intra-modality experts as well as a shared pool of inter-modality experts, enabling the model to jointly learn rich intra-modal features and cross-modal interactions. We further introduce an effective and straightforward two-stage training strategy, which facilitates the direct activation of both MoE and multi-modal capabilities. Extensive experiments across different data scales and LLM backbone demonstrate the effectiveness, efficiency and generality of our approach. Notably, our MoIIE models with 5.5B and 11.3B activated parameters match or even surpass the performance of existing advanced open-source MoE-LLMs based multi-modal models that involve more activated parameters. The code is available at https://github.com/AlenjandroWang/MoIIE.", "authors": ["Dianyi Wang", "Siyuan Wang", "Zejun Li", "Yikun Wang", "Yitong Li", "Duyu Tang", "Xiaoyu Shen", "Xuanjing Huang", "Zhongyu Wei"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-13", "url": "https://arxiv.org/abs/2508.09779", "pdf_url": "https://arxiv.org/pdf/2508.09779v2", "arxiv_id": "2508.09779", "doi": "10.48550/arXiv.2508.09779", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AlenjandroWang/MoIIE", "venue": "arXiv.org", "quality_score": 0.4126}
{"id": "9a92690921a863ebb3c3240688ef364bff2a0de2a34c7238a2d1ed5a8e3bf1c1", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Mixture-of-Experts for Incremental Graph Learning", "abstract": "Graph incremental learning is a learning paradigm that aims to adapt trained models to continuously incremented graphs and data over time without the need for retraining on the full dataset. However, regular graph machine learning methods suffer from catastrophic forgetting when applied to incremental learning settings, where previously learned knowledge is overridden by new knowledge. Previous approaches have tried to address this by treating the previously trained model as an inseparable unit and using techniques to maintain old behaviors while learning new knowledge. These approaches, however, do not account for the fact that previously acquired knowledge at different timestamps contributes differently to learning new tasks. Some prior patterns can be transferred to help learn new data, while others may deviate from the new data distribution and be detrimental. To address this, we propose a dynamic mixture-of-experts (DyMoE) approach for incremental learning. Specifically, a DyMoE GNN layer adds new expert networks specialized in modeling the incoming data blocks. We design a customized regularization loss that utilizes data sequence information so existing experts can maintain their ability to solve old tasks while helping the new expert learn the new data effectively. As the number of data blocks grows over time, the computational cost of the full mixture-of-experts (MoE) model increases. To address this, we introduce a sparse MoE approach, where only the top-$k$ most relevant experts make predictions, significantly reducing the computation time. Our model achieved 4.92\\% relative accuracy increase compared to the best baselines on class incremental learning, showing the model's exceptional power.", "authors": ["Lecheng Kong", "Theodore Vasiloudis", "Seongjun Yun", "Han Xie", "Xiang Song"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-13", "url": "https://arxiv.org/abs/2508.09974", "pdf_url": "https://arxiv.org/pdf/2508.09974v1", "arxiv_id": "2508.09974", "doi": "10.48550/arXiv.2508.09974", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.267}
{"id": "aa2e88be1e31dbc409d79b4e86dff5bf93c4f511d6a620b80e12e1b5bea1db3e", "sources": ["arxiv", "semantic_scholar"], "title": "$μ$-Parametrization for Mixture of Experts", "abstract": "Recent years have seen a growing interest and adoption of LLMs, with Mixture-of-Experts (MoE) emerging as a leading architecture in extremely large models. Currently, the largest open-source models reach over $1$T parameters. At such scales, hyperparameter tuning becomes prohibitively expensive. Precisely for this reason, the $μ$Transfer is becoming a key technique. It allows for seamless transfer of optimal hyperparameters across model scales, resulting in a huge reduction in tuning costs. However, existing work has primarily focused on dense LLMs, leaving MoE architectures unexplored. In this work, we derive a $μ$-Parameterization for MoE, providing theoretical guarantees for feature learning across model widths. Our experiments demonstrate that the optimal learning rate reliably transfers across model sizes, establishing a foundation for efficient hyperparameter tuning in large-scale MoE models.", "authors": ["Jan Małaśnicki", "Kamil Ciebiera", "Mateusz Boruń", "Maciej Pióro", "Jan Ludziejewski", "Maciej Stefaniak", "Michał Krutul", "Sebastian Jaszczur", "Marek Cygan", "Kamil Adamczewski", "Jakub Krajewski"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-13", "url": "https://arxiv.org/abs/2508.09752", "pdf_url": "https://arxiv.org/pdf/2508.09752v2", "arxiv_id": "2508.09752", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.3155}
{"id": "35b629d490a1bbf6aaa0a71fa8c36a4b49703e334374e771aa7ddf629755a037", "sources": ["arxiv", "semantic_scholar"], "title": "Fake-Mamba: Real-Time Speech Deepfake Detection Using Bidirectional Mamba as Self-Attention's Alternative", "abstract": "Advances in speech synthesis intensify security threats, motivating real-time deepfake detection research. We investigate whether bidirectional Mamba can serve as a competitive alternative to Self-Attention in detecting synthetic speech. Our solution, Fake-Mamba, integrates an XLSR front-end with bidirectional Mamba to capture both local and global artifacts. Our core innovation introduces three efficient encoders: TransBiMamba, ConBiMamba, and PN-BiMamba. Leveraging XLSR's rich linguistic representations, PN-BiMamba can effectively capture the subtle cues of synthetic speech. Evaluated on ASVspoof 21 LA, 21 DF, and In-The-Wild benchmarks, Fake-Mamba achieves 0.97%, 1.74%, and 5.85% EER, respectively, representing substantial relative gains over SOTA models XLSR-Conformer and XLSR-Mamba. The framework maintains real-time inference across utterance lengths, demonstrating strong generalization and practical viability. The code is available at https://github.com/xuanxixi/Fake-Mamba.", "authors": ["Xi Xuan", "Zimo Zhu", "Wenxin Zhang", "Yi-Cheng Lin", "Tomi Kinnunen"], "categories": ["eess.AS", "cs.AI", "cs.CL", "cs.LG", "eess.SY"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.09294", "pdf_url": "https://arxiv.org/pdf/2508.09294v1", "arxiv_id": "2508.09294", "doi": "10.1109/ASRU65441.2025.11434679", "citation_count": 7, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/xuanxixi/Fake-Mamba", "venue": "Automatic Speech Recognition & Understanding", "quality_score": 0.4108}
{"id": "851b61c33ebc08e38da9b3d9527ce74a44ff455b0503d0fea43e1503cfc64d2a", "sources": ["arxiv", "semantic_scholar"], "title": "TiMoE: Time-Aware Mixture of Language Experts", "abstract": "Large language models (LLMs) are typically trained on fixed snapshots of the web, which means that their knowledge becomes stale and their predictions risk temporal leakage: relying on information that lies in the future relative to a query. We tackle this problem by pre-training from scratch a set of GPT-style experts on disjoint two-year slices of a 2013-2024 corpus and combining them through TiMoE, a Time-aware Mixture of Language Experts. At inference time, TiMoE masks all experts whose training window ends after the query timestamp and merges the remaining log-probabilities in a shared space, guaranteeing strict causal validity while retaining the breadth of multi-period knowledge. We also release TSQA, a 10k-question benchmark whose alternatives are explicitly labelled as past, future or irrelevant, allowing fine-grained measurement of temporal hallucinations. Experiments on eight standard NLP tasks plus TSQA show that a co-adapted TiMoE variant matches or exceeds the best single-period expert and cuts future-knowledge errors by up to 15%. Our results demonstrate that modular, time-segmented pre-training paired with causal routing is a simple yet effective path toward LLMs that stay chronologically grounded without sacrificing general performance much. We open source our code at TiMoE (Github): https://github.com/epfml/TiMoE", "authors": ["Robin Faro", "Dongyang Fan", "Tamar Alphaidze", "Martin Jaggi"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.08827", "pdf_url": "https://arxiv.org/pdf/2508.08827v1", "arxiv_id": "2508.08827", "doi": "10.48550/arXiv.2508.08827", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/epfml/TiMoE", "venue": "arXiv.org", "quality_score": 0.4108}
{"id": "6c33a8e9ca12c83275fd796a28b830682e4ecf1e0de99f8161c9f492fcc70e26", "sources": ["arxiv", "semantic_scholar"], "title": "Wavelet Mixture of Experts for Time Series Forecasting", "abstract": "The field of time series forecasting is rapidly advancing, with recent large-scale Transformers and lightweight Multilayer Perceptron (MLP) models showing strong predictive performance. However, conventional Transformer models are often hindered by their large number of parameters and their limited ability to capture non-stationary features in data through smoothing. Similarly, MLP models struggle to manage multi-channel dependencies effectively. To address these limitations, we propose a novel, lightweight time series prediction model, WaveTS-B. This model combines wavelet transforms with MLP to capture both periodic and non-stationary characteristics of data in the wavelet domain. Building on this foundation, we propose a channel clustering strategy that incorporates a Mixture of Experts (MoE) framework, utilizing a gating mechanism and expert network to handle multi-channel dependencies efficiently. We propose WaveTS-M, an advanced model tailored for multi-channel time series prediction. Empirical evaluation across eight real-world time series datasets demonstrates that our WaveTS series models achieve state-of-the-art (SOTA) performance with significantly fewer parameters. Notably, WaveTS-M shows substantial improvements on multi-channel datasets, highlighting its effectiveness.", "authors": ["Zheng Zhou", "Yu-Jie Xiong", "Jia-Chen Zhang", "Chun-Ming Xia", "Xi-Jiong Xie"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.08825", "pdf_url": "https://arxiv.org/pdf/2508.08825v1", "arxiv_id": "2508.08825", "doi": "10.48550/arXiv.2508.08825", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2658}
{"id": "717eb56adf17349c1ad94524dadd2eb577ca14730ae006a564251a8cecfde42d", "sources": ["arxiv", "semantic_scholar"], "title": "TARA: Token-Aware LoRA for Composable Personalization in Diffusion Models", "abstract": "Personalized text-to-image generation aims to synthesize novel images of a specific subject or style using only a few reference images. Recent methods based on Low-Rank Adaptation (LoRA) enable efficient single-concept customization by injecting lightweight, concept-specific adapters into pre-trained diffusion models. However, combining multiple LoRA modules for multi-concept generation often leads to identity missing and visual feature leakage. In this work, we identify two key issues behind these failures: (1) token-wise interference among different LoRA modules, and (2) spatial misalignment between the attention map of a rare token and its corresponding concept-specific region. To address these issues, we propose Token-Aware LoRA (TARA), which introduces a token mask to explicitly constrain each module to focus on its associated rare token to avoid interference, and a training objective that encourages the spatial attention of a rare token to align with its concept region. Our method enables training-free multi-concept composition by directly injecting multiple independently trained TARA modules at inference time. Experimental results demonstrate that TARA enables efficient multi-concept inference and effectively preserving the visual identity of each concept by avoiding mutual interference between LoRA modules. The code and models are available at https://github.com/YuqiPeng77/TARA.", "authors": ["Yuqi Peng", "Lingtao Zheng", "Yufeng Yang", "Yi Huang", "Mingfu Yan", "Jianzhuang Liu", "Shifeng Chen"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.08812", "pdf_url": "https://arxiv.org/pdf/2508.08812v1", "arxiv_id": "2508.08812", "doi": "10.48550/arXiv.2508.08812", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/YuqiPeng77/TARA", "venue": null, "quality_score": 0.3142}
{"id": "4b598917afaaa43658a2025ae19e2de9b06d421c925e55c8e51e854cbc3ba1d1", "sources": ["arxiv", "semantic_scholar"], "title": "Selective KV-Cache Sharing to Mitigate Timing Side-Channels in LLM Inference", "abstract": "Global KV-cache sharing is an effective optimization for accelerating large language model (LLM) inference, yet it introduces an API-visible timing side channel that lets adversaries infer sensitive user inputs from shared entries, leading to cross-tenant privacy risks. To address this problem, we introduce SafeKV (Secure and Flexible KV-cache Sharing), a system-level co-design of privacy enforcement and KV-cache management. SafeKV integrates lightweight detection and isolation directly into the serving runtime to eliminate cross-tenant reuse of sensitive KV-cache blocks under our threat model, while recovering most of the performance benefits of global sharing. Our key contributions are: (1) a three-tier asynchronous detection pipeline that decouples privacy classification from inference and supports streaming workloads, (2) a unified radix-tree-based memory manager with path compression and sensitivity-aware eviction for scalable selective isolation, and (3) an RDR-guided (Reuse Diversity Ratio) runtime safeguard that detects and bounds residual leakage. On large LLM backends, SafeKV reduces the time-to-first-token (TTFT) overhead compared to full isolation by up to 40.58% and raises throughput by up to 2.66x. Overall, SafeKV restores the efficiency of KV reuse while enforcing strong, practical privacy for multi-tenant LLM inference.", "authors": ["Kexin Chu", "Zecheng Lin", "Dawei Xiang", "Zixu Shen", "Jianchang Su", "Cheng Chu", "Yiwei Yang", "Wenhui Zhang", "Wenfei Wu", "Wei Zhang"], "categories": ["cs.CR", "cs.LG", "cs.OS"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-11", "url": "https://arxiv.org/abs/2508.08438", "pdf_url": "https://arxiv.org/pdf/2508.08438v2", "arxiv_id": "2508.08438", "doi": "10.48550/arXiv.2508.08438", "citation_count": 15, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.301}
{"id": "7cc790e5a28ae4fe5d40149e9e8863c6853ee8db97dcdfe3b1aee5c7a8a0977f", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-modal Adaptive Mixture of Experts for Cold-start Recommendation", "abstract": "Recommendation systems have faced significant challenges in cold-start scenarios, where new items with a limited history of interaction need to be effectively recommended to users. Though multimodal data (e.g., images, text, audio, etc.) offer rich information to address this issue, existing approaches often employ simplistic integration methods such as concatenation, average pooling, or fixed weighting schemes, which fail to capture the complex relationships between modalities. Our study proposes a novel Mixture of Experts (MoE) framework for multimodal cold-start recommendation, named MAMEX, which dynamically leverages latent representation from different modalities. MAMEX utilizes modality-specific expert networks and introduces a learnable gating mechanism that adaptively weights the contribution of each modality based on its content characteristics. This approach enables MAMEX to emphasize the most informative modalities for each item while maintaining robustness when certain modalities are less relevant or missing. Extensive experiments on benchmark datasets show that MAMEX outperforms state-of-the-art methods in cold-start scenarios, with superior accuracy and adaptability. For reproducibility, the code has been made available on Github https://github.com/L2R-UET/MAMEX.", "authors": ["Van-Khang Nguyen", "Duc-Hoang Pham", "Huy-Son Nguyen", "Cam-Van Thi Nguyen", "Hoang-Quynh Le", "Duc-Trong Le"], "categories": ["cs.IR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-11", "url": "https://arxiv.org/abs/2508.08042", "pdf_url": "https://arxiv.org/pdf/2508.08042v1", "arxiv_id": "2508.08042", "doi": "10.1145/3746252.3760837", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/L2R-UET/MAMEX", "venue": "International Conference on Information and Knowledge Management", "quality_score": 0.4091}
{"id": "d4c8e892a8537dbc4c4f427a65c7e914e7c0c679f81812910d9492d2dc06bfcb", "sources": ["arxiv", "semantic_scholar"], "title": "MoQE: Improve Quantization Model performance via Mixture of Quantization Experts", "abstract": "Quantization method plays a crucial role in improving model efficiency and reducing deployment costs, enabling the widespread application of deep learning models on resource-constrained devices. However, the quantization process inevitably introduces accuracy degradation. In this paper, we propose Mixture of Quantization Experts( abbr. MoQE), a quantization inference framework based on the Mixture-of-Experts (MoE) architecture, aiming to jointly improve the performance of quantization models. MoQE combines multiple quantization variants of one full-precision model as specialized \"quantization experts\" and dynamically routes input data to the most suitable expert based on its characteristics. MoQE alleviates the performance degradation commonly seen in single quantization models through specialization quantization expert models. We design lightweight, structure-aware router models tailored for both CV and NLP tasks. Experimental evaluations on ResNet, LLaMA, and Qwen model families across benchmark datasets including ImageNet, WikiText, C4, and OpenWebText demonstrate that MoQE achieves performance comparable to SOTA quantization model, without incurring significant increases in inference latency.", "authors": ["Jinhao Zhang", "Yunquan Zhang", "Boyang Zhang", "Zeyu Liu", "Daning Cheng"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-09", "url": "https://arxiv.org/abs/2508.09204", "pdf_url": "https://arxiv.org/pdf/2508.09204v2", "arxiv_id": "2508.09204", "doi": "10.48550/arXiv.2508.09204", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2624}
{"id": "407ba9bbe73809721fa40cddb0e5ed6babf03a75e235e456a7cc9c28ee18b780", "sources": ["arxiv", "semantic_scholar"], "title": "SSD Offloading for LLM Mixture-of-Experts Weights Considered Harmful in Energy Efficiency", "abstract": "Large Language Models (LLMs) applying Mixture-of-Experts (MoE) scale to trillions of parameters but require vast memory, motivating a line of research to offload expert weights from fast-but-small DRAM (HBM) to denser Flash SSDs. While SSDs provide cost-effective capacity, their read energy per bit is substantially higher than that of DRAM. This paper quantitatively analyzes the energy implications of offloading MoE expert weights to SSDs during the critical decode stage of LLM inference. Our analysis, comparing SSD, CPU memory (DDR), and HBM storage scenarios for models like DeepSeek-R1, reveals that offloading MoE weights to current SSDs drastically increases per-token-generation energy consumption (e.g., by up to ~12x compared to the HBM baseline), dominating the total inference energy budget. Although techniques like prefetching effectively hide access latency, they cannot mitigate this fundamental energy penalty. We further explore future technological scaling, finding that the inherent sparsity of MoE models could potentially make SSDs energy-viable if Flash read energy improves significantly, roughly by an order of magnitude.", "authors": ["Kwanhee Kyung", "Sungmin Yun", "Jung Ho Ahn"], "categories": ["cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-09", "url": "https://arxiv.org/abs/2508.06978", "pdf_url": "https://arxiv.org/pdf/2508.06978v1", "arxiv_id": "2508.06978", "doi": "10.1109/LCA.2025.3592563", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE computer architecture letters", "quality_score": 0.2624}
{"id": "d3cd8fd2425ac79b41ce6ed439368c4552e2d6d4a87c7c64556d472a0efeeba3", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Multimodal Streaming Recommendation via Expandable Side Mixture-of-Experts", "abstract": "Streaming recommender systems (SRSs) are widely deployed in real-world applications, where user interests shift and new items arrive over time. As a result, effectively capturing users' latest preferences is challenging, as interactions reflecting recent interests are limited and new items often lack sufficient feedback. A common solution is to enrich item representations using multimodal encoders (e.g., BERT or ViT) to extract visual and textual features. However, these encoders are pretrained on general-purpose tasks: they are not tailored to user preference modeling, and they overlook the fact that user tastes toward modality-specific features such as visual styles and textual tones can also drift over time. This presents two key challenges in streaming scenarios: the high cost of fine-tuning large multimodal encoders, and the risk of forgetting long-term user preferences due to continuous model updates. To tackle these challenges, we propose Expandable Side Mixture-of-Experts (XSMoE), a memory-efficient framework for multimodal streaming recommendation. XSMoE attaches lightweight side-tuning modules consisting of expandable expert networks to frozen pretrained encoders and incrementally expands them in response to evolving user feedback. A gating router dynamically combines expert and backbone outputs, while a utilization-based pruning strategy maintains model compactness. By learning new patterns through expandable experts without overwriting previously acquired knowledge, XSMoE effectively captures both cold start and shifting preferences in multimodal features. Experiments on three real-world datasets demonstrate that XSMoE outperforms state-of-the-art baselines in both recommendation quality and computational efficiency.", "authors": ["Yunke Qu", "Liang Qu", "Tong Chen", "Quoc Viet Hung Nguyen", "Hongzhi Yin"], "categories": ["cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-08", "url": "https://arxiv.org/abs/2508.05993", "pdf_url": "https://arxiv.org/pdf/2508.05993v3", "arxiv_id": "2508.05993", "doi": "10.1145/3746252.3761390", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/qykcq/Efficient-Multimodal-Streaming-Recommendation-via-Expandable-Side-Mixture-of-Experts", "venue": "International Conference on Information and Knowledge Management", "quality_score": 0.4037}
{"id": "0a5325d41ea94ad4c8633e7005f66b884dc69b15321dbc613f4c9d863a9927e7", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache Compression for Inference Efficiency in LLMs: A Review", "abstract": "Withtherapid advancement of large language models (LLMs), the context length for inference has been continuously increasing, leading to an exponential growth in the demand for Key-Value (KV) caching. This has resulted in a significant memory bottleneck, limiting the inference efficiency and scalability of the models. Therefore, optimizing the KV cache during inference is crucial for enhancing performance and efficiency. This review systematically examines current KV cache optimization techniques, including compression strategies such as selective token strategies, quantization, and attention compression. We evaluate the effectiveness, trade-offs, and application scenarios of these methods, providing a comprehensive analysis of their impact on memory usage and inference speed. We focus on identifying the limitations and challenges of existing methods, such as compatibility issues with different models and tasks. Additionally, this review highlights future research directions, including hybrid optimization techniques, adaptive dynamic strategies, and software-hardware co-design. These approaches aim to improve inference efficiency and promote the practical application of large language models.", "authors": ["Yanyu Liu", "Jingying Fu", "Sixiang Liu", "Yitian Zou", "You Fu", "Jiehan Zhou", "Shouhua Zhang"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-08", "url": "https://arxiv.org/abs/2508.06297", "pdf_url": "https://arxiv.org/pdf/2508.06297v1", "arxiv_id": "2508.06297", "doi": "10.1145/3778534.3778567", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2258}
{"id": "b923cff857d37e4d1338ce963d24eb18ede34e2aca19c375f8cbd54545428e4d", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts Guided by Gaussian Splatters Matters: A new Approach to Weakly-Supervised Video Anomaly Detection", "abstract": "Video Anomaly Detection (VAD) is a challenging task due to the variability of anomalous events and the limited availability of labeled data. Under the Weakly-Supervised VAD (WSVAD) paradigm, only video-level labels are provided during training, while predictions are made at the frame level. Although state-of-the-art models perform well on simple anomalies (e.g., explosions), they struggle with complex real-world events (e.g., shoplifting). This difficulty stems from two key issues: (1) the inability of current models to address the diversity of anomaly types, as they process all categories with a shared model, overlooking category-specific features; and (2) the weak supervision signal, which lacks precise temporal information, limiting the ability to capture nuanced anomalous patterns blended with normal events. To address these challenges, we propose Gaussian Splatting-guided Mixture of Experts (GS-MoE), a novel framework that employs a set of expert models, each specialized in capturing specific anomaly types. These experts are guided by a temporal Gaussian splatting loss, enabling the model to leverage temporal consistency and enhance weak supervision. The Gaussian splatting approach encourages a more precise and comprehensive representation of anomalies by focusing on temporal segments most likely to contain abnormal events. The predictions from these specialized experts are integrated through a mixture-of-experts mechanism to model complex relationships across diverse anomaly patterns. Our approach achieves state-of-the-art performance, with a 91.58% AUC on the UCF-Crime dataset, and demonstrates superior results on XD-Violence and MSAD datasets. By leveraging category-specific expertise and temporal guidance, GS-MoE sets a new benchmark for VAD under weak supervision.", "authors": ["Giacomo D'Amicantonio", "Snehashis Majhi", "Quan Kong", "Lorenzo Garattoni", "Gianpiero Francesca", "François Bremond", "Egor Bondarev"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-08", "url": "https://arxiv.org/abs/2508.06318", "pdf_url": "https://arxiv.org/pdf/2508.06318v1", "arxiv_id": "2508.06318", "doi": "10.1109/ICCV51701.2025.00957", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.2612}
{"id": "c98e65c99ed275a5b33012e9d7242440fd611b931429405bacea2da9fce28231", "sources": ["arxiv", "semantic_scholar"], "title": "VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization", "abstract": "Large Multimodal Models (LMMs) excel in visual-language tasks by leveraging numerous visual tokens for fine-grained visual information, but this token redundancy results in significant computational costs. Previous research aimed at reducing visual tokens during inference typically leverages importance maps derived from attention scores among vision-only tokens or vision-language tokens to prune tokens across one or multiple pruning stages. Despite this progress, pruning frameworks and strategies remain simplistic and insufficiently explored, often resulting in substantial performance degradation. In this paper, we propose VFlowOpt, a token pruning framework that introduces an importance map derivation process and a progressive pruning module with a recycling mechanism. The hyperparameters of its pruning strategy are further optimized by a visual information flow-guided method. Specifically, we compute an importance map for image tokens based on their attention-derived context relevance and patch-level information entropy. We then decide which tokens to retain or prune and aggregate the pruned ones as recycled tokens to avoid potential information loss. Finally, we apply a visual information flow-guided method that regards the last token in the LMM as the most representative signal of text-visual interactions. This method minimizes the discrepancy between token representations in LMMs with and without pruning, thereby enabling superior pruning strategies tailored to different LMMs. Experiments demonstrate that VFlowOpt can prune 90% of visual tokens while maintaining comparable performance, leading to an 89% reduction in KV-Cache memory and 3.8 times faster inference.", "authors": ["Sihan Yang", "Runsen Xu", "Chenhang Cui", "Tai Wang", "Dahua Lin", "Jiangmiao Pang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-07", "url": "https://arxiv.org/abs/2508.05211", "pdf_url": "https://arxiv.org/pdf/2508.05211v2", "arxiv_id": "2508.05211", "doi": "10.1109/ICCV51701.2025.02218", "citation_count": 11, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.2698}
{"id": "081d33900ed6e759bd632af3cfc52905fefb0258fe3b32b085d9dfd8eddd5dfc", "sources": ["arxiv", "semantic_scholar"], "title": "HAMoBE: Hierarchical and Adaptive Mixture of Biometric Experts for Video-based Person ReID", "abstract": "Recently, research interest in person re-identification (ReID) has increasingly focused on video-based scenarios, which are essential for robust surveillance and security in varied and dynamic environments. However, existing video-based ReID methods often overlook the necessity of identifying and selecting the most discriminative features from both videos in a query-gallery pair for effective matching. To address this issue, we propose a novel Hierarchical and Adaptive Mixture of Biometric Experts (HAMoBE) framework, which leverages multi-layer features from a pre-trained large model (e.g., CLIP) and is designed to mimic human perceptual mechanisms by independently modeling key biometric features--appearance, static body shape, and dynamic gait--and adaptively integrating them. Specifically, HAMoBE includes two levels: the first level extracts low-level features from multi-layer representations provided by the frozen large model, while the second level consists of specialized experts focusing on long-term, short-term, and temporal features. To ensure robust matching, we introduce a new dual-input decision gating network that dynamically adjusts the contributions of each expert based on their relevance to the input scenarios. Extensive evaluations on benchmarks like MEVID demonstrate that our approach yields significant performance improvements (e.g., +13.0% Rank-1 accuracy).", "authors": ["Yiyang Su", "Yunping Shi", "Feng Liu", "Xiaoming Liu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-07", "url": "https://arxiv.org/abs/2508.05038", "pdf_url": "https://arxiv.org/pdf/2508.05038v2", "arxiv_id": "2508.05038", "doi": "10.1109/ICCV51701.2025.01072", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.2698}
{"id": "bfe97a60f7390c75279904814d1bd6b6c239ad98e43a15d66e9aa524a786d201", "sources": ["arxiv", "semantic_scholar"], "title": "Cross-LoRA: A Data-Free LoRA Transfer Framework across Heterogeneous LLMs", "abstract": "Traditional parameter-efficient fine-tuning (PEFT) methods such as LoRA are tightly coupled with the base model architecture, which constrains their applicability across heterogeneous pretrained large language models (LLMs). To address this limitation, we introduce Cross-LoRA, a data-free framework for transferring LoRA modules between diverse base models without requiring additional training data. Cross-LoRA consists of two key components: (a) LoRA-Align, which performs subspace alignment between source and target base models through rank-truncated singular value decomposition (SVD) and Frobenius-optimal linear transformation, ensuring compatibility under dimension mismatch; and (b) LoRA-Shift, which applies the aligned subspaces to project source LoRA weight updates into the target model parameter space. Both components are data-free, training-free, and enable lightweight adaptation on a commodity GPU in 20 minutes. Experiments on ARCs, OBOA and HellaSwag show that Cross-LoRA achieves relative gains of up to 5.26% over base models. Across other commonsense reasoning benchmarks, Cross-LoRA maintains performance comparable to that of directly trained LoRA adapters.", "authors": ["Feifan Xia", "Mingyang Liao", "Yuyang Fang", "Defang Li", "Yantong Xie", "Weikang Li", "Yang Li", "Deguo Xia", "Jizhou Huang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-07", "url": "https://arxiv.org/abs/2508.05232", "pdf_url": "https://arxiv.org/pdf/2508.05232v1", "arxiv_id": "2508.05232", "doi": "10.48550/arXiv.2508.05232", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2601}
{"id": "81818ae372c0059e5e2e99e6a4ca37ad0ff3c0f9ca9c49d8b33da88017b2de2d", "sources": ["arxiv", "semantic_scholar"], "title": "MoBE: Mixture-of-Basis-Experts for Compressing MoE-based LLMs", "abstract": "The Mixture-of-Experts (MoE) architecture has become a predominant paradigm for scaling large language models (LLMs). Despite offering strong performance and computational efficiency, large MoE-based LLMs like DeepSeek-V3-0324 and Kimi-K2-Instruct present serious challenges due to substantial memory requirements in deployment. While recent works have explored MoE compression to address this issue, existing methods often suffer from considerable accuracy drops (e.g., 7-14% relatively) even at modest compression rates. This paper introduces a novel Mixture-of-Basis-Experts (MoBE) method that achieves model compression while incurring minimal accuracy drops. Specifically, each up/gate matrix in an expert is decomposed via a rank decomposition as W = AB, where matrix A is unique to each expert. The relatively larger matrix B is further re-parameterized as a linear combination of basis matrices {Bi} shared across all experts within a given MoE layer. The factorization is learned by minimizing the reconstruction error relative to the original weight matrices. Experiments demonstrate that MoBE achieves notably lower accuracy drops compared to prior works. For instance, MoBE can reduce the parameter counts of Qwen3-235B-A22B-2507, DeepSeek-V3-0324 (671B) and Kimi-K2-Instruct (1T) by 24%-30% with only 1%-2% accuracy drop (about 2% drops when measured relatively).", "authors": ["Xiaodong Chen", "Mingming Ha", "Zhenzhong Lan", "Jing Zhang", "Jianguo Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-07", "url": "https://arxiv.org/abs/2508.05257", "pdf_url": "https://arxiv.org/pdf/2508.05257v1", "arxiv_id": "2508.05257", "doi": "10.48550/arXiv.2508.05257", "citation_count": 5, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2601}
{"id": "506161af675d196015a82a8349797a75a5a362b5c0de99314088e8328288eaf6", "sources": ["arxiv", "semantic_scholar"], "title": "KVSink: Understanding and Enhancing the Preservation of Attention Sinks in KV Cache Quantization for LLMs", "abstract": "Key-Value (KV) cache quantization has become a widely adopted optimization technique for efficient large language models (LLMs) inference by reducing KV cache memory usage and mitigating memory-bound constraints. Recent studies have emphasized the importance of preserving the original precision of KVs for the first few tokens to ensure the protection of attention sinks. While this approach has proven effective in mitigating performance degradation, its underlying principles remain insufficiently understood. Moreover, it fails to address the recent discovery that attention sinks can emerge beyond the initial token positions. In this work, we elucidate the underlying mechanisms of attention sinks during inference by examining their role in the cross-layer evolution of extreme activation outliers. Additionally, we provide a comprehensive analysis of the interplay between attention sinks and KV cache quantization. Based on our enhanced understanding, we introduce \\textit{\\textbf{KVSink}}, a plug-and-play method that effectively predicts sink tokens with negligible overhead, enabling more thorough preservation. Extensive experiments demonstrate that KVSink outperforms the existing Preserve-First-N (PFN) strategy, offering more effective preservation of attention sinks during KV cache quantization. Moreover, when applied to the well-established KVQuant method, KVSink further improves perplexity (PPL) and reduces reliance on 16-bit numerical outliers.", "authors": ["Zunhai Su", "Kehong Yuan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-06", "url": "https://arxiv.org/abs/2508.04257", "pdf_url": "https://arxiv.org/pdf/2508.04257v1", "arxiv_id": "2508.04257", "doi": "10.48550/arXiv.2508.04257", "citation_count": 17, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3138}
{"id": "2dddf2355d16c581e2b3013d481cfdf902b4229addbafa6ed890bd369f0019e0", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Attention across Multiple-context KV Cache", "abstract": "Large language models face significant cost challenges in long-sequence inference. To address this, reusing historical Key-Value (KV) Cache for improved inference efficiency has become a mainstream approach. Recent advances further enhance throughput by sparse attention mechanisms to select the most relevant KV Cache, thereby reducing sequence length. However, such techniques are limited to single-context scenarios, where historical KV Cache is computed sequentially with causal-attention dependencies. In retrieval-augmented generation (RAG) scenarios, where retrieved documents as context are unknown beforehand, each document's KV Cache is computed and stored independently (termed multiple-context KV Cache), lacking cross-attention between contexts. This renders existing methods ineffective. Although prior work partially recomputes multiple-context KV Cache to mitigate accuracy loss from missing cross-attention, it requires retaining all KV Cache throughout, failing to reduce memory overhead. This paper presents SamKV, the first exploration of attention sparsification for multiple-context KV Cache. Specifically, SamKV takes into account the complementary information of other contexts when sparsifying one context, and then locally recomputes the sparsified information. Experiments demonstrate that our method compresses sequence length to 15% without accuracy degradation compared with full-recompuation baselines, significantly boosting throughput in multi-context RAG scenarios.", "authors": ["Ziyi Cao", "Qingyi Si", "Jingbin Zhang", "Bingquan Liu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-06", "url": "https://arxiv.org/abs/2508.11661", "pdf_url": "https://arxiv.org/pdf/2508.11661v1", "arxiv_id": "2508.11661", "doi": "10.48550/arXiv.2508.11661", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.259}
{"id": "0f7d96aa4887bd4b8e8146a4d4bacd11b9c4c09047d2ce6c957d214a9b2b0d74", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring Layer-wise Information Effectiveness for Post-Training Quantization in Small Language Models", "abstract": "Large language models with billions of parameters are often over-provisioned: many layers contribute little unique information yet dominate the memory and energy footprint during inference. We present LieQ Layer-wise information effectiveness Quantization, a hardware-native, metric-driven post-training quantization framework that addresses the critical challenge of maintaining accuracy in sub-8B models, model parameters less than 8B, under extreme low-bit compression. LieQ keeps uniform bit-width within each layer while mixing precision across layers, preserving standard multiplication kernels and avoiding irregular memory access, codebooks, or irregular formats at inference time. Our method uncovers a strong correlation between layer-wise functional saliency and representational compactness, revealing that layers with higher training-induced energy concentration are functionally irreplaceable. Leveraging this insight, we propose a purely geometry-driven sensitivity proxy that enables automatic bit-width allocation under a target average-bit budget without expensive gradient updates or inference-based perplexity probing. At sub 2-bit, LieQ consistently reduces the large accuracy gap typically observed for naive 2-bit baselines on Qwen3 and LLaMA3.x families, while retaining standard-kernel efficiency. These properties make LieQ a practical path toward deploying small language models on resource-constrained edge devices. Code will available here: https://github.com/HeXiao-55/LieQ-official.git.", "authors": ["He Xiao", "Qingyao Yang", "Dirui Xie", "Wendong Xu", "Zunhai Su", "Runming yang", "Wenyong Zhou", "Haobo Liu", "Zhengwu Liu", "Ngai Wong"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-05", "url": "https://arxiv.org/abs/2508.03332", "pdf_url": "https://arxiv.org/pdf/2508.03332v2", "arxiv_id": "2508.03332", "doi": "10.48550/arXiv.2508.03332", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/HeXiao-55/LieQ-official.git", "venue": "arXiv.org", "quality_score": 0.3984}
{"id": "7bd02a9cbe6f3df75ce20bd470215926d15add77f7adae18161535364b6481c1", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Hard Sharing: Efficient Multi-Task Speech-to-Text Modeling with Supervised Mixture of Experts", "abstract": "Hard-parameter sharing is a common strategy to train a single model jointly across diverse tasks. However, this often leads to task interference, impeding overall model performance. To address the issue, we propose a simple yet effective Supervised Mixture of Experts (S-MoE). Unlike traditional Mixture of Experts models, S-MoE eliminates the need for training gating functions by utilizing special guiding tokens to route each task to its designated expert. By assigning each task to a separate feedforward network, S-MoE overcomes the limitations of hard-parameter sharing. We further apply S-MoE to a speech-to-text model, enabling the model to process mixed-bandwidth input while jointly performing automatic speech recognition (ASR) and speech translation (ST). Experimental results demonstrate the effectiveness of the proposed S-MoE, achieving a 6.35% relative improvement in Word Error Rate (WER) when applied to both the encoder and decoder.", "authors": ["Hojun Jin", "Eunsoo Hong", "Ziwon Hyung", "Sungjun Lim", "Seungjin Lee", "Keunseok Cho"], "categories": ["cs.CL", "cs.AI", "cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-08-05", "url": "https://arxiv.org/abs/2508.10009", "pdf_url": "https://arxiv.org/pdf/2508.10009v1", "arxiv_id": "2508.10009", "doi": "10.48550/arXiv.2508.10009", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.2578}
{"id": "eec741e56820d9815c52ba92655e46ea9260c404c74b5bb93cf63d93a57fb627", "sources": ["arxiv", "semantic_scholar"], "title": "MoCA: Identity-Preserving Text-to-Video Generation via Mixture of Cross Attention", "abstract": "Achieving ID-preserving text-to-video (T2V) generation remains challenging despite recent advances in diffusion-based models. Existing approaches often fail to capture fine-grained facial dynamics or maintain temporal identity coherence. To address these limitations, we propose MoCA, a novel Video Diffusion Model built on a Diffusion Transformer (DiT) backbone, incorporating a Mixture of Cross-Attention mechanism inspired by the Mixture-of-Experts paradigm. Our framework improves inter-frame identity consistency by embedding MoCA layers into each DiT block, where Hierarchical Temporal Pooling captures identity features over varying timescales, and Temporal-Aware Cross-Attention Experts dynamically model spatiotemporal relationships. We further incorporate a Latent Video Perceptual Loss to enhance identity coherence and fine-grained details across video frames. To train this model, we collect CelebIPVid, a dataset of 10,000 high-resolution videos from 1,000 diverse individuals, promoting cross-ethnicity generalization. Extensive experiments on CelebIPVid show that MoCA outperforms existing T2V methods by over 5% across Face similarity.", "authors": ["Qi Xie", "Yongjia Ma", "Donglin Di", "Xuehao Gao", "Xun Yang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-05", "url": "https://arxiv.org/abs/2508.03034", "pdf_url": "https://arxiv.org/pdf/2508.03034v2", "arxiv_id": "2508.03034", "doi": "10.1145/3743093.3770948", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "ACM Multimedia Asia", "quality_score": 0.2578}
{"id": "425782ef2ab3aa2c5a340f0e111c765b2e8279e4028c065791c980466a23d560", "sources": ["arxiv", "semantic_scholar"], "title": "Subject or Style: Adaptive and Training-Free Mixture of LoRAs", "abstract": "Fine-tuning models via Low-Rank Adaptation (LoRA) demonstrates remarkable performance in subject-driven or style-driven generation tasks. Studies have explored combinations of different LoRAs to jointly generate learned styles and content. However, current methods struggle to balance the original subject and style, and often require additional training. Recently, K-LoRA proposed a training-free LoRA fusion method. But it involves multiple hyperparameters, making it difficult to adapt to all styles and subjects. In this paper, we propose EST-LoRA, a training-free adaptive LoRA fusion method. It comprehensively considers three critical factors: \\underline{E}nergy of matrix, \\underline{S}tyle discrepancy scores and \\underline{T}ime steps. Analogous to the Mixture of Experts (MoE) architecture, the model adaptively selects between subject LoRA and style LoRA within each attention layer. This integrated selection mechanism ensures balanced contributions from both components during the generation process. Experimental results show that EST-LoRA outperforms state-of-the-art methods in both qualitative and quantitative evaluations and achieves faster generation speed compared to other efficient fusion approaches. Our code is publicly available at: https://anonymous.4open.science/r/EST-LoRA-F318.", "authors": ["Jia-Chen Zhang", "Yu-Jie Xiong"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-04", "url": "https://arxiv.org/abs/2508.02165", "pdf_url": "https://arxiv.org/pdf/2508.02165v1", "arxiv_id": "2508.02165", "doi": "10.48550/arXiv.2508.02165", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2567}
{"id": "19f66baa34e98e3e4dcc3ed5defdfaf57a5742bf7cbe5519716489959910fd19", "sources": ["arxiv", "semantic_scholar"], "title": "LeanK: Learnable K Cache Channel Pruning for Efficient Decoding", "abstract": "Large language models (LLMs) enable long-context tasks but face efficiency challenges due to the growing key-value (KV) cache. We propose LeanK, a learning-based method that prunes unimportant key (K) cache channels by leveraging static channel sparsity. With a novel two-stage training process, LeanK learns channel-wise static mask that could satisfy specific sparsity ratio and hardware alignment requirement. LeanK reduces GPU memory and accelerates decoding without sacrificing accuracy. Experiments demonstrate up to 70% K cache and 16%-18% V cache memory reduction. Custom decoding kernel enables 1.3x speedup for attention computation. We also provide insights into model channels and attention heads during long-context inference by analyzing the learned importance distribution. Our code is available at https://aka.ms/LeanK.", "authors": ["Yike Zhang", "Zhiyuan He", "Huiqiang Jiang", "Chengruidong Zhang", "Yuqing Yang", "Jianyong Wang", "Lili Qiu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-04", "url": "https://arxiv.org/abs/2508.02215", "pdf_url": "https://arxiv.org/pdf/2508.02215v1", "arxiv_id": "2508.02215", "doi": "10.48550/arXiv.2508.02215", "citation_count": 4, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3967}
{"id": "6f68b5a2916a75fea3a2a24d7c564ab45af8baf77a2eec80300c0fa893ef9ce0", "sources": ["arxiv", "semantic_scholar"], "title": "Parameter-Efficient Routed Fine-Tuning: Mixture-of-Experts Demands Mixture of Adaptation Modules", "abstract": "Mixture-of-Experts (MoE) benefits from a dynamic routing mechanism among their specialized experts, which existing Parameter- Efficient Fine-Tuning (PEFT) strategies fail to leverage. This motivates us to investigate whether adaptation modules themselves should incorporate routing mechanisms to align with MoE's multi-expert architecture. We analyze dynamics of core components when applying PEFT to MoE language models and examine how different routing strategies affect adaptation effectiveness. Extensive experiments adapting OLMoE-1B-7B and Mixtral-8x7B on various commonsense and math reasoning tasks validate the performance and efficiency of our routed approach. We identify the optimal configurations for different scenarios and provide empirical analyses with practical insights to facilitate better PEFT and MoE applications.", "authors": ["Yilun Liu", "Yunpu Ma", "Yuetian Lu", "Shuo Chen", "Zifeng Ding", "Volker Tresp"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-04", "url": "https://arxiv.org/abs/2508.02587", "pdf_url": "https://arxiv.org/pdf/2508.02587v1", "arxiv_id": "2508.02587", "doi": "10.48550/arXiv.2508.02587", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference of the European Chapter of the Association for Computational Linguistics", "quality_score": 0.2567}
{"id": "f14c3f58d493484b4a25f5635b1f1ea1747ba4409a9a72ba6dcd00f6e878879f", "sources": ["arxiv", "semantic_scholar"], "title": "Kron-LoRA: Hybrid Kronecker-LoRA Adapters for Scalable, Sustainable Fine-tuning", "abstract": "Fine-tuning massive pre-trained language models across many tasks demands adapters that are both parameter-efficient and expressive. We introduce \\textbf{Kron-LoRA}, a hybrid adapter that combines Kronecker-structured factorization with low-rank LoRA compression-an integration that, to our knowledge, has not been explored in parameter-efficient fine-tuning or in matrix approximation literature. Kron-LoRA achieves up to 4$\\times$ fewer parameters than standard LoRA while retaining similar expressivity. Experiments on DistilBERT, Mistral-7B, LLaMA-2-7B, and LLaMA-3-8B across eight benchmarks show that Kron-LoRA matches or exceeds LoRA baselines with modest memory savings and only a 5-8\\% speed overhead. In sequential fine-tuning, it also delivers competitive cross-task transfer despite using only one-quarter of the adapter parameters. Kron-LoRA thus offers a scalable, sustainable solution for multi-task adaptation of large language models.", "authors": ["Yixin Shen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-04", "url": "https://arxiv.org/abs/2508.01961", "pdf_url": "https://arxiv.org/pdf/2508.01961v2", "arxiv_id": "2508.01961", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1633}
{"id": "c34691f02cbdff0ef09dc7c09d0c5ba8af2de845da0feaaf0d66443ea5ac90b7", "sources": ["arxiv", "semantic_scholar"], "title": "EAC-MoE: Expert-Selection Aware Compressor for Mixture-of-Experts Large Language Models", "abstract": "Mixture-of-Experts (MoE) has demonstrated promising potential in scaling LLMs. However, it is hindered by two critical challenges: (1) substantial GPU memory consumption to load all experts; (2) low activated parameters cannot be equivalently translated into inference acceleration effects. In this work, we propose EAC-MoE, an Expert-Selection Aware Compressor for MoE-LLMs, which deeply aligns with the characteristics of MoE from the perspectives of quantization and pruning, and introduces two modules to address these two challenges respectively: (1) The expert selection bias caused by low-bit quantization is a major factor contributing to the performance degradation in MoE-LLMs. Based on this, we propose Quantization with Expert-Selection Calibration (QESC), which mitigates the expert selection bias by calibrating the routers within the MoE; (2) There are always certain experts that are not crucial for the corresponding tasks, yet causing inference latency. Therefore, we propose Pruning based on Expert-Selection Frequency (PESF), which significantly improves inference speed by pruning less frequently used experts for current task. Extensive experiments demonstrate that our approach significantly reduces memory usage and improves inference speed with minimal performance degradation.", "authors": ["Yuanteng Chen", "Yuantian Shao", "Peisong Wang", "Jian Cheng"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-03", "url": "https://arxiv.org/abs/2508.01625", "pdf_url": "https://arxiv.org/pdf/2508.01625v1", "arxiv_id": "2508.01625", "doi": "10.18653/v1/2025.acl-long.633", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.294}
{"id": "846eeab36f74e1e3688c5a1ef68361785c6860231b42824f501ed8b7be003e4e", "sources": ["arxiv", "semantic_scholar"], "title": "SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference", "abstract": "KV cache eviction has emerged as an effective solution to alleviate resource constraints faced by LLMs in long-context scenarios. However, existing token-level eviction methods often overlook two critical aspects: (1) their irreversible eviction strategy fails to adapt to dynamic attention patterns during decoding (the saliency shift problem), and (2) they treat both marginally important tokens and truly unimportant tokens equally, despite the collective significance of marginal tokens to model performance (the marginal information over-compression problem). To address these issues, we design two compensation mechanisms based on the high similarity of attention matrices between LLMs of different scales. We propose SmallKV, a small model assisted compensation method for KV cache compression. SmallKV can maintain attention matching between different-scale LLMs to: 1) assist the larger model in perceiving globally important information of attention; and 2) use the smaller model's attention scores to approximate those of marginal tokens in the larger model. Extensive experiments on benchmarks including GSM8K, BBH, MT-Bench, and LongBench demonstrate the effectiveness of SmallKV. Moreover, efficiency evaluations show that SmallKV achieves 1.75 - 2.56 times higher throughput than baseline methods, highlighting its potential for efficient and performant LLM inference in resource constrained environments.", "authors": ["Yi Zhao", "Yajuan Peng", "Cam-Tu Nguyen", "Zuchao Li", "Xiaoliang Wang", "Hai Zhao", "Xiaoming Fu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-03", "url": "https://arxiv.org/abs/2508.02751", "pdf_url": "https://arxiv.org/pdf/2508.02751v2", "arxiv_id": "2508.02751", "doi": "10.48550/arXiv.2508.02751", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2555}
{"id": "0d810ac5f6c9445f250291fd4286f7b0fed123c0648866893bfd00c1cb3834d5", "sources": ["arxiv", "semantic_scholar"], "title": "M3LLM: Model Context Protocol-aided Mixture of Vision Experts For Multimodal LLMs in Networks", "abstract": "Current Multimodal Large Language Models (MLLMs) rely on centralized architectures and often suffer from poor alignment between the input task and their fixed visual encoding modules, which limits performance on diverse and dynamic visual tasks. With the increasing deployment of resource-efficient models on edge devices in wireless networks, a new opportunity emerges to dynamically use distributed vision experts for improved MLLM inference quality. To enable this, we propose M3LLM, where the Model Context Protocol (MCP) coordinates a mixture of vision experts to achieve distributed MLLMs. Specifically, MCP is an open protocol that structures the input task context into interpretable representations, enabling wireless network-aware coordination between the central model backbone and edge-hosted vision experts. Based on the MCP representation, M3LLM formulates vision expert routing as a joint optimization problem that balances task-expert semantic compatibility and channel performance. To solve the resulting gradient conflicts, we develop a dual-stream Soft Actor-Critic (SAC) algorithm with decoupled reward signals and introduce an Adaptive Stability Enhancement Module (ASEM) based on hierarchical Bayesian modeling to ensure effective routing. Experiments show that M3LLM improves task accuracy, reduces communication cost, and enhances expert routing adaptability under dynamic wireless network conditions.", "authors": ["Yongjie Zeng", "Hongyang Du"], "categories": ["cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-03", "url": "https://arxiv.org/abs/2508.01805", "pdf_url": "https://arxiv.org/pdf/2508.01805v1", "arxiv_id": "2508.01805", "doi": "10.48550/arXiv.2508.01805", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2555}
{"id": "0d47d4aace05e2e4658cc5d3b9eeaa557928282d70cdcd4d6567292f72300546", "sources": ["arxiv", "semantic_scholar"], "title": "M$^3$AD: Multi-task Multi-gate Mixture of Experts for Alzheimer's Disease Diagnosis with Conversion Pattern Modeling", "abstract": "Alzheimer's disease (AD) progression follows a complex continuum from normal cognition (NC) through mild cognitive impairment (MCI) to dementia, yet most deep learning approaches oversimplify this into discrete classification tasks. This study introduces M$^3$AD, a novel multi-task multi-gate mixture of experts framework that jointly addresses diagnostic classification and cognitive transition modeling using structural MRI. We incorporate three key innovations: (1) an open-source T1-weighted sMRI preprocessing pipeline, (2) a unified learning framework capturing NC-MCI-AD transition patterns with demographic priors (age, gender, brain volume) for improved generalization, and (3) a customized multi-gate mixture of experts architecture enabling effective multi-task learning with structural MRI alone. The framework employs specialized expert networks for diagnosis-specific pathological patterns while shared experts model common structural features across the cognitive continuum. A two-stage training protocol combines SimMIM pretraining with multi-task fine-tuning for joint optimization. Comprehensive evaluation across six datasets comprising 12,037 T1-weighted sMRI scans demonstrates superior performance: 95.13% accuracy for three-class NC-MCI-AD classification and 99.15% for binary NC-AD classification, representing improvements of 4.69% and 0.55% over state-of-the-art approaches. The multi-task formulation simultaneously achieves 97.76% accuracy in predicting cognitive transition. Our framework outperforms existing methods using fewer modalities and offers a clinically practical solution for early intervention. Code: https://github.com/csyfjiang/M3AD.", "authors": ["Yufeng Jiang", "Hexiao Ding", "Hongzhao Chen", "Jing Lan", "Xinzhi Teng", "Gerald W. Y. Cheng", "Zongxi Li", "Haoran Xie", "Jung Sun Yoo", "Jing Cai"], "categories": ["eess.IV"], "fields_of_study": ["Engineering"], "published_date": "2025-08-03", "url": "https://arxiv.org/abs/2508.01819", "pdf_url": "https://arxiv.org/pdf/2508.01819v1", "arxiv_id": "2508.01819", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/csyfjiang/M3AD", "venue": null, "quality_score": 0.302}
{"id": "eba0dce9ba5346a9439c788ec9b1aa5621d90792e7a779ce5691b9056dcd408f", "sources": ["arxiv", "semantic_scholar"], "title": "PiKV: KV Cache Management System for Mixture of Experts", "abstract": "As large-scale language models continue to scale up in both size and context length, the memory and communication cost of key-value (KV) cache storage has become a major bottleneck in multi-GPU and multi-node inference. While MoE-based architectures sparsify computation across experts, the corresponding KV caches remain dense and globally synchronized, resulting in significant overhead. We introduce \\textbf{PiKV}, a parallel and distributed KV cache serving framework tailored for MoE architecture. PiKV leverages \\textit{expert-sharded KV storage} to partition caches across GPUs, \\textit{PiKV routing} to reduce token-to-KV access, and a \\textit{PiKV Scheduling} to adaptively retain query-relevant entries. To further reduce memory usage, PiKV integrates \\textit{PiKV Compression} modules the caching pipeline for acceleration. PiKV is recently publicly available as an open-source software library: \\href{https://github.com/NoakLiu/PiKV}{https://github.com/NoakLiu/PiKV}. PiKV is still a living project, aiming to become a comprehesive KV Cache management system for MoE Architectures.", "authors": ["Dong Liu", "Yanxuan Yu", "Ben Lengerich", "Ying Nian Wu"], "categories": ["cs.DC", "cs.AI", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-02", "url": "https://arxiv.org/abs/2508.06526", "pdf_url": "https://arxiv.org/pdf/2508.06526v3", "arxiv_id": "2508.06526", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/NoakLiu/PiKV", "venue": null, "quality_score": 0.3006}
{"id": "0f7706f03779f4b1e4e701bcd74cccf634b4ea2f5425ba7d131f7e288bd6b078", "sources": ["arxiv", "semantic_scholar"], "title": "Unifying Mixture of Experts and Multi-Head Latent Attention for Efficient Language Models", "abstract": "We present MoE-MLA-RoPE, a novel architecture combination that combines Mixture of Experts (MoE) with Multi-head Latent Attention (MLA) and Rotary Position Embeddings (RoPE) for efficient language modeling. Our approach addresses the fundamental trade-off between model capacity and computational efficiency through three key innovations: (1) fine-grained expert routing with 64 micro-experts and top-$k$ selection, enabling flexible specialization through 3.6 * 10^7 possible expert combinations; (2) shared expert isolation that dedicates 2 always active experts for common patterns while routing to 6 of 62 specialized experts; and (3) gradient-conflict-free load balancing that maintains expert utilization without interfering with primary loss optimization. Extensive experiments on models ranging from 17M to 202M parameters demonstrate that MoE-MLA-RoPE with compression ratio r=d/2 achieves 68% KV cache memory reduction and 3.2x inference speedup while maintaining competitive perplexity (0.8% degradation). Compared to the parameters with 53.9M parameters, MoE-MLA-RoPE improves the validation loss by 6.9% over the vanilla transformers while using 42% fewer active parameters per forward pass. FLOP-matched experiments reveal even larger gains: 11.1% improvement with 3.2x inference acceleration. Automated evaluation using GPT-4 as a judge confirms quality improvements in generation, with higher scores on coherence (8.1/10), creativity (7.9/10) and grammatical correctness (8.2/10). Our results establish that architectural novelty, not parameter scaling, defines the efficiency frontier for resource-constrained language model deployment.", "authors": ["Sushant Mehta", "Raj Dandekar", "Rajat Dandekar", "Sreedath Panat"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-02", "url": "https://arxiv.org/abs/2508.01261", "pdf_url": "https://arxiv.org/pdf/2508.01261v1", "arxiv_id": "2508.01261", "doi": "10.48550/arXiv.2508.01261", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1619}
{"id": "ac595b712866ab97603ab3db55aab55c5346a60d8caf740ede8ad2c1d9f02c1e", "sources": ["arxiv", "semantic_scholar"], "title": "HiPrune: Hierarchical Attention for Efficient Token Pruning in Vision-Language Models", "abstract": "Vision-Language Models (VLMs) encode images and videos into abundant tokens, which contain substantial redundancy and computation cost. While visual token pruning mitigates the issue, most existing methods lack insight into the intrinsic property of the vision encoder itself. In this work, we dive into the vision encoder and prove that the middle layers pay more attention to the main objects of the image qualitatively and quantitatively, while the deep layers to tokens with rich global information. Utilizing this Hierarchical attention pattern, we propose HiPrune, a training-free and model-agnostic token Pruning method. HiPrune identifies three types of visual tokens according to their attention in different phases of the vision encoder, which preserves different levels of information. By coupling with the similarity of text tokens, we propose a prompt-aware variance, HiPrune++, which further improves instruction following performance under a very low token budget. Extensive experiments across four representative VLMs show that HiPrune achieves up to 99.3% of task accuracy with only 1/3 of the tokens, while reducing inference FLOPs by 58.7%. HiPrune++ maintains up to 99.7% accuracy with 2/9 tokens, highlighting robustness under high-resolution. Our code is available at https://github.com/Danielement321/HiPrune.", "authors": ["Jizhihui Liu", "Feiyi Du", "Guangdao Zhu", "Niu Lian", "Jun Li", "Bin Chen", "Weili Guan", "Yaowei Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-01", "url": "https://arxiv.org/abs/2508.00553", "pdf_url": "https://arxiv.org/pdf/2508.00553v3", "arxiv_id": "2508.00553", "doi": null, "citation_count": 6, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/Danielement321/HiPrune", "venue": null, "quality_score": 0.2993}
{"id": "f748e5fe9fab3f261a541afde6e6069133647c3c839ab1209fc45f43c2346f2a", "sources": ["arxiv", "semantic_scholar"], "title": "Unveiling Super Experts in Mixture-of-Experts Large Language Models", "abstract": "In this study, we report, for the first time, the discovery and systematic investigation of a distinct subset of experts that play a pivotal role in the MoE LLMs' forward inference. These experts are prevalent in open-source MoE LLMs, and despite their extremely limited number, pruning them results in a substantial decline in model performance (e.g., prune just three out of 6,144 causes Qwen3-30B-A3B to generate repetitive and uninformative outputs).We refer to these experts as Super Experts (SEs). Our comprehensive analysis provides progressively deeper insights into SEs: (i) SEs are characterized by rare but extreme activation outliers in the output of the down_proj, which give rise to massive activations in the hidden states between decoder layers. Moreover, the distribution of SEs is model-specific, data-agnostic, and remains unaffected by post-training processes. (ii) By pruning SEs, we assess their significance across a variety of tasks, revealing their considerable impact on the model's overall performance, particularly in mathematical reasoning. (iii) We further investigate why compressing SEs exerts such a pronounced impact. We show that, in MoE LLMs, SEs serve as the primary source of the systematic outlier mechanism in Transformers, and that compressing them profoundly disrupts this process, ultimately causing the collapse of attention sinks. These findings advance the understanding of the internal dynamics of MoE LLMs, filling an important gap in the current knowledge. The code is provided in https://github.com/ZunhaiSu/Super-Experts-Profilling.", "authors": ["Zunhai Su", "Qingyuan Li", "Hao Zhang", "Weihao Ye", "Qibo Xue", "YuLei Qian", "Yuchen Xie", "Ngai Wong", "Kehong Yuan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-31", "url": "https://arxiv.org/abs/2507.23279", "pdf_url": "https://arxiv.org/pdf/2507.23279v3", "arxiv_id": "2507.23279", "doi": "10.48550/arXiv.2507.23279", "citation_count": 14, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ZunhaiSu/Super-Experts-Profilling", "venue": "arXiv.org", "quality_score": 0.3896}
{"id": "6dc63e19a3ff1dfab0d5017a606172426b8d5abc5fc954272508c0a903f9aedc", "sources": ["arxiv", "semantic_scholar"], "title": "ReasonCache: Accelerating Large Reasoning Model Serving through KV Cache Sharing", "abstract": "Large Reasoning Models (LRMs) are becoming integral to many AI inference systems, enhancing their capabilities with advanced reasoning. However, deploying these models in production environments presents a significant QoS challenge: the substantial memory overhead from their long, auto-regressive inference processes severely limits throughput and increases latency, thereby affecting the quality of service for concurrent users. We observe that LRMs frequently generate highly similar intermediate reasoning steps, which, in turn, correspond to highly similar KV cache states across layers. Building on this insight, we propose ReasonCache, a novel KV cache management approach designed to improve the QoS of AI inference systems. ReasonCache utilizes a Collaborative Filtering Algorithm to efficiently identify reusable KV cache blocks and enables zero-copy cache reuse. Experimental evaluation demonstrates that ReasonCache achieves a peak throughput improvement of 89.2% and an average gain of 40-60%, leading to more responsive and cost-effective AI inference services. Notably, this performance is achieved while maintaining higher accuracy compared to existing KV cache management techniques.", "authors": ["Kaiwen Chen", "Xin Tan", "Minchen Yu", "Jingzong Li", "Hong Xu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-29", "url": "https://arxiv.org/abs/2507.21433", "pdf_url": "https://arxiv.org/pdf/2507.21433v3", "arxiv_id": "2507.21433", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.159}
{"id": "ab9dd7abb2903fc9d1d3c5bbd0a2a6ead508515527fb98b7c494c5ba52b68209", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Length and Pruning Experts for Knowledge Graphs Reasoning", "abstract": "Knowledge Graph (KG) reasoning, which aims to infer new facts from structured knowledge repositories, plays a vital role in Natural Language Processing (NLP) systems. Its effectiveness critically depends on constructing informative and contextually relevant reasoning paths. However, existing graph neural networks (GNNs) often adopt rigid, query-agnostic path-exploration strategies, limiting their ability to adapt to diverse linguistic contexts and semantic nuances. To address these limitations, we propose \\textbf{MoKGR}, a mixture-of-experts framework that personalizes path exploration through two complementary components: (1) a mixture of length experts that adaptively selects and weights candidate path lengths according to query complexity, providing query-specific reasoning depth; and (2) a mixture of pruning experts that evaluates candidate paths from a complementary perspective, retaining the most informative paths for each query. Through comprehensive experiments on diverse benchmark, MoKGR demonstrates superior performance in both transductive and inductive settings, validating the effectiveness of personalized path exploration in KGs reasoning.", "authors": ["Enjun Du", "Siyi Liu", "Yongqi Zhang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-28", "url": "https://arxiv.org/abs/2507.20498", "pdf_url": "https://arxiv.org/pdf/2507.20498v1", "arxiv_id": "2507.20498", "doi": "10.48550/arXiv.2507.20498", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.25}
{"id": "1a8907430aeaa155c5de36c46a24d35e01176bb1b75a2c9b580b6dba54b0e8ef", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Large Multimodal Models with Adaptive Sparsity and KV Cache Compression", "abstract": "Large multimodal models (LMMs) have advanced significantly by integrating visual encoders with extensive language models, enabling robust reasoning capabilities. However, compressing LMMs for deployment on edge devices remains a critical challenge. In this work, we propose an adaptive search algorithm that optimizes sparsity and KV cache compression to enhance LMM efficiency. Utilizing the Tree-structured Parzen Estimator, our method dynamically adjusts pruning ratios and KV cache quantization bandwidth across different LMM layers, using model performance as the optimization objective. This approach uniquely combines pruning with key-value cache quantization and incorporates a fast pruning technique that eliminates the need for additional fine-tuning or weight adjustments, achieving efficient compression without compromising accuracy. Comprehensive evaluations on benchmark datasets, including LLaVA-1.5 7B and 13B, demonstrate our method superiority over state-of-the-art techniques such as SparseGPT and Wanda across various compression levels. Notably, our framework automatic allocation of KV cache compression resources sets a new standard in LMM optimization, delivering memory efficiency without sacrificing much performance.", "authors": ["Te Zhang", "Yuheng Li", "Junxiang Wang", "Lujun Li"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-28", "url": "https://arxiv.org/abs/2507.20613", "pdf_url": "https://arxiv.org/pdf/2507.20613v1", "arxiv_id": "2507.20613", "doi": "10.1109/ICME59968.2025.11209796", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Multimedia and Expo", "quality_score": 0.2486}
{"id": "c0f79e129eb32f2f8a3195df44b2c48a98859d4ceb63a598193d20ed60c409b3", "sources": ["arxiv", "semantic_scholar"], "title": "Clustering data with values missing at random using scale mixtures of multivariate skew-normal distributions", "abstract": "Handling missing data is a major challenge in model-based clustering, especially when the data exhibit skewness and heavy tails. We address this by extending the finite mixture of scale mixtures of multivariate skew-normal (FMSMSN) family to accommodate incomplete data under a missing at random (MAR) mechanism. Unlike previous work that is limited to one of the special cases of the FMSMSN family, our method offers a cluster analysis methodology for the entire family that accounts for skewness and excess kurtosis amidst data with missing values. The multivariate skew-normal distribution, as parameterised by \\cite{azzalini1996} and \\cite{arnoldbeaver} includes the normal distribution as a special case, which ensures that our method is flexible toward existing symmetric model-based clustering techniques under a normality assumption. We derive the distributional properties of the missing components of the data and propose an augmented EM-type algorithm tailored for incomplete observations. The modified E-step yields closed-form expressions for the conditional expectations of the missing values. The simulation experiments showcase the flexibility of the FMSMSN family in both clustering performance and parameter recovery for varying percentages of missing values, while incorporating the effects of sample size and cluster proximity. Finally, we illustrate the practical utility of the proposed method by applying special cases of the FMSMSN family to global CO2 emissions data.", "authors": ["Jason Pillay", "Cristina Tortora", "Antonio Punzo", "Andriette Bekker"], "categories": ["stat.ME", "stat.CO"], "fields_of_study": ["Mathematics"], "published_date": "2025-07-27", "url": "https://arxiv.org/abs/2507.20329", "pdf_url": "https://arxiv.org/pdf/2507.20329v1", "arxiv_id": "2507.20329", "doi": "10.1007/s00180-026-01751-5", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Computational statistics (Zeitschrift)", "quality_score": 0.2475}
{"id": "9e6f232fcd4e46e6fe9dded7fd333f8986e2c82c54d98ce1d53a88111aef4a9e", "sources": ["arxiv", "semantic_scholar"], "title": "HCAttention: Extreme KV Cache Compression via Heterogeneous Attention Computing for LLMs", "abstract": "Processing long-context inputs with large language models presents a significant challenge due to the enormous memory requirements of the Key-Value (KV) cache during inference. Existing KV cache compression methods exhibit noticeable performance degradation when memory is reduced by more than 85%. Additionally, strategies that leverage GPU-CPU collaboration for approximate attention remain underexplored in this setting. We propose HCAttention, a heterogeneous attention computation framework that integrates key quantization, value offloading, and dynamic KV eviction to enable efficient inference under extreme memory constraints. The method is compatible with existing transformer architectures and does not require model fine-tuning. Experimental results on the LongBench benchmark demonstrate that our approach preserves the accuracy of full-attention model while shrinking the KV cache memory footprint to 25% of its original size. Remarkably, it stays competitive with only 12.5% of the cache, setting a new state-of-the-art in LLM KV cache compression. To the best of our knowledge, HCAttention is the first to extend the Llama-3-8B model to process 4 million tokens on a single A100 GPU with 80GB memory.", "authors": ["Dongquan Yang", "Yifan Yang", "Xiaotian Yu", "Xianbiao Qi", "Rong Xiao"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-26", "url": "https://arxiv.org/abs/2507.19823", "pdf_url": "https://arxiv.org/pdf/2507.19823v1", "arxiv_id": "2507.19823", "doi": "10.48550/arXiv.2507.19823", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neurocomputing", "quality_score": 0.2464}
{"id": "b2d2454df6d4ff1946d683d887b026b2c0129f8b157f9acd593947fa3b365d47", "sources": ["arxiv", "semantic_scholar"], "title": "CaliDrop: KV Cache Compression with Calibration", "abstract": "Large Language Models (LLMs) require substantial computational resources during generation. While the Key-Value (KV) cache significantly accelerates this process by storing attention intermediates, its memory footprint grows linearly with sequence length, batch size, and model size, creating a bottleneck in long-context scenarios. Various KV cache compression techniques, including token eviction, quantization, and low-rank projection, have been proposed to mitigate this bottleneck, often complementing each other. This paper focuses on enhancing token eviction strategies. Token eviction leverages the observation that the attention patterns are often sparse, allowing for the removal of less critical KV entries to save memory. However, this reduction usually comes at the cost of notable accuracy degradation, particularly under high compression ratios. To address this issue, we propose \\textbf{CaliDrop}, a novel strategy that enhances token eviction through calibration. Our preliminary experiments show that queries at nearby positions exhibit high similarity. Building on this observation, CaliDrop performs speculative calibration on the discarded tokens to mitigate the accuracy loss caused by token eviction. Extensive experiments demonstrate that CaliDrop significantly improves the accuracy of existing token eviction methods.", "authors": ["Yi Su", "Quantong Qiu", "Yuechi Zhou", "Juntao Li", "Qingrong Xia", "Ping Li", "Xinyu Duan", "Zhefeng Wang", "Min Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-26", "url": "https://arxiv.org/abs/2507.19906", "pdf_url": "https://arxiv.org/pdf/2507.19906v2", "arxiv_id": "2507.19906", "doi": "10.48550/arXiv.2507.19906", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2464}
{"id": "ece3b2b28a6426e434e0edf8386631f1bf59978dff6871b552f29265325bd197", "sources": ["arxiv", "semantic_scholar"], "title": "FAEDKV: Infinite-Window Fourier Transform for Unbiased KV Cache Compression", "abstract": "The efficacy of Large Language Models (LLMs) in long-context tasks is often hampered by the substantial memory footprint and computational demands of the Key-Value (KV) cache. Current compression strategies, including token eviction and learned projections, frequently lead to biased representations -- either by overemphasizing recent/high-attention tokens or by repeatedly degrading information from earlier context -- and may require costly model retraining. We present FAEDKV (Frequency-Adaptive Infinite-Window for KV cache), a novel, training-free KV cache compression framework that ensures unbiased information retention. FAEDKV operates by transforming the KV cache into the frequency domain using a proposed Infinite-Window Fourier Transform (IWDFT). This approach allows for the equalized contribution of all tokens to the compressed representation, effectively preserving both early and recent contextual information. A preliminary frequency ablation study identifies critical spectral components for layer-wise, targeted compression. Experiments on LongBench benchmark demonstrate FAEDKV's superiority over existing methods by up to 22\\%. In addition, our method shows superior, position-agnostic retrieval accuracy on the Needle-In-A-Haystack task compared to compression based approaches.", "authors": ["Runchao Li", "Yao Fu", "Mu Sheng", "Xianxuan Long", "Haotian Yu", "Pan Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-26", "url": "https://arxiv.org/abs/2507.20030", "pdf_url": "https://arxiv.org/pdf/2507.20030v1", "arxiv_id": "2507.20030", "doi": "10.48550/arXiv.2507.20030", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2464}
{"id": "a3ecbb94330878786d94c517b30f66d3d8aea7fa932d3a18bd2e47be572a91d0", "sources": ["arxiv", "semantic_scholar"], "title": "Patch Pruning Strategy Based on Robust Statistical Measures of Attention Weight Diversity in Vision Transformers", "abstract": "Multi-head self-attention is a distinctive feature extraction mechanism of vision transformers that computes pairwise relationships among all input patches, contributing significantly to their high performance. However, it is known to incur a quadratic computational complexity with respect to the number of patches. One promising approach to address this issue is patch pruning, which improves computational efficiency by identifying and removing redundant patches. In this work, we propose a patch pruning strategy that evaluates the importance of each patch based on the variance of attention weights across multiple attention heads. This approach is inspired by the design of multi-head self-attention, which aims to capture diverse attention patterns across different subspaces of feature representations. The proposed method can be easily applied during both training and inference, and achieves improved throughput while maintaining classification accuracy in scenarios such as fine-tuning with pre-trained models. In addition, we also found that using robust statistical measures, such as the median absolute deviation in place of variance, to assess patch importance can similarly lead to strong performance. Furthermore, by introducing overlapping patch embeddings, our method achieves better performance with comparable throughput to conventional approaches that utilize all patches.", "authors": ["Yuki Igaue", "Hiroaki Aizawa"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-25", "url": "https://arxiv.org/abs/2507.19175", "pdf_url": "https://arxiv.org/pdf/2507.19175v1", "arxiv_id": "2507.19175", "doi": "10.48550/arXiv.2507.19175", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Asian Conference on Pattern Recognition", "quality_score": 0.2452}
{"id": "cf68a64e68612f1b0e779d1b1f40665760311e80e138cf9c8856e52ea2731058", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Task Dense Prediction Fine-Tuning with Mixture of Fine-Grained Experts", "abstract": "Multi-task learning (MTL) for dense prediction has shown promising results but still faces challenges in balancing shared representations with task-specific specialization. In this paper, we introduce a novel Fine-Grained Mixture of Experts (FGMoE) architecture that explores MoE-based MTL models through a combination of three key innovations and fine-tuning. First, we propose intra-task experts that partition along intermediate hidden dimensions of MLPs, enabling finer decomposition of task information while maintaining parameter efficiency. Second, we introduce shared experts that consolidate common information across different contexts of the same task, reducing redundancy, and allowing routing experts to focus on unique aspects. Third, we design a global expert that facilitates adaptive knowledge transfer across tasks based on both input feature and task requirements, promoting beneficial information sharing while preventing harmful interference. In addition, we use the fine-tuning approach to improve parameter efficiency only by training the parameters of the decoder. Extensive experimental results show that the proposed FGMoE uses fewer parameters and significantly outperforms current MoE-based competitive MTL models on two dense prediction datasets (\\textit{i.e.,} NYUD-v2, PASCAL-Context) in various metrics.", "authors": ["Yangyang Xu", "Xi Ye", "Duo Su"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-25", "url": "https://arxiv.org/abs/2507.19077", "pdf_url": "https://arxiv.org/pdf/2507.19077v1", "arxiv_id": "2507.19077", "doi": "10.1145/3746027.3755586", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ACM Multimedia", "quality_score": 0.2452}
{"id": "c99a831f32a3aff6e61c77ac535955e73a87eba918d48de43b97c9af7cdfb21d", "sources": ["arxiv", "semantic_scholar"], "title": "DriftMoE: A Mixture of Experts Approach to Handle Concept Drifts", "abstract": "Learning from non-stationary data streams subject to concept drift requires models that can adapt on-the-fly while remaining resource-efficient. Existing adaptive ensemble methods often rely on coarse-grained adaptation mechanisms or simple voting schemes that fail to optimally leverage specialized knowledge. This paper introduces DriftMoE, an online Mixture-of-Experts (MoE) architecture that addresses these limitations through a novel co-training framework. DriftMoE features a compact neural router that is co-trained alongside a pool of incremental Hoeffding tree experts. The key innovation lies in a symbiotic learning loop that enables expert specialization: the router selects the most suitable expert for prediction, the relevant experts update incrementally with the true label, and the router refines its parameters using a multi-hot correctness mask that reinforces every accurate expert. This feedback loop provides the router with a clear training signal while accelerating expert specialization. We evaluate DriftMoE's performance across nine state-of-the-art data stream learning benchmarks spanning abrupt, gradual, and real-world drifts testing two distinct configurations: one where experts specialize on data regimes (multi-class variant), and another where they focus on single-class specialization (task-based variant). Our results demonstrate that DriftMoE achieves competitive results with state-of-the-art stream learning adaptive ensembles, offering a principled and efficient approach to concept drift adaptation. All code, data pipelines, and reproducibility scripts are available in our public GitHub repository: https://github.com/miguel-ceadar/drift-moe.", "authors": ["Miguel Aspis", "Sebastián A. Cajas Ordónez", "Andrés L. Suárez-Cetrulo", "Ricardo Simón Carbajo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2025-07-24", "url": "https://arxiv.org/abs/2507.18464", "pdf_url": "https://arxiv.org/pdf/2507.18464v1", "arxiv_id": "2507.18464", "doi": "10.48550/arXiv.2507.18464", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/miguel-ceadar/drift-moe", "venue": "arXiv.org", "quality_score": 0.3772}
{"id": "435806fa0752e3dbb60f6fadb1541967db9349da9edfca6e1a3d6ed36b9ea3c4", "sources": ["arxiv"], "title": "Mammo-Mamba: A Hybrid State-Space and Transformer Architecture with Sequential Mixture of Experts for Multi-View Mammography", "abstract": "Breast cancer (BC) remains one of the leading causes of cancer-related mortality among women, despite recent advances in Computer-Aided Diagnosis (CAD) systems. Accurate and efficient interpretation of multi-view mammograms is essential for early detection, driving a surge of interest in Artificial Intelligence (AI)-powered CAD models. While state-of-the-art multi-view mammogram classification models are largely based on Transformer architectures, their computational complexity scales quadratically with the number of image patches, highlighting the need for more efficient alternatives. To address this challenge, we propose Mammo-Mamba, a novel framework that integrates Selective State-Space Models (SSMs), transformer-based attention, and expert-driven feature refinement into a unified architecture. Mammo-Mamba extends the MambaVision backbone by introducing the Sequential Mixture of Experts (SeqMoE) mechanism through its customized SecMamba block. The SecMamba is a modified MambaVision block that enhances representation learning in high-resolution mammographic images by enabling content-adaptive feature refinement. These blocks are integrated into the deeper stages of MambaVision, allowing the model to progressively adjust feature emphasis through dynamic expert gating, effectively mitigating the limitations of traditional Transformer models. Evaluated on the CBIS-DDSM benchmark dataset, Mammo-Mamba achieves superior classification performance across all key metrics while maintaining computational efficiency.", "authors": ["Farnoush Bayatmakou", "Reza Taleei", "Nicole Simone", "Arash Mohammadi"], "categories": ["eess.IV", "cs.CV", "cs.LG"], "fields_of_study": [], "published_date": "2025-07-23", "url": "https://arxiv.org/abs/2507.17662", "pdf_url": "https://arxiv.org/pdf/2507.17662v1", "arxiv_id": "2507.17662", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1546}
{"id": "da0eb5f0399a5652be860f5bc0c2d4be313f172a37cd353c0f461cbb06bb6cb2", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic-DINO: Fine-Grained Mixture of Experts Tuning for Real-time Open-Vocabulary Object Detection", "abstract": "The Mixture of Experts (MoE) architecture has excelled in Large Vision-Language Models (LVLMs), yet its potential in real-time open-vocabulary object detectors, which also leverage large-scale vision-language datasets but smaller models, remains unexplored. This work investigates this domain, revealing intriguing insights. In the shallow layers, experts tend to cooperate with diverse peers to expand the search space. While in the deeper layers, fixed collaborative structures emerge, where each expert maintains 2-3 fixed partners and distinct expert combinations are specialized in processing specific patterns. Concretely, we propose Dynamic-DINO, which extends Grounding DINO 1.5 Edge from a dense model to a dynamic inference framework via an efficient MoE-Tuning strategy. Additionally, we design a granularity decomposition mechanism to decompose the Feed-Forward Network (FFN) of base model into multiple smaller expert networks, expanding the subnet search space. To prevent performance degradation at the start of fine-tuning, we further propose a pre-trained weight allocation strategy for the experts, coupled with a specific router initialization. During inference, only the input-relevant experts are activated to form a compact subnet. Experiments show that, pretrained with merely 1.56M open-source data, Dynamic-DINO outperforms Grounding DINO 1.5 Edge, pretrained on the private Grounding20M dataset.", "authors": ["Yehao Lu", "Minghe Weng", "Zekang Xiao", "Rui Jiang", "Wei Su", "Guangcong Zheng", "Ping Lu", "Xi Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-23", "url": "https://arxiv.org/abs/2507.17436", "pdf_url": "https://arxiv.org/pdf/2507.17436v1", "arxiv_id": "2507.17436", "doi": "10.1109/ICCV51701.2025.01938", "citation_count": 7, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.3754}
{"id": "80e054d79131d4e887205f1c3dd82ed36ca872e99c90d96776aca68acf3a302b", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Market Intelligence: A Mixture of Experts Framework for Volatility-Sensitive Stock Forecasting", "abstract": "This study develops and empirically validates a Mixture of Experts (MoE) framework for stock price prediction across heterogeneous volatility regimes using real market data. The proposed model combines a Recurrent Neural Network (RNN) optimized for high-volatility stocks with a linear regression model tailored to stable equities. A volatility-aware gating mechanism dynamically weights the contributions of each expert based on asset classification. Using a dataset of 30 publicly traded U.S. stocks spanning diverse sectors, the MoE approach consistently outperforms both standalone models. Specifically, it achieves up to 33% improvement in MSE for volatile assets and 28% for stable assets relative to their respective baselines. Stratified evaluation across volatility classes demonstrates the model's ability to adapt complexity to underlying market dynamics. These results confirm that no single model suffices across market regimes and highlight the advantage of adaptive architectures in financial prediction. Future work should explore real-time gate learning, dynamic volatility segmentation, and applications to portfolio optimization.", "authors": ["Diego Vallarino"], "categories": ["q-fin.ST", "econ.EM"], "fields_of_study": ["Economics"], "published_date": "2025-07-22", "url": "https://arxiv.org/abs/2508.02686", "pdf_url": "https://arxiv.org/pdf/2508.02686v1", "arxiv_id": "2508.02686", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1539}
{"id": "92c8a90e68547206631ae6129a49b113b11ea371af8b142d6a9d4de8b3b69db7", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking LLM Inference Bottlenecks: Insights from Latent Attention and Mixture-of-Experts", "abstract": "Computational workloads composing traditional transformer models are starkly bifurcated. Multi-Head Attention (MHA) and Grouped-Query Attention are memory-bound due to low arithmetic intensity, while FeedForward Networks are compute-bound. This dichotomy has long motivated research into specialized hardware to mitigate the attention bottleneck. This paper argues that recent architectural advances in transformer models -- Multi-head Latent Attention (MLA) and Mixture of Experts (MoE) -- introduce new dominant bottlenecks, shifting the challenge away from memory-intensive attention. We make two key observations. First, the arithmetic intensity of MLA is over two orders of magnitude higher than that of MHA, moving it toward a compute-bound regime well-matched to modern accelerators such as GPUs. Second, distributing MoE experts across a pool of accelerators allows batching to tune their arithmetic intensity to that of dense layers, producing a more balanced computational profile. Consequently, the focus of hardware and system optimization should shift from attention acceleration to high-bandwidth interconnects and balancing expert workloads across accelerators.", "authors": ["Sungmin Yun", "Seonyong Park", "Hwayong Nam", "Younjoo Lee", "Gunjun Lee", "Kwanhee Kyung", "Sangpyo Kim", "Nam Sung Kim", "Jongmin Kim", "Hyungyo Kim", "Juhwan Cho", "Seungmin Baek", "Jung Ho Ahn"], "categories": ["cs.AR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-21", "url": "https://arxiv.org/abs/2507.15465", "pdf_url": "https://arxiv.org/pdf/2507.15465v3", "arxiv_id": "2507.15465", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1531}
{"id": "efa8e1d9e03dc4c8011c66ed21097fe621a8ad90d41948c994b86aa6ce237014", "sources": ["arxiv", "semantic_scholar"], "title": "R^2MoE: Redundancy-Removal Mixture of Experts for Lifelong Concept Learning", "abstract": "Enabling large-scale generative models to continuously learn new visual concepts is essential for personalizing pre-trained models to meet individual user preferences. Existing approaches for continual visual concept learning are constrained by two fundamental challenges: catastrophic forgetting and parameter expansion. In this paper, we propose Redundancy-Removal Mixture of Experts (R^2MoE), a parameter-efficient framework for lifelong visual concept learning that effectively learns new concepts while incurring minimal parameter overhead. Our framework includes three key innovative contributions: First, we propose a mixture-of-experts framework with a routing distillation mechanism that enables experts to acquire concept-specific knowledge while preserving the gating network's routing capability, thereby effectively mitigating catastrophic forgetting. Second, we propose a strategy for eliminating redundant layer-wise experts that reduces the number of expert parameters by fully utilizing previously learned experts. Third, we employ a hierarchical local attention-guided inference approach to mitigate interference between generated visual concepts. Extensive experiments have demonstrated that our method generates images with superior conceptual fidelity compared to the state-of-the-art (SOTA) method, achieving an impressive 87.8\\% reduction in forgetting rates and 63.3\\% fewer parameters on the CustomConcept 101 dataset. Our code is available at {https://github.com/learninginvision/R2MoE}", "authors": ["Xiaohan Guo", "Yusong Cai", "Zejia Liu", "Zhengning Wang", "Lili Pan", "Hongliang Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-17", "url": "https://arxiv.org/abs/2507.13107", "pdf_url": "https://arxiv.org/pdf/2507.13107v1", "arxiv_id": "2507.13107", "doi": "10.48550/arXiv.2507.13107", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/learninginvision/R2MoE}", "venue": "arXiv.org", "quality_score": 0.3648}
{"id": "1d015157979a7d473fe10ab7be3153a77f7c7b0827b71b280d89bef254c41acd", "sources": ["arxiv", "semantic_scholar"], "title": "CorrMoE: Mixture of Experts with De-stylization Learning for Cross-Scene and Cross-Domain Correspondence Pruning", "abstract": "Establishing reliable correspondences between image pairs is a fundamental task in computer vision, underpinning applications such as 3D reconstruction and visual localization. Although recent methods have made progress in pruning outliers from dense correspondence sets, they often hypothesize consistent visual domains and overlook the challenges posed by diverse scene structures. In this paper, we propose CorrMoE, a novel correspondence pruning framework that enhances robustness under cross-domain and cross-scene variations. To address domain shift, we introduce a De-stylization Dual Branch, performing style mixing on both implicit and explicit graph features to mitigate the adverse influence of domain-specific representations. For scene diversity, we design a Bi-Fusion Mixture of Experts module that adaptively integrates multi-perspective features through linear-complexity attention and dynamic expert routing. Extensive experiments on benchmark datasets demonstrate that CorrMoE achieves superior accuracy and generalization compared to state-of-the-art methods. The code and pre-trained models are available at https://github.com/peiwenxia/CorrMoE.", "authors": ["Peiwen Xia", "Tangfei Liao", "Wei Zhu", "Danhuai Zhao", "Jianjun Ke", "Kaihao Zhang", "Tong Lu", "Tao Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-16", "url": "https://arxiv.org/abs/2507.11834", "pdf_url": "https://arxiv.org/pdf/2507.11834v1", "arxiv_id": "2507.11834", "doi": "10.48550/arXiv.2507.11834", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/peiwenxia/CorrMoE", "venue": "European Conference on Artificial Intelligence", "quality_score": 0.363}
{"id": "561763d7db457c6b87ccfb2ec92db8ffae43c6f1bbd9e59e660364652c7ebfeb", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Raytraced Experts", "abstract": "We introduce a Mixture of Raytraced Experts, a stacked Mixture of Experts (MoE) architecture which can dynamically select sequences of experts, producing computational graphs of variable width and depth. Existing MoE architectures generally require a fixed amount of computation for a given sample. Our approach, in contrast, yields predictions with increasing accuracy as the computation cycles through the experts' sequence. We train our model by iteratively sampling from a set of candidate experts, unfolding the sequence akin to how Recurrent Neural Networks are trained. Our method does not require load-balancing mechanisms, and preliminary experiments show a reduction in training epochs of 10\\% to 40\\% with a comparable/higher accuracy. These results point to new research directions in the field of MoEs, allowing the design of potentially faster and more expressive models. The code is available at https://github.com/nutig/RayTracing", "authors": ["Andrea Perin", "Giacomo Lagomarsini", "Claudio Gallicchio", "Giuseppe Nuti"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-16", "url": "https://arxiv.org/abs/2507.12419", "pdf_url": "https://arxiv.org/pdf/2507.12419v1", "arxiv_id": "2507.12419", "doi": "10.48550/arXiv.2507.12419", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/nutig/RayTracing", "venue": "arXiv.org", "quality_score": 0.363}
{"id": "8f4a1c3f547790d836d0f9a0573c10755bc11f55bf596054b7ad1a800f3f7ee4", "sources": ["arxiv", "semantic_scholar"], "title": "Astro-MoE: Mixture of Experts for Multiband Astronomical Time Series", "abstract": "Multiband astronomical time series exhibit heterogeneous variability patterns, sampling cadences, and signal characteristics across bands. Standard transformers apply shared parameters to all bands, potentially limiting their ability to model this rich structure. In this work, we introduce Astro-MoE, a foundational transformer architecture that enables dynamic processing via a Mixture of Experts module. We validate our model on both simulated (ELAsTiCC-1) and real-world datasets (Pan-STARRS1).", "authors": ["Martina Cádiz-Leyton", "Guillermo Cabrera-Vives", "Pavlos Protopapas", "Daniel Moreno-Cartagena", "Ignacio Becker"], "categories": ["astro-ph.IM"], "fields_of_study": ["Physics"], "published_date": "2025-07-16", "url": "https://arxiv.org/abs/2507.12611", "pdf_url": "https://arxiv.org/pdf/2507.12611v1", "arxiv_id": "2507.12611", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1495}
{"id": "0e1bb301cc79c6ed9c3b4510337115ec7961a6c2cc44664b23e5563c5565dcbb", "sources": ["arxiv", "semantic_scholar"], "title": "KV-Latent: Dimensional-level KV Cache Reduction with Frequency-aware Rotary Positional Embedding", "abstract": "Large language models (LLMs) based on Transformer Decoders have become the preferred choice for conversational generative AI. Despite the overall superiority of the Decoder architecture, the gradually increasing Key-Value (KV) cache during inference has emerged as a primary efficiency bottleneck, both in aspects of memory consumption and data transfer bandwidth limitations. To address these challenges, we propose a paradigm called KV-Latent. By down-sampling the Key-Value vector dimensions into a latent space, we can significantly reduce the KV Cache footprint and improve inference speed, only with a small amount of extra training, less than 1\\% of pre-training takes. Besides, we enhanced the stability of Rotary Positional Embedding applied on lower-dimensional vectors by modifying its frequency sampling mechanism, avoiding noise introduced by higher frequencies while retaining position attenuation. Our experiments, including both models with Grouped Query Attention and those without, have yielded satisfactory results. Finally, we conducted comparative experiments to study the impact of separately reducing Key and Value components on model's performance. Our approach allows for the construction of more efficient language model systems, and opens the new possibility on KV Cache saving and efficient LLMs. Our code is available at https://github.com/ShiLuohe/KV-Latent.", "authors": ["Luohe Shi", "Zuchao Li", "Lefei Zhang", "Guoming Liu", "Baoyuan Qi", "Hai Zhao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-15", "url": "https://arxiv.org/abs/2507.11273", "pdf_url": "https://arxiv.org/pdf/2507.11273v1", "arxiv_id": "2507.11273", "doi": "10.18653/v1/2025.acl-long.77", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ShiLuohe/KV-Latent", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3612}
{"id": "b198c8b38d8176010f8b4b79d2fa7e28c7e5e4e3e0d5d88d360b915d9076d871", "sources": ["arxiv", "semantic_scholar"], "title": "Oneiros: KV Cache Optimization through Parameter Remapping for Multi-tenant LLM Serving", "abstract": "KV cache accelerates LLM inference by avoiding redundant computation, at the expense of memory. To support larger KV caches, prior work extends GPU memory with CPU memory via CPU-offloading. This involves swapping KV cache between GPU and CPU memory. However, because the cache updates dynamically, such swapping incurs high CPU memory traffic. We make a key observation that model parameters remain constant during runtime, unlike the dynamically updated KV cache. Building on this, we introduce Oneiros, which avoids KV cache swapping by remapping, and thereby repurposing, the memory allocated to model parameters for KV cache. This parameter remapping is especially beneficial in multi-tenant environments, where the memory used for the parameters of the inactive models can be more aggressively reclaimed. Exploiting the high CPU-GPU bandwidth offered by the modern hardware, such as the NVIDIA Grace Hopper Superchip, we show that Oneiros significantly outperforms state-of-the-art solutions, achieving a reduction of 44.8%-82.5% in tail time-between-token latency, 20.7%-99.3% in tail time-to-first-token latency, and 6.6%-86.7% higher throughput compared to vLLM. Source code of Oneiros is available at https://github.com/UT-SysML/Oneiros/.", "authors": ["Ruihao Li", "Shagnik Pal", "Vineeth Narayan Pullu", "Prasoon Sinha", "Jeeho Ryoo", "Lizy K. John", "Neeraja J. Yadwadkar"], "categories": ["cs.OS"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-15", "url": "https://arxiv.org/abs/2507.11507", "pdf_url": "https://arxiv.org/pdf/2507.11507v2", "arxiv_id": "2507.11507", "doi": "10.1145/3772052.3772215", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/UT-SysML/Oneiros/", "venue": "ACM Symposium on Cloud Computing", "quality_score": 0.3612}
{"id": "9e452af04149c17fb2edf43217870e77e4f1b29193cdfe55d4a9ad65d5b36461", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts in Large Language Models", "abstract": "This paper presents a comprehensive review of the Mixture-of-Experts (MoE) architecture in large language models, highlighting its ability to significantly enhance model performance while maintaining minimal computational overhead. Through a systematic analysis spanning theoretical foundations, core architectural designs, and large language model (LLM) applications, we examine expert gating and routing mechanisms, hierarchical and sparse MoE configurations, meta-learning approaches, multimodal and multitask learning scenarios, real-world deployment cases, and recent advances and challenges in deep learning. Our analysis identifies key advantages of MoE, including superior model capacity compared to equivalent Bayesian approaches, improved task-specific performance, and the ability to scale model capacity efficiently. We also underscore the importance of ensuring expert diversity, accurate calibration, and reliable inference aggregation, as these are essential for maximizing the effectiveness of MoE architectures. Finally, this review outlines current research limitations, open challenges, and promising future directions, providing a foundation for continued innovation in MoE architecture and its applications.", "authors": ["Danyang Zhang", "Junhao Song", "Ziqian Bi", "Xinyuan Song", "Yingfang Yuan", "Tianyang Wang", "Joe Yeong", "Junfeng Hao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-15", "url": "https://arxiv.org/abs/2507.11181", "pdf_url": "https://arxiv.org/pdf/2507.11181v2", "arxiv_id": "2507.11181", "doi": "10.48550/arXiv.2507.11181", "citation_count": 17, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3138}
{"id": "8bcdf15bf185bb22abf8c4b894fb9759d91231f6720bf79c63ef7a20f3bebf99", "sources": ["arxiv", "semantic_scholar"], "title": "LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models", "abstract": "Recent advancements in Large Language Models (LLMs) have spurred interest in numerous applications requiring robust long-range capabilities, essential for processing extensive input contexts and continuously generating extended outputs. As sequence lengths increase, the number of Key-Value (KV) pairs in LLMs escalates, creating a significant efficiency bottleneck. In this paper, we propose a new KV cache optimization paradigm called LaCache, a training-free method for efficient and accurate generative inference of LLMs. LaCache enables LLMs to simultaneously address both of the critical challenges in long-range modeling: robust long-range capabilities and continuous generation without running out-of-memory (OOM). Specifically, LaCache integrates two key innovations: (1) a ladder-shaped KV cache pattern that stores KV pairs not only sequentially (left-to-right within each layer) but also across layers (from shallow to deep), providing an extended span for capturing long-range dependencies under a fixed storage budget, thereby boosting long-range capabilities; and (2) an iterative compaction mechanism that progressively compresses older caches, freeing up space for new tokens within a fixed cache size. This token distance-based dynamic compression enables more effective continuous generation under constrained cache budgets. Experiments across various tasks, benchmarks, and LLM models consistently validate LaCache's effectiveness in enhancing LLMs' long-range capabilities. Our code is available at https://github.com/GATECH-EIC/LaCache.", "authors": ["Dachuan Shi", "Yonggan Fu", "Xiangchi Yuan", "Zhongzhi Yu", "Haoran You", "Sixu Li", "Xin Dong", "Jan Kautz", "Pavlo Molchanov", " Yingyan", " Lin"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-14", "url": "https://arxiv.org/abs/2507.14204", "pdf_url": "https://arxiv.org/pdf/2507.14204v1", "arxiv_id": "2507.14204", "doi": "10.48550/arXiv.2507.14204", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/GATECH-EIC/LaCache", "venue": "International Conference on Machine Learning", "quality_score": 0.3595}
{"id": "67743f2b5f17ccf08fd7db1534641d87e69ec19187f90c584b6f94d5b5c41c4c", "sources": ["arxiv", "semantic_scholar"], "title": "MixLoRA-DSI: Dynamically Expandable Mixture-of-LoRA Experts for Rehearsal-Free Generative Retrieval over Dynamic Corpora", "abstract": "Continually updating model-based indexes in generative retrieval with new documents remains challenging, as full retraining is computationally expensive and impractical under resource constraints. We propose MixLoRA-DSI, a novel framework that combines an expandable mixture of Low-Rank Adaptation experts with a layer-wise out-of-distribution (OOD)-driven expansion strategy. Instead of allocating new experts for each new corpus, our proposed expansion strategy enables sublinear parameter growth by selectively introducing new experts only when significant number of OOD documents are detected. Experiments on NQ320k and MS MARCO Passage demonstrate that MixLoRA-DSI outperforms full-model update baselines, with minimal parameter overhead and substantially lower training costs.", "authors": ["Tuan-Luc Huynh", "Thuy-Trang Vu", "Weiqing Wang", "Trung Le", "Dragan Gašević", "Yuan-Fang Li", "Thanh-Toan Do"], "categories": ["cs.IR", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-14", "url": "https://arxiv.org/abs/2507.09924", "pdf_url": "https://arxiv.org/pdf/2507.09924v1", "arxiv_id": "2507.09924", "doi": "10.48550/arXiv.2507.09924", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2326}
{"id": "a91209008ec95d57718a32a56b6830be3ff7a7dcb27dd32cef604efa61cf0be0", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-residual Mixture of Experts Learning for Cooperative Control in Multi-vehicle Systems", "abstract": "Autonomous vehicles (AVs) are becoming increasingly popular, with their applications now extending beyond just a mode of transportation to serving as mobile actuators of a traffic flow to control flow dynamics. This contrasts with traditional fixed-location actuators, such as traffic signals, and is referred to as Lagrangian traffic control. However, designing effective Lagrangian traffic control policies for AVs that generalize across traffic scenarios introduces a major challenge. Real-world traffic environments are highly diverse, and developing policies that perform robustly across such diverse traffic scenarios is challenging. It is further compounded by the joint complexity of the multi-agent nature of traffic systems, mixed motives among participants, and conflicting optimization objectives subject to strict physical and external constraints. To address these challenges, we introduce Multi-Residual Mixture of Expert Learning (MRMEL), a novel framework for Lagrangian traffic control that augments a given suboptimal nominal policy with a learned residual while explicitly accounting for the structure of the traffic scenario space. In particular, taking inspiration from residual reinforcement learning, MRMEL augments a suboptimal nominal AV control policy by learning a residual correction, but at the same time dynamically selects the most suitable nominal policy from a pool of nominal policies conditioned on the traffic scenarios and modeled as a mixture of experts. We validate MRMEL using a case study in cooperative eco-driving at signalized intersections in Atlanta, Dallas Fort Worth, and Salt Lake City, with real-world data-driven traffic scenarios. The results show that MRMEL consistently yields superior performance-achieving an additional 4%-9% reduction in aggregate vehicle emissions relative to the strongest baseline in each setting.", "authors": ["Vindula Jayawardana", "Sirui Li", "Yashar Farid", "Cathy Wu"], "categories": ["cs.RO", "cs.AI", "cs.LG", "cs.MA", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-07-14", "url": "https://arxiv.org/abs/2507.09836", "pdf_url": "https://arxiv.org/pdf/2507.09836v1", "arxiv_id": "2507.09836", "doi": "10.48550/arXiv.2507.09836", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2326}
{"id": "1a2a83dcc930ace82c77fbad847f38f36b236c6282e68afc79070e4223957ef5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of LoRA Experts with Multi-Modal and Multi-Granularity LLM Generative Error Correction for Accented Speech Recognition", "abstract": "Despite improvements in automatic speech recognition, performance drops with accented speech. Generative error correction (GER) leverages the linguistic knowledge of large language models (LLMs), outperforming typical language model methods. However, it lacks specificity in accented speech scenarios. Accents represent deviations from standard pronunciation, making multi-granularity pronunciation and semantic information essential for accented speech recognition. Moreover, accents exhibit considerable diversity, with each accent possessing distinct characteristics. In this study, we leverage GER to improve transcription accuracy by addressing the two primary features. We propose the multi-modal GER, which integrates pronunciation information from the speech modality, and the multi-granularity GER, which incorporates fine-grained phoneme-level pronunciation information. These methods enable the LLM to utilize the pronunciation information of accented speech and the semantic information from word-level hypotheses for accurate transcription predictions through low-rank adaptation (LoRA) fine-tuning. We employ a three-stage strategy to train separate multi-modal GER models for each accent to obtain mono-accent LoRA experts. By adopting our proposed HDMoLE method, which incorporates hierarchical routing and dynamic thresholds within the mixture of LoRA experts, we effectively merge mono-accent LoRA experts within a single multi-modal GER to overcome accent diversity challenges. Furthermore, multi-granularity GER leverages N-best word-level and phoneme-level hypotheses from the HDMoLE model to predict final transcriptions. Experiments on a multi-accent English dataset show that our methods reduce word error rate by 67.35% compared to the baseline vanilla Whisper-large-v3 model.", "authors": ["Bingshen Mu", "Kun Wei", "Pengcheng Guo", "Lei Xie"], "categories": ["cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-07-12", "url": "https://arxiv.org/abs/2507.09116", "pdf_url": "https://arxiv.org/pdf/2507.09116v3", "arxiv_id": "2507.09116", "doi": "10.1109/TASLPRO.2025.3589856", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Audio, Speech, and Language Processing", "quality_score": 0.2603}
{"id": "11d88a53b71cb6d00b0303fa74c728526b62924af26c49253e4dbf160d54c51d", "sources": ["arxiv", "semantic_scholar"], "title": "MAPEX: Modality-Aware Pruning of Experts for Remote Sensing Foundation Models", "abstract": "Remote sensing data is commonly used for tasks such as flood mapping, wildfire detection, or land-use studies. For each task, scientists carefully choose appropriate modalities or leverage data from purpose-built instruments. Recent work on remote sensing foundation models pre-trains computer vision models on large amounts of remote sensing data. These large-scale models tend to focus on specific modalities, often optical RGB or multispectral data. For many important applications, this introduces a mismatch between the application modalities and the pre-training data. Moreover, the large size of foundation models makes them expensive and difficult to fine-tune on typically small datasets for each task. We address this mismatch with MAPEX, a remote sensing foundation model based on mixture-of-modality experts. MAPEX is pre-trained on multi-modal remote sensing data with a novel modality-conditioned token routing mechanism that elicits modality-specific experts. To apply the model on a specific task, we propose a modality aware pruning technique, which only retains experts specialized for the task modalities. This yields efficient modality-specific models while simplifying fine-tuning and deployment for the modalities of interest. We experimentally validate MAPEX on diverse remote sensing datasets and show strong performance compared to fully supervised training and state-of-the-art remote sensing foundation models. Code is available at https://github.com/HSG-AIML/MAPEX.", "authors": ["Joelle Hanna", "Linus Scheibenreif", "Damian Borth"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-10", "url": "https://arxiv.org/abs/2507.07527", "pdf_url": "https://arxiv.org/pdf/2507.07527v1", "arxiv_id": "2507.07527", "doi": "10.1109/TGRS.2026.3652100", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/HSG-AIML/MAPEX", "venue": "IEEE Transactions on Geoscience and Remote Sensing", "quality_score": 0.3524}
{"id": "4206aff2c8e9c792b281a0c776c258896058fd0d544a6c1a2aff0c4d8c094794", "sources": ["arxiv", "semantic_scholar"], "title": "Krul: Efficient State Restoration for Multi-turn Conversations with Dynamic Cross-layer KV Sharing", "abstract": "Efficient state restoration in multi-turn conversations with large language models (LLMs) remains a critical challenge, primarily due to the overhead of recomputing or loading full key-value (KV) caches for all historical tokens. To address this, existing approaches compress KV caches across adjacent layers with highly similar attention patterns. However, these methods often apply a fixed compression scheme across all conversations, selecting the same layer pairs for compression without considering conversation-specific attention dynamics. This static strategy overlooks variability in attention pattern similarity across different conversations, which can lead to noticeable accuracy degradation. We present Krul, a multi-turn LLM inference system that enables accurate and efficient KV cache restoration. Krul dynamically selects compression strategies based on attention similarity across layer pairs and uses a recomputation-loading pipeline to restore the KV cache. It introduces three key innovations: 1) a preemptive compression strategy selector to preserve critical context for future conversation turns and selects a customized strategy for the conversation; 2) a token-wise heterogeneous attention similarity estimator to mitigate the attention similarity computation and storage overhead during model generation; 3) a bubble-free restoration scheduler to reduce potential bubbles brought by the imbalance of recomputing and loading stream due to compressed KV caches. Empirical evaluations on real-world tasks demonstrate that Krul achieves a 1.5x-2.68x reduction in time-to-first-token (TTFT) and a 1.33x-2.35x reduction in KV cache storage compared to state-of-the-art methods without compromising generation quality.", "authors": ["Junyi Wen", "Junyuan Liang", "Zicong Hong", "Wuhui Chen", "Ting Cai", "Zibin Zheng"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-10", "url": "https://arxiv.org/abs/2507.08045", "pdf_url": "https://arxiv.org/pdf/2507.08045v2", "arxiv_id": "2507.08045", "doi": "10.48550/arXiv.2507.08045", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.228}
{"id": "2e8d358311d6e34a7b6f09c8d45f18188dc609faa915590611d7aeab5aca80ff", "sources": ["arxiv", "semantic_scholar"], "title": "SpindleKV: A Novel KV Cache Reduction Method Balancing Both Shallow and Deep Layers", "abstract": "Large Language Models (LLMs) have achieved impressive accomplishments in recent years. However, the increasing memory consumption of KV cache has possessed a significant challenge to the inference system. Eviction methods have revealed the inherent redundancy within the KV cache, demonstrating its potential for reduction, particularly in deeper layers. However, KV cache reduction for shallower layers has been found to be insufficient. Based on our observation that, the KV cache exhibits a high degree of similarity. Based on this observation, we proposed a novel KV cache reduction method, SpindleKV, which balances both shallow and deep layers. For deep layers, we employ an attention weight based eviction method, while for shallow layers, we apply a codebook based replacement approach which is learnt by similarity and merging policy. Moreover, SpindleKV addressed the Grouped-Query Attention (GQA) dilemma faced by other attention based eviction methods. Experiments on two common benchmarks with three different LLMs shown that SpindleKV obtained better KV cache reduction effect compared to baseline methods, while preserving similar or even better model performance.", "authors": ["Zicong Tang", "Shi Luohe", "Zuchao Li", "Baoyuan Qi", "Guoming Liu", "Lefei Zhang", "Ping Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-09", "url": "https://arxiv.org/abs/2507.06517", "pdf_url": "https://arxiv.org/pdf/2507.06517v1", "arxiv_id": "2507.06517", "doi": "10.18653/v1/2025.acl-long.1380", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2386}
{"id": "e06e854e266ce4592915da5f2ec83a5f2ad5dadd74e768e7144c4ccf2314bce4", "sources": ["arxiv", "semantic_scholar"], "title": "MoFE-Time: Mixture of Frequency Domain Experts for Time-Series Forecasting Models", "abstract": "As a prominent data modality task, time series forecasting plays a pivotal role in diverse applications. With the remarkable advancements in Large Language Models (LLMs), the adoption of LLMs as the foundational architecture for time series modeling has gained significant attention. Although existing models achieve some success, they rarely both model time and frequency characteristics in a pretraining-finetuning paradigm leading to suboptimal performance in predictions of complex time series, which requires both modeling periodicity and prior pattern knowledge of signals. We propose MoFE-Time, an innovative time series forecasting model that integrates time and frequency domain features within a Mixture of Experts (MoE) network. Moreover, we use the pretraining-finetuning paradigm as our training framework to effectively transfer prior pattern knowledge across pretraining and finetuning datasets with different periodicity distributions. Our method introduces both frequency and time cells as experts after attention modules and leverages the MoE routing mechanism to construct multidimensional sparse representations of input signals. In experiments on six public benchmarks, MoFE-Time has achieved new state-of-the-art performance, reducing MSE and MAE by 6.95% and 6.02% compared to the representative methods Time-MoE. Beyond the existing evaluation benchmarks, we have developed a proprietary dataset, NEV-sales, derived from real-world business scenarios. Our method achieves outstanding results on this dataset, underscoring the effectiveness of the MoFE-Time model in practical commercial applications.", "authors": ["Yiwen Liu", "Chenyu Zhang", "Junjie Song", "Siqi Chen", "Sun Yin", "Zihan Wang", "Lingming Zeng", "Yuji Cao", "Junming Jiao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-09", "url": "https://arxiv.org/abs/2507.06502", "pdf_url": "https://arxiv.org/pdf/2507.06502v1", "arxiv_id": "2507.06502", "doi": "10.48550/arXiv.2507.06502", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2269}
{"id": "e3b1dabf1647f8d4859a0f0ad70f2482c927a1934797d212f8499cdc45494d59", "sources": ["arxiv", "semantic_scholar"], "title": "SlimCaching: Edge Caching of Mixture-of-Experts for Distributed Inference", "abstract": "Mixture-of-Experts (MoE) models improve the scalability of large language models (LLMs) by activating only a small subset of relevant experts per input. However, the sheer number of expert networks in an MoE model introduces a significant storage/memory burden for an edge device. To address this challenge, we consider a scenario where experts are dispersed across an edge network for distributed inference. Based on the popular Top-$K$ expert selection strategy, we formulate a latency minimization problem by optimizing expert caching on edge servers under storage constraints. When $K=1$, the problem reduces to a monotone submodular maximization problem with knapsack constraints, for which we design a greedy-based algorithm with a $(1 - 1/e)$-approximation guarantee. For the general case where $K\\geq1$, expert co-activation within the same MoE layer introduces non-submodularity, which renders greedy methods ineffective. To tackle this issue, we propose a successive greedy decomposition method to decompose the original problem into a series of subproblems, with each being solved by a dynamic programming approach. Furthermore, we design an accelerated algorithm based on the max-convolution technique to obtain the approximate solution with a provable guarantee in polynomial time. Simulation results on various MoE models demonstrate that our method significantly reduces inference latency compared to existing baselines.", "authors": ["Qian Chen", "Xianhao Chen", "Kaibin Huang"], "categories": ["cs.LG", "cs.DC", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-09", "url": "https://arxiv.org/abs/2507.06567", "pdf_url": "https://arxiv.org/pdf/2507.06567v3", "arxiv_id": "2507.06567", "doi": "10.1109/TMC.2026.3664415", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Mobile Computing", "quality_score": 0.25}
{"id": "d65d513064e4a46aca7b763e029a586c9da6cd87654f2b1fe6b6ff02bd781299", "sources": ["arxiv", "semantic_scholar"], "title": "Differential Mamba", "abstract": "Sequence models like Transformers and RNNs often overallocate attention to irrelevant context, leading to noisy intermediate representations. This degrades LLM capabilities by promoting hallucinations, weakening long-range and retrieval abilities, and reducing robustness. Recent work has shown that differential design can mitigate this issue in Transformers, improving their effectiveness across various applications. In this paper, we explore whether these techniques, originally developed for Transformers, can be applied to Mamba, a recent architecture based on selective state-space layers that achieves Transformer-level performance with greater efficiency. We show that a naive adaptation of differential design to Mamba is insufficient and requires careful architectural modifications. To address this, we introduce a novel differential mechanism for Mamba, empirically validated on language modeling benchmarks, demonstrating improved retrieval capabilities and superior performance over vanilla Mamba. Finally, we conduct extensive ablation studies and empirical analyses to justify our design choices and provide evidence that our approach effectively mitigates the overallocation problem in Mamba-based models. Our code is publicly available: https://github.com/NadavSc/Diff-Mamba", "authors": ["Nadav Schneider", "Itamar Zimerman", "Eliya Nachmani"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-08", "url": "https://arxiv.org/abs/2507.06204", "pdf_url": "https://arxiv.org/pdf/2507.06204v2", "arxiv_id": "2507.06204", "doi": "10.48550/arXiv.2507.06204", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/NadavSc/Diff-Mamba", "venue": null, "quality_score": 0.2668}
{"id": "8640f735b154435ead086f9af4e2234ddfa45e1c08aefde2b2354636cc891f8e", "sources": ["arxiv", "semantic_scholar"], "title": "FedPhD: Federated Pruning with Hierarchical Learning of Diffusion Models", "abstract": "Federated Learning (FL), as a distributed learning paradigm, trains models over distributed clients' data. FL is particularly beneficial for distributed training of Diffusion Models (DMs), which are high-quality image generators that require diverse data. However, challenges such as high communication costs and data heterogeneity persist in training DMs similar to training Transformers and Convolutional Neural Networks. Limited research has addressed these issues in FL environments. To address this gap and challenges, we introduce a novel approach, FedPhD, designed to efficiently train DMs in FL environments. FedPhD leverages Hierarchical FL with homogeneity-aware model aggregation and selection policy to tackle data heterogeneity while reducing communication costs. The distributed structured pruning of FedPhD enhances computational efficiency and reduces model storage requirements in clients. Our experiments across multiple datasets demonstrate that FedPhD achieves high model performance regarding Fréchet Inception Distance (FID) scores while reducing communication costs by up to $88\\%$. FedPhD outperforms baseline methods achieving at least a $34\\%$ improvement in FID, while utilizing only $56\\%$ of the total computation and communication resources.", "authors": ["Qianyu Long", "Qiyuan Wang", "Christos Anagnostopoulos", "Daning Bi"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-08", "url": "https://arxiv.org/abs/2507.06449", "pdf_url": "https://arxiv.org/pdf/2507.06449v1", "arxiv_id": "2507.06449", "doi": "10.48550/arXiv.2507.06449", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2257}
{"id": "89e16bcc24fbc58769b0e74c04fb3b178d72ba3aba52e2c195dbd9733bf06933", "sources": ["arxiv", "semantic_scholar"], "title": "QMoE: A Quantum Mixture of Experts Framework for Scalable Quantum Neural Networks", "abstract": "Quantum machine learning (QML) has emerged as a promising direction in the noisy intermediate-scale quantum (NISQ) era, offering computational and memory advantages by harnessing superposition and entanglement. However, QML models often face challenges in scalability and expressiveness due to hardware constraints. In this paper, we propose quantum mixture of experts (QMoE), a novel quantum architecture that integrates the mixture of experts (MoE) paradigm into the QML setting. QMoE comprises multiple parameterized quantum circuits serving as expert models, along with a learnable quantum routing mechanism that selects and aggregates specialized quantum experts per input. The empirical results from the proposed QMoE on quantum classification tasks demonstrate that it consistently outperforms standard quantum neural networks, highlighting its effectiveness in learning complex data patterns. Our work paves the way for scalable and interpretable quantum learning frameworks.", "authors": ["Hoang-Quan Nguyen", "Xuan-Bac Nguyen", "Sankalp Pandey", "Samee U. Khan", "Ilya Safro", "Khoa Luu"], "categories": ["quant-ph", "cs.CV"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2025-07-07", "url": "https://arxiv.org/abs/2507.05190", "pdf_url": "https://arxiv.org/pdf/2507.05190v1", "arxiv_id": "2507.05190", "doi": "10.1109/QCE65121.2025.10323", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Quantum Computing and Engineering", "quality_score": 0.2246}
{"id": "c43ef72391bece4d3e1b671c082e47d2e8ac6124c23dc9a0d4ed558d186e4be6", "sources": ["arxiv", "semantic_scholar"], "title": "MCST-Mamba: Multivariate Mamba-Based Model for Traffic Prediction", "abstract": "Accurate traffic prediction plays a vital role in intelligent transportation systems by enabling efficient routing, congestion mitigation, and proactive traffic control. However, forecasting is challenging due to the combined effects of dynamic road conditions, varying traffic patterns across different locations, and external influences such as weather and accidents. Traffic data often consists of several interrelated measurements - such as speed, flow and occupancy - yet many deep-learning approaches either predict only one of these variables or require a separate model for each. This limits their ability to capture joint patterns across channels. To address this, we introduce the Multi-Channel Spatio-Temporal (MCST) Mamba model, a forecasting framework built on the Mamba selective state-space architecture that natively handles multivariate inputs and simultaneously models all traffic features. The proposed MCST-Mamba model integrates adaptive spatio-temporal embeddings and separates the modeling of temporal sequences and spatial sensor interactions into two dedicated Mamba blocks, improving representation learning. Unlike prior methods that evaluate on a single channel, we assess MCST-Mamba across all traffic features at once, aligning more closely with how congestion arises in practice. Our results show that MCST-Mamba achieves strong predictive performance with a lower parameter count compared to baseline models.", "authors": ["Mohamed Hamad", "Mohamed Mabrok", "Nizar Zorba"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-05", "url": "https://arxiv.org/abs/2507.03927", "pdf_url": "https://arxiv.org/pdf/2507.03927v1", "arxiv_id": "2507.03927", "doi": "10.1109/GLOBECOM59602.2025.11432511", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Global Communications Conference", "quality_score": 0.2223}
{"id": "7b9dc73071cc766c5efb917c0483d7f3bf791822016a02b89b1a7755227dc26f", "sources": ["arxiv", "semantic_scholar"], "title": "Neural Inhibition Improves Dynamic Routing and Mixture of Experts", "abstract": "To be effective, efficient, and diverse, deep learning models need to dynamically choose its architecture based on signals from a population of neurons. We hypothesize dynamic routing models can be improved with neural inhibition in those neural populations. This means signals commonly shared among the various modes of data statistics can be inhibited so that the routing model can choose a specialized expert path for each data sample. Only through inhibition is the routing mechanism able to effectively select neural pathways. We believe this is an under-studied and under-verified implementation methodology for Mixture-of-Experts, dynamic routing, and transformer language models. We provide experimental evidence that the neural inhibition algorithm significantly boosts the performance of general tasks and motivates more effort to be invested in this research direction.", "authors": ["Will Y. Zou", "Jennifer Y. Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-03", "url": "https://arxiv.org/abs/2507.03221", "pdf_url": "https://arxiv.org/pdf/2507.03221v1", "arxiv_id": "2507.03221", "doi": "10.48550/arXiv.2507.03221", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.22}
{"id": "bb25080f7a735f0e561f5cf4b3f9432e0ce47cb19f4c7273f00a10684c1b7e96", "sources": ["arxiv", "semantic_scholar"], "title": "High-Layer Attention Pruning with Rescaling", "abstract": "Pruning is a highly effective approach for compressing large language models (LLMs), significantly reducing inference latency. However, conventional training-free structured pruning methods often employ a heuristic metric that indiscriminately removes some attention heads across all pruning layers, without considering their positions within the network architecture. In this work, we propose a novel pruning algorithm that strategically prunes attention heads in the model's higher layers. Since the removal of attention heads can alter the magnitude of token representations, we introduce an adaptive rescaling parameter that calibrates the representation scale post-pruning to counteract this effect. We conduct comprehensive experiments on a wide range of LLMs, including LLaMA3.1-8B, Mistral-7B-v0.3, Qwen2-7B, and Gemma2-9B. Our evaluation includes both generation and discriminative tasks across 27 datasets. The results consistently demonstrate that our method outperforms existing structured pruning methods. This improvement is particularly notable in generation tasks, where our approach significantly outperforms existing baselines. Code is available at https://github.com/SongtaoLiu0823/HARP.", "authors": ["Songtao Liu", "Peng Liu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-02", "url": "https://arxiv.org/abs/2507.01900", "pdf_url": "https://arxiv.org/pdf/2507.01900v2", "arxiv_id": "2507.01900", "doi": "10.48550/arXiv.2507.01900", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SongtaoLiu0823/HARP", "venue": null, "quality_score": 0.2586}
{"id": "88b4929ce3fa64e88bcb7a6cb98d330179ecf1becf761313b93ebec5a666881f", "sources": ["arxiv", "semantic_scholar"], "title": "Mixtures of Neural Network Experts with Application to Phytoplankton Flow Cytometry Data", "abstract": "Flow cytometry is a valuable technique that measures the optical properties of particles at a single-cell resolution. When deployed in the ocean, flow cytometry allows oceanographers to study different types of photosynthetic microbes called phytoplankton. It is of great interest to study how phytoplankton properties change in response to environmental conditions. In our work, we develop a nonlinear mixture of experts model to estimate separate regression functions for each subpopulation utilizing random-weight neural networks. Our model allows one to flexibly estimate how cell properties and relative abundances depend on environmental covariates in each segment of a heterogeneous sample, without the computational burden of backpropagation. We show that the proposed model provides superior predictive performance in simulated examples compared to a mixture of linear experts. Also, applying our model to real data, we show that our model has (1) comparable out-of-sample prediction performance, and (2) more realistic estimates of phytoplankton behavior.", "authors": ["Ethan Pawl", "François Ribalet", "Paul A. Parker", "Sangwon Hyun"], "categories": ["stat.ME", "stat.AP"], "fields_of_study": ["Mathematics"], "published_date": "2025-07-02", "url": "https://arxiv.org/abs/2507.01375", "pdf_url": "https://arxiv.org/pdf/2507.01375v3", "arxiv_id": "2507.01375", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1393}
{"id": "7cd54a395177fecee6728bc81753a16ada909157fc5aa75f158efa60414d45e8", "sources": ["arxiv", "semantic_scholar"], "title": "MoNE: Replacing Redundant Experts with Lightweight Novices for Structured Pruning of MoE", "abstract": "Mixture-of-Experts (MoE) enables efficient scaling of large language models by activating only a subset of experts per input token. However, deploying MoE-based models incurs significant memory overhead due to the need to retain all experts in memory. While structured pruning is promising to reduce memory costs, existing methods often show suboptimal performance and unstable degradation in three dimensions: model architectures, calibration data sources, and calibration sample sizes. This paper proposes Mixture-of-Novices-and-Experts (MoNE), a novel expert pruning method that replaces redundant experts with lightweight novices to achieve effective and robust model compression. MoNE evaluates expert redundancy based on two metrics: access frequency and output variance. Experts exhibiting low usage and stable outputs are pruned and replaced with lightweight novices-unbiased estimations of their original outputs-minimizing performance degradation. Extensive experiments demonstrate that MoNE consistently outperforms baseline methods with minimal accuracy degradation across the three dimensions, confirming its effectiveness and robustness. Notably, it outperforms baselines by up to 2.72 for the average zero shot accuracy across nine downstream tasks under 25% pruning ratio, with only 0.14 performance drop for Qwen2-57B-A14B. The code is available at https://github.com/zxgx/mode-pd.", "authors": ["Geng Zhang", "Yuxuan Han", "Yuxuan Lou", "Yiqi Zhang", "Wangbo Zhao", "Yang You"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-01", "url": "https://arxiv.org/abs/2507.00390", "pdf_url": "https://arxiv.org/pdf/2507.00390v2", "arxiv_id": "2507.00390", "doi": "10.48550/arXiv.2507.00390", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/zxgx/mode-pd", "venue": "arXiv.org", "quality_score": 0.3365}
{"id": "c4486bcc610bbe3596505f717f590e87a56a445448b2f9768d9831044bc9cbba", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba-FETrack V2: Revisiting State Space Model for Frame-Event based Visual Object Tracking", "abstract": "Combining traditional RGB cameras with bio-inspired event cameras for robust object tracking has garnered increasing attention in recent years. However, most existing multimodal tracking algorithms depend heavily on high-complexity Vision Transformer architectures for feature extraction and fusion across modalities. This not only leads to substantial computational overhead but also limits the effectiveness of cross-modal interactions. In this paper, we propose an efficient RGB-Event object tracking framework based on the linear-complexity Vision Mamba network, termed Mamba-FETrack V2. Specifically, we first design a lightweight Prompt Generator that utilizes embedded features from each modality, together with a shared prompt pool, to dynamically generate modality-specific learnable prompt vectors. These prompts, along with the modality-specific embedded features, are then fed into a Vision Mamba-based FEMamba backbone, which facilitates prompt-guided feature extraction, cross-modal interaction, and fusion in a unified manner. Finally, the fused representations are passed to the tracking head for accurate target localization. Extensive experimental evaluations on multiple RGB-Event tracking benchmarks, including short-term COESOT dataset and long-term datasets, i.e., FE108 and FELT V2, demonstrate the superior performance and efficiency of the proposed tracking framework. The source code and pre-trained models will be released on https://github.com/Event-AHU/Mamba_FETrack", "authors": ["Shiao Wang", "Ju Huang", "Qingchuan Ma", "Jinfeng Gao", "Chunyi Xu", "Xiao Wang", "Lan Chen", "Bo Jiang"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-30", "url": "https://arxiv.org/abs/2506.23783", "pdf_url": "https://arxiv.org/pdf/2506.23783v1", "arxiv_id": "2506.23783", "doi": "10.48550/arXiv.2506.23783", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Event-AHU/Mamba_FETrack", "venue": "arXiv.org", "quality_score": 0.3347}
{"id": "7f759f14df81d99c9f15c9070ed72eafc402ff157161665a867af981c4ff0572", "sources": ["arxiv", "semantic_scholar"], "title": "Sub-MoE: Efficient Mixture-of-Expert LLMs Compression via Subspace Expert Merging", "abstract": "Mixture of Experts (MoE) LLMs face significant obstacles due to their massive parameter scale, which imposes memory, storage, and deployment challenges. Although recent expert merging methods promise greater efficiency by consolidating multiple experts, they are fundamentally hindered by parameter conflicts arising from expert specialization. In this paper, we present Sub-MoE, a novel MoE compression framework via Subspace Expert Merging. Our key insight is to perform joint Singular Value Decomposition (SVD) on concatenated expert weights, reducing conflicting parameters by extracting shared $U$-matrices while enabling effective merging of the expert-specific $V$ components. Specifically, Sub-MoE consists of two innovative phases: (1) Adaptive Expert Clustering, which groups functionally coherent experts via K-means clustering based on cosine similarity of expert outputs; and (2) Subspace Expert Merging, which first enforces Experts Union Decomposition to derive the shared $U$-matrix across experts in the same group, then pursues frequency-based merging for individual $V$-matrices, and finalizes expert reconstruction using the merged $V$-matrix. In this way, we align and fuse experts in a shared subspace, and can be extended with intra-expert compression for further inference optimization. Extensive experiments on Mixtral, DeepSeek, and Qwen-1.5|3 MoE LLMs demonstrate that our Sub-MoE significantly outperforms existing expert pruning and merging methods. Notably, our Sub-MoE maintains 96\\%|86\\% of original performance with 25\\%|50\\% expert reduction on Mixtral-8x7B in zero-shot benchmarks. Code will be released at https://github.com/lliai/MoERazor.", "authors": ["Lujun Li", "Zhu Qiyuan", "Jiacheng Wang", "Wei Li", "Hao Gu", "Sirui Han", "Yike Guo"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-29", "url": "https://arxiv.org/abs/2506.23266", "pdf_url": "https://arxiv.org/pdf/2506.23266v1", "arxiv_id": "2506.23266", "doi": "10.48550/arXiv.2506.23266", "citation_count": 19, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/lliai/MoERazor", "venue": "arXiv.org", "quality_score": 0.3329}
{"id": "14a87f1f7e3d9cf09468b4ab4e7e29641928400bf959941ab4a679d7dc30eaaa", "sources": ["arxiv", "semantic_scholar"], "title": "QuickSilver -- Speeding up LLM Inference through Dynamic Token Halting, KV Skipping, Contextual Token Fusion, and Adaptive Matryoshka Quantization", "abstract": "Inference accounts for the majority of latency and energy consumption in large language model (LLM) deployments, often exceeding 90% of total cost. While training-time efficiency has seen extensive progress, runtime optimization remains a key bottleneck, particularly under autoregressive decoding. Existing approaches -- such as pruning, quantization, early exits, and speculative decoding -- often require retraining, architectural changes, or disrupt decoding compatibility. We introduce QuickSilver, a modular, token-level framework that enables semantic adaptivity at inference time without altering model weights or structure. QuickSilver integrates four synergistic mechanisms: (i) Dynamic Token Halting, which halts computation for tokens with converged representations; (ii) KV Cache Skipping, which selectively suppresses memory writes to reduce attention overhead; and (iii) Contextual Token Fusion, which collapses redundant tokens into shared paths to shrink sequence length. Unlike speculative decoding or MoE routing, QuickSilver operates entirely on frozen, dense models and requires no auxiliary networks. Applied to GPT-2 and Llama-2 across WikiText-103 and C4, QuickSilver achieves up to 39.6% FLOP reduction with negligible perplexity degradation (<=0.2).", "authors": ["Danush Khanna", "Aditya Kumar Guru", "Srivarshinee Sridhar", "Zidan Ahmed", "Rubhav Bahirwani", "Meetu Malhotra", "Vinija Jain", "Aman Chadha", "Amitava Das", "Kripabandhu Ghosh"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-27", "url": "https://arxiv.org/abs/2506.22396", "pdf_url": "https://arxiv.org/pdf/2506.22396v1", "arxiv_id": "2506.22396", "doi": "10.48550/arXiv.2506.22396", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2131}
{"id": "3358b6341012f9f51d0327dc3c65043c732884cc77987e744e268ef0ee5fb680", "sources": ["arxiv", "semantic_scholar"], "title": "Little by Little: Continual Learning via Incremental Mixture of Rank-1 Associative Memory Experts", "abstract": "Continual learning (CL) with large pre-trained models aims to incrementally acquire knowledge without catastrophic forgetting. Existing LoRA-based Mixture-of-Experts (MoE) methods expand capacity by adding isolated new experts while freezing old ones, but still suffer from redundancy, interference, routing ambiguity, and consequent forgetting. We investigate the issues stemming from coarse-grained expert granularity. Coarse-grained experts (e.g., high-rank LoRA) encode low-specialty information, leading to expert duplication/interference and routing degradation/confusion as experts accumulate. In this work, we propose MoRAM (Mixture of Rank-1 Associative Memory). Grounded in the view that weight matrices act as linear associative memories, MoRAM achieves CL as incremental expansion of reusable atomic rank-1 experts as memory. Each rank-1 adapter acts as a fine-grained MoE expert or an associative memory unit. By viewing rank-1 experts as key-value memory pairs, we eliminate explicit MoE-LoRA routers with self-activation, where each memory atom evaluates its relevance via its intrinsic key. The inference process thus becomes a content-addressable retrieval and recall over the incrementally accumulated memory of learning snapshots. Extensive experiments on CLIP and LLMs show that MoRAM significantly outperforms state-of-the-art methods, achieving a better plasticity-stability trade-off, stronger generalization, and reduced forgetting. Project Page: https://artificer-ai-lab.github.io/MoRAM/.", "authors": ["Haodong Lu", "Chongyang Zhao", "Minhui Xue", "Lina Yao", "Kristen Moore", "Dong Gong"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-26", "url": "https://arxiv.org/abs/2506.21035", "pdf_url": "https://arxiv.org/pdf/2506.21035v6", "arxiv_id": "2506.21035", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1349}
{"id": "6f179a1a1f3c56dade41d53706858af52ae29f7b1e33568cd48dd0388fca964b", "sources": ["arxiv", "semantic_scholar"], "title": "PLoP: Precise LoRA Placement for Efficient Finetuning of Large Models", "abstract": "Low-Rank Adaptation (LoRA) is a widely used finetuning method for large models. Its small memory footprint allows practitioners to adapt large models to specific tasks at a fraction of the cost of full finetuning. Different modifications have been proposed to enhance its efficiency by, for example, setting the learning rate, the rank, and the initialization. Another improvement axis is adapter placement strategy: when using LoRA, practitioners usually pick module types to adapt with LoRA, such as Query and Key modules. Few works have studied the problem of adapter placement, with nonconclusive results: original LoRA paper suggested placing adapters in attention modules, while other works suggested placing them in the MLP modules. Through an intuitive theoretical analysis, we introduce PLoP (Precise LoRA Placement), a lightweight method that allows automatic identification of module types where LoRA adapters should be placed, given a pretrained model and a finetuning task. We demonstrate that PLoP consistently outperforms, and in the worst case competes, with commonly used placement strategies through comprehensive experiments on supervised finetuning and reinforcement learning for reasoning.", "authors": ["Soufiane Hayou", "Nikhil Ghosh", "Bin Yu"], "categories": ["cs.LG", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-06-25", "url": "https://arxiv.org/abs/2506.20629", "pdf_url": "https://arxiv.org/pdf/2506.20629v1", "arxiv_id": "2506.20629", "doi": "10.48550/arXiv.2506.20629", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "e30e083b31e83affa023265861dbab005ef30e070a0ebb43ce7017e077e68fb6", "sources": ["arxiv", "semantic_scholar"], "title": "Joint Quantization and Pruning Neural Networks Approach: A Case Study on FSO Receivers", "abstract": "Towards fast, hardware-efficient, and low-complexity receivers, we propose a compression-aware learning approach and examine it on free-space optical (FSO) receivers for turbulence mitigation. The learning approach jointly quantize, prune, and train a convolutional neural network (CNN). In addition, we propose to have the CNN weights of power of two values so we replace the multiplication operations bit-shifting operations in every layer that has significant lower computational cost. The compression idea in the proposed approach is that the loss function is updated and both the quantization levels and the pruning limits are optimized in every epoch of training. The compressed CNN is examined for two levels of compression (1-bit and 2-bits) over different FSO systems. The numerical results show that the compression approach provides negligible decrease in performance in case of 1-bit quantization and the same performance in case of 2-bits quantization, compared to the full-precision CNNs. In general, the proposed IM/DD FSO receivers show better bit-error rate (BER) performance (without the need for channel state information (CSI)) compared to the maximum likelihood (ML) receivers that utilize imperfect CSI when the DL model is compressed whether with 1-bit or 2-bit quantization.", "authors": ["Mohanad Obeed", "Ming Jian"], "categories": ["eess.SP"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2025-06-25", "url": "https://arxiv.org/abs/2506.20084", "pdf_url": "https://arxiv.org/pdf/2506.20084v1", "arxiv_id": "2506.20084", "doi": "10.1109/ICC52391.2025.11162011", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ICC 2025 - IEEE International Conference on Communications", "quality_score": 0.2108}
{"id": "43b33c9d42a764ea0c4a357a4f4d6808ff60176352962faf588a17907ce2f138", "sources": ["arxiv", "semantic_scholar"], "title": "AnTKV: Anchor Token-Aware Sub-Bit Vector Quantization for KV Cache in Large Language Models", "abstract": "Quantization has emerged as an effective and lightweight solution to reduce the memory footprint of the KV cache in Large Language Models. Nevertheless, minimizing the accuracy degradation caused by ultra-low-bit KV cache quantization remains a significant challenge. While scalar quantization is constrained by 1-bit bound, vector quantization exploits intra-vector correlations and enables sub-bit regimes, making it more suitable for ultra-low-bit quantization. To further mitigate quantization-induced degradation, we reveal that the degradation is highly uneven across tokens in attention quality. To investigate this unevenness, we introduce anchor score to measure each token's sensitivity to quantization. Our analysis and experiments show that preserving a small subset (1\\%) of tokens with the highest Anchor Score significantly mitigates accuracy loss under aggressive quantization. We propose AnTKV, a dual-stage framework that leverages anchor token-aware vector quantization to compress the KV cache. It combines offline token-aware centroids learning and online anchor token selection to balance compression and accuracy. To enable efficient deployment, we design an online anchor token selection kernel compatible with FlashAttention. It allows LLaMA3-8B to scale to 840K tokens on a single 80GB A100, while delivering up to $3.5\\times$ higher decoding throughput over the FP16 baseline. Experiments demonstrate that AnTKV matches or surpasses prior methods at 4-bit, and significantly reduce perplexity under ultra-low-bit quantization, achieving 6.32 at 1-bit on Mistral-7B, compared to 7.25 for CQ and 15.36 for KVQuant.", "authors": ["Zeyu Li", "Chuanfu Xiao", "Yang Wang", "Xiang Liu", "Zhenheng Tang", "Baotong Lu", "Mao Yang", "Xinyu Chen", "Xiaowen Chu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-24", "url": "https://arxiv.org/abs/2506.19505", "pdf_url": "https://arxiv.org/pdf/2506.19505v2", "arxiv_id": "2506.19505", "doi": "10.48550/arXiv.2506.19505", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2097}
{"id": "aef684bab1efa7b1082a7256748b564513c2e7600c0264024a1601df4248a37b", "sources": ["arxiv", "semantic_scholar"], "title": "Radial Attention: $O(n\\log n)$ Sparse Attention with Energy Decay for Long Video Generation", "abstract": "Recent advances in diffusion models have enabled high-quality video generation, but the additional temporal dimension significantly increases computational costs, making training and inference on long videos prohibitively expensive. In this paper, we identify a phenomenon we term Spatiotemporal Energy Decay in video diffusion models: post-softmax attention scores diminish as spatial and temporal distance between tokens increase, akin to the physical decay of signal or waves over space and time in nature. Motivated by this, we propose Radial Attention, a scalable sparse attention mechanism with $\\mathcal{O}(n \\log n)$ complexity that translates energy decay into exponentially decaying compute density, which is significantly more efficient than standard $\\mathcal{O}(n^2)$ dense attention and more expressive than linear attention. Specifically, Radial Attention employs a simple, static attention mask where each token attends to spatially nearby tokens, with the attention window size shrinking with temporal distance. Moreover, it allows pre-trained video diffusion models to extend their generation length with efficient LoRA-based fine-tuning. Extensive experiments show that Radial Attention maintains video quality across Wan2.1-14B, HunyuanVideo, and Mochi 1, achieving up to a 1.9$\\times$ speedup over the original dense attention. With minimal tuning, it enables video generation up to 4$\\times$ longer while reducing training costs by up to 4.4$\\times$ compared to direct fine-tuning and accelerating inference by up to 3.7$\\times$ compared to dense attention inference. Code is released at \\href{https://github.com/mit-han-lab/radial-attention}{https://github.com/mit-han-lab/radial-attention}.", "authors": ["Xingyang Li", "Muyang Li", "Tianle Cai", "Haocheng Xi", "Shuo Yang", "Yujun Lin", "Lvmin Zhang", "Songlin Yang", "Jinbo Hu", "Kelly Peng", "Maneesh Agrawala", "Ion Stoica", "Kurt Keutzer", "Song Han"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-24", "url": "https://arxiv.org/abs/2506.19852", "pdf_url": "https://arxiv.org/pdf/2506.19852v2", "arxiv_id": "2506.19852", "doi": "10.48550/arXiv.2506.19852", "citation_count": 46, "influential_citation_count": 11, "has_code": true, "code_url": "https://github.com/mit-han-lab/radial-attention", "venue": "arXiv.org", "quality_score": 0.5396}
{"id": "5738ba5872cc269c1014fbfae82aa354da3a9ef0ad9e0074f606295062d7953c", "sources": ["arxiv", "semantic_scholar"], "title": "CommVQ: Commutative Vector Quantization for KV Cache Compression", "abstract": "Large Language Models (LLMs) are increasingly used in applications requiring long context lengths, but the key-value (KV) cache often becomes a memory bottleneck on GPUs as context grows. To address this, we propose Commutative Vector Quantization (CommVQ) to significantly reduce memory usage for long-context LLM inference. We first introduce additive quantization with a lightweight encoder and codebook to compress the KV cache, which can be decoded via simple matrix multiplication. To further reduce computational costs during decoding, we design the codebook to be commutative with Rotary Position Embedding (RoPE) and train it using an Expectation-Maximization (EM) algorithm. This enables efficient integration of decoding into the self-attention mechanism. Our approach achieves high accuracy with additive quantization and low overhead via the RoPE-commutative codebook. Experiments on long-context benchmarks and GSM8K show that our method reduces FP16 KV cache size by 87.5% with 2-bit quantization, while outperforming state-of-the-art KV cache quantization methods. Notably, it enables 1-bit KV cache quantization with minimal accuracy loss, allowing a LLaMA-3.1 8B model to run with a 128K context length on a single RTX 4090 GPU. The source code is available at: https://github.com/UMass-Embodied-AGI/CommVQ.", "authors": ["Junyan Li", "Yang Zhang", "Muhammad Yusuf Hassan", "Talha Chafekar", "Tianle Cai", "Zhile Ren", "Pengsheng Guo", "Foroozan Karimzadeh", "Colorado Reed", "Chong Wang", "Chuang Gan"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-23", "url": "https://arxiv.org/abs/2506.18879", "pdf_url": "https://arxiv.org/pdf/2506.18879v1", "arxiv_id": "2506.18879", "doi": "10.48550/arXiv.2506.18879", "citation_count": 12, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/UMass-Embodied-AGI/CommVQ", "venue": "International Conference on Machine Learning", "quality_score": 0.3223}
{"id": "639f1b8ec7f235ae79d5c1e5cc8a5c57fb2966979cff080854aea250c218f5f3", "sources": ["arxiv", "semantic_scholar"], "title": "Chain-of-Experts: Unlocking the Communication Power of Mixture-of-Experts Models", "abstract": "We propose Chain-of-Experts (CoE), a new Mixture-of-Experts (MoE) architecture that introduces sequential expert communication within each layer. Unlike traditional MoE models, where experts operate independently in parallel, CoE processes tokens iteratively across a chain of experts inside a layer. To support dynamic expert selection across iterations, CoE employs a dedicated router at each iteration step within a layer. This design allows tokens to re-evaluate and select different experts during each iteration, rather than being statically assigned. As a result, CoE introduces a flexible routing mechanism that increases the diversity of expert combinations and enriches the model's representational capacity. CoE demonstrates improved performance under fixed compute: on math reasoning tasks, it reduces validation loss from 1.20 to 1.12 compared to a standard MoE. Beyond performance, CoE offers a new scaling axis: depth through expert iteration, which complements conventional width/depth scaling. For example, using 2x iterations matches the performance of 3x expert selections (in width), while reducing memory usage by 17.6-42% relative to other scaling strategies. Our analysis reveals that CoE's benefits stem from its iterative residual structure and enhanced expert specialization empowered by iterative routing, which together unlock more expressive representations. Code is available at https://github.com/ZihanWang314/coe.", "authors": ["Zihan Wang", "Rui Pan", "Jiarui Yao", "Robert Csordas", "Linjie Li", "Lu Yin", "Jiajun Wu", "Tong Zhang", "Manling Li", "Shiwei Liu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-23", "url": "https://arxiv.org/abs/2506.18945", "pdf_url": "https://arxiv.org/pdf/2506.18945v1", "arxiv_id": "2506.18945", "doi": "10.48550/arXiv.2506.18945", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ZihanWang314/coe", "venue": "arXiv.org", "quality_score": 0.3223}
{"id": "11b95420110725c6714786d224acf0369c9db2262d38d634ab8989797f80b464", "sources": ["arxiv", "semantic_scholar"], "title": "Shift Happens: Mixture of Experts based Continual Adaptation in Federated Learning", "abstract": "Federated Learning (FL) enables collaborative model training across decentralized clients without sharing raw data, yet faces significant challenges in real-world settings where client data distributions evolve dynamically over time. This paper tackles the critical problem of covariate and label shifts in streaming FL environments, where non-stationary data distributions degrade model performance and necessitate a middleware layer that adapts FL to distributional shifts. We introduce ShiftEx, a shift-aware mixture of experts framework that dynamically creates and trains specialized global models in response to detected distribution shifts using Maximum Mean Discrepancy for covariate shifts. The framework employs a latent memory mechanism for expert reuse and implements facility location-based optimization to jointly minimize covariate mismatch, expert creation costs, and label imbalance. Through theoretical analysis and comprehensive experiments on benchmark datasets, we demonstrate 5.5-12.9 percentage point accuracy improvements and 22-95 % faster adaptation compared to state-of-the-art FL baselines across diverse shift scenarios. The proposed approach offers a scalable, privacy-preserving middleware solution for FL systems operating in non-stationary, real-world conditions while minimizing communication and computational overhead.", "authors": ["Rahul Atul Bhope", "K. R. Jayaram", "Praveen Venkateswaran", "Nalini Venkatasubramanian"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-23", "url": "https://arxiv.org/abs/2506.18789", "pdf_url": "https://arxiv.org/pdf/2506.18789v2", "arxiv_id": "2506.18789", "doi": "10.1145/3721462.3770784", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Middleware Conference", "quality_score": 0.2085}
{"id": "70975b8c57bd95b226981727407b06668d67bc171c6a48aca965f3802035deb4", "sources": ["arxiv", "semantic_scholar"], "title": "Routing Mamba: Scaling State Space Models with Mixture-of-Experts Projection", "abstract": "Linear State Space Models (SSMs) offer remarkable performance gains in efficient sequence modeling, with constant inference-time computation and memory complexity. Recent advances, such as Mamba, further enhance SSMs with input-dependent gating and hardware-aware implementations, positioning them as strong alternatives to Transformers for long sequence modeling. However, efficiently scaling the expressive power of SSMs, particularly with Mixture of Experts (MoE), remains challenging, as naive integration attempts often falter or degrade performance. In this work, we introduce Routing Mamba (RoM), a novel approach that scales SSM parameters using sparse mixtures of linear projection experts. By sharing routing decisions between projection layers and lightweight sub-modules within Mamba across experts, RoM leverages synergies among linear projection experts for effective and efficient sparse scaling of Mamba layers. At a scale of 1.3B active parameters (10B total) and 16K training sequence length, RoM achieves language modeling performance equivalent to a dense Mamba model requiring over 2.3x more active parameters, and demonstrates consistent perplexity across context lengths. Experimental results further show RoM effectively scales hybrid language models, yielding a 23% FLOPS saving compared to dense Mamba scaling for similar performance.", "authors": ["Zheng Zhan", "Liliang Ren", "Shuohang Wang", "Liyuan Liu", "Yang Liu", "Yeyun Gong", "Yanzhi Wang", "Yelong Shen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-22", "url": "https://arxiv.org/abs/2506.18145", "pdf_url": "https://arxiv.org/pdf/2506.18145v1", "arxiv_id": "2506.18145", "doi": "10.48550/arXiv.2506.18145", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2074}
{"id": "3751b2e785d871796df33c2703e6bc118396b0fa82c630e25e01b85420a8bee8", "sources": ["arxiv", "semantic_scholar"], "title": "Memba: Membrane-driven Parameter-Efficient Fine-Tuning for Mamba", "abstract": "State Space Models (SSMs) have emerged as powerful alternatives to attention-based Transformers, with Mamba demonstrating impressive efficiency and scalability. As these models grow increasingly larger, the need for Parameter-Efficient Fine-Tuning (PEFT) methods becomes critical to adapt pre-trained Mamba to downstream tasks without prohibitive computational costs. However, previous approaches simply apply traditional Transformer-tailored PEFT methods without addressing the unique temporal processing dynamics of SSMs. To address this limitation, we propose Memba, a membrane-driven PEFT approach specifically designed for Mamba. Memba introduces Leaky Integrate Membrane (LIM) neurons as bio-inspired gating mechanisms that naturally accumulate membrane potentials over time, enhancing selective information retention. By strategically combining LIM neurons with Low-Rank Adaptations (LoRA) and cross-layer membrane transfer, our approach significantly improves Mamba's temporal modeling capabilities. Extensive experiments across language and vision tasks demonstrate that Memba achieves substantial improvements over existing PEFT methods. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/Memba.", "authors": ["Donghyun Lee", "Yuhang Li", "Ruokai Yin", "Shiting Xiao", "Priyadarshini Panda"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-22", "url": "https://arxiv.org/abs/2506.18184", "pdf_url": "https://arxiv.org/pdf/2506.18184v2", "arxiv_id": "2506.18184", "doi": "10.48550/arXiv.2506.18184", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Intelligent-Computing-Lab-Yale/Memba", "venue": "arXiv.org", "quality_score": 0.3205}
{"id": "e421dbd4d7a0ac14b23020df30151a749499a8b826c6b3a67b80b7b7fd0a894a", "sources": ["arxiv", "semantic_scholar"], "title": "Quantizing for Noisy Flash Memory Channels", "abstract": "Flash memory-based processing-in-memory (flash-based PIM) offers high storage capacity and computational efficiency but faces significant reliability challenges due to noise in high-density multi-level cell (MLC) flash memories. Existing verify level optimization methods are designed for general storage scenarios and fail to address the unique requirements of flash-based PIM systems, where metrics such as mean squared error (MSE) and peak signal-to-noise ratio (PSNR) are critical. This paper introduces an integrated framework that jointly optimizes quantization and verify levels to minimize the MSE, considering both quantization and flash memory channel errors. We develop an iterative algorithm to solve the joint optimization problem. Experimental results on quantized images and SwinIR model parameters stored in flash memory show that the proposed method significantly improves the reliability of flash-based PIM systems.", "authors": ["Juyun Oh", "Taewoo Park", "Jiwoong Im", "Yuval Cassuto", "Yongjune Kim"], "categories": ["cs.IT", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering", "Mathematics"], "published_date": "2025-06-21", "url": "https://arxiv.org/abs/2506.17646", "pdf_url": "https://arxiv.org/pdf/2506.17646v1", "arxiv_id": "2506.17646", "doi": "10.1109/ISIT63088.2025.11195707", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Symposium on Information Theory", "quality_score": 0.2063}
{"id": "4e3600171c939d2a119ed9b54845cfe0527614dc368d159c567b5fe8ed81f291", "sources": ["arxiv", "semantic_scholar"], "title": "Safe Pruning LoRA: Robust Distance-Guided Pruning for Safety Alignment in Adaptation of LLMs", "abstract": "Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enhances adaptability while reducing computational costs. However, fine-tuning can compromise safety alignment, even with benign data, increasing susceptibility to harmful outputs. Existing safety alignment methods struggle to capture complex parameter shifts, leading to suboptimal safety-utility trade-offs. To address this issue, we propose Safe Pruning LoRA (SPLoRA), a novel pruning-based approach that selectively removes LoRA layers that weaken safety alignment, improving safety while preserving performance. At its core, we introduce Empirical-DIEM (E-DIEM), a dimension-insensitive similarity metric that effectively detects safety misalignment in LoRA-adapted models. We conduct extensive experiments on LLMs fine-tuned with mixed of benign and malicious data, and purely benign datasets, evaluating SPLoRA across utility, safety, and reliability metrics. Results demonstrate that SPLoRA outperforms state-of-the-art safety alignment techniques, significantly reducing safety risks while maintaining or improving model performance and reliability. Additionally, SPLoRA reduces inference overhead, making it a scalable and efficient solution for deploying safer and more reliable LLMs. The code is available at https://github.com/AoShuang92/SPLoRA.", "authors": ["Shuang Ao", "Yi Dong", "Jinwei Hu", "Sarvapali Ramchurn"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-21", "url": "https://arxiv.org/abs/2506.18931", "pdf_url": "https://arxiv.org/pdf/2506.18931v1", "arxiv_id": "2506.18931", "doi": "10.48550/arXiv.2506.18931", "citation_count": 11, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/AoShuang92/SPLoRA", "venue": "Transactions of the Association for Computational Linguistics", "quality_score": 0.3187}
{"id": "71ab6b45a93b10ccb02a6528a1305d45b567af0bf706e7ea5592e37c15e72b92", "sources": ["arxiv", "semantic_scholar"], "title": "LoLA-SpecViT: Local Attention SwiGLU Vision Transformer with LoRA for Hyperspectral Imaging", "abstract": "Hyperspectral image classification remains a challenging task due to the high dimensionality of spectral data, significant inter-band redundancy, and the limited availability of annotated samples. While recent transformer-based models have improved the global modeling of spectral-spatial dependencies, their scalability and adaptability under label-scarce conditions remain limited. In this work, we propose \\textbf{LoLA-SpecViT}(Low-rank adaptation Local Attention Spectral Vision Transformer), a lightweight spectral vision transformer that addresses these limitations through a parameter-efficient architecture tailored to the unique characteristics of hyperspectral imagery. Our model combines a 3D convolutional spectral front-end with local window-based self-attention, enhancing both spectral feature extraction and spatial consistency while reducing computational complexity. To further improve adaptability, we integrate low-rank adaptation (LoRA) into attention and projection layers, enabling fine-tuning with over 80\\% fewer trainable parameters. A novel cyclical learning rate scheduler modulates LoRA adaptation strength during training, improving convergence and generalisation. Extensive experiments on three benchmark datasets WHU-Hi LongKou, WHU-Hi HongHu, and Salinas demonstrate that LoLA-SpecViT consistently outperforms state-of-the-art baselines, achieving up to 99.91\\% accuracy with substantially fewer parameters and enhanced robustness under low-label regimes. The proposed framework provides a scalable and generalizable solution for real-world HSI applications in agriculture, environmental monitoring, and remote sensing analytics. Our code is available in the following \\href{https://github.com/FadiZidiDz/LoLA-SpecViT}{GitHub Repository}.", "authors": ["Fadi Abdeladhim Zidi", "Djamel Eddine Boukhari", "Abdellah Zakaria Sellam", "Abdelkrim Ouafi", "Cosimo Distante", "Salah Eddine Bekhouche", "Abdelmalik Taleb-Ahmed"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-21", "url": "https://arxiv.org/abs/2506.17759", "pdf_url": "https://arxiv.org/pdf/2506.17759v1", "arxiv_id": "2506.17759", "doi": "10.1016/j.jag.2025.104924", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/FadiZidiDz/LoLA-SpecViT}{GitHub", "venue": "International Journal of Applied Earth Observation and Geoinformation", "quality_score": 0.3187}
{"id": "6eb9eb7bfe7c44ecd11a805d9476a810bb24310dc0a1675f2e163267eb5e194b", "sources": ["arxiv", "semantic_scholar"], "title": "Cache Me If You Can: How Many KVs Do You Need for Effective Long-Context LMs?", "abstract": "Language models handle increasingly long contexts for tasks such as book summarization, but this leads to growing memory costs for the key-value (KV) cache. Many prior works have proposed ways of discarding KVs from memory, but their approaches are tailored to favorable settings, obscuring caveats like high peak memory and performance degradation, and a fair comparison between methods is difficult. In this paper, we propose the *KV footprint* as a unified metric, which accounts for both the amount of KV entries stored and their lifespan in memory. We evaluate methods based on the smallest footprint they attain while preserving performance in both long-context understanding and generation, with context lengths of up to 128K tokens. This metric reveals the high peak memory of prior KV eviction methods. One class of methods -- *post-fill eviction* -- has a high footprint due to being incompatible with eviction during pre-filling. We adapt these methods to be able to evict KVs during pre-filling, achieving substantially lower KV footprints. We then turn to *recency eviction* methods, wherein we propose PruLong, an end-to-end optimization method for learning which attention heads need to retain the full KV cache and which do not. PruLong saves memory while preserving long-context performance, achieving 12% smaller KV footprint than prior methods while retaining performance in challenging recall tasks. Our paper clarifies the complex tangle of long-context inference methods and paves the way for future development to minimize the KV footprint.", "authors": ["Adithya Bhaskar", "Alexander Wettig", "Tianyu Gao", "Yihe Dong", "Danqi Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-20", "url": "https://arxiv.org/abs/2506.17121", "pdf_url": "https://arxiv.org/pdf/2506.17121v1", "arxiv_id": "2506.17121", "doi": "10.48550/arXiv.2506.17121", "citation_count": 10, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/princeton-pli/PruLong", "venue": "arXiv.org", "quality_score": 0.317}
{"id": "1259c20e3ea97aaaef5f65991e9354ccde771acaac698be62adcf119f1bae52b", "sources": ["arxiv", "semantic_scholar"], "title": "LazyEviction: Lagged KV Eviction with Attention Pattern Observation for Efficient Long Reasoning", "abstract": "Large Language Models (LLMs) exhibit enhanced capabilities by Chain-of-Thought reasoning. However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a Token Importance Recurrence phenomenon: a large proportion of tokens regain high attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose LazyEviction, an observation window-based lagged eviction framework retaining latent recurring tokens by prioritized eviction based on tokens' recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache by 50%~70% while maintaining comparable accuracy, outperforming existing KV cache compression baselines. Our implementation code can be found at https://github.com/Halo-949/LazyEviction.", "authors": ["Haoyue Zhang", "Hualei Zhang", "Xiaosong Ma", "Jie Zhang", "Song Guo"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-19", "url": "https://arxiv.org/abs/2506.15969", "pdf_url": "https://arxiv.org/pdf/2506.15969v3", "arxiv_id": "2506.15969", "doi": "10.48550/arXiv.2506.15969", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Halo-949/LazyEviction", "venue": "arXiv.org", "quality_score": 0.3152}
{"id": "29a75065654237ef44f88ecf2c68fcf4a0eaed4c60829ad3ff7417711db30285", "sources": ["arxiv", "semantic_scholar"], "title": "InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video Understanding", "abstract": "Modern multimodal large language models (MLLMs) can reason over hour-long video, yet their key-value (KV) cache grows linearly with time-quickly exceeding the fixed memory of phones, AR glasses, and edge robots. Prior compression schemes either assume the whole video and user query are available offline or must first build the full cache, so memory still scales with stream length. InfiniPot-V is the first training-free, query-agnostic framework that enforces a hard, length-independent memory cap for streaming video understanding. During video encoding it monitors the cache and, once a user-set threshold is reached, runs a lightweight compression pass that (i) removes temporally redundant tokens via Temporal-axis Redundancy (TaR) metric and (ii) keeps semantically significant tokens via Value-Norm (VaN) ranking. Across four open-source MLLMs and four long-video and streaming-video benchmarks, InfiniPot-V cuts peak GPU memory by up to 94%, sustains real-time generation, and matches or surpasses full-cache accuracy-even in multi-turn dialogues. By dissolving the KV cache bottleneck without retraining or query knowledge, InfiniPot-V closes the gap for on-device streaming video assistants.", "authors": ["Minsoo Kim", "Kyuhong Shim", "Jungwook Choi", "Simyung Chang"], "categories": ["eess.IV", "cs.LG"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-06-18", "url": "https://arxiv.org/abs/2506.15745", "pdf_url": "https://arxiv.org/pdf/2506.15745v2", "arxiv_id": "2506.15745", "doi": "10.48550/arXiv.2506.15745", "citation_count": 34, "influential_citation_count": 9, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5}
{"id": "6e4a88bfc8e97dda78cb527696537e860ddf5663ee4c28db795c7ec02dedc310", "sources": ["arxiv", "semantic_scholar"], "title": "Targeted Lexical Injection: Unlocking Latent Cross-Lingual Alignment in Lugha-Llama via Early-Layer LoRA Fine-Tuning", "abstract": "Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their performance in low-resource languages (LRLs), such as Swahili, often lags due to data scarcity and underrepresentation in pre-training. A key challenge is achieving robust cross-lingual lexical alignment, crucial for tasks like translation and cross-lingual information retrieval. This paper introduces Targeted Lexical Injection (TLI), a novel and efficient fine-tuning approach. We first demonstrate that Lugha-Llama-8B-wura, a Swahili-centric LLM, exhibits strong, near-perfect lexical alignment for Swahili-English word pairs in its early internal layers (specifically Layer 2, with ~0.99998 average cosine similarity based on a pilot study), a capability not fully reflected in its final output representations (baseline ~0.32 similarity on our evaluation set). TLI leverages this insight by using Low-Rank Adaptation (LoRA) and a contrastive learning objective to fine-tune the model, specifically targeting embeddings from this empirically identified optimal early layer. Our experiments show that TLI significantly improves the output-level lexical alignment for 623 trained Swahili-English word pairs, increasing average cosine similarity from 0.3211 to 0.4113 (+28.08%, p < 1.33 x 10^-240). More importantly, these improvements generalize remarkably well to 63 unseen control word pairs, with similarity increasing from 0.3143 to 0.4033 (+28.32%, p < 7.17 x 10^-27). These findings suggest TLI enhances the model's ability to preserve and propagate its inherent early-layer cross-lingual knowledge, offering a parameter-efficient and effective strategy for improving lexical alignment in LRL-focused LLMs.", "authors": ["Stanley Ngugi"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-18", "url": "https://arxiv.org/abs/2506.15415", "pdf_url": "https://arxiv.org/pdf/2506.15415v1", "arxiv_id": "2506.15415", "doi": "10.48550/arXiv.2506.15415", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2028}
{"id": "428826ea6ea1a9b6eb1162b5959a01983e430a586757288acbfd7b058fd4877e", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-Mixer: Coordinate Modular LoRA Experts Through Serial Attention Routing", "abstract": "Recent attempts to combine low-rank adaptation (LoRA) with mixture-of-experts (MoE) for multi-task adaptation of Large Language Models (LLMs) often replace whole attention/FFN layers with switch experts or append parallel expert branches, undermining parameter efficiency and limiting task specialization. We introduce LoRA-Mixer, a modular MoE framework that routes task-specific LoRA experts into the core projection matrices of the attention module, namely input and output linear layers, rather than primarily targeting FFN blocks. The design delivers fine-grained token-level specialization by fully exploiting the attention mechanism, while remaining drop-in compatible with Transformers and state-space models (SSMs), since linear projection layers are ubiquitous. To train robust routers from limited data while promoting stable, selective decisions and high expert reuse, LoRA-Mixer employs an adaptive Routing Specialization Loss (RSL) that jointly enforces global load balance and input-aware specialization via an entropy-shaping objective. The framework supports two regimes: (i) joint optimization of adapters and router with a differentiable hard-soft top-k routing scheme, and (ii) plug-and-play routing over frozen, pre-trained LoRA modules sourced from public repositories. Across 15 benchmarks, including MedQA, GSM8K, HumanEval, and GLUE, RSL-optimized LoRA-Mixer outperforms state-of-the-art routing and LoRA-MoE baselines while using 48 percent of their trainable parameters, with gains of 3.79, 2.90, and 3.95 percentage points on GSM8K, CoLA, and ARC-C, respectively. Cross-model transfer and adapter reuse experiments further demonstrate the approach's versatility and data efficiency. Our code is available at https://github.com/hustcselwb/LoRA-Mixer.", "authors": ["Wenbing Li", "Zikai Song", "Hang Zhou", "Yunyao Zhang", "Junqing Yu", "Wei Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-17", "url": "https://arxiv.org/abs/2507.00029", "pdf_url": "https://arxiv.org/pdf/2507.00029v2", "arxiv_id": "2507.00029", "doi": "10.48550/arXiv.2507.00029", "citation_count": 13, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/hustcselwb/LoRA-Mixer", "venue": "arXiv.org", "quality_score": 0.3117}
{"id": "0d907cc216d8f42f4201f8761a51c6a63d67b81474867f20cc48449a8cb75844", "sources": ["arxiv", "semantic_scholar"], "title": "GuiLoMo: Allocating Expert Number and Rank for LoRA-MoE via Bilevel Optimization with GuidedSelection Vectors", "abstract": "Parameter-efficient fine-tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA), offer an efficient way to adapt large language models with reduced computational costs. However, their performance is limited by the small number of trainable parameters. Recent work combines LoRA with the Mixture-of-Experts (MoE), i.e., LoRA-MoE, to enhance capacity, but two limitations remain in hindering the full exploitation of its potential: 1) the influence of downstream tasks when assigning expert numbers, and 2) the uniform rank assignment across all LoRA experts, which restricts representational diversity. To mitigate these gaps, we propose GuiLoMo, a fine-grained layer-wise expert numbers and ranks allocation strategy with GuidedSelection Vectors (GSVs). GSVs are learned via a prior bilevel optimization process to capture both model- and task-specific needs, and are then used to allocate optimal expert numbers and ranks. Experiments on three backbone models across diverse benchmarks show that GuiLoMo consistently achieves superior or comparable performance to all baselines. Further analysis offers key insights into how expert numbers and ranks vary across layers and tasks, highlighting the benefits of adaptive expert configuration. Our code is available at https://github.com/Liar406/Gui-LoMo.git.", "authors": ["Hengyuan Zhang", "Xinrong Chen", "Yingmin Qiu", "Xiao Liang", "Ziyue Li", "Guanyu Wang", "Weiping Li", "Tong Mo", "Hayden Kwok-Hay So", "Ngai Wong"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-17", "url": "https://arxiv.org/abs/2506.14646", "pdf_url": "https://arxiv.org/pdf/2506.14646v2", "arxiv_id": "2506.14646", "doi": "10.48550/arXiv.2506.14646", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Liar406/Gui-LoMo.git", "venue": "arXiv.org", "quality_score": 0.3117}
{"id": "84778f180fc7e5ba43ae14f1d31ab41028080b53aa2221c95adbce4813c13757", "sources": ["arxiv", "semantic_scholar"], "title": "MoTE: Mixture of Ternary Experts for Memory-efficient Large Multimodal Models", "abstract": "Large multimodal Mixture-of-Experts (MoEs) effectively scale the model size to boost performance while maintaining fixed active parameters. However, previous works primarily utilized full-precision experts during sparse up-cycling. Despite they show superior performance on end tasks, the large amount of experts introduces higher memory footprint, which poses significant challenges for the deployment on edge devices. In this work, we propose MoTE, a scalable and memory-efficient approach to train Mixture-of-Ternary-Experts models from dense checkpoint. Instead of training fewer high-precision experts, we propose to train more low-precision experts during up-cycling. Specifically, we use the pre-trained FFN as a shared expert and train ternary routed experts with parameters in {-1, 0, 1}. Extensive experiments show that our approach has promising scaling trend along model size. MoTE achieves comparable performance to full-precision baseline MoE-LLaVA while offering lower memory footprint. Furthermore, our approach is compatible with post-training quantization methods and the advantage further amplifies when memory-constraint goes lower. Given the same amount of expert memory footprint of 3.4GB and combined with post-training quantization, MoTE outperforms MoE-LLaVA by a gain of 4.3% average accuracy on end tasks, demonstrating its effectiveness and potential for memory-constrained devices.", "authors": ["Hongyu Wang", "Jiayu Xu", "Ruiping Wang", "Yan Feng", "Yitao Zhai", "Peng Pei", "Xunliang Cai", "Xilin Chen"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-17", "url": "https://arxiv.org/abs/2506.14435", "pdf_url": "https://arxiv.org/pdf/2506.14435v2", "arxiv_id": "2506.14435", "doi": "10.48550/arXiv.2506.14435", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2017}
{"id": "76c431d816f82b3ebc9856db5ec7231db1a735ddd6176e364e659ad299b92a7f", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring Speaker Diarization with Mixture of Experts", "abstract": "In this paper, we propose a novel neural speaker diarization system using memory-aware multi-speaker embedding with sequence-to-sequence architecture (NSD-MS2S), which integrates a memory-aware multi-speaker embedding module with a sequence-to-sequence architecture. The system leverages a memory module to enhance speaker embeddings and employs a Seq2Seq framework to efficiently map acoustic features to speaker labels. Additionally, we explore the application of mixture of experts in speaker diarization, and introduce a Shared and Soft Mixture of Experts (SS-MoE) module to further mitigate model bias and enhance performance. Incorporating SS-MoE leads to the extended model NSD-MS2S-SSMoE. Experiments on multiple complex acoustic datasets, including CHiME-6, DiPCo, Mixer 6 and DIHARD-III evaluation sets, demonstrate meaningful improvements in robustness and generalization. The proposed methods achieve state-of-the-art results, showcasing their effectiveness in challenging real-world scenarios.", "authors": ["Gaobin Yang", "Maokui He", "Shutong Niu", "Ruoyu Wang", "Hang Chen", "Jun Du"], "categories": ["cs.SD", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-17", "url": "https://arxiv.org/abs/2506.14750", "pdf_url": "https://arxiv.org/pdf/2506.14750v1", "arxiv_id": "2506.14750", "doi": "10.48550/arXiv.2506.14750", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2017}
{"id": "21a305d99a7043d294bd761fc3e0a6dcfaa33e43c1d538324cc9378e3ce75e0d", "sources": ["arxiv", "semantic_scholar"], "title": "Improving LoRA with Variational Learning", "abstract": "Bayesian methods have recently been used to improve LoRA finetuning and, although they improve calibration, their effect on other metrics (such as accuracy) is marginal and can sometimes even be detrimental. Moreover, Bayesian methods also increase computational overheads and require additional tricks for them to work well. Here, we fix these issues by using a recently proposed variational algorithm called IVON. We show that IVON is easy to implement and has similar costs to AdamW, and yet it can also drastically improve many metrics by using a simple posterior pruning technique. We present extensive results on billion-scale LLMs (Llama and Qwen series) going way beyond the scale of existing applications of IVON. For example, we finetune a Llama-3.2-3B model on a set of commonsense reasoning tasks and improve accuracy over AdamW by 1.3% and reduce ECE by 5.4%, outperforming AdamW and other recent Bayesian methods like Laplace-LoRA and BLoB. Overall, our results show that variational learning with IVON can effectively improve LoRA finetuning.", "authors": ["Bai Cong", "Nico Daheim", "Yuesong Shen", "Rio Yokota", "Mohammad Emtiyaz Khan", "Thomas Möllenhoff"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-06-17", "url": "https://arxiv.org/abs/2506.14280", "pdf_url": "https://arxiv.org/pdf/2506.14280v1", "arxiv_id": "2506.14280", "doi": "10.48550/arXiv.2506.14280", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2017}
{"id": "07e6d5687d878eb0630934e97541c1ef810ff4690b3b78b86abdcb9558bd38a0", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Weight-shared Heterogeneous Group Attention Experts for Dynamic Token-wise KV Optimization", "abstract": "Transformer models face scalability challenges in causal language modeling (CLM) due to inefficient memory allocation for growing key-value (KV) caches, which strains compute and storage resources. Existing methods like Grouped Query Attention (GQA) and token-level KV optimization improve efficiency but rely on rigid resource allocation, often discarding \"low-priority\" tokens or statically grouping them, failing to address the dynamic spectrum of token importance. We propose mixSGA, a novel mixture-of-expert (MoE) approach that dynamically optimizes token-wise computation and memory allocation. Unlike prior approaches, mixSGA retains all tokens while adaptively routing them to specialized experts with varying KV group sizes, balancing granularity and efficiency. Our key novelties include: (1) a token-wise expert-choice routing mechanism guided by learned importance scores, enabling proportional resource allocation without token discard; (2) weight-sharing across grouped attention projections to minimize parameter overhead; and (3) an auxiliary loss to ensure one-hot routing decisions for training-inference consistency in CLMs. Extensive evaluations across Llama3, TinyLlama, OPT, and Gemma2 model families show mixSGA's superiority over static baselines. On instruction-following and continued pretraining tasks, mixSGA achieves higher ROUGE-L and lower perplexity under the same KV budgets.", "authors": ["Guanghui Song", "Dongping Liao", "Yiren Zhao", "Kejiang Ye", "Cheng-zhong Xu", "Xitong Gao"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-16", "url": "https://arxiv.org/abs/2506.13541", "pdf_url": "https://arxiv.org/pdf/2506.13541v1", "arxiv_id": "2506.13541", "doi": "10.48550/arXiv.2506.13541", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2005}
{"id": "d5c66741fb8caad467dbdbd21b80574faf1b2918595ecbb5db4567b1ddaea712", "sources": ["arxiv", "semantic_scholar"], "title": "Load Balancing Mixture of Experts with Similarity Preserving Routers", "abstract": "Sparse Mixture of Experts (MoE) models offer a scalable and efficient architecture for training large neural networks by activating only a subset of parameters (\"experts\") for each input. A learned router computes a distribution over these experts, and assigns input tokens to a small subset. However, without auxiliary balancing mechanisms, routers often converge to using only a few experts, severely limiting model capacity and degrading performance. Most current load balancing mechanisms encourage a distribution over experts that resembles a roughly uniform distribution of experts per token. During training, this can result in inconsistent routing behavior, resulting in the model spending its capacity to learn redundant knowledge. We address this by introducing a novel load balancing loss that preserves token-wise relational structure, encouraging consistent expert choices for similar inputs during training. Our experimental results show that applying our loss to the router results in 36% faster convergence and lower redundancy compared to a popular load balancing loss.", "authors": ["Nabil Omi", "Siddhartha Sen", "Ali Farhadi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-16", "url": "https://arxiv.org/abs/2506.14038", "pdf_url": "https://arxiv.org/pdf/2506.14038v2", "arxiv_id": "2506.14038", "doi": "10.48550/arXiv.2506.14038", "citation_count": 20, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3306}
{"id": "ecc4958fb51fcf1e958702d726b0cc8d15004f602491b0fe8c5244cd696a82fc", "sources": ["arxiv", "semantic_scholar"], "title": "EAQuant: Enhancing Post-Training Quantization for MoE Models via Expert-Aware Optimization", "abstract": "Mixture-of-Experts (MoE) models enable scalable computation and performance in large-scale deep learning but face quantization challenges due to sparse expert activation and dynamic routing. Existing post-training quantization (PTQ) methods fail to address activation outliers, routing instability, and sparse expert calibration, leading to significant performance degradation. To address this, we propose EAQuant, a PTQ framework tailored for MoE architectures. Our method introduces three expert-aware innovations: (1) smoothing aggregation to suppress activation outliers, (2) routing consistency alignment to preserve expert selection post-quantization, and (3) calibration data balance to optimize sparsely activated experts. These strategies collectively enable robust, high-precision quantization of MoE models under ultra-low-bit constraints.Extensive experiments across several extreme quantization settings (e.g., W4A4/W3A4/W3A3/W2A4) demonstrate that EAQuant significantly outperforms existing methods, achieving average accuracy improvements of 1.15 - 13.81% across three diverse MoE architectures, with particularly pronounced gains in reasoning tasks and robust performance retention under aggressive quantization. By integrating these innovations, EAQuant establishes a new state-of-the-art for high-precision, efficient MoE model compression.Our code is available at https://github.com/darren-fzq1/EAQuant.", "authors": ["Zhongqian Fu", "Tianyi Zhao", "Ning Ding", "Xianzhi Yu", "Xiaosong Li", "Yehui Tang", "Yunhe Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-16", "url": "https://arxiv.org/abs/2506.13329", "pdf_url": "https://arxiv.org/pdf/2506.13329v3", "arxiv_id": "2506.13329", "doi": "10.48550/arXiv.2506.13329", "citation_count": 5, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/darren-fzq1/EAQuant", "venue": "arXiv.org", "quality_score": 0.3099}
{"id": "28ff1d0cc25efc98fd2f31075c3b19b391cfac6a310c2d1cd5c538dfb3b44dff", "sources": ["arxiv", "semantic_scholar"], "title": "Automatic Expert Discovery in LLM Upcycling via Sparse Interpolated Mixture-of-Experts", "abstract": "We present Sparse Interpolated Mixture-of-Experts (SIMoE) instruction-tuning, an end-to-end algorithm designed to fine-tune a dense pre-trained Large Language Model (LLM) into a MoE-style model that possesses capabilities in multiple specialized domains. During instruction-tuning, SIMoE automatically identifies multiple specialized experts under a specified sparsity constraint, with each expert representing a structurally sparse subset of the seed LLM's parameters that correspond to domain-specific knowledge within the data. SIMoE simultaneously learns an input-dependent expert merging strategy via a router network, leveraging rich cross-expert knowledge for superior downstream generalization that surpasses existing baselines. Empirically, SIMoE consistently achieves state-of-the-art performance on common instruction-tuning benchmarks while maintaining an optimal performance-compute trade-off compared to all baselines.", "authors": ["Shengzhuang Chen", "Ying Wei", "Jonathan Richard Schwarz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-14", "url": "https://arxiv.org/abs/2506.12597", "pdf_url": "https://arxiv.org/pdf/2506.12597v1", "arxiv_id": "2506.12597", "doi": "10.48550/arXiv.2506.12597", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1982}
{"id": "f7b807ec3d254b552a9a0e5a8d4dc7ae5413e703d4a862dfbe2840b58867ee18", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning", "abstract": "Continual multimodal instruction tuning is crucial for adapting Multimodal Large Language Models (MLLMs) to evolving tasks. However, most existing methods adopt a fixed architecture, struggling with adapting to new tasks due to static model capacity. We propose to evolve the architecture under parameter budgets for dynamic task adaptation, which remains unexplored and imposes two challenges: 1) task architecture conflict, where different tasks require varying layer-wise adaptations, and 2) modality imbalance, where different tasks rely unevenly on modalities, leading to unbalanced updates. To address these challenges, we propose a novel Dynamic Mixture of Curriculum LoRA Experts (D-MoLE) method, which automatically evolves MLLM's architecture with controlled parameter budgets to continually adapt to new tasks while retaining previously learned knowledge. Specifically, we propose a dynamic layer-wise expert allocator, which automatically allocates LoRA experts across layers to resolve architecture conflicts, and routes instructions layer-wisely to facilitate knowledge sharing among experts. Then, we propose a gradient-based inter-modal continual curriculum, which adjusts the update ratio of each module in MLLM based on the difficulty of each modality within the task to alleviate the modality imbalance problem. Extensive experiments show that D-MoLE significantly outperforms state-of-the-art baselines, achieving a 15% average improvement over the best baseline. To the best of our knowledge, this is the first study of continual learning for MLLMs from an architectural perspective.", "authors": ["Chendi Ge", "Xin Wang", "Zeyang Zhang", "Hong Chen", "Jiapei Fan", "Longtao Huang", "Hui Xue", "Wenwu Zhu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-13", "url": "https://arxiv.org/abs/2506.11672", "pdf_url": "https://arxiv.org/pdf/2506.11672v1", "arxiv_id": "2506.11672", "doi": "10.48550/arXiv.2506.11672", "citation_count": 22, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3404}
{"id": "2c85074b605a60d2378ef915cce447eebceb7e21e6c6c96e770b152e3929f3e1", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA Users Beware: A Few Spurious Tokens Can Manipulate Your Finetuned Model", "abstract": "Large Language Models (LLMs) are commonly finetuned for a variety of use cases and domains. A common approach is to leverage Low-Rank Adaptation (LoRA) -- known to provide strong performance at low resource costs. In this study, we demonstrate that LoRA actually opens the door to short-cut vulnerabilities -- and the more resource efficient is the LoRA setup, the more vulnerable will be the finetuned model to aggressive attacks. To measure that vulnerability, we introduce Seamless Spurious Token Injection (SSTI), where we find that LoRA exclusively focuses on even just a single token that is spuriously correlated with downstream labels. In short, injection of that spurious token during finetuning ensure that the model's prediction at test-time can be manipulated on-demand. We conducted experiments across model families and datasets to evaluate the impact of SSTI during LoRA finetuning while providing possible mitigations. Our experiments conclude that none of the existing checkers and preprocessors can sanitize a dataset raising new concerns for data quality and AI safety.", "authors": ["Marcel Mateos Salles", "Praney Goyal", "Pradyut Sekhsaria", "Hai Huang", "Randall Balestriero"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-13", "url": "https://arxiv.org/abs/2506.11402", "pdf_url": "https://arxiv.org/pdf/2506.11402v2", "arxiv_id": "2506.11402", "doi": "10.48550/arXiv.2506.11402", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1971}
{"id": "0e21e3196583dd03f78ee7f6bc6f6757f16ba6ceba4a1181e47bd820eb8841d2", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Homogeneous Attention: Memory-Efficient LLMs via Fourier-Approximated KV Cache", "abstract": "Large Language Models struggle with memory demands from the growing Key-Value (KV) cache as context lengths increase. Existing compression methods homogenize head dimensions or rely on attention-guided token pruning, often sacrificing accuracy or introducing computational overhead. We propose FourierAttention, a training-free framework that exploits the heterogeneous roles of transformer head dimensions: lower dimensions prioritize local context, while upper ones capture long-range dependencies. By projecting the long-context-insensitive dimensions onto orthogonal Fourier bases, FourierAttention approximates their temporal evolution with fixed-length spectral coefficients. Evaluations on LLaMA models show that FourierAttention achieves the best long-context accuracy on LongBench and Needle-In-A-Haystack (NIAH). Besides, a custom Triton kernel, FlashFourierAttention, is designed to optimize memory via streamlined read-write operations, enabling efficient deployment without performance compromise.", "authors": ["Xiaoran Liu", "Siyang He", "Qiqi Wang", "Ruixiao Li", "Yuerong Song", "Zhigeng Liu", "Linlin Li", "Qun Liu", "Zengfeng Huang", "Qipeng Guo", "Ziwei He", "Xipeng Qiu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-13", "url": "https://arxiv.org/abs/2506.11886", "pdf_url": "https://arxiv.org/pdf/2506.11886v1", "arxiv_id": "2506.11886", "doi": "10.48550/arXiv.2506.11886", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1971}
{"id": "4708612f843032b8b41478799eeb81ebfc4e1b109c53d47621a74def05d3dcad", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Long-Context LLM Inference via KV Cache Clustering", "abstract": "Large language models (LLMs) with extended context windows have become increasingly prevalent for tackling complex tasks. However, the substantial Key-Value (KV) cache required for long-context LLMs poses significant deployment challenges. Existing approaches either discard potentially critical information needed for future generations or offer limited efficiency gains due to high computational overhead. In this paper, we introduce Chelsea, a simple yet effective framework for online KV cache clustering. Our approach is based on the observation that key states exhibit high similarity along the sequence dimension. To enable efficient clustering, we divide the sequence into chunks and propose Chunked Soft Matching, which employs an alternating partition strategy within each chunk and identifies clusters based on similarity. Chelsea then merges the KV cache within each cluster into a single centroid. Additionally, we provide a theoretical analysis of the computational complexity and the optimality of the intra-chunk partitioning strategy. Extensive experiments across various models and long-context benchmarks demonstrate that Chelsea achieves up to 80% reduction in KV cache memory usage while maintaining comparable model performance. Moreover, with minimal computational overhead, Chelsea accelerates the decoding stage of inference by up to 3.19$\\times$ and reduces end-to-end latency by up to 2.72$\\times$.", "authors": ["Jie Hu", "Shengnan Wang", "Yutong He", "Ping Gong", "Jiawei Yi", "Juncheng Zhang", "Youhui Bai", "Renhai Chen", "Gong Zhang", "Cheng Li", "Kun Yuan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-13", "url": "https://arxiv.org/abs/2506.11418", "pdf_url": "https://arxiv.org/pdf/2506.11418v1", "arxiv_id": "2506.11418", "doi": "10.48550/arXiv.2506.11418", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "0b04004eb42feae95576114729d8d2cadb39f55013ae9028e9f58d985f321f22", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs", "abstract": "In multimodal large language models (MLLMs), the length of input visual tokens is often significantly greater than that of their textual counterparts, leading to a high inference cost. Many works aim to address this issue by removing redundant visual tokens. However, current approaches either rely on attention-based pruning, which retains numerous duplicate tokens, or use similarity-based pruning, overlooking the instruction relevance, consequently causing suboptimal performance. In this paper, we go beyond attention or similarity by proposing a novel visual token pruning method named CDPruner, which maximizes the conditional diversity of retained tokens. We first define the conditional similarity between visual tokens conditioned on the instruction, and then reformulate the token pruning problem with determinantal point process (DPP) to maximize the conditional diversity of the selected subset. The proposed CDPruner is training-free and model-agnostic, allowing easy application to various MLLMs. Extensive experiments across diverse MLLMs show that CDPruner establishes new state-of-the-art on various vision-language benchmarks. By maximizing conditional diversity through DPP, the selected subset better represents the input images while closely adhering to user instructions, thereby preserving strong performance even with high reduction ratios. When applied to LLaVA, CDPruner reduces FLOPs by 95\\% and CUDA latency by 78\\%, while maintaining 94\\% of the original accuracy. Our code is available at https://github.com/Theia-4869/CDPruner.", "authors": ["Qizhe Zhang", "Mengzhen Liu", "Lichen Li", "Ming Lu", "Yuan Zhang", "Junwen Pan", "Qi She", "Shanghang Zhang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-12", "url": "https://arxiv.org/abs/2506.10967", "pdf_url": "https://arxiv.org/pdf/2506.10967v2", "arxiv_id": "2506.10967", "doi": "10.48550/arXiv.2506.10967", "citation_count": 60, "influential_citation_count": 10, "has_code": true, "code_url": "https://github.com/Theia-4869/CDPruner", "venue": "arXiv.org", "quality_score": 0.5207}
{"id": "94c788f9792e4cb65c213ff7eea8f7d3d8c0c13fdf0a55f73fc10d4a354d74b0", "sources": ["arxiv", "semantic_scholar"], "title": "AWP: Activation-Aware Weight Pruning and Quantization with Projected Gradient Descent", "abstract": "To address the enormous size of Large Language Models (LLMs), model compression methods, such as quantization and pruning, are often deployed, especially on edge devices. In this work, we focus on layer-wise post-training quantization and pruning. Drawing connections between activation-aware weight pruning and sparse approximation problems, and motivated by the success of Iterative Hard Thresholding (IHT), we propose a unified method for Activation-aware Weight pruning and quantization via Projected gradient descent (AWP). Our experiments demonstrate that AWP outperforms state-of-the-art LLM pruning and quantization methods. Theoretical convergence guarantees of the proposed method for pruning are also provided.", "authors": ["Jing Liu", "Toshiaki Koike-Akino", "Ye Wang", "Hassan Mansour", "Matthew Brand"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-11", "url": "https://arxiv.org/abs/2506.10205", "pdf_url": "https://arxiv.org/pdf/2506.10205v2", "arxiv_id": "2506.10205", "doi": "10.48550/arXiv.2506.10205", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1948}
{"id": "ef6ddfb7e376db8a114cde4762f2217f27000c45e63a1c756417676a425f16e9", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Multilingual ASR Finetuning via LoRA Language Experts", "abstract": "Recent advancements in deep learning have significantly enhanced multilingual automatic speech recognition (ASR) due to the development of advanced model architectures and available large-scale multilingual datasets. Despite that, multilingual ASR still suffers from the curse of multilinguality in that different languages tend to interfere with each other, making it difficult for the ASR model to identify multiple languages effectively while sharing model capacity across them. This paper proposes an efficient finetuning framework for customized multilingual ASR via prepared LoRA language experts based on Whisper. Through LoRA expert fusion or knowledge distillation, our approach achieves better recognition performance on target languages than standard fine-tuning methods. Experimental results demonstrate that the proposed models yield approximately 10\\% and 15\\% relative performance gains in language-aware and language-agnostic scenarios, respectively.", "authors": ["Jiahong Li", "Yiwen Shao", "Jianheng Zhuo", "Chenda Li", "Liliang Tang", "Dong Yu", "Yanmin Qian"], "categories": ["cs.CL", "cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-06-11", "url": "https://arxiv.org/abs/2506.21555", "pdf_url": "https://arxiv.org/pdf/2506.21555v1", "arxiv_id": "2506.21555", "doi": "10.48550/arXiv.2506.21555", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.2386}
{"id": "16742791c22dc07462091a4c64de8d0cf91d9849504166466f70a626693f2598", "sources": ["arxiv", "semantic_scholar"], "title": "DIVE into MoE: Diversity-Enhanced Reconstruction of Large Language Models from Dense into Mixture-of-Experts", "abstract": "Large language models (LLMs) with the Mixture-of-Experts (MoE) architecture achieve high cost-efficiency by selectively activating a subset of the parameters. Despite the inference efficiency of MoE LLMs, the training of extensive experts from scratch incurs substantial overhead, whereas reconstructing a dense LLM into an MoE LLM significantly reduces the training budget. However, existing reconstruction methods often overlook the diversity among experts, leading to potential redundancy. In this paper, we come up with the observation that a specific LLM exhibits notable diversity after being pruned on different calibration datasets, based on which we present a Diversity-Enhanced reconstruction method named DIVE. The recipe of DIVE includes domain affinity mining, pruning-based expert reconstruction, and efficient retraining. Specifically, the reconstruction includes pruning and reassembly of the feed-forward network (FFN) module. After reconstruction, we efficiently retrain the model on routers, experts and normalization modules. We implement DIVE on Llama-style LLMs with open-source training corpora. Experiments show that DIVE achieves training efficiency with minimal accuracy trade-offs, outperforming existing pruning and MoE reconstruction methods with the same number of activated parameters.", "authors": ["Yuchen Feng", "Bowen Shen", "Naibin Gu", "Jiaxuan Zhao", "Peng Fu", "Zheng Lin", "Weiping Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-11", "url": "https://arxiv.org/abs/2506.09351", "pdf_url": "https://arxiv.org/pdf/2506.09351v1", "arxiv_id": "2506.09351", "doi": "10.48550/arXiv.2506.09351", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.301}
{"id": "776cf931974ccf4083f689998c17a975ce9e6ec83de6edf7e6791edfc0b09d6f", "sources": ["arxiv", "semantic_scholar"], "title": "SparseSSM: Efficient Selective Structured State Space Models Can Be Pruned in One-Shot", "abstract": "State-space language models such as Mamba match Transformer quality while permitting linear complexity inference, yet still comprise billions of parameters that hinder deployment. Existing one-shot pruning methods are tailored to attention blocks and fail to account for the time-shared and discretized state-transition matrix at the heart of the selective state-space module (SSM). In this paper, we introduce SparseSSM, the first training-free pruning framework that extends the classic optimal brain surgeon (OBS) framework to state space architectures. Our layer-wise algorithm (i) derives an approximate second-order saliency score that aggregates Hessian-trace information across time steps, (ii) incorporates a component sensitivity analysis to guide feed-forward network (FFN) pruning, which also sheds light on where redundancy resides in mamba architecture, (iii) can be easily extended to semi-structured and structured sparsity. Empirically, we prune 50% of SSM weights without fine-tuning and observe no zero-shot accuracy loss, achieving the current state-of-the-art pruning algorithm for Mamba-based LLMs.", "authors": ["Kaiwen Tuo", "Huan Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-11", "url": "https://arxiv.org/abs/2506.09613", "pdf_url": "https://arxiv.org/pdf/2506.09613v1", "arxiv_id": "2506.09613", "doi": "10.48550/arXiv.2506.09613", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1948}
{"id": "1a257a215a5c10d00e351f1b2f20af9a48ba6ff083cac726b18dedb18695e35b", "sources": ["arxiv", "semantic_scholar"], "title": "CoQMoE: Co-Designed Quantization and Computation Orchestration for Mixture-of-Experts Vision Transformer on FPGA", "abstract": "Vision Transformers (ViTs) exhibit superior performance in computer vision tasks but face deployment challenges on resource-constrained devices due to high computational/memory demands. While Mixture-of-Experts Vision Transformers (MoE-ViTs) mitigate this through a scalable architecture with sub-linear computational growth, their hardware implementation on FPGAs remains constrained by resource limitations. This paper proposes a novel accelerator for efficiently implementing quantized MoE models on FPGAs through two key innovations: (1) A dual-stage quantization scheme combining precision-preserving complex quantizers with hardware-friendly simplified quantizers via scale reparameterization, with only 0.28 $\\%$ accuracy loss compared to full precision; (2) A resource-aware accelerator architecture featuring latency-optimized streaming attention kernels and reusable linear operators, effectively balancing performance and resource consumption. Experimental results demonstrate that our accelerator achieves nearly 155 frames per second, a 5.35$\\times$ improvement in throughput, and over $80\\%$ energy reduction compared to state-of-the-art (SOTA) FPGA MoE accelerators, while maintaining $<1\\%$ accuracy loss across vision benchmarks. Our implementation is available at https://github.com/DJ000011/CoQMoE.", "authors": ["Jiale Dong", "Hao Wu", "Zihao Wang", "Wenqi Lou", "Zhendong Zheng", "Lei Gong", "Chao Wang", "Xuehai Zhou"], "categories": ["cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-10", "url": "https://arxiv.org/abs/2506.08496", "pdf_url": "https://arxiv.org/pdf/2506.08496v1", "arxiv_id": "2506.08496", "doi": "10.48550/arXiv.2506.08496", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/DJ000011/CoQMoE", "venue": "European Conference on Parallel Processing", "quality_score": 0.2993}
{"id": "68d214d8cf2f6fc633df61f0a0880d85d96864f1744bd6e66c500e89f0443650", "sources": ["arxiv", "semantic_scholar"], "title": "MedMoE: Modality-Specialized Mixture of Experts for Medical Vision-Language Understanding", "abstract": "Different medical imaging modalities capture diagnostic information at varying spatial resolutions, from coarse global patterns to fine-grained localized structures. However, most existing vision-language frameworks in the medical domain apply a uniform strategy for local feature extraction, overlooking the modality-specific demands. In this work, we present MedMoE, a modular and extensible vision-language processing framework that dynamically adapts visual representation based on the diagnostic context. MedMoE incorporates a Mixture-of-Experts (MoE) module conditioned on the report type, which routes multi-scale image features through specialized expert branches trained to capture modality-specific visual semantics. These experts operate over feature pyramids derived from a Swin Transformer backbone, enabling spatially adaptive attention to clinically relevant regions. This framework produces localized visual representations aligned with textual descriptions, without requiring modality-specific supervision at inference. Empirical results on diverse medical benchmarks demonstrate that MedMoE improves alignment and retrieval performance across imaging modalities, underscoring the value of modality-specialized visual representations in clinical vision-language systems.", "authors": ["Shivang Chopra", "Gabriela Sanchez-Rodriguez", "Lingchao Mao", "Andrew J Feola", "Jing Li", "Zsolt Kira"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-10", "url": "https://arxiv.org/abs/2506.08356", "pdf_url": "https://arxiv.org/pdf/2506.08356v2", "arxiv_id": "2506.08356", "doi": "10.48550/arXiv.2506.08356", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "5661f8bad30d092ce9060af6e62a47cf7a834e5d7ba5c081c9e0f2879b9c803b", "sources": ["arxiv", "semantic_scholar"], "title": "MoQAE: Mixed-Precision Quantization for Long-Context LLM Inference via Mixture of Quantization-Aware Experts", "abstract": "One of the primary challenges in optimizing large language models (LLMs) for long-context inference lies in the high memory consumption of the Key-Value (KV) cache. Existing approaches, such as quantization, have demonstrated promising results in reducing memory usage. However, current quantization methods cannot take both effectiveness and efficiency into account. In this paper, we propose MoQAE, a novel mixed-precision quantization method via mixture of quantization-aware experts. First, we view different quantization bit-width configurations as experts and use the traditional mixture of experts (MoE) method to select the optimal configuration. To avoid the inefficiency caused by inputting tokens one by one into the router in the traditional MoE method, we input the tokens into the router chunk by chunk. Second, we design a lightweight router-only fine-tuning process to train MoQAE with a comprehensive loss to learn the trade-off between model accuracy and memory usage. Finally, we introduce a routing freezing (RF) and a routing sharing (RS) mechanism to further reduce the inference overhead. Extensive experiments on multiple benchmark datasets demonstrate that our method outperforms state-of-the-art KV cache quantization approaches in both efficiency and effectiveness.", "authors": ["Wei Tao", "Haocheng Lu", "Xiaoyang Qu", "Bin Zhang", "Kai Lu", "Jiguang Wan", "Jianzong Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-09", "url": "https://arxiv.org/abs/2506.07533", "pdf_url": "https://arxiv.org/pdf/2506.07533v1", "arxiv_id": "2506.07533", "doi": "10.48550/arXiv.2506.07533", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2603}
{"id": "c0af1957d2c16ded791f3274801b5c496c60c7e57c3b0cb85d34cc12e09cea48", "sources": ["arxiv", "semantic_scholar"], "title": "Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models", "abstract": "Modern large language models (LLMs) are inherently auto-regressive, requiring input to be serialized into flat sequences regardless of their structural dependencies. This serialization hinders the model's ability to leverage structural inductive biases, especially in tasks such as retrieval-augmented generation (RAG) and reasoning on data with native graph structures, where inter-segment dependencies are crucial. We introduce Graph-KV with the potential to overcome this limitation. Graph-KV leverages the KV-cache of text segments as condensed representations and governs their interaction through structural inductive biases. In this framework, 'target' segments selectively attend only to the KV-caches of their designated 'source' segments, rather than all preceding segments in a serialized sequence. This approach induces a graph-structured block mask, sparsifying attention and enabling a message-passing-like step within the LLM. Furthermore, strategically allocated positional encodings for source and target segments reduce positional bias and context window consumption. We evaluate Graph-KV across three scenarios: (1) seven RAG benchmarks spanning direct inference, multi-hop reasoning, and long-document understanding; (2) Arxiv-QA, a novel academic paper QA task with full-text scientific papers structured as citation ego-graphs; and (3) paper topic classification within a citation network. By effectively reducing positional bias and harnessing structural inductive biases, Graph-KV substantially outperforms baselines, including standard costly sequential encoding, across various settings. Code and the Graph-KV data are publicly available.", "authors": ["Haoyu Wang", "Peihao Wang", "Mufei Li", "Shikun Liu", "Siqi Miao", "Zhangyang Wang", "Pan Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-09", "url": "https://arxiv.org/abs/2506.07334", "pdf_url": "https://arxiv.org/pdf/2506.07334v2", "arxiv_id": "2506.07334", "doi": "10.48550/arXiv.2506.07334", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "cbbc1ae344d80d0348259963078665aed30b372f530c009195e23eb8efa8881e", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture Experts with Test-Time Self-Supervised Aggregation for Tabular Imbalanced Regression", "abstract": "Tabular data serve as a fundamental and ubiquitous representation of structured information in numerous real-world applications, e.g., finance and urban planning. In the realm of tabular imbalanced applications, data imbalance has been investigated in classification tasks with insufficient instances in certain labels, causing the model's ineffective generalizability. However, the imbalance issue of tabular regression tasks is underexplored, and yet is critical due to unclear boundaries for continuous labels and simplifying assumptions in existing imbalance regression work, which often rely on known and balanced test distributions. Such assumptions may not hold in practice and can lead to performance degradation. To address these issues, we propose MATI: Mixture Experts with Test-Time Self-Supervised Aggregation for Tabular Imbalance Regression, featuring two key innovations: (i) the Region-Aware Mixture Expert, which adopts a Gaussian Mixture Model to capture the underlying related regions. The statistical information of each Gaussian component is then used to synthesize and train region-specific experts to capture the unique characteristics of their respective regions. (ii) Test-Time Self-Supervised Expert Aggregation, which dynamically adjusts region expert weights based on test data features to reinforce expert adaptation across varying test distributions. We evaluated MATI on four real-world tabular imbalance regression datasets, including house pricing, bike sharing, and age prediction. To reflect realistic deployment scenarios, we adopted three types of test distributions: a balanced distribution with uniform target frequencies, a normal distribution that follows the training data, and an inverse distribution that emphasizes rare target regions. On average across these three test distributions, MATI achieved a 7.1% improvement in MAE compared to existing methods.", "authors": ["Yung-Chien Wang", "Kuang-Da Wang", "Wei-Yao Wang", "Wen-Chih Peng"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-08", "url": "https://arxiv.org/abs/2506.07033", "pdf_url": "https://arxiv.org/pdf/2506.07033v1", "arxiv_id": "2506.07033", "doi": "10.48550/arXiv.2506.07033", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1914}
{"id": "a96b960de601b28612d2bfbd962b1f365df3ca06c2ee9f227efcf602bbb878b8", "sources": ["arxiv", "semantic_scholar"], "title": "MadaKV: Adaptive Modality-Perception KV Cache Eviction for Efficient Multimodal Long-Context Inference", "abstract": "This paper introduces MadaKV, a modality-adaptive key-value (KV) cache eviction strategy designed to enhance the efficiency of multimodal large language models (MLLMs) in long-context inference. In multimodal scenarios, attention heads exhibit varying preferences for different modalities, resulting in significant disparities in modality importance across attention heads. Traditional KV cache eviction methods, which are tailored for unimodal settings, fail to capture modality-specific information, thereby yielding suboptimal performance. MadaKV addresses these challenges through two key components: modality preference adaptation and hierarchical compression compensation. By dynamically sensing modality information within attention heads and adaptively retaining critical tokens, MadaKV achieves substantial reductions in KV cache memory footprint and model inference decoding latency (1.3 to 1.5 times improvement) while maintaining high accuracy across various multimodal long-context tasks. Extensive experiments on representative MLLMs and the MileBench benchmark demonstrate the effectiveness of MadaKV compared to existing KV cache eviction methods.", "authors": ["Kunxi Li", "Zhonghua Jiang", "Zhouzhou Shen", "Zhaode Wang", "Chengfei Lv", "Shengyu Zhang", "Fan Wu", "Fei Wu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-06", "url": "https://arxiv.org/abs/2506.15724", "pdf_url": "https://arxiv.org/pdf/2506.15724v1", "arxiv_id": "2506.15724", "doi": "10.48550/arXiv.2506.15724", "citation_count": 9, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.25}
{"id": "a7214f4049d0a30f227516276a52f1b66537e05bea341c926caf35214051b115", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Lifelong Robot Learning", "abstract": "A generalist agent must continuously learn and adapt throughout its lifetime, achieving efficient forward transfer while minimizing catastrophic forgetting. Previous work within the dominant pretrain-then-finetune paradigm has explored parameter-efficient fine-tuning for single-task adaptation, effectively steering a frozen pretrained model with a small number of parameters. However, in the context of lifelong learning, these methods rely on the impractical assumption of a test-time task identifier and restrict knowledge sharing among isolated adapters. To address these limitations, we propose Dynamic Mixture of Progressive Parameter-Efficient Expert Library (DMPEL) for lifelong robot learning. DMPEL progressively builds a low-rank expert library and employs a lightweight router to dynamically combine experts into an end-to-end policy, enabling flexible and efficient lifelong forward transfer. Furthermore, by leveraging the modular structure of the fine-tuned parameters, we introduce expert coefficient replay, which guides the router to accurately retrieve frozen experts for previously encountered tasks. This technique mitigates forgetting while being significantly more storage- and computation-efficient than experience replay over the entire policy. Extensive experiments on the lifelong robot learning benchmark LIBERO demonstrate that our framework outperforms state-of-the-art lifelong learning methods in success rates during continual adaptation, while utilizing minimal trainable parameters and storage.", "authors": ["Yuheng Lei", "Sitong Mao", "Shunbo Zhou", "Hongyuan Zhang", "Xuelong Li", "Ping Luo"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-06", "url": "https://arxiv.org/abs/2506.05985", "pdf_url": "https://arxiv.org/pdf/2506.05985v3", "arxiv_id": "2506.05985", "doi": "10.48550/arXiv.2506.05985", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/HarryLui98/DMPEL", "venue": "arXiv.org", "quality_score": 0.2922}
{"id": "a24fb2c30c3c969e62c6b73a16a180ed1a0b869efb3ece8929472a2842ad31f3", "sources": ["arxiv", "semantic_scholar"], "title": "Hierarchical-Task-Aware Multi-modal Mixture of Incremental LoRA Experts for Embodied Continual Learning", "abstract": "Previous continual learning setups for embodied intelligence focused on executing low-level actions based on human commands, neglecting the ability to learn high-level planning and multi-level knowledge. To address these issues, we propose the Hierarchical Embodied Continual Learning Setups (HEC) that divide the agent's continual learning process into two layers: high-level instructions and low-level actions, and define five embodied continual learning sub-setups. Building on these setups, we introduce the Task-aware Mixture of Incremental LoRA Experts (Task-aware MoILE) method. This approach achieves task recognition by clustering visual-text embeddings and uses both a task-level router and a token-level router to select the appropriate LoRA experts. To effectively address the issue of catastrophic forgetting, we apply Singular Value Decomposition (SVD) to the LoRA parameters obtained from prior tasks, preserving key components while orthogonally training the remaining parts. The experimental results show that our method stands out in reducing the forgetting of old tasks compared to other methods, effectively supporting agents in retaining prior knowledge while continuously learning new tasks.", "authors": ["Ziqi Jia", "Anmin Wang", "Xiaoyang Qu", "Xiaowen Yang", "Jianzong Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-05", "url": "https://arxiv.org/abs/2506.04595", "pdf_url": "https://arxiv.org/pdf/2506.04595v1", "arxiv_id": "2506.04595", "doi": "10.48550/arXiv.2506.04595", "citation_count": 7, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2386}
{"id": "b34128c278b748e9407a6e57d5814dea69be051574ec3919427daefa43ef7d4f", "sources": ["arxiv", "semantic_scholar"], "title": "TaDA: Training-free recipe for Decoding with Adaptive KV Cache Compression and Mean-centering", "abstract": "The key-value (KV) cache in transformer models is a critical component for efficient decoding or inference, yet its memory demands scale poorly with sequence length, posing a major challenge for scalable deployment of large language models. Among several approaches to KV cache compression, quantization of key and value activations has been widely explored. Most KV cache quantization methods still need to manage sparse and noncontiguous outliers separately. To address this, we introduce TaDA, a training-free recipe for KV cache compression with quantization precision that adapts to error sensitivity across layers and a mean centering to eliminate separate outlier handling. Our approach yields substantial accuracy improvements for multiple models supporting various context lengths. Moreover, our approach does not need to separately manage outlier elements -- a persistent hurdle in most traditional quantization methods. Experiments on standard benchmarks demonstrate that our technique reduces KV cache memory footprint to 27% of the original 16-bit baseline while achieving comparable accuracy. Our method paves the way for scalable and high-performance reasoning in language models by potentially enabling inference for longer context length models, reasoning models, and longer chain of thoughts.", "authors": ["Vinay Joshi", "Pratik Prabhanjan Brahma", "Zicheng Liu", "Emad Barsoum"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-05", "url": "https://arxiv.org/abs/2506.04642", "pdf_url": "https://arxiv.org/pdf/2506.04642v1", "arxiv_id": "2506.04642", "doi": "10.48550/arXiv.2506.04642", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1879}
{"id": "2467e82d4449b2f96a59d64bc82af331a410d06d8234690d127f68e67b6516f9", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Few-Shot Image Classification via Prototypical Concept-Guided Mixture of LoRA Experts", "abstract": "Self-Explainable Models (SEMs) rely on Prototypical Concept Learning (PCL) to enable their visual recognition processes more interpretable, but they often struggle in data-scarce settings where insufficient training samples lead to suboptimal performance.To address this limitation, we propose a Few-Shot Prototypical Concept Classification (FSPCC) framework that systematically mitigates two key challenges under low-data regimes: parametric imbalance and representation misalignment. Specifically, our approach leverages a Mixture of LoRA Experts (MoLE) for parameter-efficient adaptation, ensuring a balanced allocation of trainable parameters between the backbone and the PCL module.Meanwhile, cross-module concept guidance enforces tight alignment between the backbone's feature representations and the prototypical concept activation patterns.In addition, we incorporate a multi-level feature preservation strategy that fuses spatial and semantic cues across various layers, thereby enriching the learned representations and mitigating the challenges posed by limited data availability.Finally, to enhance interpretability and minimize concept overlap, we introduce a geometry-aware concept discrimination loss that enforces orthogonality among concepts, encouraging more disentangled and transparent decision boundaries.Experimental results on six popular benchmarks (CUB-200-2011, mini-ImageNet, CIFAR-FS, Stanford Cars, FGVC-Aircraft, and DTD) demonstrate that our approach consistently outperforms existing SEMs by a notable margin, with 4.2%-8.7% relative gains in 5-way 5-shot classification.These findings highlight the efficacy of coupling concept learning with few-shot adaptation to achieve both higher accuracy and clearer model interpretability, paving the way for more transparent visual recognition systems.", "authors": ["Zhong Ji", "Rongshuai Wei", "Jingren Liu", "Yanwei Pang", "Jungong Han"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2025-06-05", "url": "https://arxiv.org/abs/2506.04673", "pdf_url": "https://arxiv.org/pdf/2506.04673v1", "arxiv_id": "2506.04673", "doi": "10.1109/TIP.2026.3654473", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Image Processing", "quality_score": 0.1879}
{"id": "2011e1de7cd1bb12bf02b1a95308ee588383239eabe221cad04904e175f86e7b", "sources": ["arxiv", "semantic_scholar"], "title": "Inference-Time Hyper-Scaling with KV Cache Compression", "abstract": "Inference-time scaling trades efficiency for increased reasoning accuracy by generating longer or more parallel sequences. However, in Transformer LLMs, generation cost is bottlenecked by the size of the key-value (KV) cache, rather than the number of generated tokens. Hence, we explore inference-time hyper-scaling: by compressing the KV cache, we can generate more tokens within the same compute budget and further improve the accuracy of scaled inference. The success of this approach, however, hinges on the ability of compression methods to preserve accuracy even at high compression ratios. To make hyper-scaling practical, we introduce Dynamic Memory Sparsification (DMS), a novel method for sparsifying KV caches that only requires 1K training steps to achieve 8$\\times$ compression, while maintaining better accuracy than training-free sparse attention. Instead of prematurely discarding cached tokens, DMS delays token eviction, implicitly merging representations and preserving critical information. We demonstrate the effectiveness of inference-time hyper-scaling with DMS on multiple families of LLMs, showing that it boosts accuracy for comparable inference latency and memory load. For instance, we enhance Qwen-R1 32B by 12.0 points on AIME 24, 8.6 on GPQA, and 9.7 on LiveCodeBench on average for an equivalent number of memory reads.", "authors": ["Adrian Łańcucki", "Konrad Staniszewski", "Piotr Nawrot", "Edoardo M. Ponti"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-05", "url": "https://arxiv.org/abs/2506.05345", "pdf_url": "https://arxiv.org/pdf/2506.05345v2", "arxiv_id": "2506.05345", "doi": "10.48550/arXiv.2506.05345", "citation_count": 26, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3578}
{"id": "8459752e84da8ded35c826f4a6aeeae76439609b2499451a75ddf86ed0392abe", "sources": ["arxiv", "semantic_scholar"], "title": "AhaKV: Adaptive Holistic Attention-Driven KV Cache Eviction for Efficient Inference of Large Language Models", "abstract": "Large Language Models (LLMs) have significantly advanced the field of Artificial Intelligence. However, their deployment is resource-intensive, not only due to the large number of model parameters but also because the (Key-Value) KV cache consumes a lot of memory during inference. While several works propose reducing the KV cache by evicting the unnecessary tokens, these approaches rely on accumulated attention score as eviction score to quantify the importance of the token. We identify the accumulated attention score is biased and it decreases with the position of the tokens in the mathematical expectation. As a result, the retained tokens concentrate on the initial positions, limiting model's access to global contextual information. To address this issue, we propose Adaptive holistic attention KV (AhaKV), it addresses the bias of the accumulated attention score by adaptively tuning the scale of softmax according the expectation of information entropy of attention scores. To make use of the holistic attention information in self-attention mechanism, AhaKV utilize the information of value vectors, which is overlooked in previous works, to refine the adaptive score. We show theoretically that our method is well suited for bias reduction. We deployed AhaKV on different models with a fixed cache budget. Experiments show that AhaKV successfully mitigates bias and retains crucial tokens across global context and achieve state-of-the-art results against other related work on several benchmark tasks.", "authors": ["Yifeng Gu", "Zicong Jiang", "Jianxiu Jin", "Kailing Guo", "Ziyang Zhang", "Xiangmin Xu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-04", "url": "https://arxiv.org/abs/2506.03762", "pdf_url": "https://arxiv.org/pdf/2506.03762v1", "arxiv_id": "2506.03762", "doi": "10.48550/arXiv.2506.03762", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1868}
{"id": "8cf796614b610a6a1a1403c9e7413c4eb3d57ebed8a6bd6bd60a80ca3c1f6fc8", "sources": ["arxiv", "semantic_scholar"], "title": "Homogeneous Keys, Heterogeneous Values: Exploiting Local KV Cache Asymmetry for Long-Context LLMs", "abstract": "Recent advances in Large Language Models (LLMs) have highlighted the critical importance of extending context length, yet the quadratic complexity of attention mechanisms poses significant challenges for efficient long-context modeling. KV cache compression has emerged as a key approach to address this challenge. Through extensive empirical analysis, we reveal a fundamental yet previously overlooked asymmetry in KV caches: while adjacent keys receive similar attention weights ({\\it local homogeneity}), adjacent values demonstrate distinct {\\it heterogeneous} distributions. This key-value asymmetry reveals a critical limitation in existing compression methods that treat keys and values uniformly. To address the limitation, we propose a training-free compression framework (AsymKV) that combines homogeneity-based key merging with a mathematically proven lossless value compression. Extensive experiments demonstrate that AsymKV consistently outperforms existing long-context methods across various tasks and base models. For example, on LLaMA3.1-8B, AsymKV achieves an average score of 43.95 on LongBench, surpassing SOTA methods like H$_2$O (38.89) by a large margin.Our code can be found in this link:https://github.com/the-scale-lab/Asymkv.", "authors": ["Wanyun Cui", "Mingwei Xu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-04", "url": "https://arxiv.org/abs/2506.05410", "pdf_url": "https://arxiv.org/pdf/2506.05410v2", "arxiv_id": "2506.05410", "doi": "10.48550/arXiv.2506.05410", "citation_count": 3, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/the-scale-lab/Asymkv", "venue": "arXiv.org", "quality_score": 0.2886}
{"id": "463110fea9aefefc8ddcd4a0f2c2590d1ead97db630e28c7984a698d74e57272", "sources": ["arxiv", "semantic_scholar"], "title": "SkipGPT: Dynamic Layer Pruning Reinvented with Token Awareness and Module Decoupling", "abstract": "Large language models (LLMs) achieve remarkable performance across tasks but incur substantial computational costs due to their deep, multi-layered architectures. Layer pruning has emerged as a strategy to alleviate these inefficiencies, but conventional static pruning methods overlook two critical dynamics inherent to LLM inference: (1) horizontal dynamics, where token-level heterogeneity demands context-aware pruning decisions, and (2) vertical dynamics, where the distinct functional roles of MLP and self-attention layers necessitate component-specific pruning policies. We introduce SkipGPT, a dynamic layer pruning framework designed to optimize computational resource allocation through two core innovations: (1) global token-aware routing to prioritize critical tokens, and (2) decoupled pruning policies for MLP and self-attention components. To mitigate training instability, we propose a two-stage optimization paradigm: first, a disentangled training phase that learns routing strategies via soft parameterization to avoid premature pruning decisions, followed by parameter-efficient LoRA fine-tuning to restore performance impacted by layer removal. Extensive experiments demonstrate that SkipGPT reduces over 40% of model parameters while matching or exceeding the performance of the original dense model across benchmarks. By harmonizing dynamic efficiency with preserved expressivity, SkipGPT advances the practical deployment of scalable, resource-aware LLMs. Our code is publicly available at: https://github.com/EIT-NLP/SkipGPT.", "authors": ["Anhao Zhao", "Fanghua Ye", "Yingqi Fan", "Junlong Tong", "Zhiwei Fei", "Hui Su", "Xiaoyu Shen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-04", "url": "https://arxiv.org/abs/2506.04179", "pdf_url": "https://arxiv.org/pdf/2506.04179v1", "arxiv_id": "2506.04179", "doi": "10.48550/arXiv.2506.04179", "citation_count": 10, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/EIT-NLP/SkipGPT", "venue": "arXiv.org", "quality_score": 0.2886}
{"id": "a03a2de69cde0baa4d716c3f84435ace669deee08bdad1021bbdae9db0a1bdab", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-level Mixture of Experts for Multimodal Entity Linking", "abstract": "Multimodal Entity Linking (MEL) aims to link ambiguous mentions within multimodal contexts to associated entities in a multimodal knowledge base. Existing approaches to MEL introduce multimodal interaction and fusion mechanisms to bridge the modality gap and enable multi-grained semantic matching. However, they do not address two important problems: (i) mention ambiguity, i.e., the lack of semantic content caused by the brevity and omission of key information in the mention's textual context; (ii) dynamic selection of modal content, i.e., to dynamically distinguish the importance of different parts of modal information. To mitigate these issues, we propose a Multi-level Mixture of Experts (MMoE) model for MEL. MMoE has four components: (i) the description-aware mention enhancement module leverages large language models to identify the WikiData descriptions that best match a mention, considering the mention's textual context; (ii) the multimodal feature extraction module adopts multimodal feature encoders to obtain textual and visual embeddings for both mentions and entities; (iii)-(iv) the intra-level mixture of experts and inter-level mixture of experts modules apply a switch mixture of experts mechanism to dynamically and adaptively select features from relevant regions of information. Extensive experiments demonstrate the outstanding performance of MMoE compared to the state-of-the-art. MMoE's code is available at: https://github.com/zhiweihu1103/MEL-MMoE.", "authors": ["Zhiwei Hu", "Víctor Gutiérrez-Basulto", "Zhiliang Xiang", "Ru Li", "Jeff Z. Pan"], "categories": ["cs.CV", "cs.AI", "cs.CL", "cs.LG", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-03", "url": "https://arxiv.org/abs/2507.07108", "pdf_url": "https://arxiv.org/pdf/2507.07108v1", "arxiv_id": "2507.07108", "doi": "10.1145/3711896.3737060", "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/zhiweihu1103/MEL-MMoE", "venue": "Knowledge Discovery and Data Mining", "quality_score": 0.2869}
{"id": "396c8acd26e50d3ecb003949c1fbb60b19c5498c94922ae167971479785fcb8c", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning General Large Language Models into Customized Expert Models", "abstract": "Large language models (LLMs) have revolutionized natural language processing, yet their substantial model sizes often require substantial computational resources. To preserve computing resources and accelerate inference speed, it is crucial to prune redundant parameters, especially for experienced users who often need compact expert models tailored to specific downstream scenarios. However, most existing pruning methods focus on preserving the model's general capabilities, often requiring extensive post-training or suffering from degraded performance due to coarse-grained pruning. In this work, we design a $\\underline{Cus}$tom $\\underline{Prun}$ing method ($\\texttt{Cus-Prun}$) to prune a large general model into a smaller lightweight expert model, which is positioned along the \"language\", \"domain\" and \"task\" dimensions. By identifying and pruning irrelevant neurons of each dimension, $\\texttt{Cus-Prun}$ creates expert models without any post-training. Our experiments demonstrate that $\\texttt{Cus-Prun}$ consistently outperforms other methods, achieving minimal loss in both expert and general capabilities across various models from different model families and sizes.", "authors": ["Yirao Zhao", "Guizhen Chen", "Kenji Kawaguchi", "Lidong Bing", "Wenxuan Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-03", "url": "https://arxiv.org/abs/2506.02561", "pdf_url": "https://arxiv.org/pdf/2506.02561v1", "arxiv_id": "2506.02561", "doi": "10.48550/arXiv.2506.02561", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1856}
{"id": "d2d09765f7590324ddd8904e6b268387e954e30f6c3fde1d4eaf699d3c7c9785", "sources": ["arxiv", "semantic_scholar"], "title": "Structured Pruning and Quantization for Learned Image Compression", "abstract": "The high computational costs associated with large deep learning models significantly hinder their practical deployment. Model pruning has been widely explored in deep learning literature to reduce their computational burden, but its application has been largely limited to computer vision tasks such as image classification and object detection. In this work, we propose a structured pruning method targeted for Learned Image Compression (LIC) models that aims to reduce the computational costs associated with image compression while maintaining the rate-distortion performance. We employ a Neural Architecture Search (NAS) method based on the rate-distortion loss for computing the pruning ratio for each layer of the network. We compare our pruned model with the uncompressed LIC Model with same network architecture and show that it can achieve model size reduction without any BD-Rate performance drop. We further show that our pruning method can be integrated with model quantization to achieve further model compression while maintaining similar BD-Rate performance. We have made the source code available at gitlab.com/viper-purdue/lic-pruning.", "authors": ["Md Adnan Faisal Hossain", "Fengqing Zhu"], "categories": ["eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-06-02", "url": "https://arxiv.org/abs/2506.01229", "pdf_url": "https://arxiv.org/pdf/2506.01229v1", "arxiv_id": "2506.01229", "doi": "10.1109/icip51287.2024.10648236", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "International Conference on Information Photonics", "quality_score": 0.2851}
{"id": "931533b09b8960fc84682a15fb8d0baf1f78fa4a2dff79b1d6c14ddea50c8746", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts Provably Detect and Learn the Latent Cluster Structure in Gradient-Based Learning", "abstract": "Mixture of Experts (MoE), an ensemble of specialized models equipped with a router that dynamically distributes each input to appropriate experts, has achieved successful results in the field of machine learning. However, theoretical understanding of this architecture is falling behind due to its inherent complexity. In this paper, we theoretically study the sample and runtime complexity of MoE following the stochastic gradient descent (SGD) when learning a regression task with an underlying cluster structure of single index models. On the one hand, we prove that a vanilla neural network fails in detecting such a latent organization as it can only process the problem as a whole. This is intrinsically related to the concept of information exponent which is low for each cluster, but increases when we consider the entire task. On the other hand, we show that a MoE succeeds in dividing this problem into easier subproblems by leveraging the ability of each expert to weakly recover the simpler function corresponding to an individual cluster. To the best of our knowledge, this work is among the first to explore the benefits of the MoE framework by examining its SGD dynamics in the context of nonlinear regression.", "authors": ["Ryotaro Kawata", "Kohsei Matsutani", "Yuri Kinoshita", "Naoki Nishikawa", "Taiji Suzuki"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-06-02", "url": "https://arxiv.org/abs/2506.01656", "pdf_url": "https://arxiv.org/pdf/2506.01656v2", "arxiv_id": "2506.01656", "doi": "10.48550/arXiv.2506.01656", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1845}
{"id": "ba2f8a061249fb33ede56abc4364a5074e50789dc272929c26e68b4ab3f71bf0", "sources": ["arxiv", "semantic_scholar"], "title": "FedQuad: Adaptive Layer-wise LoRA Deployment and Activation Quantization for Federated Fine-Tuning", "abstract": "Federated fine-tuning (FedFT) provides an effective paradigm for fine-tuning large language models (LLMs) in privacy-sensitive scenarios. However, practical deployment remains challenging due to the limited resources on end devices. Existing methods typically utilize parameter-efficient fine-tuning (PEFT) techniques, such as Low-Rank Adaptation (LoRA), to substantially reduce communication overhead. Nevertheless, significant memory usage for activation storage and computational demands from full backpropagation remain major barriers to efficient deployment on resource-constrained end devices. Moreover, substantial resource heterogeneity across devices results in severe synchronization bottlenecks, diminishing the overall fine-tuning efficiency. To address these issues, we propose FedQuad, a novel LoRA-based FedFT framework that adaptively adjusts the LoRA depth (the number of consecutive tunable LoRA layers from the output) according to device computational capabilities, while employing activation quantization to reduce memory overhead, thereby enabling efficient deployment on resource-constrained devices. Specifically, FedQuad first identifies the feasible and efficient combinations of LoRA depth and the number of activation quantization layers based on device-specific resource constraints. Subsequently, FedQuad employs a greedy strategy to select the optimal configurations for each device, effectively accommodating system heterogeneity. Extensive experiments demonstrate that FedQuad achieves a 1.4-5.3x convergence acceleration compared to state-of-the-art baselines when reaching target accuracy, highlighting its efficiency and deployability in resource-constrained and heterogeneous end-device environments.", "authors": ["Rukuo Li", "Jianchun Liu", "Hongli Xu", "Liusheng Huang"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-01", "url": "https://arxiv.org/abs/2506.01001", "pdf_url": "https://arxiv.org/pdf/2506.01001v1", "arxiv_id": "2506.01001", "doi": "10.1109/TMC.2025.3637064", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Mobile Computing", "quality_score": 0.2603}
{"id": "f2dec8f0ba15c7f9981170ca1bb024b65b0e4a898c928e01c21950ea7263eacb", "sources": ["arxiv", "semantic_scholar"], "title": "FLEx: Personalized Federated Learning for Mixture-of-Experts LLMs via Expert Grafting", "abstract": "Federated instruction tuning of large language models (LLMs) is challenged by significant data heterogeneity across clients, demanding robust personalization. The Mixture of Experts (MoE) architecture, where experts can specialize in distinct data patterns, presents a natural architectural solution to this challenge. The inherent sparsity of the MoE architecture, achieved by selectively activating experts, poses a significant challenge to its integration with federated learning (FL). Conventional FL frameworks, designed for dense models, naively aggregate all expert parameters irrespective of their local activation patterns. This naive approach not only undermines MoE's dynamic sparsity but also risks corrupting the world knowledge within pretrained experts. To address this, we propose FLEx (Federated LLMs with Personalized Experts), a novel framework that leverages pretrained MoE-based LLMs for efficient personalization. By aggregating only the shared non-expert parameters, FLEx significantly reduces communication overhead and preserves the world knowledge stored within the frozen pretrained experts. For personalization, we introduce a novel expert grafting mechanism that leverages dynamic sparsity to construct a client-specific expert from selected components of pretrained experts, tailored to local data. This grafted expert is then fine-tuned locally alongside the gating mechanism. This joint training enables the model to learn when to leverage the shared knowledge from frozen experts and when to employ the personalized one. Evaluations on diverse, non-IID instruction tuning datasets show that FLEx consistently outperforms federated baselines on average, while demonstrating strong knowledge preservation on the knowledge-driven benchmark MMLU. Our code is available at \\href{https://anonymous.4open.science/r/FLEx-8F12}{\\texttt{https://anonymous.4open.science/r/FLEx-8F12}}.", "authors": ["Fan Liu", "Bikang Pan", "Zhongyi Wang", "Xi Yao", "Xiaoying Tang", "Jingya Wang", "Ye Shi"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-01", "url": "https://arxiv.org/abs/2506.00965", "pdf_url": "https://arxiv.org/pdf/2506.00965v2", "arxiv_id": "2506.00965", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.2167}
{"id": "49136f72bd0be15dad4ef90d066f5f5675e0914be2e6c709c93bc046cf9c810f", "sources": ["arxiv", "semantic_scholar"], "title": "Assortment of Attention Heads: Accelerating Federated PEFT with Head Pruning and Strategic Client Selection", "abstract": "Parameter Efficient Fine-Tuning (PEFT) has become the de-facto approach in adapting Large Language Models (LLMs) for downstream tasks in Natural Language Processing. However, its adoption in privacy-preserving distributed learning frameworks, such as Federated Learning (FL), remains relatively limited. This is mainly due to challenges specific to FL, such as resource-constrained devices and diverse data distributions among clients. In this paper, we propose an efficient method to perform PEFT within the FL framework for Multi-Head Attention (MHA) based language models. We address the challenges through head pruning, a novel head-specific weighted aggregation mechanism, and a client selection strategy. Head pruning minimizes training complexity within the clients, guided by the importance score computed based on the confidence of the attention head. Weighted aggregation of heads ensures the global model captures crucial updates from diverse clients complementing our client selection strategy. We show results on the MultiNLI benchmark along with 20 Newsgroups, XL-Sum, and E2E NLG datasets. We use the MultiNLI dataset and T5-small model with LoRA as our PEFT method, attaining sparsity levels of up to 90%, resulting in a communication advantage of up to 1.8x and a reduction in training OPs of 3.9x while maintaining the accuracy drop under 2%.", "authors": ["Yeshwanth Venkatesha", "Souvik Kundu", "Priyadarshini Panda"], "categories": ["cs.CL", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-31", "url": "https://arxiv.org/abs/2506.00743", "pdf_url": "https://arxiv.org/pdf/2506.00743v1", "arxiv_id": "2506.00743", "doi": "10.48550/arXiv.2506.00743", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1159}
{"id": "f7c91fdc2753840d0ba11415af941af9ddb7d4a2e0611281fe37eefefc4d6138", "sources": ["arxiv", "semantic_scholar"], "title": "R-KV: Redundancy-aware KV Cache Compression for Reasoning Models", "abstract": "Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 16% of the KV cache. This KV-cache reduction also leads to a 90% memory saving and a 6.6X throughput over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.", "authors": ["Zefan Cai", "Wen Xiao", "Hanshi Sun", "Cheng Luo", "Yikai Zhang", "Ke Wan", "Yucheng Li", "Yeyang Zhou", "Li-Wen Chang", "Jiuxiang Gu", "Zhen Dong", "Anima Anandkumar", "Abedelkadir Asi", "Junjie Hu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24133", "pdf_url": "https://arxiv.org/pdf/2505.24133v4", "arxiv_id": "2505.24133", "doi": null, "citation_count": 19, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Advances in Neural Information Processing Systems", "quality_score": 0.4225}
{"id": "a33014674a0b557a14c2cd7ebdbffaed9c463db15fb671f6ef73e83018cc21b3", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive LoRA Merge with Parameter Pruning for Low-Resource Generation", "abstract": "This study proposes a simple yet effective LoRA merge method to achieve LLM adaptation for low-resource language generation tasks. The LoRA merge technique, which integrates multiple LoRA modules trained on different tasks, has gained attention as an effective and efficient approach for adapting LLMs to target tasks. However, previous methods are limited in adaptability as they keep the LoRA parameters frozen. Additionally, the low-resource problem has been out of their scope. We propose a LoRA merge method that updates and prunes LoRA parameters through fine-tuning with minimal target task data, which allows finer-grained adjustments of LoRA parameters and enhancement of task adaptability. Extensive experiments have been conducted taking summarization as a benchmark task. Our datasets cover various domains and multiple languages of English and Japanese. The results confirm that the proposed method achieves significant and consistent improvements in task adaptability over the previous methods.", "authors": ["Ryota Miyano", "Yuki Arase"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24174", "pdf_url": "https://arxiv.org/pdf/2505.24174v1", "arxiv_id": "2505.24174", "doi": "10.48550/arXiv.2505.24174", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2113}
{"id": "86ce658cccb3900cbaef316fa8b68be0406be301996a52e4e8532c8fcdb06f3c", "sources": ["arxiv", "semantic_scholar"], "title": "ReCalKV: Low-Rank KV Cache Compression via Head Reordering and Offline Calibration", "abstract": "Large language models (LLMs) have demonstrated remarkable performance, but their long-context reasoning remains constrained by the excessive memory required for the Key-Value (KV) cache. This makes KV cache compression a critical step toward efficient long-context inference. Recent methods have explored low-rank techniques to reduce the hidden size of the KV cache. However, they neglect the distinct roles and varying importance of Keys and Values, leading to significant performance drops under high compression. To address this, we propose ReCalKV, a post-training low-rank KV cache compression approach with tailored strategies for Keys and Values. For Keys, we propose Head-wise Similarity aware Reordering (HSR), which clusters structurally similar heads into groups, enabling more accurate low-rank approximation via grouped SVD. For Values, we propose Offline Value Calibration (OVC), which efficiently calibrates the value projection matrix using calibration data without training, ensuring an accurate representation of contextual information. Extensive experiments show that ReCalKV consistently outperforms existing low-rank compression methods, achieving high compression ratios with minimal performance loss. The code and models will be available at:https://github.com/XIANGLONGYAN/ReCalKV.", "authors": ["Xianglong Yan", "Zhiteng Li", "Tianao Zhang", "Haotong Qin", "Linghe Kong", "Yulun Zhang", "Xiaokang Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24357", "pdf_url": "https://arxiv.org/pdf/2505.24357v3", "arxiv_id": "2505.24357", "doi": "10.48550/arXiv.2505.24357", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/XIANGLONGYAN/ReCalKV", "venue": "arXiv.org", "quality_score": 0.2798}
{"id": "1e0fca43cba6b1b6b7a0f95327ffde506dec122031b8b508cc7b27340e9ee335", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba Knockout for Unraveling Factual Information Flow", "abstract": "This paper investigates the flow of factual information in Mamba State-Space Model (SSM)-based language models. We rely on theoretical and empirical connections to Transformer-based architectures and their attention mechanisms. Exploiting this relationship, we adapt attentional interpretability techniques originally developed for Transformers--specifically, the Attention Knockout methodology--to both Mamba-1 and Mamba-2. Using them we trace how information is transmitted and localized across tokens and layers, revealing patterns of subject-token information emergence and layer-wise dynamics. Notably, some phenomena vary between mamba models and Transformer based models, while others appear universally across all models inspected--hinting that these may be inherent to LLMs in general. By further leveraging Mamba's structured factorization, we disentangle how distinct \"features\" either enable token-to-token information exchange or enrich individual tokens, thus offering a unified lens to understand Mamba internal operations.", "authors": ["Nir Endy", "Idan Daniel Grosbard", "Yuval Ran-Milo", "Yonatan Slutzky", "Itay Tshuva", "Raja Giryes"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24244", "pdf_url": "https://arxiv.org/pdf/2505.24244v1", "arxiv_id": "2505.24244", "doi": "10.48550/arXiv.2505.24244", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.181}
{"id": "7ed2f2cf93ede7f4223ec2b4ff13c210835dd43e1e6c46e57cb96999dc4f3adc", "sources": ["arxiv", "semantic_scholar"], "title": "HELM: Hyperbolic Large Language Models via Mixture-of-Curvature Experts", "abstract": "Large language models (LLMs) have shown great success in text modeling tasks across domains. However, natural language exhibits inherent semantic hierarchies and nuanced geometric structure, which current LLMs do not capture completely owing to their reliance on Euclidean operations. Recent studies have also shown that not respecting the geometry of token embeddings leads to training instabilities and degradation of generative capabilities. These findings suggest that shifting to non-Euclidean geometries can better align language models with the underlying geometry of text. We thus propose to operate fully in Hyperbolic space, known for its expansive, scale-free, and low-distortion properties. We thus introduce HELM, a family of HypErbolic Large Language Models, offering a geometric rethinking of the Transformer-based LLM that addresses the representational inflexibility, missing set of necessary operations, and poor scalability of existing hyperbolic LMs. We additionally introduce a Mixture-of-Curvature Experts model, HELM-MICE, where each expert operates in a distinct curvature space to encode more fine-grained geometric structure from text, as well as a dense model, HELM-D. For HELM-MICE, we further develop hyperbolic Multi-Head Latent Attention (HMLA) for efficient, reduced-KV-cache training and inference. For both models, we develop essential hyperbolic equivalents of rotary positional encodings and RMS normalization. We are the first to train fully hyperbolic LLMs at billion-parameter scale, and evaluate them on well-known benchmarks such as MMLU and ARC, spanning STEM problem-solving, general knowledge, and commonsense reasoning. Our results show consistent gains from our HELM architectures -- up to 4% -- over popular Euclidean architectures used in LLaMA and DeepSeek, highlighting the efficacy and enhanced reasoning afforded by hyperbolic geometry in large-scale LM pretraining.", "authors": ["Neil He", "Rishabh Anand", "Hiren Madhu", "Ali Maatouk", "Smita Krishnaswamy", "Leandros Tassiulas", "Menglin Yang", "Rex Ying"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24722", "pdf_url": "https://arxiv.org/pdf/2505.24722v2", "arxiv_id": "2505.24722", "doi": "10.48550/arXiv.2505.24722", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.294}
{"id": "8b43fbd0b27d0af57a857a739af1bc0e24c09876e5e17baed1a1756313c8401b", "sources": ["arxiv", "semantic_scholar"], "title": "Decoding Knowledge Attribution in Mixture-of-Experts: A Framework of Basic-Refinement Collaboration and Efficiency Analysis", "abstract": "The interpretability of Mixture-of-Experts (MoE) models, especially those with heterogeneous designs, remains underexplored. Existing attribution methods for dense models fail to capture dynamic routing-expert interactions in sparse MoE architectures. To address this issue, we propose a cross-level attribution algorithm to analyze sparse MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) against dense models (Qwen 1.5-7B, Llama-7B, Mistral-7B). Results show MoE models achieve 37% higher per-layer efficiency via a \"mid-activation, late-amplification\" pattern: early layers screen experts, while late layers refine knowledge collaboratively. Ablation studies reveal a \"basic-refinement\" framework--shared experts handle general tasks (entity recognition), while routed experts specialize in domain-specific processing (geographic attributes). Semantic-driven routing is evidenced by strong correlations between attention heads and experts (r=0.68), enabling task-aware coordination. Notably, architectural depth dictates robustness: deep Qwen 1.5-MoE mitigates expert failures (e.g., 43% MRR drop in geographic tasks when blocking top-10 experts) through shared expert redundancy, whereas shallow OLMoE suffers severe degradation (76% drop). Task sensitivity further guides design: core-sensitive tasks (geography) require concentrated expertise, while distributed-tolerant tasks (object attributes) leverage broader participation. These insights advance MoE interpretability, offering principles to balance efficiency, specialization, and robustness.", "authors": ["Junzhuo Li", "Bo Wang", "Xiuze Zhou", "Peijie Jiang", "Jia Liu", "Xuming Hu"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24593", "pdf_url": "https://arxiv.org/pdf/2505.24593v2", "arxiv_id": "2505.24593", "doi": "10.48550/arXiv.2505.24593", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.181}
{"id": "d9821ae5eae605aaf65211697fd274c3c17d4926cb8e6e35c9ff6f72a0ad361c", "sources": ["arxiv", "semantic_scholar"], "title": "On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks", "abstract": "Mixture-of-experts networks (MoEs) have demonstrated remarkable efficiency in modern deep learning. Despite their empirical success, the theoretical foundations underlying their ability to model complex tasks remain poorly understood. In this work, we conduct a systematic study of the expressive power of MoEs in modeling complex tasks with two common structural priors: low-dimensionality and sparsity. For shallow MoEs, we prove that they can efficiently approximate functions supported on low-dimensional manifolds, overcoming the curse of dimensionality. For deep MoEs, we show that $\\mathcal{O}(L)$-layer MoEs with $E$ experts per layer can approximate piecewise functions comprising $E^L$ pieces with compositional sparsity, i.e., they can exhibit an exponential number of structured tasks. Our analysis reveals the roles of critical architectural components and hyperparameters in MoEs, including the gating mechanism, expert networks, the number of experts, and the number of layers, and offers natural suggestions for MoE variants.", "authors": ["Mingze Wang", "Weinan E"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24205", "pdf_url": "https://arxiv.org/pdf/2505.24205v2", "arxiv_id": "2505.24205", "doi": "10.48550/arXiv.2505.24205", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.181}
{"id": "d255bf616278e354c5880ede7988d764dbe78ad50eca1e67e3372dab49b16940", "sources": ["arxiv", "semantic_scholar"], "title": "MOPSA: Mixture of Prompt-Experts Based Speaker Adaptation for Elderly Speech Recognition", "abstract": "This paper proposes a novel Mixture of Prompt-Experts based Speaker Adaptation approach (MOPSA) for elderly speech recognition. It allows zero-shot, real-time adaptation to unseen speakers, and leverages domain knowledge tailored to elderly speakers. Top-K most distinctive speaker prompt clusters derived using K-means serve as experts. A router network is trained to dynamically combine clustered prompt-experts. Acoustic and language level variability among elderly speakers are modelled using separate encoder and decoder prompts for Whisper. Experiments on the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets suggest that online MOPSA adaptation outperforms the speaker-independent (SI) model by statistically significant word error rate (WER) or character error rate (CER) reductions of 0.86% and 1.47% absolute (4.21% and 5.40% relative). Real-time factor (RTF) speed-up ratios of up to 16.12 times are obtained over offline batch-mode adaptation.", "authors": ["Chengxi Deng", "Xurong Xie", "Shujie Hu", "Mengzhe Geng", "Yicong Jiang", "Jiankun Zhao", "Jiajun Deng", "Guinan Li", "Youjun Chen", "Huimeng Wang", "Haoning Xu", "Mingyu Cui", "Xunying Liu"], "categories": ["eess.AS"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2025-05-30", "url": "https://arxiv.org/abs/2505.24224", "pdf_url": "https://arxiv.org/pdf/2505.24224v1", "arxiv_id": "2505.24224", "doi": "10.21437/interspeech.2025-1263", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.181}
{"id": "0ec9179963cd3e2247198fae14fb7fa7d328845fa117fa46a33a0bbce65c3b17", "sources": ["arxiv", "semantic_scholar"], "title": "KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction", "abstract": "Transformer-based large language models (LLMs) cache context as key-value (KV) pairs during inference. As context length grows, KV cache sizes expand, leading to substantial memory overhead and increased attention latency. This paper introduces KVzip, a query-agnostic KV cache eviction method enabling effective reuse of compressed KV caches across diverse queries. KVzip quantifies the importance of a KV pair using the underlying LLM to reconstruct original contexts from cached KV pairs, subsequently evicting pairs with lower importance. Extensive empirical evaluations demonstrate that KVzip reduces KV cache size by $3$-$4\\times$ and FlashAttention decoding latency by approximately $2\\times$, with negligible performance loss in question-answering, retrieval, reasoning, and code comprehension tasks. Evaluations include various models such as LLaMA3.1, Qwen2.5, and Gemma3, with context lengths reaching up to 170K tokens. KVzip significantly outperforms existing query-aware KV eviction methods, which suffer from performance degradation even at a 90% cache budget ratio under multi-query scenarios.", "authors": ["Jang-Hyun Kim", "Jinuk Kim", "Sangwoo Kwon", "Jae W. Lee", "Sangdoo Yun", "Hyun Oh Song"], "categories": ["cs.DB", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-29", "url": "https://arxiv.org/abs/2505.23416", "pdf_url": "https://arxiv.org/pdf/2505.23416v2", "arxiv_id": "2505.23416", "doi": "10.48550/arXiv.2505.23416", "citation_count": 44, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/snu-mllab/KVzip", "venue": "arXiv.org", "quality_score": 0.4515}
{"id": "c2792773fc0616822d8806f765cf38c38f9ad6a9873c2b4ace8e9e9980239fc1", "sources": ["arxiv", "semantic_scholar"], "title": "Two Is Better Than One: Rotations Scale LoRAs", "abstract": "Scaling Low-Rank Adaptation (LoRA)-based Mixture-of-Experts (MoE) facilitates large language models (LLMs) to efficiently adapt to diverse tasks. However, traditional gating mechanisms that route inputs to the best experts may fundamentally hinder LLMs' scalability, leading to poor generalization and underfitting issues. We identify that the root cause lies in the restricted expressiveness of existing weighted-sum mechanisms, both within and outside the convex cone of LoRA representations. This motivates us to propose RadarGate, a novel geometrically inspired gating method that introduces rotational operations of LoRAs representations to boost the expressiveness and facilitate richer feature interactions among multiple LoRAs for scalable LLMs. Specifically, we first fuse each LoRA representation to other LoRAs using a learnable component and then feed the output to a rotation matrix. This matrix involves learnable parameters that define the relative angular relationship between LoRA representations. Such a simple yet effective mechanism provides an extra degree of freedom, facilitating the learning of cross-LoRA synergies and properly tracking the challenging poor generalization and underfitting issues as the number of LoRA grows. Extensive experiments on 6 public benchmarks across 21 tasks show the effectiveness of our RadarGate for scaling LoRAs. We also provide valuable insights, revealing that the rotations to each pair of representations are contrastive, encouraging closer alignment of semantically similar representations during geometrical transformation while pushing distance ones further apart. We will release our code to the community.", "authors": ["Hongcan Guo", "Guoshun Nan", "Yuan Yang", "Diyang Zhang", "Haotian Li", "Zhican Chen", "Qinchuan Zhou", "Yuhan Ran", "Xinye Cao", "Sicong Leng", "Xiaofeng Tao", "Xudong Jiang"], "categories": ["cs.LG", "cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-29", "url": "https://arxiv.org/abs/2505.23184", "pdf_url": "https://arxiv.org/pdf/2505.23184v1", "arxiv_id": "2505.23184", "doi": "10.48550/arXiv.2505.23184", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1799}
{"id": "4b7479f723b6a3e6e103c1103f87c3c02c8c1b439a5ab6c5cd2ff77839aa796d", "sources": ["arxiv", "semantic_scholar"], "title": "Noise-Robustness Through Noise: A Framework combining Asymmetric LoRA with Poisoning MoE", "abstract": "Current parameter-efficient fine-tuning methods for adapting pre-trained language models to downstream tasks are susceptible to interference from noisy data. Conventional noise-handling approaches either rely on laborious data pre-processing or employ model architecture modifications prone to error accumulation. In contrast to existing noise-process paradigms, we propose a noise-robust adaptation method via asymmetric LoRA poisoning experts (LoPE), a novel framework that enhances model robustness to noise only with generated noisy data. Drawing inspiration from the mixture-of-experts architecture, LoPE strategically integrates a dedicated poisoning expert in an asymmetric LoRA configuration. Through a two-stage paradigm, LoPE performs noise injection on the poisoning expert during fine-tuning to enhance its noise discrimination and processing ability. During inference, we selectively mask the dedicated poisoning expert to leverage purified knowledge acquired by normal experts for noise-robust output. Extensive experiments demonstrate that LoPE achieves strong performance and robustness purely through the low-cost noise injection, which completely eliminates the requirement of data cleaning.", "authors": ["Zhaokun Wang", "Jinyu Guo", "Jingwen Pu", "Lingfeng Chen", "Hongli Pu", "Jie Ou", "Libo Qin", "Wenhong Tian"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-29", "url": "https://arxiv.org/abs/2505.23868", "pdf_url": "https://arxiv.org/pdf/2505.23868v5", "arxiv_id": "2505.23868", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "5baf07e9ee1845e67957340400466c440ad52e365c0e7375a0aec96efa7914c5", "sources": ["arxiv", "semantic_scholar"], "title": "Mustafar: Promoting Unstructured Sparsity for KV Cache Pruning in LLM Inference", "abstract": "We demonstrate that unstructured sparsity significantly improves KV cache compression for LLMs, enabling sparsity levels up to 70% without compromising accuracy or requiring fine-tuning. We conduct a systematic exploration of pruning strategies and find per-token magnitude-based pruning as highly effective for both Key and Value caches under unstructured sparsity, surpassing prior structured pruning schemes. The Key cache benefits from prominent outlier elements, while the Value cache surprisingly benefits from a simple magnitude-based pruning despite its uniform distribution. KV cache size is the major bottleneck in decode performance due to high memory overhead for large context lengths. To address this, we use a bitmap-based sparse format and a custom attention kernel capable of compressing and directly computing over compressed caches pruned to arbitrary sparsity patterns, significantly accelerating memory-bound operations in decode computations and thereby compensating for the overhead of runtime pruning and compression. Our custom attention kernel coupled with the bitmap-based format delivers substantial compression of KV cache upto 45% of dense inference and thereby enables longer context length and increased tokens/sec throughput of upto 2.23x compared to dense inference. Our pruning mechanism and sparse attention kernel is available at https://github.com/dhjoo98/mustafar.", "authors": ["Donghyeon Joo", "Helya Hosseini", "Ramyad Hadidi", "Bahar Asgari"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2505.22913", "pdf_url": "https://arxiv.org/pdf/2505.22913v2", "arxiv_id": "2505.22913", "doi": "10.48550/arXiv.2505.22913", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/dhjoo98/mustafar", "venue": "arXiv.org", "quality_score": 0.2762}
{"id": "0205fbc23350292a9cf2a1242c894ab5573a4687ccd57e3b8aad170e66f44d37", "sources": ["arxiv", "semantic_scholar"], "title": "MoRE: A Mixture of Low-Rank Experts for Adaptive Multi-Task Learning", "abstract": "With the rapid development of Large Language Models (LLMs), Parameter-Efficient Fine-Tuning (PEFT) methods have gained significant attention, which aims to achieve efficient fine-tuning of LLMs with fewer parameters. As a representative PEFT method, Low-Rank Adaptation (LoRA) introduces low-rank matrices to approximate the incremental tuning parameters and achieves impressive performance over multiple scenarios. After that, plenty of improvements have been proposed for further improvement. However, these methods either focus on single-task scenarios or separately train multiple LoRA modules for multi-task scenarios, limiting the efficiency and effectiveness of LoRA in multi-task scenarios. To better adapt to multi-task fine-tuning, in this paper, we propose a novel Mixture of Low-Rank Experts (MoRE) for multi-task PEFT. Specifically, instead of using an individual LoRA for each task, we align different ranks of LoRA module with different tasks, which we named low-rank experts. Moreover, we design a novel adaptive rank selector to select the appropriate expert for each task. By jointly training low-rank experts, MoRE can enhance the adaptability and efficiency of LoRA in multi-task scenarios. Finally, we conduct extensive experiments over multiple multi-task benchmarks along with different LLMs to verify model performance. Experimental results demonstrate that compared to traditional LoRA and its variants, MoRE significantly improves the performance of LLMs in multi-task scenarios and incurs no additional inference cost. We also release the model and code to facilitate the community.", "authors": ["Dacao Zhang", "Kun Zhang", "Shimao Chu", "Le Wu", "Xin Li", "Si Wei"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2505.22694", "pdf_url": "https://arxiv.org/pdf/2505.22694v1", "arxiv_id": "2505.22694", "doi": "10.48550/arXiv.2505.22694", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2785}
{"id": "5c1bcb73bfb8d41ce08c711137fc2a3b9e5336d06682f0fb149055e386a8404b", "sources": ["arxiv", "semantic_scholar"], "title": "FIER: Fine-Grained and Efficient KV Cache Retrieval for Long-context LLM Inference", "abstract": "The Key-Value (KV) cache reading latency increases significantly with context lengths, hindering the efficiency of long-context LLM inference. To address this, previous works propose retaining a small fraction of KV cache based on token importance. For example, KV eviction uses static heuristics to retain tokens, while KV retrieval dynamically selects query-relevant tokens for more adaptive cache management. However, we observe that important tokens are often sparsely distributed across the long context. This sparsity makes existing page-level KV retrieval inaccurate, as each page may include irrelevant tokens and miss critical ones. In this work, we propose Fier, a \\underline{Fi}ne-Grained and \\underline{E}fficient KV cache \\underline{R}etrieval method. Fier uses 1-bit quantized keys to estimate the importance of each token, resulting in efficient and precise retrieval. Experiments show that Fier matches full KV performance using only 11\\% of the cache budget across various long-context tasks, reducing decoding latency by 1.2$\\times$ to 1.5$\\times$.Code is available at https://github.com/SimWangArizona/FIER", "authors": ["Dongwei Wang", "Zijie Liu", "Song Wang", "Yuxin Ren", "Jianing Deng", "Jingtong Hu", "Tianlong Chen", "Huanrui Yang"], "categories": ["cs.DB"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2508.08256", "pdf_url": "https://arxiv.org/pdf/2508.08256v2", "arxiv_id": "2508.08256", "doi": "10.48550/arXiv.2508.08256", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SimWangArizona/FIER", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2762}
{"id": "133c8b83d21acdfc50d881dcddf2a16956e0e6c16ab080da4c4164de2e878836", "sources": ["arxiv", "semantic_scholar"], "title": "EvoMoE: Expert Evolution in Mixture of Experts for Multimodal Large Language Models", "abstract": "Recent advancements have shown that the Mixture of Experts (MoE) approach significantly enhances the capacity of large language models (LLMs) and improves performance on downstream tasks. Building on these promising results, multi-modal large language models (MLLMs) have increasingly adopted MoE techniques. However, existing multi-modal MoE tuning methods typically face two key challenges: expert uniformity and router rigidity. Expert uniformity occurs because MoE experts are often initialized by simply replicating the FFN parameters from LLMs, leading to homogenized expert functions and weakening the intended diversification of the MoE architecture. Meanwhile, router rigidity stems from the prevalent use of static linear routers for expert selection, which fail to distinguish between visual and textual tokens, resulting in similar expert distributions for image and text. To address these limitations, we propose EvoMoE, an innovative MoE tuning framework. EvoMoE introduces a meticulously designed expert initialization strategy that progressively evolves multiple robust experts from a single trainable expert, a process termed expert evolution that specifically targets severe expert homogenization. Furthermore, we introduce the Dynamic Token-aware Router (DTR), a novel routing mechanism that allocates input tokens to appropriate experts based on their modality and intrinsic token values. This dynamic routing is facilitated by hypernetworks, which dynamically generate routing weights tailored for each individual token. Extensive experiments demonstrate that EvoMoE significantly outperforms other sparse MLLMs across a variety of multi-modal benchmarks, including MME, MMBench, TextVQA, and POPE. Our results highlight the effectiveness of EvoMoE in enhancing the performance of MLLMs by addressing the critical issues of expert uniformity and router rigidity.", "authors": ["Linglin Jing", "Yuting Gao", "Zhigang Wang", "Wang Lan", "Yiwen Tang", "Wenhai Wang", "Kaipeng Zhang", "Qingpei Guo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2505.23830", "pdf_url": "https://arxiv.org/pdf/2505.23830v1", "arxiv_id": "2505.23830", "doi": "10.48550/arXiv.2505.23830", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.1788}
{"id": "08880246b5298aaf12fd2711edbe84c614649e7766b1392f1f8154f80641d848", "sources": ["arxiv", "semantic_scholar"], "title": "EFIM: Efficient Serving of LLMs for Infilling Tasks with Improved KV Cache Reuse", "abstract": "Large language models (LLMs) are often used for infilling tasks, which involve predicting or generating missing information in a given text. These tasks typically require multiple interactions with similar context. To reduce the computation of repeated historical tokens, cross-request key-value (KV) cache reuse, a technique that stores and reuses intermediate computations, has become a crucial method in multi-round interactive services. However, in infilling tasks, the KV cache reuse is often hindered by the structure of the prompt format, which typically consists of a prefix and suffix relative to the insertion point. Specifically, the KV cache of the prefix or suffix part is frequently invalidated as the other part (suffix or prefix) is incrementally generated. To address the issue, we propose EFIM, a transformed prompt format of FIM to unleash the performance potential of KV cache reuse. Although the transformed prompt can solve the inefficiency, it exposes subtoken generation problems in current LLMs, where they have difficulty generating partial words accurately. Therefore, we introduce a fragment tokenization training method which splits text into multiple fragments before tokenization during data processing. Experiments on two representative LLMs show that LLM serving with EFIM can lower the latency by 52% and improve the throughput by 98% while maintaining the original infilling capability. EFIM's source code is publicly available at https://github.com/gty111/EFIM.", "authors": ["Tianyu Guo", "Hande Dong", "Yichong Leng", "Feng Liu", "Cheater Lin", "Nong Xiao", "Xianwei Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2505.21889", "pdf_url": "https://arxiv.org/pdf/2505.21889v2", "arxiv_id": "2505.21889", "doi": "10.48550/arXiv.2505.21889", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/gty111/EFIM", "venue": "European Conference on Parallel Processing", "quality_score": 0.2762}
{"id": "d0c45ed4e43af45f82f41a0619860bc541db51b4fc50c61709407786c315229b", "sources": ["arxiv", "semantic_scholar"], "title": "Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding", "abstract": "Diffusion-based large language models (Diffusion LLMs) have shown promise for non-autoregressive text generation with parallel decoding capabilities. However, the practical inference speed of open-sourced Diffusion LLMs often lags behind autoregressive models due to the lack of Key-Value (KV) Cache and quality degradation when decoding multiple tokens simultaneously. To bridge this gap, we introduce a novel block-wise approximate KV Cache mechanism tailored for bidirectional diffusion models, enabling cache reuse with negligible performance drop. Additionally, we identify the root cause of generation quality degradation in parallel decoding as the disruption of token dependencies under the conditional independence assumption. To address this, we propose a confidence-aware parallel decoding strategy that selectively decodes tokens exceeding a confidence threshold, mitigating dependency violations and maintaining generation quality. Experimental results on LLaDA and Dream models across multiple LLM benchmarks demonstrate up to \\textbf{27.6$\\times$ throughput} improvement with minimal accuracy loss, closing the performance gap with autoregressive models and paving the way for practical deployment of Diffusion LLMs.", "authors": ["Chengyue Wu", "Hao Zhang", "Shuchen Xue", "Zhijian Liu", "Shizhe Diao", "Ligeng Zhu", "Ping Luo", "Song Han", "Enze Xie"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2505.22618", "pdf_url": "https://arxiv.org/pdf/2505.22618v3", "arxiv_id": "2505.22618", "doi": "10.48550/arXiv.2505.22618", "citation_count": 306, "influential_citation_count": 80, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.9542}
{"id": "4ec8177a5522b4116daab53b6de2057a4e6f3481e1e5da2c4c0b743cb4b959f7", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Retrieval Experts for Reasoning-Guided Multimodal Knowledge Exploitation", "abstract": "Multimodal Retrieval-Augmented Generation (MRAG) has shown promise in mitigating hallucinations in Multimodal Large Language Models (MLLMs) by incorporating external knowledge. However, existing methods typically adhere to rigid retrieval paradigms by mimicking fixed retrieval trajectories and thus fail to fully exploit the knowledge of different retrieval experts through dynamic interaction based on the model's knowledge needs or evolving reasoning states. To overcome this limitation, we introduce Mixture-of-Retrieval Experts (MoRE), a novel framework that enables MLLMs to collaboratively interact with diverse retrieval experts for more effective knowledge exploitation. Specifically, MoRE learns to dynamically determine which expert to engage with, conditioned on the evolving reasoning state. To effectively train this capability, we propose Stepwise Group Relative Policy Optimization (Step-GRPO), which goes beyond sparse outcome-based supervision by encouraging MLLMs to interact with multiple retrieval experts and synthesize fine-grained rewards, thereby teaching the MLLM to fully coordinate all experts when answering a given query. Experimental results on diverse open-domain QA benchmarks demonstrate the effectiveness of MoRE, achieving average performance gains of over 7% compared to competitive baselines. Notably, MoRE exhibits strong adaptability by dynamically coordinating heterogeneous experts to precisely locate relevant information, validating its capability for robust, reasoning-driven expert collaboration. All codes and data are released on https://github.com/OpenBMB/MoRE.", "authors": ["Chunyi Peng", "Zhipeng Xu", "Zhenghao Liu", "Yishan Li", "Yukun Yan", "Shuo Wang", "Yu Gu", "Minghe Yu", "Ge Yu", "Maosong Sun"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-28", "url": "https://arxiv.org/abs/2505.22095", "pdf_url": "https://arxiv.org/pdf/2505.22095v2", "arxiv_id": "2505.22095", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/OpenBMB/MoRE", "venue": null, "quality_score": 0.2112}
{"id": "c6e0c9d671941ebbf93eb23e3aa3042ea01afd578dd850d06d0a6d044556c66c", "sources": ["arxiv", "semantic_scholar"], "title": "ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation", "abstract": "Diffusion models have demonstrated exceptional capabilities in generating high-fidelity images. However, their iterative denoising process results in significant computational overhead during inference, limiting their practical deployment in resource-constrained environments. Existing acceleration methods often adopt uniform strategies that fail to capture the temporal variations during diffusion generation, while the commonly adopted sequential pruning-then-fine-tuning strategy suffers from sub-optimality due to the misalignment between pruning decisions made on pretrained weights and the model's final parameters. To address these limitations, we introduce ALTER: All-in-One Layer Pruning and Temporal Expert Routing, a unified framework that transforms diffusion models into a mixture of efficient temporal experts. ALTER achieves a single-stage optimization that unifies layer pruning, expert routing, and model fine-tuning by employing a trainable hypernetwork, which dynamically generates layer pruning decisions and manages timestep routing to specialized, pruned expert sub-networks throughout the ongoing fine-tuning of the UNet. This unified co-optimization strategy enables significant efficiency gains while preserving high generative quality. Specifically, ALTER achieves same-level visual fidelity to the original 50-step Stable Diffusion v2.1 model while utilizing only 25.9% of its total MACs with just 20 inference steps and delivering a 3.64x speedup through 35% sparsity.", "authors": ["Xiaomeng Yang", "Lei Lu", "Qihui Fan", "Changdi Yang", "Juyi Lin", "Yanzhi Wang", "Xuan Zhang", "Shangqian Gao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-27", "url": "https://arxiv.org/abs/2505.21817", "pdf_url": "https://arxiv.org/pdf/2505.21817v1", "arxiv_id": "2505.21817", "doi": "10.48550/arXiv.2505.21817", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1776}
{"id": "ef75723a797c024eb7f3baf12df51ad90bff17ebd314a62e9e7eeb5e316586bf", "sources": ["arxiv", "semantic_scholar"], "title": "Pangu Pro MoE: Mixture of Grouped Experts for Efficient Sparsity", "abstract": "The surgence of Mixture of Experts (MoE) in Large Language Models promises a small price of execution cost for a much larger model parameter count and learning capacity, because only a small fraction of parameters are activated for each input token. However, it is commonly observed that some experts are activated far more often than others, leading to system inefficiency when running the experts on different devices in parallel. Therefore, we introduce Mixture of Grouped Experts (MoGE), which groups the experts during selection and balances the expert workload better than MoE in nature. It constrains tokens to activate an equal number of experts within each predefined expert group. When a model execution is distributed on multiple devices, this architectural design ensures a balanced computational load across devices, significantly enhancing throughput, particularly for the inference phase. Further, we build Pangu Pro MoE on Ascend NPUs, a sparse model based on MoGE with 72 billion total parameters, 16 billion of which are activated for each token. The configuration of Pangu Pro MoE is optimized for Ascend 300I Duo and 800I A2 through extensive system simulation studies. Our experiments indicate that MoGE indeed leads to better expert load balancing and more efficient execution for both model training and inference on Ascend NPUs. The inference performance of Pangu Pro MoE achieves 1148 tokens/s per card and can be further improved to 1528 tokens/s per card by speculative acceleration, outperforming comparable 32B and 72B Dense models. Furthermore, we achieve an excellent cost-to-performance ratio for model inference on Ascend 300I Duo. Our studies show that Ascend NPUs are capable of training Pangu Pro MoE with massive parallelization to make it a leading model within the sub-100B total parameter class, outperforming prominent open-source models like GLM-Z1-32B and Qwen3-32B.", "authors": ["Yehui Tang", "Xiaosong Li", "Fangcheng Liu", "Wei Guo", "Hang Zhou", "Yaoyuan Wang", "Kai Han", "Xianzhi Yu", "Jinpeng Li", "Hui Zang", "Fei Mi", "Xiaojun Meng", "Zhicheng Liu", "Hanting Chen", "Binfan Zheng", "Can Chen", "Youliang Yan", "Ruiming Tang", "Peifeng Qin", "Xinghao Chen", "Dacheng Tao", "Yunhe Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-27", "url": "https://arxiv.org/abs/2505.21411", "pdf_url": "https://arxiv.org/pdf/2505.21411v2", "arxiv_id": "2505.21411", "doi": "10.48550/arXiv.2505.21411", "citation_count": 16, "influential_citation_count": 3, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3076}
{"id": "2f5ad05c41b713f4bf8dcc6b3520f43f12cf016e8e89ba5a12e3ff5db0932376", "sources": ["arxiv", "semantic_scholar"], "title": "Uni3D-MoE: Scalable Multimodal 3D Scene Understanding via Mixture of Experts", "abstract": "Recent advancements in multimodal large language models (MLLMs) have demonstrated considerable potential for comprehensive 3D scene understanding. However, existing approaches typically utilize only one or a limited subset of 3D modalities, resulting in incomplete representations of 3D scenes and reduced interpretive accuracy. Furthermore, different types of queries inherently depend on distinct modalities, indicating that uniform processing of all modality tokens may fail to effectively capture query-specific context. To address these challenges, we propose Uni3D-MoE, a sparse Mixture-of-Experts (MoE)-based 3D MLLM designed to enable adaptive 3D multimodal fusion. Specifically, Uni3D-MoE integrates a comprehensive set of 3D modalities, including multi-view RGB and depth images, bird's-eye-view (BEV) maps, point clouds, and voxel representations. At its core, our framework employs a learnable routing mechanism within the sparse MoE-based large language model, dynamically selecting appropriate experts at the token level. Each expert specializes in processing multimodal tokens based on learned modality preferences, thus facilitating flexible collaboration tailored to diverse task-specific requirements. Extensive evaluations on standard 3D scene understanding benchmarks and specialized datasets demonstrate the efficacy of Uni3D-MoE.", "authors": ["Yue Zhang", "Yingzhao Jian", "Hehe Fan", "Yi Yang", "Roger Zimmermann"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-27", "url": "https://arxiv.org/abs/2505.21079", "pdf_url": "https://arxiv.org/pdf/2505.21079v1", "arxiv_id": "2505.21079", "doi": "10.48550/arXiv.2505.21079", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1776}
{"id": "5a01b1a38b356b8b59a9cac263dbdada2146cab97bfae6bbd01610d43f1653e5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of LoRA Experts for Low-Resourced Multi-Accent Automatic Speech Recognition", "abstract": "We aim to improve the robustness of Automatic Speech Recognition (ASR) systems against non-native speech, particularly in low-resourced multi-accent settings. We introduce Mixture of Accent-Specific LoRAs (MAS-LoRA), a fine-tuning method that leverages a mixture of Low-Rank Adaptation (LoRA) experts, each specialized in a specific accent. This method can be used when the accent is known or unknown at inference time, without the need to fine-tune the model again. Our experiments, conducted using Whisper on the L2-ARCTIC corpus, demonstrate significant improvements in Word Error Rate compared to regular LoRA and full fine-tuning when the accent is unknown. When the accent is known, the results further improve. Furthermore, MAS-LoRA shows less catastrophic forgetting than the other fine-tuning methods. To the best of our knowledge, this is the first use of a mixture of LoRA experts for non-native multi-accent ASR.", "authors": ["Raphaël Bagat", "Irina Illina", "Emmanuel Vincent"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.20006", "pdf_url": "https://arxiv.org/pdf/2505.20006v1", "arxiv_id": "2505.20006", "doi": "10.21437/Interspeech.2025-775", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.1765}
{"id": "a8220da8209246fbc8ffe5db0a333c8418485a281b71e64bab3f7370b92b35c7", "sources": ["arxiv", "semantic_scholar"], "title": "Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression", "abstract": "Visual Autoregressive (VAR) modeling has garnered significant attention for its innovative next-scale prediction approach, which yields substantial improvements in efficiency, scalability, and zero-shot generalization. Nevertheless, the coarse-to-fine methodology inherent in VAR results in exponential growth of the KV cache during inference, causing considerable memory consumption and computational redundancy. To address these bottlenecks, we introduce ScaleKV, a novel KV cache compression framework tailored for VAR architectures. ScaleKV leverages two critical observations: varying cache demands across transformer layers and distinct attention patterns at different scales. Based on these insights, ScaleKV categorizes transformer layers into two functional groups: drafters and refiners. Drafters exhibit dispersed attention across multiple scales, thereby requiring greater cache capacity. Conversely, refiners focus attention on the current token map to process local details, consequently necessitating substantially reduced cache capacity. ScaleKV optimizes the multi-scale inference pipeline by identifying scale-specific drafters and refiners, facilitating differentiated cache management tailored to each scale. Evaluation on the state-of-the-art text-to-image VAR model family, Infinity, demonstrates that our approach effectively reduces the required KV cache memory to 10% while preserving pixel-level fidelity.", "authors": ["Kunjun Li", "Zigeng Chen", "Cheng-Yen Yang", "Jenq-Neng Hwang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.19602", "pdf_url": "https://arxiv.org/pdf/2505.19602v1", "arxiv_id": "2505.19602", "doi": "10.48550/arXiv.2505.19602", "citation_count": 21, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3356}
{"id": "6a67d263a8bffd7c77c8c951d36511110f9dc1a9f31ebcafbc086cef18c92dbd", "sources": ["arxiv", "semantic_scholar"], "title": "TailorKV: A Hybrid Framework for Long-Context Inference via Tailored KV Cache Optimization", "abstract": "The Key-Value (KV) cache in generative large language models (LLMs) introduces substantial memory overhead. Existing works mitigate this burden by offloading or compressing the KV cache. However, loading the entire cache incurs significant latency due to PCIe bandwidth bottlenecks in CPU-GPU communication, while aggressive compression causes notable performance degradation. We identify that certain layers in the LLM need to maintain global information and are unsuitable for selective loading. In contrast, other layers primarily focus on a few tokens with dominant activations that potentially incur substantial quantization error. This observation leads to a key insight that loading dominant tokens and quantizing all tokens can complement each other. Building on this insight, we propose a hybrid compression method, TailorKV, which seamlessly integrates quantization and offloading. TailorKV develops an inference framework along with a hardware-friendly implementation that leverages these complementary characteristics. Extensive long-context evaluations exhibit that TailorKV achieves nearly lossless performance under aggressive compression settings, outperforming the state-of-the-art. Particularly, the Llama-3.1-8B with 128k context can be served within a single RTX 3090 GPU, reaching 82 ms per token during decoding.", "authors": ["Dingyu Yao", "Bowen Shen", "Zheng Lin", "Wei Liu", "Jian Luan", "Bin Wang", "Weiping Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.19586", "pdf_url": "https://arxiv.org/pdf/2505.19586v2", "arxiv_id": "2505.19586", "doi": "10.48550/arXiv.2505.19586", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2386}
{"id": "8b31c535813a39491d18b78637420bf7cfc9f059f04a045b27956d0cc70c160d", "sources": ["arxiv", "semantic_scholar"], "title": "AdaTP: Attention-Debiased Token Pruning for Video Large Language Models", "abstract": "Video Large Language Models (Video LLMs) have achieved remarkable results in video understanding tasks. However, they often suffer from heavy computational overhead due to the large number of visual tokens generated from multiple video frames. Existing visual token compression methods often rely on attention scores from language models as guidance. However, these scores exhibit inherent biases: global bias reflects a tendency to focus on the two ends of the visual token sequence, while local bias leads to an over-concentration on the same spatial positions across different frames. To address the issue of attention bias, we propose $\\textbf{A}$ttention-$\\textbf{D}$ebi$\\textbf{a}$sed $\\textbf{T}$oken $\\textbf{P}$runing for Video Large Language Models ($\\textbf{AdaTP}$), a novel token pruning pipeline for Video LLMs. AdaTP integrates two dedicated debiasing modules into the pipeline, targeting global attention bias and local attention bias, respectively. Without the need for additional training, our method significantly reduces the computational overhead of Video LLMs while retaining the performance of vanilla models. Extensive evaluation shows that AdaTP achieves state-of-the-art performance in various commonly used video understanding benchmarks. In particular, on LLaVA-OneVision-7B, AdaTP maintains performance without degradation while using only up to $27.3\\%$ FLOPs compared to the vanilla model. Our code will be released soon.", "authors": ["Fengyuan Sun", "Leqi Shen", "Hui Chen", "Sicheng Zhao", "Jungong Han", "Guiguang Ding"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.20100", "pdf_url": "https://arxiv.org/pdf/2505.20100v1", "arxiv_id": "2505.20100", "doi": "10.48550/arXiv.2505.20100", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.1765}
{"id": "1ee89a81a3cf44f1c28376fb4d5a80ff94beb873ca0f8817424934cd7b59a2df", "sources": ["arxiv", "semantic_scholar"], "title": "NEXT: Multi-Grained Mixture of Experts via Text-Modulation for Multi-Modal Object Re-Identification", "abstract": "Multi-modal object Re-IDentification (ReID) aims to obtain complete identity features across heterogeneous modalities. However, most existing methods rely on implicit feature fusion modules, making it difficult to model fine-grained recognition patterns under various challenges in real world. Benefiting from the powerful Multi-modal Large Language Models (MLLMs), the object appearances are effectively translated into descriptive captions. In this paper, we propose a reliable caption generation pipeline based on attribute confidence, which significantly reduces the unknown recognition rate of MLLMs and improves the quality of generated text. Additionally, to model diverse identity patterns, we propose a novel ReID framework, named NEXT, the Multi-grained Mixture of Experts via Text-Modulation for Multi-modal Object Re-Identification. Specifically, we decouple the recognition problem into semantic and structural branches to separately capture fine-grained appearance features and coarsegrained structure features. For semantic recognition, we first propose a Text-Modulated Semantic Experts (TMSE), which randomly samples high-quality captions to modulate experts capturing semantic features and mining inter-modality complementary cues. Second, to recognize structure features, we propose a Context-Shared Structure Experts (CSSE), which focuses on the holistic object structure and maintains identity structural consistency via a soft routing mechanism. Finally, we propose a Multi-Grained Features Aggregation (MGFA), which adopts a unified fusion strategy to effectively integrate multi-grained expert features into the final identity representations. Extensive experiments on two public person datasets and three vehicle datasets demonstrate the effectiveness of our method, showing that it significantly outperforms existing state-of-the-art methods.", "authors": ["Shihao Li", "Huaibo Huang", "Junxian Duan", "Aihua Zheng", "Jin Tang", "Jixin Ma"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.20001", "pdf_url": "https://arxiv.org/pdf/2505.20001v5", "arxiv_id": "2505.20001", "doi": "10.48550/arXiv.2505.20001", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1765}
{"id": "d804d7fe3d1b3a1f5345d78e0368614893e82b34e277332cdf1799e984fd63f0", "sources": ["arxiv", "semantic_scholar"], "title": "I2MoE: Interpretable Multimodal Interaction-aware Mixture-of-Experts", "abstract": "Modality fusion is a cornerstone of multimodal learning, enabling information integration from diverse data sources. However, vanilla fusion methods are limited by (1) inability to account for heterogeneous interactions between modalities and (2) lack of interpretability in uncovering the multimodal interactions inherent in the data. To this end, we propose I2MoE (Interpretable Multimodal Interaction-aware Mixture of Experts), an end-to-end MoE framework designed to enhance modality fusion by explicitly modeling diverse multimodal interactions, as well as providing interpretation on a local and global level. First, I2MoE utilizes different interaction experts with weakly supervised interaction losses to learn multimodal interactions in a data-driven way. Second, I2MoE deploys a reweighting model that assigns importance scores for the output of each interaction expert, which offers sample-level and dataset-level interpretation. Extensive evaluation of medical and general multimodal datasets shows that I2MoE is flexible enough to be combined with different fusion techniques, consistently improves task performance, and provides interpretation across various real-world scenarios. Code is available at https://github.com/Raina-Xin/I2MoE.", "authors": ["Jiayi Xin", "Sukwon Yun", "Jie Peng", "Inyoung Choi", "Jenna L. Ballard", "Tianlong Chen", "Qi Long"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Medicine", "Computer Science"], "published_date": "2025-05-25", "url": "https://arxiv.org/abs/2505.19190", "pdf_url": "https://arxiv.org/pdf/2505.19190v1", "arxiv_id": "2505.19190", "doi": "10.48550/arXiv.2505.19190", "citation_count": 26, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/Raina-Xin/I2MoE", "venue": "International Conference on Machine Learning", "quality_score": 0.3578}
{"id": "528e4d5215c80c4a346460d23d4f9c154e22c0a7cfb554df7fbef52d8ae4cb6b", "sources": ["arxiv", "semantic_scholar"], "title": "PM-KVQ: Progressive Mixed-precision KV Cache Quantization for Long-CoT LLMs", "abstract": "Recently, significant progress has been made in developing reasoning-capable Large Language Models (LLMs) through long Chain-of-Thought (CoT) techniques. However, this long-CoT reasoning process imposes substantial memory overhead due to the large Key-Value (KV) Cache memory overhead. Post-training KV Cache quantization has emerged as a promising compression technique and has been extensively studied in short-context scenarios. However, directly applying existing methods to long-CoT LLMs causes significant performance degradation due to the following two reasons: (1) Large cumulative error: Existing methods fail to adequately leverage available memory, and they directly quantize the KV Cache during each decoding step, leading to large cumulative quantization error. (2) Short-context calibration: Due to Rotary Positional Embedding (RoPE), the use of short-context data during calibration fails to account for the distribution of less frequent channels in the Key Cache, resulting in performance loss. We propose Progressive Mixed-Precision KV Cache Quantization (PM-KVQ) for long-CoT LLMs to address the above issues in two folds: (1) To reduce cumulative error, we design a progressive quantization strategy to gradually lower the bit-width of KV Cache in each block. Then, we propose block-wise memory allocation to assign a higher bit-width to more sensitive transformer blocks. (2) To increase the calibration length without additional overhead, we propose a new calibration strategy with positional interpolation that leverages short calibration data with positional interpolation to approximate the data distribution of long-context data. Extensive experiments on 7B-70B long-CoT LLMs show that PM-KVQ improves reasoning benchmark performance by up to 8% over SOTA baselines under the same memory budget. Our code is available at https://github.com/thu-nics/PM-KVQ.", "authors": ["Tengxuan Liu", "Shiyao Li", "Jiayi Yang", "Tianchen Zhao", "Feng Zhou", "Xiaohui Song", "Guohao Dai", "Shengen Yan", "Huazhong Yang", "Yu Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2505.18610", "pdf_url": "https://arxiv.org/pdf/2505.18610v1", "arxiv_id": "2505.18610", "doi": "10.48550/arXiv.2505.18610", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/thu-nics/PM-KVQ", "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "1ca5f0f0d3da940a24065d0c53de9d53316538767e4d9ef9289eeb860ea3db14", "sources": ["arxiv", "semantic_scholar"], "title": "$μ$-MoE: Test-Time Pruning as Micro-Grained Mixture-of-Experts", "abstract": "To tackle the huge computational demand of large foundation models, activation-aware compression techniques without retraining have been introduced. However, since these rely on calibration data, domain shift may arise for unknown downstream tasks. With a computationally efficient calibration, activation-aware pruning can be executed for every prompt adaptively, yet achieving reduced complexity at inference. We formulate it as a mixture of micro-experts, called $μ$-MoE. Several experiments demonstrate that $μ$-MoE can dynamically adapt to task/prompt-dependent structured sparsity on the fly.", "authors": ["Toshiaki Koike-Akino", "Jing Liu", "Ye Wang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2505.18451", "pdf_url": "https://arxiv.org/pdf/2505.18451v1", "arxiv_id": "2505.18451", "doi": "10.48550/arXiv.2505.18451", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1742}
{"id": "f7696d30b22861945aaacbd66c3dbd3c69a50eeadc4d99cbabaa882cc2da62c7", "sources": ["arxiv", "semantic_scholar"], "title": "Lookahead Q-Cache: Achieving More Consistent KV Cache Eviction via Pseudo Query", "abstract": "Large language models (LLMs) rely on key-value cache (KV cache) to accelerate decoding by reducing redundant computations. However, the KV cache memory usage grows substantially with longer text sequences, posing challenges for efficient deployment. Existing KV cache eviction methods prune tokens using prefilling-stage attention scores, causing inconsistency with actual inference queries, especially under tight memory budgets. In this paper, we propose Lookahead Q-Cache (LAQ), a novel eviction framework that generates low-cost pseudo lookahead queries to better approximate the true decoding-stage queries. By using these lookahead queries as the observation window for importance estimation, LAQ achieves more consistent and accurate KV cache eviction aligned with real inference scenarios. Experimental results on LongBench and Needle-in-a-Haystack benchmarks show that LAQ outperforms existing methods across various budget levels, achieving a 1 $\\sim$ 4 point improvement on LongBench under limited cache budget. Moreover, LAQ is complementary to existing approaches and can be flexibly combined to yield further improvements.", "authors": ["Yixuan Wang", "Shiyu Ji", "Yijun Liu", "Yuzhuang Xu", "Yang Xu", "Qingfu Zhu", "Wanxiang Che"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2505.20334", "pdf_url": "https://arxiv.org/pdf/2505.20334v2", "arxiv_id": "2505.20334", "doi": "10.48550/arXiv.2505.20334", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2386}
{"id": "7dffd02f6f0672511857e4c7a32e2b59545cb7d7bbb454c4cc773057da97b81a", "sources": ["arxiv", "semantic_scholar"], "title": "MorphServe: Efficient and Workload-Aware LLM Serving via Runtime Quantized Layer Swapping and KV Cache Resizing", "abstract": "Efficiently serving large language models (LLMs) under dynamic and bursty workloads remains a key challenge for real-world deployment. Existing serving frameworks and static model compression techniques fail to adapt to workload fluctuations, leading to either service-level objective (SLO) violations under full-precision serving or persistent accuracy degradation with static quantization. We present MorphServe, a dynamic, workload-aware LLM serving framework based on morphological adaptation. MorphServe introduces two asynchronous, token-level runtime mechanisms: quantized layer swapping, which selectively replaces less impactful layers with quantized alternatives during high-load periods, and pressure-aware KV cache resizing, which dynamically adjusts KV cache capacity in response to memory pressure. These mechanisms enable state-preserving transitions with minimum runtime overhead and are fully compatible with modern scheduling and attention techniques. Extensive experiments on Vicuna and Llama family models with real-world workloads demonstrate that MorphServe reduces average SLO violations by 92.45 percent and improves the P95 TTFT latency by 2.2x-3.9x compared to full-precision serving, without compromising generation quality. These results establish MorphServe as a practical and elastic solution for LLM deployment in dynamic environments.", "authors": ["Zhaoyuan Su", "Zeyu Zhang", "Tingfeng Lan", "Zirui Wang", "Haiying Shen", "Juncheng Yang", "Yue Cheng"], "categories": ["cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2506.02006", "pdf_url": "https://arxiv.org/pdf/2506.02006v2", "arxiv_id": "2506.02006", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1193}
{"id": "dfa0f666e683e3a95a442eac358421375ef986a1bc99a0f4ba0763e6b38717c9", "sources": ["arxiv", "semantic_scholar"], "title": "Sci-LoRA: Mixture of Scientific LoRAs for Cross-Domain Lay Paraphrasing", "abstract": "Lay paraphrasing aims to make scientific information accessible to audiences without technical backgrounds. However, most existing studies focus on a single domain, such as biomedicine. With the rise of interdisciplinary research, it is increasingly necessary to comprehend knowledge spanning multiple technical fields. To address this, we propose Sci-LoRA, a model that leverages a mixture of LoRAs fine-tuned on multiple scientific domains. In particular, Sci-LoRA dynamically generates and applies weights for each LoRA, enabling it to adjust the impact of different domains based on the input text, without requiring explicit domain labels. To balance domain-specific knowledge and generalization across various domains, Sci-LoRA integrates information at both the data and model levels. This dynamic fusion enhances the adaptability and performance across various domains. Experimental results across twelve domains on five public datasets show that Sci-LoRA significantly outperforms state-of-the-art large language models and demonstrates flexible generalization and adaptability in cross-domain lay paraphrasing.", "authors": ["Ming Cheng", "Jiaying Gong", "Hoda Eldardiry"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2505.18867", "pdf_url": "https://arxiv.org/pdf/2505.18867v1", "arxiv_id": "2505.18867", "doi": "10.48550/arXiv.2505.18867", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1742}
{"id": "741208b87e2f31c7dda7391b9f93dbbed6b9eb3419d43d4e06e5820f767987d5", "sources": ["arxiv", "semantic_scholar"], "title": "ALPS: Attention Localization and Pruning Strategy for Efficient Alignment of Large Language Models", "abstract": "Aligning general-purpose large language models (LLMs) to downstream tasks often incurs significant training adjustment costs. Prior research has explored various avenues to enhance alignment efficiency, primarily through minimal-data training or data-driven activations to identify key attention heads. However, these approaches inherently introduce data dependency, which hinders generalization and reusability. To address this issue and enhance model alignment efficiency, we propose the Attention Localization and Pruning Strategy (ALPS), an efficient algorithm that localizes the most task-sensitive attention heads and prunes by restricting attention training updates to these heads, thereby reducing alignment costs. Experimental results demonstrate that our method activates only 10% of attention parameters during fine-tuning while achieving a 2% performance improvement over baselines on three tasks. Moreover, the identified task-specific heads are transferable across datasets and mitigate knowledge forgetting. Our work and findings provide a novel perspective on efficient LLM alignment. The code is available at https://github.com/VoiceBeer/ALPS.", "authors": ["Hao Chen", "Haoze Li", "Zhiqing Xiao", "Lirong Gao", "Qi Zhang", "Xiaomeng Hu", "Ningtao Wang", "Xing Fu", "Junbo Zhao"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2505.18799", "pdf_url": "https://arxiv.org/pdf/2505.18799v4", "arxiv_id": "2505.18799", "doi": "10.48550/arXiv.2505.18799", "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/VoiceBeer/ALPS", "venue": "arXiv.org", "quality_score": 0.2692}
{"id": "1ee00e2b82e286644d4d0e55d5ff656f6158f4d4a3e3dc99538acf4d27937dd3", "sources": ["arxiv", "semantic_scholar"], "title": "On Minimax Estimation of Parameters in Softmax-Contaminated Mixture of Experts", "abstract": "The softmax-contaminated mixture of experts (MoE) model is deployed when a large-scale pre-trained model, which plays the role of a fixed expert, is fine-tuned for learning downstream tasks by including a new contamination part, or prompt, functioning as a new, trainable expert. Despite its popularity and relevance, the theoretical properties of the softmax-contaminated MoE have remained unexplored in the literature. In the paper, we study the convergence rates of the maximum likelihood estimator of gating and prompt parameters in order to gain insights into the statistical properties and potential challenges of fine-tuning with a new prompt. We find that the estimability of these parameters is compromised when the prompt acquires overlapping knowledge with the pre-trained model, in the sense that we make precise by formulating a novel analytic notion of distinguishability. Under distinguishability of the pre-trained and prompt models, we derive minimax optimal estimation rates for all the gating and prompt parameters. By contrast, when the distinguishability condition is violated, these estimation rates become significantly slower due to their dependence on the prompt convergence rate to the pre-trained model. Finally, we empirically corroborate our theoretical findings through several numerical experiments.", "authors": ["Fanqi Yan", "Huy Nguyen", "Dung Le", "Pedram Akbarian", "Nhat Ho", "Alessandro Rinaldo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2505.18455", "pdf_url": "https://arxiv.org/pdf/2505.18455v2", "arxiv_id": "2505.18455", "doi": "10.48550/arXiv.2505.18455", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1742}
{"id": "3bbe5977c777283e956bedb2a8765eb4c5dbd8fdce0eac07fcf9533b172f93c4", "sources": ["arxiv", "semantic_scholar"], "title": "Titanus: Enabling KV Cache Pruning and Quantization On-the-Fly for LLM Acceleration", "abstract": "Large language models (LLMs) have gained great success in various domains. Existing systems cache Key and Value within the attention block to avoid redundant computations. However, the size of key-value cache (KV cache) is unpredictable and can even be tens of times larger than the weights in the long context length scenario. In this work, we propose Titanus, a software-hardware co-design to efficiently compress the KV cache on-the-fly. We first propose the cascade pruning-quantization (CPQ) method to reduce the KV cache movement. The hierarchical quantization extension strategy is introduced to tackle the non-independent per-channel quantization issue. To further reduce KV cache movement, we transfer only the non-zero KV cache between the accelerator and off-chip memory. Moreover, we customize a two-stage design space exploration framework for the CPQ method. A novel pipeline and parallelism dataflow is designed to reduce the first token generation time. Experiments show that Titanus achieves 159.9x (49.6x) and 34.8x (29.2x) energy efficiency (throughput) compared to Nvidia A100 GPU and FlightLLM respectively. The code for Titanus is available at https://github.com/peilin-chen/Titanus-for-LLM-acceleration.", "authors": ["Peilin Chen", "Xiaoxuan Yang"], "categories": ["cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-23", "url": "https://arxiv.org/abs/2505.17787", "pdf_url": "https://arxiv.org/pdf/2505.17787v1", "arxiv_id": "2505.17787", "doi": "10.1145/3716368.3735145", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/peilin-chen/Titanus-for-LLM-acceleration", "venue": "ACM Great Lakes Symposium on VLSI", "quality_score": 0.2674}
{"id": "2f11cc57c5c017a5fafef07dedcbf8d3274e71cc7f7253add9c36c26f5d24d81", "sources": ["arxiv", "semantic_scholar"], "title": "NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache", "abstract": "Large Language Model (LLM) inference is typically memory-intensive, especially when processing large batch sizes and long sequences, due to the large size of key-value (KV) cache. Vector Quantization (VQ) is recently adopted to alleviate this issue, but we find that the existing approach is susceptible to distribution shift due to its reliance on calibration datasets. To address this limitation, we introduce NSNQuant, a calibration-free Vector Quantization (VQ) technique designed for low-bit compression of the KV cache. By applying a three-step transformation-1) a token-wise normalization (Normalize), 2) a channel-wise centering (Shift), and 3) a second token-wise normalization (Normalize)-with Hadamard transform, NSNQuant effectively aligns the token distribution with the standard normal distribution. This alignment enables robust, calibration-free vector quantization using a single reusable codebook. Extensive experiments show that NSNQuant consistently outperforms prior methods in both 1-bit and 2-bit settings, offering strong generalization and up to 3$\\times$ throughput gain over full-precision baselines.", "authors": ["Donghyun Son", "Euntae Choi", "Sungjoo Yoo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-23", "url": "https://arxiv.org/abs/2505.18231", "pdf_url": "https://arxiv.org/pdf/2505.18231v2", "arxiv_id": "2505.18231", "doi": "10.48550/arXiv.2505.18231", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.173}
{"id": "49eda568f4355086ac1c9df76e3fed369b34ef50e87ea506eea52e69fa3a25af", "sources": ["arxiv", "semantic_scholar"], "title": "CoMoE: Contrastive Representation for Mixture-of-Experts in Parameter-Efficient Fine-tuning", "abstract": "In parameter-efficient fine-tuning, mixture-of-experts (MoE), which involves specializing functionalities into different experts and sparsely activating them appropriately, has been widely adopted as a promising approach to trade-off between model capacity and computation overhead. However, current MoE variants fall short on heterogeneous datasets, ignoring the fact that experts may learn similar knowledge, resulting in the underutilization of MoE's capacity. In this paper, we propose Contrastive Representation for MoE (CoMoE), a novel method to promote modularization and specialization in MoE, where the experts are trained along with a contrastive objective by sampling from activated and inactivated experts in top-k routing. We demonstrate that such a contrastive objective recovers the mutual-information gap between inputs and the two types of experts. Experiments on several benchmarks and in multi-task settings demonstrate that CoMoE can consistently enhance MoE's capacity and promote modularization among the experts.", "authors": ["Jinyuan Feng", "Chaopeng Wei", "Tenghai Qiu", "Tianyi Hu", "Zhiqiang Pu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-23", "url": "https://arxiv.org/abs/2505.17553", "pdf_url": "https://arxiv.org/pdf/2505.17553v2", "arxiv_id": "2505.17553", "doi": "10.48550/arXiv.2505.17553", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.173}
{"id": "ee8b680e63fe3214a0a9ebe89ee5b5e7f4d3352faf904c7f000285d3e35cdd6b", "sources": ["arxiv", "semantic_scholar"], "title": "Hybrid Mamba-Transformer Decoder for Error-Correcting Codes", "abstract": "We introduce a novel deep learning method for decoding error correction codes based on the Mamba architecture, enhanced with Transformer layers. Our approach proposes a hybrid decoder that leverages Mamba's efficient sequential modeling while maintaining the global context capabilities of Transformers. To further improve performance, we design a novel layer-wise masking strategy applied to each Mamba layer, allowing selective attention to relevant code features at different depths. Additionally, we introduce a progressive layer-wise loss, supervising the network at intermediate stages and promoting robust feature extraction throughout the decoding process. Comprehensive experiments across a range of linear codes demonstrate that our method significantly outperforms Transformer-only decoders and standard Mamba models.", "authors": ["Shy-el Cohen", "Yoni Choukroun", "Eliya Nachmani"], "categories": ["cs.IT", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-23", "url": "https://arxiv.org/abs/2505.17834", "pdf_url": "https://arxiv.org/pdf/2505.17834v1", "arxiv_id": "2505.17834", "doi": "10.48550/arXiv.2505.17834", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "1d4131b1ee17c1835038b5b8efc622d71f0ee0fba8d5c271a77f2007b943e479", "sources": ["arxiv", "semantic_scholar"], "title": "NQKV: A KV Cache Quantization Scheme Based on Normal Distribution Characteristics", "abstract": "Large Language Models (LLMs) have demonstrated remarkable proficiency across a wide range of tasks. However, LLMs often require larger batch sizes to enhance throughput or longer context lengths to meet task demands, which significantly increases the memory resource consumption of the Key-Value (KV) cache during inference, becoming a major bottleneck in LLM deployment. To address this issue, quantization is a common and straightforward approach. Currently, quantization methods for activations are limited to 8-bit, and quantization to even lower bits can lead to substantial accuracy drops. To further save space by quantizing the KV cache to even lower bits, we analyzed the element distribution of the KV cache and designed the NQKV algorithm. Since the elements within each block of the KV cache follow a normal distribution, NQKV employs per-block quantile quantization to achieve information-theoretically optimal quantization error. Without significantly compromising model output quality, NQKV enables the OPT model to perform inference with an 2x larger batch size or a 4x longer context length, and it improves throughput by 9.3x compared to when the KV cache is not used.", "authors": ["Zhihang Cai", "Xingjun Zhang", "Zhendong Tan", "Zheng Wei"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.16210", "pdf_url": "https://arxiv.org/pdf/2505.16210v1", "arxiv_id": "2505.16210", "doi": "10.48550/arXiv.2505.16210", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1719}
{"id": "242d97dc611d2cfc709451336fb87768b1ab7cdc01f21906cdcf0d5900f25ad4", "sources": ["arxiv", "semantic_scholar"], "title": "MARché: Fast Masked Autoregressive Image Generation with Cache-Aware Attention", "abstract": "Masked autoregressive (MAR) models unify the strengths of masked and autoregressive generation by predicting tokens in a fixed order using bidirectional attention for image generation. While effective, MAR models suffer from significant computational overhead, as they recompute attention and feed-forward representations for all tokens at every decoding step, despite most tokens remaining semantically stable across steps. We propose a training-free generation framework MARché to address this inefficiency through two key components: cache-aware attention and selective KV refresh. Cache-aware attention partitions tokens into active and cached sets, enabling separate computation paths that allow efficient reuse of previously computed key/value projections without compromising full-context modeling. But a cached token cannot be used indefinitely without recomputation due to the changing contextual information over multiple steps. MARché recognizes this challenge and applies a technique called selective KV refresh. Selective KV refresh identifies contextually relevant tokens based on attention scores from newly generated tokens and updates only those tokens that require recomputation, while preserving image generation quality. MARché significantly reduces redundant computation in MAR without modifying the underlying architecture. Empirically, MARché achieves up to 1.7x speedup with negligible impact on image quality, offering a scalable and broadly applicable solution for efficient masked transformer generation.", "authors": ["Chaoyi Jiang", "Sungwoo Kim", "Lei Gao", "Hossein Entezari Zarch", "Won Woo Ro", "Murali Annavaram"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2506.12035", "pdf_url": "https://arxiv.org/pdf/2506.12035v1", "arxiv_id": "2506.12035", "doi": "10.48550/arXiv.2506.12035", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1719}
{"id": "a0986d9e8ca4ed656b616016db19bda4578f4cf203fe8e1a33092f4870931b45", "sources": ["arxiv", "semantic_scholar"], "title": "Bottlenecked Transformers: Periodic KV Cache Consolidation for Generalised Reasoning", "abstract": "Transformer LLMs have been shown to exhibit strong reasoning ability that scales with inference-time compute, most prominently through token-space \"thinking\" chains of thought. A growing line of work pushes extra computation into the model's latent space, which we term Auxiliary Latent-Space Computation (ALSC). Existing ALSC methods largely fall into three buckets: (i) token-mediated latent rollouts, (ii) residual/activation steering, and (iii) memory (KV) compression. An underexplored alternative is memory consolidation/reconsolidation, two processes in the brain that are responsible for stabilising newly formed memory traces, and, upon recall, transiently rendering established traces plastic such they can integrate new contextual information before restabilising. In Transformer LLMs, this can be seen as analogous to performing in-place rewrites of new KV segments, and rewrites of recalled past segments. In this work, we give a theoretical justification as to why memory (re)consolidation via KV cache rewrites is beneficial for improved reasoning. We do this through the lens of Information Bottleneck (IB) theory, which posits that model generalisation emerges from an optimal balance between input information compression and retention of predictive information in latent representations. We then introduce the Bottlenecked Transformer, which augments a backbone LLM with a Cache Processor, an auxiliary Transformer that performs periodic, non-causal, in-place KV rewrites at newline-delimited reasoning step boundaries. The Processor consolidates recently written KV entries and reconsolidates a small, top-k attention-selected set of prior entries. We evaluate our Bottlenecked Transformer architecture on math reasoning benchmarks. Our model sees consistent performance gains over vanilla Transformers and pause-token augmented baselines, with gains of up to +6.6pp for selected tasks/backbones.", "authors": ["Adnan Oomerjee", "Zafeirios Fountas", "Haitham Bou-Ammar", "Jun Wang"], "categories": ["cs.LG", "cs.AI", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.16950", "pdf_url": "https://arxiv.org/pdf/2505.16950v4", "arxiv_id": "2505.16950", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1193}
{"id": "e6fa98b0b56c8c245dc5939e8f3d9196060a0182b42655647c45cda05ff28097", "sources": ["arxiv", "semantic_scholar"], "title": "RAP: Runtime Adaptive Pruning for LLM Inference", "abstract": "Large language models (LLMs) excel at language understanding and generation, but their enormous computational and memory requirements hinder deployment. Compression offers a potential solution to mitigate these constraints. However, most existing methods rely on fixed heuristics and thus fail to adapt to runtime memory variations or heterogeneous KV-cache demands arising from diverse user requests. To address these limitations, we propose RAP, an elastic pruning framework driven by reinforcement learning (RL) that dynamically adjusts compression strategies in a runtime-aware manner. Specifically, RAP dynamically tracks the evolving ratio between model parameters and KV-cache across practical execution. Recognizing that FFNs house most parameters, whereas parameter -light attention layers dominate KV-cache formation, the RL agent retains only those components that maximize utility within the current memory budget, conditioned on instantaneous workload and device state. Extensive experiments results demonstrate that RAP outperforms state-of-the-art baselines, marking the first time to jointly consider model weights and KV-cache on the fly.", "authors": ["Huanrong Liu", "Chunlin Tian", "Xuyang Wei", "Qingbiao Li", "Li Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.17138", "pdf_url": "https://arxiv.org/pdf/2505.17138v5", "arxiv_id": "2505.17138", "doi": "10.48550/arXiv.2505.17138", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "0f593f4a9415f6a308e3703d2525598ab4b0d03ae7f0283a223c09508571b859", "sources": ["arxiv", "semantic_scholar"], "title": "dKV-Cache: The Cache for Diffusion Language Models", "abstract": "Diffusion Language Models (DLMs) have been seen as a promising competitor for autoregressive language models. However, diffusion language models have long been constrained by slow inference. A core challenge is that their non-autoregressive architecture and bidirectional attention preclude the key-value cache that accelerates decoding. We address this bottleneck by proposing a KV-cache-like mechanism, delayed KV-Cache, for the denoising process of DLMs. Our approach is motivated by the observation that different tokens have distinct representation dynamics throughout the diffusion process. Accordingly, we propose a delayed and conditioned caching strategy for key and value states. We design two complementary variants to cache key and value step-by-step: (1) dKV-Cache-Decode, which provides almost lossless acceleration, and even improves performance on long sequences, suggesting that existing DLMs may under-utilise contextual information during inference. (2) dKV-Cache-Greedy, which has aggressive caching with reduced lifespan, achieving higher speed-ups with quadratic time complexity at the cost of some performance degradation. dKV-Cache, in final, achieves from 2-10x speedup in inference, largely narrowing the gap between ARs and DLMs. We evaluate our dKV-Cache on several benchmarks, delivering acceleration across general language understanding, mathematical, and code-generation benchmarks. Experiments demonstrate that cache can also be used in DLMs, even in a training-free manner from current DLMs.", "authors": ["Xinyin Ma", "Runpeng Yu", "Gongfan Fang", "Xinchao Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-21", "url": "https://arxiv.org/abs/2505.15781", "pdf_url": "https://arxiv.org/pdf/2505.15781v1", "arxiv_id": "2505.15781", "doi": "10.48550/arXiv.2505.15781", "citation_count": 114, "influential_citation_count": 13, "has_code": true, "code_url": "https://github.com/horseee/dKV-Cache", "venue": "arXiv.org", "quality_score": 0.5731}
{"id": "ca613e19826381a5f057e02de83fd3f491b2579042d8f4ca5ceac594d6c1a34b", "sources": ["arxiv", "semantic_scholar"], "title": "LiveVLM: Efficient Online Video Understanding via Streaming-Oriented KV Cache and Retrieval", "abstract": "Recent developments in Video Large Language Models (Video LLMs) have enabled models to process hour-long videos and exhibit exceptional performance. Nonetheless, the Key-Value (KV) cache expands linearly over time, leading to substantial memory overhead and response delay--critical challenges in various real-world online applications, such as Deepseek services, autonomous driving and robotics. To mitigate these issues, we propose $\\textbf{LiveVLM}$, a training-free and query-agnostic framework specifically designed for online video understanding and real-time interaction. LiveVLM employs a Vision Sink Bucketing (VSB) mechanism to process video streams in real time, retain long-term video details and eliminate redundant KVs. This mechanism utilizes vision-to-vision attention scores as the metric and seeks to maximize the coverage of contextual information during compression. Noting that KV cache compressed in a query-agnostic manner inevitably retains irrelevant information for specific queries, LiveVLM incorporates a Position-agnostic KV Retrieval (PaR) mechanism to reduce interference from redundant context. The keypoint of PaR lies in decoupling positional embeddings to enhance the similarity between key tensors, thereby supporting efficient retrieval at the granularity of pages. Extensive experiments demonstrate that LiveVLM enables the foundation LLaVA-OneVision model to achieve state-of-the-art accuracy among both training-free query-agnostic methods and training-based online models.", "authors": ["Zhenyu Ning", "Guangda Liu", "Qihao Jin", "Chengwei Li", "Wenchao Ding", "Minyi Guo", "Jieru Zhao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-21", "url": "https://arxiv.org/abs/2505.15269", "pdf_url": "https://arxiv.org/pdf/2505.15269v2", "arxiv_id": "2505.15269", "doi": "10.1145/3770743.3804012", "citation_count": 37, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3949}
{"id": "3b4f9c0f6faef23099891e07a0634baf7a676d48e2535afec63a54f35d08a8a4", "sources": ["arxiv", "semantic_scholar"], "title": "MoTE: Mixture of Task-specific Experts for Pre-Trained ModelBased Class-incremental Learning", "abstract": "Class-incremental learning (CIL) requires deep learning models to continuously acquire new knowledge from streaming data while preserving previously learned information. Recently, CIL based on pre-trained models (PTMs) has achieved remarkable success. However, prompt-based approaches suffer from prompt overwriting, while adapter-based methods face challenges such as dimensional misalignment between tasks. While the idea of expert fusion in Mixture of Experts (MoE) can help address dimensional inconsistency, both expert and routing parameters are prone to being overwritten in dynamic environments, making MoE challenging to apply directly in CIL. To tackle these issues, we propose a mixture of task-specific experts (MoTE) framework that effectively mitigates the miscalibration caused by inconsistent output dimensions across tasks. Inspired by the weighted feature fusion and sparse activation mechanisms in MoE, we introduce task-aware expert filtering and reliable expert joint inference during the inference phase, mimicking the behavior of routing layers without inducing catastrophic forgetting. Extensive experiments demonstrate the superiority of our method without requiring an exemplar set. Furthermore, the number of tasks in MoTE scales linearly with the number of adapters. Building on this, we further explore the trade-off between adapter expansion and model performance and propose the Adapter-Limited MoTE. The code is available at https://github.com/Franklilinjie/MoTE.", "authors": ["Linjie Li", "Zhenyu Wu", "Yang Ji"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-21", "url": "https://arxiv.org/abs/2506.11038", "pdf_url": "https://arxiv.org/pdf/2506.11038v1", "arxiv_id": "2506.11038", "doi": "10.1016/j.knosys.2025.113795", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Franklilinjie/MoTE", "venue": "Knowledge-Based Systems", "quality_score": 0.2639}
{"id": "cc6e8d4212cd1dd0c60401e14027458d29acf384350ea7b2a9370bbe9701495f", "sources": ["arxiv", "semantic_scholar"], "title": "Not All Models Suit Expert Offloading: On Local Routing Consistency of Mixture-of-Expert Models", "abstract": "Mixture-of-Experts (MoE) enables efficient scaling of large language models (LLMs) with sparsely activated experts during inference. To effectively deploy large MoE models on memory-constrained devices, many systems introduce *expert offloading* that caches a subset of experts in fast memory, leaving others on slow memory to run on CPU or load on demand. While some research has exploited the locality of expert activations, where consecutive tokens activate similar experts, the degree of this **local routing consistency** varies across models and remains understudied. In this paper, we propose two metrics to measure local routing consistency of MoE models: (1) **Segment Routing Best Performance (SRP)**, which evaluates how well a fixed group of experts can cover the needs of a segment of tokens, and (2) **Segment Cache Best Hit Rate (SCH)**, which measures the hit rate of an expert cache utilizing a length of future information under a cache limit. We analyze 20 MoE LLMs with diverse sizes and architectures and use toy models to verify key factors related to local routing consistency. We find a strong trade-off between local routing consistency and *local* load balance, while showing that *global* load balance can coexist with local routing consistency. Meanwhile, settings like shared experts that decrease expert combination space can lead to low local routing consistency. We further reveal that domain-specialized experts contribute more to routing consistency than vocabulary-specialized ones, and that most models balance between cache effectiveness and efficiency with cache sizes approximately twice the active experts. These findings pave the way for memory-efficient MoE design and deployment without compromising inference speed. We publish the code for replicating experiments at https://github.com/ljcleo/moe-lrc .", "authors": ["Jingcong Liang", "Siyuan Wang", "Miren Tian", "Yitong Li", "Duyu Tang", "Zhongyu Wei"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-21", "url": "https://arxiv.org/abs/2505.16056", "pdf_url": "https://arxiv.org/pdf/2505.16056v4", "arxiv_id": "2505.16056", "doi": "10.48550/arXiv.2505.16056", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ljcleo/moe-lrc", "venue": "arXiv.org", "quality_score": 0.2639}
{"id": "3bc1fc60303b3f83ce8c723a1d25e083cda1f8c81fa67149bfcfdc074b89ddfa", "sources": ["arxiv", "semantic_scholar"], "title": "MoRE-Brain: Routed Mixture of Experts for Interpretable and Generalizable Cross-Subject fMRI Visual Decoding", "abstract": "Decoding visual experiences from fMRI offers a powerful avenue to understand human perception and develop advanced brain-computer interfaces. However, current progress often prioritizes maximizing reconstruction fidelity while overlooking interpretability, an essential aspect for deriving neuroscientific insight. To address this gap, we propose MoRE-Brain, a neuro-inspired framework designed for high-fidelity, adaptable, and interpretable visual reconstruction. MoRE-Brain uniquely employs a hierarchical Mixture-of-Experts architecture where distinct experts process fMRI signals from functionally related voxel groups, mimicking specialized brain networks. The experts are first trained to encode fMRI into the frozen CLIP space. A finetuned diffusion model then synthesizes images, guided by expert outputs through a novel dual-stage routing mechanism that dynamically weighs expert contributions across the diffusion process. MoRE-Brain offers three main advancements: First, it introduces a novel Mixture-of-Experts architecture grounded in brain network principles for neuro-decoding. Second, it achieves efficient cross-subject generalization by sharing core expert networks while adapting only subject-specific routers. Third, it provides enhanced mechanistic insight, as the explicit routing reveals precisely how different modeled brain regions shape the semantic and spatial attributes of the reconstructed image. Extensive experiments validate MoRE-Brain's high reconstruction fidelity, with bottleneck analyses further demonstrating its effective utilization of fMRI signals, distinguishing genuine neural decoding from over-reliance on generative priors. Consequently, MoRE-Brain marks a substantial advance towards more generalizable and interpretable fMRI-based visual decoding. Code will be publicly available soon: https://github.com/yuxiangwei0808/MoRE-Brain.", "authors": ["Yuxiang Wei", "Yanteng Zhang", "Xi Xiao", "Tianyang Wang", "Xiao Wang", "Vince D. Calhoun"], "categories": ["cs.LG", "cs.AI", "cs.CV", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-21", "url": "https://arxiv.org/abs/2505.15946", "pdf_url": "https://arxiv.org/pdf/2505.15946v3", "arxiv_id": "2505.15946", "doi": "10.48550/arXiv.2505.15946", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/yuxiangwei0808/MoRE-Brain", "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "6cec539b86f6c9ab0b82cc80983709c6dade6b8f675f56a635e9e75e121df43f", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling Law for Quantization-Aware Training", "abstract": "Large language models (LLMs) demand substantial computational and memory resources, creating deployment challenges. Quantization-aware training (QAT) addresses these challenges by reducing model precision while maintaining performance. However, the scaling behavior of QAT, especially at 4-bit precision (W4A4), is not well understood. Existing QAT scaling laws often ignore key factors such as the number of training tokens and quantization granularity, which limits their applicability. This paper proposes a unified scaling law for QAT that models quantization error as a function of model size, training data volume, and quantization group size. Through 268 QAT experiments, we show that quantization error decreases as model size increases, but rises with more training tokens and coarser quantization granularity. To identify the sources of W4A4 quantization error, we decompose it into weight and activation components. Both components follow the overall trend of W4A4 quantization error, but with different sensitivities. Specifically, weight quantization error increases more rapidly with more training tokens. Further analysis shows that the activation quantization error in the FC2 layer, caused by outliers, is the primary bottleneck of W4A4 QAT quantization error. By applying mixed-precision quantization to address this bottleneck, we demonstrate that weight and activation quantization errors can converge to similar levels. Additionally, with more training data, weight quantization error eventually exceeds activation quantization error, suggesting that reducing weight quantization error is also important in such scenarios. These findings offer key insights for improving QAT research and development.", "authors": ["Mengzhao Chen", "Chaoyi Zhang", "Jing Liu", "Yutao Zeng", "Zeyue Xue", "Zhiheng Liu", "Yunshui Li", "Jin Ma", "Jie Huang", "Xun Zhou", "Ping Luo"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14302", "pdf_url": "https://arxiv.org/pdf/2505.14302v1", "arxiv_id": "2505.14302", "doi": "10.48550/arXiv.2505.14302", "citation_count": 15, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "7ae8c1220cb2cc0737d71fe5aa106e59004bb205e05a20c5f60ac6796f2d825f", "sources": ["arxiv", "semantic_scholar"], "title": "Multimodal Mixture of Low-Rank Experts for Sentiment Analysis and Emotion Recognition", "abstract": "Multi-task learning (MTL) enables the efficient transfer of extra knowledge acquired from other tasks. The high correlation between multimodal sentiment analysis (MSA) and multimodal emotion recognition (MER) supports their joint training. However, existing methods primarily employ hard parameter sharing, ignoring parameter conflicts caused by complex task correlations. In this paper, we present a novel MTL method for MSA and MER, termed Multimodal Mixture of Low-Rank Experts (MMoLRE). MMoLRE utilizes shared and task-specific experts to distinctly model common and unique task characteristics, thereby avoiding parameter conflicts. Additionally, inspired by low-rank structures in the Mixture of Experts (MoE) framework, we design low-rank expert networks to reduce parameter and computational overhead as the number of experts increases. Extensive experiments on the CMU-MOSI and CMU-MOSEI benchmarks demonstrate that MMoLRE achieves state-of-the-art performance on the MSA task and competitive results on the MER task.", "authors": ["Shuo Zhang", "Jinsong Zhang", "Zhejun Zhang", "Lei Li"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14143", "pdf_url": "https://arxiv.org/pdf/2505.14143v1", "arxiv_id": "2505.14143", "doi": "10.1109/ICME59968.2025.11210197", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Multimedia and Expo", "quality_score": 0.1696}
{"id": "0a8969d3b38cd20a9677ee5e93ef495afe0069fb21fedeed89e89d18bd9ff7e5", "sources": ["arxiv", "semantic_scholar"], "title": "Solving MNIST with a globally trained Mixture of Quantum Experts", "abstract": "We propose a new quantum neural network for image classification, which is able to classify the parity of the MNIST dataset with full resolution with a test accuracy of up to 97.5% without any classical pre-processing or post-processing. Our architecture is based on a mixture of experts whose model function is the sum of the model functions of each expert. We encode the input with amplitude encoding, which allows us to encode full-resolution MNIST images with 10 qubits and to implement a convolution on the whole image with just a single one-qubit gate. Our training algorithm is based on training all the experts together, which significantly improves trainability with respect to training each expert independently. In fact, in the limit of infinitely many experts, our training algorithm can perfectly fit the training data. Our results demonstrate the potential of our quantum neural network to achieve high-accuracy image classification with minimal quantum resources, paving the way for more scalable and efficient quantum machine learning models.", "authors": ["Paolo Alessandro Xavier Tognini", "Leonardo Banchi", "Giacomo De Palma"], "categories": ["quant-ph"], "fields_of_study": ["Physics"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14789", "pdf_url": "https://arxiv.org/pdf/2505.14789v1", "arxiv_id": "2505.14789", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "e3b7e40f6109f648aa63a7c94682d56acbf9f8af260630ed27263a7aabccd9db", "sources": ["arxiv", "semantic_scholar"], "title": "MoMoE: Mixture of Moderation Experts Framework for AI-Assisted Online Governance", "abstract": "Large language models (LLMs) have shown great potential in flagging harmful content in online communities. Yet, existing approaches for moderation require a separate model for every community and are opaque in their decision-making, limiting real-world adoption. We introduce Mixture of Moderation Experts (MoMoE), a modular, cross-community framework that adds post-hoc explanations to scalable content moderation. MoMoE orchestrates four operators -- Allocate, Predict, Aggregate, Explain -- and is instantiated as seven community-specialized experts (MoMoE-Community) and five norm-violation experts (MoMoE-NormVio). On 30 unseen subreddits, the best variants obtain Micro-F1 scores of 0.72 and 0.67, respectively, matching or surpassing strong fine-tuned baselines while consistently producing concise and reliable explanations. Although community-specialized experts deliver the highest peak accuracy, norm-violation experts provide steadier performance across domains. These findings show that MoMoE yields scalable, transparent moderation without needing per-community fine-tuning. More broadly, they suggest that lightweight, explainable expert ensembles can guide future NLP and HCI research on trustworthy human-AI governance of online communities.", "authors": ["Agam Goyal", "Xianyang Zhan", "Yilun Chen", "Koustuv Saha", "Eshwar Chandrasekharan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14483", "pdf_url": "https://arxiv.org/pdf/2505.14483v2", "arxiv_id": "2505.14483", "doi": "10.48550/arXiv.2505.14483", "citation_count": 15, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.301}
{"id": "3ef1057d75d7c800745cf149b18e827a0e48a82ede9d8288e0694661107dc888", "sources": ["arxiv", "semantic_scholar"], "title": "Local Mixtures of Experts: Essentially Free Test-Time Training via Model Merging", "abstract": "Mixture of expert (MoE) models are a promising approach to increasing model capacity without increasing inference cost, and are core components of many state-of-the-art language models. However, current MoE models typically use only few experts due to prohibitive training and inference cost. We propose Test-Time Model Merging (TTMM) which scales the MoE paradigm to an order of magnitude more experts and uses model merging to avoid almost any test-time overhead. We show that TTMM is an approximation of test-time training (TTT), which fine-tunes an expert model for each prediction task, i.e., prompt. TTT has recently been shown to significantly improve language models, but is computationally expensive. We find that performance of TTMM improves with more experts and approaches the performance of TTT. Moreover, we find that with a 1B parameter base model, TTMM is more than 100x faster than TTT at test-time by amortizing the cost of TTT at train-time. Thus, TTMM offers a promising cost-effective approach to scale test-time training.", "authors": ["Ryo Bertolissi", "Jonas Hübotter", "Ido Hakimi", "Andreas Krause"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14136", "pdf_url": "https://arxiv.org/pdf/2505.14136v2", "arxiv_id": "2505.14136", "doi": "10.48550/arXiv.2505.14136", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "ee49eb877bd3298b6e2e56d69c876176531981920f3ae1383b05406fd3fefa69", "sources": ["arxiv", "semantic_scholar"], "title": "Generalizable Multispectral Land Cover Classification via Frequency-Aware Mixture of Low-Rank Token Experts", "abstract": "We introduce Land-MoE, a novel approach for multispectral land cover classification (MLCC). Spectral shift, which emerges from disparities in sensors and geospatial conditions, poses a significant challenge in this domain. Existing methods predominantly rely on domain adaptation and generalization strategies, often utilizing small-scale models that exhibit limited performance. In contrast, Land-MoE addresses these issues by hierarchically inserting a Frequency-aware Mixture of Low-rank Token Experts, to fine-tune Vision Foundation Models (VFMs) in a parameter-efficient manner. Specifically, Land-MoE comprises two key modules: the mixture of low-rank token experts (MoLTE) and frequency-aware filters (FAF). MoLTE leverages rank-differentiated tokens to generate diverse feature adjustments for individual instances within multispectral images. By dynamically combining learnable low-rank token experts of varying ranks, it enhances the robustness against spectral shifts. Meanwhile, FAF conducts frequency-domain modulation on the refined features. This process enables the model to effectively capture frequency band information that is strongly correlated with semantic essence, while simultaneously suppressing frequency noise irrelevant to the task. Comprehensive experiments on MLCC tasks involving cross-sensor and cross-geospatial setups demonstrate that Land-MoE outperforms existing methods by a large margin. Additionally, the proposed approach has also achieved state-of-the-art performance in domain generalization semantic segmentation tasks of RGB remote sensing images.", "authors": ["Xi Chen", "Shen Yan", "Juelin Zhu", "Chen Chen", "Yu Liu", "Maojun Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14088", "pdf_url": "https://arxiv.org/pdf/2505.14088v1", "arxiv_id": "2505.14088", "doi": "10.48550/arXiv.2505.14088", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1696}
{"id": "510c4fb70aa8710235dcc5345fdc34e7ecf94bb9bf7ffc0978c207681793504a", "sources": ["arxiv", "semantic_scholar"], "title": "FreeKV: Boosting KV Cache Retrieval for Efficient LLM Inference", "abstract": "Large language models (LLMs) are widely deployed with rapidly expanding context windows to support increasingly demanding applications. However, long contexts pose significant deployment challenges, primarily due to the KV cache whose size grows proportionally with context length. While KV cache compression methods have been proposed to address this issue, KV dropping methods incur considerable accuracy loss, and KV retrieval methods suffer from significant efficiency bottlenecks. We propose FreeKV, a training-free algorithm-system co-optimization framework to enhance KV retrieval efficiency while preserving accuracy. On the algorithm side, FreeKV introduces speculative retrieval to shift the KV selection and recall processes out of the critical path, combined with fine-grained correction to ensure accuracy. On the system side, FreeKV employs hybrid KV layouts across CPU and GPU memory to eliminate fragmented data transfers, and leverages double-buffered streamed recall to further improve efficiency, enabling effective overlap with computation, full latency hiding, and practical speedups from speculative recall. Experiments demonstrate that FreeKV achieves near-lossless accuracy across various scenarios and models, delivering up to a 13$\\times$ speedup compared to SOTA KV retrieval methods. Code is available at https://github.com/sjtu-zhao-lab/FreeKV.", "authors": ["Guangda Liu", "Chengwei Li", "Zhenyu Ning", "Jing Lin", "Yiwu Yao", "Danning Ke", "Minyi Guo", "Jieru Zhao"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-19", "url": "https://arxiv.org/abs/2505.13109", "pdf_url": "https://arxiv.org/pdf/2505.13109v5", "arxiv_id": "2505.13109", "doi": "10.48550/arXiv.2505.13109", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/sjtu-zhao-lab/FreeKV", "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "2c4f47615f5d58d87584ed95ea631f4fcee7892f2a07b88222acc430cc04e600", "sources": ["arxiv", "semantic_scholar"], "title": "Model Selection for Gaussian-gated Gaussian Mixture of Experts Using Dendrograms of Mixing Measures", "abstract": "Mixture of Experts (MoE) models constitute a widely utilized class of ensemble learning approaches in statistics and machine learning, known for their flexibility and computational efficiency. They have become integral components in numerous state-of-the-art deep neural network architectures, particularly for analyzing heterogeneous data across diverse domains. Despite their practical success, the theoretical understanding of model selection, especially concerning the optimal number of mixture components or experts, remains limited and poses significant challenges. These challenges primarily stem from the inclusion of covariates in both the Gaussian gating functions and expert networks, which introduces intrinsic interactions governed by partial differential equations with respect to their parameters. In this paper, we revisit the concept of dendrograms of mixing measures and introduce a novel extension to Gaussian-gated Gaussian MoE models that enables consistent estimation of the true number of mixture components and achieves the pointwise optimal convergence rate for parameter estimation in overfitted scenarios. Notably, this approach circumvents the need to train and compare a range of models with varying numbers of components, thereby alleviating the computational burden, particularly in high-dimensional or deep neural network settings. Experimental results on synthetic data demonstrate the effectiveness of the proposed method in accurately recovering the number of experts. It outperforms common criteria such as the Akaike information criterion, the Bayesian information criterion, and the integrated completed likelihood, while achieving optimal convergence rates for parameter estimation and accurately approximating the regression function.", "authors": ["Tuan Thai", "TrungTin Nguyen", "Dat Do", "Nhat Ho", "Christopher Drovandi"], "categories": ["stat.ML", "cs.LG", "math.ST", "stat.CO", "stat.ME"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2025-05-19", "url": "https://arxiv.org/abs/2505.13052", "pdf_url": "https://arxiv.org/pdf/2505.13052v2", "arxiv_id": "2505.13052", "doi": "10.48550/arXiv.2505.13052", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "04bdb56d675bab375f8f2b5714a515f9ca003c471906f445574e0749bfd4d433", "sources": ["arxiv", "semantic_scholar"], "title": "CompeteSMoE -- Statistically Guaranteed Mixture of Experts Training via Competition", "abstract": "Sparse mixture of experts (SMoE) offers an appealing solution to scale up the model complexity beyond the mean of increasing the network's depth or width. However, we argue that effective SMoE training remains challenging because of the suboptimal routing process where experts that perform computation do not directly contribute to the routing process. In this work, we propose competition, a novel mechanism to route tokens to experts with the highest neural response. Theoretically, we show that the competition mechanism enjoys a better sample efficiency than the traditional softmax routing. Furthermore, we develop CompeteSMoE, a simple yet effective algorithm to train large language models by deploying a router to learn the competition policy, thus enjoying strong performances at a low training overhead. Our extensive empirical evaluations on both the visual instruction tuning and language pre-training tasks demonstrate the efficacy, robustness, and scalability of CompeteSMoE compared to state-of-the-art SMoE strategies. We have made the implementation available at: https://github.com/Fsoft-AIC/CompeteSMoE. This work is an improved version of the previous study at arXiv:2402.02526", "authors": ["Nam V. Nguyen", "Huy Nguyen", "Quang Pham", "Van Nguyen", "Savitha Ramasamy", "Nhat Ho"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-19", "url": "https://arxiv.org/abs/2505.13380", "pdf_url": "https://arxiv.org/pdf/2505.13380v1", "arxiv_id": "2505.13380", "doi": "10.48550/arXiv.2505.13380", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Fsoft-AIC/CompeteSMoE", "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "ec5703e53f8a16e57c6a94d5a4658938e740b8cabb5bc8261ad8d10472eba1b9", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient training for large-scale optical neural network using an evolutionary strategy and attention pruning", "abstract": "MZI-based block optical neural networks (BONNs), which can achieve large-scale network models, have increasingly drawn attentions. However, the robustness of the current training algorithm is not high enough. Moreover, large-scale BONNs usually contain numerous trainable parameters, resulting in expensive computation and power consumption. In this article, by pruning matrix blocks and directly optimizing the individuals in population, we propose an on-chip covariance matrix adaptation evolution strategy and attention-based pruning (CAP) algorithm for large-scale BONNs. The calculated results demonstrate that the CAP algorithm can prune 60% and 80% of the parameters for MNIST and Fashion-MNIST datasets, respectively, while only degrades the performance by 3.289% and 4.693%. Considering the influence of dynamic noise in phase shifters, our proposed CAP algorithm (performance degradation of 22.327% for MNIST dataset and 24.019% for Fashion-MNIST dataset utilizing a poor fabricated chip and electrical control with a standard deviation of 0.5) exhibits strongest robustness compared with both our previously reported block adjoint training algorithm (43.963% and 41.074%) and the covariance matrix adaptation evolution strategy (25.757% and 32.871%), respectively. Moreover, when 60% of the parameters are pruned, the CAP algorithm realizes 88.5% accuracy in experiment for the simplified MNIST dataset, which is similar to the simulation result without noise (92.1%). Additionally, we simulationally and experimentally demonstrate that using MZIs with only internal phase shifters to construct BONNs is an efficient way to reduce both the system area and the required trainable parameters. Notably, our proposed CAP algorithm show excellent potential for larger-scale network models and more complex tasks.", "authors": ["Zhiwei Yang", "Zeyang Fan", "Yihang Lai", "Qi Chen", "Tian Zhang", "Jian Dai", "Kun Xu"], "categories": ["cs.LG", "physics.optics"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2025-05-19", "url": "https://arxiv.org/abs/2505.12906", "pdf_url": "https://arxiv.org/pdf/2505.12906v1", "arxiv_id": "2505.12906", "doi": "10.48550/arXiv.2505.12906", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1684}
{"id": "6a521c34fa96aa1b79ad2c515f825ffe68e186f3b997e2917c1a4c695d50600b", "sources": ["arxiv", "semantic_scholar"], "title": "KVmix: Gradient-Based Layer Importance-Aware Mixed-Precision Quantization for KV Cache", "abstract": "The high memory demands of the Key-Value (KV) Cache during the inference of Large Language Models (LLMs) severely restrict their deployment in resource-constrained platforms. Quantization can effectively alleviate the memory pressure caused by KV Cache. However, existing methods either rely on static one-size-fits-all precision allocation or fail to dynamically prioritize critical KV in long-context tasks, forcing memory-accuracy-throughput tradeoffs. In this work, we propose a novel mixed-precision quantization method for KV Cache named KVmix. KVmix leverages gradient-based importance analysis to evaluate how individual Key and Value projection matrices affect the model loss, enabling layer-specific bit-width allocation for mix-precision quantization. It dynamically prioritizes higher precision for important layers while aggressively quantizing less influential ones, achieving a tunable balance between accuracy and efficiency. KVmix also introduces a dynamic long-context optimization strategy that adaptively keeps full-precision KV pairs for recent pivotal tokens and compresses older ones, achieving high-quality sequence generation with low memory usage. Additionally, KVmix provides efficient low-bit quantization and CUDA kernels to optimize computational overhead. On LLMs such as Llama and Mistral, KVmix achieves near-lossless inference performance with extremely low quantization configuration (Key 2.19bit Value 2.38bit), while delivering a remarkable 4.9x memory compression and a 5.3x speedup in inference throughput.", "authors": ["Fei Li", "Song Liu", "Weiguo Wu", "Shiqiang Nie", "Jinyu Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-18", "url": "https://arxiv.org/abs/2506.08018", "pdf_url": "https://arxiv.org/pdf/2506.08018v3", "arxiv_id": "2506.08018", "doi": "10.48550/arXiv.2506.08018", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.1673}
{"id": "fffa573dce312eb8d326775c6e5c71c235e4ff346af7f3e215d63161a8f280b6", "sources": ["arxiv", "semantic_scholar"], "title": "Accurate KV Cache Quantization with Outlier Tokens Tracing", "abstract": "The impressive capabilities of Large Language Models (LLMs) come at the cost of substantial computational resources during deployment. While KV Cache can significantly reduce recomputation during inference, it also introduces additional memory overhead. KV Cache quantization presents a promising solution, striking a good balance between memory usage and accuracy. Previous research has shown that the Keys are distributed by channel, while the Values are distributed by token. Consequently, the common practice is to apply channel-wise quantization to the Keys and token-wise quantization to the Values. However, our further investigation reveals that a small subset of unusual tokens exhibit unique characteristics that deviate from this pattern, which can substantially impact quantization accuracy. To address this, we develop a simple yet effective method to identify these tokens accurately during the decoding process and exclude them from quantization as outlier tokens, significantly improving overall accuracy. Extensive experiments show that our method achieves significant accuracy improvements under 2-bit quantization and can deliver a 6.4 times reduction in memory usage and a 2.3 times increase in throughput.", "authors": ["Yi Su", "Yuechi Zhou", "Quantong Qiu", "Juntao Li", "Qingrong Xia", "Ping Li", "Xinyu Duan", "Zhefeng Wang", "Min Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-16", "url": "https://arxiv.org/abs/2505.10938", "pdf_url": "https://arxiv.org/pdf/2505.10938v1", "arxiv_id": "2505.10938", "doi": "10.48550/arXiv.2505.10938", "citation_count": 16, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3076}
{"id": "c9c7eba36bfc87366ff66763ae3d0fee774090bbac0a905cbcc67c1652e18c46", "sources": ["arxiv", "semantic_scholar"], "title": "SubGCache: Accelerating Graph-based RAG with Subgraph-level KV Cache", "abstract": "Graph-based retrieval-augmented generation (RAG) enables large language models (LLMs) to incorporate structured knowledge via graph retrieval as contextual input, enhancing more accurate and context-aware reasoning. We observe that for different queries, it could retrieve similar subgraphs as prompts, and thus we propose SubGCache, which aims to reduce inference latency by reusing computation across queries with similar structural prompts (i.e., subgraphs). Specifically, SubGCache clusters queries based on subgraph embeddings, constructs a representative subgraph for each cluster, and pre-computes the key-value (KV) cache of the representative subgraph. For each query with its retrieved subgraph within a cluster, it reuses the pre-computed KV cache of the representative subgraph of the cluster without computing the KV tensors again for saving computation. Experiments on two new datasets across multiple LLM backbones and graph-based RAG frameworks demonstrate that SubGCache consistently reduces inference latency with comparable and even improved generation quality, achieving up to 6.68$\\times$ reduction in time-to-first-token (TTFT).", "authors": ["Qiuyu Zhu", "Liang Zhang", "Qianxiong Xu", "Cheng Long", "Jie Zhang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-16", "url": "https://arxiv.org/abs/2505.10951", "pdf_url": "https://arxiv.org/pdf/2505.10951v2", "arxiv_id": "2505.10951", "doi": "10.48550/arXiv.2505.10951", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.165}
{"id": "481906f3c7f5be018872996a1626f0dc3202145618ae021f1a184a1ce3b3506f", "sources": ["arxiv", "semantic_scholar"], "title": "PT-MoE: An Efficient Finetuning Framework for Integrating Mixture-of-Experts into Prompt Tuning", "abstract": "Parameter-efficient fine-tuning (PEFT) methods have shown promise in adapting large language models, yet existing approaches exhibit counter-intuitive phenomena: integrating router into prompt tuning (PT) increases training efficiency yet does not improve performance universally; parameter reduction through matrix decomposition can improve performance in specific domains. Motivated by these observations and the modular nature of PT, we propose PT-MoE, a novel framework that integrates matrix decomposition with mixture-of-experts (MoE) routing for efficient PT. Results across 17 datasets demonstrate that PT-MoE achieves state-of-the-art performance in both question answering (QA) and mathematical problem solving tasks, improving F1 score by 1.49 points over PT and 2.13 points over LoRA in QA tasks, while enhancing mathematical accuracy by 10.75 points over PT and 0.44 points over LoRA, all while using 25% fewer parameters than LoRA. Our analysis reveals that while PT methods generally excel in QA tasks and LoRA-based methods in math datasets, the integration of matrix decomposition and MoE in PT-MoE yields complementary benefits: decomposition enables efficient parameter sharing across experts while MoE provides dynamic adaptation, collectively enabling PT-MoE to demonstrate cross-task consistency and generalization abilities. These findings, along with ablation studies on routing mechanisms and architectural components, provide insights for future PEFT methods.", "authors": ["Zongqian Li", "Yixuan Su", "Nigel Collier"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-14", "url": "https://arxiv.org/abs/2505.09519", "pdf_url": "https://arxiv.org/pdf/2505.09519v1", "arxiv_id": "2505.09519", "doi": "10.48550/arXiv.2505.09519", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "6095be62d43d60e31072f634932fd99484526b5f62a70effcb32daa4c69885fa", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Unstructured Pruning of Mamba State-Space Models for Resource-Constrained Environments", "abstract": "State-space models (SSMs), particularly the Mamba architecture, have emerged as powerful alternatives to Transformers for sequence modeling, offering linear-time complexity and competitive performance across diverse tasks. However, their large parameter counts pose significant challenges for deployment in resource-constrained environments. We propose a novel unstructured pruning framework tailored for Mamba models that achieves up to 70\\% parameter reduction while retaining over 95\\% of the original performance. Our approach integrates three key innovations: (1) a gradient-aware magnitude pruning technique that combines weight magnitude and gradient information to identify less critical parameters, (2) an iterative pruning schedule that gradually increases sparsity to maintain model stability, and (3) a global pruning strategy that optimizes parameter allocation across the entire model. Through extensive experiments on WikiText-103, Long Range Arena, and ETT time-series benchmarks, we demonstrate significant efficiency gains with minimal performance degradation. Our analysis of pruning effects on Mamba's components reveals critical insights into the architecture's redundancy and robustness, enabling practical deployment in resource-constrained settings while broadening Mamba's applicability.", "authors": ["Ibne Farabi Shihab", "Sanjeda Akter", "Anuj Sharma"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-13", "url": "https://arxiv.org/abs/2505.08299", "pdf_url": "https://arxiv.org/pdf/2505.08299v2", "arxiv_id": "2505.08299", "doi": "10.48550/arXiv.2505.08299", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.25}
{"id": "093117fefffbc65f88c786c6506c726ef08a62104858f991d1cfaf3fc23111cc", "sources": ["arxiv", "semantic_scholar"], "title": "SPAT: Sensitivity-based Multihead-attention Pruning on Time Series Forecasting Models", "abstract": "Attention-based architectures have achieved superior performance in multivariate time series forecasting but are computationally expensive. Techniques such as patching and adaptive masking have been developed to reduce their sizes and latencies. In this work, we propose a structured pruning method, SPAT ($\\textbf{S}$ensitivity $\\textbf{P}$runer for $\\textbf{At}$tention), which selectively removes redundant attention mechanisms and yields highly effective models. Different from previous approaches, SPAT aims to remove the entire attention module, which reduces the risk of overfitting and enables speed-up without demanding specialized hardware. We propose a dynamic sensitivity metric, $\\textbf{S}$ensitivity $\\textbf{E}$nhanced $\\textbf{N}$ormalized $\\textbf{D}$ispersion (SEND) that measures the importance of each attention module during the pre-training phase. Experiments on multivariate datasets demonstrate that SPAT-pruned models achieve reductions of 2.842% in MSE, 1.996% in MAE, and 35.274% in FLOPs. Furthermore, SPAT-pruned models outperform existing lightweight, Mamba-based and LLM-based SOTA methods in both standard and zero-shot inference, highlighting the importance of retaining only the most effective attention mechanisms. We have made our code publicly available https://anonymous.4open.science/r/SPAT-6042.", "authors": ["Suhan Guo", "Jiahong Deng", "Mengjun Yi", "Furao Shen", "Jian Zhao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-13", "url": "https://arxiv.org/abs/2505.08768", "pdf_url": "https://arxiv.org/pdf/2505.08768v1", "arxiv_id": "2505.08768", "doi": "10.48550/arXiv.2505.08768", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1616}
{"id": "de12de7aeb43f6e33fdb6e1f5ebc017a0bf8872e6c6d73d27af8b84574ec116f", "sources": ["arxiv", "semantic_scholar"], "title": "PWC-MoE: Privacy-Aware Wireless Collaborative Mixture of Experts", "abstract": "Large language models (LLMs) hosted on cloud servers alleviate the computational and storage burdens on local devices but raise privacy concerns due to sensitive data transmission and require substantial communication bandwidth, which is challenging in constrained environments. In contrast, small language models (SLMs) running locally enhance privacy but suffer from limited performance on complex tasks. To balance computational cost, performance, and privacy protection under bandwidth constraints, we propose a privacy-aware wireless collaborative mixture of experts (PWC-MoE) framework. Specifically, PWC-MoE employs a sparse privacy-aware gating network to dynamically route sensitive tokens to privacy experts located on local clients, while non-sensitive tokens are routed to non-privacy experts located at the remote base station. To achieve computational efficiency, the gating network ensures that each token is dynamically routed to and processed by only one expert. To enhance scalability and prevent overloading of specific experts, we introduce a group-wise load-balancing mechanism for the gating network that evenly distributes sensitive tokens among privacy experts and non-sensitive tokens among non-privacy experts. To adapt to bandwidth constraints while preserving model performance, we propose a bandwidth-adaptive and importance-aware token offloading scheme. This scheme incorporates an importance predictor to evaluate the importance scores of non-sensitive tokens, prioritizing the most important tokens for transmission to the base station based on their predicted importance and the available bandwidth. Experiments demonstrate that the PWC-MoE framework effectively preserves privacy and maintains high performance even in bandwidth-constrained environments, offering a practical solution for deploying LLMs in privacy-sensitive and bandwidth-limited scenarios.", "authors": ["Yang Su", "Na Yan", "Yansha Deng", "Robert Schober"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-13", "url": "https://arxiv.org/abs/2505.08719", "pdf_url": "https://arxiv.org/pdf/2505.08719v1", "arxiv_id": "2505.08719", "doi": "10.48550/arXiv.2505.08719", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1616}
{"id": "9a0819be97c3571249ee463f72c9abeef194f56b41e8c4a1e00374d2c4f6184b", "sources": ["arxiv", "semantic_scholar"], "title": "Toward Cost-Efficient Serving of Mixture-of-Experts with Asynchrony", "abstract": "Mixture-of-Experts (MoE) architectures offer the promise of larger model capacity without the prohibitive costs of fully dense designs. However, in real-world inference serving, load skew across experts often leads to suboptimal device utilization and excessive synchronization overheads. This paper introduces Asynchronous Expert Parallelism (AEP), a new paradigm that decouples layer execution from barrier-style synchronization. By dynamically queuing tokens at each layer (referred to as $μ$-queuing) and adaptively re-batching them on demand, GPUs avoid waiting for straggling experts and instead continuously process whichever layer is ready. This asynchronous approach mitigates two major inefficiencies in traditional expert-parallel systems: (1) idle GPU time while waiting for the hottest expert, and (2) small-batch executions on colder experts that waste memory bandwidth. We implement these ideas in a serving system called AMoE, which disaggregates attention from expert layers and uses a defragging scheduler to reduce batch fragmentation. Evaluations on prototype MoE models show that AMoE improves throughput by up to 2.7x compared to state-of-the-art baselines, incurring a manageable latency penalty and providing a cost-effective operating point. Furthermore, experiments demonstrate nearly linear scalability to multi-node settings, whereas the baseline system shows no throughput increase even when the number of GPUs is doubled.", "authors": ["Shaoyu Wang", "Guangrong He", "Geon-Woo Kim", "Yanqi Zhou", "Seo Jin Park"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-13", "url": "https://arxiv.org/abs/2505.08944", "pdf_url": "https://arxiv.org/pdf/2505.08944v2", "arxiv_id": "2505.08944", "doi": "10.48550/arXiv.2505.08944", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1616}
{"id": "81a206cd38404c59c5f3ff511aad7a12b5a8974665c9ef589eae52d4f69a1148", "sources": ["arxiv", "semantic_scholar"], "title": "Block-Biased Mamba for Long-Range Sequence Processing", "abstract": "Mamba extends earlier state space models (SSMs) by introducing input-dependent dynamics, and has demonstrated strong empirical performance across a range of domains, including language modeling, computer vision, and foundation models. However, a surprising weakness remains: despite being built on architectures designed for long-range dependencies, Mamba performs poorly on long-range sequential tasks. Understanding and addressing this gap is important for improving Mamba's universality and versatility. In this work, we analyze Mamba's limitations through three perspectives: expressiveness, inductive bias, and training stability. Our theoretical results show how Mamba falls short in each of these aspects compared to earlier SSMs such as S4D. To address these issues, we propose $\\text{B}_2\\text{S}_6$, a simple extension of Mamba's S6 unit that combines block-wise selective dynamics with a channel-specific bias. We prove that these changes equip the model with a better-suited inductive bias and improve its expressiveness and stability. Empirically, $\\text{B}_2\\text{S}_6$ outperforms S4 and S4D on Long-Range Arena (LRA) tasks while maintaining Mamba's performance on language modeling benchmarks.", "authors": ["Annan Yu", "N. Benjamin Erichson"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-13", "url": "https://arxiv.org/abs/2505.09022", "pdf_url": "https://arxiv.org/pdf/2505.09022v1", "arxiv_id": "2505.09022", "doi": "10.48550/arXiv.2505.09022", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "e65f928f823abff840524ed091ded664fb257604f94081829a60452a66a25f5c", "sources": ["arxiv", "semantic_scholar"], "title": "UMoE: Unifying Attention and FFN with Shared Experts", "abstract": "Sparse Mixture of Experts (MoE) architectures have emerged as a promising approach for scaling Transformer models. While initial works primarily incorporated MoE into feed-forward network (FFN) layers, recent studies have explored extending the MoE paradigm to attention layers to enhance model performance. However, existing attention-based MoE layers require specialized implementations and demonstrate suboptimal performance compared to their FFN-based counterparts. In this paper, we aim to unify MoE designs in attention and FFN layers by introducing a novel reformulation of the attention mechanism, that reveals an underlying FFN-like structure within attention modules. Our proposed architecture, UMoE, achieves superior performance through attention-based MoE layers while enabling efficient parameter sharing between FFN and attention components.", "authors": ["Yuanhang Yang", "Chaozheng Wang", "Jing Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-12", "url": "https://arxiv.org/abs/2505.07260", "pdf_url": "https://arxiv.org/pdf/2505.07260v2", "arxiv_id": "2505.07260", "doi": "10.48550/arXiv.2505.07260", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1604}
{"id": "1e095d30f3992e5b9ea18274b964ddc0d84226c35a005cc039592573cf0539b5", "sources": ["arxiv", "semantic_scholar"], "title": "The power of fine-grained experts: Granularity boosts expressivity in Mixture of Experts", "abstract": "Mixture-of-Experts (MoE) layers are increasingly central to frontier model architectures. By selectively activating parameters, they reduce computational cost while scaling total parameter count. This paper investigates the impact of the number of active experts, termed granularity, comparing architectures with many (e.g., 8 per layer in DeepSeek) to those with fewer (e.g., 1 per layer in Llama-4 models). We prove an exponential separation in network expressivity based on this design parameter, suggesting that models benefit from higher granularity. Experimental results corroborate our theoretical findings and illustrate this separation.", "authors": ["Enric Boix-Adsera", "Philippe Rigollet"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-11", "url": "https://arxiv.org/abs/2505.06839", "pdf_url": "https://arxiv.org/pdf/2505.06839v1", "arxiv_id": "2505.06839", "doi": "10.48550/arXiv.2505.06839", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "cb8fd32fa974bbb62e1539cc4282eb99ae944f5ce66bc4611ba2b48460ee9986", "sources": ["arxiv", "semantic_scholar"], "title": "PRIME: Physics-Related Intelligent Mixture of Experts for Transistor Characteristics Prediction", "abstract": "In recent years, machine learning has been extensively applied to data prediction during process ramp-up, with a particular focus on transistor characteristics for circuit design and manufacture. However, capturing the nonlinear current response across multiple operating regions remains a challenge for neural networks. To address such challenge, a novel machine learning framework, PRIME (Physics-Related Intelligent Mixture of Experts), is proposed to capture and integrate complex regional characteristics. In essence, our framework incorporates physics-based knowledge with data-driven intelligence. By leveraging a dynamic weighting mechanism in its gating network, PRIME adaptively activates the suitable expert model based on distinct input data features. Extensive evaluations are conducted on various gate-all-around (GAA) structures to examine the effectiveness of PRIME and considerable improvements (60\\%-84\\%) in prediction accuracy are shown over state-of-the-art models.", "authors": ["Zhenxing Dou", "Yijiao Wang", "Tao Zou", "Zhiwei Chen", "Fei Liu", "Peng Wang", "Weisheng Zhao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-10", "url": "https://arxiv.org/abs/2505.11523", "pdf_url": "https://arxiv.org/pdf/2505.11523v1", "arxiv_id": "2505.11523", "doi": "10.48550/arXiv.2505.11523", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1581}
{"id": "da5cb9abb8a3718245515a6d91fd5f6f15823af90b8cf5d1b65b1eccbae667d2", "sources": ["arxiv", "semantic_scholar"], "title": "Collaborative Multi-LoRA Experts with Achievement-based Multi-Tasks Loss for Unified Multimodal Information Extraction", "abstract": "Multimodal Information Extraction (MIE) has gained attention for extracting structured information from multimedia sources. Traditional methods tackle MIE tasks separately, missing opportunities to share knowledge across tasks. Recent approaches unify these tasks into a generation problem using instruction-based T5 models with visual adaptors, optimized through full-parameter fine-tuning. However, this method is computationally intensive, and multi-task fine-tuning often faces gradient conflicts, limiting performance. To address these challenges, we propose collaborative multi-LoRA experts with achievement-based multi-task loss (C-LoRAE) for MIE tasks. C-LoRAE extends the low-rank adaptation (LoRA) method by incorporating a universal expert to learn shared multimodal knowledge from cross-MIE tasks and task-specific experts to learn specialized instructional task features. This configuration enhances the model's generalization ability across multiple tasks while maintaining the independence of various instruction tasks and mitigating gradient conflicts. Additionally, we propose an achievement-based multi-task loss to balance training progress across tasks, addressing the imbalance caused by varying numbers of training samples in MIE tasks. Experimental results on seven benchmark datasets across three key MIE tasks demonstrate that C-LoRAE achieves superior overall performance compared to traditional fine-tuning methods and LoRA methods while utilizing a comparable number of training parameters to LoRA.", "authors": ["Li Yuan", "Yi Cai", "Xudong Shen", "Qing Li", "Qingbao Huang", "Zikun Deng", "Tao Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-08", "url": "https://arxiv.org/abs/2505.06303", "pdf_url": "https://arxiv.org/pdf/2505.06303v1", "arxiv_id": "2505.06303", "doi": "10.48550/arXiv.2505.06303", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Joint Conference on Artificial Intelligence", "quality_score": 0.2258}
{"id": "9680750d4f37690e075e925e6933f50ea17fd4ba7b90be74b5719aad3fcb33fd", "sources": ["arxiv", "semantic_scholar"], "title": "Low-bit Model Quantization for Deep Neural Networks: A Survey", "abstract": "With unprecedented rapid development, deep neural networks (DNNs) have deeply influenced almost all fields. However, their heavy computation costs and model sizes are usually unacceptable in real-world deployment. Model quantization, an effective weight-lighting technique, has become an indispensable procedure in the whole deployment pipeline. The essence of quantization acceleration is the conversion from continuous floating-point numbers to discrete integer ones, which significantly speeds up the memory I/O and calculation, i.e., addition and multiplication. However, performance degradation also comes with the conversion because of the loss of precision. Therefore, it has become increasingly popular and critical to investigate how to perform the conversion and how to compensate for the information loss. This article surveys the recent five-year progress towards low-bit quantization on DNNs. We discuss and compare the state-of-the-art quantization methods and classify them into 8 main categories and 24 sub-categories according to their core techniques. Furthermore, we shed light on the potential research opportunities in the field of model quantization. A curated list of model quantization is provided at https://github.com/Kai-Liu001/Awesome-Model-Quantization.", "authors": ["Kai Liu", "Qian Zheng", "Kaiwen Tao", "Zhiteng Li", "Haotong Qin", "Wenbo Li", "Yong Guo", "Xianglong Liu", "Linghe Kong", "Guihai Chen", "Yulun Zhang", "Xiaokang Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-08", "url": "https://arxiv.org/abs/2505.05530", "pdf_url": "https://arxiv.org/pdf/2505.05530v1", "arxiv_id": "2505.05530", "doi": "10.48550/arXiv.2505.05530", "citation_count": 21, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/Kai-Liu001/Awesome-Model-Quantization", "venue": "arXiv.org", "quality_score": 0.3356}
{"id": "f628dc1e0eb3d8e36fa2bd07fa34082dcba2dd55e72f33124afbe0fa4703f309", "sources": ["arxiv"], "title": "Divide-and-Conquer: Cold-Start Bundle Recommendation via Mixture of Diffusion Experts", "abstract": "Cold-start bundle recommendation focuses on modeling new bundles with insufficient information to provide recommendations. Advanced bundle recommendation models usually learn bundle representations from multiple views (e.g., interaction view) at both the bundle and item levels. Consequently, the cold-start problem for bundles is more challenging than that for traditional items due to the dual-level multi-view complexity. In this paper, we propose a novel Mixture of Diffusion Experts (MoDiffE) framework, which employs a divide-and-conquer strategy for cold-start bundle recommendation and follows three steps:(1) Divide: The bundle cold-start problem is divided into independent but similar sub-problems sequentially by level and view, which can be summarized as the poor representation of feature-missing bundles in prior-embedding models. (2) Conquer: Beyond prior-embedding models that fundamentally provide the embedded representations, we introduce a diffusion-based method to solve all sub-problems in a unified way, which directly generates diffusion representations using diffusion models without depending on specific features. (3) Combine: A cold-aware hierarchical Mixture of Experts (MoE) is employed to combine results of the sub-problems for final recommendations, where the two models for each view serve as experts and are adaptively fused for different bundles in a multi-layer manner. Additionally, MoDiffE adopts a multi-stage decoupled training pipeline and introduces a cold-start gating augmentation method to enable the training of gating for cold bundles. Through extensive experiments on three real-world datasets, we demonstrate that MoDiffE significantly outperforms existing solutions in handling cold-start bundle recommendation. It achieves up to a 0.1027 absolute gain in Recall@20 in cold-start scenarios and up to a 47.43\\% relative improvement in all-bundle scenarios.", "authors": ["Ming Li", "Lin Li", "Xiaohui Tao", "Dong Zhang", "Jimmy Xiangji Huang"], "categories": ["cs.IR"], "fields_of_study": [], "published_date": "2025-05-08", "url": "https://arxiv.org/abs/2505.05035", "pdf_url": "https://arxiv.org/pdf/2505.05035v1", "arxiv_id": "2505.05035", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0992}
{"id": "c41be3f4bbe3302034dacbd9cfac24a167632e826dbd65ad60697a6e13c83f92", "sources": ["arxiv", "semantic_scholar"], "title": "MoDE: Mixture of Diffusion Experts for Any Occluded Face Recognition", "abstract": "With the continuous impact of epidemics, people have become accustomed to wearing masks. However, most current occluded face recognition (OFR) algorithms lack prior knowledge of occlusions, resulting in poor performance when dealing with occluded faces of varying types and severity in reality. Recognizing occluded faces is still a significant challenge, which greatly affects the convenience of people's daily lives. In this paper, we propose an identity-gated mixture of diffusion experts (MoDE) for OFR. Each diffusion-based generative expert estimates one possible complete image for occluded faces. Considering the random sampling process of the diffusion model, which introduces inevitable differences and variations between the inpainted faces and the real ones. To ensemble effective information from multi-reconstructed faces, we introduce an identity-gating network to evaluate the contribution of each reconstructed face to the identity and adaptively integrate the predictions in the decision space. Moreover, our MoDE is a plug-and-play module for most existing face recognition models. Extensive experiments on three public face datasets and two datasets in the wild validate our advanced performance for various occlusions in comparison with the competing methods.", "authors": ["Qiannan Fan", "Zhuoyang Li", "Jitong Li", "Chenyang Cao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-07", "url": "https://arxiv.org/abs/2505.04306", "pdf_url": "https://arxiv.org/pdf/2505.04306v1", "arxiv_id": "2505.04306", "doi": "10.1109/IJCNN64981.2025.11228891", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Joint Conference on Neural Network", "quality_score": 0.1547}
{"id": "6f1897dea02108c42b26a63a509bfb9bfa5750ed504c3823e0c22129db9bd37a", "sources": ["arxiv", "semantic_scholar"], "title": "A Sensitivity-Driven Expert Allocation Method in LoRA-MoE for Efficient Fine-Tuning", "abstract": "As deep learning models expand, the pre-training-fine-tuning paradigm has become the standard approach for handling various downstream tasks. However, shared parameters can lead to diminished performance when dealing with complex datasets involving multiple tasks. While introducing Mixture-of-Experts (MoE) methods has alleviated this issue to some extent, it also significantly increases the number of parameters required for fine-tuning and training time, introducing greater parameter redundancy. To address these challenges, we propose a method for allocating expert numbers based on parameter sensitivity LoRA-SMoE (A Sensitivity-Driven Expert Allocation Method in LoRA-MoE for Efficient Fine-Tuning). This method rapidly assesses the sensitivity of different tasks to parameters by sampling a small amount of data and using gradient information. It then adaptively allocates expert numbers within a given budget. The process maintains comparable memory consumption to LoRA (Low-Rank Adaptation) while ensuring an efficient and resource-friendly fine-tuning procedure. Experimental results demonstrate that compared to SOTA fine-tuning methods, our LoRA-SMoE approach can enhance model performance while reducing the number of trainable parameters. This significantly improves model performance in resource-constrained environments. Additionally, due to its efficient parameter sensitivity evaluation mechanism, LoRA-SMoE requires minimal computational overhead to optimize expert allocation, making it particularly suitable for scenarios with limited computational resources. All the code in this study will be made publicly available following the acceptance of the paper for publication. Source code is at https://github.com/EMLS-ICTCAS/LoRA-SMoE", "authors": ["Junzhou Xu", "Boyu Diao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-06", "url": "https://arxiv.org/abs/2505.06272", "pdf_url": "https://arxiv.org/pdf/2505.06272v1", "arxiv_id": "2505.06272", "doi": "10.1109/CCGRID64434.2025.00059", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/EMLS-ICTCAS/LoRA-SMoE", "venue": "IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing", "quality_score": 0.2373}
{"id": "b9a7ba7d8c25ad23df3b166601dd8faffee67b60de1bf3d3afd126c7fdf9ad09", "sources": ["arxiv", "semantic_scholar"], "title": "An Active Inference Model of Covert and Overt Visual Attention", "abstract": "The ability to selectively attend to relevant stimuli while filtering out distractions is essential for agents that process complex, high-dimensional sensory input. This paper introduces a model of covert and overt visual attention through the framework of active inference, utilizing dynamic optimization of sensory precisions to minimize free-energy. The model determines visual sensory precisions based on both current environmental beliefs and sensory input, influencing attentional allocation in both covert and overt modalities. To test the effectiveness of the model, we analyze its behavior in the Posner cueing task and a simple target focus task using two-dimensional(2D) visual data. Reaction times are measured to investigate the interplay between exogenous and endogenous attention, as well as valid and invalid cueing. The results show that exogenous and valid cues generally lead to faster reaction times compared to endogenous and invalid cues. Furthermore, the model exhibits behavior similar to inhibition of return, where previously attended locations become suppressed after a specific cue-target onset asynchrony interval. Lastly, we investigate different aspects of overt attention and show that involuntary, reflexive saccades occur faster than intentional ones, but at the expense of adaptability.", "authors": ["Tin Mišić", "Karlo Koledić", "Fabio Bonsignorio", "Ivan Petrović", "Ivan Marković"], "categories": ["cs.CV", "cs.AI", "q-bio.NC"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2025-05-06", "url": "https://arxiv.org/abs/2505.03856", "pdf_url": "https://arxiv.org/pdf/2505.03856v1", "arxiv_id": "2505.03856", "doi": "10.48550/arXiv.2505.03856", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/unizgfer-lamor/ainf-visual-attention", "venue": "arXiv.org", "quality_score": 0.2373}
{"id": "2121311e1a12305bb035612ef04d0ba93b88e9cb8e773c0b439d1216ebc589f1", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Heterogeneous Mixture of Scene Experts for Large-scale Neural Radiance Fields", "abstract": "Recent NeRF methods on large-scale scenes have underlined the importance of scene decomposition for scalable NeRFs. Although achieving reasonable scalability, there are several critical problems remaining unexplored, i.e., learnable decomposition, modeling scene heterogeneity, and modeling efficiency. In this paper, we introduce Switch-NeRF++, a Heterogeneous Mixture of Hash Experts (HMoHE) network that addresses these challenges within a unified framework. It is a highly scalable NeRF that learns heterogeneous decomposition and heterogeneous NeRFs efficiently for large-scale scenes in an end-to-end manner. In our framework, a gating network learns to decompose scenes and allocates 3D points to specialized NeRF experts. This gating network is co-optimized with the experts by our proposed Sparsely Gated Mixture of Experts (MoE) NeRF framework. We incorporate a hash-based gating network and distinct heterogeneous hash experts. The hash-based gating efficiently learns the decomposition of the large-scale scene. The distinct heterogeneous hash experts consist of hash grids of different resolution ranges, enabling effective learning of the heterogeneous representation of different scene parts. These design choices make our framework an end-to-end and highly scalable NeRF solution for real-world large-scale scene modeling to achieve both quality and efficiency. We evaluate our accuracy and scalability on existing large-scale NeRF datasets and a new dataset with very large-scale scenes ($>6.5km^2$) from UrbanBIS. Extensive experiments demonstrate that our approach can be easily scaled to various large-scale scenes and achieve state-of-the-art scene rendering accuracy. Furthermore, our method exhibits significant efficiency, with an 8x acceleration in training and a 16x acceleration in rendering compared to Switch-NeRF. Codes will be released at https://github.com/MiZhenxing/Switch-NeRF.", "authors": ["Zhenxing Mi", "Ping Yin", "Xue Xiao", "Dan Xu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2025-05-04", "url": "https://arxiv.org/abs/2505.02005", "pdf_url": "https://arxiv.org/pdf/2505.02005v2", "arxiv_id": "2505.02005", "doi": "10.1109/TPAMI.2025.3603305", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/MiZhenxing/Switch-NeRF", "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.2337}
{"id": "865fc294938d2a135ccef74d6a83d905234c1b0a8b636627e104cf8a0bf6c1b8", "sources": ["arxiv", "semantic_scholar"], "title": "BadPatches: Routing-aware Backdoor Attacks on Vision Mixture of Experts", "abstract": "Mixture of Experts (MoE) architectures have gained popularity for reducing computational costs in deep neural networks by activating only a subset of parameters during inference. While this efficiency makes MoE attractive for vision tasks, the patch-based processing in vision models introduces new methods for adversaries to perform backdoor attacks. In this work, we investigate the vulnerability of vision MoE models for image classification, specifically the patch-based MoE (pMoE) models and MoE-based vision transformers, against backdoor attacks. We propose a novel routing-aware trigger application method BadPatches, which is designed for patch-based processing in vision MoE models. BadPatches applies triggers on image patches rather than on the entire image. We show that BadPatches achieves high attack success rates (ASRs) with lower poisoning rates than routing-agnostic triggers and is successful at poisoning rates as low as 0.01% with an ASR above 80% on pMoE. Moreover, BadPatches is still effective when an adversary does not have complete knowledge of the patch routing configuration of the considered models. Next, we explore how trigger design affects pMoE patch routing. Finally, we investigate fine-pruning as a defense. Results show that only the fine-tuning stage of fine-pruning removes the backdoor from the model.", "authors": ["Cedric Chan", "Jona te Lintelo", "Stjepan Picek"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-03", "url": "https://arxiv.org/abs/2505.01811", "pdf_url": "https://arxiv.org/pdf/2505.01811v3", "arxiv_id": "2505.01811", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0955}
{"id": "8b8a43d8b3be39541d857a30645bde8fbe8f19fd2bca5c29339ea5f94777de23", "sources": ["arxiv", "semantic_scholar"], "title": "MoEQuant: Enhancing Quantization for Mixture-of-Experts Large Language Models via Expert-Balanced Sampling and Affinity Guidance", "abstract": "Mixture-of-Experts (MoE) large language models (LLMs), which leverage dynamic routing and sparse activation to enhance efficiency and scalability, have achieved higher performance while reducing computational costs. However, these models face significant memory overheads, limiting their practical deployment and broader adoption. Post-training quantization (PTQ), a widely used method for compressing LLMs, encounters severe accuracy degradation and diminished generalization performance when applied to MoE models. This paper investigates the impact of MoE's sparse and dynamic characteristics on quantization and identifies two primary challenges: (1) Inter-expert imbalance, referring to the uneven distribution of samples across experts, which leads to insufficient and biased calibration for less frequently utilized experts; (2) Intra-expert imbalance, arising from MoE's unique aggregation mechanism, which leads to varying degrees of correlation between different samples and their assigned experts. To address these challenges, we propose MoEQuant, a novel quantization framework tailored for MoE LLMs. MoE-Quant includes two novel techniques: 1) Expert-Balanced Self-Sampling (EBSS) is an efficient sampling method that efficiently constructs a calibration set with balanced expert distributions by leveraging the cumulative probabilities of tokens and expert balance metrics as guiding factors. 2) Affinity-Guided Quantization (AGQ), which incorporates affinities between experts and samples into the quantization process, thereby accurately assessing the impact of individual samples on different experts within the MoE layer. Experiments demonstrate that MoEQuant achieves substantial performance gains (more than 10 points accuracy gain in the HumanEval for DeepSeekMoE-16B under 4-bit quantization) and boosts efficiency.", "authors": ["Xing Hu", "Zhixuan Chen", "Dawei Yang", "Zukang Xu", "Chen Xu", "Zhihang Yuan", "Sifan Zhou", "Jiangyong Yu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-02", "url": "https://arxiv.org/abs/2505.03804", "pdf_url": "https://arxiv.org/pdf/2505.03804v1", "arxiv_id": "2505.03804", "doi": "10.48550/arXiv.2505.03804", "citation_count": 27, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3618}
{"id": "5d0aa1154539eca7849b73516b9e1e1b2209b9ac08c93c668f9e976bf5707a21", "sources": ["arxiv", "semantic_scholar"], "title": "CoCoAFusE: Beyond Mixtures of Experts via Model Fusion", "abstract": "Many learning problems involve multiple patterns and varying degrees of uncertainty dependent on the covariates. Advances in Deep Learning (DL) have addressed these issues by learning highly nonlinear input-output dependencies. However, model interpretability and Uncertainty Quantification (UQ) have often straggled behind. In this context, we introduce the Competitive/Collaborative Fusion of Experts (CoCoAFusE), a novel, Bayesian Covariates-Dependent Modeling technique. CoCoAFusE builds on the very philosophy behind Mixtures of Experts (MoEs), blending predictions from several simple sub-models (or \"experts\") to achieve high levels of expressiveness while retaining a substantial degree of local interpretability. Our formulation extends that of a classical Mixture of Experts by contemplating the fusion of the experts' distributions in addition to their more usual mixing (i.e., superimposition). Through this additional feature, CoCoAFusE better accommodates different scenarios for the intermediate behavior between generating mechanisms, resulting in tighter credible bounds on the response variable. Indeed, only resorting to mixing, as in classical MoEs, may lead to multimodality artifacts, especially over smooth transitions. Instead, CoCoAFusE can avoid these artifacts even under the same structure and priors for the experts, leading to greater expressiveness and flexibility in modeling. This new approach is showcased extensively on a suite of motivating numerical examples and a collection of real-data ones, demonstrating its efficacy in tackling complex regression problems where uncertainty is a key quantity of interest.", "authors": ["Aurelio Raffa Ugolini", "Mara Tanelli", "Valentina Breschi"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-02", "url": "https://arxiv.org/abs/2505.01105", "pdf_url": "https://arxiv.org/pdf/2505.01105v1", "arxiv_id": "2505.01105", "doi": "10.48550/arXiv.2505.01105", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.149}
{"id": "5e30efb942712f326734ad86b70b49da828429274c808faaa835bf7957ce2a95", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routing", "abstract": "Recent advances in large language models highlighted the excessive quadratic cost of self-attention. Despite the significant research efforts, subquadratic attention methods still suffer from inferior performance in practice. We hypothesize that dynamic, learned content-based sparsity can lead to more efficient attention mechanisms. We present Mixture of Sparse Attention (MoSA), a novel approach inspired by Mixture of Experts (MoE) with expert choice routing. MoSA dynamically selects tokens for each attention head, allowing arbitrary sparse attention patterns. By selecting $k$ tokens from a sequence of length $T$, MoSA reduces the computational complexity of each attention head from $O(T^2)$ to $O(k^2 + T)$. This enables using more heads within the same computational budget, allowing higher specialization. We show that among the tested sparse attention variants, MoSA is the only one that can outperform the dense baseline, sometimes with up to 27% better perplexity for an identical compute budget. MoSA can also reduce the resource usage compared to dense self-attention. Despite using torch implementation without an optimized kernel, perplexity-matched MoSA models are simultaneously faster in wall-clock time, require less memory for training, and drastically reduce the size of the KV-cache compared to the dense transformer baselines.", "authors": ["Piotr Piękos", "Róbert Csordás", "Jürgen Schmidhuber"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-01", "url": "https://arxiv.org/abs/2505.00315", "pdf_url": "https://arxiv.org/pdf/2505.00315v1", "arxiv_id": "2505.00315", "doi": "10.48550/arXiv.2505.00315", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "49d84ea516a58a2027c12fd553db197cd551e197155be319c39e951a8572ef28", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Routing in Sparse Mixture of Experts with Graph of Tokens", "abstract": "Sparse Mixture of Experts (SMoE) has emerged as a key to achieving unprecedented scalability in deep learning. By activating only a small subset of parameters per sample, SMoE achieves an exponential increase in parameter counts while maintaining a constant computational overhead. However, SMoE models are susceptible to routing fluctuations--changes in the routing of a given input to its target expert--at the late stage of model training, leading to model non-robustness. In this work, we unveil the limitation of SMoE through the perspective of the probabilistic graphical model (PGM). Through this PGM framework, we highlight the independence in the expert-selection of tokens, which exposes the model to routing fluctuation and non-robustness. Alleviating this independence, we propose the novel Similarity-Aware (S)MoE, which considers interactions between tokens during expert selection. We then derive a new PGM underlying an (S)MoE-Attention block, going beyond just a single (S)MoE layer. Leveraging the token similarities captured by the attention matrix, we propose the innovative Attention-Aware (S)MoE, which employs the attention matrix to guide the routing of tokens to appropriate experts in (S)MoE. We theoretically prove that Similarity/Attention-Aware routing help reduce the entropy of expert selection, resulting in more stable token routing mechanisms. We empirically validate our models on various tasks and domains, showing significant improvements in reducing routing fluctuations, enhancing accuracy, and increasing model robustness over the baseline MoE-Transformer with token routing via softmax gating.", "authors": ["Tam Nguyen", "Ngoc N. Tran", "Khai Nguyen", "Richard G. Baraniuk"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-01", "url": "https://arxiv.org/abs/2505.00792", "pdf_url": "https://arxiv.org/pdf/2505.00792v1", "arxiv_id": "2505.00792", "doi": "10.48550/arXiv.2505.00792", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "16110654a1cd7c508f5a842f5ef4006a6990ad286162dc5ec8b498156db163ab", "sources": ["arxiv", "semantic_scholar"], "title": "MoxE: Mixture of xLSTM Experts with Entropy-Aware Routing for Efficient Language Modeling", "abstract": "This paper introduces MoxE, a novel architecture that synergistically combines the Extended Long Short-Term Memory (xLSTM) with the Mixture of Experts (MoE) framework to address critical scalability and efficiency challenges in large language models (LLMs). The proposed method effectively leverages xLSTM's innovative memory structures while strategically introducing sparsity through MoE to substantially reduce computational overhead. At the heart of our approach is a novel entropy-based routing mechanism, designed to dynamically route tokens to specialized experts, thereby ensuring efficient and balanced resource utilization. This entropy awareness enables the architecture to effectively manage both rare and common tokens, with mLSTM blocks being favored to handle rare tokens. To further enhance generalization, we introduce a suite of auxiliary losses, including entropy-based and group-wise balancing losses, ensuring robust performance and efficient training. Theoretical analysis and empirical evaluations rigorously demonstrate that MoxE achieves significant efficiency gains and enhanced effectiveness compared to existing approaches, marking a notable advancement in scalable LLM architectures.", "authors": ["Abdoul Majid O. Thiombiano", "Brahim Hnich", "Ali Ben Mrad", "Mohamed Wiem Mkaouer"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-01", "url": "https://arxiv.org/abs/2505.01459", "pdf_url": "https://arxiv.org/pdf/2505.01459v1", "arxiv_id": "2505.01459", "doi": "10.48550/arXiv.2505.01459", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "29292aecbbd735a330a4e9d3bc30b04ac40399d8df5171283360577394c2262b", "sources": ["arxiv", "semantic_scholar"], "title": "TT-LoRA MoE: Unifying Parameter-Efficient Fine-Tuning and Sparse Mixture-of-Experts", "abstract": "We propose Tensor-Trained Low-Rank Adaptation Mixture of Experts (TT-LoRA MoE), a novel computational framework integrating Parameter-Efficient Fine-Tuning (PEFT) with sparse MoE routing to address scalability challenges in large model deployments. Unlike traditional MoE approaches, which face substantial computational overhead as expert counts grow, TT-LoRA MoE decomposes training into two distinct, optimized stages. First, we independently train lightweight, tensorized low-rank adapters (TT-LoRA experts), each specialized for specific tasks. Subsequently, these expert adapters remain frozen, eliminating inter-task interference and catastrophic forgetting in multi-task setting. A sparse MoE router, trained separately, dynamically leverages base model representations to select exactly one specialized adapter per input at inference time, automating expert selection without explicit task specification. Comprehensive experiments confirm our architecture retains the memory efficiency of low-rank adapters, seamlessly scales to large expert pools, and achieves robust task-level optimization. This structured decoupling significantly enhances computational efficiency and flexibility: uses only 2% of LoRA, 0.3% of Adapters and 0.03% of AdapterFusion parameters and outperforms AdapterFusion by 4 value in multi-tasking, enabling practical and scalable multi-task inference deployments.", "authors": ["Pradip Kunwar", "Minh N. Vu", "Maanak Gupta", "Mahmoud Abdelsalam", "Manish Bhattarai"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-29", "url": "https://arxiv.org/abs/2504.21190", "pdf_url": "https://arxiv.org/pdf/2504.21190v1", "arxiv_id": "2504.21190", "doi": "10.1145/3712285.3759888", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Software Composition", "quality_score": 0.1945}
{"id": "e80619d8313201505538629c1866f50d758a15a45a8828bb5147683eb1ed9fbb", "sources": ["arxiv", "semantic_scholar"], "title": "CachePrune: Teaching LLMs What Not to Follow via KV-Cache Editing", "abstract": "Large Language Models (LLMs) are susceptible to indirect prompt injection attacks, where the model inadvertently responds to instructions injected into the prompt context. This vulnerability stems from LLMs' inability to distinguish between data and instructions within a prompt. We propose CachePrune, which defends against this attack by identifying and pruning neurons associated with instruction-following during KV cache encoding of the prompt context. The pruning steers the LLM toward interpreting the context purely as data rather than as instructions to follow. To identify these neurons, we introduce a neural attribution mechanism guided by a preferential attribution loss, and theoretically connect this loss to an upper bound of the Direct Preference Optimization (DPO) objective. Further, we improve the fidelity of neural attribution by leveraging an observed triggering effect in instruction-following. Our approach does not interfere with prompt formatting or incur test-time overhead during response generation. Experiments show that CachePrune significantly reduces the attack success rate while preserving the LLM's ability to follow user instructions.", "authors": ["Rui Wang", "Junda Wu", "Yu Xia", "Tong Yu", "Ruiyi Zhang", "Ryan Rossi", "Subrata Mitra", "Lina Yao", "Julian McAuley"], "categories": ["cs.CR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-29", "url": "https://arxiv.org/abs/2504.21228", "pdf_url": "https://arxiv.org/pdf/2504.21228v3", "arxiv_id": "2504.21228", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0926}
{"id": "dc25c1f9140a8edb6f438b34499280bec7eca986e174ac498257bf407b7ee2f0", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient LLMs with AMP: Attention Heads and MLP Pruning", "abstract": "Deep learning drives a new wave in computing systems and triggers the automation of increasingly complex problems. In particular, Large Language Models (LLMs) have significantly advanced cognitive tasks, often matching or even surpassing human-level performance. However, their extensive parameters result in high computational costs and slow inference, posing challenges for deployment in resource-limited settings. Among the strategies to overcome the aforementioned challenges, pruning emerges as a successful mechanism since it reduces model size while maintaining predictive ability. In this paper, we introduce AMP: Attention Heads and MLP Pruning, a novel structured pruning method that efficiently compresses LLMs by removing less critical structures within Multi-Head Attention (MHA) and Multilayer Perceptron (MLP). By projecting the input data onto weights, AMP assesses structural importance and overcomes the limitations of existing techniques, which often fall short in flexibility or efficiency. In particular, AMP surpasses the current state-of-the-art on commonsense reasoning tasks by up to 1.49 percentage points, achieving a 30% pruning ratio with minimal impact on zero-shot task performance. Moreover, AMP also improves inference speeds, making it well-suited for deployment in resource-constrained environments. We confirm the flexibility of AMP on different families of LLMs, including LLaMA and Phi.", "authors": ["Leandro Giusti Mugnaini", "Bruno Lopes Yamamoto", "Lucas Lauton de Alcantara", "Victor Zacarias", "Edson Bollis", "Lucas Pellicer", "Anna Helena Reali Costa", "Artur Jordao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-29", "url": "https://arxiv.org/abs/2504.21174", "pdf_url": "https://arxiv.org/pdf/2504.21174v1", "arxiv_id": "2504.21174", "doi": "10.1109/IJCNN64981.2025.11227985", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Joint Conference on Neural Network", "quality_score": 0.1505}
{"id": "f48cfa0d97a3b93fca8161275a87b2be6e1ab30c97125b0b85c03c6a9c4c1db3", "sources": ["arxiv", "semantic_scholar"], "title": "MicarVLMoE: A Modern Gated Cross-Aligned Vision-Language Mixture of Experts Model for Medical Image Captioning and Report Generation", "abstract": "Medical image reporting (MIR) aims to generate structured clinical descriptions from radiological images. Existing methods struggle with fine-grained feature extraction, multimodal alignment, and generalization across diverse imaging types, often relying on vanilla transformers and focusing primarily on chest X-rays. We propose MicarVLMoE, a vision-language mixture-of-experts model with gated cross-aligned fusion, designed to address these limitations. Our architecture includes: (i) a multiscale vision encoder (MSVE) for capturing anatomical details at varying resolutions, (ii) a multihead dual-branch latent attention (MDLA) module for vision-language alignment through latent bottleneck representations, and (iii) a modulated mixture-of-experts (MoE) decoder for adaptive expert specialization. We extend MIR to CT scans, retinal imaging, MRI scans, and gross pathology images, reporting state-of-the-art results on COVCTR, MMR, PGROSS, and ROCO datasets. Extensive experiments and ablations confirm improved clinical accuracy, cross-modal alignment, and model interpretability. Code is available at https://github.com/AI-14/micar-vl-moe.", "authors": ["Amaan Izhar", "Nurul Japar", "Norisma Idris", "Ting Dang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-29", "url": "https://arxiv.org/abs/2504.20343", "pdf_url": "https://arxiv.org/pdf/2504.20343v1", "arxiv_id": "2504.20343", "doi": "10.1109/IJCNN64981.2025.11228227", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AI-14/micar-vl-moe", "venue": "IEEE International Joint Conference on Neural Network", "quality_score": 0.2249}
{"id": "1c212f3150e4877f613b00325df873c238b03f7d9d43e49899daa78527c2ef4b", "sources": ["arxiv", "semantic_scholar"], "title": "ARTEMIS: Autoregressive End-to-End Trajectory Planning with Mixture of Experts for Autonomous Driving", "abstract": "This paper presents ARTEMIS, an end-to-end autonomous driving framework that combines autoregressive trajectory planning with Mixture-of-Experts (MoE). Traditional modular methods suffer from error propagation, while existing end-to-end models typically employ static one-shot inference paradigms that inadequately capture the dynamic changes of the environment. ARTEMIS takes a different method by generating trajectory waypoints sequentially, preserves critical temporal dependencies while dynamically routing scene-specific queries to specialized expert networks. It effectively relieves trajectory quality degradation issues encountered when guidance information is ambiguous, and overcomes the inherent representational limitations of singular network architectures when processing diverse driving scenarios. Additionally, we use a lightweight batch reallocation strategy that significantly improves the training speed of the Mixture-of-Experts model. Through experiments on the NAVSIM dataset, ARTEMIS exhibits superior competitive performance, achieving 87.0 PDMS and 83.1 EPDMS with ResNet-34 backbone, demonstrates state-of-the-art performance on multiple metrics.", "authors": ["Renju Feng", "Ning Xi", "Duanfeng Chu", "Rukang Wang", "Zejian Deng", "Anzheng Wang", "Liping Lu", "Jinxiang Wang", "Yanjun Huang"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-28", "url": "https://arxiv.org/abs/2504.19580", "pdf_url": "https://arxiv.org/pdf/2504.19580v2", "arxiv_id": "2504.19580", "doi": "10.1109/LRA.2025.3632616", "citation_count": 53, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "IEEE Robotics and Automation Letters", "quality_score": 0.4331}
{"id": "805f66666bb534c562048399987d9009ad2961f4da4a223d59fdccab43d4d003", "sources": ["arxiv", "semantic_scholar"], "title": "BadMoE: Backdooring Mixture-of-Experts LLMs via Optimizing Routing Triggers and Infecting Dormant Experts", "abstract": "Mixture-of-Experts (MoE) have emerged as a powerful architecture for large language models (LLMs), enabling efficient scaling of model capacity while maintaining manageable computational costs. The key advantage lies in their ability to route different tokens to different ``expert'' networks within the model, enabling specialization and efficient handling of diverse input. However, the vulnerabilities of MoE-based LLMs still have barely been studied, and the potential for backdoor attacks in this context remains largely unexplored. This paper presents the first backdoor attack against MoE-based LLMs where the attackers poison ``dormant experts'' (i.e., underutilized experts) and activate them by optimizing routing triggers, thereby gaining control over the model's output. We first rigorously prove the existence of a few ``dominating experts'' in MoE models, whose outputs can determine the overall MoE's output. We also show that dormant experts can serve as dominating experts to manipulate model predictions. Accordingly, our attack, namely BadMoE, exploits the unique architecture of MoE models by 1) identifying dormant experts unrelated to the target task, 2) constructing a routing-aware loss to optimize the activation triggers of these experts, and 3) promoting dormant experts to dominating roles via poisoned training data. Extensive experiments show that BadMoE successfully enforces malicious prediction on attackers' target tasks while preserving overall model utility, making it a more potent and stealthy attack than existing methods.", "authors": ["Qingyue Wang", "Qi Pang", "Xixun Lin", "Shuai Wang", "Daoyuan Wu"], "categories": ["cs.CR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-24", "url": "https://arxiv.org/abs/2504.18598", "pdf_url": "https://arxiv.org/pdf/2504.18598v2", "arxiv_id": "2504.18598", "doi": "10.48550/arXiv.2504.18598", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "af7333c08b363d78eb07663fda9ea06499d0a9b46ce2960016d6dbeeb2e8e10b", "sources": ["arxiv", "semantic_scholar"], "title": "A LoRA-Based Approach to Fine-Tuning LLMs for Educational Guidance in Resource-Constrained Settings", "abstract": "The current study describes a cost-effective method for adapting large language models (LLMs) for academic advising with study-abroad contexts in mind and for application in low-resource methods for acculturation. With the Mistral-7B-Instruct model applied with a Low-Rank Adaptation (LoRA) method and a 4-bit quantization method, the model underwent training in two distinct stages related to this study's purpose to enhance domain specificity while maintaining computational efficiency. In Phase 1, the model was conditioned with a synthetic dataset via the Gemini Pro API, and in Phase 2, it was trained with manually curated datasets from the StudyAbroadGPT project to achieve enhanced, contextualized responses. Technical innovations entailed memory-efficient quantization, parameter-efficient adaptation, and continuous training analytics via Weights & Biases. After training, this study demonstrated a reduction in training loss by 52.7%, 92% accuracy in domain-specific recommendations, achieved 95% markdown-based formatting support, and a median run-rate of 100 samples per second on off-the-shelf GPU equipment. These findings support the effective application of instruction-tuned LLMs within educational advisers, especially in low-resource institutional scenarios. Limitations included decreased generalizability and the application of a synthetically generated dataset, but this framework is scalable for adding new multilingual-augmented and real-time academic advising processes. Future directions may include plans for the integration of retrieval-augmented generation, applying dynamic quantization routines, and connecting to real-time academic databases to increase adaptability and accuracy.", "authors": ["Md Millat Hosen"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-22", "url": "https://arxiv.org/abs/2504.15610", "pdf_url": "https://arxiv.org/pdf/2504.15610v3", "arxiv_id": "2504.15610", "doi": "10.48550/arXiv.2504.15610", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "4e7a9b4c5c70709125d965c424337a2f2c1d04619becb6d0ff702ff09ff580c5", "sources": ["arxiv", "semantic_scholar"], "title": "KeyDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference in Resource-Constrained Environments", "abstract": "We demonstrate that geometrically distinctive keys during LLM inference tend to have high attention scores. Based on the phenomenon we propose KeyDiff, a training-free KV cache eviction method based solely on key similarity. Unlike other KV cache eviction methods, KeyDiff can process arbitrarily long prompts within strict resource constraints and efficiently generate responses. We provide a theoretical basis for KeyDiff by relating key diversity with attention scores. These results imply KeyDiff can efficiently identify the most important tokens to retain. Notably KeyDiff does not rely on attention scores, allowing the use of optimized attention mechanisms like FlashAttention. Under a strict memory allowance, we demonstrate the effectiveness of KeyDiff for the Llama and Qwen model families by observing a performance gap of less than 0.04% with 8K cache budget ($\\sim$23% KV cache reduction) from the non-evicting baseline on LongBench for Llama 3.1-8B and Llama 3.2-3B. We also observe near baseline performance for Deepseek-R1-Distill-Llama-8B on the Math500 reasoning benchmark and decrease end-to-end inference latency by up to 30% compared to the other token-eviction methods.", "authors": ["Junyoung Park", "Dalton Jones", "Matthew J Morse", "Raghavv Goel", "Mingu Lee", "Chris Lott"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-21", "url": "https://arxiv.org/abs/2504.15364", "pdf_url": "https://arxiv.org/pdf/2504.15364v4", "arxiv_id": "2504.15364", "doi": "10.48550/arXiv.2504.15364", "citation_count": 29, "influential_citation_count": 11, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5396}
{"id": "1bfd92a50d515c270f55f3ded7404bbb6f03b7218f31a87f50b96be8a53b3a5e", "sources": ["arxiv", "semantic_scholar"], "title": "Improving the Serving Performance of Multi-LoRA Large Language Models via Efficient LoRA and KV Cache Management", "abstract": "Multiple Low-Rank Adapters (Multi-LoRAs) are gaining popularity for task-specific Large Language Model (LLM) applications. For multi-LoRA serving, caching hot KV caches and LoRA adapters in high bandwidth memory of accelerations can improve inference performance. However, existing Multi-LoRA inference systems fail to optimize serving performance like Time-To-First-Toke (TTFT), neglecting usage dependencies when caching LoRAs and KVs. We therefore propose FASTLIBRA, a Multi-LoRA caching system to optimize the serving performance. FASTLIBRA comprises a dependency-aware cache manager and a performance-driven cache swapper. The cache manager maintains the usage dependencies between LoRAs and KV caches during the inference with a unified caching pool. The cache swapper determines the swap-in or out of LoRAs and KV caches based on a unified cost model, when the HBM is idle or busy, respectively. Experimental results show that ELORA reduces the TTFT by 63.4% on average, compared to state-of-the-art works.", "authors": ["Hang Zhang", "Jiuchen Shi", "Yixiao Wang", "Quan Chen", "Yizhou Shan", "Minyi Guo"], "categories": ["cs.AR", "cs.AI", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-19", "url": "https://arxiv.org/abs/2505.03756", "pdf_url": "https://arxiv.org/pdf/2505.03756v1", "arxiv_id": "2505.03756", "doi": "10.48550/arXiv.2505.03756", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "8dfff456fc1107fee65f2c5f5365937c93a7c3db0e12938c1cb89069025feefa", "sources": ["arxiv", "semantic_scholar"], "title": "CAOTE: KV Cache Selection for LLMs via Attention Output Error-Based Token Eviction", "abstract": "While long context support of large language models has extended their abilities, it also incurs challenges in memory and compute which becomes crucial bottlenecks in resource-restricted devices. Token eviction, a widely adopted post-training methodology designed to alleviate the bottlenecks by evicting less important tokens from the cache, typically uses attention scores as proxy metrics for token importance. However, one major limitation of attention score as a token-wise importance metrics is that it lacks the information about contribution of tokens to the attention output. In this paper, we propose a simple eviction criterion based on the contribution of cached tokens to attention outputs. Our method, CAOTE, optimizes for eviction error due to token eviction, by seamlessly integrating attention scores and value vectors. This is the first method which uses value tokens on top of attention-based eviction scores in closed-form. Additionally, CAOTE can act as a meta-heuristic method with flexible usage with any token eviction method. We show that CAOTE, when combined with the state-of-the-art attention score-based methods, always improves accuracies on the downstream task, indicating the importance of leveraging information from values during token eviction process.", "authors": ["Raghavv Goel", "Junyoung Park", "Mukul Gagrani", "Dalton Jones", "Matthew Morse", "Harper Langston", "Mingu Lee", "Chris Lott"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-18", "url": "https://arxiv.org/abs/2504.14051", "pdf_url": "https://arxiv.org/pdf/2504.14051v6", "arxiv_id": "2504.14051", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "7e450dd5509e16e0bf8ea6f3469f0f4ba5d674de8f9a22ff4ab6523b9d1fb7d4", "sources": ["arxiv", "semantic_scholar"], "title": "Unveiling Hidden Collaboration within Mixture-of-Experts in Large Language Models", "abstract": "Mixture-of-Experts based large language models (MoE LLMs) have shown significant promise in multitask adaptability by dynamically routing inputs to specialized experts. Despite their success, the collaborative mechanisms among experts are still not well understood, limiting both the interpretability and optimization of these models. In this paper, we focus on two critical issues: (1) identifying expert collaboration patterns, and (2) optimizing MoE LLMs through expert pruning. To address the first issue, we propose a hierarchical sparse dictionary learning (HSDL) method that uncovers the collaboration patterns among experts. For the second issue, we introduce the Contribution-Aware Expert Pruning (CAEP) algorithm, which effectively prunes low-contribution experts. Our extensive experiments demonstrate that expert collaboration patterns are closely linked to specific input types and exhibit semantic significance across various tasks. Moreover, pruning experiments show that our approach improves overall performance by 2.5\\% on average, outperforming existing methods. These findings offer valuable insights into enhancing the efficiency and interpretability of MoE LLMs, offering a clearer understanding of expert interactions and improving model optimization.", "authors": ["Yuanbo Tang", "Yan Tang", "Naifan Zhang", "Meixuan Chen", "Yang Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-16", "url": "https://arxiv.org/abs/2504.12359", "pdf_url": "https://arxiv.org/pdf/2504.12359v1", "arxiv_id": "2504.12359", "doi": "10.48550/arXiv.2504.12359", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "448267b868fd41d891bc312a4f5fc457ee243f098359528ec8b6ff0ae3b7da66", "sources": ["arxiv", "semantic_scholar"], "title": "Cost-Efficient LLM Serving in the Cloud: VM Selection with KV Cache Offloading", "abstract": "LLM inference is essential for applications like text summarization, translation, and data analysis, but the high cost of GPU instances from Cloud Service Providers (CSPs) like AWS is a major burden. This paper proposes InferSave, a cost-efficient VM selection framework for cloud based LLM inference. InferSave optimizes KV cache offloading based on Service Level Objectives (SLOs) and workload charac teristics, estimating GPU memory needs, and recommending cost-effective VM instances. Additionally, the Compute Time Calibration Function (CTCF) improves instance selection accuracy by adjusting for discrepancies between theoretical and actual GPU performance. Experiments on AWS GPU instances show that selecting lower-cost instances without KV cache offloading improves cost efficiency by up to 73.7% for online workloads, while KV cache offloading saves up to 20.19% for offline workloads.", "authors": ["Kihyun Kim", "Jinwoo Kim", "Hyunsun Chung", "Myung-Hoon Cha", "Hong-Yeon Kim", "Youngjae Kim"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-16", "url": "https://arxiv.org/abs/2504.11816", "pdf_url": "https://arxiv.org/pdf/2504.11816v1", "arxiv_id": "2504.11816", "doi": "10.48550/arXiv.2504.11816", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1306}
{"id": "f51138ccd4d1567692bca2c17e99ef77bfd3690db032472e77935fb7a4179595", "sources": ["arxiv", "semantic_scholar"], "title": "Activated LoRA: Fine-tuned LLMs for Intrinsics", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a highly efficient framework for finetuning the weights of large foundation models, and has become the go-to method for data-driven customization of LLMs. Despite the promise of highly customized behaviors and capabilities, switching between relevant LoRAs in a multiturn setting is inefficient, as the key-value (KV) cache of the entire turn history must be recomputed with the LoRA weights before generation can begin. To address this problem, we propose Activated LoRA (aLoRA), an adapter architecture which modifies the LoRA framework to only adapt weights for the tokens in the sequence after the aLoRA is invoked. This change crucially allows aLoRA to accept the base model's KV cache of the input string, meaning that aLoRA can be instantly activated whenever needed in a chain without recomputing the prior keys and values. This enables building what we call intrinsics, i.e. specialized models invoked to perform well-defined operations on portions of an input chain or conversation that otherwise uses the base model by default. We train a set of aLoRA-based intrinsics models, demonstrating competitive accuracy with standard LoRA while significantly improving inference efficiency. We contributed our Activated LoRA implementation to the Huggingface PEFT library https://github.com/huggingface/peft.", "authors": ["Kristjan Greenewald", "Luis Lastras", "Thomas Parnell", "Vraj Shah", "Lucian Popa", "Giulio Zizzo", "Chulaka Gunasekara", "Ambrish Rawat", "David Cox"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-16", "url": "https://arxiv.org/abs/2504.12397", "pdf_url": "https://arxiv.org/pdf/2504.12397v5", "arxiv_id": "2504.12397", "doi": "10.48550/arXiv.2504.12397", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/huggingface/peft", "venue": "arXiv.org", "quality_score": 0.2019}
{"id": "e660e328ca40649b9022b636a792b9a7553aa5565cacaa2fa6c3793bbdc2e46b", "sources": ["arxiv", "semantic_scholar"], "title": "Dense Backpropagation Improves Training for Sparse Mixture-of-Experts", "abstract": "Mixture of Experts (MoE) pretraining is more scalable than dense Transformer pretraining, because MoEs learn to route inputs to a sparse set of their feedforward parameters. However, this means that MoEs only receive a sparse backward update, leading to training instability and suboptimal performance. We present a lightweight approximation method that gives the MoE router a dense gradient update while continuing to sparsely activate its parameters. Our method, which we refer to as Default MoE, substitutes missing expert activations with default outputs consisting of an exponential moving average of expert outputs previously seen over the course of training. This allows the router to receive signals from every expert for each token, leading to significant improvements in training performance. Our Default MoE outperforms standard TopK routing in a variety of settings without requiring significant computational overhead. Code: https://github.com/vatsal0/default-moe.", "authors": ["Ashwinee Panda", "Vatsal Baherwani", "Zain Sarwar", "Benjamin Therien", "Sambit Sahu", "Tom Goldstein", "Supriyo Chakraborty"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-16", "url": "https://arxiv.org/abs/2504.12463", "pdf_url": "https://arxiv.org/pdf/2504.12463v3", "arxiv_id": "2504.12463", "doi": "10.48550/arXiv.2504.12463", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/vatsal0/default-moe", "venue": "arXiv.org", "quality_score": 0.2019}
{"id": "f3835918741c9c20edddf9d2a720857006501b65a5fde577938dd5a8df20b2c1", "sources": ["arxiv", "semantic_scholar"], "title": "Shared Disk KV Cache Management for Efficient Multi-Instance Inference in RAG-Powered LLMs", "abstract": "Recent large language models (LLMs) face increasing inference latency as input context length and model size continue to grow. In particular, the retrieval-augmented generation (RAG) technique, which enhances LLM responses by incorporating external knowledge, exacerbates this issue by significantly increasing the number of input tokens. This expansion in token length leads to a substantial rise in computational overhead, particularly during the prefill stage, resulting in prolonged time-to-first-token (TTFT). To address this issue, this paper proposes a method to reduce TTFT by leveraging a disk-based key-value (KV) cache to lessen the computational burden during the prefill stage. We also introduce a disk-based shared KV cache management system, called Shared RAG-DCache, for multi-instance LLM RAG service environments. This system, together with an optimal system configuration, improves both throughput and latency under given resource constraints. Shared RAG-DCache exploits the locality of documents related to user queries in RAG, as well as the queueing delay in LLM inference services. It proactively generates and stores disk KV caches for query-related documents and shares them across multiple LLM instances to enhance inference performance. In experiments on a single host equipped with 2 GPUs and 1 CPU, Shared RAG-DCache achieved a 15~71% increase in throughput and up to a 12~65% reduction in latency, depending on the resource configuration.", "authors": ["Hyungwoo Lee", "Kihyun Kim", "Jinwoo Kim", "Jungmin So", "Myung-Hoon Cha", "Hong-Yeon Kim", "James J. Kim", "Youngjae Kim"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-16", "url": "https://arxiv.org/abs/2504.11765", "pdf_url": "https://arxiv.org/pdf/2504.11765v1", "arxiv_id": "2504.11765", "doi": "10.48550/arXiv.2504.11765", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "f9c7fd1a693c6aef1682ea7db00025c6907a4b3cde8c33727617eb6f96f11ae5", "sources": ["arxiv", "semantic_scholar"], "title": "Super-LoRa: Enhancing LoRa Throughput via Payload Superposition", "abstract": "This paper presents Super-LoRa, a novel approach to enhancing the throughput of LoRa networks by leveraging the inherent robustness of LoRa modulation against interference. By superimposing multiple payload symbols, Super-LoRa significantly increases the data rate while maintaining lower transmitter and receiver complexity. Our solution is evaluated through both simulations and real-world experiments, showing a potential throughput improvement of up to 5x compared to standard LoRa. This advancement positions Super-LoRa as a viable solution for data-intensive IoT applications such as smart cities and precision agriculture, which demand higher data transmission rates.", "authors": ["Salah Abdeljabar", "Mohamed-Slim Alouini"], "categories": ["eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-04-16", "url": "https://arxiv.org/abs/2504.11927", "pdf_url": "https://arxiv.org/pdf/2504.11927v1", "arxiv_id": "2504.11927", "doi": "10.1109/JIOT.2025.3562270", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Internet of Things Journal", "quality_score": 0.2258}
{"id": "28964b37769dedfe044a33e59f0126474b252d9f4767f6c83a356b25d175da27", "sources": ["arxiv", "semantic_scholar"], "title": "Trend Filtered Mixture of Experts for Automated Gating of High-Frequency Flow Cytometry Data", "abstract": "Ocean microbes are critical to both ocean ecosystems and the global climate. Flow cytometry, which measures cell optical properties in fluid samples, is routinely used in oceanographic research. Despite decades of accumulated data, identifying key microbial populations (a process known as ``gating'') remains a significant analytical challenge. To address this, we focus on gating multidimensional, high-frequency flow cytometry data collected {\\it continuously} on board oceanographic research vessels, capturing time- and space-wise variations in the dynamic ocean. Our paper proposes a novel mixture-of-experts model in which both the gating function and the experts are given by trend filtering. The model leverages two key assumptions: (1) Each snapshot of flow cytometry data is a mixture of multivariate Gaussians and (2) the parameters of these Gaussians vary smoothly over time. Our method uses regularization and a constraint to ensure smoothness and that cluster means match biologically distinct microbe types. We demonstrate, using flow cytometry data from the North Pacific Ocean, that our proposed model accurately matches human-annotated gating and corrects significant errors.", "authors": ["Sangwon Hyun", "Tim Coleman", "Francois Ribalet", "Jacob Bien"], "categories": ["stat.ME", "stat.AP", "stat.ML"], "fields_of_study": ["Mathematics"], "published_date": "2025-04-16", "url": "https://arxiv.org/abs/2504.12287", "pdf_url": "https://arxiv.org/pdf/2504.12287v1", "arxiv_id": "2504.12287", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "873062fa8369f6ba8222e6fbe81df3a2fba17cfc50f51123d1c5cca935615b6e", "sources": ["arxiv", "semantic_scholar"], "title": "KeepKV: Achieving Periodic Lossless KV Cache Compression for Efficient LLM Inference", "abstract": "Efficient inference of large language models (LLMs) is hindered by an ever-growing key-value (KV) cache, making KV cache compression a critical research direction. Traditional methods selectively evict less important KV cache entries, which leads to information loss and hallucinations. Recently, merging-based strategies have been explored to retain more information by merging KV pairs that would be discarded; however, these existing approaches inevitably introduce inconsistencies in attention distributions before and after merging, causing degraded generation quality. To overcome this challenge, we propose KeepKV, a novel adaptive KV cache merging method designed to preserve performance under strict memory constraints, achieving single-step lossless compression and providing error bounds for multi-step compression. KeepKV introduces the Electoral Votes mechanism that records merging history and adaptively adjusts attention scores. Moreover, it further leverages a novel Zero Inference-Perturbation Merging method, compensating for attention loss resulting from cache merging. Extensive experiments on various benchmarks and LLM architectures demonstrate that KeepKV substantially reduces memory usage while successfully retaining essential context information, achieving over 2x inference throughput improvement and maintaining superior generation quality even with only 10% KV cache budgets.", "authors": ["Yuxuan Tian", "Zihan Wang", "Yebo Peng", "Aomufei Yuan", "Zhiming Wang", "Bairen Yi", "Xin Liu", "Yong Cui", "Tong Yang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-14", "url": "https://arxiv.org/abs/2504.09936", "pdf_url": "https://arxiv.org/pdf/2504.09936v2", "arxiv_id": "2504.09936", "doi": "10.1609/aaai.v40i39.40611", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.1283}
{"id": "df09c98b49bebf1971fa3d014d40bc29b604d4965d379d84e456b79015127c86", "sources": ["arxiv", "semantic_scholar"], "title": "Head-Aware KV Cache Compression for Efficient Visual Autoregressive Modeling", "abstract": "Visual Autoregressive (VAR) models adopt a next-scale prediction paradigm, offering high-quality content generation with substantially fewer decoding steps. However, existing VAR models suffer from significant attention complexity and severe memory overhead due to the accumulation of key-value (KV) caches across scales. In this paper, we tackle this challenge by introducing KV cache compression into the next-scale generation paradigm. We begin with a crucial observation: attention heads in VAR models can be divided into two functionally distinct categories: Contextual Heads focus on maintaining semantic consistency, while Structural Heads are responsible for preserving spatial coherence. This structural divergence causes existing one-size-fits-all compression methods to perform poorly on VAR models. To address this, we propose HACK, a training-free Head-Aware KV cache Compression frameworK. HACK utilizes an offline classification scheme to separate head types, enabling it to apply pattern-specific compression strategies with asymmetric cache budgets for each category. By doing so, HACK effectively constrains the average KV cache length within a fixed budget $B$, reducing the theoretical attention complexity from $\\mathcal{O}(n^4)$ to $\\mathcal{O}(Bn^2)$. Extensive experiments on multiple VAR models across text-to-image and class-conditional tasks validate the effectiveness and generalizability of HACK. It achieves up to 70% KV cache compression without degrading output quality, resulting in memory savings and faster inference. For example, HACK provides a $1.75\\times$ memory reduction and a $1.57\\times$ speedup on Infinity-8B.", "authors": ["Ziran Qin", "Youru Lv", "Mingbao Lin", "Hang Guo", "Zeren Zhang", "Danping Zou", "Weiyao Lin"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-12", "url": "https://arxiv.org/abs/2504.09261", "pdf_url": "https://arxiv.org/pdf/2504.09261v2", "arxiv_id": "2504.09261", "doi": "10.48550/arXiv.2504.09261", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.294}
{"id": "ab82d3c10e8827a8b9b5c24a9038e9671e73c123faead5571ea05524d1911fc1", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Group Experts for Learning Invariant Representations", "abstract": "Sparsely activated Mixture-of-Experts (MoE) models effectively increase the number of parameters while maintaining consistent computational costs per token. However, vanilla MoE models often suffer from limited diversity and specialization among experts, constraining their performance and scalability, especially as the number of experts increases. In this paper, we present a novel perspective on vanilla MoE with top-$k$ routing inspired by sparse representation. This allows us to bridge established theoretical insights from sparse representation into MoE models. Building on this foundation, we propose a group sparse regularization approach for the input of top-$k$ routing, termed Mixture of Group Experts (MoGE). MoGE indirectly regularizes experts by imposing structural constraints on the routing inputs, while preserving the original MoE architecture. Furthermore, we organize the routing input into a 2D topographic map, spatially grouping neighboring elements. This structure enables MoGE to capture representations invariant to minor transformations, thereby significantly enhancing expert diversity and specialization. Comprehensive evaluations across various Transformer models for image classification and language modeling tasks demonstrate that MoGE substantially outperforms its MoE counterpart, with minimal additional memory and computation overhead. Our approach provides a simple yet effective solution to scale the number of experts and reduce redundancy among them. The source code is included in the supplementary material and will be publicly released.", "authors": ["Lei Kang", "Jia Li", "Mi Tian", "Hua Huang"], "categories": ["cs.LG", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-12", "url": "https://arxiv.org/abs/2504.09265", "pdf_url": "https://arxiv.org/pdf/2504.09265v2", "arxiv_id": "2504.09265", "doi": "10.1145/3743093.3770949", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ACM Multimedia Asia", "quality_score": 0.126}
{"id": "8448b9bd64aaff5044046948356554a7c625bad9ef6faafa12af22d3d2f0eb09", "sources": ["arxiv", "semantic_scholar"], "title": "PQS (Prune, Quantize, and Sort): Low-Bitwidth Accumulation of Dot Products in Neural Network Computations", "abstract": "We present PQS, which uses three techniques together - Prune, Quantize, and Sort - to achieve low-bitwidth accumulation of dot products in neural network computations. In conventional quantized (e.g., 8-bit) dot products, partial results are accumulated into wide (e.g., 32-bit) accumulators to avoid overflows when accumulating intermediate partial sums. However, such wide accumulators increase memory bandwidth usage and reduce energy efficiency. We show that iterative N:M pruning in floating point followed by quantization to 8 (or fewer) bits, and accumulation of partial products in a sorted order (\"small to large\") allows for accurate, compressed models with short dot product lengths that do not require wide accumulators. We design, analyze, and implement the PQS algorithm to eliminate accumulation overflows at inference time for several neural networks. Our method offers a 2.5x reduction in accumulator bitwidth while achieving model accuracy on par with floating-point baselines for multiple image classification tasks.", "authors": ["Vikas Natesh", "H. T. Kung"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-12", "url": "https://arxiv.org/abs/2504.09064", "pdf_url": "https://arxiv.org/pdf/2504.09064v1", "arxiv_id": "2504.09064", "doi": "10.48550/arXiv.2504.09064", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.126}
{"id": "cf673f5aadf9c9cfa4c2c6e75737abbc32c624731a3ace17f85ef08896511d4c", "sources": ["arxiv", "semantic_scholar"], "title": "UniCAIM: A Unified CAM/CIM Architecture with Static-Dynamic KV Cache Pruning for Efficient Long-Context LLM Inference", "abstract": "Transformer-based large language models (LLMs) have achieved impressive performance in various natural language processing (NLP) applications. However, the high memory and computation cost induced by the KV cache limits the inference efficiency, especially for long input sequences. Compute-in-memory (CIM)-based accelerators have been proposed for LLM acceleration with KV cache pruning. However, as existing accelerators only support static pruning with a fixed pattern or dynamic pruning with primitive implementations, they suffer from either high accuracy degradation or low efficiency. In this paper, we propose a ferroelectric FET (FeFET)-based unified content addressable memory (CAM) and CIM architecture, dubbed as UniCAIM. UniCAIM features simultaneous support for static and dynamic pruning with 3 computation modes: 1) in the CAM mode, UniCAIM enables approximate similarity measurement in O(1) time for dynamic KV cache pruning with high energy efficiency; 2) in the charge-domain CIM mode, static pruning can be supported based on accumulative similarity score, which is much more flexible compared to fixed patterns; 3) in the current-domain mode, exact attention computation can be conducted with a subset of selected KV cache. We further propose a novel CAM/CIM cell design that leverages the multi-level characteristics of FeFETs for signed multibit storage of the KV cache and in-place attention computation. With extensive experimental results, we demonstrate UniCAIM can reduce the area-energy-delay product (AEDP) by 8.2-831x over the state-ofthe-art CIM-based LLM accelerators at the circuit level, along with high accuracy comparable with dense attention at the application level, showing its great potential for efficient long-context LLM inference.", "authors": ["Weikai Xu", "Wenxuan Zeng", "Qianqian Huang", "Meng Li", "Ru Huang"], "categories": ["cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-10", "url": "https://arxiv.org/abs/2504.07479", "pdf_url": "https://arxiv.org/pdf/2504.07479v1", "arxiv_id": "2504.07479", "doi": "10.1109/DAC63849.2025.11133273", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Design Automation Conference", "quality_score": 0.1505}
{"id": "f6204ae197813662598fe2d02c496875416ff76d8e66d714a9310daab82d078b", "sources": ["arxiv", "semantic_scholar"], "title": "Cluster-Driven Expert Pruning for Mixture-of-Experts Large Language Models", "abstract": "Mixture-of-Experts (MoE) architectures have emerged as a promising paradigm for scaling large language models (LLMs) with sparse activation of task-specific experts. Despite their computational efficiency during inference, the massive overall parameter footprint of MoE models (e.g., GPT-4) introduces critical challenges for practical deployment. Current pruning approaches often fail to address two inherent characteristics of MoE systems: 1).intra-layer expert homogeneity where experts within the same MoE layer exhibit functional redundancy, and 2). inter-layer similarity patterns where deeper layers tend to contain progressively more homogeneous experts. To tackle these issues, we propose Cluster-driven Expert Pruning (C-Prune), a novel two-stage framework for adaptive task-specific compression of MoE LLMs. C-Prune operates through layer-wise expert clustering, which groups functionally similar experts within each MoE layer using parameter similarity metrics, followed by global cluster pruning, which eliminates redundant clusters across all layers through a unified importance scoring mechanism that accounts for cross-layer homogeneity. We validate C-Prune through extensive experiments on multiple MoE models and benchmarks. The results demonstrate that C-Prune effectively reduces model size while outperforming existing MoE pruning methods.", "authors": ["Hongcheng Guo", "Juntao Yao", "Boyang Wang", "Junjia Du", "Shaosheng Cao", "Donglin Di", "Shun Zhang", "Zhoujun Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-10", "url": "https://arxiv.org/abs/2504.07807", "pdf_url": "https://arxiv.org/pdf/2504.07807v1", "arxiv_id": "2504.07807", "doi": "10.48550/arXiv.2504.07807", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "d749b0792a9e62db098cd176174df4d31ad41b09daf8de63389991a6c9b5b081", "sources": ["arxiv", "semantic_scholar"], "title": "ms-Mamba: Multi-scale Mamba for Time-Series Forecasting", "abstract": "The problem of Time-series Forecasting is generally addressed by recurrent, Transformer-based and the recently proposed Mamba-based architectures. However, existing architectures generally process their input at a single temporal scale, which may be sub-optimal for many tasks where information changes over multiple time scales. In this paper, we introduce a novel architecture called Multi-scale Mamba (ms-Mamba) to address this gap. ms-Mamba incorporates multiple temporal scales by using multiple Mamba blocks with different sampling rates ($Δ$s). Our experiments on many benchmarks demonstrate that ms-Mamba outperforms state-of-the-art approaches, including the recently proposed Transformer-based and Mamba-based models. For example, on the Solar-Energy dataset, ms-Mamba outperforms its closest competitor S-Mamba (0.229 vs. 0.240 in terms of mean-squared error) while using fewer parameters (3.53M vs. 4.77M), less memory (13.46MB vs. 18.18MB), and less operations (14.93G vs. 20.53G MACs), averaged across four forecast lengths. Codes and models will be made available.", "authors": ["Yusuf Meric Karadag", "Ismail Talaz", "Ipek Gursel Dino", "Sinan Kalkan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-10", "url": "https://arxiv.org/abs/2504.07654", "pdf_url": "https://arxiv.org/pdf/2504.07654v2", "arxiv_id": "2504.07654", "doi": "10.48550/arXiv.2504.07654", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neurocomputing", "quality_score": 0.2113}
{"id": "8dc4fec18aa780fb5ecacd14dac8491c9df35f9931fa05567f580a55d184c1d0", "sources": ["arxiv", "semantic_scholar"], "title": "Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations", "abstract": "Mixture-of-Experts (MoE) models achieve a favorable trade-off between performance and inference efficiency by activating only a subset of experts. However, the memory overhead of storing all experts remains a major limitation, especially in large-scale MoE models such as DeepSeek-R1(671B). In this study, we investigate domain specialization and expert redundancy in large-scale MoE models and uncover a consistent behavior we term few-shot expert localization, with only a few in-domain demonstrations, the model consistently activates a sparse and stable subset of experts on tasks within the same domain. Building on this observation, we propose a simple yet effective pruning framework, EASY-EP, that leverages a few domain-specific demonstrations to identify and retain only the most relevant experts. EASY-EP comprises two key components: output-aware expert importance assessment and expert-level token contribution estimation. The former evaluates the importance of each expert for the current token by considering the gating scores and L2 norm of the outputs of activated experts, while the latter assesses the contribution of tokens based on representation similarities before and after routed experts. Experiments on DeepSeek-R1 and DeepSeek-V3-0324 show that our method can achieve comparable performances and $2.99\\times$ throughput under the same memory budget with full model with only half the experts.", "authors": ["Zican Dong", "Han Peng", "Peiyu Liu", "Wayne Xin Zhao", "Dong Wu", "Feng Xiao", "Zhifeng Wang"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-09", "url": "https://arxiv.org/abs/2504.06792", "pdf_url": "https://arxiv.org/pdf/2504.06792v2", "arxiv_id": "2504.06792", "doi": "10.48550/arXiv.2504.06792", "citation_count": 12, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "84a0f2c25ed5512820364521717c9a2056c5014877130204b09ea1762e5799ff", "sources": ["arxiv", "semantic_scholar"], "title": "S'MoRE: Structural Mixture of Residual Experts for Parameter-Efficient LLM Fine-tuning", "abstract": "Fine-tuning pre-trained large language models (LLMs) presents a dual challenge of balancing parameter efficiency and model capacity. Existing methods like low-rank adaptations (LoRA) are efficient but lack flexibility, while Mixture-of-Experts (MoE) enhance model capacity at the cost of more & under-utilized parameters. To address these limitations, we propose Structural Mixture of Residual Experts (S'MoRE), a novel framework that seamlessly integrates the efficiency of LoRA with the flexibility of MoE. Conceptually, S'MoRE employs hierarchical low-rank decomposition of expert weights, yielding residuals of varying orders interconnected in a multi-layer structure. By routing input tokens through sub-trees of residuals, S'MoRE emulates the capacity of numerous experts by instantiating and assembling just a few low-rank matrices. We craft the inter-layer propagation of S'MoRE's residuals as a special type of Graph Neural Network (GNN), and prove that under similar parameter budget, S'MoRE improves structural flexibility of traditional MoE (or Mixture-of-LoRA) by exponential order. Comprehensive theoretical analysis and empirical results demonstrate that S'MoRE achieves superior fine-tuning performance, offering a transformative approach for efficient LLM adaptation. Our implementation is available at: https://github.com/ZimpleX/SMoRE-LLM.", "authors": ["Hanqing Zeng", "Yinglong Xia", "Zhuokai Zhao", "Chuan Jiang", "Qiang Zhang", "Jiayi Liu", "Qunshu Zhang", "Lizhu Zhang", "Xiangjun Fan", "Benyu Zhang"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-08", "url": "https://arxiv.org/abs/2504.06426", "pdf_url": "https://arxiv.org/pdf/2504.06426v2", "arxiv_id": "2504.06426", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ZimpleX/SMoRE-LLM", "venue": null, "quality_score": 0.1435}
{"id": "75d12bd872db1b070a161dde0c3dfa249d16a923c95b7f1192e21e91b983d557", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerating LLM Inference Throughput via Asynchronous KV Cache Prefetching", "abstract": "Large Language Models (LLMs) exhibit pronounced memory-bound characteristics during inference due to High Bandwidth Memory (HBM) bandwidth constraints. In this paper, we propose an L2 Cache-oriented asynchronous KV Cache prefetching method to break through the memory bandwidth bottleneck in LLM inference through computation-load overlap. By strategically scheduling idle memory bandwidth during active computation windows, our method proactively prefetches required KV Cache into GPU L2 cache, enabling high-speed L2 cache hits for subsequent accesses and effectively hiding HBM access latency within computational cycles. Extensive experiments on NVIDIA H20 GPUs demonstrate that the proposed method achieves 2.15x improvement in attention kernel efficiency and up to 1.97x end-to-end throughput enhancement, surpassing state-of-the-art baseline FlashAttention-3. Notably, our solution maintains orthogonality to existing optimization techniques and can be integrated with current inference frameworks, providing a scalable latency-hiding solution for next-generation LLM inference engines.", "authors": ["Yanhao Dong", "Yubo Miao", "Weinan Li", "Xiao Zheng", "Chao Wang", "Jiesheng Wu", "Feng Lyu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-08", "url": "https://arxiv.org/abs/2504.06319", "pdf_url": "https://arxiv.org/pdf/2504.06319v2", "arxiv_id": "2504.06319", "doi": "10.48550/arXiv.2504.06319", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.1945}
{"id": "06b8b4179b9315c59552d1452830e61053d5705866aeb28678cf360d6c47bd70", "sources": ["arxiv", "semantic_scholar"], "title": "CDM-QTA: Quantized Training Acceleration for Efficient LoRA Fine-Tuning of Diffusion Model", "abstract": "Fine-tuning large diffusion models for custom applications demands substantial power and time, which poses significant challenges for efficient implementation on mobile devices. In this paper, we develop a novel training accelerator specifically for Low-Rank Adaptation (LoRA) of diffusion models, aiming to streamline the process and reduce computational complexity. By leveraging a fully quantized training scheme for LoRA fine-tuning, we achieve substantial reductions in memory usage and power consumption while maintaining high model fidelity. The proposed accelerator features flexible dataflow, enabling high utilization for irregular and variable tensor shapes during the LoRA process. Experimental results show up to 1.81x training speedup and 5.50x energy efficiency improvements compared to the baseline, with minimal impact on image generation quality.", "authors": ["Jinming Lu", "Minghao She", "Wendong Mao", "Zhongfeng Wang"], "categories": ["cs.GR", "cs.AI", "cs.AR", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-08", "url": "https://arxiv.org/abs/2504.07998", "pdf_url": "https://arxiv.org/pdf/2504.07998v1", "arxiv_id": "2504.07998", "doi": "10.1109/ISCAS56072.2025.11043305", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Symposium on Circuits and Systems", "quality_score": 0.1215}
{"id": "efa79dc7ec7185a59de79d252e6f192a4b5dc6344fc0c207d23ec6ef46cb1b3a", "sources": ["arxiv", "semantic_scholar"], "title": "Finding Fantastic Experts in MoEs: A Unified Study for Expert Dropping Strategies and Observations", "abstract": "Sparsely activated Mixture-of-Experts (SMoE) has shown promise in scaling up the learning capacity of neural networks. However, vanilla SMoEs have issues such as expert redundancy and heavy memory requirements, making them inefficient and non-scalable, especially for resource-constrained scenarios. Expert-level sparsification of SMoEs involves pruning the least important experts to address these limitations. In this work, we aim to address three questions: (1) What is the best recipe to identify the least knowledgeable subset of experts that can be dropped with minimal impact on performance? (2) How should we perform expert dropping (one-shot or iterative), and what correction measures can we undertake to minimize its drastic impact on SMoE subnetwork capabilities? (3) What capabilities of full-SMoEs are severely impacted by the removal of the least dominant experts, and how can we recover them? Firstly, we propose MoE Experts Compression Suite (MC-Suite), which is a collection of some previously explored and multiple novel recipes to provide a comprehensive benchmark for estimating expert importance from diverse perspectives, as well as unveil numerous valuable insights for SMoE experts. Secondly, unlike prior works with a one-shot expert pruning approach, we explore the benefits of iterative pruning with the re-estimation of the MC-Suite criterion. Moreover, we introduce the benefits of task-agnostic fine-tuning as a correction mechanism during iterative expert dropping, which we term MoE Lottery Subnetworks. Lastly, we present an experimentally validated conjecture that, during expert dropping, SMoEs' instruction-following capabilities are predominantly hurt, which can be restored to a robust level subject to external augmentation of instruction-following capabilities using k-shot examples and supervised fine-tuning.", "authors": ["Ajay Jaiswal", "Jianyu Wang", "Yixiao Li", "Pingzhi Li", "Tianlong Chen", "Zhangyang Wang", "Chong Wang", "Ruoming Pang", "Xianzhi Du"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-08", "url": "https://arxiv.org/abs/2504.05586", "pdf_url": "https://arxiv.org/pdf/2504.05586v2", "arxiv_id": "2504.05586", "doi": "10.48550/arXiv.2504.05586", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "56cf77ee61063d7e4528728a8d79f25f2c0304e886182488b4550fa969322c9c", "sources": ["arxiv", "semantic_scholar"], "title": "LagKV: Lag-Relative Information of the KV Cache Tells Which Tokens Are Important", "abstract": "The increasing size of the Key-Value (KV) cache during the Large Language Models long-context inference is the main obstacle for its balance between the deployment cost and task accuracy. To reduce the KV cache size in such scenarios, most previous efforts leveraged on the attention weight to evict non-critical cache tokens. But there is a trade-off in those methods, they usually require major modification of the inference infrastructure and significant computation overhead. Based on the fact that the Large Language models are autoregressive models, we propose LagKV, a KV compression strategy only relying on straight forward comparison among KV themselves. It is a totally attention free method which offers easy integration to the main stream inference platform and comparable performance comparing to other complicated KV compression methods. Results on RULER benchmark show that, our approach outperforms SnapKV and StreamingLLM in different compression ratios. Especially in the 64-digit passkey retrieval task, our method outperforms the attention weight based method $H_2O$ over $50\\%$ with same compression ratios. Our code is available at https://github.com/AI-Lab-China-Merchants-Bank/LagKV.", "authors": ["Manlai Liang", "JiaMing Zhang", "Xiong Li", "Jinlong Li"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-07", "url": "https://arxiv.org/abs/2504.04704", "pdf_url": "https://arxiv.org/pdf/2504.04704v2", "arxiv_id": "2504.04704", "doi": "10.48550/arXiv.2504.04704", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AI-Lab-China-Merchants-Bank/LagKV", "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "10cc706a711241965430ace176de0a5857c3f04c43e4ec08ed0a617117312e0a", "sources": ["arxiv", "semantic_scholar"], "title": "LEO-MINI: An Efficient Multimodal Large Language Model using Conditional Token Reduction and Mixture of Multi-Modal Experts", "abstract": "Redundancy of visual tokens in multi-modal large language models (MLLMs) significantly reduces their computational efficiency. Recent approaches, such as resamplers and summarizers, have sought to reduce the number of visual tokens, but at the cost of visual reasoning ability. To address this, we propose LEO-MINI, a novel MLLM that significantly reduces the number of visual tokens and simultaneously boosts visual reasoning capabilities. For efficiency, LEO-MINI incorporates CoTR, a novel token reduction module to consolidate a large number of visual tokens into a smaller set of tokens, using the similarity between visual tokens, text tokens, and a compact learnable query. For effectiveness, to scale up the model's ability with minimal computational overhead, LEO-MINI employs MMoE, a novel mixture of multi-modal experts module. MMOE employs a set of LoRA experts with a novel router to switch between them based on the input text and visual tokens instead of only using the input hidden state. MMoE also includes a general LoRA expert that is always activated to learn general knowledge for LLM reasoning. For extracting richer visual features, MMOE employs a set of vision experts trained on diverse domain-specific data. To demonstrate LEO-MINI's improved efficiency and performance, we evaluate it against existing efficient MLLMs on various benchmark vision-language tasks.", "authors": ["Yimu Wang", "Mozhgan Nasr Azadani", "Sean Sedwards", "Krzysztof Czarnecki"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-07", "url": "https://arxiv.org/abs/2504.04653", "pdf_url": "https://arxiv.org/pdf/2504.04653v2", "arxiv_id": "2504.04653", "doi": "10.48550/arXiv.2504.04653", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.1505}
{"id": "0c48f0f8c149a1692d521150eaa54396d9b887c7ac76bf6e072d1f79bd75f4fe", "sources": ["arxiv", "semantic_scholar"], "title": "Saliency-driven Dynamic Token Pruning for Large Language Models", "abstract": "Despite the recent success of large language models (LLMs), LLMs are particularly challenging in long-sequence inference scenarios due to the quadratic computational complexity of the attention mechanism. Inspired by the interpretability theory of feature attribution in neural network models, we observe that not all tokens have the same contribution. Based on this observation, we propose a novel token pruning framework, namely Saliency-driven Dynamic Token Pruning (SDTP), to gradually and dynamically prune redundant tokens based on the input context. Specifically, a lightweight saliency-driven prediction module is designed to estimate the importance score of each token with its hidden state, which is added to different layers of the LLM to hierarchically prune redundant tokens. Furthermore, a ranking-based optimization strategy is proposed to minimize the ranking divergence of the saliency score and the predicted importance score. Extensive experiments have shown that our framework is generalizable to various models and datasets. By hierarchically pruning 65\\% of the input tokens, our method greatly reduces 33\\% $\\sim$ 47\\% FLOPs and achieves speedup up to 1.75$\\times$ during inference, while maintaining comparable performance. We further demonstrate that SDTP can be combined with KV cache compression method for further compression.", "authors": ["Yao Tao", "Yehui Tang", "Yun Wang", "Mingjian Zhu", "Hailin Hu", "Yunhe Wang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-06", "url": "https://arxiv.org/abs/2504.04514", "pdf_url": "https://arxiv.org/pdf/2504.04514v2", "arxiv_id": "2504.04514", "doi": "10.48550/arXiv.2504.04514", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "db8aa52022ef5e0981163c9fb5d72b77cb70b591bc70ae66e24d2bee635177bc", "sources": ["arxiv", "semantic_scholar"], "title": "HeterMoE: Efficient Training of Mixture-of-Experts Models on Heterogeneous GPUs", "abstract": "The Mixture-of-Experts (MoE) architecture has become increasingly popular as a method to scale up large language models (LLMs). To save costs, heterogeneity-aware training solutions have been proposed to utilize GPU clusters made up of both newer and older-generation GPUs. However, existing solutions are agnostic to the performance characteristics of different MoE model components (i.e., attention and expert) and do not fully utilize each GPU's compute capability. In this paper, we introduce HeterMoE, a system to efficiently train MoE models on heterogeneous GPUs. Our key insight is that newer GPUs significantly outperform older generations on attention due to architectural advancements, while older GPUs are still relatively efficient for experts. HeterMoE disaggregates attention and expert computation, where older GPUs are only assigned with expert modules. Through the proposed zebra parallelism, HeterMoE overlaps the computation on different GPUs, in addition to employing an asymmetric expert assignment strategy for fine-grained load balancing to minimize GPU idle time. Our evaluation shows that HeterMoE achieves up to 2.3x speed-up compared to existing MoE training systems, and 1.4x compared to an optimally balanced heterogeneity-aware solution. HeterMoE efficiently utilizes older GPUs by maintaining 95% training throughput on average, even with half of the GPUs in a homogeneous A40 cluster replaced with V100.", "authors": ["Yongji Wu", "Xueshen Liu", "Shuowei Jin", "Ceyu Xu", "Feng Qian", "Z. Morley Mao", "Matthew Lentz", "Danyang Zhuo", "Ion Stoica"], "categories": ["cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-04", "url": "https://arxiv.org/abs/2504.03871", "pdf_url": "https://arxiv.org/pdf/2504.03871v1", "arxiv_id": "2504.03871", "doi": "10.48550/arXiv.2504.03871", "citation_count": 8, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "cacbd2621f3b847bb2426e8ad5df2052919f1dffcd93d8aa7fcb54a4682dd10b", "sources": ["arxiv", "semantic_scholar"], "title": "HyperRAG: Enhancing Quality-Efficiency Tradeoffs in Retrieval-Augmented Generation with Reranker KV-Cache Reuse", "abstract": "Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for enhancing the performance of large language models (LLMs) by integrating external knowledge into the generation process. A key component of RAG pipelines is the reranker, which selects the most relevant documents from a pool of retrieved candidates and significantly improves the quality of the generated responses. While rerankers refine the selection of retrieved documents in RAG pipelines, they introduce computational challenges that hinder high throughput and low latency. To address this problem, we propose HyperRAG, a system that optimizes the trade-off between quality and efficiency in RAG pipelines by leveraging KV-cache reuse for efficient reranker inference. By reusing document-side KV-cache, HyperRAG achieves both high-quality generation and system-level efficiency. To fully realize the benefits of KV-cache reuse, HyperRAG incorporates a range of system-level optimizations designed to enhance efficiency and scalability. Experiments show that HyperRAG achieves a 2 - 3 throughput improvement with decoder-only rerankers while also delivering higher downstream performance compared with traditional RAG service.", "authors": ["Yuwei An", "Yihua Cheng", "Seo Jin Park", "Junchen Jiang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-03", "url": "https://arxiv.org/abs/2504.02921", "pdf_url": "https://arxiv.org/pdf/2504.02921v1", "arxiv_id": "2504.02921", "doi": "10.48550/arXiv.2504.02921", "citation_count": 11, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "1d3ab51b2afe4dd5f4bae01d817b6c5f54c360ed657fad002ca53600c21236b9", "sources": ["arxiv", "semantic_scholar"], "title": "FlowKV: A Disaggregated Inference Framework with Low-Latency KV Cache Transfer and Load-Aware Scheduling", "abstract": "Disaggregated inference has become an essential framework that separates the prefill (P) and decode (D) stages in large language model inference to improve throughput. However, the KV cache transfer faces significant delays between prefill and decode nodes. The block-wise calling method and discontinuous KV cache memory allocation increase the number of calls to the transmission kernel. Additionally, existing frameworks often fix the roles of P and D nodes, leading to computational imbalances. In this paper, we propose FlowKV, a novel disaggregated inference framework, which reduces the average transmission latency of KV cache by 96%, from 0.944s to 0.053s, almost eliminating the transfer time relative to the total request latency by optimizing the KV cache transfer. FlowKV introduces the Load-Aware Scheduler for balanced request scheduling and flexible PD node allocation. This design maximizes hardware resource utilization, achieving peak system throughput across various scenarios, including normal, computational imbalance, and extreme overload conditions. Experimental results demonstrate that FlowKV significantly accelerates inference by 15.2%-48.9% on LongBench dataset compared to the baseline and supports applications with heterogeneous GPUs.", "authors": ["Weiqing Li", "Guochao Jiang", "Xiangyong Ding", "Zhangcheng Tao", "Chuzhan Hao", "Chenfeng Xu", "Yuewei Zhang", "Hao Wang"], "categories": ["cs.DC", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-03", "url": "https://arxiv.org/abs/2504.03775", "pdf_url": "https://arxiv.org/pdf/2504.03775v1", "arxiv_id": "2504.03775", "doi": "10.48550/arXiv.2504.03775", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "47af5df2fc82f91a4763f6998802caaca63bb8f3a65e4f487638ab855ff153d1", "sources": ["arxiv", "semantic_scholar"], "title": "MiLo: Efficient Quantized MoE Inference with Mixture of Low-Rank Compensators", "abstract": "A critical approach for efficiently deploying Mixture-of-Experts (MoE) models with massive parameters is quantization. However, state-of-the-art MoE models suffer from non-negligible accuracy loss with extreme quantization, such as under 4 bits. To address this, we introduce MiLo, a novel method that augments highly quantized MoEs with a mixture of low-rank compensators. These compensators consume only a small amount of additional memory but significantly recover accuracy loss from extreme quantization. MiLo also identifies that MoEmodels exhibit distinctive characteristics across weights due to their hybrid dense-sparse architectures, and employs adaptive rank selection policies along with iterative optimizations to close the accuracy gap. MiLo does not rely on calibration data, allowing it to generalize to different MoE models and datasets without overfitting to a calibration set. To avoid the hardware inefficiencies of extreme quantization, such as 3-bit, MiLo develops Tensor Core-friendly 3-bit kernels, enabling measured latency speedups on 3-bit quantized MoE models. Our evaluation shows that MiLo outperforms existing methods on SoTA MoE models across various tasks.", "authors": ["Beichen Huang", "Yueming Yuan", "Zelei Shao", "Minjia Zhang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-03", "url": "https://arxiv.org/abs/2504.02658", "pdf_url": "https://arxiv.org/pdf/2504.02658v2", "arxiv_id": "2504.02658", "doi": "10.48550/arXiv.2504.02658", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Machine Learning and Systems", "quality_score": 0.1945}
{"id": "a16b5e4f1d4c16d1b66c4567d5b3708866b47d888149e464c301656aad6d3338", "sources": ["arxiv", "semantic_scholar"], "title": "Token Pruning in Audio Transformers: Optimizing Performance and Decoding Patch Importance", "abstract": "Vision Transformers (ViTs) have achieved state-of-the-art performance across various computer vision tasks, but their high computational cost remains a challenge. Token pruning has been proposed to reduce this cost by selectively removing less important tokens. While effective in vision tasks by discarding non-object regions, applying this technique to audio tasks presents unique challenges, as distinguishing relevant from irrelevant regions in time-frequency representations is less straightforward. In this study, for the first time, we applied token pruning to ViT-based audio classification models using Mel-spectrograms and analyzed the trade-offs between model performance and computational cost: TopK token pruning can reduce MAC operations of AudioMAE and AST by 30-40%, with less than a 1% drop in accuracy. Our analysis reveals that while high-intensity or high-variation tokens contribute significantly to model accuracy, low-intensity or low variation tokens also remain important when token pruning is applied; pruning solely based on the intensity or variation of signals in a patch leads to a noticeable drop in accuracy. We support our claim by measuring high correlation between attention scores and these statistical features and by showing retained tokens consistently receive distinct attention compared to pruned ones. We also show that AudioMAE retains more low-intensity tokens than AST. This can be explained by AudioMAE's self-supervised reconstruction objective, which encourages attention to all patches, whereas AST's supervised training focuses on label-relevant tokens.", "authors": ["Taehan Lee", "Hyukjun Lee"], "categories": ["cs.SD", "cs.AI", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-04-02", "url": "https://arxiv.org/abs/2504.01690", "pdf_url": "https://arxiv.org/pdf/2504.01690v2", "arxiv_id": "2504.01690", "doi": "10.3233/FAIA251136", "citation_count": 10, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/andylee-24/token-pruning-audio-transformer", "venue": "European Conference on Artificial Intelligence", "quality_score": 0.2603}
{"id": "932e8837bb89c358a42dbc7541c9df4f21d0879a4399999c4c32f9b1b6fc06ff", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Mamba: Time Series Modeling with Adaptive Pooling Acceleration and Receptive Field Enhancements", "abstract": "\"This work has been submitted to the lEEE for possible publication. Copyright may be transferred without noticeafter which this version may no longer be accessible.\" Time series modeling serves as the cornerstone of real-world applications, such as weather forecasting and transportation management. Recently, Mamba has become a promising model that combines near-linear computational complexity with high prediction accuracy in time series modeling, while facing challenges such as insufficient modeling of nonlinear dependencies in attention and restricted receptive fields caused by convolutions. To overcome these limitations, this paper introduces an innovative framework, Attention Mamba, featuring a novel Adaptive Pooling block that accelerates attention computation and incorporates global information, effectively overcoming the constraints of limited receptive fields. Furthermore, Attention Mamba integrates a bidirectional Mamba block, efficiently capturing long-short features and transforming inputs into the Value representations for attention mechanisms. Extensive experiments conducted on diverse datasets underscore the effectiveness of Attention Mamba in extracting nonlinear dependencies and enhancing receptive fields, establishing superior performance among leading counterparts. Our codes will be available on GitHub.", "authors": ["Sijie Xiong", "Shuqing Liu", "Cheng Tang", "Fumiya Okubo", "Haoling Xiong", "Atsushi Shimada"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-02", "url": "https://arxiv.org/abs/2504.02013", "pdf_url": "https://arxiv.org/pdf/2504.02013v1", "arxiv_id": "2504.02013", "doi": "10.1109/SMC58881.2025.11343445", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Systems, Man and Cybernetics", "quality_score": 0.1146}
{"id": "5d16f716be0b56bea9ff1e5b70644679e981a6d6d5079943a0840727f6337b35", "sources": ["arxiv", "semantic_scholar"], "title": "DynMoLE: Boosting Mixture of LoRA Experts Fine-Tuning with a Hybrid Routing Mechanism", "abstract": "Instruction-based fine-tuning of large language models (LLMs) has achieved remarkable success in various natural language processing (NLP) tasks. Parameter-efficient fine-tuning (PEFT) methods, such as Mixture of LoRA Experts (MoLE), combine the efficiency of Low-Rank Adaptation (LoRA) with the versatility of Mixture of Experts (MoE) models, demonstrating significant potential for handling multiple downstream tasks. However, the existing routing mechanisms for MoLE often involve a trade-off between computational efficiency and predictive accuracy, and they fail to fully address the diverse expert selection demands across different transformer layers. In this work, we propose DynMoLE, a hybrid routing strategy that dynamically adjusts expert selection based on the Tsallis entropy of the router's probability distribution. This approach mitigates router uncertainty, enhances stability, and promotes more equitable expert participation, leading to faster convergence and improved model performance. Additionally, we introduce an auxiliary loss based on Tsallis entropy to further guide the model toward convergence with reduced uncertainty, thereby improving training stability and performance. Our extensive experiments on commonsense reasoning benchmarks demonstrate that DynMoLE achieves substantial performance improvements, outperforming LoRA by 9.6% and surpassing the state-of-the-art MoLE method, MoLA, by 2.3%. We also conduct a comprehensive ablation study to evaluate the contributions of DynMoLE's key components.", "authors": ["Dengchun Li", "Naizheng Wang", "Zihao Zhang", "Haoyang Yin", "Lei Duan", "Meng Xiao", "Mingjie Tang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-01", "url": "https://arxiv.org/abs/2504.00661", "pdf_url": "https://arxiv.org/pdf/2504.00661v1", "arxiv_id": "2504.00661", "doi": "10.48550/arXiv.2504.00661", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "6f20bd5ec37cc7ed87f700ee8ddcd454b1cbc713638cfbb02416e0939b039d41", "sources": ["arxiv", "semantic_scholar"], "title": "SentenceKV: Efficient LLM Inference via Sentence-Level Semantic KV Caching", "abstract": "Large language models face significant computational and memory challenges when processing long contexts. During inference, efficient management of the key-value (KV) cache, which stores intermediate activations for autoregressive generation, is critical to reducing memory overhead and improving computational efficiency. Traditional token-level efficient KV caching methods overlook semantic information, treating tokens independently without considering their semantic relationships. Meanwhile, existing semantic-preserving KV cache management approaches often suffer from substantial memory usage and high time-to-first-token. To address these limitations, we propose SentenceKV, a novel sentence-level semantic KV caching approach designed to enhance inference efficiency while preserving semantic coherence. During prefilling, SentenceKV groups tokens based on sentence-level semantic similarity, compressing sentence representations into concise semantic vectors stored directly on the GPU, while individual KV pairs are offloaded to CPU. During decoding, SentenceKV generates tokens by selectively retrieving semantically relevant sentence-level KV entries, leveraging the semantic similarity between the prefilling-stage semantic vectors and decoding-stage queries. This ensures efficient and contextually accurate predictions, minimizing the loading of redundant or irrelevant data into GPU memory and significantly reducing memory overhead while maintaining stable inference latency, even for extremely long contexts. Extensive evaluations on benchmarks including PG-19, LongBench, and Needle-In-A-Haystack demonstrate that SentenceKV significantly outperforms state-of-the-art methods in both efficiency and memory usage, without compromising model accuracy.", "authors": ["Yuxuan Zhu", "Ali Falahati", "David H. Yang", "Mohammad Mohammadi Amiri"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-01", "url": "https://arxiv.org/abs/2504.00970", "pdf_url": "https://arxiv.org/pdf/2504.00970v2", "arxiv_id": "2504.00970", "doi": "10.48550/arXiv.2504.00970", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "799ddc73873c25cf1566a9b24988eb955558f1d99647163ec9c5c45e33cd1273", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Experts for Distributed Edge Computing with Channel-Aware Gating Function", "abstract": "In a distributed mixture-of-experts (MoE) system, a server collaborates with multiple specialized expert clients to perform inference. The server extracts features from input data and dynamically selects experts based on their areas of specialization to produce the final output. Although MoE models are widely valued for their flexibility and performance benefits, adapting distributed MoEs to operate effectively in wireless networks has remained unexplored. In this work, we introduce a novel channel-aware gating function for wireless distributed MoE, which incorporates channel conditions into the MoE gating mechanism. To train the channel-aware gating, we simulate various signal-to-noise ratios (SNRs) for each expert's communication channel and add noise to the features distributed to the experts based on these SNRs. The gating function then utilizes both features and SNRs to optimize expert selection. Unlike conventional MoE models which solely consider the alignment of features with the specializations of experts, our approach additionally considers the impact of channel conditions on expert performance. Experimental results demonstrate that the proposed channel-aware gating scheme outperforms traditional MoE models.", "authors": ["Qiuchen Song", "Shusen Jing", "Shuai Zhang", "Songyang Zhang", "Chuan Huang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-01", "url": "https://arxiv.org/abs/2504.00819", "pdf_url": "https://arxiv.org/pdf/2504.00819v1", "arxiv_id": "2504.00819", "doi": "10.1109/ICCWorkshops67674.2025.11162260", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2258}
{"id": "32177ce3063c6ca9d861f089892cee097dba8a9a073d8925f69d275416122663", "sources": ["arxiv", "semantic_scholar"], "title": "SQuat: Subspace-orthogonal KV Cache Quantization", "abstract": "The key-value (KV) cache accelerates LLMs decoding by storing KV tensors from previously generated tokens. It reduces redundant computation at the cost of increased memory usage. To mitigate this overhead, existing approaches compress KV tensors into lower-bit representations; however, quantization errors can accumulate as more tokens are generated, potentially resulting in undesired outputs. In this paper, we introduce SQuat (Subspace-orthogonal KV cache quantization). It first constructs a subspace spanned by query tensors to capture the most critical task-related information. During key tensor quantization, it enforces that the difference between the (de)quantized and original keys remains orthogonal to this subspace, minimizing the impact of quantization errors on the attention mechanism's outputs. SQuat requires no model fine-tuning, no additional calibration dataset for offline learning, and is grounded in a theoretical framework we develop. Through numerical experiments, we show that our method reduces peak memory by 2.17 to 2.82, improves throughput by 2.45 to 3.60, and achieves more favorable benchmark scores than existing KV cache quantization algorithms.", "authors": ["Hao Wang", "Ligong Han", "Kai Xu", "Akash Srivastava"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-03-31", "url": "https://arxiv.org/abs/2503.24358", "pdf_url": "https://arxiv.org/pdf/2503.24358v2", "arxiv_id": "2503.24358", "doi": "10.48550/arXiv.2503.24358", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "7080114fef59332bd84aa733267953113f973e328290d68654306dd300d2e850", "sources": ["arxiv", "semantic_scholar"], "title": "AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference", "abstract": "Recent advancements in Large Visual Language Models (LVLMs) have gained significant attention due to their remarkable reasoning capabilities and proficiency in generalization. However, processing a large number of visual tokens and generating long-context outputs impose substantial computational overhead, leading to excessive demands for key-value (KV) cache. To address this critical bottleneck, we propose AirCache, a novel KV cache compression method aimed at accelerating LVLMs inference. This work systematically investigates the correlations between visual and textual tokens within the attention mechanisms of LVLMs. Our empirical analysis reveals considerable redundancy in cached visual tokens, wherein strategically eliminating these tokens preserves model performance while significantly accelerating context generation. Inspired by these findings, we introduce an elite observation window for assessing the importance of visual components in the KV cache, focusing on stable inter-modal relevancy modeling with enhanced multi-perspective consistency. Additionally, we develop an adaptive layer-wise budget allocation strategy that capitalizes on the strength and skewness of token importance distribution, showcasing superior efficiency compared to uniform allocation. Comprehensive evaluations across multiple LVLMs and benchmarks demonstrate that our method achieves comparable performance to the full cache while retaining only 10% of visual KV cache, thereby reducing decoding latency by 29% to 66% across various batch size and prompt length of inputs. Notably, as cache retention rates decrease, our method exhibits increasing performance advantages over existing approaches.", "authors": ["Kai Huang", "Hao Zou", "Bochen Wang", "Ye Xi", "Zhen Xie", "Hao Wang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-31", "url": "https://arxiv.org/abs/2503.23956", "pdf_url": "https://arxiv.org/pdf/2503.23956v3", "arxiv_id": "2503.23956", "doi": "10.1109/ICCV51701.2025.02221", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.1747}
{"id": "9a7e6f6d2167315b34c580e8835650866eea9472f991b78767be58f40c414cec", "sources": ["arxiv", "semantic_scholar"], "title": "TransMamba: A Sequence-Level Hybrid Transformer-Mamba Language Model", "abstract": "Transformers are the cornerstone of modern large language models, but their quadratic computational complexity limits efficiency in long-sequence processing. Recent advancements in Mamba, a state space model (SSM) with linear complexity, offer promising efficiency gains but suffer from unstable contextual learning and multitask generalization. Some works conduct layer-level hybrid structures that combine Transformer and Mamba layers, aiming to make full use of both advantages. This paper proposes TransMamba, a novel sequence-level hybrid framework that unifies Transformer and Mamba through shared parameter matrices (QKV and CBx), and thus could dynamically switch between attention and SSM mechanisms at different token lengths and layers. We design the Memory Converter to bridge Transformer and Mamba by converting attention outputs into SSM-compatible states, ensuring seamless information flow at TransPoints where the transformation happens. The TransPoint scheduling is also thoroughly explored for balancing effectiveness and efficiency. We conducted extensive experiments demonstrating that TransMamba achieves superior training efficiency and performance compared to single and hybrid baselines, and validated the deeper consistency between Transformer and Mamba paradigms at sequence level, offering a scalable solution for next-generation language modeling. Code and data are available at https://github.com/Yixing-Li/TransMamba", "authors": ["Yixing Li", "Ruobing Xie", "Zhen Yang", "Xingwu Sun", "Shuaipeng Li", "Weidong Han", "Zhanhui Kang", "Yu Cheng", "Chengzhong Xu", "Di Wang", "Jie Jiang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-31", "url": "https://arxiv.org/abs/2503.24067", "pdf_url": "https://arxiv.org/pdf/2503.24067v2", "arxiv_id": "2503.24067", "doi": "10.1609/aaai.v40i38.40451", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Yixing-Li/TransMamba", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.1735}
{"id": "941489c5cfec19b712b713099529f526289734b876d2030d9a9c04b30c27c82e", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Routers", "abstract": "Supervised fine-tuning (SFT) is a milestone in aligning large language models with human instructions and adapting them to downstream tasks. In particular, Low-Rank Adaptation (LoRA) has gained widespread attention due to its parameter efficiency. However, its impact on improving the performance of large models remains limited. Recent studies suggest that combining LoRA with Mixture-of-Experts (MoE) can significantly enhance fine-tuning performance. MoE adapts to the diversity and complexity of datasets by dynamically selecting the most suitable experts, thereby improving task accuracy and efficiency. Despite impressive results, recent studies reveal issues in the MoE routing mechanism, such as incorrect assignments and imbalanced expert allocation. Inspired by the principles of Redundancy and Fault Tolerance Theory. We innovatively integrate the concept of Mixture of Experts into the routing mechanism and propose an efficient fine-tuning method called Mixture of Routers (MoR). It employs multiple sub-routers for joint selection and uses a learnable main router to determine the weights of the sub-routers. The results show that MoR outperforms baseline models on most tasks, achieving an average performance improvement of 1%. MoR can serve as a plug-and-play, parameter-efficient fine-tuning method suitable for a wide range of applications. Our code is available here: https://anonymous.4open.science/r/MoR-DFC6.", "authors": ["Jia-Chen Zhang", "Yu-Jie Xiong", "Xi-He Qiu", "Chun-Ming Xia", "Fei Dai", "Zheng Zhou"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-30", "url": "https://arxiv.org/abs/2503.23362", "pdf_url": "https://arxiv.org/pdf/2503.23362v3", "arxiv_id": "2503.23362", "doi": "10.48550/arXiv.2503.23362", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "c1888bed0e12191641d4f5dc0ac93cc3bc2948b1543494f192d5df4ac5f7d8d5", "sources": ["arxiv", "semantic_scholar"], "title": "Cocktail: Chunk-Adaptive Mixed-Precision Quantization for Long-Context LLM Inference", "abstract": "Recently, large language models (LLMs) have been able to handle longer and longer contexts. However, a context that is too long may cause intolerant inference latency and GPU memory usage. Existing methods propose mixed-precision quantization to the key-value (KV) cache in LLMs based on token granularity, which is time-consuming in the search process and hardware inefficient during computation. This paper introduces a novel approach called Cocktail, which employs chunk-adaptive mixed-precision quantization to optimize the KV cache. Cocktail consists of two modules: chunk-level quantization search and chunk-level KV cache computation. Chunk-level quantization search determines the optimal bitwidth configuration of the KV cache chunks quickly based on the similarity scores between the corresponding context chunks and the query, maintaining the model accuracy. Furthermore, chunk-level KV cache computation reorders the KV cache chunks before quantization, avoiding the hardware inefficiency caused by mixed-precision quantization in inference computation. Extensive experiments demonstrate that Cocktail outperforms state-of-the-art KV cache quantization methods on various models and datasets.", "authors": ["Wei Tao", "Bin Zhang", "Xiaoyang Qu", "Jiguang Wan", "Jianzong Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-30", "url": "https://arxiv.org/abs/2503.23294", "pdf_url": "https://arxiv.org/pdf/2503.23294v1", "arxiv_id": "2503.23294", "doi": "10.23919/DATE64628.2025.10992912", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Design, Automation and Test in Europe", "quality_score": 0.2258}
{"id": "74eeffa0149b50068dbf10e025e6f90099412f035dfa4efebc0b7dd181cff255", "sources": ["arxiv", "semantic_scholar"], "title": "S2MoE: Robust Sparse Mixture of Experts via Stochastic Learning", "abstract": "Sparse Mixture of Experts (SMoE) enables efficient training of large language models by routing input tokens to a select number of experts. However, training SMoE remains challenging due to the issue of representation collapse. Recent studies have focused on improving the router to mitigate this problem, but existing approaches face two key limitations: (1) expert embeddings are significantly smaller than the model's dimension, contributing to representation collapse, and (2) routing each input to the Top-K experts can cause them to learn overly similar features. In this work, we propose a novel approach called Robust Sparse Mixture of Experts via Stochastic Learning (S2MoE), which is a mixture of experts designed to learn from both deterministic and non-deterministic inputs via Learning under Uncertainty. Extensive experiments across various tasks demonstrate that S2MoE achieves performance comparable to other routing methods while reducing computational inference costs by 28%.", "authors": ["Giang Do", "Hung Le", "Truyen Tran"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-29", "url": "https://arxiv.org/abs/2503.23007", "pdf_url": "https://arxiv.org/pdf/2503.23007v1", "arxiv_id": "2503.23007", "doi": "10.48550/arXiv.2503.23007", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.11}
{"id": "d7951e7279f6f18512bd799197f69f78ed68d6d11b0e4a6d8b8a2fbafa082829", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Sparse Mixture of Experts from a Unified Perspective", "abstract": "Sparse Mixture of Experts (SMoE) models scale the capacity of models while maintaining constant computational overhead. SMoE methods fall into two categories: Token Choice, which routes each token to a fixed number of experts, and Expert Choice, which assigns a fixed number of tokens to each expert. However, the use of fixed budgets for tokens or experts causes both approaches to select irrelevant token-expert pairs or overlook critical assignments, which degrades overall performance. To fill that gap, we rethink SMoE from a unified perspective through the lens of linear programming, which provides a general formulation for SMoE models. Furthermore, we introduce Unified Sparse Mixture of Experts (USMoE), a novel framework comprising a unified mechanism and a unified score to overcome these limitations. We provide both theoretical justification and empirical evidence demonstrating USMoE's effectiveness. Extensive evaluations across diverse data settings (clean and corrupted), multiple domains (including texts and vision tasks), and different learning approaches (training-free and training-based) show that USMoE not only delivers significant performance improvements over existing SMoE methods, but also enables more flexible expert selection budgets, reducing inference costs without compromising model performance. Our implementation is publicly available at https://github.com/giangdip2410/USMoE.", "authors": ["Giang Do", "Hung Le", "Truyen Tran"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-29", "url": "https://arxiv.org/abs/2503.22996", "pdf_url": "https://arxiv.org/pdf/2503.22996v3", "arxiv_id": "2503.22996", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/giangdip2410/USMoE", "venue": "ICML 2026", "quality_score": 0.17}
{"id": "07c988b277de24aae828dd3e12ba0137021e203328a2da302a8e00795e275e6f", "sources": ["arxiv", "semantic_scholar"], "title": "MoLAE: Mixture of Latent Experts for Parameter-Efficient Language Models", "abstract": "Mixture of Experts (MoE) has become a key architectural paradigm for efficiently scaling Large Language Models (LLMs) by selectively activating a subset of parameters for each input token. However, standard MoE architectures face significant challenges, including high memory consumption and communication overhead during distributed training. In this paper, we introduce Mixture of Latent Experts (MoLAE), a novel parameterization that addresses these limitations by reformulating expert operations through a shared projection into a lower-dimensional latent space, followed by expert-specific transformations. This factorized approach substantially reduces parameter count and computational requirements, particularly in existing LLMs where hidden dimensions significantly exceed MoE intermediate dimensions. We provide a rigorous mathematical framework for transforming pre-trained MoE models into MoLAE architecture, characterizing conditions for optimal factorization, and developing a systematic two-step algorithm for this conversion. Our comprehensive theoretical analysis demonstrates that MoLAE significantly improves efficiency across multiple dimensions while preserving model capabilities. Experimental results confirm that MoLAE achieves comparable performance to standard MoE with substantially reduced resource requirements.", "authors": ["Zehua Liu", "Han Wu", "Ruifeng She", "Xiaojin Fu", "Xiongwei Han", "Tao Zhong", "Mingxuan Yuan"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-29", "url": "https://arxiv.org/abs/2503.23100", "pdf_url": "https://arxiv.org/pdf/2503.23100v2", "arxiv_id": "2503.23100", "doi": null, "citation_count": 5, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2386}
{"id": "00be495fad774c40dbf1ea2ba4236258f2754331cfbe2cd5606e4a8fa327c537", "sources": ["arxiv", "semantic_scholar"], "title": "LLaVA-CMoE: Towards Continual Mixture of Experts for Large Vision-Language Models", "abstract": "Mixture of Experts (MoE) architectures have recently advanced the scalability and adaptability of large language models (LLMs) for continual multimodal learning. However, efficiently extending these models to accommodate sequential tasks remains challenging. As new tasks arrive, naive model expansion leads to rapid parameter growth, while modifying shared routing components often causes catastrophic forgetting, undermining previously learned knowledge. To address these issues, we propose LLaVA-CMoE, a continual learning framework for LLMs that requires no replay data of previous tasks and ensures both parameter efficiency and robust knowledge retention. Our approach introduces a Probe-Guided Knowledge Extension mechanism, which uses probe experts to dynamically determine when and where new experts should be added, enabling adaptive and minimal parameter expansion tailored to task complexity. Furthermore, we present a Probabilistic Task Locator that assigns each task a dedicated, lightweight router. To handle the practical issue that task labels are unknown during inference, we leverage a VAE-based reconstruction strategy to identify the most suitable router by matching input distributions, allowing automatic and accurate expert allocation. This design mitigates routing conflicts and catastrophic forgetting, enabling robust continual learning without explicit task labels. Extensive experiments on the CoIN benchmark, covering eight diverse VQA tasks, demonstrate that LLaVA-CMoE delivers strong continual learning performance with a compact model size, significantly reducing forgetting and parameter overhead compared to prior methods. These results showcase the effectiveness and scalability of our approach for parameter-efficient continual learning in large language models. Our code will be open-sourced soon.", "authors": ["Hengyuan Zhao", "Ziqin Wang", "Qixin Sun", "Kaiyou Song", "Yilin Li", "Xiaolin Hu", "Qingpei Guo", "Si Liu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-27", "url": "https://arxiv.org/abs/2503.21227", "pdf_url": "https://arxiv.org/pdf/2503.21227v3", "arxiv_id": "2503.21227", "doi": "10.48550/arXiv.2503.21227", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1665}
{"id": "0d27debade3560afc8c57108a471e12ff8fb79ace92495200bdeb92e3f24aa75", "sources": ["arxiv", "semantic_scholar"], "title": "MoRE-LLM: Mixture of Rule Experts Guided by a Large Language Model", "abstract": "To ensure the trustworthiness and interpretability of AI systems, it is essential to align machine learning models with human domain knowledge. This can be a challenging and time-consuming endeavor that requires close communication between data scientists and domain experts. Recent leaps in the capabilities of Large Language Models (LLMs) can help alleviate this burden. In this paper, we propose a Mixture of Rule Experts guided by a Large Language Model (MoRE-LLM) which combines a data-driven black-box model with knowledge extracted from an LLM to enable domain knowledge-aligned and transparent predictions. While the introduced Mixture of Rule Experts (MoRE) steers the discovery of local rule-based surrogates during training and their utilization for the classification task, the LLM is responsible for enhancing the domain knowledge alignment of the rules by correcting and contextualizing them. Importantly, our method does not rely on access to the LLM during test time and ensures interpretability while not being prone to LLM-based confabulations. We evaluate our method on several tabular data sets and compare its performance with interpretable and non-interpretable baselines. Besides performance, we evaluate our grey-box method with respect to the utilization of interpretable rules. In addition to our quantitative evaluation, we shed light on how the LLM can provide additional context to strengthen the comprehensibility and trustworthiness of the model's reasoning process.", "authors": ["Alexander Koebler", "Ingo Thon", "Florian Buettner"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-26", "url": "https://arxiv.org/abs/2503.22731", "pdf_url": "https://arxiv.org/pdf/2503.22731v1", "arxiv_id": "2503.22731", "doi": "10.1109/ICDM59182.2024.00085", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Industrial Conference on Data Mining", "quality_score": 0.1066}
{"id": "60bd71a1ce0fe99df39a73a24826bcb58b9902e394ab644554eb0ce33559863b", "sources": ["arxiv", "semantic_scholar"], "title": "Vision as LoRA", "abstract": "We introduce Vision as LoRA (VoRA), a novel paradigm for transforming an LLM into an MLLM. Unlike prevalent MLLM architectures that rely on external vision modules for vision encoding, VoRA internalizes visual capabilities by integrating vision-specific LoRA layers directly into the LLM. This design allows the added parameters to be seamlessly merged into the LLM during inference, eliminating structural complexity and minimizing computational overhead. Moreover, inheriting the LLM's ability of handling flexible context, VoRA can process inputs at arbitrary resolutions. To further strengthen VoRA's visual capabilities, we introduce a block-wise distillation method that transfers visual priors from a pre-trained ViT into the LoRA layers, effectively accelerating training by injecting visual knowledge. Additionally, we apply bi-directional attention masks to better capture the context information of an image. We successfully demonstrate that with additional pre-training data, VoRA can perform comparably with conventional encode-based MLLMs. All training data, codes, and model weights will be released at https://github.com/Hon-Wong/VoRA.", "authors": ["Han Wang", "Yongjie Ye", "Bingru Li", "Yuxiang Nie", "Jinghui Lu", "Jingqun Tang", "Yanjie Wang", "Can Huang"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-26", "url": "https://arxiv.org/abs/2503.20680", "pdf_url": "https://arxiv.org/pdf/2503.20680v1", "arxiv_id": "2503.20680", "doi": "10.48550/arXiv.2503.20680", "citation_count": 31, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Hon-Wong/VoRA", "venue": "arXiv.org", "quality_score": 0.3763}
{"id": "3c0276fa35cffc86cf7fd18d3728cdf76407389aaed2f8afd6be25f97080aa00", "sources": ["arxiv", "semantic_scholar"], "title": "LogQuant: Log-Distributed 2-Bit Quantization of KV Cache with Superior Accuracy Preservation", "abstract": "We introduce LogQuant, a groundbreaking 2-bit quantization technique for KV Cache in large language model (LLM) inference, delivering substantial memory savings while preserving superior performance. Previous methods either assume that later tokens are more important or attempt to predict important tokens based on earlier attention patterns. Both approaches, however, can result in performance bottlenecks or frequent mispredictions. LogQuant takes a different approach. By applying a log-based filtering mechanism, it selectively compresses the KV Cache across the entire context, achieving better performance with the same or even reduced memory footprint compared to existing methods. In benchmark tests, it enhances throughput by 25% and boosts batch size by 60% without increasing memory consumption. For challenging tasks such as Math and Code Completion, LogQuant improves accuracy by 40% to 200% at the same compression ratio, outperforming comparable techniques.LogQuant integrates effortlessly with popular inference frameworks like Python's transformers library. Implementation can be available in https://github.com/Concyclics/LogQuantKV.", "authors": ["Han Chen", "Zicong Jiang", "Zining Zhang", "Bingsheng He", "Pingyi Luo", "Mian Lu", "Yuqiang Chen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-25", "url": "https://arxiv.org/abs/2503.19950", "pdf_url": "https://arxiv.org/pdf/2503.19950v1", "arxiv_id": "2503.19950", "doi": "10.48550/arXiv.2503.19950", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Concyclics/LogQuantKV", "venue": "arXiv.org", "quality_score": 0.1629}
{"id": "97ae5ccaf533d1090c1bf5ac038fd73987999df215cd3be9dc3d21cfe54da95e", "sources": ["arxiv", "semantic_scholar"], "title": "Attention IoU: Examining Biases in CelebA using Attention Maps", "abstract": "Computer vision models have been shown to exhibit and amplify biases across a wide array of datasets and tasks. Existing methods for quantifying bias in classification models primarily focus on dataset distribution and model performance on subgroups, overlooking the internal workings of a model. We introduce the Attention-IoU (Attention Intersection over Union) metric and related scores, which use attention maps to reveal biases within a model's internal representations and identify image features potentially causing the biases. First, we validate Attention-IoU on the synthetic Waterbirds dataset, showing that the metric accurately measures model bias. We then analyze the CelebA dataset, finding that Attention-IoU uncovers correlations beyond accuracy disparities. Through an investigation of individual attributes through the protected attribute of Male, we examine the distinct ways biases are represented in CelebA. Lastly, by subsampling the training set to change attribute correlations, we demonstrate that Attention-IoU reveals potential confounding variables not present in dataset labels.", "authors": ["Aaron Serianni", "Tyler Zhu", "Olga Russakovsky", "Vikram V. Ramaswamy"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-25", "url": "https://arxiv.org/abs/2503.19846", "pdf_url": "https://arxiv.org/pdf/2503.19846v2", "arxiv_id": "2503.19846", "doi": "10.48550/arXiv.2503.19846", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/aaronserianni/attention-iou", "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.1747}
{"id": "8f7942ec9ae7708e3882e3a6ca111780a00a05b529b5a262c57e9f6e6ace220a", "sources": ["arxiv", "semantic_scholar"], "title": "Oaken: Fast and Efficient LLM Serving with Online-Offline Hybrid KV Cache Quantization", "abstract": "Modern Large Language Model serving system batches multiple requests to achieve high throughput, while batching attention operations is challenging, rendering memory bandwidth a critical bottleneck. The community relies on high-end GPUs with multiple high-bandwidth memory channels. Unfortunately, HBM's high bandwidth often comes at the expense of limited memory capacity, which reduces core utilization and increases costs. Recent advancements enabling longer contexts for LLMs have substantially increased the key-value cache size, further intensifying the pressures on memory capacity. The literature has explored KV cache quantization techniques, which commonly use low bitwidth for most values, selectively using higher bitwidth for outlier values. While this approach helps achieve high accuracy and low bitwidth simultaneously, it comes with the limitation that cost for online outlier detection is excessively high, negating the advantages. We propose Oaken, an acceleration solution that achieves high accuracy and high performance simultaneously through co-designing algorithm and hardware. To effectively find a sweet spot in the accuracy-performance trade-off space of KV cache quantization, Oaken employs an online-offline hybrid approach, setting outlier thresholds offline, which are then used to determine the quantization scale online. To translate the proposed algorithmic technique into tangible performance gains, Oaken also comes with custom quantization engines and memory management units that can be integrated with any LLM accelerators. We built an Oaken accelerator on top of an LLM accelerator, LPU, and conducted a comprehensive evaluation. Our experiments show that for a batch size of 256, Oaken achieves up to 1.58x throughput improvement over NVIDIA A100 GPU, incurring a minimal accuracy loss of only 0.54\\% on average, compared to state-of-the-art KV cache quantization techniques.", "authors": ["Minsu Kim", "Seongmin Hong", "RyeoWook Ko", "Soongyu Choi", "Hunjong Lee", "Junsoo Kim", "Joo-Young Kim", "Jongse Park"], "categories": ["cs.AR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-24", "url": "https://arxiv.org/abs/2503.18599", "pdf_url": "https://arxiv.org/pdf/2503.18599v2", "arxiv_id": "2503.18599", "doi": "10.1145/3695053.3731019", "citation_count": 18, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Symposium on Computer Architecture", "quality_score": 0.3197}
{"id": "d643c6c1ce78e71427df527c20cb43516c79a313b12c4aa7767a7469ce6d89c9", "sources": ["arxiv", "semantic_scholar"], "title": "xKV: Cross-Layer KV-Cache Compression via Aligned Singular Vector Extraction", "abstract": "Long-context Large Language Models (LLMs) enable powerful applications but incur high memory costs due to the key-value states (KV-Cache). Recent studies attempt to share KV-Cache across layers, but these approaches either require expensive pretraining or rely on per-token cross-layer cosine similarity that is often limited in practice. We show, via Centered Kernel Alignment (CKA), that the dominant singular vectors of KV-Cache are well aligned across layers. Motivated by this observation, we propose xKV, a post-training compression method that jointly factorizes grouped-layer KV-Cache into a shared low-rank subspace, substantially reducing KV-Cache memory. Across widely used LLMs, xKV achieves up to 8x KV-Cache compression while preserving accuracy on long-context tasks and in multi-turn settings. To further improve efficiency, we introduce Selective Reconstruction (SR) at decode time. Combined with SR, xKV achieves up to 4.23x end-to-end speedup over the full attention baseline, and surpasses notable baselines with 30% higher throughput under a similar accuracy level. Overall, xKV provides a plug-and-play approach to reduce both memory and latency for long-context LLM inference. Our code is publicly available at: https://github.com/abdelfattah-lab/xKV.", "authors": ["Chi-Chih Chang", "Wei-Cheng Lin", "Chien-Yu Lin", "Hung-Yueh Chiang", "Yash Akhauri", "Xilai Dai", "Huiqiang Jiang", "Yucheng Li", "Luis Ceze", "Kai-Chiang Wu", "Mohamed S. Abdelfattah"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-24", "url": "https://arxiv.org/abs/2503.18893", "pdf_url": "https://arxiv.org/pdf/2503.18893v2", "arxiv_id": "2503.18893", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/abdelfattah-lab/xKV", "venue": null, "quality_score": 0.1232}
{"id": "1ff78f867bcba55cf48326ae953524c01aff55475d1b478ce4cb52d2c6d8dfe4", "sources": ["arxiv", "semantic_scholar"], "title": "BitDecoding: Unlocking Tensor Cores for Long-Context LLMs with Low-Bit KV Cache", "abstract": "The growth of long-context Large Language Models (LLMs) significantly increases memory and bandwidth pressure during autoregressive decoding due to the expanding Key-Value (KV) cache. While accuracy-preserving KV-cache quantization (e.g., 4-bit or 2-bit) reduces memory footprint, existing systems decode inefficiently by relying solely on CUDA cores, underutilizing Tensor Cores-the dominant compute resource on GPUs. We present BitDecoding, the first inference system to efficiently decode low-bit KV caches by cooperatively leveraging CUDA cores and Tensor Cores. BitDecoding smartly induces Tensor-Core-friendly layouts, introduces warp-level dequantization parallelism, and provides unified system support through query transformation, high-performance tensor- and channel-wise quantization, and a software-pipelined dequantization kernel enabling mixed-precision execution. Architecture-aware optimizations further leverage Hopper's warpgroup tensor instructions and Blackwell's NVFP4 (MXFP4) tensor formats. Evaluated on Blackwell, Hopper, and Ampere GPUs, BitDecoding achieves an average 7.5x decoding speedup over FP16 FlashDecoding-v2, up to 8.6x on Blackwell with NVFP4, and up to 4.3x over state-of-the-art approaches. On LLaMA-3.1-8B with a 128K context, BitDecoding reduces single-batch decoding latency by 3x. BitDecoding is open-sourced at https://github.com/OpenBitSys/BitDecoding.", "authors": ["Dayou Du", "Shijie Cao", "Jianyi Cheng", "Luo Mai", "Ting Cao", "Mao Yang"], "categories": ["cs.AR", "cs.AI", "cs.CL", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-24", "url": "https://arxiv.org/abs/2503.18773", "pdf_url": "https://arxiv.org/pdf/2503.18773v3", "arxiv_id": "2503.18773", "doi": "10.1109/HPCA68181.2026.11408481", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/OpenBitSys/BitDecoding", "venue": "International Symposium on High-Performance Computer Architecture", "quality_score": 0.2113}
{"id": "bef0bfdf0fb4c25984c70ec0ef8fe04169cc5a545eda82666a4b97568133cb81", "sources": ["arxiv", "semantic_scholar"], "title": "TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model", "abstract": "Vision-Language Models (VLMs) demand substantial computational resources during inference, largely due to the extensive visual input tokens for representing visual information. Previous studies have noted that visual tokens tend to receive less attention than text tokens, suggesting their lower importance during inference and potential for pruning. However, their methods encounter several challenges: reliance on greedy heuristic criteria for token importance and incompatibility with FlashAttention and KV cache. To address these issues, we introduce \\textbf{TopV}, a compatible \\textbf{TO}ken \\textbf{P}runing with inference Time Optimization for fast and low-memory \\textbf{V}LM, achieving efficient pruning without additional training or fine-tuning. Instead of relying on attention scores, we formulate token pruning as an optimization problem, accurately identifying important visual tokens while remaining compatible with FlashAttention. Additionally, since we only perform this pruning once during the prefilling stage, it effectively reduces KV cache size. Our optimization framework incorporates a visual-aware cost function considering factors such as Feature Similarity, Relative Spatial Distance, and Absolute Central Distance, to measure the importance of each source visual token, enabling effective pruning of low-importance tokens. Extensive experiments demonstrate that our method outperforms previous token pruning methods, validating the effectiveness and efficiency of our approach.", "authors": ["Cheng Yang", "Yang Sui", "Jinqi Xiao", "Lingyi Huang", "Yu Gong", "Chendi Li", "Jinghua Yan", "Yu Bai", "Ponnuswamy Sadayappan", "Xia Hu", "Bo Yuan"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-24", "url": "https://arxiv.org/abs/2503.18278", "pdf_url": "https://arxiv.org/pdf/2503.18278v2", "arxiv_id": "2503.18278", "doi": "10.1109/CVPR52734.2025.01844", "citation_count": 46, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.418}
{"id": "7d96d384f56289c8e2a642d50e47d006739b807ea48f30b4d77773c930da5990", "sources": ["arxiv", "semantic_scholar"], "title": "SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking", "abstract": "Most state-of-the-art trackers adopt one-stream paradigm, using a single Vision Transformer for joint feature extraction and relation modeling of template and search region images. However, relation modeling between different image patches exhibits significant variations. For instance, background regions dominated by target-irrelevant information require reduced attention allocation, while foreground, particularly boundary areas, need to be be emphasized. A single model may not effectively handle all kinds of relation modeling simultaneously. In this paper, we propose a novel tracker called SPMTrack based on mixture-of-experts tailored for visual tracking task (TMoE), combining the capability of multiple experts to handle diverse relation modeling more flexibly. Benefiting from TMoE, we extend relation modeling from image pairs to spatio-temporal context, further improving tracking accuracy with minimal increase in model parameters. Moreover, we employ TMoE as a parameter-efficient fine-tuning method, substantially reducing trainable parameters, which enables us to train SPMTrack of varying scales efficiently and preserve the generalization ability of pretrained models to achieve superior performance. We conduct experiments on seven datasets, and experimental results demonstrate that our method significantly outperforms current state-of-the-art trackers. The source code is available at https://github.com/WenRuiCai/SPMTrack.", "authors": ["Wenrui Cai", "Qingjie Liu", "Yunhong Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-24", "url": "https://arxiv.org/abs/2503.18338", "pdf_url": "https://arxiv.org/pdf/2503.18338v2", "arxiv_id": "2503.18338", "doi": "10.1109/CVPR52734.2025.01572", "citation_count": 26, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/WenRuiCai/SPMTrack", "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.3578}
{"id": "798e704a26a47af4a355eac1c9afbb359fa3a434df41c0aedea8b4e37aa9ffff", "sources": ["arxiv", "semantic_scholar"], "title": "WindowKV: Task-Adaptive Group-Wise KV Cache Window Selection for Efficient LLM Inference", "abstract": "With the advancements in long-context inference capabilities of large language models (LLMs), the KV cache has become one of the foundational components. However, its substantial GPU memory consumption makes KV cache compression a key technique for enabling efficient LLM inference in industrial scenarios. While recent studies have focused on optimizing the memory occupied by the KV cache, they overlook two critical factors: preserving semantic coherence and considering task-specific characteristic during compression. To address these limitations, we propose a novel task-adaptive KV cache window selection method, WindowKV. WindowKV dynamically selects local semantic windows consisting of consecutive tokens, according to task-specific characteristics, ensuring the retained KV cache captures continuous, essential context. Additionally, we introduce an intra-group layer KV cache indices sharing strategy to reduce computational overhead, achieving a balance between performance and efficiency. We rigorously evaluate WindowKV on the LongBench benchmark, and the results demonstrate that it maintains a performance comparable to full KV cache retention while using only 12% of the original KV cache, significantly reducing memory requirements. Furthermore, our method also achieves state-of-the-art results in the Needle-in-a-Haystack evaluation, highlighting its effectiveness and robustness.", "authors": ["Youhui Zuo", "Sibo Wei", "Chen Zhang", "Zhuorui Liu", "Wenpeng Lu", "Dawei Song"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-23", "url": "https://arxiv.org/abs/2503.17922", "pdf_url": "https://arxiv.org/pdf/2503.17922v2", "arxiv_id": "2503.17922", "doi": "10.48550/arXiv.2503.17922", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1031}
{"id": "cab68b4a58abd849a0e4a34615c89ebd9329cf222dbd5d18cb7cb2331421051e", "sources": ["arxiv", "semantic_scholar"], "title": "ExpertRAG: Efficient RAG with Mixture of Experts -- Optimizing Context Retrieval for Adaptive LLM Responses", "abstract": "ExpertRAG is a novel theoretical framework that integrates Mixture-of-Experts (MoE) architectures with Retrieval Augmented Generation (RAG) to advance the efficiency and accuracy of knowledge-intensive language modeling. We propose a dynamic retrieval gating mechanism coupled with expert routing, enabling the model to selectively consult an external knowledge store or rely on specialized internal experts based on the query's needs. The paper lays out the theoretical foundations of ExpertRAG, including a probabilistic formulation that treats retrieval and expert selection as latent decisions, and mathematical justifications for its efficiency in both computation and knowledge utilization. We derive formulae to quantify the expected computational cost savings from selective retrieval and the capacity gains from sparse expert utilization. A comparative analysis positions ExpertRAG against standard RAG (with always-on retrieval) and pure MoE models (e.g., Switch Transformer, Mixtral) to highlight its unique balance between parametric knowledge and non-parametric retrieval. We also outline an experimental validation strategy, proposing benchmarks and evaluation protocols to test ExpertRAG's performance on factual recall, generalization, and inference efficiency. The proposed framework, although presented theoretically, is supported by insights from prior work in RAG and MoE, and is poised to provide more factual, efficient, and adaptive generation by leveraging the best of both paradigms. In summary, ExpertRAG contributes a new perspective on scaling and augmenting language models, backed by a thorough analysis and a roadmap for empirical validation.", "authors": ["Esmail Gumaan"], "categories": ["cs.IR", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-23", "url": "https://arxiv.org/abs/2504.08744", "pdf_url": "https://arxiv.org/pdf/2504.08744v1", "arxiv_id": "2504.08744", "doi": "10.48550/arXiv.2504.08744", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "85a9c8de0ced368ec42b0df6c5ac74e2eccd9e61bfe86fcc6fc686c7695880a7", "sources": ["arxiv", "semantic_scholar"], "title": "Plug-and-Play 1.x-Bit KV Cache Quantization for Video Large Language Models", "abstract": "Video large language models (VideoLLMs) have demonstrated the capability to process longer video inputs and enable complex reasoning and analysis. However, due to the thousands of visual tokens from the video frames, the key-value (KV) cache can significantly increase memory requirements, becoming a bottleneck for inference speed and memory usage. KV cache quantization is a widely used approach to address this problem. In this paper, we find that 2-bit KV quantization of VideoLLMs can hardly hurt the model performance, while the limit of KV cache quantization in even lower bits has not been investigated. To bridge this gap, we introduce VidKV, a plug-and-play KV cache quantization method to compress the KV cache to lower than 2 bits. Specifically, (1) for key, we propose a mixed-precision quantization strategy in the channel dimension, where we perform 2-bit quantization for anomalous channels and 1-bit quantization combined with FFT for normal channels; (2) for value, we implement 1.58-bit quantization while selectively filtering semantically salient visual tokens for targeted preservation, for a better trade-off between precision and model performance. Importantly, our findings suggest that the value cache of VideoLLMs should be quantized in a per-channel fashion instead of the per-token fashion proposed by prior KV cache quantization works for LLMs. Empirically, extensive results with LLaVA-OV-7B and Qwen2.5-VL-7B on six benchmarks show that VidKV effectively compresses the KV cache to 1.5-bit and 1.58-bit precision with almost no performance drop compared to the FP16 counterparts.", "authors": ["Keda Tao", "Haoxuan You", "Yang Sui", "Can Qin", "Huan Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-20", "url": "https://arxiv.org/abs/2503.16257", "pdf_url": "https://arxiv.org/pdf/2503.16257v2", "arxiv_id": "2503.16257", "doi": "10.48550/arXiv.2503.16257", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "0696cac74c3e231eb3492dd1323d953e71bda4588a21b9897a875135def897d5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Lookup Experts", "abstract": "Mixture-of-Experts (MoE) activates only a subset of experts during inference, allowing the model to maintain low inference FLOPs and latency even as the parameter count scales up. However, since MoE dynamically selects the experts, all the experts need to be loaded into VRAM. Their large parameter size still limits deployment, and offloading, which load experts into VRAM only when needed, significantly increase inference latency. To address this, we propose Mixture of Lookup Experts (MoLE), a new MoE architecture that is efficient in both communication and VRAM usage. In MoLE, the experts are Feed-Forward Networks (FFNs) during training, taking the output of the embedding layer as input. Before inference, these experts can be re-parameterized as lookup tables (LUTs) that retrieves expert outputs based on input ids, and offloaded to storage devices. Therefore, we do not need to perform expert computations during inference. Instead, we directly retrieve the expert's computation results based on input ids and load them into VRAM, and thus the resulting communication overhead is negligible. Experiments show that, with the same FLOPs and VRAM usage, MoLE achieves inference speeds comparable to dense models and significantly faster than MoE with experts offloading, while maintaining performance on par with MoE.", "authors": ["Shibo Jie", "Yehui Tang", "Kai Han", "Yitong Li", "Duyu Tang", "Zhi-Hong Deng", "Yunhe Wang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-20", "url": "https://arxiv.org/abs/2503.15798", "pdf_url": "https://arxiv.org/pdf/2503.15798v2", "arxiv_id": "2503.15798", "doi": "10.48550/arXiv.2503.15798", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2785}
{"id": "93baf5d0c2cb2a395c32ba04281b1803de22c5f23b1b0dbb7ff58d9fad41b231", "sources": ["arxiv", "semantic_scholar"], "title": "Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts", "abstract": "Diffusion models have emerged as mainstream framework in visual generation. Building upon this success, the integration of Mixture of Experts (MoE) methods has shown promise in enhancing model scalability and performance. In this paper, we introduce Race-DiT, a novel MoE model for diffusion transformers with a flexible routing strategy, Expert Race. By allowing tokens and experts to compete together and select the top candidates, the model learns to dynamically assign experts to critical tokens. Additionally, we propose per-layer regularization to address challenges in shallow layer learning, and router similarity loss to prevent mode collapse, ensuring better expert utilization. Extensive experiments on ImageNet validate the effectiveness of our approach, showcasing significant performance gains while promising scaling properties.", "authors": ["Yike Yuan", "Ziyu Wang", "Zihao Huang", "Defa Zhu", "Xun Zhou", "Jingyi Yu", "Qiyang Min"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-20", "url": "https://arxiv.org/abs/2503.16057", "pdf_url": "https://arxiv.org/pdf/2503.16057v3", "arxiv_id": "2503.16057", "doi": "10.48550/arXiv.2503.16057", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.25}
{"id": "907582e7045cb6bdc5cacd8130d68efa12d51da1c22bb27b46623dc3f3736820", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Pruning: Automated Fairness Repair of Language Models via Surrogate Simulated Annealing", "abstract": "This paper explores pruning attention heads as a post-processing bias mitigation method for large language models (LLMs). Modern AI systems such as LLMs are expanding into sensitive social contexts where fairness concerns become especially crucial. Since LLMs develop decision-making patterns by training on massive datasets of human-generated content, they naturally encode and perpetuate societal biases. While modifying training datasets and algorithms is expensive and requires significant resources; post-processing techniques-such as selectively deactivating neurons and attention heads in pre-trained LLMs-can provide feasible and effective approaches to improve fairness. However, identifying the optimal subset of parameters to prune presents a combinatorial challenge within LLMs' immense parameter space, requiring solutions that efficiently balance competing objectives across the frontiers of model fairness and utility. To address the computational challenges, we explore a search-based program repair approach via randomized simulated annealing. Given the prohibitive evaluation costs in billion-parameter LLMs, we develop surrogate deep neural networks that efficiently model the relationship between attention head states (active/inactive) and their corresponding fairness/utility metrics. This allows us to perform optimization over the surrogate models and efficiently identify optimal subsets of attention heads for selective pruning rather than directly searching through the LLM parameter space. This paper introduces Attention Pruning, a fairness-aware surrogate simulated annealing approach to prune attention heads in LLMs that disproportionately contribute to bias while minimally impacting overall model utility. Our experiments show that Attention Pruning achieves up to $40\\%$ reduction in gender bias and outperforms the state-of-the-art bias mitigation strategies.", "authors": ["Vishnu Asutosh Dasu", "Md Rafi ur Rashid", "Vipul Gupta", "Saeid Tizpaz-Niari", "Gang Tan"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-20", "url": "https://arxiv.org/abs/2503.15815", "pdf_url": "https://arxiv.org/pdf/2503.15815v3", "arxiv_id": "2503.15815", "doi": "10.48550/arXiv.2503.15815", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "0c403e5639e006c5d8916f31b1e6f74a4fee97444e3a51081acf4cabb849927c", "sources": ["arxiv", "semantic_scholar"], "title": "SpeCache: Speculative Key-Value Caching for Efficient Generation of LLMs", "abstract": "Transformer-based large language models (LLMs) have already achieved remarkable results on long-text tasks, but the limited GPU memory (VRAM) resources struggle to accommodate the linearly growing demand for key-value (KV) cache as the sequence length increases, which has become a bottleneck for the application of LLMs on long sequences. Existing KV cache compression methods include eviction, merging, or quantization of the KV cache to reduce its size. However, compression results in irreversible information forgetting, potentially affecting the accuracy of subsequent decoding. In this paper, we propose SpeCache, which takes full advantage of the large and easily expandable CPU memory to offload the complete KV cache, and dynamically fetches KV pairs back in each decoding step based on their importance measured by low-bit KV cache copy in VRAM. To avoid inference latency caused by CPU-GPU communication, SpeCache speculatively predicts the KV pairs that the next token might attend to, allowing us to prefetch them before the next decoding step which enables parallelization of prefetching and computation. Experiments on LongBench and Needle-in-a-Haystack benchmarks verify that SpeCache effectively reduces VRAM usage while avoiding information forgetting for long sequences without re-training, even with a 10x high KV cache compression ratio.", "authors": ["Shibo Jie", "Yehui Tang", "Kai Han", "Zhi-Hong Deng", "Jing Han"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-20", "url": "https://arxiv.org/abs/2503.16163", "pdf_url": "https://arxiv.org/pdf/2503.16163v1", "arxiv_id": "2503.16163", "doi": "10.48550/arXiv.2503.16163", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1945}
{"id": "3d71a955bc96015f1149b688d76114ef3a51fef9fd8fbf8c5def835264f2fd7d", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning-Based TinyML Optimization of Machine Learning Models for Anomaly Detection in Electric Vehicle Charging Infrastructure", "abstract": "With the growing need for real-time processing on IoT devices, optimizing machine learning (ML) models' size, latency, and computational efficiency is essential. This paper investigates a pruning method for anomaly detection in resource-constrained environments, specifically targeting Electric Vehicle Charging Infrastructure (EVCI). Using the CICEVSE2024 dataset, we trained and optimized three models-Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), and XGBoost-through hyperparameter tuning with Optuna, further refining them using SHapley Additive exPlanations (SHAP)-based feature selection (FS) and unstructured pruning techniques. The optimized models achieved significant reductions in model size and inference times, with only a marginal impact on their performance. Notably, our findings indicate that, in the context of EVCI, pruning and FS can enhance computational efficiency while retaining critical anomaly detection capabilities.", "authors": ["Fatemeh Dehrouyeh", "Ibrahim Shaer", "Soodeh Nikan", "Firouz Badrkhani Ajaei", "Abdallah Shami"], "categories": ["cs.LG", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-03-19", "url": "https://arxiv.org/abs/2503.14799", "pdf_url": "https://arxiv.org/pdf/2503.14799v1", "arxiv_id": "2503.14799", "doi": "10.1109/ICC52391.2025.11161868", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Western-OC2-Lab/EVCI-Pruning", "venue": null, "quality_score": 0.1193}
{"id": "2855c87523a948d10949b51d4fbfe01fbc80a9c9b4cd8bc32e407a1812454abb", "sources": ["arxiv", "semantic_scholar"], "title": "Tiled Flash Linear Attention: More Efficient Linear RNN and xLSTM Kernels", "abstract": "Linear RNNs with gating recently demonstrated competitive performance compared to Transformers in language modeling. Although their linear compute scaling in sequence length offers theoretical runtime advantages over Transformers, realizing these benefits in practice requires optimized custom kernels, as Transformers rely on the highly efficient Flash Attention kernels (Dao, 2024). Leveraging the chunkwise-parallel formulation of linear RNNs, Flash Linear Attention (FLA) (Yang & Zhang, 2024) shows that linear RNN kernels are faster than Flash Attention, by parallelizing over chunks of the input sequence. However, since the chunk size of FLA is limited, many intermediate states must be materialized in GPU memory. This leads to low arithmetic intensity and causes high memory consumption and IO cost, especially for long-context pre-training. In this work, we present Tiled Flash Linear Attention (TFLA), a novel kernel algorithm for linear RNNs, that enables arbitrary large chunk sizes and high arithmetic intensity by introducing an additional level of sequence parallelization within each chunk. First, we apply TFLA to the xLSTM with matrix memory, the mLSTM (Beck et al., 2024). Second, we propose an mLSTM variant with sigmoid input gate and reduced computation for even faster kernel runtimes at equal language modeling performance. In our speed benchmarks, we show that our new mLSTM kernels based on TFLA outperform highly optimized Flash Attention, Linear Attention and Mamba kernels, setting a new state of the art for efficient long-context sequence modeling primitives.", "authors": ["Maximilian Beck", "Korbinian Pöppel", "Phillip Lippe", "Sepp Hochreiter"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-18", "url": "https://arxiv.org/abs/2503.14376", "pdf_url": "https://arxiv.org/pdf/2503.14376v3", "arxiv_id": "2503.14376", "doi": "10.48550/arXiv.2503.14376", "citation_count": 13, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/NX-AI/mlstm_kernels", "venue": "arXiv.org", "quality_score": 0.2865}
{"id": "4f4dfb73d325e1f641f20fe7539ef6ae35ef6977abb2b7eefa300fd1d8899666", "sources": ["arxiv", "semantic_scholar"], "title": "Towards More Economical Context-Augmented LLM Generation by Reusing Stored KV Cache", "abstract": "Across large language model (LLM) applications, we observe an emerging trend for reusing KV caches to save the prefill delays of processing repeated input texts in different LLM inputs. This has led to a broad design space, including colocating stored KV caches with (or close to) GPUs to various KV cache compression. However, a key question remains unanswered: can these delay reductions also be economically favorable? Specifically, we ask whether a developer can use public cloud services to store precomputed KV caches and reuse them to save delay without incurring more costs in terms of compute, storage, and network. To answer this question, we propose an validated analytical model for the cloud cost (in compute, storage, and network) of storing and reusing KV caches based on various workload parameters, such as reuse frequency, generated text lengths, model sizes, etc. Preliminary results show that KV cache reusing is able to save both delay and cloud cost across a range of workloads with long context. And we call more efforts on building more economical context augmented LLM by KV cache reusing.", "authors": ["Hanchen Li", "Yuhan Liu", "Yihua Cheng", "Kuntai Du", "Junchen Jiang"], "categories": ["cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-18", "url": "https://arxiv.org/abs/2503.14647", "pdf_url": "https://arxiv.org/pdf/2503.14647v1", "arxiv_id": "2503.14647", "doi": "10.48550/arXiv.2503.14647", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "2976a00b1b1b3fcab66ff8b00fc3594d89fe0aea9fe155188d476e7e446bfa57", "sources": ["arxiv", "semantic_scholar"], "title": "KVShare: An LLM Service System with Efficient and Effective Multi-Tenant KV Cache Reuse", "abstract": "Recent advances in long-text understanding have pushed the context length of large language models (LLMs) up to one million tokens. It boosts LLMs's accuracy and reasoning capacity but causes exorbitant computational costs and unsatisfactory Time to First Token (TTFT). KV cache reuse, which reuses the exact same KV cache of prefixes and templates or shares similar ones but with extra selective recomputation, offers a promising way to tackle this issue. However, prior studies overlook the cross-request KV reuse and the attention deviations introduced by new tokens during the decoding stage. In this paper, we present a KV cache management module that shares the KV cache across requests under multi-tenant scenarios without sacrificing model accuracy. Our system, KVShare, enables accurate and efficient LLM serving by 1) a Dual-Stage High Deviation algorithm (DHD) that conditionally selects a small portion of KV cache to be recomputed during both prefill and decode phases, and 2) a cache-aware scheduler that prioritizes requests based on their KV cache hit rates and orchestrates continuous batching to achieve enhanced system efficiency and faster TTFT. Multi-task experiments conducted on models such as Qwen2.5-7B,Llama3.1-8B and Yi1.5-9B demonstrate that KVShare reduces TTFT by up to 9.39x and increases 1.2x of the throughput compared to the full KV recompute. Moreover, KVShare achieves 20.38% boost in terms of accuracy compared to SOTA methods.", "authors": ["Huan Yang", "Renji Zhang", "Mingzhe Huang", "Weijun Wang", "Yin Tang", "Yuanchun Li", "Yunxin Liu", "Deyu Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-17", "url": "https://arxiv.org/abs/2503.16525", "pdf_url": "https://arxiv.org/pdf/2503.16525v2", "arxiv_id": "2503.16525", "doi": null, "citation_count": 15, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.301}
{"id": "ff2781071bc31f06f22e6aceb209163252b14737ea1bb1cb120c1d29545ded2b", "sources": ["arxiv", "semantic_scholar"], "title": "Mitigating KV Cache Competition to Enhance User Experience in LLM Inference", "abstract": "In Large Language Model (LLM) serving, the KV-cache (KVC) bottleneck causes high tail Time-to-First-Token (TTFT) and Time-Between-Tokens (TBT), impairing user experience, particularly in time-sensitive applications. However, satisfying both TTFT and TBT service-level objectives (SLOs) is challenging. To address this, we propose a system, named CacheOPT for mitigating KV Cache competition, based on key insights from our measurements, incorporating novel components. First, it estimates a request's output length, bounding the deviation with a high specified probability, adjusted based on the request arrival rate. Second, it allocates the estimated KVC demand to a request, and reuses other requests' allocated KVC to avoid preemptions while reducing waiting time. Third, it proactively allocates KVC before instead of at the time a request exhausts its allocation and reserves KVC globally to prevent preemptions. Fourth, it chooses a request that has long TBT SLO, long job remaining time and short preemption time to preempt. Fifth, it selects the shortest-latency strategy between swapping and recomputation for preemptions. Experiments show that CacheOPT achieves up to 3.29$\\times$ and 2.83$\\times$ lower tail TBT and tail TTFT, 47\\% and 53\\% higher TTFT and TBT SLO attainments, and supports up to 1.58$\\times$ higher request arrival rate than the state-of-the-art methods.", "authors": ["Haiying Shen", "Tanmoy Sen", "Masahiro Tanaka"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-17", "url": "https://arxiv.org/abs/2503.13773", "pdf_url": "https://arxiv.org/pdf/2503.13773v2", "arxiv_id": "2503.13773", "doi": "10.48550/arXiv.2503.13773", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0963}
{"id": "10b23f10f4bfe4721ebe4b846842cb3e6ce78808707b389802068c727fac9b92", "sources": ["arxiv", "semantic_scholar"], "title": "CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences", "abstract": "Large language models (LLMs) excel at processing long sequences, boosting demand for key-value (KV) caching. While recent efforts to evict KV cache have alleviated the inference burden, they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a novel approach that frames KV cache eviction as a \"cake-slicing problem.\" CAKE assesses layer-specific preferences by considering attention dynamics in both spatial and temporal dimensions, allocates rational cache size for layers accordingly, and manages memory constraints in a cascading manner. This approach enables a global view of cache allocation, adaptively distributing resources across diverse attention mechanisms while maintaining memory budgets. CAKE also employs a new eviction indicator that considers the shifting importance of tokens over time, addressing limitations in existing methods that overlook temporal dynamics. Comprehensive experiments on LongBench and NeedleBench show that CAKE maintains model performance with only 3.2% of the KV cache and consistently outperforms current baselines across various models and memory constraints, particularly in low-memory settings. Additionally, CAKE achieves over 10x speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. Our code is available at https://github.com/antgroup/cakekv.", "authors": ["Ziran Qin", "Yuchen Cao", "Mingbao Lin", "Wen Hu", "Shixuan Fan", "Ke Cheng", "Weiyao Lin", "Jianguo Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-16", "url": "https://arxiv.org/abs/2503.12491", "pdf_url": "https://arxiv.org/pdf/2503.12491v2", "arxiv_id": "2503.12491", "doi": "10.48550/arXiv.2503.12491", "citation_count": 48, "influential_citation_count": 6, "has_code": true, "code_url": "https://github.com/antgroup/cakekv", "venue": "International Conference on Learning Representations", "quality_score": 0.4225}
{"id": "0306e9892799a89dfc744c73555ada04cee01a6c62c627534fded81e6b62e856", "sources": ["arxiv", "semantic_scholar"], "title": "MoECollab: Democratizing LLM Development Through Collaborative Mixture of Experts", "abstract": "Large Language Model (LLM) development has become increasingly centralized, limiting participation to well-resourced organizations. This paper introduces MoECollab, a novel framework leveraging Mixture of Experts (MoE) architecture to enable distributed, collaborative LLM development. By decomposing monolithic models into specialized expert modules coordinated by a trainable gating network, our framework allows diverse contributors to participate regardless of computational resources. We provide a complete technical implementation with mathematical foundations for expert dynamics, gating mechanisms, and integration strategies. Experiments on multiple datasets demonstrate that our approach achieves accuracy improvements of 3-7% over baseline models while reducing computational requirements by 34%. Expert specialization yields significant domain-specific gains, with improvements from 51% to 88% F1 score in general classification and from 23% to 44% accuracy in news categorization. We formalize the routing entropy optimization problem and demonstrate how proper regularization techniques lead to 14% higher expert utilization rates. These results validate MoECollab as an effective approach for democratizing LLM development through architecturally-supported collaboration.", "authors": [" Harshit"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-16", "url": "https://arxiv.org/abs/2503.12592", "pdf_url": "https://arxiv.org/pdf/2503.12592v1", "arxiv_id": "2503.12592", "doi": "10.48550/arXiv.2503.12592", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "528bd37ceb4e0214bf9a023e6f9f8e509866d6f82ee5d92ed913c63ec5bdb6d9", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Mixture of Low-Rank Experts for Robust Audio Spoofing Detection", "abstract": "In audio spoofing detection, most studies rely on clean datasets, making models susceptible to real-world post-processing attacks, such as channel compression and noise. To overcome this challenge, we propose the Adaptive MixtUre Low-rank ExperTs (AMULET) framework, which enhances resilience by leveraging attack-specific knowledge and dynamically adapting to varied attack conditions. Specifically, AMULET employs Attack-Specific Experts (ASEs) fine-tuned with Low-Rank Adaptation (LoRA), allowing each expert to focus on distinct post-processing patterns using just 1.13\\% of the parameters required for full fine-tuning. Furthermore, we introduce Adaptive Expert Fusion (AEF), which adaptively selects and integrates expert knowledge to enhance the robustness of spoofing detection. Experimental results demonstrate that AMULET significantly enhances robustness by improving noise resilience and exhibiting greater adaptability to unseen post-processing methods compared to models trained with full fine-tuning. Additionally, our framework outperforms both single expert and other expert aggregation strategies under various mixed attacks, demonstrating its superior robustness and adaptability in managing complex real-world scenarios.", "authors": ["Qixian Chen", "Yuxiong Xu", "Sara Mandelli", "Sheng Li", "Bin Li"], "categories": ["eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-03-15", "url": "https://arxiv.org/abs/2503.12010", "pdf_url": "https://arxiv.org/pdf/2503.12010v3", "arxiv_id": "2503.12010", "doi": "10.1109/LSP.2025.3595511", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Signal Processing Letters", "quality_score": 0.094}
{"id": "0ef83b6989e96277094eda2d6760971267c9b2c15989c3f32dc83f31379e7519", "sources": ["arxiv", "semantic_scholar"], "title": "Time and Memory Trade-off of KV-Cache Compression in Tensor Transformer Decoding", "abstract": "The key-value (KV) cache in the tensor version of transformers presents a significant bottleneck during inference. While previous work analyzes the fundamental space complexity barriers in standard attention mechanisms [Haris and Onak, 2025], our work generalizes the space complexity barriers result to tensor attention version. Our theoretical contributions rely on a reduction from communication complexity and deduce the memory lower bound for tensor-structured attention mechanisms when $d = Ω(\\log n)$. Furthermore, we introduce two types of tensor attention cache and present a trade-off between time and memory for two scenarios. Overall, our work provides a theoretical foundation for us to understand the time-memory tradeoff of KV-Cache compression in tensor attention decoding and offers more perspectives in developing more memory-efficient tensor attention Transformer architectures.", "authors": ["Yifang Chen", "Xiaoyu Li", "Yingyu Liang", "Zhenmei Shi", "Zhao Song", "Yu Tian"], "categories": ["cs.LG", "cs.AI", "cs.CC", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-14", "url": "https://arxiv.org/abs/2503.11108", "pdf_url": "https://arxiv.org/pdf/2503.11108v2", "arxiv_id": "2503.11108", "doi": "10.48550/arXiv.2503.11108", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "4f5e9ddbf0ab14a14aff50d05a645ef399cae213a1e6a1f1ab1ef3722a0a6300", "sources": ["arxiv", "semantic_scholar"], "title": "Key, Value, Compress: A Systematic Exploration of KV Cache Compression Techniques", "abstract": "Large language models (LLMs) have demonstrated exceptional capabilities in generating text, images, and video content. However, as context length grows, the computational cost of attention increases quadratically with the number of tokens, presenting significant efficiency challenges. This paper presents an analysis of various Key-Value (KV) cache compression strategies, offering a comprehensive taxonomy that categorizes these methods by their underlying principles and implementation techniques. Furthermore, we evaluate their impact on performance and inference latency, providing critical insights into their effectiveness. Our findings highlight the trade-offs involved in KV cache compression and its influence on handling long-context scenarios, paving the way for more efficient LLM implementations.", "authors": ["Neusha Javidnia", "Bita Darvish Rouhani", "Farinaz Koushanfar"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-14", "url": "https://arxiv.org/abs/2503.11816", "pdf_url": "https://arxiv.org/pdf/2503.11816v3", "arxiv_id": "2503.11816", "doi": "10.1109/CICC63670.2025.10983416", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Custom Integrated Circuits Conference", "quality_score": 0.1945}
{"id": "10e12c9bf40ca11a8467db9fa73960bed0a38197a37dbf2be4c936a179ce8240", "sources": ["arxiv", "semantic_scholar"], "title": "X-EcoMLA: Upcycling Pre-Trained Attention into MLA for Efficient and Extreme KV Compression", "abstract": "Multi-head latent attention (MLA) is designed to optimize KV cache memory through low-rank key-value joint compression. Rather than caching keys and values separately, MLA stores their compressed latent representations, reducing memory overhead while maintaining the performance. While MLA improves memory efficiency without compromising language model accuracy, its major limitation lies in its integration during the pre-training phase, requiring models to be trained from scratch. This raises a key question: can we use MLA's benefits fully or partially in models that have already been pre-trained with different attention mechanisms? In this paper, we propose X-EcoMLA to deploy post training distillation to enable the upcycling of Transformer-based attention into an efficient hybrid MLA variant through lightweight post-training adaptation, bypassing the need for extensive pre-training. We demonstrate that leveraging the dark knowledge of a well-trained model can enhance training accuracy and enable extreme KV cache compression in MLA without compromising model performance. The experimental results show that our proposed method can effectively compress the KV cache while preserving the performance on the benchmarks; specifically, for Llama3.2-1B-Instruct baseline, a 6.4x compression achieves the same average score by using only 3.6B training tokens and 70 GPU hours on AMD MI300, whereas a 10.6x compression have less than 0.1% average score drop with 7B training tokens and 140 GPU hours. The code for this work is available at https://github.com/AMD-AGI/AMD-Hybrid-Models.", "authors": ["Guihong Li", "Mehdi Rezagholizadeh", "Mingyu Yang", "Vikram Appia", "Emad Barsoum"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-14", "url": "https://arxiv.org/abs/2503.11132", "pdf_url": "https://arxiv.org/pdf/2503.11132v4", "arxiv_id": "2503.11132", "doi": "10.48550/arXiv.2503.11132", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/AMD-AGI/AMD-Hybrid-Models", "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "282d95691beca0aa6b9b9b83c2f9197e199b2a3f6731e9bb1dfc928ed0ca9594", "sources": ["arxiv", "semantic_scholar"], "title": "MoLEx: Mixture of Layer Experts for Finetuning with Sparse Upcycling", "abstract": "Large-scale pre-training of deep models, followed by fine-tuning them, has become the cornerstone of natural language processing (NLP). The prevalence of data coupled with computational resources has led to large models with a considerable number of parameters. While the massive size of these models has led to remarkable success in many NLP tasks, a detriment is the expense required to retrain all the base model's parameters for the adaptation to each task or domain. Parameter Efficient Fine-Tuning (PEFT) provides an effective solution for this challenge by minimizing the number of parameters required to be fine-tuned while maintaining the quality of the model. While existing methods have achieved impressive results, they mainly focus on adapting a subset of parameters, weight reparameterization, and prompt engineering. In this paper, we study layers as extractors of different types of linguistic information that are valuable when used in conjunction. We then propose the Mixture of Layer Experts (MoLEx), a novel sparse mixture of experts (SMoE) whose experts are layers in the pre-trained model. It performs a conditional computation of a mixture of layers during fine-tuning to provide the model with more structural knowledge about the data. By providing an avenue for information exchange between layers, MoLEx enables the model to make a more well-informed prediction for the downstream task, leading to better fine-tuning results with the same number of effective parameters. As experts can be processed in parallel, MoLEx introduces minimal additional computational overhead. We empirically corroborate the advantages of MoLEx when combined with popular PEFT baseline methods on a variety of downstream fine-tuning tasks, including the popular GLUE benchmark as well as the End-to-End Challenge (E2E). The code is publicly available at https://github.com/rachtsy/molex.", "authors": ["Rachel S. Y. Teo", "Tan M. Nguyen"], "categories": ["cs.CL", "cs.AI", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-14", "url": "https://arxiv.org/abs/2503.11144", "pdf_url": "https://arxiv.org/pdf/2503.11144v1", "arxiv_id": "2503.11144", "doi": "10.48550/arXiv.2503.11144", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/rachtsy/molex", "venue": "International Conference on Learning Representations", "quality_score": 0.2258}
{"id": "8dca46a8b566de6c987183db7c4ef5d88e8a03fa191ba18ae8ec7cdf8807bcde", "sources": ["arxiv", "semantic_scholar"], "title": "FedALT: Federated Fine-Tuning through Adaptive Local Training with Rest-of-World LoRA", "abstract": "Fine-tuning large language models (LLMs) in federated settings enables privacy-preserving adaptation but suffers from cross-client interference due to model aggregation. Existing federated LoRA fine-tuning methods, primarily based on FedAvg, struggle with data heterogeneity, leading to harmful cross-client interference and suboptimal personalization. In this work, we propose \\textbf{FedALT}, a novel personalized federated LoRA fine-tuning algorithm that fundamentally departs from FedAvg. Instead of using an aggregated model to initialize local training, each client continues training its individual LoRA while incorporating shared knowledge through a separate Rest-of-World (RoW) LoRA component. To effectively balance local adaptation and global information, FedALT introduces an adaptive mixer that dynamically learns input-specific weightings between the individual and RoW LoRA components, drawing conceptual foundations from the Mixture-of-Experts (MoE) paradigm. Through extensive experiments on NLP benchmarks, we demonstrate that FedALT significantly outperforms state-of-the-art personalized federated LoRA fine-tuning methods, achieving superior local adaptation without sacrificing computational efficiency.", "authors": ["Jieming Bian", "Lei Wang", "Letian Zhang", "Jie Xu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-14", "url": "https://arxiv.org/abs/2503.11880", "pdf_url": "https://arxiv.org/pdf/2503.11880v3", "arxiv_id": "2503.11880", "doi": "10.48550/arXiv.2503.11880", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2785}
{"id": "351597aa36f59947aa57ad13cd51c6a11c6f759249cfa065c1f9f99218c3f3b8", "sources": ["arxiv", "semantic_scholar"], "title": "ZSMerge: Zero-Shot KV Cache Compression for Memory-Efficient Long-Context LLMs", "abstract": "The linear growth of key-value (KV) cache memory and quadratic computational in attention mechanisms complexity pose significant bottlenecks for large language models (LLMs) in long-context processing. While existing KV cache optimization methods address these challenges through token pruning or feature merging, they often incur irreversible information loss or require costly parameter retraining. To this end, we propose ZSMerge, a dynamic KV cache compression framework designed for efficient cache management, featuring three key operations: (1) fine-grained memory allocation guided by multi-dimensional token importance metrics at head-level granularity, (2) a residual merging mechanism that preserves critical context through compensated attention scoring, and (3) a zero-shot adaptation mechanism compatible with diverse LLM architectures without requiring retraining. ZSMerge significantly enhances memory efficiency and inference speed with negligible performance degradation across LLMs. When applied to LLaMA2-7B, it demonstrates a 20:1 compression ratio for key-value cache retention (reducing memory footprint to 5\\% of baseline) while sustaining comparable generation quality, coupled with triple throughput gains at extreme 54k-token contexts that eliminate out-of-memory failures. The code is available at https://github.com/SusCom-Lab/ZSMerge.", "authors": ["Xin Liu", "Xudong Wang", "Pei Liu", "Guoming Tang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-13", "url": "https://arxiv.org/abs/2503.10714", "pdf_url": "https://arxiv.org/pdf/2503.10714v3", "arxiv_id": "2503.10714", "doi": null, "citation_count": 9, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SusCom-Lab/ZSMerge", "venue": null, "quality_score": 0.25}
{"id": "a7ab578bd9ecd6f9cdb2d32986490b691ef95853762d2dc2cc26671792911f9f", "sources": ["arxiv", "semantic_scholar"], "title": "AttentionRAG: Attention-Guided Context Pruning in Retrieval-Augmented Generation", "abstract": "While RAG demonstrates remarkable capabilities in LLM applications, its effectiveness is hindered by the ever-increasing length of retrieved contexts, which introduces information redundancy and substantial computational overhead. Existing context pruning methods, such as LLMLingua, lack contextual awareness and offer limited flexibility in controlling compression rates, often resulting in either insufficient pruning or excessive information loss. In this paper, we propose AttentionRAG, an attention-guided context pruning method for RAG systems. The core idea of AttentionRAG lies in its attention focus mechanism, which reformulates RAG queries into a next-token prediction paradigm. This mechanism isolates the query's semantic focus to a single token, enabling precise and efficient attention calculation between queries and retrieved contexts. Extensive experiments on LongBench and Babilong benchmarks show that AttentionRAG achieves up to 6.3$\\times$ context compression while outperforming LLMLingua methods by around 10\\% in key metrics.", "authors": ["Yixiong Fang", "Tianran Sun", "Yuling Shi", "Xiaodong Gu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-13", "url": "https://arxiv.org/abs/2503.10720", "pdf_url": "https://arxiv.org/pdf/2503.10720v2", "arxiv_id": "2503.10720", "doi": "10.48550/arXiv.2503.10720", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "aadab589c2f867aba55be65b6d7357a3f13d0181231772b0ba4aa22b6fc590c5", "sources": ["arxiv", "semantic_scholar"], "title": "Automated Tomato Maturity Estimation Using an Optimized Residual Model with Pruning and Quantization Techniques", "abstract": "Tomato maturity plays a pivotal role in optimizing harvest timing and ensuring product quality, but current methods struggle to achieve high accuracy along computational efficiency simultaneously. Existing deep learning approaches, while accurate, are often too computationally demanding for practical use in resource-constrained agricultural settings. In contrast, simpler techniques fail to capture the nuanced features needed for precise classification. This study aims to develop a computationally efficient tomato classification model using the ResNet-18 architecture optimized through transfer learning, pruning, and quantization techniques. Our objective is to address the dual challenge of maintaining high accuracy while enabling real-time performance on low-power edge devices. Then, these models were deployed on an edge device to investigate their performance for tomato maturity classification. The quantized model achieved an accuracy of 97.81%, with an average classification time of 0.000975 seconds per image. The pruned and auto-tuned model also demonstrated significant improvements in deployment metrics, further highlighting the benefits of optimization techniques. These results underscore the potential for a balanced solution that meets the accuracy and efficiency demands of modern agricultural production, paving the way for practical, real-world deployment in resource-limited environments.", "authors": ["Muhammad Waseem", "Chung-Hsuan Huang", "Muhammad Muzzammil Sajjad", "Laraib Haider Naqvi", "Yaqoob Majeed", "Tanzeel Ur Rehman", "Tayyaba Nadeem"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-13", "url": "https://arxiv.org/abs/2503.10940", "pdf_url": "https://arxiv.org/pdf/2503.10940v1", "arxiv_id": "2503.10940", "doi": "10.48550/arXiv.2503.10940", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "1f23c37637f9b5fc4464d25f8cd1a20e94d4476508a8a075b0474d7b97d50d12", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba time series forecasting with uncertainty quantification", "abstract": "State space models, such as Mamba, have recently garnered attention in time series forecasting due to their ability to capture sequence patterns. However, in electricity consumption benchmarks, Mamba forecasts exhibit a mean error of approximately 8\\%. Similarly, in traffic occupancy benchmarks, the mean error reaches 18\\%. This discrepancy leaves us to wonder whether the prediction is simply inaccurate or falls within error given spread in historical data. To address this limitation, we propose a method to quantify the predictive uncertainty of Mamba forecasts. Here, we propose a dual-network framework based on the Mamba architecture for probabilistic forecasting, where one network generates point forecasts while the other estimates predictive uncertainty by modeling variance. We abbreviate our tool, Mamba with probabilistic time series forecasting, as Mamba-ProbTSF and the code for its implementation is available on GitHub (https://github.com/PessoaP/Mamba-ProbTSF). Evaluating this approach on synthetic and real-world benchmark datasets, we find Kullback-Leibler divergence between the learned distributions and the data--which, in the limit of infinite data, should converge to zero if the model correctly captures the underlying probability distribution--reduced to the order of $10^{-3}$ for synthetic data and $10^{-1}$ for real-world benchmark, demonstrating its effectiveness. We find that in both the electricity consumption and traffic occupancy benchmark, the true trajectory stays within the predicted uncertainty interval at the two-sigma level about 95\\% of the time. We end with a consideration of potential limitations, adjustments to improve performance, and considerations for applying this framework to processes for purely or largely stochastic dynamics where the stochastic changes accumulate, as observed for example in pure Brownian motion or molecular dynamics trajectories.", "authors": ["Pedro Pessoa", "Paul Campitelli", "Douglas P. Shepherd", "S. Banu Ozkan", "Steve Pressé"], "categories": ["stat.ML", "cs.LG", "nlin.CD"], "fields_of_study": ["Mathematics", "Computer Science", "Physics", "Medicine"], "published_date": "2025-03-13", "url": "https://arxiv.org/abs/2503.10873", "pdf_url": "https://arxiv.org/pdf/2503.10873v2", "arxiv_id": "2503.10873", "doi": "10.1088/2632-2153/adec3b", "citation_count": 9, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/PessoaP/Mamba-ProbTSF", "venue": null, "quality_score": 0.25}
{"id": "d0a7878652c13cf5e169b302ad81d5c0ce3fc31ab22a4f3d394e442d638d749d", "sources": ["arxiv", "semantic_scholar"], "title": "MILLION: Mastering Long-Context LLM Inference Via Outlier-Immunized KV Product Quantization", "abstract": "Large language models (LLMs) are increasingly utilized for complex tasks requiring longer context lengths, with some models supporting up to 128K or 1M tokens. This trend, however, presents significant challenges in inference speed and memory management. Quantization emerges as a promising approach to address the widening gap between LLM size and memory capacity. However, traditional quantization schemes often yield suboptimal compression results for KV caches due to two key factors: i) On-the-fly quantization and de-quantization, causing significant performance overhead; ii) Prevalence of outliers in KV values, challenging low-bitwidth uniform quantization. To this end, we propose MILLION, a novel quantization framework achieving low-bitwidth KV cache through product quantization. First, we conduct a thorough analysis of KV cache distribution, revealing the limitations of existing quantization schemes. Second, we introduce a non-uniform quantization algorithm based on product quantization, which efficiently compresses data while preserving accuracy. Third, we develop a high-performance GPU inference framework with efficient attention kernel and pipeline design for MILLION that leverages sparse computation and asynchronous quantization, significantly enhancing inference speed. Comprehensive evaluation results demonstrate that MILLION can achieve 4 bits quantization with trivial perplexity and accuracy loss, and achieve 2.09x end-to-end performance gains at 32K context length. Code is released at https://github.com/ZongwuWang/MILLION.", "authors": ["Zongwu Wang", "Peng Xu", "Fangxin Liu", "Yiwei Hu", "Qingxiao Sun", "Gezi Li", "Cheng Li", "Xuan Wang", "Li Jiang", "Haibing Guan"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-12", "url": "https://arxiv.org/abs/2504.03661", "pdf_url": "https://arxiv.org/pdf/2504.03661v2", "arxiv_id": "2504.03661", "doi": "10.1109/DAC63849.2025.11132862", "citation_count": 6, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/ZongwuWang/MILLION", "venue": "Design Automation Conference", "quality_score": 0.2113}
{"id": "14459e69d6b9bd1dbfdd73036c89319c63b788c2313a99aa68193a32b26bee88", "sources": ["arxiv", "semantic_scholar"], "title": "Double-Stage Feature-Level Clustering-Based Mixture of Experts Framework", "abstract": "The Mixture-of-Experts (MoE) model has succeeded in deep learning (DL). However, its complex architecture and advantages over dense models in image classification remain unclear. In previous studies, MoE performance has often been affected by noise and outliers in the input space. Some approaches incorporate input clustering for training MoE models, but most clustering algorithms lack access to labeled data, limiting their effectiveness. This paper introduces the Double-stage Feature-level Clustering and Pseudo-labeling-based Mixture of Experts (DFCP-MoE) framework, which consists of input feature extraction, feature-level clustering, and a computationally efficient pseudo-labeling strategy. This approach reduces the impact of noise and outliers while leveraging a small subset of labeled data to label a large portion of unlabeled inputs. We propose a conditional end-to-end joint training method that improves expert specialization by training the MoE model on well-labeled, clustered inputs. Unlike traditional MoE and dense models, the DFCP-MoE framework effectively captures input space diversity, leading to competitive inference results. We validate our approach on three benchmark datasets for multi-class classification tasks.", "authors": ["Bakary Badjie", "José Cecílio", "António Casimiro"], "categories": ["cs.LG", "cs.AI", "cs.CV", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-12", "url": "https://arxiv.org/abs/2503.09504", "pdf_url": "https://arxiv.org/pdf/2503.09504v1", "arxiv_id": "2503.09504", "doi": "10.48550/arXiv.2503.09504", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "0fa4b5a78be8ac6a8aa0eb5247c3043be2934d5fa859850a792a69057004c3e8", "sources": ["arxiv", "semantic_scholar"], "title": "LLMs Know What to Drop: Self-Attention Guided KV Cache Eviction for Efficient Long-Context Inference", "abstract": "Efficient long-context inference is critical as large language models (LLMs) adopt context windows of ranging from 128K to 1M tokens. However, the growing key-value (KV) cache and the high computational complexity of attention create significant bottlenecks in memory usage and latency. In this paper, we find that attention in diverse long-context tasks exhibits sparsity, and LLMs implicitly \"know\" which tokens can be dropped or evicted at the head level after the pre-filling stage. Based on this insight, we propose Self-Attention Guided Eviction~(SAGE-KV), a simple and effective KV eviction cache method for long-context inference. After prefilling, our method performs a one-time top-k selection at both the token and head levels to compress the KV cache, enabling efficient inference with the reduced cache. Evaluations on LongBench and three long-context LLMs (Llama3.1-8B-Instruct-128k, Llama3-8B-Prolong-512k-Instruct, and Qwen2.5-7B-Instruct-128k) show that SAGE-KV maintains accuracy comparable to full attention while significantly improving efficiency. Specifically, SAGE-KV achieves 4x higher memory efficiency with improved accuracy over the static KV cache selection method StreamLLM, and 2x higher memory efficiency with better accuracy than the dynamic KV cache selection method Quest.", "authors": ["Guangtao Wang", "Shubhangi Upasani", "Chen Wu", "Darshan Gandhi", "Jonathan Li", "Changran Hu", "Bo Li", "Urmish Thakker"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-11", "url": "https://arxiv.org/abs/2503.08879", "pdf_url": "https://arxiv.org/pdf/2503.08879v1", "arxiv_id": "2503.08879", "doi": "10.48550/arXiv.2503.08879", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.294}
{"id": "1596f07205dd948fddb5b42c665758566a47f1322862e975b6319f9d000bf930", "sources": ["arxiv", "semantic_scholar"], "title": "FastCache: Optimizing Multimodal LLM Serving through Lightweight KV-Cache Compression Framework", "abstract": "Multi-modal Large Language Models (MLLMs) serving systems commonly employ KV-cache compression to reduce memory footprint. However, existing compression methods introduce significant processing overhead and queuing delays, particularly in concurrent serving scenarios. We present \\texttt{FastCache}, a novel serving framework that effectively addresses these challenges through two key innovations: (1) a dynamic batching strategy that optimizes request scheduling across prefill, compression, and decode stages, and (2) an efficient KV-cache memory pool mechanism that eliminates memory fragmentation while maintaining high GPU utilization. Our comprehensive experiments on the GQA and MileBench datasets demonstrate that \\texttt{FastCache} achieves up to 19.3$\\times$ reduction in Time-To-First-Token (TTFT) and 12.1$\\times$ improvement in throughput compared to state-of-the-art baselines. The system maintains stable performance under high-concurrency scenarios (up to 40 req/s) while reducing average memory consumption by 20\\%. These results establish \\texttt{FastCache} as an efficient solution for real-world LLM serving systems with KV-cache compression.", "authors": ["Jianian Zhu", "Hang Wu", "Haojie Wang", "Yinghui Li", "Biao Hou", "Ruixuan Li", "Jidong Zhai"], "categories": ["cs.MM", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-11", "url": "https://arxiv.org/abs/2503.08461", "pdf_url": "https://arxiv.org/pdf/2503.08461v1", "arxiv_id": "2503.08461", "doi": "10.48550/arXiv.2503.08461", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "aee1deedb016f0672e5708d467ce67f9ab7d915a9e5cb47902148d73c3c5407d", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-Loco: Mixture of Experts for Multitask Locomotion", "abstract": "We present MoE-Loco, a Mixture of Experts (MoE) framework for multitask locomotion for legged robots. Our method enables a single policy to handle diverse terrains, including bars, pits, stairs, slopes, and baffles, while supporting quadrupedal and bipedal gaits. Using MoE, we mitigate the gradient conflicts that typically arise in multitask reinforcement learning, improving both training efficiency and performance. Our experiments demonstrate that different experts naturally specialize in distinct locomotion behaviors, which can be leveraged for task migration and skill composition. We further validate our approach in both simulation and real-world deployment, showcasing its robustness and adaptability.", "authors": ["Runhan Huang", "Shaoting Zhu", "Yilun Du", "Hang Zhao"], "categories": ["cs.RO", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-11", "url": "https://arxiv.org/abs/2503.08564", "pdf_url": "https://arxiv.org/pdf/2503.08564v2", "arxiv_id": "2503.08564", "doi": "10.1109/IROS60139.2025.11246585", "citation_count": 37, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "quality_score": 0.3949}
{"id": "53224c1d23355dad95bc4e9419fa10f5be1996b3ab174f872dd85191877947d4", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerate 3D Object Detection Models via Zero-Shot Attention Key Pruning", "abstract": "Query-based methods with dense features have demonstrated remarkable success in 3D object detection tasks. However, the computational demands of these models, particularly with large image sizes and multiple transformer layers, pose significant challenges for efficient running on edge devices. Existing pruning and distillation methods either need retraining or are designed for ViT models, which are hard to migrate to 3D detectors. To address this issue, we propose a zero-shot runtime pruning method for transformer decoders in 3D object detection models. The method, termed tgGBC (trim keys gradually Guided By Classification scores), systematically trims keys in transformer modules based on their importance. We expand the classification score to multiply it with the attention map to get the importance score of each key and then prune certain keys after each transformer layer according to their importance scores. Our method achieves a 1.99x speedup in the transformer decoder of the latest ToC3D model, with only a minimal performance loss of less than 1%. Interestingly, for certain models, our method even enhances their performance. Moreover, we deploy 3D detectors with tgGBC on an edge device, further validating the effectiveness of our method. The code can be found at https://github.com/iseri27/tg_gbc.", "authors": ["Lizhen Xu", "Xiuxiu Bai", "Xiaojun Jia", "Jianwu Fang", "Shanmin Pang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-11", "url": "https://arxiv.org/abs/2503.08101", "pdf_url": "https://arxiv.org/pdf/2503.08101v3", "arxiv_id": "2503.08101", "doi": "10.1109/ICCV51701.2025.02143", "citation_count": 2, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/iseri27/tg_gbc", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.1505}
{"id": "f751abc4ace01f2922a39137eff50c257feda533eca6fd0b9b35fb4ea3d4b7e8", "sources": ["arxiv", "semantic_scholar"], "title": "TimeStep Master: Asymmetrical Mixture of Timestep LoRA Experts for Versatile and Efficient Diffusion Models in Vision", "abstract": "Diffusion models have driven the advancement of vision generation over the past years. However, it is often difficult to apply these large models in downstream tasks, due to massive fine-tuning cost. Recently, Low-Rank Adaptation (LoRA) has been applied for efficient tuning of diffusion models. Unfortunately, the capabilities of LoRA-tuned diffusion models are limited, since the same LoRA is used for different timesteps of the diffusion process. To tackle this problem, we introduce a general and concise TimeStep Master (TSM) paradigm with two key fine-tuning stages. In the fostering stage (1-stage), we apply different LoRAs to fine-tune the diffusion model at different timestep intervals. This results in different TimeStep LoRA experts that can effectively capture different noise levels. In the assembling stage (2-stage), we design a novel asymmetrical mixture of TimeStep LoRA experts, via core-context collaboration of experts at multi-scale intervals. For each timestep, we leverage TimeStep LoRA expert within the smallest interval as the core expert without gating, and use experts within the bigger intervals as the context experts with time-dependent gating. Consequently, our TSM can effectively model the noise level via the expert in the finest interval, and adaptively integrate contexts from the experts of other scales, boosting the versatility of diffusion models. To show the effectiveness of our TSM paradigm, we conduct extensive experiments on three typical and popular LoRA-related tasks of diffusion models, including domain adaptation, post-pretraining, and model distillation. Our TSM achieves the state-of-the-art results on all these tasks, throughout various model structures (UNet, DiT and MM-DiT) and visual data modalities (Image, Video), showing its remarkable generalization capacity.", "authors": ["Shaobin Zhuang", "Yiwei Guo", "Yanbo Ding", "Kunchang Li", "Xinyuan Chen", "Yaohui Wang", "Fangyikang Wang", "Ying Zhang", "Chen Li", "Yali Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-10", "url": "https://arxiv.org/abs/2503.07416", "pdf_url": "https://arxiv.org/pdf/2503.07416v1", "arxiv_id": "2503.07416", "doi": "10.48550/arXiv.2503.07416", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.25}
{"id": "20e33b2fc1f1f1723e1a32de69ed919e34229e17c8bae838c523000709073a4c", "sources": ["arxiv", "semantic_scholar"], "title": "SEAP: Training-free Sparse Expert Activation Pruning Unlock the Brainpower of Large Language Models", "abstract": "Large Language Models have achieved remarkable success across various natural language processing tasks, yet their high computational cost during inference remains a major bottleneck. This paper introduces Sparse Expert Activation Pruning (SEAP), a training-free pruning method that selectively retains task-relevant parameters to reduce inference overhead. Inspired by the clustering patterns of hidden states and activations in LLMs, SEAP identifies task-specific expert activation patterns and prunes the model while preserving task performance and enhancing computational efficiency. Experimental results demonstrate that SEAP significantly reduces computational overhead while maintaining competitive accuracy. Notably, at 50% pruning, SEAP surpasses both WandA and FLAP by over 20%, and at 20% pruning, it incurs only a 2.2% performance drop compared to the dense model. These findings highlight SEAP's scalability and effectiveness, making it a promising approach for optimizing large-scale LLMs.", "authors": ["Xun Liang", "Hanyu Wang", "Huayi Lai", "Simin Niu", "Shichao Song", "Jiawei Yang", "Jihao Zhao", "Feiyu Xiong", "Bo Tang", "Zhiyu Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-10", "url": "https://arxiv.org/abs/2503.07605", "pdf_url": "https://arxiv.org/pdf/2503.07605v1", "arxiv_id": "2503.07605", "doi": "10.48550/arXiv.2503.07605", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "779b660ed000ca57cde3e6021a73601681471c832cded9b0f3a185686b96ad85", "sources": ["arxiv", "semantic_scholar"], "title": "ResMoE: Space-efficient Compression of Mixture of Experts LLMs via Residual Restoration", "abstract": "Mixture-of-Experts (MoE) Transformer, the backbone architecture of multiple phenomenal language models, leverages sparsity by activating only a fraction of model parameters for each input token. The sparse structure, while allowing constant time costs, results in space inefficiency: we still need to load all the model parameters during inference. We introduce ResMoE, an innovative MoE approximation framework that utilizes Wasserstein barycenter to extract a common expert (barycenter expert) and approximate the residuals between this barycenter expert and the original ones. ResMoE enhances the space efficiency for inference of large-scale MoE Transformers in a one-shot and data-agnostic manner without retraining while maintaining minimal accuracy loss, thereby paving the way for broader accessibility to large language models. We demonstrate the effectiveness of ResMoE through extensive experiments on Switch Transformer, Mixtral, and DeepSeekMoE models. The results show that ResMoE can reduce the number of parameters in an expert by up to 75% while maintaining comparable performance. The code is available at https://github.com/iDEA-iSAIL-Lab-UIUC/ResMoE.", "authors": ["Mengting Ai", "Tianxin Wei", "Yifan Chen", "Zhichen Zeng", "Ritchie Zhao", "Girish Varatkar", "Bita Darvish Rouhani", "Xianfeng Tang", "Hanghang Tong", "Jingrui He"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-10", "url": "https://arxiv.org/abs/2503.06881", "pdf_url": "https://arxiv.org/pdf/2503.06881v1", "arxiv_id": "2503.06881", "doi": "10.1145/3690624.3709196", "citation_count": 20, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/iDEA-iSAIL-Lab-UIUC/ResMoE", "venue": "Knowledge Discovery and Data Mining", "quality_score": 0.3306}
{"id": "28216b6509d8af594756b82b6e1aaa3594364102aeb694f87ea8de563dc9688e", "sources": ["arxiv", "semantic_scholar"], "title": "A Comprehensive Survey of Mixture-of-Experts: Algorithms, Theory, and Applications", "abstract": "Artificial intelligence (AI) has achieved astonishing successes in many domains, especially with the recent breakthroughs in the development of foundational large models. These large models, leveraging their extensive training data, provide versatile solutions for a wide range of downstream tasks. However, as modern datasets become increasingly diverse and complex, the development of large AI models faces two major challenges: (1) the enormous consumption of computational resources and deployment difficulties, and (2) the difficulty in fitting heterogeneous and complex data, which limits the usability of the models. Mixture of Experts (MoE) models has recently attracted much attention in addressing these challenges, by dynamically selecting and activating the most relevant sub-models to process input data. It has been shown that MoEs can significantly improve model performance and efficiency with fewer resources, particularly excelling in handling large-scale, multimodal data. Given the tremendous potential MoE has demonstrated across various domains, it is urgent to provide a comprehensive summary of recent advancements of MoEs in many important fields. Existing surveys on MoE have their limitations, e.g., being outdated or lacking discussion on certain key areas, and we aim to address these gaps. In this paper, we first introduce the basic design of MoE, including gating functions, expert networks, routing mechanisms, training strategies, and system design. We then explore the algorithm design of MoE in important machine learning paradigms such as continual learning, meta-learning, multi-task learning, and reinforcement learning. Additionally, we summarize theoretical studies aimed at understanding MoE and review its applications in computer vision and natural language processing. Finally, we discuss promising future research directions.", "authors": ["Siyuan Mu", "Sen Lin"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-10", "url": "https://arxiv.org/abs/2503.07137", "pdf_url": "https://arxiv.org/pdf/2503.07137v4", "arxiv_id": "2503.07137", "doi": "10.48550/arXiv.2503.07137", "citation_count": 102, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5032}
{"id": "4708b8eb0ccdb11086c47033fbb82bace570e6b2f28cf4f65eae287d64aec6bf", "sources": ["arxiv", "semantic_scholar"], "title": "MoFE: Mixture of Frozen Experts Architecture", "abstract": "We propose the Mixture of Frozen Experts (MoFE) architecture, which integrates Parameter-efficient Fine-tuning (PEFT) and the Mixture of Experts (MoE) architecture to enhance both training efficiency and model scalability. By freezing the Feed Forward Network (FFN) layers within the MoE framework, MoFE significantly reduces the number of trainable parameters, improving training efficiency while still allowing for effective knowledge transfer from the expert models. This facilitates the creation of models proficient in multiple domains. We conduct experiments to evaluate the trade-offs between performance and efficiency, compare MoFE with other PEFT methodologies, assess the impact of domain expertise in the constituent models, and determine the optimal training strategy. The results show that, although there may be some trade-offs in performance, the efficiency gains are substantial, making MoFE a reasonable solution for real-world, resource-constrained environments.", "authors": ["Jean Seo", "Jaeyoon Kim", "Hyopil Shin"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-09", "url": "https://arxiv.org/abs/2503.06491", "pdf_url": "https://arxiv.org/pdf/2503.06491v1", "arxiv_id": "2503.06491", "doi": "10.48550/arXiv.2503.06491", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.1505}
{"id": "00ffe4589800c9cdf3564ef998f16258962977f93f1a727a28daaf19d5559111", "sources": ["arxiv", "semantic_scholar"], "title": "Gamma: Toward Generic Image Assessment with Mixture of Assessment Experts", "abstract": "Image assessment aims to evaluate the quality and aesthetics of images and has been applied across various scenarios, such as natural and AIGC scenes. Existing methods mostly address these sub-tasks or scenes individually. While some works attempt to develop unified image assessment models, they have struggled to achieve satisfactory performance or cover a broad spectrum of assessment scenarios. In this paper, we present \\textbf{Gamma}, a \\textbf{G}eneric im\\textbf{A}ge assess\\textbf{M}ent model using \\textbf{M}ixture of \\textbf{A}ssessment Experts, which can effectively assess images from diverse scenes through mixed-dataset training. Achieving unified training in image assessment presents significant challenges due to annotation biases across different datasets. To address this issue, we first propose a Mixture of Assessment Experts (MoAE) module, which employs shared and adaptive experts to dynamically learn common and specific knowledge for different datasets, respectively. In addition, we introduce a Scene-based Differential Prompt (SDP) strategy, which uses scene-specific prompts to provide prior knowledge and guidance during the learning process, further boosting adaptation for various scenes. Our Gamma model is trained and evaluated on 12 datasets spanning 6 image assessment scenarios. Extensive experiments show that our unified Gamma outperforms other state-of-the-art mixed-training methods by significant margins while covering more scenes. Codes are available at https://github.com/zht8506/Gamma.", "authors": ["Hantao Zhou", "Rui Yang", "Longxiang Tang", "Guanyi Qin", "Runze Hu", "Xiu Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-09", "url": "https://arxiv.org/abs/2503.06678", "pdf_url": "https://arxiv.org/pdf/2503.06678v2", "arxiv_id": "2503.06678", "doi": "10.1145/3746027.3755699", "citation_count": 5, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/zht8506/Gamma", "venue": "ACM Multimedia", "quality_score": 0.1945}
{"id": "92f5956f80ea026c89bd995cb5a896ee90c6a0a824f027cc8ada9e3804353941", "sources": ["arxiv", "semantic_scholar"], "title": "Swift Hydra: Self-Reinforcing Generative Framework for Anomaly Detection with Multiple Mamba Models", "abstract": "Despite a plethora of anomaly detection models developed over the years, their ability to generalize to unseen anomalies remains an issue, particularly in critical systems. This paper aims to address this challenge by introducing Swift Hydra, a new framework for training an anomaly detection method based on generative AI and reinforcement learning (RL). Through featuring an RL policy that operates on the latent variables of a generative model, the framework synthesizes novel and diverse anomaly samples that are capable of bypassing a detection model. These generated synthetic samples are, in turn, used to augment the detection model, further improving its ability to handle challenging anomalies. Swift Hydra also incorporates Mamba models structured as a Mixture of Experts (MoE) to enable scalable adaptation of the number of Mamba experts based on data complexity, effectively capturing diverse feature distributions without increasing the model's inference time. Empirical evaluations on ADBench benchmark demonstrate that Swift Hydra outperforms other state-of-the-art anomaly detection models while maintaining a relatively short inference time. From these results, our research highlights a new and auspicious paradigm of integrating RL and generative AI for advancing anomaly detection.", "authors": ["Nguyen Do", "Truc Nguyen", "Malik Hassanaly", "Raed Alharbi", "Jung Taek Seo", "My T. Thai"], "categories": ["stat.ML", "cs.AI", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2025-03-09", "url": "https://arxiv.org/abs/2503.06413", "pdf_url": "https://arxiv.org/pdf/2503.06413v2", "arxiv_id": "2503.06413", "doi": "10.48550/arXiv.2503.06413", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.1193}
{"id": "8e87f44736d07a25083a6da7dcbe1f270b62e93993ea5a11fee0544b322e2916", "sources": ["arxiv", "semantic_scholar"], "title": "A Novel Trustworthy Video Summarization Algorithm Through a Mixture of LoRA Experts", "abstract": "With the exponential growth of user-generated content on video-sharing platforms, the challenge of facilitating efficient searching and browsing of videos has garnered significant attention. To enhance users' ability to swiftly locate and review pertinent videos, the creation of concise and informative video summaries has become increasingly important. Video-llama is an effective tool for generating video summarization, but it cannot effectively unify and optimize the modeling of temporal and spatial features and requires a lot of computational resources and time. Therefore, we propose MiLoRA-ViSum to more efficiently capture complex temporal dynamics and spatial relationships inherent in video data and to control the number of parameters for training. By extending traditional Low-Rank Adaptation (LoRA) into a sophisticated mixture-of-experts paradigm, MiLoRA-ViSum incorporates a dual temporal-spatial adaptation mechanism tailored specifically for video summarization tasks. This approach dynamically integrates specialized LoRA experts, each fine-tuned to address distinct temporal or spatial dimensions. Extensive evaluations of the VideoXum and ActivityNet datasets demonstrate that MiLoRA-ViSum achieves the best summarization performance compared to state-of-the-art models, while maintaining significantly lower computational costs. The proposed mixture-of-experts strategy, combined with the dual adaptation mechanism, highlights the model's potential to enhance video summarization capabilities, particularly in large-scale applications requiring both efficiency and precision.", "authors": ["Wenzhuo Du", "Gerun Wang", "Guancheng Chen", "Hang Zhao", "Xin Li", "Jian Gao"], "categories": ["cs.CV", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-08", "url": "https://arxiv.org/abs/2503.06064", "pdf_url": "https://arxiv.org/pdf/2503.06064v1", "arxiv_id": "2503.06064", "doi": "10.48550/arXiv.2503.06064", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Electronics", "quality_score": 0.0859}
{"id": "943e37a7ca780e6fe764c9867094871fe95fb92e4e705abdd1d89caea1568324", "sources": ["arxiv", "semantic_scholar"], "title": "Capacity-Aware Inference: Mitigating the Straggler Effect in Mixture of Experts", "abstract": "The Mixture of Experts (MoE) is an effective architecture for scaling large language models by leveraging sparse expert activation to balance performance and efficiency. However, under expert parallelism, MoE suffers from inference inefficiencies due to imbalanced token-to-expert assignment, where underloaded experts complete computations early but must wait for overloaded experts, leading to global delays. We define this phenomenon as the \\textbf{\\textit{Straggler Effect}}, as the most burdened experts dictate the overall inference latency. To address this, we first propose \\textit{\\textbf{Capacity-Aware Token Drop}}, which enforces expert capacity limits by discarding excess tokens from overloaded experts, effectively reducing load imbalance with minimal performance impact (e.g., $30\\%$ speedup with only $0.9\\%$ degradation on OLMoE). Next, given the presence of low-load experts remaining well below the capacity threshold, we introduce \\textit{\\textbf{Capacity-Aware Expanded Drop}}, which allows tokens to include additional local experts in their candidate set before enforcing strict local capacity constraints, thereby improving load balance and enhancing the utilization of underused experts. Extensive experiments on both language and multimodal MoE models demonstrate the effectiveness of our approach, yielding substantial gains in expert utilization, model performance, and inference efficiency, e.g., applying Expanded Drop to Mixtral-8$\\times$7B-Instruct yields a {0.2\\%} average performance improvement and a {1.85$\\times$} inference speedup. The code is released at: https://github.com/CASE-Lab-UMD/Capacity-Aware-MoE.", "authors": ["Shwai He", "Weilin Cai", "Jiayi Huang", "Ang Li"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-07", "url": "https://arxiv.org/abs/2503.05066", "pdf_url": "https://arxiv.org/pdf/2503.05066v5", "arxiv_id": "2503.05066", "doi": "10.48550/arXiv.2503.05066", "citation_count": 12, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/CASE-Lab-UMD/Capacity-Aware-MoE", "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "c1167ce873a337a0aa457bfa6a49edfadc4c1d633404aba03ac65f244bbcc7ca", "sources": ["arxiv", "semantic_scholar"], "title": "Convergence Rates for Softmax Gating Mixture of Experts", "abstract": "Mixture of experts (MoE) has recently emerged as an effective framework to advance the efficiency and scalability of machine learning models by softly dividing complex tasks among multiple specialized sub-models termed experts. Central to the success of MoE is an adaptive softmax gating mechanism which takes responsibility for determining the relevance of each expert to a given input and then dynamically assigning experts their respective weights. Despite its widespread use in practice, a comprehensive study on the effects of the softmax gating on the MoE has been lacking in the literature. To bridge this gap in this paper, we perform a convergence analysis of parameter estimation and expert estimation under the MoE equipped with the standard softmax gating or its variants, including a dense-to-sparse gating and a hierarchical softmax gating, respectively. Furthermore, our theories also provide useful insights into the design of sample-efficient expert structures. In particular, we demonstrate that it requires polynomially many data points to estimate experts satisfying our proposed \\emph{strong identifiability} condition, namely a commonly used two-layer feed-forward network. In stark contrast, estimating linear experts, which violate the strong identifiability condition, necessitates exponentially many data points as a result of intrinsic parameter interactions expressed in the language of partial differential equations. All the theoretical results are substantiated with a rigorous guarantee.", "authors": ["Huy Nguyen", "Nhat Ho", "Alessandro Rinaldo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-03-05", "url": "https://arxiv.org/abs/2503.03213", "pdf_url": "https://arxiv.org/pdf/2503.03213v1", "arxiv_id": "2503.03213", "doi": "10.1109/TIT.2025.3647061", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Information Theory", "quality_score": 0.1193}
{"id": "fc1b030e7981b282ae38760927958fe7f43f1771807f5106e786aaa0665239e5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts Made Intrinsically Interpretable", "abstract": "Neurons in large language models often exhibit \\emph{polysemanticity}, simultaneously encoding multiple unrelated concepts and obscuring interpretability. Instead of relying on post-hoc methods, we present \\textbf{MoE-X}, a Mixture-of-Experts (MoE) language model designed to be \\emph{intrinsically} interpretable. Our approach is motivated by the observation that, in language models, wider networks with sparse activations are more likely to capture interpretable factors. However, directly training such large sparse networks is computationally prohibitive. MoE architectures offer a scalable alternative by activating only a subset of experts for any given input, inherently aligning with interpretability objectives. In MoE-X, we establish this connection by rewriting the MoE layer as an equivalent sparse, large MLP. This approach enables efficient scaling of the hidden size while maintaining sparsity. To further enhance interpretability, we enforce sparse activation within each expert and redesign the routing mechanism to prioritize experts with the highest activation sparsity. These designs ensure that only the most salient features are routed and processed by the experts. We evaluate MoE-X on chess and natural language tasks, showing that it achieves performance comparable to dense models while significantly improving interpretability. MoE-X achieves a perplexity better than GPT-2, with interpretability surpassing even sparse autoencoder (SAE)-based approaches.", "authors": ["Xingyi Yang", "Constantin Venhoff", "Ashkan Khakzar", "Christian Schroeder de Witt", "Puneet K. Dokania", "Adel Bibi", "Philip Torr"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-05", "url": "https://arxiv.org/abs/2503.07639", "pdf_url": "https://arxiv.org/pdf/2503.07639v1", "arxiv_id": "2503.07639", "doi": "10.48550/arXiv.2503.07639", "citation_count": 21, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3356}
{"id": "3bfb60a1c0efff37443728553eb2e6eaccfaca28741932f201a86a3b7ed69dcd", "sources": ["arxiv", "semantic_scholar"], "title": "Put the Space of LoRA Initialization to the Extreme to Preserve Pre-trained Knowledge", "abstract": "Low-Rank Adaptation (LoRA) is the leading parameter-efficient fine-tuning method for Large Language Models (LLMs), but it still suffers from catastrophic forgetting. Recent work has shown that specialized LoRA initialization can alleviate catastrophic forgetting. There are currently two approaches to LoRA initialization aimed at preventing knowledge forgetting during fine-tuning: (1) making residual weights close to pre-trained weights, and (2) ensuring the space of LoRA initialization is orthogonal to pre-trained knowledge. The former is what current methods strive to achieve, while the importance of the latter is not sufficiently recognized. We find that the space of LoRA initialization is the key to preserving pre-trained knowledge rather than the residual weights. Existing methods like MiLoRA propose making the LoRA initialization space orthogonal to pre-trained weights. However, MiLoRA utilizes the null space of pre-trained weights. Compared to pre-trained weights, the input activations of pre-trained knowledge take into account the parameters of all previous layers as well as the input data, while pre-trained weights only contain information from the current layer. Moreover, we find that the effective ranks of input activations are much smaller than those of pre-trained weights. Thus, the null space of activations is more accurate and contains less pre-trained knowledge information compared to that of weights. Based on these, we introduce LoRA-Null, our proposed method that initializes LoRA in the null space of activations. Experimental results show that LoRA-Null effectively preserves the pre-trained world knowledge of LLMs while achieving good fine-tuning performance, as evidenced by extensive experiments. Code is available at {https://github.com/HungerPWAY/LoRA-Null}.", "authors": ["Pengwei Tang", "Xiaolin Hu", "Yong Liu", "Lizhong Ding", "Dongjie Zhang", "Xing Wu", "Debing Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-04", "url": "https://arxiv.org/abs/2503.02659", "pdf_url": "https://arxiv.org/pdf/2503.02659v2", "arxiv_id": "2503.02659", "doi": "10.1609/aaai.v40i39.40608", "citation_count": 8, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/HungerPWAY/LoRA-Null}", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2386}
{"id": "d54f9ae2ae0ac7185a0b32a5157c2f0f2ad153c4692921a0d5225bd0d5d030f3", "sources": ["arxiv", "semantic_scholar"], "title": "Q-Filters: Leveraging QK Geometry for Efficient KV Cache Compression", "abstract": "Autoregressive language models rely on a Key-Value (KV) Cache, which avoids re-computing past hidden states during generation, making it faster. As model sizes and context lengths grow, the KV Cache becomes a significant memory bottleneck, which calls for compression methods that limit its size during generation. In this paper, we discover surprising properties of Query (Q) and Key (K) vectors that allow us to efficiently approximate attention scores without computing the attention maps. We propose Q-Filters, a training-free KV Cache compression method that filters out less crucial Key-Value pairs based on a single context-agnostic projection. Contrarily to many alternatives, Q-Filters is compatible with FlashAttention, as it does not require direct access to attention weights. Experimental results in long-context settings demonstrate that Q-Filters is competitive with attention-based compression methods such as SnapKV in retrieval tasks while consistently outperforming efficient compression schemes such as Streaming-LLM in generation setups. Notably, Q-Filters achieves a 99% accuracy in the needle-in-a-haystack task with a x32 compression level while reducing the generation perplexity drop by up to 65% in text generation compared to Streaming-LLM.", "authors": ["Nathan Godey", "Alessio Devoto", "Yu Zhao", "Simone Scardapane", "Pasquale Minervini", "Éric de la Clergerie", "Benoît Sagot"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-04", "url": "https://arxiv.org/abs/2503.02812", "pdf_url": "https://arxiv.org/pdf/2503.02812v1", "arxiv_id": "2503.02812", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "b4176d958905de4c96e229097c768bc266965b3315dcdc77af7154ce4274600f", "sources": ["arxiv", "semantic_scholar"], "title": "MX-Font++: Mixture of Heterogeneous Aggregation Experts for Few-shot Font Generation", "abstract": "Few-shot Font Generation (FFG) aims to create new font libraries using limited reference glyphs, with crucial applications in digital accessibility and equity for low-resource languages, especially in multilingual artificial intelligence systems. Although existing methods have shown promising performance, transitioning to unseen characters in low-resource languages remains a significant challenge, especially when font glyphs vary considerably across training sets. MX-Font considers the content of a character from the perspective of a local component, employing a Mixture of Experts (MoE) approach to adaptively extract the component for better transition. However, the lack of a robust feature extractor prevents them from adequately decoupling content and style, leading to sub-optimal generation results. To alleviate these problems, we propose Heterogeneous Aggregation Experts (HAE), a powerful feature extraction expert that helps decouple content and style downstream from being able to aggregate information in channel and spatial dimensions. Additionally, we propose a novel content-style homogeneity loss to enhance the untangling. Extensive experiments on several datasets demonstrate that our MX-Font++ yields superior visual results in FFG and effectively outperforms state-of-the-art methods. Code and data are available at https://github.com/stephensun11/MXFontpp.", "authors": ["Weihang Wang", "Duolin Sun", "Jielei Zhang", "Longwen Gao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-04", "url": "https://arxiv.org/abs/2503.02799", "pdf_url": "https://arxiv.org/pdf/2503.02799v1", "arxiv_id": "2503.02799", "doi": "10.1109/ICASSP49660.2025.10888465", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/stephensun11/MXFontpp", "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.1747}
{"id": "db7153449da1f99a7a68d775abe1405e4ce25da28532a1561a8d4740f7c7cdf5", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba base PKD for efficient knowledge compression", "abstract": "Deep neural networks (DNNs) have remarkably succeeded in various image processing tasks. However, their large size and computational complexity present significant challenges for deploying them in resource-constrained environments. This paper presents an innovative approach for integrating Mamba Architecture within a Progressive Knowledge Distillation (PKD) process to address the challenge of reducing model complexity while maintaining accuracy in image classification tasks. The proposed framework distills a large teacher model into progressively smaller student models, designed using Mamba blocks. Each student model is trained using Selective-State-Space Models (S-SSM) within the Mamba blocks, focusing on important input aspects while reducing computational complexity. The work's preliminary experiments use MNIST and CIFAR-10 as datasets to demonstrate the effectiveness of this approach. For MNIST, the teacher model achieves 98% accuracy. A set of seven student models as a group retained 63% of the teacher's FLOPs, approximating the teacher's performance with 98% accuracy. The weak student used only 1% of the teacher's FLOPs and maintained 72% accuracy. Similarly, for CIFAR-10, the students achieved 1% less accuracy compared to the teacher, with the small student retaining 5% of the teacher's FLOPs to achieve 50% accuracy. These results confirm the flexibility and scalability of Mamba Architecture, which can be integrated into PKD, succeeding in the process of finding students as weak learners. The framework provides a solution for deploying complex neural networks in real-time applications with a reduction in computational cost.", "authors": ["José Medina", "Amnir Hadachi", "Paul Honeine", "Abdelaziz Bensrhair"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.01727", "pdf_url": "https://arxiv.org/pdf/2503.01727v2", "arxiv_id": "2503.01727", "doi": "10.48550/arXiv.2503.01727", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0802}
{"id": "28cdd3d1c94dd8511fd964506dbcbd2fab71ff7776042b7342eb40762a50194d", "sources": ["arxiv", "semantic_scholar"], "title": "EliteKV: Scalable KV Cache Compression via RoPE Frequency Selection and Joint Low-Rank Projection", "abstract": "Rotary Position Embedding (RoPE) enables each attention head to capture multi-frequency information along the sequence dimension and is widely applied in foundation models. However, the nonlinearity introduced by RoPE complicates optimization of the key state in the Key-Value (KV) cache for RoPE-based attention. Existing KV cache compression methods typically store key state before rotation and apply the transformation during decoding, introducing additional computational overhead. This paper introduces EliteKV, a flexible modification framework for RoPE-based models supporting variable KV cache compression ratios. EliteKV first identifies the intrinsic frequency preference of each head using RoPElite, selectively restoring linearity to certain dimensions of key within attention computation. Building on this, joint low-rank compression of key and value enables partial cache sharing. Experimental results show that with minimal uptraining on only $0.6\\%$ of the original training data, RoPE-based models achieve a $75\\%$ reduction in KV cache size while preserving performance within a negligible margin. Furthermore, EliteKV consistently performs well across models of different scales within the same family.", "authors": ["Yuhao Zhou", "Sirui Song", "Boyang Liu", "Zhiheng Xi", "Senjie Jin", "Xiaoran Fan", "Zhihao Zhang", "Wei Li", "Xuanjing Huang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.01586", "pdf_url": "https://arxiv.org/pdf/2503.01586v1", "arxiv_id": "2503.01586", "doi": "10.48550/arXiv.2503.01586", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "7beca9f6615193752ad577020beb7d8d035a82cbfbff8a8f6ae28c359c017c30", "sources": ["arxiv", "semantic_scholar"], "title": "Forgetting Transformer: Softmax Attention with a Forget Gate", "abstract": "An essential component of modern recurrent sequence models is the forget gate. While Transformers do not have an explicit recurrent form, we show that a forget gate can be naturally incorporated into Transformers by down-weighting the unnormalized attention scores in a data-dependent way. We name this attention mechanism Forgetting Attention and the resulting model the Forgetting Transformer (FoX). We show that FoX outperforms the Transformer on long-context language modeling, length extrapolation, and short-context downstream tasks, while performing on par with the Transformer on long-context downstream tasks. Moreover, it is compatible with the FlashAttention algorithm and does not require any positional embeddings. Several analyses, including the needle-in-the-haystack test, show that FoX also retains the Transformer's superior long-context capabilities over recurrent sequence models such as Mamba-2, HGRN2, and DeltaNet. We also introduce a \"Pro\" block design that incorporates some common architectural components in recurrent sequence models and find it significantly improves the performance of both FoX and the Transformer. Our code is available at https://github.com/zhixuan-lin/forgetting-transformer.", "authors": ["Zhixuan Lin", "Evgenii Nikishin", "Xu Owen He", "Aaron Courville"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.02130", "pdf_url": "https://arxiv.org/pdf/2503.02130v2", "arxiv_id": "2503.02130", "doi": "10.48550/arXiv.2503.02130", "citation_count": 44, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/zhixuan-lin/forgetting-transformer", "venue": "International Conference on Learning Representations", "quality_score": 0.4515}
{"id": "a4fda86f9547c8e830baa040f1927e2c1f91953935561c84ea7d687e909ab1ac", "sources": ["arxiv", "semantic_scholar"], "title": "WeightedKV: Attention Scores Weighted Key-Value Cache Merging for Large Language Models", "abstract": "Large Language Models (LLMs) use key-value (KV) cache to reduce redundant computation in autoregressive generation. However, the KV cache size increases linearly during generation, leading to excessive memory usage, especially for long texts. Most KV cache compression methods evict the unimportant KV pairs to maintain a fixed cache size, which leads to the permanent loss of tokens during generation. However, singular value decomposition shows that \\textit{values} do not exhibit a strong low-rank property as \\textit{keys} do, suggesting that information is distributed more evenly across \\textit{values}, in contrast to its more redundant distribution within \\textit{keys}. Therefore, methods that evict both \\textit{keys} and \\textit{values} risk losing crucial information and compromise context integrity, ultimately degrading the output quality. To address this problem, we propose WeightedKV, a novel, training-free approach that discards the \\textit{keys} of less important tokens, while merging their \\textit{values} into neighboring tokens via a convex combination weighted by their average attention scores. In this way, the retained \\textit{keys} serve as anchors that guide the generation process, while the merged \\textit{values} provide a rich contextual backdrop. We assess our method on four widely used language modeling datasets, demonstrating superior performance compared to all baseline methods, particularly with a lower budget ratio.", "authors": ["Jian Yuan", "Ziwei He", "Haoli Bai", "Jingwen Leng", "Bo Jiang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.01330", "pdf_url": "https://arxiv.org/pdf/2503.01330v1", "arxiv_id": "2503.01330", "doi": "10.1109/ICASSP49660.2025.10889583", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.1505}
{"id": "140000793eaa6056d369fa1bcbf912d905dd6223b84133efa42570bdbf695cc5", "sources": ["arxiv", "semantic_scholar"], "title": "DeRS: Towards Extremely Efficient Upcycled Mixture-of-Experts Models", "abstract": "Upcycled Mixture-of-Experts (MoE) models have shown great potential in various tasks by converting the original Feed-Forward Network (FFN) layers in pre-trained dense models into MoE layers. However, these models still suffer from significant parameter inefficiency due to the introduction of multiple experts. In this work, we propose a novel DeRS (Decompose, Replace, and Synthesis) paradigm to overcome this shortcoming, which is motivated by our observations about the unique redundancy mechanisms of upcycled MoE experts. Specifically, DeRS decomposes the experts into one expert-shared base weight and multiple expert-specific delta weights, and subsequently represents these delta weights in lightweight forms. Our proposed DeRS paradigm can be applied to enhance parameter efficiency in two different scenarios, including: 1) DeRS Compression for inference stage, using sparsification or quantization to compress vanilla upcycled MoE models; and 2) DeRS Upcycling for training stage, employing lightweight sparse or low-rank matrixes to efficiently upcycle dense models into MoE models. Extensive experiments across three different tasks show that the proposed methods can achieve extreme parameter efficiency while maintaining the performance for both training and compression of upcycled MoE models.", "authors": ["Yongqi Huang", "Peng Ye", "Chenyu Huang", "Jianjian Cao", "Lin Zhang", "Baopu Li", "Gang Yu", "Tao Chen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.01359", "pdf_url": "https://arxiv.org/pdf/2503.01359v1", "arxiv_id": "2503.01359", "doi": "10.1109/CVPR52734.2025.00940", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.2386}
{"id": "6f51f14d7149fb667564a95f9f96f30759e56b6e275471523d5c75d49169172c", "sources": ["arxiv", "semantic_scholar"], "title": "Dialogue Without Limits: Constant-Sized KV Caches for Extended Responses in LLMs", "abstract": "Autoregressive Transformers rely on Key-Value (KV) caching to accelerate inference. However, the linear growth of the KV cache with context length leads to excessive memory consumption and bandwidth constraints. This bottleneck is particularly problematic in real-time applications -- such as chatbots and interactive assistants -- where low latency and high memory efficiency are critical. Existing methods drop distant tokens or compress states in a lossy manner, sacrificing accuracy by discarding vital context or introducing bias. We propose MorphKV, an inference-time technique that maintains a constant-sized KV cache while preserving accuracy. MorphKV balances long-range dependencies and local coherence during text generation. It eliminates early-token bias while retaining high-fidelity context by adaptively ranking tokens through correlation-aware selection. Unlike heuristic retention or lossy compression, MorphKV iteratively refines the KV cache via lightweight updates guided by attention patterns of recent tokens. This approach captures inter-token correlation with greater accuracy, crucial for tasks like content creation and code generation. Our studies on long-response tasks show 52.9$\\%$ memory savings and 18.2$\\%$ higher accuracy on average compared to state-of-the-art prior works, enabling efficient real-world deployment.", "authors": ["Ravi Ghadia", "Avinash Kumar", "Gaurav Jain", "Prashant Nair", "Poulami Das"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-02", "url": "https://arxiv.org/abs/2503.00979", "pdf_url": "https://arxiv.org/pdf/2503.00979v2", "arxiv_id": "2503.00979", "doi": "10.48550/arXiv.2503.00979", "citation_count": 18, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3197}
{"id": "983ad5667abd5f0e6b1b57743f4507b03ff3f1e018c48f4621017d864ddd28fb", "sources": ["arxiv", "semantic_scholar"], "title": "Streaming Video Question-Answering with In-context Video KV-Cache Retrieval", "abstract": "We propose ReKV, a novel training-free approach that enables efficient streaming video question-answering (StreamingVQA), by seamlessly integrating with existing Video Large Language Models (Video-LLMs). Traditional VideoQA systems struggle with long videos, as they must process entire videos before responding to queries, and repeat this process for each new question. In contrast, our approach analyzes long videos in a streaming manner, allowing for prompt responses as soon as user queries are received. Building on a common Video-LLM, we first incorporate a sliding-window attention mechanism, ensuring that input frames attend to a limited number of preceding frames, thereby reducing computational overhead. To prevent information loss, we store processed video key-value caches (KV-Caches) in RAM and disk, reloading them into GPU memory as needed. Additionally, we introduce a retrieval method that leverages an external retriever or the parameters within Video-LLMs to retrieve only query-relevant KV-Caches, ensuring both efficiency and accuracy in question answering. ReKV enables the separation of video encoding and question-answering across different processes and GPUs, significantly enhancing the efficiency of StreamingVQA. Through comprehensive experimentation, we validate the efficacy and practicality of our approach, which significantly boosts efficiency and enhances applicability over existing VideoQA models.", "authors": ["Shangzhe Di", "Zhelun Yu", "Guanghao Zhang", "Haoyuan Li", "Tao Zhong", "Hao Cheng", "Bolin Li", "Wanggui He", "Fangxun Shu", "Hao Jiang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-01", "url": "https://arxiv.org/abs/2503.00540", "pdf_url": "https://arxiv.org/pdf/2503.00540v1", "arxiv_id": "2503.00540", "doi": "10.48550/arXiv.2503.00540", "citation_count": 80, "influential_citation_count": 26, "has_code": true, "code_url": "https://github.com/Becomebright/ReKV", "venue": "International Conference on Learning Representations", "quality_score": 0.7157}
{"id": "5a883b016a7de0717b2bc4b29f0a6fa4867902997e5f794e0c6d4f41ae5205a2", "sources": ["arxiv", "semantic_scholar"], "title": "Efficiently Editing Mixture-of-Experts Models with Compressed Experts", "abstract": "Mixture-of-Experts (MoE) models have become a key approach for scaling large language models efficiently by activating only a subset of experts during training and inference. Typically, the number of activated experts presents a trade-off: fewer experts reduce computational costs, while more experts improve performance. Recent studies reveal that not all activated experts contribute equally to model performance, with some providing minimal utility, particularly when finetuning pretrained MoE models for specialized downstream tasks. The co-existence of significant and redundant parameters in experts provides us an opportunity to reduce the number of activated experts while maintaining model performance. In this work, we propose the concept of compressed experts, lightweight modules that serve as compact representations of full experts. Our approach preserves the most important experts while replacing other auxiliary activated experts with compressed experts. The reduction of active parameters significantly lowers inference costs while achieving comparable performance. Extensive experiments on models including Phi-MoE and OLMoE demonstrate that compressed experts recover over 90% of full expert performance across various tasks while reducing more than 30% active parameters and saving 20% in inference costs. This approach enables efficient deployment of MoE models in resource-constrained settings and facilitates scaling to larger models with manageable overhead. Our code is available at https://github.com/yifei-he/Compressed-Experts.", "authors": ["Yifei He", "Yang Liu", "Chen Liang", "Hany Hassan Awadalla"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-01", "url": "https://arxiv.org/abs/2503.00634", "pdf_url": "https://arxiv.org/pdf/2503.00634v2", "arxiv_id": "2503.00634", "doi": "10.48550/arXiv.2503.00634", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/yifei-he/Compressed-Experts", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.1945}
{"id": "46d1846c8ea1b98324446e9cebc49a26f982ecbef246bc8f9cce2867ef0ead2f", "sources": ["arxiv", "semantic_scholar"], "title": "Visual Attention Exploration in Vision-Based Mamba Models", "abstract": "State space models (SSMs) have emerged as an efficient alternative to transformer-based models, offering linear complexity that scales better than transformers. One of the latest advances in SSMs, Mamba, introduces a selective scan mechanism that assigns trainable weights to input tokens, effectively mimicking the attention mechanism. Mamba has also been successfully extended to the vision domain by decomposing 2D images into smaller patches and arranging them as 1D sequences. However, it remains unclear how these patches interact with (or attend to) each other in relation to their original 2D spatial location. Additionally, the order used to arrange the patches into a sequence also significantly impacts their attention distribution. To better understand the attention between patches and explore the attention patterns, we introduce a visual analytics tool specifically designed for vision-based Mamba models. This tool enables a deeper understanding of how attention is distributed across patches in different Mamba blocks and how it evolves throughout a Mamba model. Using the tool, we also investigate the impact of different patch-ordering strategies on the learned attention, offering further insights into the model's behavior.", "authors": ["Junpeng Wang", "Chin-Chia Michael Yeh", "Uday Singh Saini", "Mahashweta Das"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-28", "url": "https://arxiv.org/abs/2502.20764", "pdf_url": "https://arxiv.org/pdf/2502.20764v1", "arxiv_id": "2502.20764", "doi": "10.1109/PacificVis64226.2025.00031", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Pacific Visualization Symposium", "quality_score": 0.0768}
{"id": "ff17259b8874b47d7c678783211431dbdf8be106625b31c5a481da3e5d5b2cb3", "sources": ["arxiv", "semantic_scholar"], "title": "CoSMoEs: Compact Sparse Mixture of Experts", "abstract": "Sparse Mixture of Expert (MoE) models are popular foundational architectures at large scale, however, under-explored at smaller sizes. Here, we show how to enable Compact Sparse Mixture of Experts (CoSMoEs) for on-device inference. Specifically, we tackle the three main on-device dimensions: Quality, Memory and Latency. Along the quality axis, we show that in a fair evaluation (removing confounding factors) MoE architectures outperform FLOP-aligned dense models at on-device scale. We introduce weight-decomposed experts, further improving the MoE model performance. Regarding model memory and latency, we significantly improve model offloading efficiency and, in turn, reduce model inference latency.", "authors": ["Patrick Huber", "Akshat Shrivastava", "Ernie Chang", "Chinnadhurai Sankar", "Ahmed Aly", "Adithya Sagar"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-28", "url": "https://arxiv.org/abs/2503.00245", "pdf_url": "https://arxiv.org/pdf/2503.00245v1", "arxiv_id": "2503.00245", "doi": "10.48550/arXiv.2503.00245", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0768}
{"id": "12dba8eda96b1f488130fa2324d11b8877c7d31a64987bcc813e350665a8761e", "sources": ["arxiv", "semantic_scholar"], "title": "K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs", "abstract": "Recent studies have explored combining different LoRAs to jointly generate learned style and content. However, existing methods either fail to effectively preserve both the original subject and style simultaneously or require additional training. In this paper, we argue that the intrinsic properties of LoRA can effectively guide diffusion models in merging learned subject and style. Building on this insight, we propose K-LoRA, a simple yet effective training-free LoRA fusion approach. In each attention layer, K-LoRA compares the Top-K elements in each LoRA to be fused, determining which LoRA to select for optimal fusion. This selection mechanism ensures that the most representative features of both subject and style are retained during the fusion process, effectively balancing their contributions. Experimental results demonstrate that the proposed method effectively integrates the subject and style information learned by the original LoRAs, outperforming state-of-the-art training-based approaches in both qualitative and quantitative results.", "authors": ["Ziheng Ouyang", "Zhen Li", "Qibin Hou"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-25", "url": "https://arxiv.org/abs/2502.18461", "pdf_url": "https://arxiv.org/pdf/2502.18461v2", "arxiv_id": "2502.18461", "doi": "10.1109/CVPR52734.2025.01217", "citation_count": 53, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.4515}
{"id": "f65fd157c9b059d2d2297af15aa2c42b20716051d55134cf4ad9641cd3044113", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Threshold-Free KV Cache Pruning", "abstract": "To reduce memory consumption during LLM inference, prior works have proposed numerous methods that focus on KV cache pruning based on various criteria. While these techniques often accomplish lossless memory reduction on many datasets, they often rely on an under-emphasized condition: a dataset/domain-specific budget size threshold needs to be pre-determined to achieve the optimal performance. However, such input-specific tuning may be considerably limited in real-world scenarios, as open-domain inputs span diverse domains, lengths and difficulty levels, without clear boundaries for pre-tuning. Thus, the dependence of an input-sensitive threshold can be an inherent limitation that may cause large degradation on arbitrary inputs. In this work, we propose a new objective that lifts the threshold constraints for robust KV pruning, calling for \"threshold-free\" methods that automatically adjust budget sizes while ensuring full-cache performance. We then propose a novel method ReFreeKV as the first solution fulfilling this objective, validated by intensive experiments on 13 datasets of diverse context lengths, task types, and model sizes.", "authors": ["Xuanfan Ni", "Liyan Xu", "Chenyang Lyu", "Longyue Wang", "Mo Yu", "Lemao Liu", "Fandong Meng", "Jie Zhou", "Piji Li"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-24", "url": "https://arxiv.org/abs/2502.16886", "pdf_url": "https://arxiv.org/pdf/2502.16886v3", "arxiv_id": "2502.16886", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0459}
{"id": "cdfaa2819124c85b97135890e4c40498267d8641e49157b976c6ec621c1ac9c5", "sources": ["arxiv", "semantic_scholar"], "title": "Make LoRA Great Again: Boosting LoRA with Adaptive Singular Values and Mixture-of-Experts Optimization Alignment", "abstract": "While Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning for Large Language Models (LLMs), its performance often falls short of Full Fine-Tuning (Full FT). Current methods optimize LoRA by initializing with static singular value decomposition (SVD) subsets, leading to suboptimal leveraging of pre-trained knowledge. Another path for improving LoRA is incorporating a Mixture-of-Experts (MoE) architecture. However, weight misalignment and complex gradient dynamics make it challenging to adopt SVD prior to the LoRA MoE architecture. To mitigate these issues, we propose \\underline{G}reat L\\underline{o}R\\underline{A} Mixture-of-Exper\\underline{t} (GOAT), a framework that (1) adaptively integrates relevant priors using an SVD-structured MoE, and (2) aligns optimization with full fine-tuned MoE by deriving a theoretical scaling factor. We demonstrate that proper scaling, without modifying the architecture or training algorithms, boosts LoRA MoE's efficiency and performance. Experiments across 25 datasets, including natural language understanding, commonsense reasoning, image classification, and natural language generation, demonstrate GOAT's state-of-the-art performance, closing the gap with Full FT.", "authors": ["Chenghao Fan", "Zhenyi Lu", "Sichen Liu", "Chengfeng Gu", "Xiaoye Qu", "Wei Wei", "Yu Cheng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-24", "url": "https://arxiv.org/abs/2502.16894", "pdf_url": "https://arxiv.org/pdf/2502.16894v4", "arxiv_id": "2502.16894", "doi": "10.48550/arXiv.2502.16894", "citation_count": 26, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3578}
{"id": "aad428d9c0d2da86af80c5a4a49830b90a41d2641029fb865a72abb5270f32cb", "sources": ["arxiv", "semantic_scholar"], "title": "KV-Edit: Training-Free Image Editing for Precise Background Preservation", "abstract": "Background consistency remains a significant challenge in image editing tasks. Despite extensive developments, existing works still face a trade-off between maintaining similarity to the original image and generating content that aligns with the target. Here, we propose KV-Edit, a training-free approach that uses KV cache in DiTs to maintain background consistency, where background tokens are preserved rather than regenerated, eliminating the need for complex mechanisms or expensive training, ultimately generating new content that seamlessly integrates with the background within user-provided regions. We further explore the memory consumption of the KV cache during editing and optimize the space complexity to $O(1)$ using an inversion-free method. Our approach is compatible with any DiT-based generative model without additional training. Experiments demonstrate that KV-Edit significantly outperforms existing approaches in terms of both background and image quality, even surpassing training-based methods. Project webpage is available at https://xilluill.github.io/projectpages/KV-Edit", "authors": ["Tianrui Zhu", "Shiyi Zhang", "Jiawei Shao", "Yansong Tang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-24", "url": "https://arxiv.org/abs/2502.17363", "pdf_url": "https://arxiv.org/pdf/2502.17363v3", "arxiv_id": "2502.17363", "doi": "10.1109/ICCV51701.2025.01542", "citation_count": 38, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.3978}
{"id": "8867b3f9a0d69ab8980810cfb28ba55d5e16002946ef85350ebbb69a8da8f4b9", "sources": ["arxiv", "semantic_scholar"], "title": "MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference", "abstract": "Long-context Multimodal Large Language Models (MLLMs) that incorporate long text-image and text-video modalities, demand substantial resources as their multimodal Key-Value (KV) caches grow with increasing input lengths, challenging inference efficiency. Existing methods for KV cache compression, in both text-only and multimodal LLMs, have neglected attention density variations across layers, thus often adopting uniform or progressive reduction strategies for layer-wise cache allocation. In this work, we propose MEDA, a dynamic layer-wise KV cache allocation method for efficient multimodal long-context inference. As its core, MEDA utilizes cross-modal attention entropy to determine the KV cache size at each MLLMs layer. Given the dynamically allocated KV cache size at each layer, MEDA also employs a KV pair selection scheme to identify which KV pairs to select and a KV pair merging strategy that merges the selected and non-selected ones to preserve information from the entire context. MEDA achieves up to 72% KV cache memory reduction and 2.82 times faster decoding speed, while maintaining or enhancing performance on various multimodal tasks in long-context settings, including multi-images and long-video scenarios. Our code is released at https://github.com/AIoT-MLSys-Lab/MEDA.", "authors": ["Zhongwei Wan", "Hui Shen", "Xin Wang", "Che Liu", "Zheda Mai", "Mi Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-24", "url": "https://arxiv.org/abs/2502.17599", "pdf_url": "https://arxiv.org/pdf/2502.17599v2", "arxiv_id": "2502.17599", "doi": "10.48550/arXiv.2502.17599", "citation_count": 27, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/AIoT-MLSys-Lab/MEDA", "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.3618}
{"id": "4f681abacf84d81d9c9fcb02d93bcb9d39739f39631260faa63cd6fcb95c4b12", "sources": ["arxiv", "semantic_scholar"], "title": "KVCrush: Key value cache size-reduction using similarity in head-behaviour", "abstract": "Key-value (KV) caching has emerged as a crucial optimization technique for accelerating inference in large language models (LLMs). By allowing the attention operation to scale linearly rather than quadratically with the total sequence length, KV caching significantly enhances generation throughput. However, due to large context lengths in the modern LLMs, the memory footprint of the KV is a huge bottleneck for model deployment directly impacting the model's batch size, hindering its ability to deliver high-throughput. Existing research addresses this challenge using several techniques, such as discarding low-attention tokens, quantization, and matrix approximation which typically lead to a negative impact on the model accuracy. In this paper, We propose KVCrush technology which can be combined with many KV compression technologies to improve the model accuracy at a much smaller memory. KVCrush provides an alternate representation scheme for key-value states, along with a low-overhead token pruning algorithm that accounts for the token distribution in the KV cache, which in turn allows for a a smaller footprint while maintaining the accuracy of the model. Based on our results, KVCrush reduces LongBench KV Cache size by 4x with less than 1% accuracy drop and achieves state-of-the-art average accuracy with minimal overhead, incurring less than 0.5% total inference latency. KVCrush not only outperforms the accuracy of state-of-the-art importance-based token retention schemes but is also compatible with typical practical LLM deployments using KV cache paging schemes such as vLLM and mixed precision quantization.", "authors": ["Gopi Krishna Jha", "Sameh Gobriel", "Liubov Talamanova", "Nilesh Jain"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-24", "url": "https://arxiv.org/abs/2503.00022", "pdf_url": "https://arxiv.org/pdf/2503.00022v2", "arxiv_id": "2503.00022", "doi": "10.48550/arXiv.2503.00022", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0722}
{"id": "5fba36270d4081e99a89ad58c94d0fbdf099b4aaad9a854c475513ba9a6b9fa1", "sources": ["arxiv", "semantic_scholar"], "title": "Automatic Joint Structured Pruning and Quantization for Efficient Neural Network Training and Compression", "abstract": "Structured pruning and quantization are fundamental techniques used to reduce the size of deep neural networks (DNNs) and typically are applied independently. Applying these techniques jointly via co-optimization has the potential to produce smaller, high-quality models. However, existing joint schemes are not widely used because of (1) engineering difficulties (complicated multi-stage processes), (2) black-box optimization (extensive hyperparameter tuning to control the overall compression), and (3) insufficient architecture generalization. To address these limitations, we present the framework GETA, which automatically and efficiently performs joint structured pruning and quantization-aware training on any DNNs. GETA introduces three key innovations: (i) a quantization-aware dependency graph (QADG) that constructs a pruning search space for generic quantization-aware DNN, (ii) a partially projected stochastic gradient method that guarantees layerwise bit constraints are satisfied, and (iii) a new joint learning strategy that incorporates interpretable relationships between pruning and quantization. We present numerical experiments on both convolutional neural networks and transformer architectures that show that our approach achieves competitive (often superior) performance compared to existing joint pruning and quantization methods.", "authors": ["Xiaoyi Qu", "David Aponte", "Colby Banbury", "Daniel P. Robinson", "Tianyu Ding", "Kazuhito Koishida", "Ilya Zharkov", "Tianyi Chen"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-23", "url": "https://arxiv.org/abs/2502.16638", "pdf_url": "https://arxiv.org/pdf/2502.16638v1", "arxiv_id": "2502.16638", "doi": "10.1109/CVPR52734.2025.01419", "citation_count": 20, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.3306}
{"id": "30bd7b89a3b1c191eb36c59bc8750e7b081bbb4863a51c5863721e7900e353f5", "sources": ["arxiv", "semantic_scholar"], "title": "SVDq: 1.25-bit and 410x Key Cache Compression for LLM Attention", "abstract": "For the efficient inference of Large Language Models (LLMs), the effective compression of key-value (KV) cache is essential. Three main types of KV cache compression techniques, namely sparsity, channel compression, and quantization, have been identified. This study presents SVDq, a Singular Value Decomposition (SVD) - based mixed precision quantization method for K cache. Initially, K cache is transformed into latent channels using SVD basis representations. Since the values in latent channels decay rapidly and become negligible after only a few latent channels, our method then incorporates importance-aware quantization and compression for latent channels. This enables the effective allocation of higher precision to more significant channels. Theoretically, we prove that SVDq results in quantization errors (x0.1 or even lower) that are much lower than those of per-channel key quantization in the original space. Our findings based on RULER and LongBench benchmarks demonstrate that SVDq can achieve an equivalent key cache precision as low as 1.25-bit. When combined with key sparsity, it can reach a key compression ratio of up to 410x for attention computation, all while maintaining comparable model performance. Notably, our method is nearly lossless for LongBench datasets. This indicates that SVDq enables high-precision low-bit quantization, providing a more efficient solution for KV cache compression in LLMs.", "authors": ["Hong Yankun", "Li Xing", "Zhen Hui-Ling", "Yu Xianzhi", "Liu Wulong", "Yuan Mingxuan"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-21", "url": "https://arxiv.org/abs/2502.15304", "pdf_url": "https://arxiv.org/pdf/2502.15304v1", "arxiv_id": "2502.15304", "doi": "10.48550/arXiv.2502.15304", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "4eff49d05343b7c0b1f68a407156f5e9e4e1402c0e56c02114b710321c9f1233", "sources": ["arxiv", "semantic_scholar"], "title": "KVLink: Accelerating Large Language Models via Efficient KV Cache Reuse", "abstract": "We describe KVLink, an approach for efficient key-value (KV) cache reuse in large language models (LLMs). In many LLM applications, different inputs can share overlapping context, such as the same retrieved document appearing in multiple queries. However, the LLMs still need to encode the entire context for each query, leading to redundant computation. In this paper, we investigate a new strategy to eliminate such inefficiency, where the KV cache of each document is precomputed independently. During inference, the KV caches of retrieved documents are concatenated, allowing the model to reuse cached representations instead of recomputing them. To mitigate the performance degradation when using KV caches computed independently for each document, KVLink introduces two key techniques: adjusting positional embeddings of the KV cache at inference to match the global position after concatenation, and using trainable special tokens to restore self-attention across independently encoded documents. Experiments across 7 datasets demonstrate that KVLink improves question answering accuracy by an average of 4% over state-of-the-art methods. Furthermore, by leveraging precomputed KV caches, our approach reduces time-to-first-token by up to 96% compared to standard LLM inference, making it a scalable and efficient solution for context reuse. Additionally, KVLink can be combined with KV cache compression to further save cache loading and storage overhead while outperforming the baselines.", "authors": ["Jingbo Yang", "Bairu Hou", "Wei Wei", "Yujia Bao", "Shiyu Chang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-21", "url": "https://arxiv.org/abs/2502.16002", "pdf_url": "https://arxiv.org/pdf/2502.16002v4", "arxiv_id": "2502.16002", "doi": "10.48550/arXiv.2502.16002", "citation_count": 41, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4058}
{"id": "0fbd3aa50c7b9d0bda1e781f429d9032c2594e10968dc8496bcc36e33e8a2bf1", "sources": ["arxiv", "semantic_scholar"], "title": "Optimizing Pre-Training Data Mixtures with Mixtures of Data Expert Models", "abstract": "We propose a method to optimize language model pre-training data mixtures through efficient approximation of the cross-entropy loss corresponding to each candidate mixture via a Mixture of Data Experts (MDE). We use this approximation as a source of additional features in a regression model, trained from observations of model loss for a small number of mixtures. Experiments with Transformer decoder-only language models in the range of 70M to 1B parameters on the SlimPajama dataset show that our method achieves significantly better performance than approaches that train regression models using only the mixture rates as input features. Combining this improved optimization method with an objective that takes into account cross-entropy on end task data leads to superior performance on few-shot downstream evaluations. We also provide theoretical insights on why aggregation of data expert predictions can provide good approximations to model losses for data mixtures.", "authors": ["Lior Belenki", "Alekh Agarwal", "Tianze Shi", "Kristina Toutanova"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-21", "url": "https://arxiv.org/abs/2502.15950", "pdf_url": "https://arxiv.org/pdf/2502.15950v1", "arxiv_id": "2502.15950", "doi": "10.48550/arXiv.2502.15950", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1505}
{"id": "f5ff19edc1593f061e3004775987678305e23cd8e14d8191838c37463bb547f0", "sources": ["arxiv", "semantic_scholar"], "title": "CoKV: Optimizing KV Cache Allocation via Cooperative Game", "abstract": "Large language models (LLMs) have achieved remarkable success on various aspects of human life. However, one of the major challenges in deploying these models is the substantial memory consumption required to store key-value pairs (KV), which imposes significant resource demands. Recent research has focused on KV cache budget allocation, with several approaches proposing head-level budget distribution by evaluating the importance of individual attention heads. These methods, however, assess the importance of heads independently, overlooking their cooperative contributions within the model, which may result in a deviation from their true impact on model performance. In light of this limitation, we propose CoKV, a novel method that models the cooperation between heads in model inference as a cooperative game. By evaluating the contribution of each head within the cooperative game, CoKV can allocate the cache budget more effectively. Extensive experiments show that CoKV achieves state-of-the-art performance on the LongBench benchmark using LLama-3-8B-Instruct and Mistral-7B models.", "authors": ["Qiheng Sun", "Hongwei Zhang", "Haocheng Xia", "Jiayao Zhang", "Jinfei Liu", "Kui Ren"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-21", "url": "https://arxiv.org/abs/2502.17501", "pdf_url": "https://arxiv.org/pdf/2502.17501v1", "arxiv_id": "2502.17501", "doi": "10.48550/arXiv.2502.17501", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "e2cbb2cd204a49a00613e72e94b3e8f56736322dc24b1a8affe8668d117eaf09", "sources": ["arxiv", "semantic_scholar"], "title": "A Stronger Mixture of Low-Rank Experts for Fine-Tuning Foundation Models", "abstract": "In order to streamline the fine-tuning of foundation models, Low-Rank Adapters (LoRAs) have been substantially adopted across various fields, including instruction tuning and domain adaptation. The underlying concept of LoRA involves decomposing a full-rank matrix into the product of two lower-rank matrices, which reduces storage consumption and accelerates the training process. Furthermore, to address the limited expressive capacity of LoRA, the Mixture-of-Expert (MoE) has been introduced for incorporating multiple LoRA adapters. The integration of LoRA experts leads to a visible improvement across several downstream scenes. However, the mixture of LoRAs (MoE-LoRA) still exhibits its low robustness during tuning and inferring. Inspired by the Riemannian Preconditioners which train LoRA as a sub-space projector, we propose a new training strategy for MoE-LoRA, to stabilize and boost its feature learning procedure by multi-space projections. Examinations on SGD and AdamW optimizers demonstrate the effectiveness of our methodology. Source code is available at https://github.com/THUDM/MoELoRA_Riemannian.", "authors": ["Mengyang Sun", "Yihao Wang", "Tao Feng", "Dan Zhang", "Yifan Zhu", "Jie Tang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-20", "url": "https://arxiv.org/abs/2502.15828", "pdf_url": "https://arxiv.org/pdf/2502.15828v1", "arxiv_id": "2502.15828", "doi": "10.48550/arXiv.2502.15828", "citation_count": 9, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/THUDM/MoELoRA_Riemannian", "venue": "International Conference on Machine Learning", "quality_score": 0.25}
{"id": "1a8eb412485ee06b6202cb646da84fbe0e9ec1806e5fb89aab39bbc5e0414dde", "sources": ["arxiv", "semantic_scholar"], "title": "PLPHP: Per-Layer Per-Head Vision Token Pruning for Efficient Large Vision-Language Models", "abstract": "Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across a range of multimodal tasks. However, their inference efficiency is constrained by the large number of visual tokens processed during decoding. To address this challenge, we propose Per-Layer Per-Head Vision Token Pruning (PLPHP), a two-level fine-grained pruning method including Layer-Level Retention Rate Allocation and Head-Level Vision Token Pruning. Motivated by the Vision Token Re-attention phenomenon across decoder layers, we dynamically adjust token retention rates layer by layer. Layers that exhibit stronger attention to visual information preserve more vision tokens, while layers with lower vision attention are aggressively pruned. Furthermore, PLPHP applies pruning at the attention head level, enabling different heads within the same layer to independently retain critical context. Experiments on multiple benchmarks demonstrate that PLPHP delivers an 18% faster decoding speed and reduces the Key-Value Cache (KV Cache) size by over 50%, all at the cost of 0.46% average performance drop, while also achieving notable performance improvements in multi-image tasks. These results highlight the effectiveness of fine-grained token pruning and contribute to advancing the efficiency and scalability of LVLMs. Our source code will be made publicly available.", "authors": ["Yu Meng", "Kaiyuan Li", "Chenran Huang", "Chen Gao", "Xinlei Chen", "Yong Li", "Xiaoping Zhang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-20", "url": "https://arxiv.org/abs/2502.14504", "pdf_url": "https://arxiv.org/pdf/2502.14504v1", "arxiv_id": "2502.14504", "doi": "10.48550/arXiv.2502.14504", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "d0bd89cb61c21221f6042682bd0352ead9104324bbe5dacbdc40a9757f311135", "sources": ["arxiv", "semantic_scholar"], "title": "Quantize What Counts: More for Keys, Less for Values", "abstract": "Large Language Models (LLMs) suffer inference-time memory bottlenecks dominated by the attention Key-Value (KV) cache, which scales with model size and context length. While KV-cache quantization alleviates this cost, bit allocation between keys and values is often tuned heuristically, lacking theoretical grounding and generalizability. This paper proposes two theorems that anchor mixed-precision KV quantization in the intrinsic geometry of Transformer models. First, key projections systematically have larger spectral and Frobenius norms than value matrices, implying higher information density along the key path. Second, for any given memory budget, prioritizing precision for keys over values strictly reduces quantization error and better preserves accuracy. Empirical evaluations across various prominent LLMs and benchmarks show that key-favored allocations (e.g., 4-bit keys, 2-bit values) retain up to 98.3\\% accuracy compared to uniform allocations (e.g., 4-bit for both), while conserving memory. These results transform bit allocation from ad hoc tuning into a theoretically grounded, geometry-driven design principle for efficient LLM inference. Source code is available at https://github.com/mohsenhariri/spectral-kv.", "authors": ["Mohsen Hariri", "Alan Luo", "Weicong Chen", "Shaochen Zhong", "Tianyi Zhang", "Qifan Wang", "Xia Hu", "Xiaotian Han", "Vipin Chaudhary"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-20", "url": "https://arxiv.org/abs/2502.15075", "pdf_url": "https://arxiv.org/pdf/2502.15075v3", "arxiv_id": "2502.15075", "doi": null, "citation_count": 6, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/mohsenhariri/spectral-kv", "venue": "The 64th Annual Meeting of the Association for Computational Linguistics (ACL), 2026", "quality_score": 0.2386}
{"id": "01285cd5fdb2afe90b07fad7d4d8d2db898b52dd2ee46fe140bd471d185ae2e0", "sources": ["arxiv", "semantic_scholar"], "title": "FairKV: Balancing Per-Head KV Cache for Fast Multi-GPU Inference", "abstract": "KV cache techniques in Transformer models aim to reduce redundant computations at the expense of substantially increased memory usage, making KV cache compression an important and popular research topic. Recently, state-of-the-art KV cache compression methods implement imbalanced, per-head allocation algorithms that dynamically adjust the KV cache budget for each attention head, achieving excellent performance in single-GPU scenarios. However, we observe that such imbalanced compression leads to significant load imbalance when deploying multi-GPU inference, as some GPUs become overburdened while others remain underutilized. In this paper, we propose FairKV, a method designed to ensure fair memory usage among attention heads in systems employing imbalanced KV cache compression. The core technique of FairKV is Fair-Copying, which replicates a small subset of memory-intensive attention heads across GPUs using data parallelism to mitigate load imbalance. Our experiments on popular models, including LLaMA 70b and Mistral 24b model, demonstrate that FairKV increases throughput by 1.66x compared to standard tensor parallelism inference. Our code will be released as open source upon acceptance.", "authors": ["Bingzhe Zhao", "Ke Cheng", "Aomufei Yuan", "Yuxuan Tian", "Ruiguang Zhong", "Chengchen Hu", "Tong Yang", "Lian Yu"], "categories": ["cs.DC", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-19", "url": "https://arxiv.org/abs/2502.15804", "pdf_url": "https://arxiv.org/pdf/2502.15804v2", "arxiv_id": "2502.15804", "doi": "10.48550/arXiv.2502.15804", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1027}
{"id": "e06734bfdaf168779249dc1f6345aa98d19a594b68b6ad21c7e1edec4755c234", "sources": ["arxiv", "semantic_scholar"], "title": "RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compression", "abstract": "Transformer-based Large Language Models rely critically on the KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy containing two consecutive stages. In the first stage, it performs coarse-grain permanent KV cache eviction on the input sequence tokens. In the second stage, it adopts a hybrid sparse attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensionality reductions. We show that RocketKV provides a compression ratio of up to 400$\\times$, end-to-end speedup of up to 3.7$\\times$ as well as peak memory reduction of up to 32.6% in the decode phase on an NVIDIA A100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks. We also propose a variant of RocketKV for multi-turn scenarios, which consistently outperforms other existing methods and achieves accuracy nearly on par with an oracle top-k attention scheme. The source code is available here: https://github.com/NVlabs/RocketKV.", "authors": ["Payman Behnam", "Yaosheng Fu", "Ritchie Zhao", "Po-An Tsai", "Zhiding Yu", "Alexey Tumanov"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-19", "url": "https://arxiv.org/abs/2502.14051", "pdf_url": "https://arxiv.org/pdf/2502.14051v3", "arxiv_id": "2502.14051", "doi": "10.48550/arXiv.2502.14051", "citation_count": 28, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/NVlabs/RocketKV", "venue": "International Conference on Machine Learning", "quality_score": 0.3656}
{"id": "7b75d8760a7e5fd9e0e015031526e70f0009e186e047d8af82f58cf46d50811b", "sources": ["arxiv", "semantic_scholar"], "title": "Train Small, Infer Large: Memory-Efficient LoRA Training for Large Language Models", "abstract": "Large Language Models (LLMs) have significantly advanced natural language processing with exceptional task generalization capabilities. Low-Rank Adaption (LoRA) offers a cost-effective fine-tuning solution, freezing the original model parameters and training only lightweight, low-rank adapter matrices. However, the memory footprint of LoRA is largely dominated by the original model parameters. To mitigate this, we propose LoRAM, a memory-efficient LoRA training scheme founded on the intuition that many neurons in over-parameterized LLMs have low training utility but are essential for inference. LoRAM presents a unique twist: it trains on a pruned (small) model to obtain pruned low-rank matrices, which are then recovered and utilized with the original (large) model for inference. Additionally, minimal-cost continual pre-training, performed by the model publishers in advance, aligns the knowledge discrepancy between pruned and original models. Our extensive experiments demonstrate the efficacy of LoRAM across various pruning strategies and downstream tasks. For a model with 70 billion parameters, LoRAM enables training on a GPU with only 20G HBM, replacing an A100-80G GPU for LoRA training and 15 GPUs for full fine-tuning. Specifically, QLoRAM implemented by structured pruning combined with 4-bit quantization, for LLaMA-3.1-70B (LLaMA-2-70B), reduces the parameter storage cost that dominates the memory usage in low-rank matrix training by 15.81$\\times$ (16.95$\\times$), while achieving dominant performance gains over both the original LLaMA-3.1-70B (LLaMA-2-70B) and LoRA-trained LLaMA-3.1-8B (LLaMA-2-13B). Code is available at https://github.com/junzhang-zj/LoRAM.", "authors": ["Jun Zhang", "Jue Wang", "Huan Li", "Lidan Shou", "Ke Chen", "Yang You", "Guiming Xie", "Xuejian Gong", "Kunlong Zhou"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-19", "url": "https://arxiv.org/abs/2502.13533", "pdf_url": "https://arxiv.org/pdf/2502.13533v2", "arxiv_id": "2502.13533", "doi": "10.48550/arXiv.2502.13533", "citation_count": 10, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/junzhang-zj/LoRAM", "venue": "International Conference on Learning Representations", "quality_score": 0.2603}
{"id": "c8eb99ce696f2680fdc20313eccb4971eeab465449bbeed2c1da749d8effa10c", "sources": ["arxiv", "semantic_scholar"], "title": "BaKlaVa -- Budgeted Allocation of KV cache for Long-context Inference", "abstract": "In Large Language Model (LLM) inference, Key-Value (KV) caches (KV-caches) are essential for reducing time complexity. However, they result in a linear increase in GPU memory as the context length grows. While recent work explores KV-cache eviction and compression policies to reduce memory usage, they often consider uniform KV-caches across all attention heads, leading to suboptimal performance. We introduce BaKlaVa, a method to allocate optimal memory for individual KV-caches across the model by estimating the importance of each KV-cache. Our empirical analysis demonstrates that not all KV-caches are equally critical for LLM performance. Using a one-time profiling approach, BaKlaVa assigns optimal memory budgets to each KV-cache. We evaluated our method on LLaMA-3-8B, and Qwen2.5-7B models, achieving up to a 70\\% compression ratio while keeping baseline performance and delivering up to an order-of-magnitude accuracy improvement at higher compression levels.", "authors": ["Ahmed Burak Gulhan", "Krishna Teja Chitty-Venkata", "Murali Emani", "Mahmut Kandemir", "Venkatram Vishwanath"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-18", "url": "https://arxiv.org/abs/2502.13176", "pdf_url": "https://arxiv.org/pdf/2502.13176v2", "arxiv_id": "2502.13176", "doi": "10.48550/arXiv.2502.13176", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0653}
{"id": "95258c5f9c9fd766b95f9b228b4510de2a99b8a8fd9577d2fd87b31d655188b5", "sources": ["arxiv", "semantic_scholar"], "title": "A$^2$ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization", "abstract": "Long context large language models (LLMs) pose significant challenges for efficient serving due to the large memory footprint and high access overhead of KV cache. Retrieval-based KV cache reduction methods can mitigate these challenges, typically by offloading the complete KV cache to CPU and retrieving necessary tokens on demand during inference. However, these methods still suffer from unsatisfactory accuracy degradation and extra retrieval overhead. To address these limitations, this paper proposes A$^2$ATS, a novel retrieval-based KV cache reduction method. A$^2$ATS aims to obtain an accurate approximation of attention scores by applying the vector quantization technique to key states, thereby enabling efficient and precise retrieval of the top-K tokens. First, we propose Windowed Rotary Position Embedding, which decouples the positional dependency from query and key states after position embedding. Then, we propose query-aware vector quantization that optimizes the objective of attention score approximation directly. Finally, we design the heterogeneous inference architecture for KV cache offloading, enabling long context serving with larger batch sizes. Experimental results demonstrate that A$^2$ATS can achieve a lower performance degradation with similar or lower overhead compared to existing methods, thereby increasing long context serving throughput by up to $2.7 \\times$.", "authors": ["Junhui He", "Junna Xing", "Nan Wang", "Rui Xu", "Shangyu Wu", "Peng Zhou", "Qiang Liu", "Chun Jason Xue", "Qingan Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-18", "url": "https://arxiv.org/abs/2502.12665", "pdf_url": "https://arxiv.org/pdf/2502.12665v2", "arxiv_id": "2502.12665", "doi": "10.48550/arXiv.2502.12665", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1747}
{"id": "86b018a4775b526fb8f97627af089b5c26247c475fbde3ce9fb5db57a7715e55", "sources": ["arxiv", "semantic_scholar"], "title": "MoBA: Mixture of Block Attention for Long-Context LLMs", "abstract": "Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.", "authors": ["Enzhe Lu", "Zhejun Jiang", "Jingyuan Liu", "Yulun Du", "Tao Jiang", "Chao Hong", "Shaowei Liu", "Weiran He", "Enming Yuan", "Yuzhi Wang", "Zhiqi Huang", "Huan Yuan", "Suting Xu", "Xinran Xu", "Guokun Lai", "Yanru Chen", "Huabin Zheng", "Junjie Yan", "Jianlin Su", "Yuxin Wu", "Neo Y. Zhang", "Zhilin Yang", "Xinyu Zhou", "Mingxing Zhang", "Jiezhong Qiu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-18", "url": "https://arxiv.org/abs/2502.13189", "pdf_url": "https://arxiv.org/pdf/2502.13189v1", "arxiv_id": "2502.13189", "doi": "10.48550/arXiv.2502.13189", "citation_count": 160, "influential_citation_count": 19, "has_code": true, "code_url": "https://github.com/MoonshotAI/MoBA", "venue": "arXiv.org", "quality_score": 0.6505}
{"id": "ff1a696a17d6dd565af44b0b69618be7dfefce910fe27d9f4ea523179a442250", "sources": ["arxiv", "semantic_scholar"], "title": "Every Expert Matters: Towards Effective Knowledge Distillation for Mixture-of-Experts Language Models", "abstract": "With the emergence of Mixture-of-Experts (MoE), the efficient scaling of model size has accelerated the development of large language models in recent years. However, their high memory requirements prevent their use in resource-constrained environments. While knowledge distillation (KD) has been a proven method for model compression, its application to MoE teacher models remains underexplored. Through our investigation, we discover that non-activated experts in MoE models possess valuable knowledge that benefits student models. We further demonstrate that existing KD methods are not optimal for compressing MoE models, as they fail to leverage this knowledge effectively. To address this, we propose two intuitive MoE-specific KD methods for the first time: Knowledge Augmentation (KA) and Student-Aware Router (SAR), both designed to effectively extract knowledge from all experts. Specifically, KA augments knowledge by sampling experts multiple times, while SAR uses all experts and adjusts the expert weights through router training to provide optimal knowledge. Extensive experiments show that our methods outperform conventional KD methods, demonstrating their effectiveness for MoE teacher models.", "authors": ["Gyeongman Kim", "Gyouk Chu", "Eunho Yang"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-18", "url": "https://arxiv.org/abs/2502.12947", "pdf_url": "https://arxiv.org/pdf/2502.12947v1", "arxiv_id": "2502.12947", "doi": "10.48550/arXiv.2502.12947", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "ed7a6f30a8e9e4007994fe567d4208c84420c1ae17971820437eb2be1f2e889d", "sources": ["arxiv", "semantic_scholar"], "title": "Fate: Fast Edge Inference of Mixture-of-Experts Models via Cross-Layer Gate", "abstract": "Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.", "authors": ["Zhiyuan Fang", "Zicong Hong", "Yuegui Huang", "Yufeng Lyu", "Wuhui Chen", "Yue Yu", "Fan Yu", "Zibin Zheng"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-17", "url": "https://arxiv.org/abs/2502.12224", "pdf_url": "https://arxiv.org/pdf/2502.12224v2", "arxiv_id": "2502.12224", "doi": "10.1145/3774904.3792527", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "3c5af1f2fddc873f54fe9306529c57dfd7509cbfba5a609e2309d32166522ce7", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Tunable Experts -- Behavior Modification of DeepSeek-R1 at Inference Time", "abstract": "We present the Mixture-of-Tunable-Experts (MoTE), a method that extends the Mixture-of-Experts architecture of Large Language Models (LLMs). Without additional training, MoTE enables meaningful and focused behavior changes in LLMs on-the-fly during inference time. By analyzing the digital LLM brain of DeepSeek-R1 using a technique we dub 'functional Token Resonance Imaging' (fTRI) -- inspired by fMRI and using prompts designed to elicit specific behavior (e.g., 'What happened {time}{place}?') -- we empirically identify distinctive experts associated with behaviors like refusal responses. Using MoTE we are able to intervene and control such specific behavior. We switched off the top 10 most refusal-relevant experts (0.07% of R1's 14,848 routed experts), achieving a 52% refusal reduction on sensitive reference prompts without performance degradation on MT-Bench. Random expert deactivation resulted in smaller behavioral shifts with increased noise, whereas forced expert activation led to significantly higher refusal rates. Our approach shares similarities with sparse autoencoders (SAEs) in terms of explainability and steerability. Unlike SAEs, MoTE does not require large training efforts, as within MoEs with a vast number of experts, specialization already emerged naturally during pretraining. Our findings suggest that significant functional mechanisms in Mixture-of-Experts architectures can at least partially be localized in a small number of specific experts, rather than being distributed throughout the model's weights. Expert subgroups can be tuned to trigger significant behavior variations, providing insights into the inner workings of LLMs.", "authors": ["Robert Dahlke", "Henrik Klagges", "Dan Zecha", "Benjamin Merkel", "Sven Rohr", "Fabian Klemm"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-16", "url": "https://arxiv.org/abs/2502.11096", "pdf_url": "https://arxiv.org/pdf/2502.11096v1", "arxiv_id": "2502.11096", "doi": "10.48550/arXiv.2502.11096", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "1c7be8376626809892db3a0babfe87639d42b31556ca71532ba692690e5d4953", "sources": ["arxiv", "semantic_scholar"], "title": "CalibQuant: 1-Bit KV Cache Quantization for Multimodal LLMs", "abstract": "Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance across diverse applications. However, their computational overhead during deployment remains a critical bottleneck. While Key-Value (KV) caching effectively trades memory for computation to enhance inference efficiency, the growing memory footprint from extensive KV caches significantly reduces throughput and restricts prolonged deployment on memory-constrained GPU devices. To address this challenge, we propose CalibQuant, a simple yet highly effective visual quantization strategy that drastically reduces both memory and computational overhead. Specifically, CalibQuant introduces an extreme 1-bit quantization scheme, complemented by novel post-scaling and calibration techniques tailored to the intrinsic patterns of KV caches, thereby ensuring high efficiency without compromising model performance. Leveraging Triton for runtime optimization, we achieve a 10x throughput increase on InternVL models. Our method is designed to be plug-and-play, seamlessly integrating with various existing MLLMs without requiring architectural changes. Extensive experiments confirm that our approach significantly reduces memory usage while maintaining computational efficiency and preserving multimodal capabilities. Codes are available at https://github.com/insuhan/calibquant.", "authors": ["Insu Han", "Zeliang Zhang", "Zhiyuan Wang", "Yifan Zhu", "Susan Liang", "Jiani Liu", "Haiting Lin", "Mingjie Zhao", "Chenliang Xu", "Kun Wan", "Wentian Zhao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-15", "url": "https://arxiv.org/abs/2502.14882", "pdf_url": "https://arxiv.org/pdf/2502.14882v2", "arxiv_id": "2502.14882", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/insuhan/calibquant", "venue": null, "quality_score": 0.1505}
{"id": "d18a58822cf02ba6426274f52cbc1db54879c2b9d627301f8a0a15dcd7e6778e", "sources": ["arxiv", "semantic_scholar"], "title": "Convex envelope method for T, p flash calculations for mixtures with an arbitrary number of components and arbitrary aggregate states", "abstract": "$T, p$ flash calculations determine the correct number of phases at phase equilibrium and their compositions for fixed temperature and pressure. They are essential for chemical process simulation and optimization. The convex envelope method (CEM) is an existing approach that employs the tangent plane criterion to determine liquid phase equilibria for mixtures with an arbitrary number of components without providing the number of phases beforehand. This work extends the CEM to include also vapor and solid phases. Thus, any phase equilibrium of a given mixture with an arbitrary number of components and phases can be calculated over the whole composition space. The CEM results are presented for various vapor-liquid and solid-liquid phase equilibria examples of up to four components. We show how the CEM can be used for parameter fitting of $g^E$-models. As an outlook, we demonstrate how the CEM can be combined with a machine learning-based tool for property prediction to construct phase equilibria.", "authors": ["Quirin Göttl", "Natalie Rosen", "Jakob Burger"], "categories": ["physics.chem-ph"], "fields_of_study": ["Physics", "Computer Science"], "published_date": "2025-02-13", "url": "https://arxiv.org/abs/2502.09402", "pdf_url": "https://arxiv.org/pdf/2502.09402v2", "arxiv_id": "2502.09402", "doi": "10.1016/j.compchemeng.2025.109326", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Computers and Chemical Engineering", "quality_score": 0.1505}
{"id": "56530bb57e1ab1f3cbf633e3080f33cdfe6b164a85448b321529352243b59a5b", "sources": ["arxiv", "semantic_scholar"], "title": "Heterogeneous Mixture of Experts for Remote Sensing Image Super-Resolution", "abstract": "Remote sensing image super-resolution (SR) aims to reconstruct high-resolution remote sensing images from low-resolution inputs, thereby addressing limitations imposed by sensors and imaging conditions. However, the inherent characteristics of remote sensing images, including diverse ground object types and complex details, pose significant challenges to achieving high-quality reconstruction. Existing methods typically employ a uniform structure to process various types of ground objects without distinction, making it difficult to adapt to the complex characteristics of remote sensing images. To address this issue, we introduce a Mixture of Experts (MoE) model and design a set of heterogeneous experts. These experts are organized into multiple expert groups, where experts within each group are homogeneous while being heterogeneous across groups. This design ensures that specialized activation parameters can be employed to handle the diverse and intricate details of ground objects effectively. To better accommodate the heterogeneous experts, we propose a multi-level feature aggregation strategy to guide the routing process. Additionally, we develop a dual-routing mechanism to adaptively select the optimal expert for each pixel. Experiments conducted on the UCMerced and AID datasets demonstrate that our proposed method achieves superior SR reconstruction accuracy compared to state-of-the-art methods. The code will be available at https://github.com/Mr-Bamboo/MFG-HMoE.", "authors": ["Bowen Chen", "Keyan Chen", "Mohan Yang", "Zhengxia Zou", "Zhenwei Shi"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-02-12", "url": "https://arxiv.org/abs/2502.09654", "pdf_url": "https://arxiv.org/pdf/2502.09654v2", "arxiv_id": "2502.09654", "doi": "10.1109/LGRS.2025.3557928", "citation_count": 27, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Mr-Bamboo/MFG-HMoE", "venue": "IEEE Geoscience and Remote Sensing Letters", "quality_score": 0.3618}
{"id": "8b4c8576db45835154b0e9c69f7985dc8de2bc1a1569d3cdbfc5b0916a6e74e5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Message Passing Experts with Routing Entropy Regularization for Node Classification", "abstract": "Graph neural networks (GNNs) have achieved significant progress in graph-based learning tasks, yet their performance often deteriorates when facing heterophilous structures where connected nodes differ substantially in features and labels. To address this limitation, we propose GNNMoE, a novel entropy-driven mixture of message-passing experts framework that enables node-level adaptive representation learning. GNNMoE decomposes message passing into propagation and transformation operations and integrates them through multiple expert networks guided by a hybrid routing mechanism. And a routing entropy regularization dynamically adjusts soft weighting and soft top-$k$ routing, allowing GNNMoE to flexibly adapt to diverse neighborhood contexts. Extensive experiments on twelve benchmark datasets demonstrate that GNNMoE consistently outperforms SOTA node classification methods, while maintaining scalability and interpretability. This work provides a unified and principled approach for achieving fine-grained, personalized node representation learning.", "authors": ["Xuanze Chen", "Jiajun Zhou", "Yadong Li", "Jinsong Chen", "Shanqing Yu", "Qi Xuan"], "categories": ["cs.LG", "cs.SI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-12", "url": "https://arxiv.org/abs/2502.08083", "pdf_url": "https://arxiv.org/pdf/2502.08083v2", "arxiv_id": "2502.08083", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "f76bc4fb6e548f9adb73c8a9e8f5a78650c42008c2858f1f737feb47c1a1c476", "sources": ["arxiv", "semantic_scholar"], "title": "MoHAVE: Mixture of Hierarchical Audio-Visual Experts for Robust Speech Recognition", "abstract": "Audio-visual speech recognition (AVSR) has become critical for enhancing speech recognition in noisy environments by integrating both auditory and visual modalities. However, existing AVSR systems struggle to scale up without compromising computational efficiency. In this study, we introduce MoHAVE (Mixture of Hierarchical Audio-Visual Experts), a novel robust AVSR framework designed to address these scalability constraints. By leveraging a Mixture-of-Experts (MoE) architecture, MoHAVE activates modality-specific expert groups, ensuring dynamic adaptation to various audio-visual inputs with minimal computational overhead. Key contributions of MoHAVE include: (1) a sparse MoE framework that efficiently scales AVSR model capacity, (2) a hierarchical gating mechanism that dynamically utilizes the expert groups based on input context, enhancing adaptability and robustness, and (3) remarkable performance across robust AVSR benchmarks, including LRS3 and MuAViC transcription and translation tasks, setting a new standard for scalable speech recognition systems.", "authors": ["Sungnyun Kim", "Kangwook Jang", "Sangmin Bae", "Sungwoo Cho", "Se-Young Yun"], "categories": ["eess.AS", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-02-11", "url": "https://arxiv.org/abs/2502.10447", "pdf_url": "https://arxiv.org/pdf/2502.10447v2", "arxiv_id": "2502.10447", "doi": "10.48550/arXiv.2502.10447", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1747}
{"id": "ff48cc11ec4ed0ca76a8b75f451b8de00d265781b2908c4ada5433f38a2c19ea", "sources": ["arxiv", "semantic_scholar"], "title": "MoENAS: Mixture-of-Expert based Neural Architecture Search for jointly Accurate, Fair, and Robust Edge Deep Neural Networks", "abstract": "There has been a surge in optimizing edge Deep Neural Networks (DNNs) for accuracy and efficiency using traditional optimization techniques such as pruning, and more recently, employing automatic design methodologies. However, the focus of these design techniques has often overlooked critical metrics such as fairness, robustness, and generalization. As a result, when evaluating SOTA edge DNNs' performance in image classification using the FACET dataset, we found that they exhibit significant accuracy disparities (14.09%) across 10 different skin tones, alongside issues of non-robustness and poor generalizability. In response to these observations, we introduce Mixture-of-Experts-based Neural Architecture Search (MoENAS), an automatic design technique that navigates through a space of mixture of experts to discover accurate, fair, robust, and general edge DNNs. MoENAS improves the accuracy by 4.02% compared to SOTA edge DNNs and reduces the skin tone accuracy disparities from 14.09% to 5.60%, while enhancing robustness by 3.80% and minimizing overfitting to 0.21%, all while keeping model size close to state-of-the-art models average size (+0.4M). With these improvements, MoENAS establishes a new benchmark for edge DNN design, paving the way for the development of more inclusive and robust edge DNNs.", "authors": ["Lotfi Abdelkrim Mecharbat", "Alberto Marchisio", "Muhammad Shafique", "Mohammad M. Ghassemi", "Tuka Alhanai"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-11", "url": "https://arxiv.org/abs/2502.07422", "pdf_url": "https://arxiv.org/pdf/2502.07422v1", "arxiv_id": "2502.07422", "doi": "10.48550/arXiv.2502.07422", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "644a2757f8d1b215cb6650e5ebadd2a83ae552497a73f7b0bc9217d742a0a335", "sources": ["arxiv", "semantic_scholar"], "title": "Training Sparse Mixture Of Experts Text Embedding Models", "abstract": "Transformer-based text embedding models have improved their performance on benchmarks like MIRACL and BEIR by increasing their parameter counts. However, this scaling approach introduces significant deployment challenges, including increased inference latency and memory usage. These challenges are particularly severe in retrieval-augmented generation (RAG) applications, where large models' increased memory requirements constrain dataset ingestion capacity, and their higher latency directly impacts query-time performance. While causal language models have addressed similar efficiency challenges using Mixture of Experts (MoE) architectures, this approach hasn't been successfully adapted to the general text embedding setting. In this paper, we introduce Nomic Embed v2, the first general purpose MoE text embedding model. Our model outperforms models in the same parameter class on both monolingual and multilingual benchmarks while also maintaining competitive performance with models twice its size. We open-source all code, models, and evaluation data to ensure full reproducibility of our training pipeline at \\href{https://github.com/nomic-ai/contrastors}{https://github.com/nomic-ai/contrastors}.", "authors": ["Zach Nussbaum", "Brandon Duderstadt"], "categories": ["cs.CL", "cs.AI", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-11", "url": "https://arxiv.org/abs/2502.07972", "pdf_url": "https://arxiv.org/pdf/2502.07972v3", "arxiv_id": "2502.07972", "doi": null, "citation_count": 29, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/nomic-ai/contrastors}{https://github.com/nomic-ai/contrastors}", "venue": "arXiv.org", "quality_score": 0.3693}
{"id": "df93b10c624c1f524bec43edf5a85220bc826b0feba7d6f1161f921e724566ad", "sources": ["arxiv", "semantic_scholar"], "title": "Online Scheduling for LLM Inference with KV Cache Constraints", "abstract": "Large Language Model (LLM) inference, where a trained model generates text one word at a time in response to user prompts, is a computationally intensive process requiring efficient scheduling to optimize latency and resource utilization. A key challenge in LLM inference is the management of the Key-Value (KV) cache, which reduces redundant computations but introduces memory constraints. In this work, we model LLM inference with KV cache constraints theoretically and propose a novel batching and scheduling algorithm that minimizes inference latency while effectively managing the KV cache's memory. More specifically, we make the following contributions. First, to evaluate the performance of online algorithms for scheduling in LLM inference, we introduce a hindsight optimal benchmark, formulated as an integer program that computes the minimum total inference latency under full future information. Second, we prove that no deterministic online algorithm can achieve a constant competitive ratio when the arrival process is arbitrary. Third, motivated by the computational intractability of solving the integer program at scale, we propose a polynomial-time online scheduling algorithm and show that under certain conditions it can achieve a constant competitive ratio. We also demonstrate our algorithm's strong empirical performance by comparing it to the hindsight optimal in a synthetic dataset. Finally, we conduct empirical evaluations on a real-world public LLM inference dataset, simulating the Llama2-70B model on A100 GPUs, and show that our algorithm significantly outperforms the benchmark algorithms. Overall, our results offer a path toward more sustainable and cost-effective LLM deployment.", "authors": ["Patrick Jaillet", "Jiashuo Jiang", "Konstantina Mellou", "Marco Molinaro", "Chara Podimata", "Zijie Zhou"], "categories": ["cs.LG", "cs.AI", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-02-10", "url": "https://arxiv.org/abs/2502.07115", "pdf_url": "https://arxiv.org/pdf/2502.07115v5", "arxiv_id": "2502.07115", "doi": "10.48550/arXiv.2502.07115", "citation_count": 16, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3076}
{"id": "61f6d2be557e9aadef65da23d466300f257a88728ec19212632d12dd0df2db9f", "sources": ["arxiv", "semantic_scholar"], "title": "MoETuner: Optimized Mixture of Expert Serving with Balanced Expert Placement and Token Routing", "abstract": "Mixture-of-Experts (MoE) model architecture has emerged as a promising solution for scaling transformer models efficiently, offering sparse activation that reduces computational costs while increasing model capacity. However, as MoE models scale, they need to be distributed across GPU devices, thus face critical performance bottlenecks due to their large memory footprint. Expert parallelism distributes experts across GPUs, however, faces key challenges including an unbalanced token routing and expert activation, resulting in communication tail latency and processing inefficiencies. While existing solutions address some of these issues, they fail to resolve the dual challenges of load imbalance and communication skew. The imbalance in token processing load across experts causes uneven processing times on different GPUs, while communication skew between GPUs leads to unbalanced inter-GPU data transfers. These factors degrade the performance of MoE models by increasing tail latency and reducing overall throughput. To address these limitations, we propose an Integer Linear Programming (ILP) formulation to optimize expert placement by jointly considering token load, communication, and computation costs. We exploit the property that there is a token routing dependency across layers, where tokens routed to a specific expert in one layer are likely to be routed to a limited set of experts in the subsequent layer. Our solution, MoETuner, offers an optimal expert-to-GPU assignment that minimizes inter-GPU token routing costs and balances token processing across devices, thereby reducing tail latency and end-to-end execution time. Experimental results demonstrate 9.3% and 17.5% of end-to-end speedups for single-node and multi-node inference respectively, showcasing the potential of our ILP-based optimization for offering expert parallel solutions for next-generation MoEs.", "authors": ["Seokjin Go", "Divya Mahajan"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-10", "url": "https://arxiv.org/abs/2502.06643", "pdf_url": "https://arxiv.org/pdf/2502.06643v1", "arxiv_id": "2502.06643", "doi": "10.48550/arXiv.2502.06643", "citation_count": 27, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3618}
{"id": "bc8cec293f76e242807729867e762f47bc2d86d756357c93608e609628abbc28", "sources": ["arxiv", "semantic_scholar"], "title": "QP-SNN: Quantized and Pruned Spiking Neural Networks", "abstract": "Brain-inspired Spiking Neural Networks (SNNs) leverage sparse spikes to encode information and operate in an asynchronous event-driven manner, offering a highly energy-efficient paradigm for machine intelligence. However, the current SNN community focuses primarily on performance improvement by developing large-scale models, which limits the applicability of SNNs in resource-limited edge devices. In this paper, we propose a hardware-friendly and lightweight SNN, aimed at effectively deploying high-performance SNN in resource-limited scenarios. Specifically, we first develop a baseline model that integrates uniform quantization and structured pruning, called QP-SNN baseline. While this baseline significantly reduces storage demands and computational costs, it suffers from performance decline. To address this, we conduct an in-depth analysis of the challenges in quantization and pruning that lead to performance degradation and propose solutions to enhance the baseline's performance. For weight quantization, we propose a weight rescaling strategy that utilizes bit width more effectively to enhance the model's representation capability. For structured pruning, we propose a novel pruning criterion using the singular value of spatiotemporal spike activities to enable more accurate removal of redundant kernels. Extensive experiments demonstrate that integrating two proposed methods into the baseline allows QP-SNN to achieve state-of-the-art performance and efficiency, underscoring its potential for enhancing SNN deployment in edge intelligence computing.", "authors": ["Wenjie Wei", "Malu Zhang", "Zijian Zhou", "Ammar Belatreche", "Yimeng Shan", "Yu Liang", "Honglin Cao", "Jieyuan Zhang", "Yang Yang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-09", "url": "https://arxiv.org/abs/2502.05905", "pdf_url": "https://arxiv.org/pdf/2502.05905v2", "arxiv_id": "2502.05905", "doi": "10.48550/arXiv.2502.05905", "citation_count": 21, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3356}
{"id": "f62d7021545dcfdfced101bb71663e8bc79eb3fc0ffb16b00df3d2f39f54e8f5", "sources": ["arxiv", "semantic_scholar"], "title": "Continual Adaptation for Autonomous Driving with the Mixture of Progressive Experts Network", "abstract": "Learning-based autonomous driving requires continuous integration of diverse knowledge in complex traffic , yet existing methods exhibit significant limitations in adaptive capabilities. Addressing this gap demands autonomous driving systems that enable continual adaptation through dynamic adjustments to evolving environmental interactions. This underscores the necessity for enhanced continual learning capabilities to improve system adaptability. To address these challenges, the paper introduces a dynamic progressive optimization framework that facilitates adaptation to variations in dynamic environments, achieved by integrating reinforcement learning and supervised learning for data aggregation. Building on this framework, we propose the Mixture of Progressive Experts (MoPE) network. The proposed method selectively activates multiple expert models based on the distinct characteristics of each task and progressively refines the network architecture to facilitate adaptation to new tasks. Simulation results show that the MoPE model outperforms behavior cloning methods, achieving up to a 7.8% performance improvement in intricate urban road environments.", "authors": ["Yixin Cui", "Shuo Yang", "Chi Wan", "Xincheng Li", "Jiaming Xing", "Yuanjian Zhang", "Yanjun Huang", "Hong Chen"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-09", "url": "https://arxiv.org/abs/2502.05943", "pdf_url": "https://arxiv.org/pdf/2502.05943v2", "arxiv_id": "2502.05943", "doi": "10.48550/arXiv.2502.05943", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "db70fd2025d2ff6addfcdae81dbf37ebf0dad74cb445955dbba5e78294723217", "sources": ["arxiv", "semantic_scholar"], "title": "Klotski: Efficient Mixture-of-Expert Inference via Expert-Aware Multi-Batch Pipeline", "abstract": "Mixture of Experts (MoE), with its distinctive sparse structure, enables the scaling of language models up to trillions of parameters without significantly increasing computational costs. However, the substantial parameter size presents a challenge for inference, as the expansion in GPU memory cannot keep pace with the growth in parameters. Although offloading techniques utilise memory from the CPU and disk and parallelise the I/O and computation for efficiency, the computation for each expert in MoE models is often less than the I/O, resulting in numerous bubbles in the pipeline. Therefore, we propose Klotski, an efficient MoE inference engine that significantly reduces pipeline bubbles through a novel expert-aware multi-batch pipeline paradigm. The proposed paradigm uses batch processing to extend the computation time of the current layer to overlap with the loading time of the next layer. Although this idea has been effectively applied to dense models, more batches may activate more experts in the MoE, leading to longer loading times and more bubbles. Thus, unlike traditional approaches, we balance computation and I/O time and minimise bubbles by orchestrating their inference orders based on their heterogeneous computation and I/O requirements and activation patterns under different batch numbers. Moreover, to adapt to different hardware environments and models, we design a constraint-sensitive I/O-compute planner and a correlation-aware expert prefetcher for a schedule that minimises pipeline bubbles. Experimental results demonstrate that Klotski achieves a superior throughput-latency trade-off compared to state-of-the-art techniques, with throughput improvements of up to 85.12x.", "authors": ["Zhiyuan Fang", "Yuegui Huang", "Zicong Hong", "Yufeng Lyu", "Wuhui Chen", "Yue Yu", "Fan Yu", "Zibin Zheng"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-09", "url": "https://arxiv.org/abs/2502.06888", "pdf_url": "https://arxiv.org/pdf/2502.06888v1", "arxiv_id": "2502.06888", "doi": "10.1145/3676641.3716261", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Architectural Support for Programming Languages and Operating Systems", "quality_score": 0.3197}
{"id": "f4cb8a05734dac2b58c704102435388564fd5df0ea29130836ac3571141c4374", "sources": ["arxiv", "semantic_scholar"], "title": "Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient", "abstract": "Mixture of Experts (MoE) architectures have significantly increased computational efficiency in both research and real-world applications of large-scale machine learning models. However, their scalability and efficiency under memory constraints remain relatively underexplored. In this work, we present joint scaling laws for dense and MoE models, incorporating key factors such as the number of active parameters, dataset size, and the number of experts. Our findings provide a principled framework for selecting the optimal MoE configuration under fixed memory and compute budgets. Surprisingly, we show that MoE models can be more memory-efficient than dense models, contradicting conventional wisdom. To derive and validate the theoretical predictions of our scaling laws, we conduct over 280 experiments with up to 2.7B active parameters and up to 5B total parameters. These results offer actionable insights for designing and deploying MoE models in practical large-scale training scenarios.", "authors": ["Jan Ludziejewski", "Maciej Pióro", "Jakub Krajewski", "Maciej Stefaniak", "Michał Krutul", "Jan Małaśnicki", "Marek Cygan", "Piotr Sankowski", "Kamil Adamczewski", "Piotr Miłoś", "Sebastian Jaszczur"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-07", "url": "https://arxiv.org/abs/2502.05172", "pdf_url": "https://arxiv.org/pdf/2502.05172v2", "arxiv_id": "2502.05172", "doi": "10.48550/arXiv.2502.05172", "citation_count": 26, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3578}
{"id": "3df09cc1cfc54f7cbbe15e8f8f872746469fa00378aea5bee37d967f1c9dee5e", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Foundational Models for Dynamical System Reconstruction: Hierarchical Meta-Learning via Mixture of Experts", "abstract": "As foundational models reshape scientific discovery, a bottleneck persists in dynamical system reconstruction (DSR): the ability to learn across system hierarchies. Many meta-learning approaches have been applied successfully to single systems, but falter when confronted with sparse, loosely related datasets requiring multiple hierarchies to be learned. Mixture of Experts (MoE) offers a natural paradigm to address these challenges. Despite their potential, we demonstrate that naive MoEs are inadequate for the nuanced demands of hierarchical DSR, largely due to their gradient descent-based gating update mechanism which leads to slow updates and conflicted routing during training. To overcome this limitation, we introduce MixER: Mixture of Expert Reconstructors, a novel sparse top-1 MoE layer employing a custom gating update algorithm based on $K$-means and least squares. Extensive experiments validate MixER's capabilities, demonstrating efficient training and scalability to systems of up to ten parametric ordinary differential equations. However, our layer underperforms state-of-the-art meta-learners in high-data regimes, particularly when each expert is constrained to process only a fraction of a dataset composed of highly related data points. Further analysis with synthetic and neuroscientific time series suggests that the quality of the contextual representations generated by MixER is closely linked to the presence of hierarchical structure in the data.", "authors": ["Roussel Desmond Nzoyem", "Grant Stevens", "Amarpal Sahota", "David A. W. Barton", "Tom Deakin"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-07", "url": "https://arxiv.org/abs/2502.05335", "pdf_url": "https://arxiv.org/pdf/2502.05335v2", "arxiv_id": "2502.05335", "doi": "10.48550/arXiv.2502.05335", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "681800811b8111617e85ae2e778de378783a417d689a3225cd289db9f6d58921", "sources": ["arxiv", "semantic_scholar"], "title": "KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference", "abstract": "KV cache quantization can improve Large Language Models (LLMs) inference throughput and latency in long contexts and large batch-size scenarios while preserving LLMs effectiveness. However, current methods have three unsolved issues: overlooking layer-wise sensitivity to KV cache quantization, high overhead of online fine-grained decision-making, and low flexibility to different LLMs and constraints. Therefore, we theoretically analyze the inherent correlation of layer-wise transformer attention patterns to KV cache quantization errors and study why key cache is generally more important than value cache for quantization error reduction. We further propose a simple yet effective framework KVTuner to adaptively search for the optimal hardware-friendly layer-wise KV quantization precision pairs for coarse-grained KV cache with multi-objective optimization and directly utilize the offline searched configurations during online inference. To reduce the computational cost of offline calibration, we utilize the intra-layer KV precision pair pruning and inter-layer clustering to reduce the search space. Experimental results show that we can achieve nearly lossless 3.25-bit mixed precision KV cache quantization for LLMs like Llama-3.1-8B-Instruct and 4.0-bit for sensitive models like Qwen2.5-7B-Instruct on mathematical reasoning tasks. The maximum inference throughput can be improved by 21.25\\% compared with KIVI-KV8 quantization over various context lengths. Our code and searched configurations are available at https://github.com/cmd2001/KVTuner.", "authors": ["Xing Li", "Zeyu Xing", "Yiming Li", "Linping Qu", "Hui-Ling Zhen", "Wulong Liu", "Yiwu Yao", "Sinno Jialin Pan", "Mingxuan Yuan"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.04420", "pdf_url": "https://arxiv.org/pdf/2502.04420v5", "arxiv_id": "2502.04420", "doi": "10.48550/arXiv.2502.04420", "citation_count": 30, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/cmd2001/KVTuner", "venue": "International Conference on Machine Learning", "quality_score": 0.3728}
{"id": "f6c629974384131aca34247125f115bd79854a81ffce765807627af3fd0f3e9a", "sources": ["arxiv", "semantic_scholar"], "title": "CriticalKV: Optimizing KV Cache Eviction from an Output Perturbation Perspective", "abstract": "Large language models have revolutionized natural language processing but face significant challenges of high storage and runtime costs, due to the transformer architecture's reliance on self-attention, particularly the large KV cache for long-sequence inference. Recent efforts to reduce KV cache size by pruning less critical entries based on attention weights remain empirical and lack formal grounding. This paper presents a formal study on identifying critical KV cache entries by analyzing attention output perturbation. Our analysis reveals that, beyond attention weights, the value states within KV entries and pretrained parameter matrices are also crucial. Based on this, we propose a perturbation-constrained selection algorithm that optimizes the worst-case output perturbation to identify critical entries. We demonstrate that our algorithm is a universal, plug-and-play enhancement that incurs negligible computational overhead. When integrated with three state-of-the-art cache eviction methods on three distinct LLMs, our algorithm significantly reduces the compression loss by more than \\textit{half} on average across 29 datasets from the Ruler and LongBench benchmarks. Further perturbation analysis, at both the head and layer levels, confirms the principles underlying our effectiveness. This work offers a new, formally grounded perspective to cache eviction , opening promising avenues for future research. The code is publicly available at https://github.com/FFY0/DefensiveKV.", "authors": ["Yuan Feng", "Junlin Lv", "Haoyu Guo", "Yukun Cao", "S Kevin Zhou", "Xike Xie"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.03805", "pdf_url": "https://arxiv.org/pdf/2502.03805v2", "arxiv_id": "2502.03805", "doi": null, "citation_count": 21, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/FFY0/DefensiveKV", "venue": null, "quality_score": 0.3356}
{"id": "402bc382a416770a700eca48f59ee86da229b497adb8b92a9bb1b25a54ae3703", "sources": ["arxiv", "semantic_scholar"], "title": "AttentionPredictor: Temporal Patterns Matter for KV Cache Compression", "abstract": "With the development of large language models (LLMs), efficient inference through Key-Value (KV) cache compression has attracted considerable attention, especially for long-context generation. To compress the KV cache, recent methods identify critical KV tokens through static modeling of attention scores. However, these methods often struggle to accurately determine critical tokens as they neglect the temporal patterns in attention scores, resulting in a noticeable degradation in LLM performance. To address this challenge, we propose AttentionPredictor, which is the first learning-based method to directly predict attention patterns for KV cache compression and critical token identification. Specifically, AttentionPredictor learns a lightweight, unified convolution model to dynamically capture spatiotemporal patterns and predict the next-token attention scores. An appealing feature of AttentionPredictor is that it accurately predicts the attention score and shares the unified prediction model, which consumes negligible memory, among all transformer layers. Moreover, we propose a cross-token critical cache prefetching framework that hides the token estimation time overhead to accelerate the decoding stage. By retaining most of the attention information, AttentionPredictor achieves 13$\\times$ KV cache compression and 5.6$\\times$ speedup in a cache offloading scenario with comparable LLM performance, significantly outperforming the state-of-the-arts. The code is available at https://github.com/MIRALab-USTC/LLM-AttentionPredictor.", "authors": ["Qingyue Yang", "Jie Wang", "Xing Li", "Zhihai Wang", "Chen Chen", "Lei Chen", "Xianzhi Yu", "Wulong Liu", "Jianye Hao", "Mingxuan Yuan", "Bin Li"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.04077", "pdf_url": "https://arxiv.org/pdf/2502.04077v3", "arxiv_id": "2502.04077", "doi": null, "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/MIRALab-USTC/LLM-AttentionPredictor", "venue": null, "quality_score": 0.2113}
{"id": "0ed97bb2ecb49eaa29db9fe646fd6d0ee28cc2accc95c683b5d3b5802fdf6aa9", "sources": ["arxiv", "semantic_scholar"], "title": "Rank Also Matters: Hierarchical Configuration for Mixture of Adapter Experts in LLM Fine-Tuning", "abstract": "Large language models (LLMs) have demonstrated remarkable success across various tasks, accompanied by a continuous increase in their parameter size. Parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), address the challenges of fine-tuning LLMs by significantly reducing the number of trainable parameters. Recent studies have integrated LoRA with Mixture of Experts (MoE) architectures, leveraging multiple adapter experts and gating mechanisms to further improve fine-tuning performance. However, existing approaches primarily focus on adjusting the allocations of adapter experts per layer to optimize the introduced trainable parameter size, while neglecting a critical factor of adapters' rank. To this end, we propose a hierarchical scheme for expert allocation and rank configuration, HILO, which dynamically adjusts the number and rank of adapter experts across layers, matching the varying representational complexity of model layers in adapter-granularity. Extensive experiments on multiple benchmark tasks demonstrate that HILO outperforms existing methods in accuracy while introducing fewer trainable parameters, providing an efficient and practical solution for fine-tuning LLMs.", "authors": ["Peizhuang Cong", "Wenpu Liu", "Wenhan Yu", "Haochen Zhao", "Tong Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.03884", "pdf_url": "https://arxiv.org/pdf/2502.03884v1", "arxiv_id": "2502.03884", "doi": "10.48550/arXiv.2502.03884", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "4917c072a1c726824841e82f02534c1daa57a85836c653ad10d12b418cb3c124", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of neural operator experts for learning boundary conditions and model selection", "abstract": "While Fourier-based neural operators are best suited to learning mappings between functions on periodic domains, several works have introduced techniques for incorporating non trivial boundary conditions. However, all previously introduced methods have restrictions that limit their applicability. In this work, we introduce an alternative approach to imposing boundary conditions inspired by volume penalization from numerical methods and Mixture of Experts (MoE) from machine learning. By introducing competing experts, the approach additionally allows for model selection. To demonstrate the method, we combine a spatially conditioned MoE with the Fourier based, Modal Operator Regression for Physics (MOR-Physics) neural operator and recover a nonlinear operator on a disk and quarter disk. Next, we extract a large eddy simulation (LES) model from direct numerical simulation of channel flow and show the domain decomposition provided by our approach. Finally, we train our LES model with Bayesian variational inference and obtain posterior predictive samples of flow far past the DNS simulation time horizon.", "authors": ["Dwyer Deighan", "Jonas A. Actor", "Ravi G. Patel"], "categories": ["cs.LG", "math.NA", "physics.flu-dyn"], "fields_of_study": ["Computer Science", "Mathematics", "Physics"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.04562", "pdf_url": "https://arxiv.org/pdf/2502.04562v1", "arxiv_id": "2502.04562", "doi": "10.48550/arXiv.2502.04562", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "34945b27884ef4ef7fd735328da26b2598beb14bc81ec907b7f7ee3bc2eb7e15", "sources": ["arxiv", "semantic_scholar"], "title": "QuantSpec: Self-Speculative Decoding with Hierarchical Quantized KV Cache", "abstract": "Large Language Models (LLMs) are increasingly being deployed on edge devices for long-context settings, creating a growing need for fast and efficient long-context inference. In these scenarios, the Key-Value (KV) cache is the primary bottleneck in terms of both GPU memory and latency, as the full KV cache must be loaded for each decoding step. While speculative decoding is a widely accepted technique to accelerate autoregressive decoding, existing methods often struggle to achieve significant speedups due to inefficient KV cache optimization strategies and result in low acceptance rates. To address these challenges, we propose a novel self-speculative decoding framework, QuantSpec, where the draft model shares the architecture of the target model but employs a hierarchical 4-bit quantized KV cache and 4-bit quantized weights for acceleration. QuantSpec maintains high acceptance rates ($>$90%) and reliably provides consistent end-to-end speedups upto $\\sim2.5\\times$, outperforming other self-speculative decoding methods that use sparse KV cache for long-context LLM inference. QuantSpec also reduces the memory requirements by $\\sim 1.3\\times$ compared to these alternatives.", "authors": ["Rishabh Tiwari", "Haocheng Xi", "Aditya Tomar", "Coleman Hooper", "Sehoon Kim", "Maxwell Horton", "Mahyar Najibi", "Michael W. Mahoney", "Kurt Keutzer", "Amir Gholami"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-05", "url": "https://arxiv.org/abs/2502.10424", "pdf_url": "https://arxiv.org/pdf/2502.10424v1", "arxiv_id": "2502.10424", "doi": "10.48550/arXiv.2502.10424", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3197}
{"id": "496d2491fd6ff78460926a240d7b90a3cfd9633165bc2fe2d5febbdfcdf932aa", "sources": ["arxiv", "semantic_scholar"], "title": "RepLoRA: Reparameterizing Low-Rank Adaptation via the Perspective of Mixture of Experts", "abstract": "Low-rank Adaptation (LoRA) has emerged as a powerful method for fine-tuning large-scale foundation models. Despite its popularity, the theoretical understanding of LoRA has remained limited. This paper presents a theoretical analysis of LoRA by examining its connection to the Mixture of Experts models. Under this framework, we show that simple reparameterizations of the LoRA matrices can notably accelerate the low-rank matrix estimation process. In particular, we prove that reparameterization can reduce the data needed to achieve a desired estimation error from an exponential to a polynomial scale. Motivated by this insight, we propose Reparameterized Low-Rank Adaptation (RepLoRA), which incorporates lightweight MLPs to reparameterize the LoRA matrices. Extensive experiments across multiple domains demonstrate that RepLoRA consistently outperforms vanilla LoRA. Notably, with limited data, RepLoRA surpasses LoRA by a margin of up to 40.0% and achieves LoRA's performance with only 30.0% of the training data, highlighting both the theoretical and empirical robustness of our PEFT method.", "authors": ["Tuan Truong", "Chau Nguyen", "Huy Nguyen", "Minh Le", "Trung Le", "Nhat Ho"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-05", "url": "https://arxiv.org/abs/2502.03044", "pdf_url": "https://arxiv.org/pdf/2502.03044v2", "arxiv_id": "2502.03044", "doi": "10.48550/arXiv.2502.03044", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2258}
{"id": "f08ca77ddc3768dcb77334afd657656a3c6dad3998e1778a0c9901bf7941a082", "sources": ["arxiv", "semantic_scholar"], "title": "Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach", "abstract": "Mixture of Experts (MoE) have shown remarkable success in leveraging specialized expert networks for complex machine learning tasks. However, their susceptibility to adversarial attacks presents a critical challenge for deployment in robust applications. This paper addresses the critical question of how to incorporate robustness into MoEs while maintaining high natural accuracy. We begin by analyzing the vulnerability of MoE components, finding that expert networks are notably more susceptible to adversarial attacks than the router. Based on this insight, we propose a targeted robust training technique that integrates a novel loss function to enhance the adversarial robustness of MoE, requiring only the robustification of one additional expert without compromising training or inference efficiency. Building on this, we introduce a dual-model strategy that linearly combines a standard MoE model with our robustified MoE model using a smoothing parameter. This approach allows for flexible control over the robustness-accuracy trade-off. We further provide theoretical foundations by deriving certified robustness bounds for both the single MoE and the dual-model. To push the boundaries of robustness and accuracy, we propose a novel joint training strategy JTDMoE for the dual-model. This joint training enhances both robustness and accuracy beyond what is achievable with separate models. Experimental results on CIFAR-10 and TinyImageNet datasets using ResNet18 and Vision Transformer (ViT) architectures demonstrate the effectiveness of our proposed methods. The code is publicly available at https://github.com/TIML-Group/Robust-MoE-Dual-Model.", "authors": ["Xu Zhang", "Kaidi Xu", "Ziqing Hu", "Ren Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-05", "url": "https://arxiv.org/abs/2502.06832", "pdf_url": "https://arxiv.org/pdf/2502.06832v3", "arxiv_id": "2502.06832", "doi": "10.48550/arXiv.2502.06832", "citation_count": 8, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/TIML-Group/Robust-MoE-Dual-Model", "venue": "International Conference on Machine Learning", "quality_score": 0.2386}
{"id": "b49675c943e50fedc85ff7d9b5ff890d3f265c8951d750a8f479f9dcae1c3a65", "sources": ["arxiv", "semantic_scholar"], "title": "PolarQuant: Quantizing KV Caches with Polar Transformation", "abstract": "Large language models (LLMs) require significant memory to store Key-Value (KV) embeddings in their KV cache, especially when handling long-range contexts. Quantization of these KV embeddings is a common technique to reduce memory consumption. This work introduces PolarQuant, a novel quantization method employing random preconditioning and polar transformation. Our method transforms the KV embeddings into polar coordinates using an efficient recursive algorithm and then quantizes resulting angles. Our key insight is that, after random preconditioning, the angles in the polar representation exhibit a tightly bounded and highly concentrated distribution with an analytically computable form. This nice distribution eliminates the need for explicit normalization, a step required by traditional quantization methods which introduces significant memory overhead because quantization parameters (e.g., zero point and scale) must be stored in full precision per each data block. PolarQuant bypasses this normalization step, enabling substantial memory savings. The long-context evaluation demonstrates that PolarQuant compresses the KV cache by over x4.2 while achieving the best quality scores compared to the state-of-the-art methods.", "authors": ["Insu Han", "Praneeth Kacham", "Amin Karbasi", "Vahab Mirrokni", "Amir Zandieh"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-04", "url": "https://arxiv.org/abs/2502.02617", "pdf_url": "https://arxiv.org/pdf/2502.02617v1", "arxiv_id": "2502.02617", "doi": "10.48550/arXiv.2502.02617", "citation_count": 19, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3253}
{"id": "2939777c985e47fe58ac4561113014a57ebdcc6ecf136f56d4dbe982b4a4e291", "sources": ["arxiv", "semantic_scholar"], "title": "Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning", "abstract": "Leveraging attention sparsity to accelerate long-context large language models (LLMs) has been a hot research topic. However, current algorithms such as sparse attention or key-value (KV) cache compression tend to use a fixed budget, which presents a significant challenge during deployment because it fails to account for the dynamic nature of real-world scenarios, where the optimal balance between accuracy and efficiency can vary greatly. In this paper, we find that borrowing top-$p$ sampling (nucleus sampling) to sparse attention can surprisingly achieve adaptive budgeting. Based on this, we propose Twilight, a framework to bring adaptive sparsity to any existing sparse attention algorithm without sacrificing their accuracy. Empirical results show that Twilight can adaptively prune at most 98% of redundant tokens, leading to $15.4\\times$ acceleration in self-attention operations and $3.9\\times$ acceleration in end-to-end per token latency in long context LLM decoding.", "authors": ["Chaofan Lin", "Jiaming Tang", "Shuo Yang", "Hanshuo Wang", "Tian Tang", "Boyu Tian", "Ion Stoica", "Song Han", "Mingyu Gao"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-04", "url": "https://arxiv.org/abs/2502.02770", "pdf_url": "https://arxiv.org/pdf/2502.02770v5", "arxiv_id": "2502.02770", "doi": "10.48550/arXiv.2502.02770", "citation_count": 34, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3891}
{"id": "c8773793b1525074d72e759320012397452fbd62bac77148b71348784e0bbd20", "sources": ["arxiv", "semantic_scholar"], "title": "Semantic Integrity Matters: Benchmarking and Preserving High-Density Reasoning in KV Cache Compression", "abstract": "While Key-Value (KV) cache compression is essential for efficient LLM inference, current evaluations disproportionately focus on sparse retrieval tasks, potentially masking the degradation of High-Density Reasoning where Chain-of-Thought (CoT) coherence is critical. We introduce KVFundaBench to systematically evaluate this gap, revealing a sharp dichotomy: while retrieval tasks remain robust, reasoning tasks exhibit severe Task-Dependent Degradation under aggressive compression due to disrupted CoT links. Extending our analysis to the DeepSeek-R1 model, we uncover that its specialized attention patterns offer unique insights into the fragility of reasoning chains. Guided by these findings -- specifically the necessity of preserving few-shot examples as indivisible Semantic Units -- we propose ShotKV. This approach explicitly separates prefill and decoding phases to prioritize semantic integrity. Empirical results demonstrate that ShotKV achieves 9%-18% accuracy improvements on long-context generation tasks and effectively generalizes to document QA, all while delivering an 11% latency reduction compared to full cache inference.", "authors": ["Xiang Liu", "Zhenheng Tang", "Hong Chen", "Peijie Dong", "Zeyu Li", "Xiuze Zhou", "Bo Li", "Xuming Hu", "Xiaowen Chu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-04", "url": "https://arxiv.org/abs/2502.01941", "pdf_url": "https://arxiv.org/pdf/2502.01941v4", "arxiv_id": "2502.01941", "doi": null, "citation_count": 18, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3197}
{"id": "9e12c2dcae7e0cf5237e12fc231f6676dae01616962595f7e11d50040381956f", "sources": ["arxiv", "semantic_scholar"], "title": "FastKV: Decoupling of Context Reduction and KV Cache Compression for Prefill-Decoding Acceleration", "abstract": "While large language models (LLMs) excel at handling long-context sequences, they require substantial prefill computation and key-value (KV) cache, which can heavily burden computational efficiency and memory usage in both prefill and decoding stages. Recent works that compress KV caches with prefill acceleration reduce this cost but inadvertently tie the prefill compute reduction to the decoding KV budget. This coupling arises from overlooking the layer-dependent variation of critical context, often leading to accuracy degradation. To address this issue, we introduce FastKV, a KV cache compression framework designed to reduce latency in both prefill and decoding by leveraging the stabilization of token importance in later layers. FastKV performs full-context computation until a Token-Selective Propagation (TSP) layer, which forwards only the most informative tokens to subsequent layers. From these propagated tokens, FastKV independently selects salient KV entries for caching, thereby decoupling KV budget from the prefill compute reduction based on the TSP decision. This independent control of the TSP rate and KV retention rate enables flexible optimization of efficiency and accuracy. Experimental results show that FastKV achieves speedups of up to 1.82$\\times$ in prefill and 2.87$\\times$ in decoding compared to the full-context baseline, while matching the accuracy of the decoding-only baselines. Our code is available at https://github.com/dongwonjo/FastKV.", "authors": ["Dongwon Jo", "Jiwon Song", "Yulhwa Kim", "Jae-Joon Kim"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-03", "url": "https://arxiv.org/abs/2502.01068", "pdf_url": "https://arxiv.org/pdf/2502.01068v7", "arxiv_id": "2502.01068", "doi": null, "citation_count": 6, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/dongwonjo/FastKV", "venue": null, "quality_score": 0.2386}
{"id": "59e1f0636e75bcabf2a41710b55fa07fb1aa82dc04ad47976cc1deaad876c4b5", "sources": ["arxiv", "semantic_scholar"], "title": "CE-LoRA: Computation-Efficient LoRA Fine-Tuning for Language Models", "abstract": "Large Language Models (LLMs) demonstrate exceptional performance across various tasks but demand substantial computational resources even for fine-tuning computation. Although Low-Rank Adaptation (LoRA) significantly alleviates memory consumption during fine-tuning, its impact on computational cost reduction is limited. This paper identifies the computation of activation gradients as the primary bottleneck in LoRA's backward propagation and introduces the Computation-Efficient LoRA (CE-LoRA) algorithm, which enhances computational efficiency while preserving memory efficiency. CE-LoRA leverages two key techniques: Approximated Matrix Multiplication, which replaces dense multiplications of large and complete matrices with sparse multiplications involving only critical rows and columns, and the Double-LoRA technique, which reduces error propagation in activation gradients. Theoretically, CE-LoRA converges at the same rate as LoRA, $ \\mathcal{O}(1/\\sqrt{T}) $, where $T$ is the number of iteartions. Empirical evaluations confirm that CE-LoRA significantly reduces computational costs compared to LoRA without notable performance degradation.", "authors": ["Guanduo Chen", "Yutong He", "Yipeng Hu", "Kun Yuan", "Binhang Yuan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-03", "url": "https://arxiv.org/abs/2502.01378", "pdf_url": "https://arxiv.org/pdf/2502.01378v1", "arxiv_id": "2502.01378", "doi": "10.48550/arXiv.2502.01378", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "064bb09840f671304f752fe2863627c72cf376ab0352adff8dc77f049ba99ab5", "sources": ["arxiv", "semantic_scholar"], "title": "Progressive Binarization with Semi-Structured Pruning for LLMs", "abstract": "Large language models (LLMs) have achieved remarkable progress in natural language processing, but their high computational and memory costs hinder deployment on resource-constrained devices. Binarization represents the most extreme form of quantization, yet binarized models still contain redundancy that can be further removed. Pruning provides a natural way to eliminate such redundancy, but naïve combination with binarization often results in severe performance degradation. In this paper, we propose Progressive Binarization with Semi-Structured Pruning (PBS$^2$P), a novel post-training framework that seamlessly integrates binarization and semi-structured pruning. We first propose Stepwise semi-structured Pruning with Binarization Optimization (SPBO), which progressively introduces sparsity while optimizing binarization parameters to jointly reduce pruning and quantization error, yielding more stable and accurate compression. Additionally, we propose a Coarse-to-Fine Search (CFS) that first allocates pruning ratios and then refines element selection, further enhancing overall performance. Extensive experiments across multiple LLM families show that PBS$^2$P consistently outperforms state-of-the-art (SOTA) binary post-training quantization methods in both perplexity and downstream accuracy. The code and models will be available at https://github.com/XIANGLONGYAN/PBS2P.", "authors": ["Xianglong Yan", "Tianao Zhang", "Zhiteng Li", "Haotong Qin", "Yulun Zhang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-03", "url": "https://arxiv.org/abs/2502.01705", "pdf_url": "https://arxiv.org/pdf/2502.01705v4", "arxiv_id": "2502.01705", "doi": "10.48550/arXiv.2502.01705", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/XIANGLONGYAN/PBS2P", "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "1e4cfdea0f607b20643d008eb36f661f0d8e0092dd2c83dc7c70baed1a7ee4e1", "sources": ["arxiv", "semantic_scholar"], "title": "Sigmoid Self-Attention has Lower Sample Complexity than Softmax Self-Attention: A Mixture-of-Experts Perspective", "abstract": "At the core of the popular Transformer architecture is the self-attention mechanism, which dynamically assigns softmax weights to each input token so that the model can focus on the most salient information. However, the softmax structure slows down the attention computation due to its row-wise nature, and it inherently introduces competition among tokens: as the weight assigned to one token increases, the weights of others decrease. This competitive dynamic may narrow the focus of self-attention to a limited set of features, potentially overlooking other informative characteristics. Recent experimental studies have shown that using the element-wise sigmoid function helps eliminate token competition and reduce the computational overhead. Despite these promising empirical results, a rigorous comparison between sigmoid and softmax self-attention mechanisms remains absent in the literature. This paper closes this gap by theoretically demonstrating that sigmoid self-attention is more sample-efficient than its softmax counterpart. Toward that goal, we represent the self-attention matrix as a mixture of experts and show that ``experts'' in sigmoid self-attention require significantly less data to achieve the same approximation error as those in softmax self-attention.", "authors": ["Fanqi Yan", "Huy Nguyen", "Pedram Akbarian", "Nhat Ho", "Alessandro Rinaldo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-01", "url": "https://arxiv.org/abs/2502.00281", "pdf_url": "https://arxiv.org/pdf/2502.00281v2", "arxiv_id": "2502.00281", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "7bd9852df209bc448000ffdd45a1eded2cb02d642f3b40527c4faae9c64790eb", "sources": ["arxiv", "semantic_scholar"], "title": "PolarQuant: Leveraging Polar Transformation for Efficient Key Cache Quantization and Decoding Acceleration", "abstract": "The KV cache in large language models is a dominant factor in memory usage, limiting their broader applicability. Quantizing the cache to lower bit widths is an effective way to reduce computational costs; however, previous methods struggle with quantizing key vectors due to outliers, resulting in excessive overhead. We propose a novel quantization approach called PolarQuant, which efficiently addresses the outlier challenge. We observe that outliers typically appear in only one of two dimensions, which are rotated together by a specific angle when rotary position embeddings are applied. When represented as two-dimensional vectors, these dimensions exhibit well-structured patterns, with radii and angles smoothly distributed in polar coordinates. This alleviates the challenge of outliers on per-channel quantization, making them well-suited for quantization. Thus, PolarQuant divides key vectors into groups of two-dimensional sub-vectors, encoding them as the corresponding quantized radius and the polar angle, rather than quantizing original key vectors directly. PolarQuant achieves the superior efficiency in KV cache quantization and accelerates the decoding process by turning the query-key inner product into a table lookup, all while maintaining the downstream performance of full-precision models.", "authors": ["Songhao Wu", "Ang Lv", "Xiao Feng", "Yufei Zhang", "Xun Zhang", "Guojun Yin", "Wei Lin", "Rui Yan"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-01", "url": "https://arxiv.org/abs/2502.00527", "pdf_url": "https://arxiv.org/pdf/2502.00527v2", "arxiv_id": "2502.00527", "doi": "10.48550/arXiv.2502.00527", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "346f4a159d84f9622e39f91eb04cb6449aeef3342efac517b6d1721228437102", "sources": ["arxiv", "semantic_scholar"], "title": "ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference", "abstract": "Large Language Models (LLMs) require significant GPU memory when processing long texts, with the key value (KV) cache consuming up to 70\\% of total memory during inference. Although existing compression methods reduce memory by evaluating the importance of individual tokens, they overlook critical semantic relationships between tokens, resulting in fragmented context and degraded performance. We introduce ChunkKV, which fundamentally reimagines KV cache compression by treating semantic chunks - rather than isolated tokens - as basic compression units. This approach preserves complete linguistic structures and contextual integrity, ensuring that essential meaning is retained even under aggressive compression. Our innovation includes a novel layer-wise index reuse technique that exploits the higher cross-layer similarity of preserved indices in ChunkKV, reducing computational overhead and improving throughput by 26.5\\%. Comprehensive evaluations on challenging benchmarks: LongBench, Needle-In-A-HayStack, GSM8K, and JailbreakV demonstrate that ChunkKV outperforms state-of-the-art methods by up to 8.7\\% in precision while maintaining the same compression ratio. These results confirm that semantic-aware compression significantly enhances both efficiency and performance for long-context LLM inference, providing a simple yet effective solution to the memory bottleneck problem. The code is available at \\href{https://github.com/NVIDIA/kvpress}{link}.", "authors": ["Xiang Liu", "Zhenheng Tang", "Peijie Dong", "Zeyu Li", "Yue Liu", "Bo Li", "Xuming Hu", "Xiaowen Chu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-01", "url": "https://arxiv.org/abs/2502.00299", "pdf_url": "https://arxiv.org/pdf/2502.00299v5", "arxiv_id": "2502.00299", "doi": "10.48550/arXiv.2502.00299", "citation_count": 41, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/NVIDIA/kvpress}{link}", "venue": "arXiv.org", "quality_score": 0.4058}
{"id": "52f300f4cb995f7935d7beb84d6df2c6353051c117aafa4d6b7628358eb4743a", "sources": ["arxiv", "semantic_scholar"], "title": "Federated Sketching LoRA: A Flexible Framework for Heterogeneous Collaborative Fine-Tuning of LLMs", "abstract": "Fine-tuning large language models (LLMs) on resource-constrained clients remains a challenging problem. Recent works have fused low-rank adaptation (LoRA) techniques with federated fine-tuning to mitigate challenges associated with client model sizes and data scarcity. Still, the heterogeneity of resources remains a critical bottleneck: while higher-rank modules generally enhance performance, varying client capabilities constrain LoRA's feasible rank range. Existing approaches attempting to resolve this issue either lack analytical justification or impose additional computational overhead, leaving a wide gap for efficient and theoretically-grounded solutions. To address these challenges, we propose federated sketching LoRA (FSLoRA), which leverages a sketching mechanism to enable clients to selectively update submatrices of global LoRA modules maintained by the server. By adjusting the sketching ratios, which determine the ranks of the submatrices on the clients, FSLoRA flexibly adapts to client-specific communication and computational constraints. We provide a rigorous convergence analysis of FSLoRA that characterizes how the sketching ratios affect the convergence rate. Through extensive experiments, we demonstrate that FSLoRA outperforms baselines and significantly improves training efficiency while preserving stable convergence.", "authors": ["Wenzhi Fang", "Dong-Jun Han", "Liangqi Yuan", "Seyyedali Hosseinalipour", "Christopher G. Brinton"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-31", "url": "https://arxiv.org/abs/2501.19389", "pdf_url": "https://arxiv.org/pdf/2501.19389v4", "arxiv_id": "2501.19389", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1193}
{"id": "01ddcc583befbec6dc18d1be81f581ed019f04953fd59ce21666d66b257f0da8", "sources": ["arxiv", "semantic_scholar"], "title": "Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models", "abstract": "Efficient real-world deployments of large language models (LLMs) rely on Key-Value (KV) caching for processing and generating long outputs, reducing the need for repetitive computation. For large contexts, Key-Value caches can take up tens of gigabytes of device memory, as they store vector representations for each token and layer. Recent work has shown that the cached vectors can be compressed through quantization, pruning or merging, but these techniques often compromise quality towards higher compression rates. In this work, we aim to improve Key & Value compression by exploiting two observations: 1) the inherent dependencies between keys and values across different layers, and 2) high-compression mechanisms for internal network states. We propose AQUA-KV, an adaptive quantization for Key-Value caches that relies on compact adapters to exploit existing dependencies between Keys and Values, and aims to \"optimally\" compress the information that cannot be predicted. AQUA-KV significantly improves compression rates, while maintaining high accuracy on state-of-the-art LLM families. On Llama 3.2 LLMs, we achieve near-lossless inference at 2-2.5 bits per value with under $1\\%$ relative error in perplexity and LongBench scores. AQUA-KV is one-shot, simple, and efficient: it can be calibrated on a single GPU within 1-6 hours, even for 70B models.", "authors": ["Alina Shutova", "Vladimir Malinovskii", "Vage Egiazarian", "Denis Kuznedelev", "Denis Mazur", "Nikita Surkov", "Ivan Ermakov", "Dan Alistarh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-31", "url": "https://arxiv.org/abs/2501.19392", "pdf_url": "https://arxiv.org/pdf/2501.19392v4", "arxiv_id": "2501.19392", "doi": "10.48550/arXiv.2501.19392", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2785}
{"id": "bb32a1411b51806f6ef088a80ee6bb1192fa382c5db69157410d465885d3566f", "sources": ["arxiv", "semantic_scholar"], "title": "CLoQ: Enhancing Fine-Tuning of Quantized LLMs via Calibrated LoRA Initialization", "abstract": "Fine-tuning large language models (LLMs) using low-rank adaptation (LoRA) has become a highly efficient approach for downstream tasks, particularly in scenarios with limited computational resources. However, applying LoRA techniques to quantized LLMs poses unique challenges due to the reduced representational precision of quantized weights. In this paper, we introduce CLoQ (Calibrated LoRA initialization for Quantized LLMs), a simplistic initialization strategy designed to overcome these challenges. Our approach focuses on minimizing the layer-wise discrepancy between the original LLM and its quantized counterpart with LoRA components during initialization. By leveraging a small calibration dataset, CLoQ quantizes a pre-trained LLM and determines the optimal LoRA components for each layer, ensuring a strong foundation for subsequent fine-tuning. A key contribution of this work is a novel theoretical result that enables the accurate and closed-form construction of these optimal LoRA components. We validate the efficacy of CLoQ across multiple tasks such as language generation, arithmetic reasoning, and commonsense reasoning, demonstrating that it consistently outperforms existing LoRA fine-tuning methods for quantized LLMs, especially at ultra low-bit widths.", "authors": ["Yanxia Deng", "Aozhong Zhang", "Selcuk Gurses", "Naigang Wang", "Zi Yang", "Penghang Yin"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-30", "url": "https://arxiv.org/abs/2501.18475", "pdf_url": "https://arxiv.org/pdf/2501.18475v2", "arxiv_id": "2501.18475", "doi": "10.48550/arXiv.2501.18475", "citation_count": 4, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Transactions on Machine Learning Research (TMLR), 2025", "quality_score": 0.2386}
{"id": "603ff637571b8b054aceb70081e7b909d76f6135280f5880cdc05a56f4f64a41", "sources": ["arxiv", "semantic_scholar"], "title": "Heuristic-Informed Mixture of Experts for Link Prediction in Multilayer Networks", "abstract": "Link prediction algorithms for multilayer networks are in principle required to effectively account for the entire layered structure while capturing the unique contexts offered by each layer. However, many existing approaches excel at predicting specific links in certain layers but struggle with others, as they fail to effectively leverage the diverse information encoded across different network layers. In this paper, we present MoE-ML-LP, the first Mixture-of-Experts (MoE) framework specifically designed for multilayer link prediction. Building on top of multilayer heuristics for link prediction, MoE-ML-LP synthesizes the decisions taken by diverse experts, resulting in significantly enhanced predictive capabilities. Our extensive experimental evaluation on real-world and synthetic networks demonstrates that MoE-ML-LP consistently outperforms several baselines and competing methods, achieving remarkable improvements of +60% in Mean Reciprocal Rank, +82% in Hits@1, +55% in Hits@5, and +41% in Hits@10. Furthermore, MoE-ML-LP features a modular architecture that enables the seamless integration of newly developed experts without necessitating the re-training of the entire framework, fostering efficiency and scalability to new experts, paving the way for future advancements in link prediction.", "authors": ["Lucio La Cava", "Domenico Mandaglio", "Lorenzo Zangari", "Andrea Tagarelli"], "categories": ["cs.LG", "cs.SI", "physics.soc-ph"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2025-01-29", "url": "https://arxiv.org/abs/2501.17557", "pdf_url": "https://arxiv.org/pdf/2501.17557v1", "arxiv_id": "2501.17557", "doi": "10.48550/arXiv.2501.17557", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "1563b96b4a4f16dc7a17bfffcb6966b52ac301ba424fa66ed3792a013734f41f", "sources": ["arxiv", "semantic_scholar"], "title": "In-Context Meta LoRA Generation", "abstract": "Low-rank Adaptation (LoRA) has demonstrated remarkable capabilities for task specific fine-tuning. However, in scenarios that involve multiple tasks, training a separate LoRA model for each one results in considerable inefficiency in terms of storage and inference. Moreover, existing parameter generation methods fail to capture the correlations among these tasks, making multi-task LoRA parameter generation challenging. To address these limitations, we propose In-Context Meta LoRA (ICM-LoRA), a novel approach that efficiently achieves task-specific customization of large language models (LLMs). Specifically, we use training data from all tasks to train a tailored generator, Conditional Variational Autoencoder (CVAE). CVAE takes task descriptions as inputs and produces task-aware LoRA weights as outputs. These LoRA weights are then merged with LLMs to create task-specialized models without the need for additional fine-tuning. Furthermore, we utilize in-context meta-learning for knowledge enhancement and task mapping, to capture the relationship between tasks and parameter distributions. As a result, our method achieves more accurate LoRA parameter generation for diverse tasks using CVAE. ICM-LoRA enables more accurate LoRA parameter reconstruction than current parameter reconstruction methods and is useful for implementing task-specific enhancements of LoRA parameters. At the same time, our method occupies 283MB, only 1\\% storage compared with the original LoRA.", "authors": ["Yihua Shao", "Minxi Yan", "Yang Liu", "Siyu Chen", "Wenjie Chen", "Xinwei Long", "Ziyang Yan", "Lei Li", "Chenyu Zhang", "Nicu Sebe", "Hao Tang", "Yan Wang", "Hao Zhao", "Mengzhu Wang", "Jingcai Guo"], "categories": ["cs.CL", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-29", "url": "https://arxiv.org/abs/2501.17635", "pdf_url": "https://arxiv.org/pdf/2501.17635v3", "arxiv_id": "2501.17635", "doi": "10.48550/arXiv.2501.17635", "citation_count": 17, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Joint Conference on Artificial Intelligence", "quality_score": 0.3138}
{"id": "c18b88a445677f0d0821287478d4a952f328fe3376666c79c7e82bfac6868742", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Mamba: Enhancing Multi-Modal State-Space Models with Modality-Aware Sparsity", "abstract": "State Space Models (SSMs) have emerged as efficient alternatives to Transformers for sequential modeling, but their inability to leverage modality-specific features limits their performance in multi-modal pretraining. Here, we propose Mixture-of-Mamba, a novel SSM architecture that introduces modality-aware sparsity through modality-specific parameterization of the Mamba block. Building on Mixture-of-Transformers (W. Liang et al. arXiv:2411.04996; 2024), we extend the benefits of modality-aware sparsity to SSMs while preserving their computational efficiency. We evaluate Mixture-of-Mamba across three multi-modal pretraining settings: Transfusion (interleaved text and continuous image tokens with diffusion loss), Chameleon (interleaved text and discrete image tokens), and an extended three-modality framework incorporating speech. Mixture-of-Mamba consistently reaches the same loss values at earlier training steps with significantly reduced computational costs. In the Transfusion setting, Mixture-of-Mamba achieves equivalent image loss using only 34.76% of the training FLOPs at the 1.4B scale. In the Chameleon setting, Mixture-of-Mamba reaches similar image loss with just 42.50% of the FLOPs at the 1.4B scale, and similar text loss with just 65.40% of the FLOPs. In the three-modality setting, MoM matches speech loss at 24.80% of the FLOPs at the 1.4B scale. Our ablation study highlights the synergistic effects of decoupling projection components, where joint decoupling yields greater gains than individual modifications. These results establish modality-aware sparsity as a versatile and effective design principle, extending its impact from Transformers to SSMs and setting new benchmarks in multi-modal pretraining. Our code can be accessed at https://github.com/Weixin-Liang/Mixture-of-Mamba", "authors": ["Weixin Liang", "Junhong Shen", "Genghan Zhang", "Ning Dong", "Luke Zettlemoyer", "Lili Yu"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-27", "url": "https://arxiv.org/abs/2501.16295", "pdf_url": "https://arxiv.org/pdf/2501.16295v1", "arxiv_id": "2501.16295", "doi": "10.48550/arXiv.2501.16295", "citation_count": 8, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Weixin-Liang/Mixture-of-Mamba", "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "ed643910cf65800d394575f8e61182bfc544341c61424a3bc38c615aaa7d4389", "sources": ["arxiv", "semantic_scholar"], "title": "AKVQ-VL: Attention-Aware KV Cache Adaptive 2-Bit Quantization for Vision-Language Models", "abstract": "Vision-language models (VLMs) show remarkable performance in multimodal tasks. However, excessively long multimodal inputs lead to oversized Key-Value (KV) caches, resulting in significant memory consumption and I/O bottlenecks. Previous KV quantization methods for Large Language Models (LLMs) may alleviate these issues but overlook the attention saliency differences of multimodal tokens, resulting in suboptimal performance. In this paper, we investigate the attention-aware token saliency patterns in VLM and propose AKVQ-VL. AKVQ-VL leverages the proposed Text-Salient Attention (TSA) and Pivot-Token-Salient Attention (PSA) patterns to adaptively allocate bit budgets. Moreover, achieving extremely low-bit quantization requires effectively addressing outliers in KV tensors. AKVQ-VL utilizes the Walsh-Hadamard transform (WHT) to construct outlier-free KV caches, thereby reducing quantization difficulty. Evaluations of 2-bit quantization on 12 long-context and multimodal tasks demonstrate that AKVQ-VL maintains or even improves accuracy, outperforming LLM-oriented methods. AKVQ-VL can reduce peak memory usage by 2.13x, support up to 3.25x larger batch sizes and 2.46x throughput.", "authors": ["Zunhai Su", "Wang Shen", "Linge Li", "Zhe Chen", "Hanyu Wei", "Huangqi Yu", "Kehong Yuan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-25", "url": "https://arxiv.org/abs/2501.15021", "pdf_url": "https://arxiv.org/pdf/2501.15021v1", "arxiv_id": "2501.15021", "doi": "10.1109/ICME59968.2025.11209367", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Multimedia and Expo", "quality_score": 0.2386}
{"id": "ce6b294930c22a5faed694be982034520f50e1b4761771fec6225bac4a9c2e1a", "sources": ["arxiv", "semantic_scholar"], "title": "Each Rank Could be an Expert: Single-Ranked Mixture of Experts LoRA for Multi-Task Learning", "abstract": "Low-Rank Adaptation (LoRA) is widely used for adapting large language models (LLMs) to specific domains due to its efficiency and modularity. Meanwhile, vanilla LoRA struggles with task conflicts in multi-task scenarios. Recent works adopt Mixture of Experts (MoE) by treating each LoRA module as an expert, thereby mitigating task interference through multiple specialized LoRA modules. While effective, these methods often isolate knowledge within individual tasks, failing to fully exploit the shared knowledge across related tasks. In this paper, we establish a connection between single LoRA and multi-LoRA MoE, integrating them into a unified framework. We demonstrate that the dynamic routing of multiple LoRAs is functionally equivalent to rank partitioning and block-level activation within a single LoRA. We further empirically demonstrate that finer-grained LoRA partitioning, within the same total and activated parameter constraints, leads to better performance gains across heterogeneous tasks. Building on these findings, we propose Single-ranked Mixture of Experts LoRA (\\textbf{SMoRA}), which embeds MoE into LoRA by \\textit{treating each rank as an independent expert}. With a \\textit{dynamic rank-wise activation} mechanism, SMoRA promotes finer-grained knowledge sharing while mitigating task conflicts. Experiments demonstrate that SMoRA activates fewer parameters yet achieves better performance in multi-task scenarios.", "authors": ["Ziyu Zhao", "Yixiao Zhou", "Zhi Zhang", "Didi Zhu", "Tao Shen", "Zexi Li", "Jinluan Yang", "Xuwu Wang", "Jing Su", "Kun Kuang", "Zhongyu Wei", "Fei Wu", "Yu Cheng"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-25", "url": "https://arxiv.org/abs/2501.15103", "pdf_url": "https://arxiv.org/pdf/2501.15103v2", "arxiv_id": "2501.15103", "doi": "10.1145/3770854.3780222", "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2865}
{"id": "b6d2c2860ca13d3720aae655b1cf5b4a67e5457043e32afbb34976753520d937", "sources": ["arxiv", "semantic_scholar"], "title": "Task-KV: Task-aware KV Cache Optimization via Semantic Differentiation of Attention Heads", "abstract": "KV cache is a widely used acceleration technique for large language models (LLMs) inference. However, its memory requirement grows rapidly with input length. Previous studies have reduced the size of KV cache by either removing the same number of unimportant tokens for all attention heads or by allocating differentiated KV cache budgets for pre-identified attention heads. However, due to the importance of attention heads varies across different tasks, the pre-identified attention heads fail to adapt effectively to various downstream tasks. To address this issue, we propose Task-KV, a method that leverages the semantic differentiation of attention heads to allocate differentiated KV cache budgets across various tasks. We demonstrate that attention heads far from the semantic center (called heterogeneous heads) make an significant contribution to task outputs and semantic understanding. In contrast, other attention heads play the role of aggregating important information and focusing reasoning. Task-KV allocates full KV cache budget to heterogeneous heads to preserve comprehensive semantic information, while reserving a small number of recent tokens and attention sinks for non-heterogeneous heads. Furthermore, we innovatively introduce middle activations to preserve key contextual information aggregated from non-heterogeneous heads. To dynamically perceive semantic differences among attention heads, we design a semantic separator to distinguish heterogeneous heads from non-heterogeneous ones based on their distances from the semantic center. Experimental results on multiple benchmarks and different model architectures demonstrate that Task-KV significantly outperforms existing baseline methods.", "authors": ["Xingyang He", "Jie Liu", "Shaowei Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-25", "url": "https://arxiv.org/abs/2501.15113", "pdf_url": "https://arxiv.org/pdf/2501.15113v1", "arxiv_id": "2501.15113", "doi": "10.48550/arXiv.2501.15113", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "40106908688606cdf052aea8136abddbdc4db863604bed31de34de47848b6c12", "sources": ["arxiv", "semantic_scholar"], "title": "RotateKV: Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations", "abstract": "Key-Value (KV) cache facilitates efficient large language models (LLMs) inference by avoiding recomputation of past KVs. As the batch size and context length increase, the oversized KV caches become a significant memory bottleneck, highlighting the need for efficient compression. Existing KV quantization rely on fine-grained quantization or the retention of a significant portion of high bit-widths caches, both of which compromise compression ratio and often fail to maintain robustness at extremely low average bit-widths. In this work, we explore the potential of rotation technique for 2-bit KV quantization and propose RotateKV, which achieves accurate and robust performance through the following innovations: (i) Outlier-Aware Rotation, which utilizes channel-reordering to adapt the rotations to varying channel-wise outlier distributions without sacrificing the computational efficiency of the fast Walsh-Hadamard transform (FWHT); (ii) Pre-RoPE Grouped-Head Rotation, which mitigates the impact of rotary position embedding (RoPE) on proposed outlier-aware rotation and further smooths outliers across heads; (iii) Attention-Sink-Aware Quantization, which leverages the massive activations to precisely identify and protect attention sinks. RotateKV achieves less than 0.3 perplexity (PPL) degradation with 2-bit quantization on WikiText-2 using LLaMA-2-13B, maintains strong CoT reasoning and long-context capabilities, with less than 1.7\\% degradation on GSM8K, outperforming existing methods even at lower average bit-widths. RotateKV also showcases a 3.97x reduction in peak memory usage, supports 5.75x larger batch sizes, and achieves a 2.32x speedup in decoding stage.", "authors": ["Zunhai Su", "Zhe Chen", "Wang Shen", "Hanyu Wei", "Linge Li", "Huangqi Yu", "Kehong Yuan"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-25", "url": "https://arxiv.org/abs/2501.16383", "pdf_url": "https://arxiv.org/pdf/2501.16383v2", "arxiv_id": "2501.16383", "doi": "10.48550/arXiv.2501.16383", "citation_count": 24, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Joint Conference on Artificial Intelligence", "quality_score": 0.3495}
{"id": "45031621fa7941a7f02cd712214bbf3ac2e0940d443a3df635865d61e13227f1", "sources": ["arxiv", "semantic_scholar"], "title": "ToMoE: Converting Dense Large Language Models to Mixture-of-Experts through Dynamic Structural Pruning", "abstract": "Large Language Models (LLMs) have demonstrated remarkable abilities in tackling a wide range of complex tasks. However, their huge computational and memory costs raise significant challenges in deploying these models on resource-constrained devices or efficiently serving them. Prior approaches have attempted to alleviate these problems by permanently removing less important model structures, yet these methods often result in substantial performance degradation due to the permanent deletion of model parameters. In this work, we tried to mitigate this issue by reducing the number of active parameters without permanently removing them. Specifically, we introduce a differentiable dynamic pruning method that pushes dense models to maintain a fixed number of active parameters by converting their MLP layers into a Mixture of Experts (MoE) architecture. Our method, even without fine-tuning, consistently outperforms previous structural pruning techniques across diverse model families, including Phi-2, LLaMA-2, LLaMA-3, and Qwen-2.5.", "authors": ["Shangqian Gao", "Ting Hua", "Reza Shirkavand", "Chi-Heng Lin", "Zheng Tang", "Zhengao Li", "Longge Yuan", "Fangyi Li", "Zeyu Zhang", "Alireza Ganjdanesh", "Lou Qian", "Xu Jie", "Yen-Chang Hsu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-25", "url": "https://arxiv.org/abs/2501.15316", "pdf_url": "https://arxiv.org/pdf/2501.15316v2", "arxiv_id": "2501.15316", "doi": "10.48550/arXiv.2501.15316", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.25}
{"id": "6d5449da6048f26c394ed005bf6086eabc56ca3f44e380755d0340d145db36d7", "sources": ["arxiv", "semantic_scholar"], "title": "Mean-field limit from general mixtures of experts to quantum neural networks", "abstract": "In this work, we study the asymptotic behavior of Mixture of Experts (MoE) trained via gradient flow on supervised learning problems. Our main result establishes the propagation of chaos for a MoE as the number of experts diverges. We demonstrate that the corresponding empirical measure of their parameters is close to a probability measure that solves a nonlinear continuity equation, and we provide an explicit convergence rate that depends solely on the number of experts. We apply our results to a MoE generated by a quantum neural network.", "authors": ["Anderson Melchor Hernandez", "Davide Pastorello", "Giacomo De Palma"], "categories": ["math-ph", "cs.LG", "math.PR"], "fields_of_study": ["Computer Science", "Physics", "Mathematics"], "published_date": "2025-01-24", "url": "https://arxiv.org/abs/2501.14660", "pdf_url": "https://arxiv.org/pdf/2501.14660v2", "arxiv_id": "2501.14660", "doi": "10.1007/s11005-026-02065-9", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Letters in Mathematical Physics", "quality_score": 0.0367}
{"id": "23a7834c082961cf5eb2f3ce888994710f605d7e7e495d9e9379f79d66e46ebf", "sources": ["arxiv", "semantic_scholar"], "title": "Hierarchical Time-Aware Mixture of Experts for Multi-Modal Sequential Recommendation", "abstract": "Multi-modal sequential recommendation (SR) leverages multi-modal data to learn more comprehensive item features and user preferences than traditional SR methods, which has become a critical topic in both academia and industry. Existing methods typically focus on enhancing multi-modal information utility through adaptive modality fusion to capture the evolving of user preference from user-item interaction sequences. However, most of them overlook the interference caused by redundant interest-irrelevant information contained in rich multi-modal data. Additionally, they primarily rely on implicit temporal information based solely on chronological ordering, neglecting explicit temporal signals that could more effectively represent dynamic user interest over time. To address these limitations, we propose a Hierarchical time-aware Mixture of experts for multi-modal Sequential Recommendation (HM4SR) with a two-level Mixture of Experts (MoE) and a multi-task learning strategy. Specifically, the first MoE, named Interactive MoE, extracts essential user interest-related information from the multi-modal data of each item. Then, the second MoE, termed Temporal MoE, captures user dynamic interests by introducing explicit temporal embeddings from timestamps in modality encoding. To further address data sparsity, we propose three auxiliary supervision tasks: sequence-level category prediction (CP) for item feature understanding, contrastive learning on ID (IDCL) to align sequence context with user interests, and placeholder contrastive learning (PCL) to integrate temporal information with modalities for dynamic interest modeling. Extensive experiments on four public datasets verify the effectiveness of HM4SR compared to several state-of-the-art approaches.", "authors": ["Shengzhe Zhang", "Liyi Chen", "Dazhong Shen", "Chao Wang", "Hui Xiong"], "categories": ["cs.IR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-24", "url": "https://arxiv.org/abs/2501.14269", "pdf_url": "https://arxiv.org/pdf/2501.14269v2", "arxiv_id": "2501.14269", "doi": "10.1145/3696410.3714676", "citation_count": 31, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "The Web Conference", "quality_score": 0.3763}
{"id": "55cf3eddb4f6140ee421babe19a7211ca574a285772a60de0b5fca4b4c4658a1", "sources": ["arxiv", "semantic_scholar"], "title": "MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods", "abstract": "Mamba is an efficient sequence model that rivals Transformers and demonstrates significant potential as a foundational architecture for various tasks. Quantization is commonly used in neural networks to reduce model size and computational latency. However, applying quantization to Mamba remains underexplored, and existing quantization methods, which have been effective for CNN and Transformer models, appear inadequate for Mamba models (e.g., Quarot suffers a 21% accuracy drop on Vim-T$^\\dagger$ even under W8A8). We have pioneered the exploration of this issue and identified several key challenges. First, significant outliers are present in gate projections, output projections, and matrix multiplications. Second, Mamba's unique parallel scan further amplifies these outliers, leading to uneven and heavy-tailed data distributions. Third, even with the application of the Hadamard transform, the variance across channels in weights and activations still remains inconsistent. To these ends, we propose MambaQuant, a post-training quantization (PTQ) framework consisting of: 1) Karhunen-Loeve Transformation (KLT) enhanced rotation, rendering the rotation matrix adaptable to diverse channel distributions. 2) Smooth-Fused rotation, which equalizes channel variances and can merge additional parameters into model weights. Experiments show that MambaQuant can quantize both weights and activations into 8-bit with less than 1% accuracy loss for Mamba-based vision and language tasks. To the best of our knowledge, MambaQuant is the first comprehensive PTQ design for the Mamba family, paving the way for further advancements in its application.", "authors": ["Zukang Xu", "Yuxuan Yue", "Xing Hu", "Zhihang Yuan", "Zixu Jiang", "Zhixuan Chen", "Jiangyong Yu", "Chen Xu", "Sifan Zhou", "Dawei Yang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-23", "url": "https://arxiv.org/abs/2501.13484", "pdf_url": "https://arxiv.org/pdf/2501.13484v3", "arxiv_id": "2501.13484", "doi": "10.48550/arXiv.2501.13484", "citation_count": 28, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4225}
{"id": "a28f3b69b6ee26dafe4f68a2f16cbb4fcad4e8d26ebe75ee1a8a9137d2d64e6f", "sources": ["arxiv", "semantic_scholar"], "title": "BLR-MoE: Boosted Language-Routing Mixture of Experts for Domain-Robust Multilingual E2E ASR", "abstract": "Recently, the Mixture of Expert (MoE) architecture, such as LR-MoE, is often used to alleviate the impact of language confusion on the multilingual ASR (MASR) task. However, it still faces language confusion issues, especially in mismatched domain scenarios. In this paper, we decouple language confusion in LR-MoE into confusion in self-attention and router. To alleviate the language confusion in self-attention, based on LR-MoE, we propose to apply attention-MoE architecture for MASR. In our new architecture, MoE is utilized not only on feed-forward network (FFN) but also on self-attention. In addition, to improve the robustness of the LID-based router on language confusion, we propose expert pruning and router augmentation methods. Combining the above, we get the boosted language-routing MoE (BLR-MoE) architecture. We verify the effectiveness of the proposed BLR-MoE in a 10,000-hour MASR dataset.", "authors": ["Guodong Ma", "Wenxuan Wang", "Lifeng Zhou", "Yuting Yang", "Yuke Li", "Binbin Du"], "categories": ["cs.CL", "cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-01-22", "url": "https://arxiv.org/abs/2501.12602", "pdf_url": "https://arxiv.org/pdf/2501.12602v1", "arxiv_id": "2501.12602", "doi": "10.1109/ICASSP49660.2025.10887955", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.1747}
{"id": "da690bfa99c1e0d1ae88ee190ae03a6006720f1a507d5d237d8c1be5f00b58d4", "sources": ["arxiv", "semantic_scholar"], "title": "D-LoRa: a Distributed Parameter Adaptation Scheme for LoRa Network", "abstract": "The deployment of LoRa networks necessitates joint performance optimization, including packet delivery rate, energy efficiency, and throughput. Additionally, multiple LoRa parameters for packet transmission must be dynamically configured to tailor the performance metrics prioritization across varying channel environments. Because of the coupling relationship between LoRa parameters and metrics, existing works have opted to focus on certain parameters or specific metrics to circumvent the intricate coupling relationship, leading to limited adaptability. Therefore, we propose D-LoRa, a distributed parameter adaptation scheme, based on reinforcement learning towards network performance. We decompose the joint performance optimization problem into multiple independent Multi-Armed Bandit (MAB) problems with different reward functions. We have also built a comprehensive analytical model for the LoRa network that considers path loss, quasi-orthogonality of spreading factor, and packet collision. Experimental results show that our scheme can increase packet delivery rate by up to 28.8% and demonstrates superior adaptability across different performance metrics.", "authors": ["Ruiqi Wang", "Tongyu Song", "Jing Ren", "Xiong Wang", "Shizhong Xu", "Sheng Wang"], "categories": ["cs.NI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-22", "url": "https://arxiv.org/abs/2501.12589", "pdf_url": "https://arxiv.org/pdf/2501.12589v2", "arxiv_id": "2501.12589", "doi": "10.1109/GLOBECOM59602.2025.11431835", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Global Communications Conference", "quality_score": 0.1505}
{"id": "e4da9af6de7024886b9f951b464f9e93c45e7b99ee557ba2b8a96d66e0f5a26d", "sources": ["arxiv", "semantic_scholar"], "title": "A One-Dimensional Energy Balance Model Parameterization for the Formation of CO2 Ice on the Surfaces of Eccentric Extrasolar Planets", "abstract": "Eccentric planets may spend a significant portion of their orbits at large distances from their host stars, where low temperatures can cause atmospheric CO2 to condense out onto the surface, similar to the polar ice caps on Mars. The radiative effects on the climates of these planets throughout their orbits would depend on the wavelength-dependent albedo of surface CO2 ice that may accumulate at or near apoastron and vary according to the spectral energy distribution of the host star. To explore these possible effects, we incorporated a CO2 ice-albedo parameterization into a one-dimensional energy balance climate model. With the inclusion of this parameterization, our simulations demonstrated that F-dwarf planets require 29% more orbit-averaged flux to thaw out of global water ice cover compared with simulations that solely use a traditional pure water ice-albedo parameterization. When no eccentricity is assumed, and host stars are varied, F-dwarf planets with higher bond albedos relative to their M-dwarf planet counterparts require 30% more orbit-averaged flux to exit a water snowball state. Additionally, the intense heat experienced at periastron aids eccentric planets in exiting a snowball state with a smaller increase in instellation compared with planets on circular orbits; this enables eccentric planets to exhibit warmer conditions along a broad range of instellation. This study emphasizes the significance of incorporating an albedo parameterization for the formation of CO2 ice into climate models to accurately assess the habitability of eccentric planets, as we show that, even at moderate eccentricities, planets with Earth-like atmospheres can reach surface temperatures cold enough for the condensation of CO2 onto their surfaces, as can planets receiving low amounts of instellation on circular orbits.", "authors": ["Vidya Venkatesan", "Aomawa L. Shields", "Russell Deitrick", "Eric T. Wolf", "Andrew Rushby", "Department of Physics", " Astronomy", "University of California", " Irvine", " California", " USA", "School of Earth", "Ocean Sciences", "University of Victoria", " Victoria", " Canada", "Laboratory for Atmospheric", "Space Physics", "University of Colorado Boulder", " Boulder", " Colorado", " USA", "Sellers Exoplanet Environment Collaboration", "NASA Goddard Space Flight Center", " Greenbelt", " Maryland", " USA", "Blue Marble Space Institute of Science", " Seattle", " Washington", " USA", "Department of Earth", "Planetary Sciences", "Birkbeck University of London", " London", "United Kingdom"], "categories": ["astro-ph.EP"], "fields_of_study": ["Physics", "Medicine"], "published_date": "2025-01-20", "url": "https://arxiv.org/abs/2501.11667", "pdf_url": "https://arxiv.org/pdf/2501.11667v1", "arxiv_id": "2501.11667", "doi": "10.1089/ast.2023.0103", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Astrobiology", "quality_score": 0.1945}
{"id": "427fb51889667f7709246e9dd094e1fffba36e7ff756c4a89c0e145411ec14ec", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts (MoE): A Big Data Perspective", "abstract": "As the era of big data arrives, traditional artificial intelligence algorithms have difficulty processing the demands of massive and diverse data. Mixture of experts (MoE) has shown excellent performance and broad application prospects. This paper provides an in-depth review and analysis of the latest progress in this field from multiple perspectives, including the basic principles, algorithmic models, key technical challenges, and application practices of MoE. First, we introduce the basic concept of MoE and its core idea and elaborate on its advantages over traditional single models. Then, we discuss the basic architecture of MoE and its main components, including the gating network, expert networks, and learning algorithms. Next, we review the applications of MoE in addressing key technical issues in big data. For each challenge, we provide specific MoE solutions and their innovations. Furthermore, we summarize the typical use cases of MoE in various application domains. This fully demonstrates the powerful capability of MoE in big data processing. We also analyze the advantages of MoE in big data environments. Finally, we explore the future development trends of MoE. We believe that MoE will become an important paradigm of artificial intelligence in the era of big data. In summary, this paper systematically elaborates on the principles, techniques, and applications of MoE in big data processing, providing theoretical and practical references to further promote the application of MoE in real scenarios.", "authors": ["Wensheng Gan", "Zhenyao Ning", "Zhenlian Qi", "Philip S. Yu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-18", "url": "https://arxiv.org/abs/2501.16352", "pdf_url": "https://arxiv.org/pdf/2501.16352v1", "arxiv_id": "2501.16352", "doi": "10.48550/arXiv.2501.16352", "citation_count": 36, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Information Fusion", "quality_score": 0.3921}
{"id": "60392609483323c0644957492382f1af8c0e6d14baf646f6a0eea901cf0e2808", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Modal Attention Networks for Enhanced Segmentation and Depth Estimation of Subsurface Defects in Pulse Thermography", "abstract": "AI-driven pulse thermography (PT) has become a crucial tool in non-destructive testing (NDT), enabling automatic detection of hidden anomalies in various industrial components. Current state-of-the-art techniques feed segmentation and depth estimation networks compressed PT sequences using either Principal Component Analysis (PCA) or Thermographic Signal Reconstruction (TSR). However, treating these two modalities independently constrains the performance of PT inspection models as these representations possess complementary semantic features. To address this limitation, this work proposes PT-Fusion, a multi-modal attention-based fusion network that fuses both PCA and TSR modalities for defect segmentation and depth estimation of subsurface defects in PT setups. PT-Fusion introduces novel feature fusion modules, Encoder Attention Fusion Gate (EAFG) and Attention Enhanced Decoding Block (AEDB), to fuse PCA and TSR features for enhanced segmentation and depth estimation of subsurface defects. In addition, a novel data augmentation technique is proposed based on random data sampling from thermographic sequences to alleviate the scarcity of PT datasets. The proposed method is benchmarked against state-of-the-art PT inspection models, including U-Net, attention U-Net, and 3D-CNN on the Université Laval IRT-PVC dataset. The results demonstrate that PT-Fusion outperforms the aforementioned models in defect segmentation and depth estimation accuracies with a margin of 10%.", "authors": ["Mohammed Salah", "Naoufel Werghi", "Davor Svetinovic", "Yusra Abdulrahman"], "categories": ["cs.CV", "cs.AI", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2025-01-17", "url": "https://arxiv.org/abs/2501.09994", "pdf_url": "https://arxiv.org/pdf/2501.09994v1", "arxiv_id": "2501.09994", "doi": "10.48550/arXiv.2501.09994", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "f60cb7622def57ababf99ef52811d2cde48d0510f437b0815db43ca71c41c7a6", "sources": ["arxiv", "semantic_scholar"], "title": "LLM-Based Routing in Mixture of Experts: A Novel Framework for Trading", "abstract": "Recent advances in deep learning and large language models (LLMs) have facilitated the deployment of the mixture-of-experts (MoE) mechanism in the stock investment domain. While these models have demonstrated promising trading performance, they are often unimodal, neglecting the wealth of information available in other modalities, such as textual data. Moreover, the traditional neural network-based router selection mechanism fails to consider contextual and real-world nuances, resulting in suboptimal expert selection. To address these limitations, we propose LLMoE, a novel framework that employs LLMs as the router within the MoE architecture. Specifically, we replace the conventional neural network-based router with LLMs, leveraging their extensive world knowledge and reasoning capabilities to select experts based on historical price data and stock news. This approach provides a more effective and interpretable selection mechanism. Our experiments on multimodal real-world stock datasets demonstrate that LLMoE outperforms state-of-the-art MoE models and other deep neural network approaches. Additionally, the flexible architecture of LLMoE allows for easy adaptation to various downstream tasks.", "authors": ["Kuan-Ming Liu", "Ming-Chih Lo"], "categories": ["cs.LG", "q-fin.TR"], "fields_of_study": ["Computer Science", "Economics"], "published_date": "2025-01-16", "url": "https://arxiv.org/abs/2501.09636", "pdf_url": "https://arxiv.org/pdf/2501.09636v2", "arxiv_id": "2501.09636", "doi": "10.48550/arXiv.2501.09636", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "0f65f1a55c198b9e459f9a4ab087525ef29fb5da5b8ed71fa5efc7f8b526a64f", "sources": ["arxiv", "semantic_scholar"], "title": "PRESERVE: Prefetching Model Weights and KV-Cache in Distributed LLM Serving", "abstract": "Large language models (LLMs) are typically served from clusters of GPUs/NPUs that consist of large number of devices. Unfortunately, communication between these devices incurs significant overhead, increasing the inference latency and cost while limiting the scalability. Prior work addressed this issue by overlapping communication with compute, but has severe limitations due to the data dependencies between these operations. In this paper, we propose PRESERVE, a novel framework that prefetches model weights and KV-cache from off-chip HBM memory to the on-chip cache of AI accelerators during the communication operations, which offers various advantages and performance improvements compared to prior methods. Through extensive experiments conducted on commercial AI accelerators, we demonstrate up to 1.6x end-to-end speedup on state-of-the-art, open-source LLMs. Additionally, we perform a design space exploration that identifies the optimal hardware configuration for the proposed method, showing a further 1.25x improvement in performance per cost by selecting the optimal L2 cache size. Our results show that PRESERVE has the potential to mitigate the memory bottlenecks and communication overheads, offering a solution to improve the performance and scalability of the LLM inference systems.", "authors": ["Ahmet Caner Yüzügüler", "Jiawei Zhuang", "Lukas Cavigelli"], "categories": ["cs.AI", "cs.AR", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-14", "url": "https://arxiv.org/abs/2501.08192", "pdf_url": "https://arxiv.org/pdf/2501.08192v2", "arxiv_id": "2501.08192", "doi": "10.48550/arXiv.2501.08192", "citation_count": 12, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "91a360b5ec0c35c8a02a259212d97adbc565ef9100c1fa003e95a98896ff9135", "sources": ["arxiv", "semantic_scholar"], "title": "GRAPHMOE: Amplifying Cognitive Depth of Mixture-of-Experts Network via Introducing Self-Rethinking Mechanism", "abstract": "Traditional Mixture-of-Experts (MoE) networks benefit from utilizing multiple smaller expert models as opposed to a single large network. However, these experts typically operate independently, leaving a question open about whether interconnecting these models could enhance the performance of MoE networks. In response, we introduce GRAPHMOE, a novel method aimed at augmenting the cognitive depth of language models via a self-rethinking mechanism constructed on Pseudo GraphMoE networks. GRAPHMOE employs a recurrent routing strategy to simulate iterative thinking steps, thereby facilitating the flow of information among expert nodes. We implement the GRAPHMOE architecture using Low-Rank Adaptation techniques (LoRA) and conduct extensive experiments on various benchmark datasets. The experimental results reveal that GRAPHMOE outperforms other LoRA based models, achieving state-of-the-art (SOTA) performance. Additionally, this study explores a novel recurrent routing strategy that may inspire further advancements in enhancing the reasoning capabilities of language models.", "authors": ["Bo Lv", "Chen Tang", "Zifan Zheng", "Bohao Yang", "Kun Zhao", "Ning Liao", "Xiaoxing Wang", "Feiyu Xiong", "Zhiyu Li", "Nayu Liu", "Jingchi Jiang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-14", "url": "https://arxiv.org/abs/2501.07890", "pdf_url": "https://arxiv.org/pdf/2501.07890v4", "arxiv_id": "2501.07890", "doi": "10.48550/arXiv.2501.07890", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "8d7425a0e4df3f87db76293c2a65067154bb582c3f98e41de7e771b2f0432796", "sources": ["arxiv", "semantic_scholar"], "title": "Fast sampling and model selection for Bayesian mixture models", "abstract": "We study Bayesian estimation of mixture models and argue in favor of fitting the marginal posterior distribution over component assignments directly, rather than Gibbs sampling from the joint posterior on components and parameters as is commonly done. Some previous authors have found the former approach to have slow mixing, but we show that, implemented correctly, it can achieve excellent performance. In particular, we describe a new Monte Carlo algorithm for sampling from the marginal posterior of a general integrable mixture that makes use of rejection-free sampling from the prior over component assignments to achieve excellent mixing times in typical applications, outperforming standard Gibbs sampling, in some cases by a wide margin. We demonstrate the approach with a selection of applications to Gaussian, Poisson, and categorical models.", "authors": ["M. E. J. Newman"], "categories": ["stat.CO", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-01-13", "url": "https://arxiv.org/abs/2501.07668", "pdf_url": "https://arxiv.org/pdf/2501.07668v2", "arxiv_id": "2501.07668", "doi": "10.1007/s11222-025-10753-0", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Statistics and computing", "quality_score": 0.1193}
{"id": "2db85e009ff307f0db80206caa3de4b5ad61de6c2ebc18dbc590260dcd4d2bdc", "sources": ["arxiv", "semantic_scholar"], "title": "MPCache: MPC-Friendly KV Cache Eviction for Efficient Private LLM Inference", "abstract": "Private large language model (LLM) inference based on secure multi-party computation (MPC) achieves formal data privacy protection but suffers from significant latency overhead, especially for long input sequences. While key-value (KV) cache eviction and sparse attention algorithms have been proposed for efficient LLM inference in plaintext, they are not designed for MPC and cannot benefit private LLM inference directly. In this paper, we propose an accurate and MPC-friendly KV cache eviction framework, dubbed MPCache, building on the observation that historical tokens in a long sequence may have different effects on the downstream decoding. Hence, MPCache combines a look-once static eviction algorithm to discard unimportant KV cache and a query-aware dynamic selection algorithm to activate only a small subset of KV cache for attention computation. MPCache further incorporates a series of optimizations for efficient dynamic KV cache selection, including MPC-friendly similarity approximation, hierarchical KV cache clustering, and cross-layer index-sharing strategy. Extensive experiments demonstrate that MPCache consistently outperforms prior-art KV cache eviction baselines across different generation tasks and achieves 1.8 ~ 2.01x and 3.39 ~ 8.37x decoding latency and communication reduction on different sequence lengths, respectively.", "authors": ["Wenxuan Zeng", "Ye Dong", "Jinjin Zhou", "Jin Tan", "Lei Wang", "Tao Wei", "Runsheng Wang", "Meng Li"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-12", "url": "https://arxiv.org/abs/2501.06807", "pdf_url": "https://arxiv.org/pdf/2501.06807v2", "arxiv_id": "2501.06807", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "cbbfa67c67093ef7329f8f2b9f0144d7b266c7e649ba0e724d9421665e037031", "sources": ["arxiv", "semantic_scholar"], "title": "Mell: Memory-Efficient Large Language Model Serving via Multi-GPU KV Cache Management", "abstract": "Serving large language models (LLMs) for massive users is challenged by the significant memory footprint of the transient state, known as the key-value (KV) cache, which scales with sequence length and number of requests. Instead of renting or buying more expensive GPUs, the load imbalance of the KV cache across GPUs, coupled with recent advances in inter-GPU communication, provides an opportunity to serve more requests via request migration. However, high migration overhead and unpredictable request patterns make it challenging. Therefore, this paper proposes MELL, a memory-efficient LLM serving system via multi-GPU KV cache management. It saves the number of GPUs needed in the system by considering the dynamic KV cache load and the costly request migration. Specifically, we first develop an adaptive request migration mechanism to balance the computational and communication overheads and adapt to diverse resource conditions. Then, we design an online algorithm tailored to a multi-LLM request and multi-GPU scheduling problem with migration enabled. It aims to minimise the required GPUs while limiting the number of migrations. Finally, we implement a prototype of MELL and demonstrate that it reduces the number of GPUs by 31% and increases the GPU utilization by 43% at most compared to existing LLM serving systems.", "authors": ["Liu Qianli", "Hong Zicong", "Chen Fahao", "Li Peng", "Guo Song"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-12", "url": "https://arxiv.org/abs/2501.06709", "pdf_url": "https://arxiv.org/pdf/2501.06709v1", "arxiv_id": "2501.06709", "doi": "10.1109/INFOCOM55648.2025.11044533", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Conference on Computer Communications", "quality_score": 0.1747}
{"id": "1b5f8b2948dd94d0c9d0e0c05627c148a817f4414a5d718059fabebae34994cc", "sources": ["arxiv", "semantic_scholar"], "title": "Flash Window Attention: speedup the attention computation for Swin Transformer", "abstract": "To address the high resolution of image pixels, the Swin Transformer introduces window attention. This mechanism divides an image into non-overlapping windows and restricts attention computation to within each window, significantly enhancing computational efficiency. To further optimize this process, one might consider replacing standard attention with flash attention, which has proven to be more efficient in language models. However, a direct substitution is ineffective. Flash attention is designed for long sequences, whereas window attention deals with shorter sequences but must handle numerous of them in parallel. In this report, we present an optimized solution called Flash Window Attention, tailored specifically for window attention. Flash Window Attention improves attention computation efficiency by up to 300% and enhances end-to-end runtime efficiency by up to 30%. Our code is available online.", "authors": ["Zhendong Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-11", "url": "https://arxiv.org/abs/2501.06480", "pdf_url": "https://arxiv.org/pdf/2501.06480v2", "arxiv_id": "2501.06480", "doi": "10.48550/arXiv.2501.06480", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "e71417cefc7b2b209ac392a5a2454911c234a7fa73db7c9fab8a4ff6ae0c4a62", "sources": ["arxiv", "semantic_scholar"], "title": "Design of a 6-bit Threshold Inverter Quantization (TIQ) Flash Analog to Digital Converter (ADC)", "abstract": "An ADC is used to convert analog signals into binary signals. Compared with many other types of ADCs, flash converters are incredibly quick. A typical Flash ADC consists of 2n resistors, 2n-1 op-amp comparators, and an encoder which requires more area. The resistors and comparators can be eliminated by using threshold inverter quantization (TIQ) comparators. As a voltage comparator, TIQ technique uses two cascaded CMOS inverters. So that there will be no variation in the fabrication process, and temperature. A 6-bit flash ADC based on threshold inverter quantization (TIQ) comparator was designed and software implementation was performed employing a fat tree encoder with 0.25 micrometer CMOS technology. The design consists of 2n-1 TIQ comparator arrays, a gain booster, a 1-out-of-n code generator, and a fat tree encoder. This TIQ flash ADC is simulated with the Tanner EDA Tool. Here the supply voltage is 2.5 V, input frequency of 10 kHz and 10 MHz. The power consumption of the ADC is 6.25 mW, and the propagation delay is 1.07 microseconds for 10 kHz input frequency. For 10 MHz input frequency, power consumption is 12.12 mW and propagation delay is 947.14 ms.", "authors": ["Noyon Kumar Sarkar", "Moumita Roy", "Md. Tariq Hasan"], "categories": ["cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-08", "url": "https://arxiv.org/abs/2501.04627", "pdf_url": "https://arxiv.org/pdf/2501.04627v1", "arxiv_id": "2501.04627", "doi": "10.48550/arXiv.2501.04627", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0183}
{"id": "bc419da60704cd54b63e1669ef239d710d6151af359409e38c261d254eda2477", "sources": ["arxiv", "semantic_scholar"], "title": "PTQ4VM: Post-Training Quantization for Visual Mamba", "abstract": "Visual Mamba is an approach that extends the selective space state model, Mamba, to vision tasks. It processes image tokens sequentially in a fixed order, accumulating information to generate outputs. Despite its growing popularity for delivering high-quality outputs at a low computational cost across various tasks, Visual Mamba is highly susceptible to quantization, which makes further performance improvements challenging. Our analysis reveals that the fixed token access order in Visual Mamba introduces unique quantization challenges, which we categorize into three main issues: 1) token-wise variance, 2) channel-wise outliers, and 3) a long tail of activations. To address these challenges, we propose Post-Training Quantization for Visual Mamba (PTQ4VM), which introduces two key strategies: Per-Token Static (PTS) quantization and Joint Learning of Smoothing Scale and Step Size (JLSS). To the our best knowledge, this is the first quantization study on Visual Mamba. PTQ4VM can be applied to various Visual Mamba backbones, converting the pretrained model to a quantized format in under 15 minutes without notable quality degradation. Extensive experiments on large-scale classification and regression tasks demonstrate its effectiveness, achieving up to 1.83x speedup on GPUs with negligible accuracy loss compared to FP16. Our code is available at https://github.com/YoungHyun197/ptq4vm.", "authors": ["Younghyun Cho", "Changhun Lee", "Seonggon Kim", "Eunhyeok Park"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-29", "url": "https://arxiv.org/abs/2412.20386", "pdf_url": "https://arxiv.org/pdf/2412.20386v2", "arxiv_id": "2412.20386", "doi": "10.1109/WACV61041.2025.00122", "citation_count": 11, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/YoungHyun197/ptq4vm", "venue": "IEEE Workshop/Winter Conference on Applications of Computer Vision", "quality_score": 0.2698}
{"id": "7155ec02f1ccc630d34ba5d19765b3110dc5e44f993c62a6b61230f5f5088b97", "sources": ["arxiv", "semantic_scholar"], "title": "A Survey on Large Language Model Acceleration based on KV Cache Management", "abstract": "Large Language Models (LLMs) have revolutionized a wide range of domains such as natural language processing, computer vision, and multi-modal tasks due to their ability to comprehend context and perform logical reasoning. However, the computational and memory demands of LLMs, particularly during inference, pose significant challenges when scaling them to real-world, long-context, and real-time applications. Key-Value (KV) cache management has emerged as a critical optimization technique for accelerating LLM inference by reducing redundant computations and improving memory utilization. This survey provides a comprehensive overview of KV cache management strategies for LLM acceleration, categorizing them into token-level, model-level, and system-level optimizations. Token-level strategies include KV cache selection, budget allocation, merging, quantization, and low-rank decomposition, while model-level optimizations focus on architectural innovations and attention mechanisms to enhance KV reuse. System-level approaches address memory management, scheduling, and hardware-aware designs to improve efficiency across diverse computing environments. Additionally, the survey provides an overview of both text and multimodal datasets and benchmarks used to evaluate these strategies. By presenting detailed taxonomies and comparative analyses, this work aims to offer useful insights for researchers and practitioners to support the development of efficient and scalable KV cache management techniques, contributing to the practical deployment of LLMs in real-world applications. The curated paper list for KV cache management is in: \\href{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}.", "authors": ["Haoyang Li", "Yiming Li", "Anxin Tian", "Tianhao Tang", "Zhanchao Xu", "Xuejia Chen", "Nicole Hu", "Wei Dong", "Qing Li", "Lei Chen"], "categories": ["cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-27", "url": "https://arxiv.org/abs/2412.19442", "pdf_url": "https://arxiv.org/pdf/2412.19442v3", "arxiv_id": "2412.19442", "doi": "10.48550/arXiv.2412.19442", "citation_count": 126, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}", "venue": null, "quality_score": 0.526}
{"id": "57b094f50fb4fb41da34ecc5f7bbdd8c1811ce420a6c0c15438c1308ed5f6841", "sources": ["arxiv", "semantic_scholar"], "title": "Graph Mixture of Experts and Memory-augmented Routers for Multivariate Time Series Anomaly Detection", "abstract": "Multivariate time series (MTS) anomaly detection is a critical task that involves identifying abnormal patterns or events in data that consist of multiple interrelated time series. In order to better model the complex interdependence between entities and the various inherent characteristics of each entity, the GNN based methods are widely adopted by existing methods. In each layer of GNN, node features aggregate information from their neighboring nodes to update their information. In doing so, from shallow layer to deep layer in GNN, original individual node features continue to be weakened and more structural information,i.e., from short-distance neighborhood to long-distance neighborhood, continues to be enhanced. However, research to date has largely ignored the understanding of how hierarchical graph information is represented and their characteristics that can benefit anomaly detection. Existing methods simply leverage the output from the last layer of GNN for anomaly estimation while neglecting the essential information contained in the intermediate GNN layers. To address such limitations, in this paper, we propose a Graph Mixture of Experts (Graph-MoE) network for multivariate time series anomaly detection, which incorporates the mixture of experts (MoE) module to adaptively represent and integrate hierarchical multi-layer graph information into entity representations. It is worth noting that our Graph-MoE can be integrated into any GNN-based MTS anomaly detection method in a plug-and-play manner. In addition, the memory-augmented routers are proposed in this paper to capture the correlation temporal information in terms of the global historical features of MTS to adaptively weigh the obtained entity representations to achieve successful anomaly estimation. Extensive experiments on five challenging datasets prove the superiority of our approach and each proposed module.", "authors": ["Xiaoyu Huang", "Weidong Chen", "Bo Hu", "Zhendong Mao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-26", "url": "https://arxiv.org/abs/2412.19108", "pdf_url": "https://arxiv.org/pdf/2412.19108v2", "arxiv_id": "2412.19108", "doi": "10.48550/arXiv.2412.19108", "citation_count": 27, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3618}
{"id": "f31e349cf3a0a457a22c6d88c251648893d42d6b3fbc2bad30478ba39013c2d9", "sources": ["arxiv", "semantic_scholar"], "title": "Unified Stochastic Framework for Neural Network Quantization and Pruning", "abstract": "Quantization and pruning are two essential techniques for compressing neural networks, yet they are often treated independently, with limited theoretical analysis connecting them. This paper introduces a unified framework for post-training quantization and pruning using stochastic path-following algorithms. Our approach builds on the Stochastic Path Following Quantization (SPFQ) method, extending its applicability to pruning and low-bit quantization, including challenging 1-bit regimes. By incorporating a scaling parameter and generalizing the stochastic operator, the proposed method achieves robust error correction and yields rigorous theoretical error bounds for both quantization and pruning as well as their combination.", "authors": ["Haoyu Zhang", "Rayan Saab"], "categories": ["cs.LG", "math.NA", "math.PR"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-12-24", "url": "https://arxiv.org/abs/2412.18184", "pdf_url": "https://arxiv.org/pdf/2412.18184v3", "arxiv_id": "2412.18184", "doi": "10.48550/arXiv.2412.18184", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "e259e18a41ccf60bbe999351fd17bd2bfd5a4b1f0eefe488ab3d40f794739042", "sources": ["arxiv", "semantic_scholar"], "title": "Part-Of-Speech Sensitivity of Routers in Mixture of Experts Models", "abstract": "This study investigates the behavior of model-integrated routers in Mixture of Experts (MoE) models, focusing on how tokens are routed based on their linguistic features, specifically Part-of-Speech (POS) tags. The goal is to explore across different MoE architectures whether experts specialize in processing tokens with similar linguistic traits. By analyzing token trajectories across experts and layers, we aim to uncover how MoE models handle linguistic information. Findings from six popular MoE models reveal expert specialization for specific POS categories, with routing paths showing high predictive accuracy for POS, highlighting the value of routing paths in characterizing tokens.", "authors": ["Elie Antoine", "Frédéric Béchet", "Philippe Langlais"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-22", "url": "https://arxiv.org/abs/2412.16971", "pdf_url": "https://arxiv.org/pdf/2412.16971v1", "arxiv_id": "2412.16971", "doi": "10.48550/arXiv.2412.16971", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Computational Linguistics", "quality_score": 0.2113}
{"id": "c7503a552bf3e3f2edd4824607e2a0a4e2d41f7691cb3bc0ee21199d471128d2", "sources": ["arxiv", "semantic_scholar"], "title": "Theory of Mixture-of-Experts for Mobile Edge Computing", "abstract": "In mobile edge computing (MEC) networks, mobile users generate diverse machine learning tasks dynamically over time. These tasks are typically offloaded to the nearest available edge server, by considering communication and computational efficiency. However, its operation does not ensure that each server specializes in a specific type of tasks and leads to severe overfitting or catastrophic forgetting of previous tasks. To improve the continual learning (CL) performance of online tasks, we are the first to introduce mixture-of-experts (MoE) theory in MEC networks and save MEC operation from the increasing generalization error over time. Our MoE theory treats each MEC server as an expert and dynamically adapts to changes in server availability by considering data transfer and computation time. Unlike existing MoE models designed for offline tasks, ours is tailored for handling continuous streams of tasks in the MEC environment. We introduce an adaptive gating network in MEC to adaptively identify and route newly arrived tasks of unknown data distributions to available experts, enabling each expert to specialize in a specific type of tasks upon convergence. We derived the minimum number of experts required to match each task with a specialized, available expert. Our MoE approach consistently reduces the overall generalization error over time, unlike the traditional MEC approach. Interestingly, when the number of experts is sufficient to ensure convergence, adding more experts delays the convergence time and worsens the generalization error. Finally, we perform extensive experiments on real datasets in deep neural networks (DNNs) to verify our theoretical results.", "authors": ["Hongbo Li", "Lingjie Duan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-20", "url": "https://arxiv.org/abs/2412.15690", "pdf_url": "https://arxiv.org/pdf/2412.15690v2", "arxiv_id": "2412.15690", "doi": "10.1109/INFOCOM55648.2025.11044567", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Conference on Computer Communications", "quality_score": 0.3076}
{"id": "5bdf2be45b353b719217b34750529199b511514c761bf9c678eda94f4dffe6b7", "sources": ["arxiv", "semantic_scholar"], "title": "DynamicKV: Task-Aware Adaptive KV Cache Compression for Long Context LLMs", "abstract": "Efficient KV cache management in LLMs is crucial for long-context tasks like RAG and summarization. Existing KV cache compression methods enforce a fixed pattern, neglecting task-specific characteristics and reducing the retention of essential information. However, we observe distinct activation patterns across layers in various tasks, highlighting the need for adaptive strategies tailored to each task's unique demands. Based on this insight, we propose DynamicKV, a method that dynamically optimizes token retention by adjusting the number of tokens retained at each layer to adapt to the specific task. DynamicKV establishes global and per-layer maximum KV cache budgets, temporarily retaining the maximum budget for the current layer, and periodically updating the KV cache sizes of all preceding layers during inference. Our method retains only 1.7% of the KV cache size while achieving ~85% of the Full KV cache performance on LongBench. Notably, even under extreme compression (0.9%), DynamicKV surpasses state-of-the-art (SOTA) methods by 11% in the Needle-in-a-Haystack test using Mistral-7B-Instruct-v0.2. The code will be released.", "authors": ["Xiabin Zhou", "Wenbin Wang", "Minyan Zeng", "Jiaxian Guo", "Xuebo Liu", "Li Shen", "Min Zhang", "Liang Ding"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-19", "url": "https://arxiv.org/abs/2412.14838", "pdf_url": "https://arxiv.org/pdf/2412.14838v4", "arxiv_id": "2412.14838", "doi": "10.48550/arXiv.2412.14838", "citation_count": 29, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3693}
{"id": "2c654e33f5302ded0957efdf975fade8ca2653565d0a671bb75923c88156dc47", "sources": ["arxiv", "semantic_scholar"], "title": "ReMoE: Fully Differentiable Mixture-of-Experts with ReLU Routing", "abstract": "Sparsely activated Mixture-of-Experts (MoE) models are widely adopted to scale up model capacity without increasing the computation budget. However, vanilla TopK routers are trained in a discontinuous, non-differentiable way, limiting their performance and scalability. To address this issue, we propose ReMoE, a fully differentiable MoE architecture that offers a simple yet effective drop-in replacement for the conventional TopK+Softmax routing, utilizing ReLU as the router instead. We further propose methods to regulate the router's sparsity while balancing the load among experts. ReMoE's continuous nature enables efficient dynamic allocation of computation across tokens and layers, while also exhibiting domain specialization. Our experiments demonstrate that ReMoE consistently outperforms vanilla TopK-routed MoE across various model sizes, expert counts, and levels of granularity. Furthermore, ReMoE exhibits superior scalability with respect to the number of experts, surpassing traditional MoE architectures. The implementation based on Megatron-LM is available at https://github.com/thu-ml/ReMoE.", "authors": ["Ziteng Wang", "Jun Zhu", "Jianfei Chen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-19", "url": "https://arxiv.org/abs/2412.14711", "pdf_url": "https://arxiv.org/pdf/2412.14711v2", "arxiv_id": "2412.14711", "doi": "10.48550/arXiv.2412.14711", "citation_count": 52, "influential_citation_count": 6, "has_code": true, "code_url": "https://github.com/thu-ml/ReMoE", "venue": "International Conference on Learning Representations", "quality_score": 0.4311}
{"id": "cece0873cc662e42318c3d562e7c1cc0b07ecd427b13cc39d1ff20ea356d2852", "sources": ["arxiv", "semantic_scholar"], "title": "A Survey on Inference Optimization Techniques for Mixture of Experts Models", "abstract": "The emergence of large-scale Mixture of Experts (MoE) models represents a significant advancement in artificial intelligence, offering enhanced model capacity and computational efficiency through conditional computation. However, deploying and running inference on these models presents significant challenges in computational resources, latency, and energy efficiency. This comprehensive survey analyzes optimization techniques for MoE models across the entire system stack. We first establish a taxonomical framework that categorizes optimization approaches into model-level, system-level, and hardware-level optimizations. At the model level, we examine architectural innovations including efficient expert design, attention mechanisms, various compression techniques such as pruning, quantization, and knowledge distillation, as well as algorithm improvement including dynamic routing strategies and expert merging methods. At the system level, we investigate distributed computing approaches, load balancing mechanisms, and efficient scheduling algorithms that enable scalable deployment. Furthermore, we delve into hardware-specific optimizations and co-design strategies that maximize throughput and energy efficiency. This survey provides both a structured overview of existing solutions and identifies key challenges and promising research directions in MoE inference optimization. To facilitate ongoing updates and the sharing of cutting-edge advances in MoE inference optimization research, we have established a repository accessible at https://github.com/MoE-Inf/awesome-moe-inference/.", "authors": ["Jiacheng Liu", "Peng Tang", "Wenfeng Wang", "Yuhang Ren", "Xiaofeng Hou", "Pheng-Ann Heng", "Minyi Guo", "Chao Li"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-18", "url": "https://arxiv.org/abs/2412.14219", "pdf_url": "https://arxiv.org/pdf/2412.14219v2", "arxiv_id": "2412.14219", "doi": "10.1145/3794845", "citation_count": 44, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/MoE-Inf/awesome-moe-inference/", "venue": "ACM Computing Surveys", "quality_score": 0.4133}
{"id": "97febda891357fcd012bd54078ffde5d70884c0545b1593ba4b46fe4fefb2048", "sources": ["arxiv", "semantic_scholar"], "title": "More Tokens, Lower Precision: Towards the Optimal Token-Precision Trade-off in KV Cache Compression", "abstract": "As large language models (LLMs) process increasing context windows, the memory usage of KV cache has become a critical bottleneck during inference. The mainstream KV compression methods, including KV pruning and KV quantization, primarily focus on either token or precision dimension separately. However, these works leaving the trade-off between these two orthogonal dimensions largely under-explored. In this paper, we comprehensively investigate the token-precision trade-off in KV cache compression.Experiments demonstrate that storing more tokens in the KV cache with lower precision,a strategy we term quantized pruning, can significantly enhance the long-context performance of LLMs. In-depth analysis of the token-precision trade-off across key aspects demonstrates that, quantized pruning achieves substantial improvements in retrieval-related tasks and consistently performs well across varying input lengths. Furthermore, quantized pruning demonstrates notable stability and effectiveness across different KV pruning methods, quantization strategies, and model scales. These findings offer valuable insights into optimizing KV cache compression through balanced token-precision trade-off strategies. Our code is available at https://github.com/zhzihao/QPruningKV.", "authors": ["Jiebin Zhang", "Dawei Zhu", "Yifan Song", "Wenhao Wu", "Chuqiao Kuang", "Xiaoguang Li", "Lifeng Shang", "Qun Liu", "Sujian Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-17", "url": "https://arxiv.org/abs/2412.12706", "pdf_url": "https://arxiv.org/pdf/2412.12706v2", "arxiv_id": "2412.12706", "doi": "10.48550/arXiv.2412.12706", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/zhzihao/QPruningKV", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2258}
{"id": "db131fd47bca7d1d6a52fd83b0a7e4bb99a65800a2c6a6c2df18e20e6c4324bd", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning", "abstract": "Diffusion Policies have become widely used in Imitation Learning, offering several appealing properties, such as generating multimodal and discontinuous behavior. As models are becoming larger to capture more complex capabilities, their computational demands increase, as shown by recent scaling laws. Therefore, continuing with the current architectures will present a computational roadblock. To address this gap, we propose Mixture-of-Denoising Experts (MoDE) as a novel policy for Imitation Learning. MoDE surpasses current state-of-the-art Transformer-based Diffusion Policies while enabling parameter-efficient scaling through sparse experts and noise-conditioned routing, reducing both active parameters by 40% and inference costs by 90% via expert caching. Our architecture combines this efficient scaling with noise-conditioned self-attention mechanism, enabling more effective denoising across different noise levels. MoDE achieves state-of-the-art performance on 134 tasks in four established imitation learning benchmarks (CALVIN and LIBERO). Notably, by pretraining MoDE on diverse robotics data, we achieve 4.01 on CALVIN ABC and 0.95 on LIBERO-90. It surpasses both CNN-based and Transformer Diffusion Policies by an average of 57% across 4 benchmarks, while using 90% fewer FLOPs and fewer active parameters compared to default Diffusion Transformer architectures. Furthermore, we conduct comprehensive ablations on MoDE's components, providing insights for designing efficient and scalable Transformer architectures for Diffusion Policies. Code and demonstrations are available at https://mbreuss.github.io/MoDE_Diffusion_Policy/.", "authors": ["Moritz Reuss", "Jyothish Pari", "Pulkit Agrawal", "Rudolf Lioutikov"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-17", "url": "https://arxiv.org/abs/2412.12953", "pdf_url": "https://arxiv.org/pdf/2412.12953v1", "arxiv_id": "2412.12953", "doi": "10.48550/arXiv.2412.12953", "citation_count": 49, "influential_citation_count": 10, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.5207}
{"id": "d2f2ed27c97fbc8ed532eb1a8eef181d2cefc660906df96bb47abe9c449d748a", "sources": ["arxiv", "semantic_scholar"], "title": "CAMEL: Cross-Attention Enhanced Mixture-of-Experts and Language Bias for Code-Switching Speech Recognition", "abstract": "Code-switching automatic speech recognition (ASR) aims to transcribe speech that contains two or more languages accurately. To better capture language-specific speech representations and address language confusion in code-switching ASR, the mixture-of-experts (MoE) architecture and an additional language diarization (LD) decoder are commonly employed. However, most researches remain stagnant in simple operations like weighted summation or concatenation to fuse languagespecific speech representations, leaving significant opportunities to explore the enhancement of integrating language bias information. In this paper, we introduce CAMEL, a cross-attention-based MoE and language bias approach for code-switching ASR. Specifically, after each MoE layer, we fuse language-specific speech representations with cross-attention, leveraging its strong contextual modeling abilities. Additionally, we design a source attention-based mechanism to incorporate the language information from the LD decoder output into text embeddings. Experimental results demonstrate that our approach achieves state-of-the-art performance on the SEAME, ASRU200, and ASRU700+LibriSpeech460 Mandarin-English code-switching ASR datasets.", "authors": ["He Wang", "Xucheng Wan", "Naijun Zheng", "Kai Liu", "Huan Zhou", "Guojian Li", "Lei Xie"], "categories": ["cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-12-17", "url": "https://arxiv.org/abs/2412.12760", "pdf_url": "https://arxiv.org/pdf/2412.12760v2", "arxiv_id": "2412.12760", "doi": "10.1109/ICASSP49660.2025.10890869", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.1945}
{"id": "61b1b3365483e7d64ac31dde2ead6aa5140c66f1a844f7bf50057cf955b701ee", "sources": ["arxiv", "semantic_scholar"], "title": "SMOSE: Sparse Mixture of Shallow Experts for Interpretable Reinforcement Learning in Continuous Control Tasks", "abstract": "Continuous control tasks often involve high-dimensional, dynamic, and non-linear environments. State-of-the-art performance in these tasks is achieved through complex closed-box policies that are effective, but suffer from an inherent opacity. Interpretable policies, while generally underperforming compared to their closed-box counterparts, advantageously facilitate transparent decision-making within automated systems. Hence, their usage is often essential for diagnosing and mitigating errors, supporting ethical and legal accountability, and fostering trust among stakeholders. In this paper, we propose SMOSE, a novel method to train sparsely activated interpretable controllers, based on a top-1 Mixture-of-Experts architecture. SMOSE combines a set of interpretable decisionmakers, trained to be experts in different basic skills, and an interpretable router that assigns tasks among the experts. The training is carried out via state-of-the-art Reinforcement Learning algorithms, exploiting load-balancing techniques to ensure fair expert usage. We then distill decision trees from the weights of the router, significantly improving the ease of interpretation. We evaluate SMOSE on six benchmark environments from MuJoCo: our method outperforms recent interpretable baselines and narrows the gap with noninterpretable state-of-the-art algorithms", "authors": ["Mátyás Vincze", "Laura Ferrarotti", "Leonardo Lucio Custode", "Bruno Lepri", "Giovanni Iacca"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-17", "url": "https://arxiv.org/abs/2412.13053", "pdf_url": "https://arxiv.org/pdf/2412.13053v1", "arxiv_id": "2412.13053", "doi": "10.1609/aaai.v39i20.35394", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.1505}
{"id": "128c571a36de6ef96d26f607c2f037086e54c753dd66306d34658f626713e3a7", "sources": ["arxiv", "semantic_scholar"], "title": "QPruner: Probabilistic Decision Quantization for Structured Pruning in Large Language Models", "abstract": "The rise of large language models (LLMs) has significantly advanced various natural language processing (NLP) tasks. However, the resource demands of these models pose substantial challenges. Structured pruning is an effective approach to reducing model size, but it often results in significant accuracy degradation, necessitating parameter updates to adapt. Unfortunately, such fine-tuning requires substantial memory, which limits its applicability. To address these challenges, we introduce quantization into the structured pruning framework to reduce memory consumption during both fine-tuning and inference. However, the combined errors from pruning and quantization increase the difficulty of fine-tuning, requiring a more refined quantization scheme. To this end, we propose QPruner, a novel framework that employs structured pruning to reduce model size, followed by a layer-wise mixed-precision quantization scheme. Quantization precisions are assigned to each layer based on their importance to the target task, and Bayesian optimization is employed to refine precision allocation strategies, ensuring a balance between model accuracy and memory efficiency. Extensive experiments on benchmark datasets demonstrate that QPruner significantly outperforms existing methods in memory savings while maintaining or improving model performance.", "authors": ["Changhai Zhou", "Yuhua Zhou", "Shijie Han", "Qian Qiao", "Hongguang Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-16", "url": "https://arxiv.org/abs/2412.11629", "pdf_url": "https://arxiv.org/pdf/2412.11629v1", "arxiv_id": "2412.11629", "doi": "10.48550/arXiv.2412.11629", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.0}
{"id": "408759f6f0135e2049040e15f8d8150bd59bf41656b0284e53f3fb5399a88ef0", "sources": ["arxiv", "semantic_scholar"], "title": "CSR:Achieving 1 Bit Key-Value Cache via Sparse Representation", "abstract": "The emergence of long-context text applications utilizing large language models (LLMs) has presented significant scalability challenges, particularly in memory footprint. The linear growth of the Key-Value (KV) cache responsible for storing attention keys and values to minimize redundant computations can lead to substantial increases in memory consumption, potentially causing models to fail to serve with limited memory resources. To address this issue, we propose a novel approach called Cache Sparse Representation (CSR), which converts the KV cache by transforming the dense Key-Value cache tensor into sparse indexes and weights, offering a more memory-efficient representation during LLM inference. Furthermore, we introduce NeuralDict, a novel neural network-based method for automatically generating the dictionary used in our sparse representation. Our extensive experiments demonstrate that CSR achieves performance comparable to state-of-the-art KV cache quantization algorithms while maintaining robust functionality in memory-constrained environments.", "authors": ["Hongxuan Zhang", "Yao Zhao", "Jiaqi Zheng", "Chenyi Zhuang", "Jinjie Gu", "Guihai Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-16", "url": "https://arxiv.org/abs/2412.11741", "pdf_url": "https://arxiv.org/pdf/2412.11741v1", "arxiv_id": "2412.11741", "doi": "10.48550/arXiv.2412.11741", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.0753}
{"id": "57160748b09dbe176e22c1a7d5321829c3bf64dca684e8e16c0fc9b33a119c56", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Adversarial Robustness of Model-Level Mixture-of-Experts Architectures for Semantic Segmentation", "abstract": "Vulnerability to adversarial attacks is a well-known deficiency of deep neural networks. Larger networks are generally more robust, and ensembling is one method to increase adversarial robustness: each model's weaknesses are compensated by the strengths of others. While an ensemble uses a deterministic rule to combine model outputs, a mixture of experts (MoE) includes an additional learnable gating component that predicts weights for the outputs of the expert models, thus determining their contributions to the final prediction. MoEs have been shown to outperform ensembles on specific tasks, yet their susceptibility to adversarial attacks has not been studied yet. In this work, we evaluate the adversarial vulnerability of MoEs for semantic segmentation of urban and highway traffic scenes. We show that MoEs are, in most cases, more robust to per-instance and universal white-box adversarial attacks and can better withstand transfer attacks. Our code is available at \\url{https://github.com/KASTEL-MobilityLab/mixtures-of-experts/}.", "authors": ["Svetlana Pavlitska", "Enrico Eisen", "J. Marius Zöllner"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-16", "url": "https://arxiv.org/abs/2412.11608", "pdf_url": "https://arxiv.org/pdf/2412.11608v1", "arxiv_id": "2412.11608", "doi": "10.1109/ICMLA61862.2024.00226", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/KASTEL-MobilityLab/mixtures-of-experts/}", "venue": "International Conference on Machine Learning and Applications", "quality_score": 0.2258}
{"id": "3a28e7851917491eecbfff3a5b64ae7c286fcdb538e6e1ffb743d18bc4c35e3a", "sources": ["arxiv", "semantic_scholar"], "title": "Investigating Mixture of Experts in Dense Retrieval", "abstract": "While Dense Retrieval Models (DRMs) have advanced Information Retrieval (IR), one limitation of these neural models is their narrow generalizability and robustness. To cope with this issue, one can leverage the Mixture-of-Experts (MoE) architecture. While previous IR studies have incorporated MoE architectures within the Transformer layers of DRMs, our work investigates an architecture that integrates a single MoE block (SB-MoE) after the output of the final Transformer layer. Our empirical evaluation investigates how SB-MoE compares, in terms of retrieval effectiveness, to standard fine-tuning. In detail, we fine-tune three DRMs (TinyBERT, BERT, and Contriever) across four benchmark collections with and without adding the MoE block. Moreover, since MoE showcases performance variations with respect to its parameters (i.e., the number of experts), we conduct additional experiments to investigate this aspect further. The findings show the effectiveness of SB-MoE especially for DRMs with a low number of parameters (i.e., TinyBERT), as it consistently outperforms the fine-tuned underlying model on all four benchmarks. For DRMs with a higher number of parameters (i.e., BERT and Contriever), SB-MoE requires larger numbers of training samples to yield better retrieval performance.", "authors": ["Effrosyni Sokli", "Pranav Kasela", "Georgios Peikos", "Gabriella Pasi"], "categories": ["cs.IR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-16", "url": "https://arxiv.org/abs/2412.11864", "pdf_url": "https://arxiv.org/pdf/2412.11864v1", "arxiv_id": "2412.11864", "doi": "10.48550/arXiv.2412.11864", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Italian Information Retrieval Workshop", "quality_score": 0.1747}
{"id": "15c05e32f39480998fe608112954b5ff384103408cbece4953d8d47e9add42b8", "sources": ["arxiv", "semantic_scholar"], "title": "GraphMoRE: Mitigating Topological Heterogeneity via Mixture of Riemannian Experts", "abstract": "Real-world graphs have inherently complex and diverse topological patterns, known as topological heterogeneity. Most existing works learn graph representation in a single constant curvature space that is insufficient to match the complex geometric shapes, resulting in low-quality embeddings with high distortion. This also constitutes a critical challenge for graph foundation models, which are expected to uniformly handle a wide variety of diverse graph data. Recent studies have indicated that product manifold gains the possibility to address topological heterogeneity. However, the product manifold is still homogeneous, which is inadequate and inflexible for representing the mixed heterogeneous topology. In this paper, we propose a novel Graph Mixture of Riemannian Experts (GraphMoRE) framework to effectively tackle topological heterogeneity by personalized fine-grained topology geometry pattern preservation. Specifically, to minimize the embedding distortion, we propose a topology-aware gating mechanism to select the optimal embedding space for each node. By fusing the outputs of diverse Riemannian experts with learned gating weights, we construct personalized mixed curvature spaces for nodes, effectively embedding the graph into a heterogeneous manifold with varying curvatures at different points. Furthermore, to fairly measure pairwise distances between different embedding spaces, we present a concise and effective alignment strategy. Extensive experiments on real-world and synthetic datasets demonstrate that our method achieves superior performance with lower distortion, highlighting its potential for modeling complex graphs with topological heterogeneity, and providing a novel architectural perspective for graph foundation models.", "authors": ["Zihao Guo", "Qingyun Sun", "Haonan Yuan", "Xingcheng Fu", "Min Zhou", "Yisen Gao", "Jianxin Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-15", "url": "https://arxiv.org/abs/2412.11085", "pdf_url": "https://arxiv.org/pdf/2412.11085v1", "arxiv_id": "2412.11085", "doi": "10.48550/arXiv.2412.11085", "citation_count": 26, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3578}
{"id": "6111cc453d1ec928613f2182b5e4ce273218e55a0188db1b5f8449ec8da4e400", "sources": ["arxiv", "semantic_scholar"], "title": "DeMo: Decoupled Feature-Based Mixture of Experts for Multi-Modal Object Re-Identification", "abstract": "Multi-modal object Re-IDentification (ReID) aims to retrieve specific objects by combining complementary information from multiple modalities. Existing multi-modal object ReID methods primarily focus on the fusion of heterogeneous features. However, they often overlook the dynamic quality changes in multi-modal imaging. In addition, the shared information between different modalities can weaken modality-specific information. To address these issues, we propose a novel feature learning framework called DeMo for multi-modal object ReID, which adaptively balances decoupled features using a mixture of experts. To be specific, we first deploy a Patch-Integrated Feature Extractor (PIFE) to extract multi-granularity and multi-modal features. Then, we introduce a Hierarchical Decoupling Module (HDM) to decouple multi-modal features into non-overlapping forms, preserving the modality uniqueness and increasing the feature diversity. Finally, we propose an Attention-Triggered Mixture of Experts (ATMoE), which replaces traditional gating with dynamic attention weights derived from decoupled features. With these modules, our DeMo can generate more robust multi-modal features. Extensive experiments on three multi-modal object ReID benchmarks fully verify the effectiveness of our methods. The source code is available at https://github.com/924973292/DeMo.", "authors": ["Yuhao Wang", "Yang Liu", "Aihua Zheng", "Pingping Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-14", "url": "https://arxiv.org/abs/2412.10650", "pdf_url": "https://arxiv.org/pdf/2412.10650v1", "arxiv_id": "2412.10650", "doi": "10.48550/arXiv.2412.10650", "citation_count": 33, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/924973292/DeMo", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3829}
{"id": "c4b3319ffdb7e95843c9a3a3b15ec24be282e6840f591b0df9e53c893208fffb", "sources": ["arxiv", "semantic_scholar"], "title": "SCBench: A KV Cache-Centric Analysis of Long-Context Methods", "abstract": "Long-context LLMs have enabled numerous downstream applications but also introduced significant challenges related to computational and memory efficiency. To address these challenges, optimizations for long-context inference have been developed, centered around the KV cache. However, existing benchmarks often evaluate in single-request, neglecting the full lifecycle of the KV cache in real-world use. This oversight is particularly critical, as KV cache reuse has become widely adopted in LLMs inference frameworks, such as vLLM and SGLang, as well as by LLM providers, including OpenAI, Microsoft, Google, and Anthropic. To address this gap, we introduce SCBench(SharedContextBench), a comprehensive benchmark for evaluating long-context methods from a KV cachecentric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, 4) KV cache loading. Specifically, SCBench uses test examples with shared context, ranging 12 tasks with two shared context modes, covering four categories of long-context capabilities: string retrieval, semantic retrieval, global information, and multi-task. With it, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs, Mamba-Attention hybrids, and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on 8 long-context LLMs. Our findings show that sub-O(n) memory methods suffer in multi-turn scenarios, while sparse encoding with O(n) memory and sub-O(n^2) pre-filling computation perform robustly. Dynamic sparsity yields more expressive KV caches than static patterns, and layer-level sparsity in hybrid architectures reduces memory usage with strong performance. Additionally, we identify attention distribution shift issues in long-generation scenarios. https://aka.ms/SCBench.", "authors": ["Yucheng Li", "Huiqiang Jiang", "Qianhui Wu", "Xufang Luo", "Surin Ahn", "Chengruidong Zhang", "Amir H. Abdi", "Dongsheng Li", "Jianfeng Gao", "Yuqing Yang", "Lili Qiu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-13", "url": "https://arxiv.org/abs/2412.10319", "pdf_url": "https://arxiv.org/pdf/2412.10319v2", "arxiv_id": "2412.10319", "doi": "10.48550/arXiv.2412.10319", "citation_count": 51, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4771}
{"id": "385de9d55f3efe5dd7ec6d60accdedfdc00681f4bf2d9407fc5bf4f04dab22c0", "sources": ["arxiv", "semantic_scholar"], "title": "HashEvict: A Pre-Attention KV Cache Eviction Strategy using Locality-Sensitive Hashing", "abstract": "Transformer-based large language models (LLMs) use the key-value (KV) cache to significantly accelerate inference by storing the key and value embeddings of past tokens. However, this cache consumes significant GPU memory. In this work, we introduce HashEvict, an algorithm that uses locality-sensitive hashing (LSH) to compress the KV cache. HashEvict quickly locates tokens in the cache that are cosine dissimilar to the current query token. This is achieved by computing the Hamming distance between binarized Gaussian projections of the current token query and cached token keys, with a projection length much smaller than the embedding dimension. We maintain a lightweight binary structure in GPU memory to facilitate these calculations. Unlike existing compression strategies that compute attention to determine token retention, HashEvict makes these decisions pre-attention, thereby reducing computational costs. Additionally, HashEvict is dynamic - at every decoding step, the key and value of the current token replace the embeddings of a token expected to produce the lowest attention score. We demonstrate that HashEvict can compress the KV cache by 30%-70% while maintaining high performance across reasoning, multiple-choice, long-context retrieval and summarization tasks.", "authors": ["Minghui Liu", "Tahseen Rabbani", "Tony O'Halloran", "Ananth Sankaralingam", "Mary-Anne Hartley", "Furong Huang", "Cornelia Fermüller", "Yiannis Aloimonos"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.DS", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-13", "url": "https://arxiv.org/abs/2412.16187", "pdf_url": "https://arxiv.org/pdf/2412.16187v3", "arxiv_id": "2412.16187", "doi": "10.48550/arXiv.2412.16187", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "33c3c972647609acfc7c2cddb96f35b01e3e730722385bdc4092466cd0238fa9", "sources": ["arxiv", "semantic_scholar"], "title": "OP-LoRA: The Blessing of Dimensionality", "abstract": "Low-rank adapters (LoRA) enable finetuning of large models with only a small number of parameters. However, they often suffer from an ill-conditioned loss landscape, leading to difficult optimization. Prior work addresses these challenges by aligning adapter updates with full finetuning gradients via custom optimizers, but these methods lack the flexibility to accommodate new adapter architectures and are computationally expensive. We instead introduce OP-LoRA, a novel method which replaces each LoRA adapter with weights predicted by an extra MLP, which is discarded after training. This temporarily allows additional parameters during training to improve optimization, yet requires less wall time than custom optimizers and zero extra cost at inference time because the MLP is discarded. Crucially, extending OP-LoRA to other adapters is as simple as modifying the size of the prediction head for each new adapter type. We show that OP-LoRA allows the optimization to adaptively increase or decrease step size, improving performance and decreasing sensitivity to learning rate. On both small and large-scale LoRA tuning tasks, we observe consistent performance gains of OP-LoRA relative to LoRA and its variants. We achieve especially notable improvements in image generation, with OP-LoRA CMMD scores improving by up to 15 points relative to LoRA. This allows OP-LoRA to achieve the performance of LoRA with half of the inference parameters.", "authors": ["Piotr Teterwak", "Kate Saenko", "Bryan A. Plummer", "Ser-Nam Lim"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-13", "url": "https://arxiv.org/abs/2412.10362", "pdf_url": "https://arxiv.org/pdf/2412.10362v2", "arxiv_id": "2412.10362", "doi": "10.48550/arXiv.2412.10362", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "7461e120dd221dfb237aea57e049507034fd68df4ae1bae40103a2885c686d9d", "sources": ["arxiv", "semantic_scholar"], "title": "Lexico: Extreme KV Cache Compression via Sparse Coding over Universal Dictionaries", "abstract": "We introduce Lexico, a novel KV cache compression method that leverages sparse coding with a universal dictionary. Our key finding is that key-value cache in modern LLMs can be accurately approximated using sparse linear combination from a small, input-agnostic dictionary of ~4k atoms, enabling efficient compression across different input prompts, tasks and models. Using orthogonal matching pursuit for sparse approximation, Lexico achieves flexible compression ratios through direct sparsity control. On GSM8K, across multiple model families (Mistral, Llama 3, Qwen2.5), Lexico maintains 90-95% of the original performance while using only 15-25% of the full KV-cache memory, outperforming both quantization and token eviction methods. Notably, Lexico remains effective in low memory regimes where 2-bit quantization fails, achieving up to 1.7x better compression on LongBench and GSM8K while maintaining high accuracy.", "authors": ["Junhyuck Kim", "Jongho Park", "Jaewoong Cho", "Dimitris Papailiopoulos"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-12", "url": "https://arxiv.org/abs/2412.08890", "pdf_url": "https://arxiv.org/pdf/2412.08890v1", "arxiv_id": "2412.08890", "doi": "10.48550/arXiv.2412.08890", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3076}
{"id": "5437bfedf37d362723010600e5c8911d6bce25749a490e470e1f19af1343c1f2", "sources": ["arxiv", "semantic_scholar"], "title": "ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty", "abstract": "Large Language models (LLMs) have become a research hotspot. To accelerate the inference of LLMs, storing computed caches in memory has become the standard technique. However, as the inference length increases, growing KV caches might lead to out-of-memory issues. Many existing methods address this issue through KV cache compression, primarily by preserving key tokens throughout all layers to reduce information loss. Most of them allocate a uniform budget size for each layer to retain. However, we observe that the minimum budget sizes needed to retain essential information vary across layers and models based on the perspectives of attention and hidden state output. Building on this observation, this paper proposes a simple yet effective KV cache compression method that leverages layer uncertainty to allocate budget size for each layer. Experimental results show that the proposed method can reduce memory usage of the KV caches to only $\\sim$20\\% when compared to Full KV inference while achieving nearly lossless performance.", "authors": ["Meizhi Zhong", "Xikai Liu", "Chen Zhang", "Yikun Lei", "Yan Gao", "Yao Hu", "Kehai Chen", "Min Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-12", "url": "https://arxiv.org/abs/2412.09036", "pdf_url": "https://arxiv.org/pdf/2412.09036v1", "arxiv_id": "2412.09036", "doi": "10.48550/arXiv.2412.09036", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Computational Linguistics", "quality_score": 0.25}
{"id": "88d33069cfce8a2ae7d7dbc80fbc9e5c5071dffa4b85d9c4bd67d00277ff2d6b", "sources": ["arxiv", "semantic_scholar"], "title": "MoSLD: An Extremely Parameter-Efficient Mixture-of-Shared LoRAs for Multi-Task Learning", "abstract": "Recently, LoRA has emerged as a crucial technique for fine-tuning large pre-trained models, yet its performance in multi-task learning scenarios often falls short. In contrast, the MoE architecture presents a natural solution to this issue. However, it introduces challenges such as mutual interference of data across multiple domains and knowledge forgetting of various tasks. Additionally, MoE significantly increases the number of parameters, posing a computational cost challenge. Therefore, in this paper, we propose MoSLD, a mixture-of-shared-LoRAs model with a dropout strategy. MoSLD addresses these challenges by sharing the upper projection matrix in LoRA among different experts, encouraging the model to learn general knowledge across tasks, while still allowing the lower projection matrix to focus on the unique features of each task. The application of dropout alleviates the imbalanced update of parameter matrix and mitigates parameter overfitting in LoRA. Extensive experiments demonstrate that our model exhibits excellent performance in both single-task and multi-task scenarios, with robust out-of-domain generalization capabilities.", "authors": ["Lulu Zhao", "Weihao Zeng", "Xiaofeng Shi", "Hua Zhou"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-12", "url": "https://arxiv.org/abs/2412.08946", "pdf_url": "https://arxiv.org/pdf/2412.08946v1", "arxiv_id": "2412.08946", "doi": "10.48550/arXiv.2412.08946", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Computational Linguistics", "quality_score": 0.1747}
{"id": "d749211ee303a7fcc0941ecfe492dbb01be2754ffdb19989454dcfa012735752", "sources": ["arxiv", "semantic_scholar"], "title": "Optimising TinyML with Quantization and Distillation of Transformer and Mamba Models for Indoor Localisation on Edge Devices", "abstract": "This paper proposes small and efficient machine learning models (TinyML) for resource-constrained edge devices, specifically for on-device indoor localisation. Typical approaches for indoor localisation rely on centralised remote processing of data transmitted from lower powered devices such as wearables. However, there are several benefits for moving this to the edge device itself, including increased battery life, enhanced privacy, reduced latency and lowered operational costs, all of which are key for common applications such as health monitoring. The work focuses on model compression techniques, including quantization and knowledge distillation, to significantly reduce the model size while maintaining high predictive performance. We base our work on a large state-of-the-art transformer-based model and seek to deploy it within low-power MCUs. We also propose a state-space-based architecture using Mamba as a more compact alternative to the transformer. Our results show that the quantized transformer model performs well within a 64 KB RAM constraint, achieving an effective balance between model size and localisation precision. Additionally, the compact Mamba model has strong performance under even tighter constraints, such as a 32 KB of RAM, without the need for model compression, making it a viable option for more resource-limited environments. We demonstrate that, through our framework, it is feasible to deploy advanced indoor localisation models onto low-power MCUs with restricted memory limitations. The application of these TinyML models in healthcare has the potential to revolutionize patient monitoring by providing accurate, real-time location data while minimizing power consumption, increasing data privacy, improving latency and reducing infrastructure costs.", "authors": ["Thanaphon Suwannaphong", "Ferdian Jovan", "Ian Craddock", "Ryan McConville"], "categories": ["cs.LG", "cs.SE"], "fields_of_study": ["Medicine", "Computer Science"], "published_date": "2024-12-12", "url": "https://arxiv.org/abs/2412.09289", "pdf_url": "https://arxiv.org/pdf/2412.09289v1", "arxiv_id": "2412.09289", "doi": "10.1038/s41598-025-94205-9", "citation_count": 38, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Scientific Reports", "quality_score": 0.3978}
{"id": "3f4724d36413390db808198cca30a4e012d431cbb715e1c3e6de8992312d28b7", "sources": ["arxiv", "semantic_scholar"], "title": "EMS: Adaptive Evict-then-Merge Strategy for Head-wise KV Cache Compression Based on Global-Local Importance", "abstract": "As large language models (LLMs) continue to advance, the demand for higher quality and faster processing of long contexts across various applications is growing. KV cache is widely adopted as it stores previously generated key and value tokens, effectively reducing redundant computations during inference. However, as memory overhead becomes a significant concern, efficient compression of KV cache has gained increasing attention. Most existing methods perform compression from two perspectives: identifying important tokens and designing compression strategies. However, these approaches often produce biased distributions of important tokens due to the influence of accumulated attention scores or positional encoding. Furthermore, they overlook the sparsity and redundancy across different heads, which leads to difficulties in preserving the most effective information at the head level. To this end, we propose EMS to overcome these limitations, while achieving better KV cache compression under extreme compression ratios. Specifically, we introduce a Global-Local score that combines accumulated attention scores from both global and local KV tokens to better identify the token importance. For the compression strategy, we design an adaptive and unified Evict-then-Merge framework that accounts for the sparsity and redundancy of KV tokens across different heads. Additionally, we implement the head-wise parallel compression through a zero-class mechanism to enhance efficiency. Extensive experiments demonstrate our SOTA performance even under extreme compression ratios. EMS consistently achieves the lowest perplexity, improves scores by over 1.28 points across four LLMs on LongBench under a 256 cache budget, and preserves 95% retrieval accuracy with a cache budget less than 2% of the context length in the Needle-in-a-Haystack task.", "authors": ["Yingxin Li", "Ye Li", "Yuan Meng", "Xinzhu Ma", "Zihan Geng", "Shutao Xia", "Zhi Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-11", "url": "https://arxiv.org/abs/2412.08521", "pdf_url": "https://arxiv.org/pdf/2412.08521v2", "arxiv_id": "2412.08521", "doi": "10.48550/arXiv.2412.08521", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "78cabdd50f5a370521ba18143b1f55f9ba016e913a0f1041b38fb923dad254fe", "sources": ["arxiv", "semantic_scholar"], "title": "TurboAttention: Efficient Attention Approximation For High Throughputs LLMs", "abstract": "Large language model (LLM) inference demands significant amount of computation and memory, especially in the key attention mechanism. While techniques, such as quantization and acceleration algorithms, like FlashAttention, have improved efficiency of the overall inference, they address different aspects of the problem: quantization focuses on weight-activation operations, while FlashAttention improves execution but requires high-precision formats. Recent Key-value (KV) cache quantization reduces memory bandwidth but still needs floating-point dequantization for attention operation. We present TurboAttention, a comprehensive approach to enable quantized execution of attention that simultaneously addresses both memory and computational efficiency. Our solution introduces two key innovations: FlashQ, a headwise attention quantization technique that enables both compression of KV cache and quantized execution of activation-activation multiplication, and Sparsity-based Softmax Approximation (SAS), which eliminates the need for dequantization to FP32 during exponentiation operation in attention. Experimental results demonstrate that TurboAttention achieves 1.2-1.8x speedup in attention, reduces the KV cache size by over 4.4x, and enables up to 2.37x maximum throughput over the FP16 baseline while outperforming state-of-the-art quantization and compression techniques across various datasets and models.", "authors": ["Hao Kang", "Srikant Bharadwaj", "James Hensman", "Tushar Krishna", "Victor Ruhle", "Saravan Rajmohan"], "categories": ["cs.LG", "cs.AI", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-11", "url": "https://arxiv.org/abs/2412.08585", "pdf_url": "https://arxiv.org/pdf/2412.08585v3", "arxiv_id": "2412.08585", "doi": "10.48550/arXiv.2412.08585", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "80aa1b7f3d8b444da13b3347408b90c9a06893e10e4395d0154db3dc68764b28", "sources": ["arxiv", "semantic_scholar"], "title": "The Computational Limits of State-Space Models and Mamba via the Lens of Circuit Complexity", "abstract": "In this paper, we analyze the computational limitations of Mamba and State-space Models (SSMs) by using the circuit complexity framework. Despite Mamba's stateful design and recent attention as a strong candidate to outperform Transformers, we have demonstrated that both Mamba and SSMs with $\\mathrm{poly}(n)$-precision and constant-depth layers reside within the $\\mathsf{DLOGTIME}$-uniform $\\mathsf{TC}^0$ complexity class. This result indicates Mamba has the same computational capabilities as Transformer theoretically, and it cannot solve problems like arithmetic formula problems, boolean formula value problems, and permutation composition problems if $\\mathsf{TC}^0 \\neq \\mathsf{NC}^1$. Therefore, it challenges the assumption Mamba is more computationally expressive than Transformers. Our contributions include rigorous proofs showing that Selective SSM and Mamba architectures can be simulated by $\\mathsf{DLOGTIME}$-uniform $\\mathsf{TC}^0$ circuits, and they cannot solve problems outside $\\mathsf{TC}^0$.", "authors": ["Yifang Chen", "Xiaoyu Li", "Yingyu Liang", "Zhenmei Shi", "Zhao Song"], "categories": ["cs.CC", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-09", "url": "https://arxiv.org/abs/2412.06148", "pdf_url": "https://arxiv.org/pdf/2412.06148v2", "arxiv_id": "2412.06148", "doi": "10.48550/arXiv.2412.06148", "citation_count": 28, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3656}
{"id": "3df141600987c5048c6ff0e185bf677b2581b388d7c73b09227dacc3bec07f37", "sources": ["arxiv", "semantic_scholar"], "title": "UniPaint: Unified Space-time Video Inpainting via Mixture-of-Experts", "abstract": "In this paper, we present UniPaint, a unified generative space-time video inpainting framework that enables spatial-temporal inpainting and interpolation. Different from existing methods that treat video inpainting and video interpolation as two distinct tasks, we leverage a unified inpainting framework to tackle them and observe that these two tasks can mutually enhance synthesis performance. Specifically, we first introduce a plug-and-play space-time video inpainting adapter, which can be employed in various personalized models. The key insight is to propose a Mixture of Experts (MoE) attention to cover various tasks. Then, we design a spatial-temporal masking strategy during the training stage to mutually enhance each other and improve performance. UniPaint produces high-quality and aesthetically pleasing results, achieving the best quantitative results across various tasks and scale setups. The code and checkpoints are available at $\\href{https://github.com/mmmmm-w/UniPaint}{this \\ repository}$.", "authors": ["Zhen Wan", "Chenyang Qi", "Zhiheng Liu", "Tao Gui", "Yue Ma"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-09", "url": "https://arxiv.org/abs/2412.06340", "pdf_url": "https://arxiv.org/pdf/2412.06340v2", "arxiv_id": "2412.06340", "doi": "10.1109/ICCVW69036.2025.00198", "citation_count": 9, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/mmmmm-w/UniPaint}{this", "venue": null, "quality_score": 0.25}
{"id": "f7c4d075cf4be9533cf7223d27f482dc36a86d48c9a8058cac63ec91833539df", "sources": ["arxiv", "semantic_scholar"], "title": "Federated Split Learning with Model Pruning and Gradient Quantization in Wireless Networks", "abstract": "As a paradigm of distributed machine learning, federated learning typically requires all edge devices to train a complete model locally. However, with the increasing scale of artificial intelligence models, the limited resources on edge devices often become a bottleneck for efficient fine-tuning. To address this challenge, federated split learning (FedSL) implements collaborative training across the edge devices and the server through model splitting. In this paper, we propose a lightweight FedSL scheme, that further alleviates the training burden on resource-constrained edge devices by pruning the client-side model dynamicly and using quantized gradient updates to reduce computation overhead. Additionally, we apply random dropout to the activation values at the split layer to reduce communication overhead. We conduct theoretical analysis to quantify the convergence performance of the proposed scheme. Finally, simulation results verify the effectiveness and advantages of the proposed lightweight FedSL in wireless network environments.", "authors": ["Junhe Zhang", "Wanli Ni", "Dongyu Wang"], "categories": ["cs.LG", "cs.DC", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-09", "url": "https://arxiv.org/abs/2412.06414", "pdf_url": "https://arxiv.org/pdf/2412.06414v2", "arxiv_id": "2412.06414", "doi": "10.1109/TVT.2024.3515083", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Vehicular Technology", "quality_score": 0.294}
{"id": "84fe1f28c0c14e05d7ca4554ebd1029d5114eadc495d39d0c1e03cd246c365b8", "sources": ["arxiv", "semantic_scholar"], "title": "XKV: Personalized KV Cache Memory Reduction for Long-Context LLM Inference", "abstract": "Recently the generative Large Language Model (LLM) has achieved remarkable success in numerous applications. Notably its inference generates output tokens one-by-one, leading to many redundant computations. The widely-used KV-Cache framework makes a compromise between time and space complexities. However, caching data generates the increasingly growing memory demand, that can quickly exhaust the limited memory capacity of the modern accelerator like GPUs, particularly in long-context inference tasks. Existing studies reduce memory consumption by evicting some of cached data that have less important impact on inference accuracy. But the benefit in practice is far from ideal due to the static cache allocation across different LLM network layers. This paper observes that the layer-specific cached data have very different impacts on accuracy. We quantify this difference, and give experimental and theoretical validation. We accordingly make a formal analysis and shows that customizing the cache size for each layer in a personalized manner can yield a significant memory reduction, while still providing comparable accuracy. We simulate the cache allocation as a combinatorial optimization problem and give a global optimal solution. In particular, we devise a mini- and sampling-based inference over a lightweight variant of the LLM model, so as to quickly capture the difference and then feed it into the personalized algorithms. Extensive experiments on real-world datasets demonstrate that our proposals can reduce KV cache memory consumption by 61.6% on average, improve computational efficiency by 2.1x and then increase the throughput by up to 5.5x.", "authors": ["Weizhuo Li", "Zhigang Wang", "Yu Gu", "Ge Yu"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-08", "url": "https://arxiv.org/abs/2412.05896", "pdf_url": "https://arxiv.org/pdf/2412.05896v1", "arxiv_id": "2412.05896", "doi": "10.48550/arXiv.2412.05896", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "a0c6f4b47ab608749c8727f4261828a1a32b808bd090094f0c1a754a5ed5ee7a", "sources": ["arxiv", "semantic_scholar"], "title": "Towards 3D Acceleration for low-power Mixture-of-Experts and Multi-Head Attention Spiking Transformers", "abstract": "Spiking Neural Networks(SNNs) provide a brain-inspired and event-driven mechanism that is believed to be critical to unlock energy-efficient deep learning. The mixture-of-experts approach mirrors the parallel distributed processing of nervous systems, introducing conditional computation policies and expanding model capacity without scaling up the number of computational operations. Additionally, spiking mixture-of-experts self-attention mechanisms enhance representation capacity, effectively capturing diverse patterns of entities and dependencies between visual or linguistic tokens. However, there is currently a lack of hardware support for highly parallel distributed processing needed by spiking transformers, which embody a brain-inspired computation. This paper introduces the first 3D hardware architecture and design methodology for Mixture-of-Experts and Multi-Head Attention spiking transformers. By leveraging 3D integration with memory-on-logic and logic-on-logic stacking, we explore such brain-inspired accelerators with spatially stackable circuitry, demonstrating significant optimization of energy efficiency and latency compared to conventional 2D CMOS integration.", "authors": ["Boxun Xu", "Junyoung Hwang", "Pruek Vanna-iampikul", "Yuxuan Yin", "Sung Kyu Lim", "Peng Li"], "categories": ["cs.NE", "cs.AI", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-07", "url": "https://arxiv.org/abs/2412.05540", "pdf_url": "https://arxiv.org/pdf/2412.05540v1", "arxiv_id": "2412.05540", "doi": "10.48550/arXiv.2412.05540", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "5315e23b751426e092dfbfb9a9912d8cf08c72ad3f11f9676cf15531fd68a411", "sources": ["arxiv", "semantic_scholar"], "title": "Batch-Max: Higher LLM Throughput using Larger Batch Sizes and KV Cache Compression", "abstract": "Several works have developed eviction policies to remove key-value (KV) pairs from the KV cache for more efficient inference. The focus has been on compressing the KV cache after the input prompt has been processed for faster token generation. In settings with limited GPU memory, and when the input context is longer than the generation length, we show that by also compressing the KV cache during the input processing phase, larger batch sizes can be used resulting in significantly higher throughput while still maintaining the original model's accuracy.", "authors": ["Michael R. Metel", "Boxing Chen", "Mehdi Rezagholizadeh"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-07", "url": "https://arxiv.org/abs/2412.05693", "pdf_url": "https://arxiv.org/pdf/2412.05693v3", "arxiv_id": "2412.05693", "doi": "10.48550/arXiv.2412.05693", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "6813a2a6c368c1f29aabd6d46ace92d7b65536a5fa4f83b9e236d594cd326f28", "sources": ["arxiv", "semantic_scholar"], "title": "Cross-Self KV Cache Pruning for Efficient Vision-Language Inference", "abstract": "KV cache pruning has emerged as a promising technique for reducing memory and computation costs in long-context auto-regressive generation. Existing methods for vision-language models (VLMs) typically rely on self-attention scores from large language models (LLMs) to identify and prune irrelevant tokens. However, these approaches overlook the inherent distributional discrepancies between modalities, often leading to inaccurate token importance estimation and the over-pruning of critical visual tokens. To address this, we propose decomposing attention scores into intra-modality attention (within the same modality) and inter-modality attention (across modalities), enabling more precise KV cache pruning by independently managing these distinct attention types. Additionally, we introduce an n-softmax function to counteract distribution shifts caused by pruning, preserving the original smoothness of attention scores and ensuring stable performance. Our final training-free method, \\textbf{C}ross-\\textbf{S}elf \\textbf{P}runing (CSP), achieves competitive performance compared to models with full KV caches while significantly outperforming previous pruning methods. Extensive evaluations on MileBench, a benchmark encompassing 29 multimodal datasets, demonstrate CSP's effectiveness, achieving up to a 41\\% performance improvement on challenging tasks like conversational embodied dialogue while reducing the KV cache budget by 13.6\\%. The code is available at https://github.com/TerryPei/CSP", "authors": ["Xiaohuan Pei", "Tao Huang", "Chang Xu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-05", "url": "https://arxiv.org/abs/2412.04652", "pdf_url": "https://arxiv.org/pdf/2412.04652v1", "arxiv_id": "2412.04652", "doi": "10.48550/arXiv.2412.04652", "citation_count": 10, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/TerryPei/CSP", "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "bbec98800bfdf2f7bafeae8bc037c59283de81cfb8666049943fbebb8f71d9ea", "sources": ["arxiv", "semantic_scholar"], "title": "Customize Segment Anything Model for Multi-Modal Semantic Segmentation with Mixture of LoRA Experts", "abstract": "The recent Segment Anything Model (SAM) represents a significant breakthrough in scaling segmentation models, delivering strong performance across various downstream applications in the RGB modality. However, directly applying SAM to emerging visual modalities, such as depth and event data results in suboptimal performance in multi-modal segmentation tasks. In this paper, we make the first attempt to adapt SAM for multi-modal semantic segmentation by proposing a Mixture of Low-Rank Adaptation Experts (MoE-LoRA) tailored for different input visual modalities. By training only the MoE-LoRA layers while keeping SAM's weights frozen, SAM's strong generalization and segmentation capabilities can be preserved for downstream tasks. Specifically, to address cross-modal inconsistencies, we propose a novel MoE routing strategy that adaptively generates weighted features across modalities, enhancing multi-modal feature integration. Additionally, we incorporate multi-scale feature extraction and fusion by adapting SAM's segmentation head and introducing an auxiliary segmentation head to combine multi-scale features for improved segmentation performance effectively. Extensive experiments were conducted on three multi-modal benchmarks: DELIVER, MUSES, and MCubeS. The results consistently demonstrate that the proposed method significantly outperforms state-of-the-art approaches across diverse scenarios. Notably, under the particularly challenging condition of missing modalities, our approach exhibits a substantial performance gain, achieving an improvement of 32.15% compared to existing methods.", "authors": ["Chenyang Zhu", "Bin Xiao", "Lin Shi", "Shoukun Xu", "Xu Zheng"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-05", "url": "https://arxiv.org/abs/2412.04220", "pdf_url": "https://arxiv.org/pdf/2412.04220v1", "arxiv_id": "2412.04220", "doi": "10.48550/arXiv.2412.04220", "citation_count": 26, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3578}
{"id": "1cea911a8c53ad0635d39a7710f84df33b54615f060423bc825ad57e528fffd2", "sources": ["arxiv", "semantic_scholar"], "title": "Monet: Mixture of Monosemantic Experts for Transformers", "abstract": "Understanding the internal computations of large language models (LLMs) is crucial for aligning them with human values and preventing undesirable behaviors like toxic content generation. However, mechanistic interpretability is hindered by polysemanticity -- where individual neurons respond to multiple, unrelated concepts. While Sparse Autoencoders (SAEs) have attempted to disentangle these features through sparse dictionary learning, they have compromised LLM performance due to reliance on post-hoc reconstruction loss. To address this issue, we introduce Mixture of Monosemantic Experts for Transformers (Monet) architecture, which incorporates sparse dictionary learning directly into end-to-end Mixture-of-Experts pretraining. Our novel expert decomposition method enables scaling the expert count to 262,144 per layer while total parameters scale proportionally to the square root of the number of experts. Our analyses demonstrate mutual exclusivity of knowledge across experts and showcase the parametric knowledge encapsulated within individual experts. Moreover, Monet allows knowledge manipulation over domains, languages, and toxicity mitigation without degrading general performance. Our pursuit of transparent LLMs highlights the potential of scaling expert counts to enhance mechanistic interpretability and directly resect the internal knowledge to fundamentally adjust model behavior. The source code and pretrained checkpoints are available at https://github.com/dmis-lab/Monet.", "authors": ["Jungwoo Park", "Young Jin Ahn", "Kee-Eung Kim", "Jaewoo Kang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-05", "url": "https://arxiv.org/abs/2412.04139", "pdf_url": "https://arxiv.org/pdf/2412.04139v4", "arxiv_id": "2412.04139", "doi": "10.48550/arXiv.2412.04139", "citation_count": 14, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/dmis-lab/Monet", "venue": "International Conference on Learning Representations", "quality_score": 0.294}
{"id": "203c0c1deb9e1e59d5fda5dfdf03dda6c033bb63847cd8dfff163fb44c965a9b", "sources": ["arxiv", "semantic_scholar"], "title": "DiffKV: Differentiated Memory Management for Large Language Models with Parallel KV Compaction", "abstract": "Large language models (LLMs) demonstrate remarkable capabilities but face substantial serving costs due to their high memory demands, with the key-value (KV) cache being a primary bottleneck. State-of-the-art KV cache compression techniques, such as quantization and pruning, apply uniform treatment to both keys and values, and discard unimportant tokens entirely, overlooking the fine-grained distinctions in the significance of individual KV cache components. To address such limitations, we introduce \\textit{DiffKV}, a novel framework for efficient KV cache compression that exploits three levels of differentiation in the KV cache: (1) the differing impact of keys and values on attention computation, (2) the varying importance of tokens, and (3) the diverse dynamic sparsity patterns across attention heads. These levels of differentiation introduce irregular memory usage patterns across different requests and attention heads, posing significant scalability challenges for memory management. To address these challenges, DiffKV proposes an on-GPU memory manager that compacts fragmented free memory list into contiguous regions in parallel, effectively translating sparsity in the KV cache into performance gains. We evaluate DiffKV on several mainstream LLMs, including the emerging thinking models that generate extended chains of thought. DiffKV is able to compress the KV cache by $2.7\\times$ to $5.7\\times$ with near-lossless accuracy on complex workloads requiring sophisticated reasoning and long-generation capabilities, and enhances throughput by $1.9\\times$ to $5.4\\times$. Source codes of DiffKV are available at https://github.com/zyqCSL/DiffKV.", "authors": ["Yanqi Zhang", "Yuwei Hu", "Runyuan Zhao", "John C. S. Lui", "Haibo Chen"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-04", "url": "https://arxiv.org/abs/2412.03131", "pdf_url": "https://arxiv.org/pdf/2412.03131v3", "arxiv_id": "2412.03131", "doi": "10.1145/3731569.3764810", "citation_count": 12, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/zyqCSL/DiffKV", "venue": "Symposium on Operating Systems Principles", "quality_score": 0.2785}
{"id": "3efbfd2d625150ac4982bbcb6168858e8b0d2f3a30ff38e17af049283dde8b60", "sources": ["arxiv", "semantic_scholar"], "title": "ClusterKV: Manipulating LLM KV Cache in Semantic Space for Recallable Compression", "abstract": "Large Language Models (LLMs) have been widely deployed in a variety of applications, and the context length is rapidly increasing to handle tasks such as long-document QA and complex logical reasoning. However, long context poses significant challenges for inference efficiency, including high memory costs of key-value (KV) cache and increased latency due to extensive memory accesses. Recent works have proposed compressing KV cache to approximate computation, but these methods either evict tokens permanently, never recalling them for later inference, or recall previous tokens at the granularity of pages divided by textual positions. Both approaches degrade the model accuracy and output quality. To achieve efficient and accurate recallable KV cache compression, we introduce ClusterKV, which recalls tokens at the granularity of semantic clusters. We design and implement efficient algorithms and systems for clustering, selection, indexing and caching. Experiment results show that ClusterKV attains negligible accuracy loss across various tasks with 32k context lengths, using only a 1k to 2k KV cache budget, and achieves up to a 2$\\times$ speedup in latency and a 2.5$\\times$ improvement in decoding throughput. Compared to SoTA recallable KV compression methods, ClusterKV demonstrates higher model accuracy and output quality, while maintaining or exceeding inference efficiency. Our code is available at https://github.com/sjtu-zhao-lab/ClusterKV.", "authors": ["Guangda Liu", "Chengwei Li", "Jieru Zhao", "Chenqi Zhang", "Minyi Guo"], "categories": ["cs.LG", "cs.AI", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-04", "url": "https://arxiv.org/abs/2412.03213", "pdf_url": "https://arxiv.org/pdf/2412.03213v2", "arxiv_id": "2412.03213", "doi": "10.1109/DAC63849.2025.11132479", "citation_count": 54, "influential_citation_count": 6, "has_code": true, "code_url": "https://github.com/sjtu-zhao-lab/ClusterKV", "venue": "Design Automation Conference", "quality_score": 0.4351}
{"id": "375d94c58f6c755216d4063d729bc380c3769de5fddc26579ea65b4c6da1feb0", "sources": ["arxiv", "semantic_scholar"], "title": "PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation", "abstract": "Recently, large vision-language models (LVLMs) have rapidly gained popularity for their strong generation and reasoning capabilities given diverse multimodal inputs. However, these models incur significant computational and memory overhead during inference, which greatly hinders the efficient deployment in practical scenarios. The extensive key-value (KV) cache, necessitated by the lengthy input and output sequences, notably contributes to the high inference cost. Based on this, recent works have investigated ways to reduce the KV cache size for higher efficiency. Although effective, they generally overlook the distinct importance distributions of KV vectors across layers and maintain the same cache size for each layer during the next token prediction. This results in the significant contextual information loss for certain layers, leading to notable performance decline. To address this, we present PrefixKV, where \"Prefix\" means the top-ranked KV based on importance rather than position in the original sequence. It reframes the challenge of determining KV cache sizes for all layers into the task of searching for the optimal global prefix configuration. With an adaptive layer-wise KV retention recipe based on binary search, the maximum contextual information can thus be preserved in each layer, facilitating the generation. Extensive experiments demonstrate that our method achieves the state-of-the-art performance compared with others. It exhibits superior inference efficiency and generation quality trade-offs, showing promising potential for practical applications. Code is available at https://github.com/THU-MIG/PrefixKV.", "authors": ["Ao Wang", "Hui Chen", "Jiaxin Li", "Jianchao Tan", "Kefeng Zhang", "Xunliang Cai", "Zijia Lin", "Jungong Han", "Guiguang Ding"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-04", "url": "https://arxiv.org/abs/2412.03409", "pdf_url": "https://arxiv.org/pdf/2412.03409v4", "arxiv_id": "2412.03409", "doi": "10.48550/arXiv.2412.03409", "citation_count": 10, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/THU-MIG/PrefixKV", "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "5e1fea87bbdc7af39ce76c70f48e314778f1acf427dbb1e4487bed75f892bcdd", "sources": ["arxiv", "semantic_scholar"], "title": "Convolutional Neural Networks and Mixture of Experts for Intrusion Detection in 5G Networks and beyond", "abstract": "The advent of 6G/NextG networks comes along with a series of benefits, including extreme capacity, reliability, and efficiency. However, these networks may become vulnerable to new security threats. Therefore, 6G/NextG networks must be equipped with advanced Artificial Intelligence algorithms, in order to evade these attacks. Existing studies on the intrusion detection task rely on the train of shallow machine learning classifiers, including Logistic Regression, Decision Trees, and so on, yielding suboptimal performance. Others are based on deep neural networks consisting of static components, which are not conditional on the input. This limits their representation power and efficiency. To resolve these issues, we present the first study integrating Mixture of Experts (MoE) for identifying malicious traffic. Specifically, we use network traffic data and convert the 1D array of features into a 2D matrix. Next, we pass this matrix through convolutional neural network (CNN) layers followed by batch normalization and max pooling layers. After obtaining the representation vector via the CNN layers, a sparsely gated MoE layer is used. This layer consists of a set of experts (dense layers) and a router, where the router assigns weights to the output of each expert. Sparsity is achieved by choosing the most relevant experts of the total ones. Finally, we perform a series of ablation experiments to prove the effectiveness of our proposed model. Experiments are conducted on the 5G-NIDD dataset, a network intrusion detection dataset generated from a real 5G test network. Results show that our introduced approach reaches weighted F1-score up to 99.95% achieving comparable performance to existing approaches. Findings also show that our proposed model achieves multiple advantages over state-of-the-art approaches.", "authors": ["Loukas Ilias", "George Doukas", "Vangelis Lamprou", "Christos Ntanos", "Dimitris Askounis"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2024-12-04", "url": "https://arxiv.org/abs/2412.03483", "pdf_url": "https://arxiv.org/pdf/2412.03483v2", "arxiv_id": "2412.03483", "doi": "10.3389/frai.2025.1708953", "citation_count": 13, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2865}
{"id": "3c195f1a828da9d05356bf0e68a5773a5b4b3b6358bf1a47d8da7e5d287864ef", "sources": ["arxiv", "semantic_scholar"], "title": "Compressing KV Cache for Long-Context LLM Inference with Inter-Layer Attention Similarity", "abstract": "The rapid expansion of context window sizes in Large Language Models~(LLMs) has enabled them to tackle increasingly complex tasks involving lengthy documents. However, this progress comes at the cost of a substantial increase in memory usage during inference, primarily due to the linear growth of the key-value~(KV) cache. Existing KV cache compression methods often discard less relevant tokens, which can lead to significant performance degradation when critical information is lost. In this paper, we propose \\textsc{PoD}~(Proximal tokens over Distant tokens), a novel KV cache compression framework that allocates memory according to token importance, retaining less important tokens in a more compact, shared form rather than discarding them entirely. Our approach is motivated by two key observations: (1) proximal tokens -- those at the beginning and end of the context -- are significantly more important for next-token prediction, and (2) attention scores for distant tokens are highly redundant across consecutive layers. Leveraging these insights, \\textsc{PoD} preserves the full KV cache for proximal tokens, while for distant tokens, it shares key states across layers. Since attention scores are determined by both queries and keys, sharing key states enables multiple layers to reuse a single set of keys for distant tokens, substantially reducing KV cache memory without discarding essential context. We further introduce a lightweight post-training adaptation to enable the model to adjust to this new attention-sharing structure. Extensive experiments on both synthetic~(Needle in a Haystack) and real-world long-context benchmarks demonstrate that \\textsc{PoD} reduces KV cache memory usage by up to 35\\% without compromising performance. Our method is orthogonal to existing token-selection-based techniques and can be combined with them for further KV cache compression.", "authors": ["Da Ma", "Lu Chen", "Situo Zhang", "Yuxun Miao", "Su Zhu", "Zhi Chen", "Hongshen Xu", "Hanqi Li", "Shuai Fan", "Lei Pan", "Kai Yu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-03", "url": "https://arxiv.org/abs/2412.02252", "pdf_url": "https://arxiv.org/pdf/2412.02252v2", "arxiv_id": "2412.02252", "doi": "10.48550/arXiv.2412.02252", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "25328f3dd5b532f24d0ad4099a111f78ab680f3f833b16ddfa83516f44dea4a1", "sources": ["arxiv", "semantic_scholar"], "title": "VA-MoE: Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting", "abstract": "This paper presents Variables Adaptive Mixture of Experts (VAMoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VAMoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct subpatterns of atmospheric variables (temperature, humidity, wind speed). Moreover, the proposed method employs a variable adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on real world ERA5 dataset demonstrate that VAMoE performs comparable against SoTA models in both short term (1 days) and long term (5 days) forecasting tasks, with only about 25% of trainable parameters and 50% of the initial training data.", "authors": ["Hao Chen", "Han Tao", "Guo Song", "Jie Zhang", "Yunlong Yu", "Yonghan Dong", "Lei Bai"], "categories": ["cs.LG", "physics.ao-ph"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2024-12-03", "url": "https://arxiv.org/abs/2412.02503", "pdf_url": "https://arxiv.org/pdf/2412.02503v2", "arxiv_id": "2412.02503", "doi": "10.1109/ICCV51701.2025.00742", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.2113}
{"id": "78bbf146619d9e58aecb8ca374746be62c18f330d0bfe954317ef400238bfdc6", "sources": ["arxiv", "semantic_scholar"], "title": "RILQ: Rank-Insensitive LoRA-based Quantization Error Compensation for Boosting 2-bit Large Language Model Accuracy", "abstract": "Low-rank adaptation (LoRA) has become the dominant method for parameter-efficient LLM fine-tuning, with LoRA-based quantization error compensation (LQEC) emerging as a powerful tool for recovering accuracy in compressed LLMs. However, LQEC has underperformed in sub-4-bit scenarios, with no prior investigation into understanding this limitation. We propose RILQ (Rank-Insensitive LoRA-based Quantization Error Compensation) to understand fundamental limitation and boost 2-bit LLM accuracy. Based on rank analysis revealing model-wise activation discrepancy loss's rank-insensitive nature, RILQ employs this loss to adjust adapters cooperatively across layers, enabling robust error compensation with low-rank adapters. Evaluations on LLaMA-2 and LLaMA-3 demonstrate RILQ's consistent improvements in 2-bit quantized inference across various state-of-the-art quantizers and enhanced accuracy in task-specific fine-tuning. RILQ maintains computational efficiency comparable to existing LoRA methods, enabling adapter-merged weight-quantized LLM inference with significantly enhanced accuracy, making it a promising approach for boosting 2-bit LLM performance. Our code is available at https://github.com/aiha-lab/RILQ.", "authors": ["Geonho Lee", "Janghwan Lee", "Sukjin Hong", "Minsoo Kim", "Euijai Ahn", "Du-Seong Chang", "Jungwook Choi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-02", "url": "https://arxiv.org/abs/2412.01129", "pdf_url": "https://arxiv.org/pdf/2412.01129v3", "arxiv_id": "2412.01129", "doi": "10.48550/arXiv.2412.01129", "citation_count": 9, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/aiha-lab/RILQ", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.25}
{"id": "88af4582e15415f7c972ebda90b12d42fb191c9118af42bb5302a920daef3c5d", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning in VLMs", "abstract": "Large vision-language models (LVLMs) generally contain significantly more visual tokens than their textual counterparts, resulting in a considerable computational burden. Recent efforts have been made to tackle this issue by pruning visual tokens early within the language model. Most existing works use attention scores between text and visual tokens to assess the importance of visual tokens. However, in this study, we first analyze the text-visual attention in the language model and find that this score is not an ideal indicator for token pruning. Based on the analysis, We propose VisPruner, a plug-and-play method that utilizes visual cues for more effective token pruning in LVLMs. Specifically, we first use visual attention to select a limited number of significant tokens. Then, we remove duplicate tokens from the remaining ones based on their similarity. By retaining diverse tokens alongside the initially selected important tokens, we maximally preserve the visual information of the input image. Experimental results demonstrate that our VisPruner sustains strong performance across various VLM architectures and reduction ratios, significantly outperforming existing methods based on text-visual attention. Notably, without any training, VisPruner can reduce the FLOPs of LLaVA-1.5-7B by 91% and inference latency by 75%, while maintaining comparable performance. Our code is available at https://github.com/Theia-4869/VisPruner.", "authors": ["Qizhe Zhang", "Aosong Cheng", "Ming Lu", "Renrui Zhang", "Zhiyong Zhuo", "Jiajun Cao", "Shaobo Guo", "Qi She", "Shanghang Zhang"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-02", "url": "https://arxiv.org/abs/2412.01818", "pdf_url": "https://arxiv.org/pdf/2412.01818v2", "arxiv_id": "2412.01818", "doi": "10.1109/ICCV51701.2025.01939", "citation_count": 102, "influential_citation_count": 18, "has_code": true, "code_url": "https://github.com/Theia-4869/VisPruner", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.6394}
{"id": "ae7435fff06d3e991e36e5bd7163541808b9456cd8872ca2a035345b79c972b9", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts for Node Classification", "abstract": "Nodes in the real-world graphs exhibit diverse patterns in numerous aspects, such as degree and homophily. However, most existent node predictors fail to capture a wide range of node patterns or to make predictions based on distinct node patterns, resulting in unsatisfactory classification performance. In this paper, we reveal that different node predictors are good at handling nodes with specific patterns and only apply one node predictor uniformly could lead to suboptimal result. To mitigate this gap, we propose a mixture of experts framework, MoE-NP, for node classification. Specifically, MoE-NP combines a mixture of node predictors and strategically selects models based on node patterns. Experimental results from a range of real-world datasets demonstrate significant performance improvements from MoE-NP.", "authors": ["Yu Shi", "Yiqi Wang", "WeiXuan Lang", "Jiaxin Zhang", "Pan Dong", "Aiping Li"], "categories": ["cs.SI", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-30", "url": "https://arxiv.org/abs/2412.00418", "pdf_url": "https://arxiv.org/pdf/2412.00418v3", "arxiv_id": "2412.00418", "doi": "10.1145/3731715.3733392", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Multimedia Retrieval", "quality_score": 0.1747}
{"id": "03d59323fbc13c4b9287a41d3777dece3268b6a68405cade0c9f6c02c7009009", "sources": ["arxiv", "semantic_scholar"], "title": "Pruned Convolutional Attention Network Based Wideband Spectrum Sensing with Sub-Nyquist Sampling", "abstract": "Wideband spectrum sensing (WSS) is critical for orchestrating multitudinous wireless transmissions via spectrum sharing, but may incur excessive costs of hardware, power and computation due to the high sampling rate. In this article, a deep learning based WSS framework embedding the multicoset preprocessing is proposed to enable the low-cost sub-Nyquist sampling. A pruned convolutional attention WSS network (PCA-WSSNet) is designed to organically integrate the multicoset preprocessing and the convolutional attention mechanism as well as to reduce the model complexity remarkably via the selective weight pruning without the performance loss. Furthermore, a transfer learning (TL) strategy benefiting from the model pruning is developed to improve the robustness of PCA-WSSNet with few adaptation samples of new scenarios. Simulation results show the performance superiority of PCA-WSSNet over the state of the art. Compared with direct TL, the pruned TL strategy can simultaneously improve the prediction accuracy in unseen scenarios, reduce the model size, and accelerate the model inference.", "authors": ["Peihao Dong", "Jibin Jia", "Shen Gao", "Fuhui Zhou", "Qihui Wu"], "categories": ["eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-11-30", "url": "https://arxiv.org/abs/2412.00562", "pdf_url": "https://arxiv.org/pdf/2412.00562v1", "arxiv_id": "2412.00562", "doi": "10.1109/TVT.2024.3507159", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Vehicular Technology", "quality_score": 0.1747}
{"id": "66c423e69bba5501ebefc471e1e8ccda2dc089ccdba7160d704c5116f37032f2", "sources": ["arxiv", "semantic_scholar"], "title": "Is Oracle Pruning the True Oracle?", "abstract": "Oracle pruning, which selects unimportant weights by minimizing the pruned train loss, has been taken as the foundation for most neural network pruning methods for over 35 years, while few (if not none) have thought about how much the foundation really holds. This paper, for the first time, attempts to examine its validity on modern deep models through empirical correlation analyses and provide reflections on the field of neural network pruning. Specifically, for a typical pruning algorithm with three stages (pertaining, pruning, and retraining), we analyze the model performance correlation before and after retraining. Extensive experiments (37K models are trained) across a wide spectrum of models (LeNet5, VGG, ResNets, ViT, MLLM) and datasets (MNIST and its variants, CIFAR10/CIFAR100, ImageNet-1K, MLLM data) are conducted. The results lead to a surprising conclusion: on modern deep learning models, the performance before retraining is barely correlated with the performance after retraining. Namely, the weights selected by oracle pruning can hardly guarantee a good performance after retraining. This further implies that existing works using oracle pruning to derive pruning criteria may be groundless from the beginning. Further studies suggest the rising task complexity is one factor that makes oracle pruning invalid nowadays. Finally, given the evidence, we argue that the retraining stage in a pruning algorithm should be accounted for when developing any pruning criterion.", "authors": ["Sicheng Feng", "Keda Tao", "Huan Wang"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-28", "url": "https://arxiv.org/abs/2412.00143", "pdf_url": "https://arxiv.org/pdf/2412.00143v1", "arxiv_id": "2412.00143", "doi": "10.48550/arXiv.2412.00143", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "0a82ebe970ad076f709ceb9928998b2c1746c4088c020de15ba56a93f623ba81", "sources": ["arxiv", "semantic_scholar"], "title": "On the Role of Discrete Representation in Sparse Mixture of Experts", "abstract": "Sparse mixture of experts (SMoE) is an effective solution for scaling up model capacity without increasing the computational costs. A crucial component of SMoE is the router, responsible for directing the input to relevant experts; however, it also presents a major weakness, leading to routing inconsistencies and representation collapse issues. Instead of fixing the router like previous works, we propose an alternative that assigns experts to input via indirection, which employs the discrete representation of input that points to the expert. The discrete representations are learnt via vector quantization, resulting in a new architecture dubbed Vector-Quantized Mixture of Experts (VQMoE). We provide theoretical support and empirical evidence demonstrating the VQMoE's ability to overcome the challenges present in traditional routers. Through extensive evaluations on both large language models and vision tasks for pre-training and fine-tuning, we show that VQMoE achieves a 28% improvement in robustness compared to other SMoE routing methods, while maintaining strong performance in fine-tuning tasks.", "authors": ["Giang Do", "Kha Pham", "Hung Le", "Truyen Tran"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-28", "url": "https://arxiv.org/abs/2411.19402", "pdf_url": "https://arxiv.org/pdf/2411.19402v2", "arxiv_id": "2411.19402", "doi": null, "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "02efbc98688fdc517d8dbff5a74cf85c98b6d2f73b43dbc3a65f98fcb786f6a6", "sources": ["arxiv", "semantic_scholar"], "title": "MiniKV: Pushing the Limits of LLM Inference via 2-Bit Layer-Discriminative KV Cache", "abstract": "How to efficiently serve LLMs in practice has become exceptionally challenging due to their prohibitive memory and computation requirements. In this study, we investigate optimizing the KV cache, whose memory footprint poses a critical bottleneck in LLM inference, especially when dealing with long context tasks. To tackle the challenge, we introduce MiniKV, a KV cache optimization method that simultaneously preserves long context task accuracy while significantly reducing KV cache size via a novel 2-bit layer-discriminative KV cache. More importantly, we develop specialized CUDA kernels to make MiniKV compatible with FlashAttention. Experiments on a wide range of long context tasks show that MiniKV effectively achieves 86% KV cache compression ratio while recovering over 98.5% of accuracy, outperforming state-of-the-art methods while achieving excellent measured system performance improvements.", "authors": ["Akshat Sharma", "Hangliang Ding", "Jianping Li", "Neel Dani", "Minjia Zhang"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-27", "url": "https://arxiv.org/abs/2411.18077", "pdf_url": "https://arxiv.org/pdf/2411.18077v3", "arxiv_id": "2411.18077", "doi": "10.48550/arXiv.2411.18077", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "db6f0d9ff1ace3b25075bb1ceb1fc757ef7ad010fdf8acf25b53688f4801514c", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Cache-Conditional Experts for Efficient Mobile Device Inference", "abstract": "Mixture of Experts (MoE) LLMs have recently gained attention for their ability to enhance performance by selectively engaging specialized subnetworks or \"experts\" for each input. However, deploying MoEs on memory-constrained devices remains challenging, particularly when generating tokens sequentially with a batch size of one, as opposed to typical high-throughput settings involving long sequences or large batches. In this work, we optimize MoE on memory-constrained devices where only a subset of expert weights fit in DRAM. We introduce a novel cache-aware routing strategy that leverages expert reuse during token generation to improve cache locality. We evaluate our approach on language modeling, MMLU, and GSM8K benchmarks and present on-device results demonstrating 2$\\times$ speedups on mobile devices, offering a flexible, training-free solution to extend MoE's applicability across real-world applications.", "authors": ["Andrii Skliar", "Ties van Rozendaal", "Romain Lepert", "Todor Boinovski", "Mart van Baalen", "Markus Nagel", "Paul Whatmough", "Babak Ehteshami Bejnordi"], "categories": ["cs.LG", "cs.AI", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-27", "url": "https://arxiv.org/abs/2412.00099", "pdf_url": "https://arxiv.org/pdf/2412.00099v2", "arxiv_id": "2412.00099", "doi": "10.48550/arXiv.2412.00099", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2603}
{"id": "b95278f315847201dc4d457335669c7bbf472e05dbc0fe82753bb7d9ed45abf5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts in Image Classification: What's the Sweet Spot?", "abstract": "Mixture-of-Experts (MoE) models have shown promising potential for parameter-efficient scaling across domains. However, their application to image classification remains limited, often requiring billion-scale datasets to be competitive. In this work, we explore the integration of MoE layers into image classification architectures using open datasets. We conduct a systematic analysis across different MoE configurations and model scales. We find that moderate parameter activation per sample provides the best trade-off between performance and efficiency. However, as the number of activated parameters increases, the benefits of MoE diminish. Our analysis yields several practical insights for vision MoE design. First, MoE layers most effectively strengthen tiny and mid-sized models, while gains taper off for large-capacity networks and do not redefine state-of-the-art ImageNet performance. Second, a Last-2 placement heuristic offers the most robust cross-architecture choice, with Every-2 slightly better for Vision Transform (ViT), and both remaining effective as data and model scale increase. Third, larger datasets (e.g., ImageNet-21k) allow more experts, up to 16, for ConvNeXt to be utilized effectively without changing placement, as increased data reduces overfitting and promotes broader expert specialization. Finally, a simple linear router performs best, suggesting that additional routing complexity yields no consistent benefit.", "authors": ["Mathurin Videau", "Alessandro Leite", "Marc Schoenauer", "Olivier Teytaud"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-27", "url": "https://arxiv.org/abs/2411.18322", "pdf_url": "https://arxiv.org/pdf/2411.18322v2", "arxiv_id": "2411.18322", "doi": "10.48550/arXiv.2411.18322", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "cba7a512aeeeaa24e27c0e0dabcda41b8d3cbd470c0d78a3f410bd85bece7540", "sources": ["arxiv", "semantic_scholar"], "title": "SEUF: Is Unlearning One Expert Enough for Mixture-of-Experts LLMs?", "abstract": "Recent advancements in LLMs unlearning have shown remarkable success in removing unwanted data-model influences while preserving the model's utility for legitimate knowledge. Despite these strides, sparse Mixture-of-Experts (MoE) LLMs--a key subset of the LLM family--have remained unexplored in the context of unlearning. As MoE LLMs are celebrated for their exceptional performance, we ask:How can unlearning be performed effectively and efficiently on MoE LLMs? Our pilot study shows that the dynamic routing nature of MoE LLMs introduces unique challenges, leading to excessive forgetting, uncontrolled knowledge erasure and substantial utility drops when existing unlearning methods are applied. To address this, we propose a novel Selected-Expert Unlearning Framework (SEUF). Through expert attribution, unlearning is concentrated on the most actively engaged experts for the specified knowledge. Concurrently, an anchor loss is applied to the router to stabilize the active state of this targeted expert, ensuring focused and controlled unlearning. SEUF is compatible with various standard unlearning algorithms. Extensive experiments demonstrate that SEUF enhances both forget quality up to 5% and model utility by 35% on MoE LLMs across various benchmarks and LLM architectures (compared to standard unlearning algorithms), while only unlearning 0.06% of the model parameters.", "authors": ["Haomin Zhuang", "Yihua Zhang", "Kehan Guo", "Jinghan Jia", "Gaowen Liu", "Sijia Liu", "Xiangliang Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-27", "url": "https://arxiv.org/abs/2411.18797", "pdf_url": "https://arxiv.org/pdf/2411.18797v2", "arxiv_id": "2411.18797", "doi": "10.18653/v1/2025.acl-long.424", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2785}
{"id": "b334ec4420adc3db0bd456966dec1b5d6af179ada79ecf0f7b3fff561efb5b6b", "sources": ["arxiv", "semantic_scholar"], "title": "KVPR: Efficient LLM Inference with I/O-Aware KV Cache Partial Recomputation", "abstract": "Inference for Large Language Models (LLMs) is computationally demanding. To reduce the cost of auto-regressive decoding, Key-Value (KV) cache is used to store intermediate activations, which significantly lowers the computational overhead for token generation. However, the memory required for the KV cache grows rapidly, often exceeding the capacity of GPU memory. A cost-effective alternative is to offload KV cache to CPU memory, which alleviates GPU memory pressure, but shifts the bottleneck to the limited bandwidth of the PCIe connection between the CPU and GPU. Existing methods attempt to address these issues by overlapping GPU computation with I/O or employing CPU-GPU heterogeneous execution, but they are hindered by excessive data movement and dependence on CPU capabilities. Fully overlapping PCIe communication latency gets challenging as the size of the KV cache grows and/or the GPU compute capabilities increase. In this paper, we introduce KVPR, an efficient I/O-aware LLM inference method where the CPU first transfers a partial set of activations, from which the GPU can start recomputing the KV cache values. While the GPU recomputes the partial KV cache, the remaining portion of the KV cache is transferred concurrently from the CPU. This approach overlaps GPU recomputation with KV cache transfer to minimize idle GPU time and maximize inference performance. KVPR is fully automated by integrating a profiler module that utilizes input characteristics and system hardware information, a scheduler module to optimize the distribution of computation and communication workloads, and a runtime module to efficiently execute the derived execution plan. Experimental results show that KVPR achieves up to 35.8% lower latency and 46.2% higher throughput during decoding compared to state-of-the-art approaches. The code is available at https://github.com/chaoyij/KVPR.", "authors": ["Chaoyi Jiang", "Lei Gao", "Hossein Entezari Zarch", "Murali Annavaram"], "categories": ["cs.LG", "cs.DC", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-26", "url": "https://arxiv.org/abs/2411.17089", "pdf_url": "https://arxiv.org/pdf/2411.17089v2", "arxiv_id": "2411.17089", "doi": "10.18653/v1/2025.findings-acl.997", "citation_count": 13, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/chaoyij/KVPR", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2865}
{"id": "fc388ddd7805685035930d647db8de10e875f5009c0ab0e290896bcdb9eb262a", "sources": ["arxiv", "semantic_scholar"], "title": "MH-MoE: Multi-Head Mixture-of-Experts", "abstract": "Multi-Head Mixture-of-Experts (MH-MoE) demonstrates superior performance by using the multi-head mechanism to collectively attend to information from various representation spaces within different experts. In this paper, we present a novel implementation of MH-MoE that maintains both FLOPs and parameter parity with sparse Mixture of Experts models. Experimental results on language models show that the new implementation yields quality improvements over both vanilla MoE and fine-grained MoE models. Additionally, our experiments demonstrate that MH-MoE is compatible with 1-bit Large Language Models (LLMs) such as BitNet.", "authors": ["Shaohan Huang", "Xun Wu", "Shuming Ma", "Furu Wei"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-25", "url": "https://arxiv.org/abs/2411.16205", "pdf_url": "https://arxiv.org/pdf/2411.16205v3", "arxiv_id": "2411.16205", "doi": "10.48550/arXiv.2411.16205", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "02a146351417013aa4e102511e8d19fa3f2cc2ca033e4b5d0f87a4c40112fa69", "sources": ["arxiv", "semantic_scholar"], "title": "A Method for Building Large Language Models with Predefined KV Cache Capacity", "abstract": "This paper introduces a novel approach, the Bounded-Cache Transformer (BCT), for building large language models with a predefined Key-Value (KV) cache capacity. The BCT addresses the excessive memory consumption issue in traditional KV caches by implementing a bounded-length KV cache, which is particularly suitable for the attention layers in Transformer decode-only architectures. By dynamically updating the key-value vector sequences, the BCT achieves efficient inference within limited cache capacity, significantly reducing memory usage while maintaining model performance and system throughput. Experimental results demonstrate that the BCT significantly reduces memory usage while maintaining the model's inference quality, offering a new solution for efficient inference in large language models.", "authors": ["Zhonghua Yi", "Ge Niu", "Lei Wang", "Wei Tang", "Liqiu Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-24", "url": "https://arxiv.org/abs/2411.15785", "pdf_url": "https://arxiv.org/pdf/2411.15785v2", "arxiv_id": "2411.15785", "doi": "10.48550/arXiv.2411.15785", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "30236fa6175bf92e1bfd2ee0523ed6e99d9d81185ddb24e523e07e9912a3b32e", "sources": ["arxiv", "semantic_scholar"], "title": "LeMoLE: LLM-Enhanced Mixture of Linear Experts for Time Series Forecasting", "abstract": "Recent research has shown that large language models (LLMs) can be effectively used for real-world time series forecasting due to their strong natural language understanding capabilities. However, aligning time series into semantic spaces of LLMs comes with high computational costs and inference complexity, particularly for long-range time series generation. Building on recent advancements in using linear models for time series, this paper introduces an LLM-enhanced mixture of linear experts for precise and efficient time series forecasting. This approach involves developing a mixture of linear experts with multiple lookback lengths and a new multimodal fusion mechanism. The use of a mixture of linear experts is efficient due to its simplicity, while the multimodal fusion mechanism adaptively combines multiple linear experts based on the learned features of the text modality from pre-trained large language models. In experiments, we rethink the need to align time series to LLMs by existing time-series large language models and further discuss their efficiency and effectiveness in time series forecasting. Our experimental results show that the proposed LeMoLE model presents lower prediction errors and higher computational efficiency than existing LLM models.", "authors": ["Lingzheng Zhang", "Lifeng Shen", "Yimin Zheng", "Shiyuan Piao", "Ziyue Li", "Fugee Tsung"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-24", "url": "https://arxiv.org/abs/2412.00053", "pdf_url": "https://arxiv.org/pdf/2412.00053v1", "arxiv_id": "2412.00053", "doi": "10.48550/arXiv.2412.00053", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "66244004a5e9fdd53bc9371f592ab28799def00ca06ba2697c8f55ecc46001a6", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-FAIR: Federated LoRA Fine-Tuning with Aggregation and Initialization Refinement", "abstract": "Foundation models (FMs) achieve strong performance across diverse tasks with task-specific fine-tuning, yet full parameter fine-tuning is often computationally prohibitive for large models. Parameter-efficient fine-tuning (PEFT) methods like Low-Rank Adaptation (LoRA) reduce this cost by introducing low-rank matrices for tuning fewer parameters. While LoRA allows for efficient fine-tuning, it requires significant data for adaptation, making Federated Learning (FL) an appealing solution due to its privacy-preserving collaborative framework. However, combining LoRA with FL introduces two key challenges: the \\textbf{Server-Side Aggregation Bias}, where server-side averaging of LoRA matrices diverges from the ideal global update, and the \\textbf{Client-Side Initialization Lag}, emphasizing the need for consistent initialization across rounds. Existing approaches address these challenges individually, limiting their effectiveness. We propose LoRA-FAIR, a novel method that tackles both issues by introducing a correction term on the server, enhancing aggregation efficiency and accuracy. LoRA-FAIR maintains computational and communication efficiency, yielding superior performance over state-of-the-art methods. Experimental results on ViT and MLP-Mixer models across large-scale datasets demonstrate that LoRA-FAIR consistently achieves performance improvements in FL settings.", "authors": ["Jieming Bian", "Lei Wang", "Letian Zhang", "Jie Xu"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-22", "url": "https://arxiv.org/abs/2411.14961", "pdf_url": "https://arxiv.org/pdf/2411.14961v3", "arxiv_id": "2411.14961", "doi": "10.1109/ICCV51701.2025.00356", "citation_count": 32, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/jmbian/LoRA-FAIR", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.3796}
{"id": "71754185e75e908cd2d61c6d01cef294b005fcf5a33391e6581eb991af1530ec", "sources": ["arxiv", "semantic_scholar"], "title": "AutoMixQ: Self-Adjusting Quantization for High Performance Memory-Efficient Fine-Tuning", "abstract": "Fine-tuning large language models (LLMs) under resource constraints is a significant challenge in deep learning. Low-Rank Adaptation (LoRA), pruning, and quantization are all effective methods for improving resource efficiency. However, combining them directly often results in suboptimal performance, especially with uniform quantization across all model layers. This is due to the complex, uneven interlayer relationships introduced by pruning, necessitating more refined quantization strategies. To address this, we propose AutoMixQ, an end-to-end optimization framework that selects optimal quantization configurations for each LLM layer. AutoMixQ leverages lightweight performance models to guide the selection process, significantly reducing time and computational resources compared to exhaustive search methods. By incorporating Pareto optimality, AutoMixQ balances memory usage and performance, approaching the upper bounds of model capability under strict resource constraints. Our experiments on widely used benchmarks show that AutoMixQ reduces memory consumption while achieving superior performance. For example, at a 30\\% pruning rate in LLaMA-7B, AutoMixQ achieved 66.21\\% on BoolQ compared to 62.45\\% for LoRA and 58.96\\% for LoftQ, while reducing memory consumption by 35.5\\% compared to LoRA and 27.5\\% compared to LoftQ.", "authors": ["Changhai Zhou", "Shiyang Zhang", "Yuhua Zhou", "Zekai Liu", "Shichao Weng"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-21", "url": "https://arxiv.org/abs/2411.13814", "pdf_url": "https://arxiv.org/pdf/2411.13814v1", "arxiv_id": "2411.13814", "doi": "10.48550/arXiv.2411.13814", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "5416b975d46237d4d056c15cb231a8063181b6565ab14588563889f7a68f67e1", "sources": ["arxiv", "semantic_scholar"], "title": "SeqProFT: Sequence-only Protein Property Prediction with LoRA Finetuning", "abstract": "Protein language models (PLMs) have demonstrated remarkable capabilities in learning relationships between protein sequences and functions. However, finetuning these large models requires substantial computational resources, often with suboptimal task-specific results. This study investigates how parameter-efficient finetuning via LoRA can enhance protein property prediction while significantly reducing computational demands. By applying LoRA to ESM-2 and ESM-C models of varying sizes and evaluating 10 diverse protein property prediction tasks, we demonstrate that smaller models with LoRA adaptation can match or exceed the performance of larger models without adaptation. Additionally, we integrate contact map information through a multi-head attention mechanism, improving model comprehension of structural features. Our systematic analysis reveals that LoRA finetuning enables faster convergence, better performance, and more efficient resource utilization, providing practical guidance for protein research applications in resource-constrained environments. The code is available at https://github.com/jiankliu/SeqProFT.", "authors": ["Shuo Zhang", "Jian K. Liu"], "categories": ["cs.LG", "q-bio.QM"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2024-11-18", "url": "https://arxiv.org/abs/2411.11530", "pdf_url": "https://arxiv.org/pdf/2411.11530v2", "arxiv_id": "2411.11530", "doi": "10.1109/TAI.2025.3636109", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/jiankliu/SeqProFT", "venue": "IEEE Transactions on Artificial Intelligence", "quality_score": 0.1747}
{"id": "4aea46bfda6462d9d5ebcce83fecb11c63bfa1208959fdd12662df3716808ee0", "sources": ["arxiv", "semantic_scholar"], "title": "SageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization", "abstract": "Although quantization for linear layers has been widely used, its application to accelerate the attention process remains limited. To further enhance the efficiency of attention computation compared to SageAttention while maintaining precision, we propose SageAttention2, which utilizes significantly faster 4-bit matrix multiplication (Matmul) alongside additional precision-enhancing techniques. First, we propose to quantize matrices $(Q, K)$ to INT4 in a hardware-friendly thread-level granularity and quantize matrices $(\\widetilde P, V)$ to FP8. Second, we propose a method to smooth $Q$, enhancing the accuracy of INT4 $QK^\\top$. Third, we propose a two-level accumulation strategy for $\\widetilde PV$ to enhance the accuracy of FP8 $\\widetilde PV$. The operations per second (OPS) of SageAttention2 surpass FlashAttention2 and xformers by about 3x and 4.5x on RTX4090, respectively. Moreover, SageAttention2 matches the speed of FlashAttention3(fp8) on the Hopper GPUs, while delivering much higher accuracy. Comprehensive experiments confirm that our approach incurs negligible end-to-end metrics loss across diverse models, including those for language, image, and video generation. The code is available at https://github.com/thu-ml/SageAttention.", "authors": ["Jintao Zhang", "Haofeng Huang", "Pengle Zhang", "Jia Wei", "Jun Zhu", "Jianfei Chen"], "categories": ["cs.LG", "cs.AI", "cs.CV", "cs.NE", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-17", "url": "https://arxiv.org/abs/2411.10958", "pdf_url": "https://arxiv.org/pdf/2411.10958v7", "arxiv_id": "2411.10958", "doi": null, "citation_count": 105, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/thu-ml/SageAttention", "venue": "International Conference on Machine Learning", "quality_score": 0.5063}
{"id": "8a00ff3c705bf5865da83c7e584dbccfdfce3455dbe096e7ac1441b281475403", "sources": ["arxiv", "semantic_scholar"], "title": "Awaker2.5-VL: Stably Scaling MLLMs with Parameter-Efficient Mixture of Experts", "abstract": "As the research of Multimodal Large Language Models (MLLMs) becomes popular, an advancing MLLM model is typically required to handle various textual and visual tasks (e.g., VQA, Detection, OCR, and ChartQA) simultaneously for real-world applications. However, due to the significant differences in representation and distribution among data from various tasks, simply mixing data of all tasks together leads to the well-known``multi-task conflict\" issue, resulting in performance degradation across various tasks. To address this issue, we propose Awaker2.5-VL, a Mixture of Experts~(MoE) architecture suitable for MLLM, which acquires the multi-task capabilities through multiple sparsely activated experts. To speed up the training and inference of Awaker2.5-VL, each expert in our model is devised as a low-rank adaptation (LoRA) structure. Extensive experiments on multiple latest benchmarks demonstrate the effectiveness of Awaker2.5-VL. The code and model weight are released in our Project Page: https://github.com/MetabrainAGI/Awaker.", "authors": ["Jinqiang Long", "Yanqi Dai", "Guoxing Yang", "Hongpeng Lin", "Nanyi Fei", "Yizhao Gao", "Zhiwu Lu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-16", "url": "https://arxiv.org/abs/2411.10669", "pdf_url": "https://arxiv.org/pdf/2411.10669v1", "arxiv_id": "2411.10669", "doi": "10.48550/arXiv.2411.10669", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/MetabrainAGI/Awaker", "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "f9e59f21f1e503cf14cad5504156762d5683d8ccf69c529e234cd62e844264f2", "sources": ["arxiv", "semantic_scholar"], "title": "On the Surprising Effectiveness of Attention Transfer for Vision Transformers", "abstract": "Conventional wisdom suggests that pre-training Vision Transformers (ViT) improves downstream performance by learning useful representations. Is this actually true? We investigate this question and find that the features and representations learned during pre-training are not essential. Surprisingly, using only the attention patterns from pre-training (i.e., guiding how information flows between tokens) is sufficient for models to learn high quality features from scratch and achieve comparable downstream performance. We show this by introducing a simple method called attention transfer, where only the attention patterns from a pre-trained teacher ViT are transferred to a student, either by copying or distilling the attention maps. Since attention transfer lets the student learn its own features, ensembling it with a fine-tuned teacher also further improves accuracy on ImageNet. We systematically study various aspects of our findings on the sufficiency of attention maps, including distribution shift settings where they underperform fine-tuning. We hope our exploration provides a better understanding of what pre-training accomplishes and leads to a useful alternative to the standard practice of fine-tuning", "authors": ["Alexander C. Li", "Yuandong Tian", "Beidi Chen", "Deepak Pathak", "Xinlei Chen"], "categories": ["cs.LG", "cs.AI", "cs.CV", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-14", "url": "https://arxiv.org/abs/2411.09702", "pdf_url": "https://arxiv.org/pdf/2411.09702v1", "arxiv_id": "2411.09702", "doi": "10.48550/arXiv.2411.09702", "citation_count": 15, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/alexlioralexli/attention-transfer", "venue": "Neural Information Processing Systems", "quality_score": 0.301}
{"id": "7b424e1504b4ff8fa684a506e35a1846c668a1c2ee320547c2bb16529b260060", "sources": ["arxiv", "semantic_scholar"], "title": "Imitation Learning from Observations: An Autoregressive Mixture of Experts Approach", "abstract": "This paper presents a novel approach to imitation learning from observations, where an autoregressive mixture of experts model is deployed to fit the underlying policy. The parameters of the model are learned via a two-stage framework. By leveraging the existing dynamics knowledge, the first stage of the framework estimates the control input sequences and hence reduces the problem complexity. At the second stage, the policy is learned by solving a regularized maximum-likelihood estimation problem using the estimated control input sequences. We further extend the learning procedure by incorporating a Lyapunov stability constraint to ensure asymptotic stability of the identified model, for accurate multi-step predictions. The effectiveness of the proposed framework is validated using two autonomous driving datasets collected from human demonstrations, demonstrating its practical applicability in modelling complex nonlinear dynamics.", "authors": ["Renzi Wang", "Flavia Sofia Acerbo", "Tong Duy Son", "Panagiotis Patrinos"], "categories": ["cs.LG", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-11-12", "url": "https://arxiv.org/abs/2411.08232", "pdf_url": "https://arxiv.org/pdf/2411.08232v1", "arxiv_id": "2411.08232", "doi": "10.48550/arXiv.2411.08232", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "ed3e15ba23397efc5af6c18b089a9469616a5290dbef458467458e4cc811a208", "sources": ["arxiv", "semantic_scholar"], "title": "Anchor Attention, Small Cache: Code Generation with Large Language Models", "abstract": "The development of large language models (LLMs) has revolutionized automated code generation. However, their high demand of computation resources has hindered a broader deployment and raised environmental concerns. A common strategy for diminishing computational demands is to cache Key-Value (KV) states from the attention mechanism which is adopted predominately by mainstream LLMs. It can mitigate the need of repeated attention computations, but brings significant memory overhead. Current practices in NLP often use sparse attention which may, unfortunately, lead to substantial inaccuracies, or hallucinations, in code generation tasks. In this paper, we analyze the attention weights distribution within code generation models via an empirical study, uncovering a sparsity pattern, i.e., the aggregation of information at specific anchor points. Based on this observation, we propose a novel approach, AnchorCoder, which features token-wise anchor attention designed to extract and compress the contextual information, and layer-wise anchor attention enabling cross-layer communication to mitigate the issue of excessive superposition caused by the compression. The extensive experiments across multiple benchmark datasets confirm the effectiveness of AnchorCoder, which can consistently achieve a significant (at least 70%) reduction in KV cache requirements, while preserving the majority of model's performance.", "authors": ["Xiangyu Zhang", "Yu Zhou", "Guang Yang", "Harald C. Gall", "Taolue Chen"], "categories": ["cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-11", "url": "https://arxiv.org/abs/2411.06680", "pdf_url": "https://arxiv.org/pdf/2411.06680v1", "arxiv_id": "2411.06680", "doi": "10.1109/TSE.2025.3570680", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Software Engineering", "quality_score": 0.1747}
{"id": "72d02195208e4104efb69aba565e85fbf8146f34a979d824e85bbe12e9ba8288", "sources": ["arxiv", "semantic_scholar"], "title": "WDMoE: Wireless Distributed Mixture of Experts for Large Language Models", "abstract": "Large Language Models (LLMs) have achieved significant success in various natural language processing tasks, but the role of wireless networks in supporting LLMs has not been thoroughly explored. In this paper, we propose a wireless distributed Mixture of Experts (WDMoE) architecture to enable collaborative deployment of LLMs across edge servers at the base station (BS) and mobile devices in wireless networks. Specifically, we decompose the MoE layer in LLMs by placing the gating network and the preceding neural network layer at BS, while distributing the expert networks among the devices. This deployment leverages the parallel inference capabilities of expert networks on mobile devices, effectively utilizing the limited computing and caching resources of these devices. Accordingly, we develop a performance metric for WDMoE-based LLMs, which accounts for both model capability and latency. To minimize the latency while maintaining accuracy, we jointly optimize expert selection and bandwidth allocation based on the performance metric. Moreover, we build a hardware testbed using NVIDIA Jetson kits to validate the effectiveness of WDMoE. Both theoretical simulations and practical hardware experiments demonstrate that the proposed method can significantly reduce the latency without compromising LLM performance.", "authors": ["Nan Xue", "Yaping Sun", "Zhiyong Chen", "Meixia Tao", "Xiaodong Xu", "Liang Qian", "Shuguang Cui", "Wenjun Zhang", "Ping Zhang"], "categories": ["cs.LG", "cs.AI", "cs.DC", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-11-11", "url": "https://arxiv.org/abs/2411.06681", "pdf_url": "https://arxiv.org/pdf/2411.06681v1", "arxiv_id": "2411.06681", "doi": "10.1109/TWC.2025.3585163", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Wireless Communications", "quality_score": 0.3076}
{"id": "c8c84e2c47c5257edcd64c3d102167ae41a76d6aa3d6b6f04ab8ec778132dd20", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Mixtures of Experts with EM: A Mirror Descent Perspective", "abstract": "Classical Mixtures of Experts (MoE) are Machine Learning models that involve partitioning the input space, with a separate \"expert\" model trained on each partition. Recently, MoE-based model architectures have become popular as a means to reduce training and inference costs. There, the partitioning function and the experts are both learnt jointly via gradient descent-type methods on the log-likelihood. In this paper we study theoretical guarantees of the Expectation Maximization (EM) algorithm for the training of MoE models. We first rigorously analyze EM for MoE where the conditional distribution of the target and latent variable conditioned on the feature variable belongs to an exponential family of distributions and show its equivalence to projected Mirror Descent with unit step size and a Kullback-Leibler Divergence regularizer. This perspective allows us to derive new convergence results and identify conditions for local linear convergence; In the special case of mixture of $2$ linear or logistic experts, we additionally provide guarantees for linear convergence based on the signal-to-noise ratio. Experiments on synthetic and (small-scale) real-world data supports that EM outperforms the gradient descent algorithm both in terms of convergence rate and the achieved accuracy.", "authors": ["Quentin Fruytier", "Aryan Mokhtari", "Sujay Sanghavi"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-11-09", "url": "https://arxiv.org/abs/2411.06056", "pdf_url": "https://arxiv.org/pdf/2411.06056v2", "arxiv_id": "2411.06056", "doi": "10.48550/arXiv.2411.06056", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "d0030776c7ac83b2e180de0f87bbc9ba6beeb162ea1b8e982c02278ecea84e4a", "sources": ["arxiv", "semantic_scholar"], "title": "RefreshKV: Updating Small KV Cache During Long-form Generation", "abstract": "Generating long sequences of tokens given a long-context input is a very compute-intensive inference scenario for large language models (LLMs). One prominent inference speed-up approach is to construct a smaller key-value (KV) cache, relieving LLMs from computing attention over a long sequence of tokens. While such methods work well to generate short sequences, their performance degrades rapidly for long-form generation. Most KV compression happens once, prematurely removing tokens that can be useful later in the generation. We propose a new inference method, RefreshKV, that flexibly alternates between full context attention and attention over a subset of input tokens during generation. After each full attention step, we update the smaller KV cache based on the attention pattern over the entire input. Applying our method to off-the-shelf LLMs achieves comparable speedup to eviction-based methods while improving performance for various long-form generation tasks. Lastly, we show that continued pretraining with our inference setting brings further gains in performance.", "authors": ["Fangyuan Xu", "Tanya Goyal", "Eunsol Choi"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-08", "url": "https://arxiv.org/abs/2411.05787", "pdf_url": "https://arxiv.org/pdf/2411.05787v2", "arxiv_id": "2411.05787", "doi": "10.18653/v1/2025.acl-long.1211", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2113}
{"id": "3c8e10cb933c78eb58d2e40ab8e2f8d7db5d5bc764e26f7f189abefbf68bd091", "sources": ["arxiv", "semantic_scholar"], "title": "MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba", "abstract": "An ecosystem of Transformer-based models has been established by building large models with extensive data. Parameter-efficient fine-tuning (PEFT) is a crucial technology for deploying these models to downstream tasks with minimal cost while achieving effective performance. Recently, Mamba, a State Space Model (SSM)-based model, has attracted attention as a potential alternative to Transformers. While many large-scale Mamba-based models have been proposed, efficiently adapting pre-trained Mamba-based models to downstream tasks remains unexplored. In this paper, we conduct an exploratory analysis of PEFT methods for Mamba. We investigate the effectiveness of existing PEFT methods for Transformers when applied to Mamba. We also modify these methods to better align with the Mamba architecture. Additionally, we propose new Mamba-specific PEFT methods that leverage the distinctive structure of Mamba. Our experiments indicate that PEFT performs more effectively for Mamba than Transformers. Lastly, we demonstrate how to effectively combine multiple PEFT methods and provide a framework that outperforms previous works. To ensure reproducibility, we will release the code after publication.", "authors": ["Masakazu Yoshimura", "Teruaki Hayashi", "Yota Maeda"], "categories": ["cs.CL", "cs.AI", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-06", "url": "https://arxiv.org/abs/2411.03855", "pdf_url": "https://arxiv.org/pdf/2411.03855v3", "arxiv_id": "2411.03855", "doi": "10.48550/arXiv.2411.03855", "citation_count": 14, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.294}
{"id": "5614bba67ca77f195e66ebde8923e9974bed286c3fcb54a0a67d94be710f1ba9", "sources": ["arxiv", "semantic_scholar"], "title": "DroidSpeak: KV Cache Sharing for Cross-LLM Communication and Multi-LLM Serving", "abstract": "Compound AI systems, such as agentic systems, are an emerging trend in large-scale enterprise settings, with multiple LLMs specialized for different users, tasks, and/or roles working together. In these scenarios, different models often process inputs that share the same context prefix. Although much work was done in the past to enable the reuse of prefix KV caches across inputs for a single model, how to enable one model to reuse the prefix KV caches of a different model remains an open question. We introduce DroidSpeak, the first distributed LLM inference system that enables KV cache reuse across distributed nodes running inference of different LLMs, so long as the LLMs have the same architecture. We present the first study that aims at understanding the impact of sharing KV caches across different LLMs, and if/when such sharing affects quality. Inspired by the findings, we present DroidSpeak, which selectively recomputes a few layers of the KV cache produced by another LLM and reuses the remaining layers, with negligible quality loss. Moreover, carefully pipelining the layer-wise re-computation and the loading of reused KV cache further improves the inference performance. Experiments on diverse datasets and model pairs demonstrate that DroidSpeak achieves up to 4x throughput improvement and about 3.1x faster prefill (time to first token), with negligible loss of quality in F1 scores, Rouge-L or code similarity score, compared to the baseline which does not allow any sharing across models.", "authors": ["Yuhan Liu", "Yuyang Huang", "Jiayi Yao", "Shaoting Feng", "Zhuohan Gu", "Kuntai Du", "Hanchen Li", "Yihua Cheng", "Junchen Jiang", "Shan Lu", "Madan Musuvathi", "Esha Choukse"], "categories": ["cs.MA", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-05", "url": "https://arxiv.org/abs/2411.02820", "pdf_url": "https://arxiv.org/pdf/2411.02820v4", "arxiv_id": "2411.02820", "doi": null, "citation_count": 27, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3618}
{"id": "3f1ad7d262d5d199f86c37ec1898a6dbd2a4a25d6430a330b48a0bd90245efa6", "sources": ["arxiv", "semantic_scholar"], "title": "TokenSelect: Efficient Long-Context Inference and Length Extrapolation for LLMs via Dynamic Token-Level KV Cache Selection", "abstract": "Rapid advances in Large Language Models (LLMs) have spurred demand for processing extended context sequences in contemporary applications. However, this progress faces two challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues limit LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using QK dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a few critical KV cache tokens in attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we design the Selection Cache based on observations of consecutive Query similarity and implemented the efficient Paged Dot Product Kernel, significantly reducing the selection overhead. A comprehensive evaluation of TokenSelect demonstrates up to $23.84\\times$ speedup in attention computation and up to $2.28\\times$ acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.", "authors": ["Wei Wu", "Zhuoshi Pan", "Chao Wang", "Liyi Chen", "Yunchu Bai", "Tianfu Wang", "Kun Fu", "Zheng Wang", "Hui Xiong"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-05", "url": "https://arxiv.org/abs/2411.02886", "pdf_url": "https://arxiv.org/pdf/2411.02886v4", "arxiv_id": "2411.02886", "doi": "10.48550/arXiv.2411.02886", "citation_count": 34, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.386}
{"id": "49b156f9d04c5f2e2d93bda604099ab6f5a4beee91a61c17a4b52602aee716bf", "sources": ["arxiv", "semantic_scholar"], "title": "FedMoE-DA: Federated Mixture of Experts via Domain Aware Fine-grained Aggregation", "abstract": "Federated learning (FL) is a collaborative machine learning approach that enables multiple clients to train models without sharing their private data. With the rise of deep learning, large-scale models have garnered significant attention due to their exceptional performance. However, a key challenge in FL is the limitation imposed by clients with constrained computational and communication resources, which hampers the deployment of these large models. The Mixture of Experts (MoE) architecture addresses this challenge with its sparse activation property, which reduces computational workload and communication demands during inference and updates. Additionally, MoE facilitates better personalization by allowing each expert to specialize in different subsets of the data distribution. To alleviate the communication burdens between the server and clients, we propose FedMoE-DA, a new FL model training framework that leverages the MoE architecture and incorporates a novel domain-aware, fine-grained aggregation strategy to enhance the robustness, personalizability, and communication efficiency simultaneously. Specifically, the correlation between both intra-client expert models and inter-client data heterogeneity is exploited. Moreover, we utilize peer-to-peer (P2P) communication between clients for selective expert model synchronization, thus significantly reducing the server-client transmissions. Experiments demonstrate that our FedMoE-DA achieves excellent performance while reducing the communication pressure on the server.", "authors": ["Ziwei Zhan", "Wenkuan Zhao", "Yuanqing Li", "Weijie Liu", "Xiaoxi Zhang", "Chee Wei Tan", "Chuan Wu", "Deke Guo", "Xu Chen"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-04", "url": "https://arxiv.org/abs/2411.02115", "pdf_url": "https://arxiv.org/pdf/2411.02115v2", "arxiv_id": "2411.02115", "doi": "10.1109/MSN63567.2024.00027", "citation_count": 11, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Mobile Ad-hoc and Sensor Networks", "quality_score": 0.2698}
{"id": "c7626fe32d91e3b1e792a871dccf879261c2702f09be7d50a4f1b376b7c35904", "sources": ["arxiv", "semantic_scholar"], "title": "MoCE: Adaptive Mixture of Contextualization Experts for Byte-based Neural Machine Translation", "abstract": "Byte-based machine translation systems have shown significant potential in massively multilingual settings. Unicode encoding, which maps each character to specific byte(s), eliminates the emergence of unknown words, even in new languages. This avoids out-of-vocabulary risk in multilingual translation and enables broad language scalability. However, byte-level tokenization results in sequences that are hard to interpret due to limited semantic information per byte. Local contextualization has proven effective in assigning initial semantics to tokens, improving sentence comprehension. Nevertheless, variations in encoding rules across languages necessitate an adaptive approach for effective contextualization. To this end, we propose Mixture of Contextualization Experts (MoCE), adaptively selecting and mixing attention heads, which are treated as contextualization experts. This enhances the flexibility of contextualization scales and allows models to search for better contextualization combinations. Experiment results show that our method outperforms existing methods without extensive manual adjustment of hyper-parameters and surpasses subword-based models with fewer parameters in Ted-59 dataset. Our code is available at https://github.com/ictnlp/MoCE.", "authors": ["Langlin Huang", "Mengyu Bu", "Yang Feng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-03", "url": "https://arxiv.org/abs/2411.01474", "pdf_url": "https://arxiv.org/pdf/2411.01474v2", "arxiv_id": "2411.01474", "doi": "10.48550/arXiv.2411.01474", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ictnlp/MoCE", "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.0753}
{"id": "8fa9ac56973e30e080326213af1d133ec6abab41db3d7171c2ac068e0d2bb282", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-I$^2$: Compressing Mixture of Experts Models through Inter-Expert Pruning and Intra-Expert Low-Rank Decomposition", "abstract": "The emergence of Mixture of Experts (MoE) LLMs has significantly advanced the development of language models. Compared to traditional LLMs, MoE LLMs outperform traditional LLMs by achieving higher performance with considerably fewer activated parameters. Despite this efficiency, their enormous parameter size still leads to high deployment costs. In this paper, we introduce a two-stage compression method tailored for MoE to reduce the model size and decrease the computational cost. First, in the inter-expert pruning stage, we analyze the importance of each layer and propose the Layer-wise Genetic Search and Block-wise KT-Reception Field with the non-uniform pruning ratio to prune the individual expert. Second, in the intra-expert decomposition stage, we apply the low-rank decomposition to further compress the parameters within the remaining experts. Extensive experiments on Qwen1.5-MoE-A2.7B, DeepSeek-V2-Lite, and Mixtral-8$\\times$7B demonstrate that our proposed methods can both reduce the model size and enhance inference efficiency while maintaining performance in various zero-shot tasks. The code will be available at \\url{https://github.com/xiaochengsky/MoEI-2.git}", "authors": ["Cheng Yang", "Yang Sui", "Jinqi Xiao", "Lingyi Huang", "Yu Gong", "Yuanlin Duan", "Wenqi Jia", "Miao Yin", "Yu Cheng", "Bo Yuan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-01", "url": "https://arxiv.org/abs/2411.01016", "pdf_url": "https://arxiv.org/pdf/2411.01016v1", "arxiv_id": "2411.01016", "doi": "10.18653/v1/2024.findings-emnlp.612", "citation_count": 51, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/xiaochengsky/MoEI-2.git}", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.429}
{"id": "2b95e97030e9d213781c6299ead13588b9901953d9cb1ca10a592698215ca240", "sources": ["arxiv", "semantic_scholar"], "title": "Magnitude Pruning of Large Pretrained Transformer Models with a Mixture Gaussian Prior", "abstract": "Large pretrained transformer models have revolutionized modern AI applications with their state-of-the-art performance in natural language processing (NLP). However, their substantial parameter count poses challenges for real-world deployment. To address this, researchers often reduce model size by pruning parameters based on their magnitude or sensitivity. Previous research has demonstrated the limitations of magnitude pruning, especially in the context of transfer learning for modern NLP tasks. In this paper, we introduce a new magnitude-based pruning algorithm called mixture Gaussian prior pruning (MGPP), which employs a mixture Gaussian prior for regularization. MGPP prunes non-expressive weights under the guidance of the mixture Gaussian prior, aiming to retain the model's expressive capability. Extensive evaluations across various NLP tasks, including natural language understanding, question answering, and natural language generation, demonstrate the superiority of MGPP over existing pruning methods, particularly in high sparsity settings. Additionally, we provide a theoretical justification for the consistency of the sparse transformer, shedding light on the effectiveness of the proposed pruning method.", "authors": ["Mingxuan Zhang", "Yan Sun", "Faming Liang"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics", "Medicine"], "published_date": "2024-11-01", "url": "https://arxiv.org/abs/2411.00969", "pdf_url": "https://arxiv.org/pdf/2411.00969v1", "arxiv_id": "2411.00969", "doi": "10.6339/24-jds1156", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Journal of Data Science", "quality_score": 0.1505}
{"id": "f3d556c4c5777dc57978458f1ee1e13e7fbeb2f8800e56e89a1638d091c29a22", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Tracker: Detecting Prompt Injection Attacks in LLMs", "abstract": "Large Language Models (LLMs) have revolutionized various domains but remain vulnerable to prompt injection attacks, where malicious inputs manipulate the model into ignoring original instructions and executing designated action. In this paper, we investigate the underlying mechanisms of these attacks by analyzing the attention patterns within LLMs. We introduce the concept of the distraction effect, where specific attention heads, termed important heads, shift focus from the original instruction to the injected instruction. Building on this discovery, we propose Attention Tracker, a training-free detection method that tracks attention patterns on instruction to detect prompt injection attacks without the need for additional LLM inference. Our method generalizes effectively across diverse models, datasets, and attack types, showing an AUROC improvement of up to 10.0% over existing methods, and performs well even on small LLMs. We demonstrate the robustness of our approach through extensive evaluations and provide insights into safeguarding LLM-integrated systems from prompt injection vulnerabilities.", "authors": ["Kuo-Han Hung", "Ching-Yun Ko", "Ambrish Rawat", "I-Hsin Chung", "Winston H. Hsu", "Pin-Yu Chen"], "categories": ["cs.CR", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-01", "url": "https://arxiv.org/abs/2411.00348", "pdf_url": "https://arxiv.org/pdf/2411.00348v2", "arxiv_id": "2411.00348", "doi": "10.48550/arXiv.2411.00348", "citation_count": 95, "influential_citation_count": 11, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.5396}
{"id": "c23664fe71ba7ba62c22576881343c33e9a5206a93f54ef68b050c4becbd3010", "sources": ["arxiv", "semantic_scholar"], "title": "LIBMoE: A Library for comprehensive benchmarking Mixture of Experts in Large Language Models", "abstract": "Mixture of experts (MoE) architectures have become a cornerstone for scaling up and are a key component in most large language models such as GPT-OSS, DeepSeek-V3, Llama-4, and Gemini-2.5. However, systematic research on MoE remains severely constrained by the prohibitive computational costs of training and evaluation, restricting large-scale studies accessible to most researchers. We introduce LibMoE, a unified framework for reproducible, efficient, and extensible MoE research that supports both pretraining and sparse-upcycling regimes. Beyond unified implementations, the framework provides transparent analytical tools for probing routing and expert dynamics. Leveraging this foundation, we conduct a comprehensive analysis along three dimensions: (i) routing dynamics, covering expert selection patterns, routing stability and optimality, and how routing entropy reveals task specialization and expert diversity; (ii) the effect of lightweight initialization on load balancing, demonstrating how subtle changes in router initialization shape early expert utilization; and (iii) training regime differences, revealing how sparse upcycling and full pretraining exhibit distinct routing patterns and stability profiles. By lowering the barrier to entry and standardizing evaluation, along with our comprehensive analysis, LibMoE broadens access to MoE research and establishes a reliable benchmark to guide future innovations. GitHub: \\href{https://github.com/Fsoft-AIC/LibMoE}{https://github.com/Fsoft-AIC/LibMoE}.", "authors": ["Nam V. Nguyen", "Thong T. Doan", "Luong Tran", "Van Nguyen", "Quang Pham"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-01", "url": "https://arxiv.org/abs/2411.00918", "pdf_url": "https://arxiv.org/pdf/2411.00918v4", "arxiv_id": "2411.00918", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Fsoft-AIC/LibMoE}{https://github.com/Fsoft-AIC/LibMoE}", "venue": null, "quality_score": 0.1945}
{"id": "aa1812876a6cecf7de2c6800a9cd5039dd3c771a368396b2c8b4d937dcd57649", "sources": ["arxiv", "semantic_scholar"], "title": "Empower Vision Applications with LoRA LMM", "abstract": "Large Multimodal Models (LMMs) have shown significant progress in various complex vision tasks with the solid linguistic and reasoning capacity inherited from large language models (LMMs). Low-rank adaptation (LoRA) offers a promising method to integrate external knowledge into LMMs, compensating for their limitations on domain-specific tasks. However, the existing LoRA model serving is excessively computationally expensive and causes extremely high latency. In this paper, we present an end-to-end solution that empowers diverse vision tasks and enriches vision applications with LoRA LMMs. Our system, VaLoRA, enables accurate and efficient vision tasks by 1) an accuracy-aware LoRA adapter generation approach that generates LoRA adapters rich in domain-specific knowledge to meet application-specific accuracy requirements, 2) an adaptive-tiling LoRA adapters batching operator that efficiently computes concurrent heterogeneous LoRA adapters, and 3) a flexible LoRA adapter orchestration mechanism that manages application requests and LoRA adapters to achieve the lowest average response latency. We prototype VaLoRA on five popular vision tasks on three LMMs. Experiment results reveal that VaLoRA improves 24-62% of the accuracy compared to the original LMMs and reduces 20-89% of the latency compared to the state-of-the-art LoRA model serving systems.", "authors": ["Liang Mi", "Weijun Wang", "Wenming Tu", "Qingfeng He", "Rui Kong", "Xinyu Fang", "Yazhu Dong", "Yikang Zhang", "Yunchun Li", "Meng Li", "Haipeng Dai", "Guihai Chen", "Yunxin Liu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-01", "url": "https://arxiv.org/abs/2411.00915", "pdf_url": "https://arxiv.org/pdf/2411.00915v5", "arxiv_id": "2411.00915", "doi": "10.1145/3689031.3717472", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "European Conference on Computer Systems", "quality_score": 0.2113}
{"id": "80f43077421a7ae41ad9cea1cf0cfef6a58b264402f8412b935dddd5ec187111", "sources": ["arxiv", "semantic_scholar"], "title": "In-Context LoRA for Diffusion Transformers", "abstract": "Recent research arXiv:2410.15027 has explored the use of diffusion transformers (DiTs) for task-agnostic image generation by simply concatenating attention tokens across images. However, despite substantial computational resources, the fidelity of the generated images remains suboptimal. In this study, we reevaluate and streamline this framework by hypothesizing that text-to-image DiTs inherently possess in-context generation capabilities, requiring only minimal tuning to activate them. Through diverse task experiments, we qualitatively demonstrate that existing text-to-image DiTs can effectively perform in-context generation without any tuning. Building on this insight, we propose a remarkably simple pipeline to leverage the in-context abilities of DiTs: (1) concatenate images instead of tokens, (2) perform joint captioning of multiple images, and (3) apply task-specific LoRA tuning using small datasets (e.g., 20~100 samples) instead of full-parameter tuning with large datasets. We name our models In-Context LoRA (IC-LoRA). This approach requires no modifications to the original DiT models, only changes to the training data. Remarkably, our pipeline generates high-fidelity image sets that better adhere to prompts. While task-specific in terms of tuning data, our framework remains task-agnostic in architecture and pipeline, offering a powerful tool for the community and providing valuable insights for further research on product-level task-agnostic generation systems. We release our code, data, and models at https://github.com/ali-vilab/In-Context-LoRA", "authors": ["Lianghua Huang", "Wei Wang", "Zhi-Fan Wu", "Yupeng Shi", "Huanzhang Dou", "Chen Liang", "Yutong Feng", "Yu Liu", "Jingren Zhou"], "categories": ["cs.CV", "cs.GR"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-31", "url": "https://arxiv.org/abs/2410.23775", "pdf_url": "https://arxiv.org/pdf/2410.23775v3", "arxiv_id": "2410.23775", "doi": "10.48550/arXiv.2410.23775", "citation_count": 150, "influential_citation_count": 21, "has_code": true, "code_url": "https://github.com/ali-vilab/In-Context-LoRA", "venue": "arXiv.org", "quality_score": 0.6712}
{"id": "4549ae1f88ae5f7c52d2898cba4b32c1791a2d5d17eb9b7aa70f7ac8679bafe3", "sources": ["arxiv", "semantic_scholar"], "title": "CopRA: A Progressive LoRA Training Strategy", "abstract": "Low-Rank Adaptation (LoRA) is a parameter-efficient technique for rapidly fine-tuning foundation models. In standard LoRA training dynamics, models tend to quickly converge to a local optimum near the initialization. However, this local optimum may not be ideal for out-of-distribution data or tasks such as merging and pruning. In this work, we propose a novel progressive training strategy for LoRA with random layer dropping. This strategy also optimizes the Shapley value of LoRA parameters in each layer, treating each layer as a player in a cooperative game. We refer to this method as Cooperative LoRA (CopRA). Our experimental results demonstrate that parameters trained with CopRA exhibit linear mode connectivity, which enables efficient model merging. This also paves the way for federated learning and multi-task learning via LoRA merging. Additionally, by optimizing the Shapley value, CopRA shows superior performance in pruning tasks.", "authors": ["Zhan Zhuang", "Xiequn Wang", "Yulong Zhang", "Wei Li", "Yu Zhang", "Ying Wei"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-30", "url": "https://arxiv.org/abs/2410.22911", "pdf_url": "https://arxiv.org/pdf/2410.22911v1", "arxiv_id": "2410.22911", "doi": "10.48550/arXiv.2410.22911", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "1bc1cb3163afd3b41b92d8bd6803c86ba2955762531f4d2f6249da9e1ad0fb6e", "sources": ["arxiv", "semantic_scholar"], "title": "Stealing User Prompts from Mixture of Experts", "abstract": "Mixture-of-Experts (MoE) models improve the efficiency and scalability of dense language models by routing each token to a small number of experts in each layer. In this paper, we show how an adversary that can arrange for their queries to appear in the same batch of examples as a victim's queries can exploit Expert-Choice-Routing to fully disclose a victim's prompt. We successfully demonstrate the effectiveness of this attack on a two-layer Mixtral model, exploiting the tie-handling behavior of the torch.topk CUDA implementation. Our results show that we can extract the entire prompt using $O({VM}^2)$ queries (with vocabulary size $V$ and prompt length $M$) or 100 queries on average per token in the setting we consider. This is the first attack to exploit architectural flaws for the purpose of extracting user prompts, introducing a new class of LLM vulnerabilities.", "authors": ["Itay Yona", "Ilia Shumailov", "Jamie Hayes", "Nicholas Carlini"], "categories": ["cs.CR", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-30", "url": "https://arxiv.org/abs/2410.22884", "pdf_url": "https://arxiv.org/pdf/2410.22884v1", "arxiv_id": "2410.22884", "doi": "10.48550/arXiv.2410.22884", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "19f7179d5ffe2b7122f763fa1a31f9b8b8ee58a82747149993609985086a16ca", "sources": ["arxiv", "semantic_scholar"], "title": "BUZZ: Beehive-structured Sparse KV Cache with Segmented Heavy Hitters for Efficient LLM Inference", "abstract": "Large language models (LLMs) are essential in natural language processing but often struggle with inference speed and computational efficiency, limiting real-time deployment. The key-value (KV) cache mechanism reduces computational overhead in transformer models, but challenges in maintaining contextual understanding remain. In this paper, we propose BUZZ, a novel KV caching algorithm that leverages structured contextual information to minimize cache memory usage while enhancing inference speed. BUZZ employs a beehive-structured sparse cache, incorporating a sliding window to capture recent information and dynamically segmenting historical tokens into chunks to prioritize important tokens in local neighborhoods. We evaluate BUZZ on four real-world datasets: CNN/Daily Mail, XSUM, Wikitext, and 10-QA. Our results demonstrate that BUZZ (1) reduces cache memory usage by $\\textbf{2.5}\\times$ in LLM inference while maintaining over 99% accuracy in long-text summarization, and (2) surpasses state-of-the-art performance in multi-document question answering by $\\textbf{7.69%}$ under the same memory limit, where full cache methods encounter out-of-memory issues. Additionally, BUZZ achieves significant inference speedup with a $\\log{n}$ time complexity. The code is available at https://github.com/JunqiZhao888/buzz-llm.", "authors": ["Junqi Zhao", "Zhijin Fang", "Shu Li", "Shaohui Yang", "Shichao He"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-30", "url": "https://arxiv.org/abs/2410.23079", "pdf_url": "https://arxiv.org/pdf/2410.23079v1", "arxiv_id": "2410.23079", "doi": "10.1145/3810942", "citation_count": 6, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/JunqiZhao888/buzz-llm", "venue": "ACM Transactions on Intelligent Systems and Technology", "quality_score": 0.2113}
{"id": "47f7a8f8cd1f99fea319264072f01ff4031b23d17f51462f7e4d9ba33cce2a27", "sources": ["arxiv", "semantic_scholar"], "title": "MoLE: Enhancing Human-centric Text-to-image Diffusion via Mixture of Low-rank Experts", "abstract": "Text-to-image diffusion has attracted vast attention due to its impressive image-generation capabilities. However, when it comes to human-centric text-to-image generation, particularly in the context of faces and hands, the results often fall short of naturalness due to insufficient training priors. We alleviate the issue in this work from two perspectives. 1) From the data aspect, we carefully collect a human-centric dataset comprising over one million high-quality human-in-the-scene images and two specific sets of close-up images of faces and hands. These datasets collectively provide a rich prior knowledge base to enhance the human-centric image generation capabilities of the diffusion model. 2) On the methodological front, we propose a simple yet effective method called Mixture of Low-rank Experts (MoLE) by considering low-rank modules trained on close-up hand and face images respectively as experts. This concept draws inspiration from our observation of low-rank refinement, where a low-rank module trained by a customized close-up dataset has the potential to enhance the corresponding image part when applied at an appropriate scale. To validate the superiority of MoLE in the context of human-centric image generation compared to state-of-the-art, we construct two benchmarks and perform evaluations with diverse metrics and human studies. Datasets, model, and code are released at https://sites.google.com/view/mole4diffuser/.", "authors": ["Jie Zhu", "Yixiong Chen", "Mingyu Ding", "Ping Luo", "Leye Wang", "Jingdong Wang"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-30", "url": "https://arxiv.org/abs/2410.23332", "pdf_url": "https://arxiv.org/pdf/2410.23332v1", "arxiv_id": "2410.23332", "doi": "10.48550/arXiv.2410.23332", "citation_count": 16, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3076}
{"id": "4b551aa46c17b888670ffdd0b177ee56180d8063fcf6288745ce8d55db18085f", "sources": ["arxiv", "semantic_scholar"], "title": "MALoRA: Mixture of Asymmetric Low-Rank Adaptation for Enhanced Multi-Task Learning", "abstract": "Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA have significantly improved the adaptation of LLMs to downstream tasks in a resource-efficient manner. However, in multi-task scenarios, challenges such as training imbalance and the seesaw effect frequently emerge. Mixture-of-LoRA (MoLoRA), which combines LoRA with sparse Mixture-of-Experts, mitigates some of these issues by promoting task-specific learning across experts. Despite this, MoLoRA remains inefficient in terms of training speed, parameter utilization, and overall multi-task performance. In this paper, we propose Mixture of Asymmetric Low-Rank Adaptaion (MALoRA), a flexible fine-tuning framework that leverages asymmetric optimization across LoRA experts. MALoRA reduces the number of trainable parameters by 30% to 48%, increases training speed by 1.2x, and matches the computational efficiency of single-task LoRA models. Additionally, MALoRA addresses overfitting issues commonly seen in high-rank configurations, enhancing performance stability. Extensive experiments across diverse multi-task learning scenarios demonstrate that MALoRA consistently outperforms all baseline methods in both inter-domain and intra-domain tasks.", "authors": ["Xujia Wang", "Haiyan Zhao", "Shuo Wang", "Hanqing Wang", "Zhiyuan Liu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-30", "url": "https://arxiv.org/abs/2410.22782", "pdf_url": "https://arxiv.org/pdf/2410.22782v1", "arxiv_id": "2410.22782", "doi": "10.48550/arXiv.2410.22782", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.2258}
{"id": "eaa3ba15c226ca6645524b2e9ea226a885aee149fca02260ffbee38ce19de604", "sources": ["arxiv", "semantic_scholar"], "title": "VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration", "abstract": "Vision-Language Models (VLMs) have demonstrated impressive performance across a versatile set of tasks. A key challenge in accelerating VLMs is storing and accessing the large Key-Value (KV) cache that encodes long visual contexts, such as images or videos. While existing KV cache compression methods are effective for Large Language Models (LLMs), directly migrating them to VLMs yields suboptimal accuracy and speedup. To bridge the gap, we propose VL-Cache, a novel KV cache compression recipe tailored for accelerating VLM inference. In this paper, we first investigate the unique sparsity pattern of VLM attention by distinguishing visual and text tokens in prefill and decoding phases. Based on these observations, we introduce a layer-adaptive sparsity-aware cache budget allocation method that effectively distributes the limited cache budget across different layers, further reducing KV cache size without compromising accuracy. Additionally, we develop a modality-aware token scoring policy to better evaluate the token importance. Empirical results on multiple benchmark datasets demonstrate that retaining only 10% of KV cache achieves accuracy comparable to that with full cache. In a speed benchmark, our method accelerates end-to-end latency of generating 100 tokens by up to 2.33x and speeds up decoding by up to 7.08x, while reducing the memory footprint of KV cache in GPU by 90%.", "authors": ["Dezhan Tu", "Danylo Vashchilenko", "Yuzhe Lu", "Panpan Xu"], "categories": ["cs.CV", "cs.AI", "cs.CL", "cs.DC", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-29", "url": "https://arxiv.org/abs/2410.23317", "pdf_url": "https://arxiv.org/pdf/2410.23317v1", "arxiv_id": "2410.23317", "doi": "10.48550/arXiv.2410.23317", "citation_count": 40, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4225}
{"id": "5cf1900c6a2aae3784bb8b8f32679d9e569f0aa28dc48c5af25bd62b2a402c63", "sources": ["arxiv", "semantic_scholar"], "title": "Neural Experts: Mixture of Experts for Implicit Neural Representations", "abstract": "Implicit neural representations (INRs) have proven effective in various tasks including image, shape, audio, and video reconstruction. These INRs typically learn the implicit field from sampled input points. This is often done using a single network for the entire domain, imposing many global constraints on a single function. In this paper, we propose a mixture of experts (MoE) implicit neural representation approach that enables learning local piece-wise continuous functions that simultaneously learns to subdivide the domain and fit locally. We show that incorporating a mixture of experts architecture into existing INR formulations provides a boost in speed, accuracy, and memory requirements. Additionally, we introduce novel conditioning and pretraining methods for the gating network that improves convergence to the desired solution. We evaluate the effectiveness of our approach on multiple reconstruction tasks, including surface reconstruction, image reconstruction, and audio signal reconstruction and show improved performance compared to non-MoE methods.", "authors": ["Yizhak Ben-Shabat", "Chamin Hewa Koneputugodage", "Sameera Ramasinghe", "Stephen Gould"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-29", "url": "https://arxiv.org/abs/2410.21643", "pdf_url": "https://arxiv.org/pdf/2410.21643v1", "arxiv_id": "2410.21643", "doi": "10.48550/arXiv.2410.21643", "citation_count": 13, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.2865}
{"id": "381a3f983bd798062d2c403f4b6daf229655f6b51250d99794ed8621d437fa75", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient and Effective Weight-Ensembling Mixture of Experts for Multi-Task Model Merging", "abstract": "Multi-task learning (MTL) leverages a shared model to accomplish multiple tasks and facilitate knowledge transfer. Recent research on task arithmetic-based MTL demonstrates that merging the parameters of independently fine-tuned models can effectively achieve MTL. However, existing merging methods primarily seek a static optimal solution within the original model parameter space, which often results in performance degradation due to the inherent diversity among tasks and potential interferences. To address this challenge, in this paper, we propose a Weight-Ensembling Mixture of Experts (WEMoE) method for multi-task model merging. Specifically, we first identify critical (or sensitive) modules by analyzing parameter variations in core modules of Transformer-based models before and after finetuning. Then, our WEMoE statically merges non-critical modules while transforming critical modules into a mixture-of-experts (MoE) structure. During inference, expert modules in the MoE are dynamically merged based on input samples, enabling a more flexible and adaptive merging approach. Building on WEMoE, we further introduce an efficient-and-effective WEMoE (E-WEMoE) method, whose core mechanism involves eliminating non-essential elements in the critical modules of WEMoE and implementing shared routing across multiple MoE modules, thereby significantly reducing both the trainable parameters, the overall parameter count, and computational overhead of the merged model by WEMoE. Experimental results across various architectures and tasks demonstrate that both WEMoE and E-WEMoE outperform state-of-the-art (SOTA) model merging methods in terms of MTL performance, generalization, and robustness.", "authors": ["Li Shen", "Anke Tang", "Enneng Yang", "Guibing Guo", "Yong Luo", "Lefei Zhang", "Xiaochun Cao", "Bo Du", "Dacheng Tao"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2024-10-29", "url": "https://arxiv.org/abs/2410.21804", "pdf_url": "https://arxiv.org/pdf/2410.21804v1", "arxiv_id": "2410.21804", "doi": "10.1109/TPAMI.2025.3629605", "citation_count": 23, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.3451}
{"id": "c15ae60809da1e01306d06e63db563d2cea2e21eb0b6b0dd7da71a79f53a3bf7", "sources": ["arxiv", "semantic_scholar"], "title": "ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference", "abstract": "With the widespread deployment of long-context large language models (LLMs), there has been a growing demand for efficient support of high-throughput inference. However, as the key-value (KV) cache expands with the sequence length, the increasing memory footprint and the need to access it for each token generation both result in low throughput when serving long-context LLMs. While various dynamic sparse attention methods have been proposed to speed up inference while maintaining generation quality, they either fail to sufficiently reduce GPU memory consumption or introduce significant decoding latency by offloading the KV cache to the CPU. We present ShadowKV, a high-throughput long-context LLM inference system that stores the low-rank key cache and offloads the value cache to reduce the memory footprint for larger batch sizes and longer sequences. To minimize decoding latency, ShadowKV employs an accurate KV selection strategy that reconstructs minimal sparse KV pairs on-the-fly. By evaluating ShadowKV on a broad range of benchmarks, including RULER, LongBench, and Needle In A Haystack, and models like Llama-3.1-8B, Llama-3-8B-1M, GLM-4-9B-1M, Yi-9B-200K, Phi-3-Mini-128K, and Qwen2-7B-128K, we demonstrate that it can support up to 6$\\times$ larger batch sizes and boost throughput by up to 3.04$\\times$ on an A100 GPU without sacrificing accuracy, even surpassing the performance achievable with infinite batch size under the assumption of infinite GPU memory. The code is available at https://github.com/bytedance/ShadowKV.", "authors": ["Hanshi Sun", "Li-Wen Chang", "Wenlei Bao", "Size Zheng", "Ningxin Zheng", "Xin Liu", "Harry Dong", "Yuejie Chi", "Beidi Chen"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-28", "url": "https://arxiv.org/abs/2410.21465", "pdf_url": "https://arxiv.org/pdf/2410.21465v3", "arxiv_id": "2410.21465", "doi": "10.48550/arXiv.2410.21465", "citation_count": 100, "influential_citation_count": 12, "has_code": true, "code_url": "https://github.com/bytedance/ShadowKV", "venue": "International Conference on Machine Learning", "quality_score": 0.557}
{"id": "8fb0461312feb85e7db6a36163063628e4495dd5b4dfaca9d9a1ed0f8cc64316", "sources": ["arxiv", "semantic_scholar"], "title": "KD-LoRA: A Hybrid Approach to Efficient Fine-Tuning with LoRA and Knowledge Distillation", "abstract": "Large language models (LLMs) have demonstrated remarkable performance across various downstream tasks. However, the high computational and memory requirements of LLMs are a major bottleneck. To address this, parameter-efficient fine-tuning (PEFT) methods such as low-rank adaptation (LoRA) have been proposed to reduce computational costs while ensuring minimal loss in performance. Additionally, knowledge distillation (KD) has been a popular choice for obtaining compact student models from teacher models. In this work, we present KD-LoRA, a novel fine-tuning method that combines LoRA with KD. Our results demonstrate that KD-LoRA achieves performance comparable to full fine-tuning (FFT) and LoRA while significantly reducing resource requirements. Specifically, KD-LoRA retains 98% of LoRA's performance on the GLUE benchmark, while being 40% more compact. Additionally, KD-LoRA reduces GPU memory usage by 30% compared to LoRA, while decreasing inference time by 30% compared to both FFT and LoRA. We evaluate KD-LoRA across three encoder-only models: BERT, RoBERTa, and DeBERTaV3. Code is available at https://github.com/rambodazimi/KD-LoRA.", "authors": ["Rambod Azimi", "Rishav Rishav", "Marek Teichmann", "Samira Ebrahimi Kahou"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-28", "url": "https://arxiv.org/abs/2410.20777", "pdf_url": "https://arxiv.org/pdf/2410.20777v1", "arxiv_id": "2410.20777", "doi": "10.48550/arXiv.2410.20777", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/rambodazimi/KD-LoRA", "venue": null, "quality_score": 0.2258}
{"id": "759537927a20b1e1e037a330cffea0716dbc2ce020717878e165c97847d13108", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA Done RITE: Robust Invariant Transformation Equilibration for LoRA Optimization", "abstract": "Low-rank adaption (LoRA) is a widely used parameter-efficient finetuning method for LLM that reduces memory requirements. However, current LoRA optimizers lack transformation invariance, meaning the actual updates to the weights depends on how the two LoRA factors are scaled or rotated. This deficiency leads to inefficient learning and sub-optimal solutions in practice. This paper introduces LoRA-RITE, a novel adaptive matrix preconditioning method for LoRA optimization, which can achieve transformation invariance and remain computationally efficient. We provide theoretical analysis to demonstrate the benefit of our method and conduct experiments on various LLM tasks with different models including Gemma 2B, 7B, and mT5-XXL. The results demonstrate consistent improvements against existing optimizers. For example, replacing Adam with LoRA-RITE during LoRA fine-tuning of Gemma-2B yielded 4.6\\% accuracy gain on Super-Natural Instructions and 3.5\\% accuracy gain across other four LLM benchmarks (HellaSwag, ArcChallenge, GSM8K, OpenBookQA).", "authors": ["Jui-Nan Yen", "Si Si", "Zhao Meng", "Felix Yu", "Sai Surya Duvvuri", "Inderjit S. Dhillon", "Cho-Jui Hsieh", "Sanjiv Kumar"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-27", "url": "https://arxiv.org/abs/2410.20625", "pdf_url": "https://arxiv.org/pdf/2410.20625v2", "arxiv_id": "2410.20625", "doi": "10.48550/arXiv.2410.20625", "citation_count": 24, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/gkevinyen5418/LoRA-RITE", "venue": "International Conference on Learning Representations", "quality_score": 0.3891}
{"id": "6bd06da62bf0abb3929873add54cc05cb4bdfd993f7a8c16e523698edbbf6549", "sources": ["arxiv", "semantic_scholar"], "title": "Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning", "abstract": "Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark. Codes are available at https://github.com/FYYFU/HeadKV", "authors": ["Yu Fu", "Zefan Cai", "Abedelkadir Asi", "Wayne Xiong", "Yue Dong", "Wen Xiao"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-25", "url": "https://arxiv.org/abs/2410.19258", "pdf_url": "https://arxiv.org/pdf/2410.19258v4", "arxiv_id": "2410.19258", "doi": "10.48550/arXiv.2410.19258", "citation_count": 82, "influential_citation_count": 12, "has_code": true, "code_url": "https://github.com/FYYFU/HeadKV", "venue": "International Conference on Learning Representations", "quality_score": 0.557}
{"id": "38c22f524195fd729da2c606fceff4cd5412d6eb5967ce6dfe4f9866a6e0b9db", "sources": ["arxiv", "semantic_scholar"], "title": "KVSharer: Efficient Inference via Layer-Wise Dissimilar KV Cache Sharing", "abstract": "The development of large language models (LLMs) has significantly expanded model sizes, resulting in substantial GPU memory requirements during inference. The key and value storage of the attention map in the KV (key-value) cache accounts for more than 80\\% of this memory consumption. Nowadays, most existing KV cache compression methods focus on intra-layer compression within a single Transformer layer but few works consider layer-wise compression. In this paper, we propose a plug-and-play method called \\textit{KVSharer}, which shares the KV cache between layers to achieve layer-wise compression. Rather than intuitively sharing based on higher similarity, we discover a counterintuitive phenomenon: sharing dissimilar KV caches better preserves the model performance. Experiments show that \\textit{KVSharer} can reduce KV cache computation by 30\\%, thereby lowering memory consumption without significantly impacting model performance and it can also achieve at least 1.3 times generation acceleration. Additionally, we verify that \\textit{KVSharer} is compatible with existing intra-layer KV cache compression methods, and combining both can further save memory.", "authors": ["Yifei Yang", "Zouying Cao", "Qiguang Chen", "Libo Qin", "Dongjie Yang", "Hai Zhao", "Zhi Chen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-24", "url": "https://arxiv.org/abs/2410.18517", "pdf_url": "https://arxiv.org/pdf/2410.18517v1", "arxiv_id": "2410.18517", "doi": "10.48550/arXiv.2410.18517", "citation_count": 33, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3829}
{"id": "b6ce544498e483b779476a719349f9993940ee1e9453d2a4109a0848b879e812", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Parrots: Experts improve memorization more than reasoning", "abstract": "The Mixture-of-Experts (MoE) architecture enables a significant increase in the total number of model parameters with minimal computational overhead. However, it is not clear what performance tradeoffs, if any, exist between MoEs and standard dense transformers. In this paper, we show that as we increase the number of experts (while fixing the number of active parameters), the memorization performance consistently increases while the reasoning capabilities saturate. We begin by analyzing the theoretical limitations of MoEs at reasoning. We prove that there exist graph problems that cannot be solved by any number of experts of a certain width; however, the same task can be easily solved by a dense model with a slightly larger width. On the other hand, we find that on memory-intensive tasks, MoEs can effectively leverage a small number of active parameters with a large number of experts to memorize the data. We empirically validate these findings on synthetic graph problems and memory-intensive closed book retrieval tasks. Lastly, we pre-train a series of MoEs and dense transformers and evaluate them on commonly used benchmarks in math and natural language. We find that increasing the number of experts helps solve knowledge-intensive tasks, but fails to yield the same benefits for reasoning tasks.", "authors": ["Samy Jelassi", "Clara Mohri", "David Brandfonbrener", "Alex Gu", "Nikhil Vyas", "Nikhil Anand", "David Alvarez-Melis", "Yuanzhi Li", "Sham M. Kakade", "Eran Malach"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-24", "url": "https://arxiv.org/abs/2410.19034", "pdf_url": "https://arxiv.org/pdf/2410.19034v2", "arxiv_id": "2410.19034", "doi": "10.48550/arXiv.2410.19034", "citation_count": 21, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3356}
{"id": "275e4cb00ced98582c029bcb7081315552f1c507270171c5c0521b2764451f16", "sources": ["arxiv", "semantic_scholar"], "title": "ExpertFlow: Efficient Mixture-of-Experts Inference via Predictive Expert Caching and Token Scheduling", "abstract": "Sparse Mixture-of-Experts (MoE) models can outperform dense large language models at similar computation by activating only a small set of experts per token. However, stacking many expert modules introduces substantial parameter memory, which makes MoE models difficult to deploy in memory-constrained environments such as single-GPU devices. Offloading alleviates this issue by storing inactive experts in CPU memory and loading them on demand, but existing methods remain limited: static caches disregard input-dependent routing, and methods that train separate models to predict expert usage ahead of time are often inaccurate or require significant training cost. We propose ExpertFlow, a lightweight MoE inference system that addresses this routing dependency through three coordinated components: 1) a transformer-based routing path predictor that estimates expert usage across all MoE layers in a single forward pass, 2) a token scheduler that groups tokens with similar predicted routes to improve expert utilization, and 3) a predictive expert cache that loads only the required experts while correcting mispredictions at runtime. Together, these components enable efficient expert loading and execution, reducing GPU memory usage by up to 93.72% and improving inference throughput by up to 10x over strong offloading baselines on a single GPU.", "authors": ["Xin He", "Shunkang Zhang", "Kaijie Tang", "Shaohuai Shi", "Yuxin Wang", "Zihao Zeng", "Zhenheng Tang", "Xiaowen Chu", "Haiyan Yin", "Ivor W. Tsang", "Yew Soon Ong"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-23", "url": "https://arxiv.org/abs/2410.17954", "pdf_url": "https://arxiv.org/pdf/2410.17954v2", "arxiv_id": "2410.17954", "doi": "10.1145/3770743.3804292", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "5f15f148b705b2700376f9bc46dd3672a9c73e96d9c65fbe390d1c3e6dfda139", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-C: Parameter-Efficient Fine-Tuning of Robust CNN for IoT Devices", "abstract": "Efficient fine-tuning of pre-trained convolutional neural network (CNN) models using local data is essential for providing high-quality services to users using ubiquitous and resource-limited Internet of Things (IoT) devices. Low-Rank Adaptation (LoRA) fine-tuning has attracted widespread attention from industry and academia because it is simple, efficient, and does not incur any additional reasoning burden. However, most of the existing advanced methods use LoRA to fine-tune Transformer, and there are few studies on using LoRA to fine-tune CNN. The CNN model is widely deployed on IoT devices for application due to its advantages in comprehensive resource occupancy and performance. Moreover, IoT devices are widely deployed outdoors and usually process data affected by the environment (such as fog, snow, rain, etc.). The goal of this paper is to use LoRA technology to efficiently improve the robustness of the CNN model. To this end, this paper first proposes a strong, robust CNN fine-tuning method for IoT devices, LoRA-C, which performs low-rank decomposition in convolutional layers rather than kernel units to reduce the number of fine-tuning parameters. Then, this paper analyzes two different rank settings in detail and observes that the best performance is usually achieved when $α/{r}$ is a constant in either standard data or corrupted data. This discovery provides experience for the widespread application of LoRA-C. Finally, this paper conducts many experiments based on pre-trained models. Experimental results on CIFAR-10, CIFAR-100, CIFAR-10-C, and Icons50 datasets show that the proposed LoRA-Cs outperforms standard ResNets. Specifically, on the CIFAR-10-C dataset, the accuracy of LoRA-C-ResNet-101 achieves 83.44% accuracy, surpassing the standard ResNet-101 result by +9.5%.", "authors": ["Chuntao Ding", "Xu Cao", "Jianhang Xie", "Linlin Fan", "Shangguang Wang", "Zhichao Lu"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-22", "url": "https://arxiv.org/abs/2410.16954", "pdf_url": "https://arxiv.org/pdf/2410.16954v2", "arxiv_id": "2410.16954", "doi": "10.48550/arXiv.2410.16954", "citation_count": 15, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/alexyyds2024/lora-C", "venue": "arXiv.org", "quality_score": 0.301}
{"id": "3184c868169a45bab0ccd75ad95b6e1a4b80c28a23cb0a548706c242deee7424", "sources": ["arxiv", "semantic_scholar"], "title": "Self-calibration for Language Model Quantization and Pruning", "abstract": "Quantization and pruning are fundamental approaches for model compression, enabling efficient inference for language models. In a post-training setting, state-of-the-art quantization and pruning methods require calibration data, a small set of unlabeled examples. Conventionally, this is randomly sampled web text, aiming to reflect the model training data. However, this poses two key problems: (1) unrepresentative calibration examples can harm model performance, and (2) organizations increasingly avoid releasing model training data. In this paper, we propose self-calibration as a solution. Our approach requires no external data, instead leveraging the model itself to generate synthetic calibration data, with a view to better approximating the pre-training data distribution. We extensively compare the performance of self-calibration with several baselines, across a variety of models, compression methods, and tasks. Our approach proves consistently competitive in maximizing downstream task performance, frequently outperforming even using real data.", "authors": ["Miles Williams", "George Chrysostomou", "Nikolaos Aletras"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-22", "url": "https://arxiv.org/abs/2410.17170", "pdf_url": "https://arxiv.org/pdf/2410.17170v2", "arxiv_id": "2410.17170", "doi": "10.48550/arXiv.2410.17170", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.2386}
{"id": "e9bf9078c917b4084d5a8c691dc6f16999513f094e7f978295323134170f9f9e", "sources": ["arxiv", "semantic_scholar"], "title": "Residual vector quantization for KV cache compression in large language model", "abstract": "KV cache compression methods have mainly relied on scalar quantization techniques to reduce the memory requirements during decoding. In this work, we apply residual vector quantization, which has been widely used for high fidelity audio compression, to compress KV cache in large language models (LLM). We adapt the standard recipe with minimal changes to compress the output of any key or value projection matrix in a pretrained LLM: we scale the vector by its standard deviation, divide channels into groups and then quantize each group with the same residual vector quantizer. We learn the codebook using exponential moving average and there are no other learnable parameters including the input and output projections normally used in a vector quantization set up. We find that a residual depth of 8 recovers most of the performance of the unquantized model. We also find that grouping non-contiguous channels together works better than grouping contiguous channels for compressing key matrix and the method further benefits from a light weight finetuning of LLM together with the quantization. Overall, the proposed technique is competitive with existing quantization methods while being much simpler and results in 5.5x compression compared to half precision.", "authors": ["Ankur Kumar"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-21", "url": "https://arxiv.org/abs/2410.15704", "pdf_url": "https://arxiv.org/pdf/2410.15704v1", "arxiv_id": "2410.15704", "doi": "10.48550/arXiv.2410.15704", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "cab82ea1bcfee92c149f3c7ef6308d30bdcb153f45503c95c1dbc094ec5b494a", "sources": ["arxiv", "semantic_scholar"], "title": "CartesianMoE: Boosting Knowledge Sharing among Experts via Cartesian Product Routing in Mixture-of-Experts", "abstract": "Large language models (LLM) have been attracting much attention from the community recently, due to their remarkable performance in all kinds of downstream tasks. According to the well-known scaling law, scaling up a dense LLM enhances its capabilities, but also significantly increases the computational complexity. Mixture-of-Experts (MoE) models address that by allowing the model size to grow without substantially raising training or inference costs. Yet MoE models face challenges regarding knowledge sharing among experts, making their performance somehow sensitive to routing accuracy. To tackle that, previous works introduced shared experts and combined their outputs with those of the top $K$ routed experts in an ``addition'' manner. In this paper, inspired by collective matrix factorization to learn shared knowledge among data, we propose CartesianMoE, which implements more effective knowledge sharing among experts in more like a ``multiplication'' manner. Extensive experimental results indicate that CartesianMoE outperforms previous MoE models for building LLMs, in terms of both perplexity and downstream task performance. And we also find that CartesianMoE achieves better expert routing robustness.", "authors": ["Zhenpeng Su", "Xing Wu", "Zijia Lin", "Yizhe Xiong", "Minxuan Lv", "Guangyuan Ma", "Hui Chen", "Songlin Hu", "Guiguang Ding"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-21", "url": "https://arxiv.org/abs/2410.16077", "pdf_url": "https://arxiv.org/pdf/2410.16077v3", "arxiv_id": "2410.16077", "doi": "10.48550/arXiv.2410.16077", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.1945}
{"id": "dcf280977f3479fb14ce722bdbeca66f1a7f82ca3c16074733998b71198460a0", "sources": ["arxiv", "semantic_scholar"], "title": "Lossless KV Cache Compression to 2%", "abstract": "Large language models have revolutionized data processing in numerous domains, with their ability to handle extended context reasoning receiving notable recognition. To speed up inference, maintaining a key-value (KV) cache memory is essential. Nonetheless, the growing demands for KV cache memory create significant hurdles for efficient implementation. This work introduces a novel architecture, Cross-Layer Latent Attention (CLLA), aimed at compressing the KV cache to less than 2% of its original size while maintaining comparable performance levels. CLLA integrates multiple aspects of KV cache compression, including attention head/dimension reduction, layer sharing, and quantization techniques, into a cohesive framework. Our extensive experiments demonstrate that CLLA achieves lossless performance on most tasks while utilizing minimal KV cache, marking a significant advancement in practical KV cache compression.", "authors": ["Zhen Yang", "J. N. Han", "Kan Wu", "Ruobing Xie", "An Wang", "Xingwu Sun", "Zhanhui Kang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-20", "url": "https://arxiv.org/abs/2410.15252", "pdf_url": "https://arxiv.org/pdf/2410.15252v1", "arxiv_id": "2410.15252", "doi": "10.48550/arXiv.2410.15252", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "dbfd1bd6ae82edeea05523b3e15ad753660e11888e270fa5ee7b94d185b50770", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-IR: Taming Low-Rank Experts for Efficient All-in-One Image Restoration", "abstract": "Prompt-based all-in-one image restoration (IR) frameworks have achieved remarkable performance by incorporating degradation-specific information into prompt modules. Nevertheless, handling the complex and diverse degradations encountered in real-world scenarios remains a significant challenge. To tackle this, we propose LoRA-IR, a flexible framework that dynamically leverages compact low-rank experts to facilitate efficient all-in-one image restoration. Specifically, LoRA-IR consists of two training stages: degradation-guided pre-training and parameter-efficient fine-tuning. In the pre-training stage, we enhance the pre-trained CLIP model by introducing a simple mechanism that scales it to higher resolutions, allowing us to extract robust degradation representations that adaptively guide the IR network. In the fine-tuning stage, we refine the pre-trained IR network through low-rank adaptation (LoRA). Built upon a Mixture-of-Experts (MoE) architecture, LoRA-IR dynamically integrates multiple low-rank restoration experts through a degradation-guided router. This dynamic integration mechanism significantly enhances our model's adaptability to diverse and unknown degradations in complex real-world scenarios. Extensive experiments demonstrate that LoRA-IR achieves SOTA performance across 14 IR tasks and 29 benchmarks, while maintaining computational efficiency. Code and pre-trained models will be available at: https://github.com/shallowdream204/LoRA-IR.", "authors": ["Yuang Ai", "Huaibo Huang", "Ran He"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-20", "url": "https://arxiv.org/abs/2410.15385", "pdf_url": "https://arxiv.org/pdf/2410.15385v2", "arxiv_id": "2410.15385", "doi": "10.48550/arXiv.2410.15385", "citation_count": 22, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/shallowdream204/LoRA-IR", "venue": "arXiv.org", "quality_score": 0.3404}
{"id": "a6007c3827e272850ff0e2ed9f750f946706285e60292992838bc0a74acdbe5f", "sources": ["arxiv", "semantic_scholar"], "title": "MomentumSMoE: Integrating Momentum into Sparse Mixture of Experts", "abstract": "Sparse Mixture of Experts (SMoE) has become the key to unlocking unparalleled scalability in deep learning. SMoE has the potential to exponentially increase parameter count while maintaining the efficiency of the model by only activating a small subset of these parameters for a given sample. However, it has been observed that SMoE suffers from unstable training and has difficulty adapting to new distributions, leading to the model's lack of robustness to data contamination. To overcome these limitations, we first establish a connection between the dynamics of the expert representations in SMoEs and gradient descent on a multi-objective optimization problem. Leveraging our framework, we then integrate momentum into SMoE and propose a new family of SMoEs named MomentumSMoE. We theoretically prove and numerically demonstrate that MomentumSMoE is more stable and robust than SMoE. In particular, we verify the advantages of MomentumSMoE over SMoE on a variety of practical tasks including ImageNet-1K object recognition and WikiText-103 language modeling. We demonstrate the applicability of MomentumSMoE to many types of SMoE models, including those in the Sparse MoE model for vision (V-MoE) and the Generalist Language Model (GLaM). We also show that other advanced momentum-based optimization methods, such as Adam, can be easily incorporated into the MomentumSMoE framework for designing new SMoE models with even better performance, almost negligible additional computation cost, and simple implementations.", "authors": ["Rachel S. Y. Teo", "Tan M. Nguyen"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-18", "url": "https://arxiv.org/abs/2410.14574", "pdf_url": "https://arxiv.org/pdf/2410.14574v1", "arxiv_id": "2410.14574", "doi": "10.48550/arXiv.2410.14574", "citation_count": 7, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/rachtsy/MomentumSMoE", "venue": "Neural Information Processing Systems", "quality_score": 0.2258}
{"id": "57841403dccdd6be129816aece33129c395688403484e4eeec90396db3d6b900", "sources": ["arxiv", "semantic_scholar"], "title": "AsymKV: Enabling 1-Bit Quantization of KV Cache with Layer-Wise Asymmetric Quantization Configurations", "abstract": "Large language models have shown exceptional capabilities in a wide range of tasks, such as text generation and video generation, among others. However, due to their massive parameter count, these models often require substantial storage space, imposing significant constraints on the machines deploying LLMs. To overcome this limitation, one research direction proposes to compress the models using integer replacements for floating-point numbers, in a process known as Quantization. Some recent studies suggest quantizing the key and value cache (KV Cache) of LLMs, and designing quantization techniques that treat the key and value matrices equivalently. This work delves deeper into the asymmetric structural roles of KV Cache, a phenomenon where the transformer's output loss is more sensitive to the quantization of key matrices. We conduct a systematic examination of the attention output error resulting from key and value quantization. The phenomenon inspires us to propose an asymmetric quantization strategy. Our approach allows for 1-bit quantization of the KV cache by implementing distinct configurations for key and value matrices. We carry out experiments across a variety of datasets, demonstrating that our proposed model allows for the quantization of up to 75% decoder layers with 1 bit, while simultaneously maintaining performance levels comparable to those of the models with floating parameters.", "authors": ["Qian Tao", "Wenyuan Yu", "Jingren Zhou"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-17", "url": "https://arxiv.org/abs/2410.13212", "pdf_url": "https://arxiv.org/pdf/2410.13212v1", "arxiv_id": "2410.13212", "doi": "10.48550/arXiv.2410.13212", "citation_count": 24, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Computational Linguistics", "quality_score": 0.3495}
{"id": "a95257ebe3fa2c50bb91993cf1d4384fc59455102d9dca6c7a91345c7c7741f3", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding Expert Structures on Minimax Parameter Estimation in Contaminated Mixture of Experts", "abstract": "We conduct the convergence analysis of parameter estimation in the contaminated mixture of experts. This model is motivated from the prompt learning problem where ones utilize prompts, which can be formulated as experts, to fine-tune a large-scale pre-trained model for learning downstream tasks. There are two fundamental challenges emerging from the analysis: (i) the proportion in the mixture of the pre-trained model and the prompt may converge to zero during the training, leading to the prompt vanishing issue; (ii) the algebraic interaction among parameters of the pre-trained model and the prompt can occur via some partial differential equations and decelerate the prompt learning. In response, we introduce a distinguishability condition to control the previous parameter interaction. Additionally, we also investigate various types of expert structure to understand their effects on the convergence behavior of parameter estimation. In each scenario, we provide comprehensive convergence rates of parameter estimation along with the corresponding minimax lower bounds. Finally, we run several numerical experiments to empirically justify our theoretical findings.", "authors": ["Fanqi Yan", "Huy Nguyen", "Dung Le", "Pedram Akbarian", "Nhat Ho"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-16", "url": "https://arxiv.org/abs/2410.12258", "pdf_url": "https://arxiv.org/pdf/2410.12258v2", "arxiv_id": "2410.12258", "doi": "10.48550/arXiv.2410.12258", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "63c30ac52b100750ebc5ee507e7404996724a7589c652a328e9c90408a92a948", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Scale Experts for Alignment-free RGBT Video Object Detection and A Unified Benchmark", "abstract": "Existing RGB-Thermal Video Object Detection (RGBT VOD) methods predominantly rely on the manual alignment of image pairs, that is both labor-intensive and time-consuming. This dependency significantly restricts the scalability and practical applicability of these methods in real-world scenarios. To address this critical limitation, we propose a novel framework termed the Mixture of Scale Experts Network (MSENet). MSENet integrates multiple experts trained at different perceptual scales, enabling the capture of scale discrepancies between RGB and thermal image pairs without the need for explicit alignment. Specifically, to address the issue of unaligned scales, MSENet introduces a set of experts designed to perceive the correlation between RGBT image pairs across various scales. These experts are capable of identifying and quantifying the scale differences inherent in the image pairs. Subsequently, a dynamic routing mechanism is incorporated to assign adaptive weights to each expert, allowing the network to dynamically select the most appropriate experts based on the specific characteristics of the input data. Furthermore, to address the issue of weakly unaligned positions, we integrate deformable convolution into the network. Deformable convolution is employed to learn position displacements between the RGB and thermal modalities, thereby mitigating the impact of spatial misalignment. To provide a comprehensive evaluation platform for alignment-free RGBT VOD, we introduce a new benchmark dataset. This dataset includes eleven common object categories, with a total of 60,988 images and 271,835 object instances. The dataset encompasses a wide range of scenes from both daily life and natural environments, ensuring high content diversity and complexity.", "authors": ["Qishun Wang", "Zhengzheng Tu", "Kunpeng Wang", "Le Gu", "Chuanwang Guo"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-16", "url": "https://arxiv.org/abs/2410.12143", "pdf_url": "https://arxiv.org/pdf/2410.12143v2", "arxiv_id": "2410.12143", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "f645130556e4a32b1c72746fdfbd69b065c3ad96d53f1de354f34a2296a37a42", "sources": ["arxiv", "semantic_scholar"], "title": "In-context KV-Cache Eviction for LLMs via Attention-Gate", "abstract": "The KV-Cache technique has become the standard for the inference of large language models (LLMs). Yet, it is widely criticized that KV-Cache can become a bottleneck of the LLM inference system. This paper enables a novel dynamic KV-Cache eviction policy by injecting a lightweight module called Attention-Gate to the model. It accepts the global context as input and yields eviction flags for each token. The self-attention modules in the model proceed according to the flags and cache only a subset of the KV states for next token prediction. The Attention-Gates can yield various flags for different heads and layers and be easily tuned on top of a pre-trained LLM via continual pre-training or supervised fine-tuning. The computational and memory overhead introduced by Attention-Gates can be minimal. We empirically evaluate the proposed approach across multiple scenarios, showing that effective eviction of redundant tokens can not only improve efficiency but also enhance performance.", "authors": ["Zihao Zeng", "Bokai Lin", "Tianqi Hou", "Hao Zhang", "Zhijie Deng"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-15", "url": "https://arxiv.org/abs/2410.12876", "pdf_url": "https://arxiv.org/pdf/2410.12876v3", "arxiv_id": "2410.12876", "doi": "10.48550/arXiv.2410.12876", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "d660d620abae7cca36f1728ab11883af688e70c97596d209e97149384cc694dc", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-Pruner: Pruning Mixture-of-Experts Large Language Model using the Hints from Its Router", "abstract": "Mixture-of-Experts (MoE) architectures face challenges such as high memory consumption and redundancy in experts. Pruning MoE can reduce network weights while maintaining model performance. Motivated by the recent observation of emergent large magnitude features in Large Language Models (LLM) and MoE routing policy, we propose MoE-Pruner, a method that prunes weights with the smallest magnitudes multiplied by the corresponding input activations and router weights, on each output neuron. Our pruning method is one-shot, requiring no retraining or weight updates. We evaluate our method on Mixtral-8x7B and Mixtral-8x22B across multiple language benchmarks. Experimental results show that our pruning method significantly outperforms state-of-the-art LLM pruning methods. Furthermore, our pruned MoE models can benefit from a pretrained teacher model through expert-wise knowledge distillation, improving performance post-pruning. Experimental results demonstrate that the Mixtral-8x7B model with 50% sparsity maintains 99% of the performance of the original model after the expert-wise knowledge distillation.", "authors": ["Yanyue Xie", "Zhi Zhang", "Ding Zhou", "Cong Xie", "Ziang Song", "Xin Liu", "Yanzhi Wang", "Xue Lin", "An Xu"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-15", "url": "https://arxiv.org/abs/2410.12013", "pdf_url": "https://arxiv.org/pdf/2410.12013v1", "arxiv_id": "2410.12013", "doi": "10.48550/arXiv.2410.12013", "citation_count": 49, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4247}
{"id": "28054630648b6dd2612849fccb22a9319779bc16f4deb7352cc6f3934880ccc1", "sources": ["arxiv", "semantic_scholar"], "title": "Quadratic Gating Mixture of Experts: Statistical Insights into Self-Attention", "abstract": "Mixture of Experts (MoE) models are well known for effectively scaling model capacity while preserving computational overheads. In this paper, we establish a rigorous relation between MoE and the self-attention mechanism, showing that each row of a self-attention matrix can be written as a quadratic gating mixture of linear experts. Motivated by this connection, we conduct a comprehensive convergence analysis of MoE models with two different quadratic gating functions, namely the quadratic polynomial gate and the quadratic monomial gate, offering useful insights into the design of gating and experts for the MoE framework. First, our analysis indicates that the use of the quadratic monomial gate yields an improved sample efficiency for estimating parameters and experts compared to the quadratic polynomial gate. Second, parameter and expert estimation rates become significantly faster when employing non-linear experts in place of linear experts. Combining these theoretical insights with the above link between MoE and self-attention, we propose a novel \\emph{active-attention} mechanism where we apply a non-linear activation function to the value matrix in the formula of self-attention. Finally, we demonstrate that the proposed active-attention outperforms the standard self-attention through several extensive experiments in various tasks, including image classification, language modeling, and multivariate time series forecasting.", "authors": ["Pedram Akbarian", "Huy Nguyen", "Xing Han", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2024-10-15", "url": "https://arxiv.org/abs/2410.11222", "pdf_url": "https://arxiv.org/pdf/2410.11222v3", "arxiv_id": "2410.11222", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "114de5d1d2cddd5e0a1a2f91d48728ddc4a496ea960271f330b72bd79c336ed9", "sources": ["arxiv", "semantic_scholar"], "title": "MoH: Multi-Head Attention as Mixture-of-Head Attention", "abstract": "In this work, we upgrade the multi-head attention mechanism, the core of the Transformer model, to improve efficiency while maintaining or surpassing the previous accuracy level. We show that multi-head attention can be expressed in the summation form. Drawing on the insight that not all attention heads hold equal significance, we propose Mixture-of-Head attention (MoH), a new architecture that treats attention heads as experts in the Mixture-of-Experts (MoE) mechanism. MoH has two significant advantages: First, MoH enables each token to select the appropriate attention heads, enhancing inference efficiency without compromising accuracy or increasing the number of parameters. Second, MoH replaces the standard summation in multi-head attention with a weighted summation, introducing flexibility to the attention mechanism and unlocking extra performance potential. Extensive experiments on ViT, DiT, and LLMs demonstrate that MoH outperforms multi-head attention by using only 50%-90% of the attention heads. Moreover, we demonstrate that pre-trained multi-head attention models, such as LLaMA3-8B, can be further continue-tuned into our MoH models. Notably, MoH-LLaMA3-8B achieves an average accuracy of 64.0% across 14 benchmarks, outperforming LLaMA3-8B by 2.4% by utilizing only 75% of the attention heads. We believe the proposed MoH is a promising alternative to multi-head attention and provides a strong foundation for developing advanced and efficient attention-based models.", "authors": ["Peng Jin", "Bo Zhu", "Li Yuan", "Shuicheng Yan"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-15", "url": "https://arxiv.org/abs/2410.11842", "pdf_url": "https://arxiv.org/pdf/2410.11842v3", "arxiv_id": "2410.11842", "doi": "10.48550/arXiv.2410.11842", "citation_count": 53, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/SkyworkAI/MoH", "venue": "International Conference on Machine Learning", "quality_score": 0.4515}
{"id": "d216c12b39284ef83703e973679cd5711bf7e704095099fea63a35d69d36efd1", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix", "abstract": "Large Language Models (LLMs) have shown immense potential in enhancing various aspects of our daily lives, from conversational AI to search and AI assistants. However, their growing capabilities come at the cost of extremely large model sizes, making deployment on edge devices challenging due to memory and computational constraints. This paper introduces a novel approach to LLM weight pruning that directly optimizes for approximating the attention matrix, a core component of transformer architectures. Unlike existing methods that focus on linear approximations, our approach accounts for the non-linear nature of the Softmax attention mechanism. We provide theoretical guarantees for the convergence of our Gradient Descent-based optimization method to a near-optimal pruning mask solution. Our empirical results demonstrate the effectiveness of our non-linear pruning approach in maintaining model performance while significantly reducing computational costs, which is beyond the current state-of-the-art methods, i.e., SparseGPT and Wanda, by a large margin. This work establishes a new theoretical foundation for pruning algorithm design in LLMs, potentially paving the way for more efficient LLM inference on resource-constrained devices.", "authors": ["Yingyu Liang", "Jiangxuan Long", "Zhenmei Shi", "Zhao Song", "Yufa Zhou"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-15", "url": "https://arxiv.org/abs/2410.11261", "pdf_url": "https://arxiv.org/pdf/2410.11261v2", "arxiv_id": "2410.11261", "doi": "10.48550/arXiv.2410.11261", "citation_count": 32, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3796}
{"id": "37e25fa96cc3e1f53211126c273833326a2b43b66f1b3d0c8abb4ef7234dbba9", "sources": ["arxiv", "semantic_scholar"], "title": "GaVaMoE: Gaussian-Variational Gated Mixture of Experts for Explainable Recommendation", "abstract": "Large language model-based explainable recommendation (LLM-based ER) systems show promise in generating human-like explanations for recommendations. However, they face challenges in modeling user-item collaborative preferences, personalizing explanations, and handling sparse user-item interactions. To address these issues, we propose GaVaMoE, a novel Gaussian-Variational Gated Mixture of Experts framework for explainable recommendation. GaVaMoE introduces two key components: (1) a rating reconstruction module that employs Variational Autoencoder (VAE) with a Gaussian Mixture Model (GMM) to capture complex user-item collaborative preferences, serving as a pre-trained multi-gating mechanism; and (2) a set of fine-grained expert models coupled with the multi-gating mechanism for generating highly personalized explanations. The VAE component models latent factors in user-item interactions, while the GMM clusters users with similar behaviors. Each cluster corresponds to a gate in the multi-gating mechanism, routing user-item pairs to appropriate expert models. This architecture enables GaVaMoE to generate tailored explanations for specific user types and preferences, mitigating data sparsity by leveraging user similarities. Extensive experiments on three real-world datasets demonstrate that GaVaMoE significantly outperforms existing methods in explanation quality, personalization, and consistency. Notably, GaVaMoE exhibits robust performance in scenarios with sparse user-item interactions, maintaining high-quality explanations even for users with limited historical data.", "authors": ["Fei Tang", "Yongliang Shen", "Hang Zhang", "Zeqi Tan", "Wenqi Zhang", "Zhibiao Huang", "Kaitao Song", "Weiming Lu", "Yueting Zhuang"], "categories": ["cs.IR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-15", "url": "https://arxiv.org/abs/2410.11841", "pdf_url": "https://arxiv.org/pdf/2410.11841v2", "arxiv_id": "2410.11841", "doi": "10.48550/arXiv.2410.11841", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "5f4ad782dfe90783ffed675e3cdcca255adb4ba40d07ee82f1e740303b1f4769", "sources": ["arxiv", "semantic_scholar"], "title": "AlphaLoRA: Assigning LoRA Experts Based on Layer Training Quality", "abstract": "Parameter-efficient fine-tuning methods, such as Low-Rank Adaptation (LoRA), are known to enhance training efficiency in Large Language Models (LLMs). Due to the limited parameters of LoRA, recent studies seek to combine LoRA with Mixture-of-Experts (MoE) to boost performance across various tasks. However, inspired by the observed redundancy in traditional MoE structures, previous studies identify similar redundancy among LoRA experts within the MoE architecture, highlighting the necessity for non-uniform allocation of LoRA experts across different layers. In this paper, we leverage Heavy-Tailed Self-Regularization (HT-SR) Theory to design a fine-grained allocation strategy. Our analysis reveals that the number of experts per layer correlates with layer training quality, which exhibits significant variability across layers. Based on this, we introduce AlphaLoRA, a theoretically principled and training-free method for allocating LoRA experts to further mitigate redundancy. Experiments on three models across ten language processing and reasoning benchmarks demonstrate that AlphaLoRA achieves comparable or superior performance over all baselines. Our code is available at https://github.com/morelife2017/alphalora.", "authors": ["Peijun Qing", "Chongyang Gao", "Yefan Zhou", "Xingjian Diao", "Yaoqing Yang", "Soroush Vosoughi"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-14", "url": "https://arxiv.org/abs/2410.10054", "pdf_url": "https://arxiv.org/pdf/2410.10054v1", "arxiv_id": "2410.10054", "doi": "10.48550/arXiv.2410.10054", "citation_count": 26, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/morelife2017/alphalora", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3578}
{"id": "2b8f0f1390a9f68775d801554c2e47cfaaf44b8cf3862f2cd4bbea8674ab43fa", "sources": ["arxiv", "semantic_scholar"], "title": "Tighter Risk Bounds for Mixtures of Experts", "abstract": "In this work, we provide upper bounds on the risk of mixtures of experts by imposing local differential privacy (LDP) on their gating mechanism. These theoretical guarantees are tailored to mixtures of experts that utilize the one-out-of-$n$ gating mechanism, as opposed to the conventional $n$-out-of-$n$ mechanism. The bounds exhibit logarithmic dependence on the number of experts, and encapsulate the dependence on the gating mechanism in the LDP parameter, making them significantly tighter than existing bounds, under reasonable conditions. Experimental results support our theory, demonstrating that our approach enhances the generalization ability of mixtures of experts and validating the feasibility of imposing LDP on the gating mechanism.", "authors": ["Wissam Akretche", "Frédéric LeBlanc", "Mario Marchand"], "categories": ["cs.LG", "cs.CR", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-14", "url": "https://arxiv.org/abs/2410.10397", "pdf_url": "https://arxiv.org/pdf/2410.10397v1", "arxiv_id": "2410.10397", "doi": "10.48550/arXiv.2410.10397", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "ea95cee591d30840dc6f7a170bf98b148252da4f4cefb213e12bd233220cdd16", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts Made Personalized: Federated Prompt Learning for Vision-Language Models", "abstract": "Federated prompt learning benefits federated learning with CLIP-like Vision-Language Model's (VLM's) robust representation learning ability through prompt learning. However, current federated prompt learning methods are habitually restricted to the traditional FL paradigm, where the participating clients are generally only allowed to download a single globally aggregated model from the server. While justifiable for training full-sized models under federated settings, in this work, we argue that this paradigm is ill-suited for lightweight prompts. By facilitating the clients to download multiple pre-aggregated prompts as fixed non-local experts, we propose Personalized Federated Mixture of Adaptive Prompts (pFedMoAP), a novel FL framework that personalizes the prompt learning process through the lens of Mixture of Experts (MoE). pFedMoAP implements a local attention-based gating network that learns to generate enhanced text features for better alignment with local image data, benefiting from both local and downloaded non-local adaptive prompt experts. Extensive experiments on 9 datasets under various federated settings demonstrate the efficacy of the proposed pFedMoAP algorithm. The code is available at https://github.com/ljaiverson/pFedMoAP.", "authors": ["Jun Luo", "Chen Chen", "Shandong Wu"], "categories": ["cs.LG", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2024-10-14", "url": "https://arxiv.org/abs/2410.10114", "pdf_url": "https://arxiv.org/pdf/2410.10114v4", "arxiv_id": "2410.10114", "doi": "10.48550/arXiv.2410.10114", "citation_count": 21, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/ljaiverson/pFedMoAP", "venue": "International Conference on Learning Representations", "quality_score": 0.3356}
{"id": "4dc1e9a97c3ece740abccdfe0c89482efc96b4b04fbfeb6a20c27152497773c4", "sources": ["arxiv", "semantic_scholar"], "title": "Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family Experts", "abstract": "Adapting medical Large Language Models to local languages can reduce barriers to accessing healthcare services, but data scarcity remains a significant challenge, particularly for low-resource languages. To address this, we first construct a high-quality medical dataset and conduct analysis to ensure its quality. In order to leverage the generalization capability of multilingual LLMs to efficiently scale to more resource-constrained languages, we explore the internal information flow of LLMs from a multilingual perspective using Mixture of Experts (MoE) modularity. Technically, we propose a novel MoE routing method that employs language-specific experts and cross-lingual routing. Inspired by circuit theory, our routing analysis revealed a Spread Out in the End information flow mechanism: while earlier layers concentrate cross-lingual information flow, the later layers exhibit language-specific divergence. This insight directly led to the development of the Post-MoE architecture, which applies sparse routing only in the later layers while maintaining dense others. Experimental results demonstrate that this approach enhances the generalization of multilingual models to other languages while preserving interpretability. Finally, to efficiently scale the model to 50 languages, we introduce the concept of language family experts, drawing on linguistic priors, which enables scaling the number of languages without adding additional parameters.", "authors": ["Guorui Zheng", "Xidong Wang", "Juhao Liang", "Nuo Chen", "Yuping Zheng", "Benyou Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-14", "url": "https://arxiv.org/abs/2410.10626", "pdf_url": "https://arxiv.org/pdf/2410.10626v2", "arxiv_id": "2410.10626", "doi": "10.48550/arXiv.2410.10626", "citation_count": 16, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3076}
{"id": "219cd01328a1fd824ceb9e0c50a50c117a8856d2dc00349c53f9fbc85d3ca7f4", "sources": ["arxiv", "semantic_scholar"], "title": "Gaussian Mixture Vector Quantization with Aggregated Categorical Posterior", "abstract": "The vector quantization is a widely used method to map continuous representation to discrete space and has important application in tokenization for generative mode, bottlenecking information and many other tasks in machine learning. Vector Quantized Variational Autoencoder (VQ-VAE) is a type of variational autoencoder using discrete embedding as latent. We generalize the technique further, enriching the probabilistic framework with a Gaussian mixture as the underlying generative model. This framework leverages a codebook of latent means and adaptive variances to capture complex data distributions. This principled framework avoids various heuristics and strong assumptions that are needed with the VQ-VAE to address training instability and to improve codebook utilization. This approach integrates the benefits of both discrete and continuous representations within a variational Bayesian framework. Furthermore, by introducing the \\textit{Aggregated Categorical Posterior Evidence Lower Bound} (ALBO), we offer a principled alternative optimization objective that aligns variational distributions with the generative model. Our experiments demonstrate that GM-VQ improves codebook utilization and reduces information loss without relying on handcrafted heuristics.", "authors": ["Mingyuan Yan", "Jiawei Wu", "Rushi Shah", "Dianbo Liu"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-14", "url": "https://arxiv.org/abs/2410.10180", "pdf_url": "https://arxiv.org/pdf/2410.10180v1", "arxiv_id": "2410.10180", "doi": "10.48550/arXiv.2410.10180", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "80ca1d8225dfe729aa0866a33b90bc0413ed8833082ed9e1ef83031db582a0bf", "sources": ["arxiv", "semantic_scholar"], "title": "MoIN: Mixture of Introvert Experts to Upcycle an LLM", "abstract": "The goal of this paper is to improve (upcycle) an existing large language model without the prohibitive requirements of continued pre-training of the full-model. The idea is to split the pre-training data into semantically relevant groups and train an expert on each subset. An expert takes the form of a lightweight adapter added on the top of a frozen base model. During inference, an incoming query is first routed to the most relevant expert which is then loaded onto the base model for the forward pass. Unlike typical Mixture of Experts (MoE) models, the experts in our method do not work with other experts for a single query. Hence, we dub them \"introvert\" experts. Freezing the base model and keeping the experts as lightweight adapters allows extreme parallelism during training and inference. Training of all experts can be done in parallel without any communication channels between them. Similarly, the inference can also be heavily parallelized by distributing experts on different GPUs and routing each request to the GPU containing its relevant expert. We implement a proof-of-concept version of this method and show the validity of our approach.", "authors": ["Ajinkya Tejankar", "KL Navaneet", "Ujjawal Panchal", "Kossar Pourahmadi", "Hamed Pirsiavash"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-13", "url": "https://arxiv.org/abs/2410.09687", "pdf_url": "https://arxiv.org/pdf/2410.09687v1", "arxiv_id": "2410.09687", "doi": "10.48550/arXiv.2410.09687", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "285351230c960aa4b9e835a71487ab5fba3d07250a5c2e834b4ace13e33656d9", "sources": ["arxiv", "semantic_scholar"], "title": "AT-MoE: Adaptive Task-planning Mixture of Experts via LoRA Approach", "abstract": "The advent of Large Language Models (LLMs) has ushered in a new era of artificial intelligence, with the potential to transform various sectors through automation and insightful analysis. The Mixture of Experts (MoE) architecture has been proposed as a solution to enhance model performance in complex tasks. Yet, existing MoE models struggle with task-specific learning and interpretability, especially in fields like medicine where precision is critical. This paper introduces the Adaptive Task-planing Mixture of Experts(AT-MoE), an innovative architecture designed to address these limitations. We first train task-specific experts via LoRA approach to enhance problem-solving capabilities and interpretability in specialized areas. Subsequently, we introduce a layer-wise adaptive grouped routing module that optimizes module fusion based on complex task instructions, ensuring optimal task resolution. The grouped routing module first perform overall weight allocation from the dimension of the expert group, and then conduct local weight normalization adjustments within the group. This design maintains multi-dimensional balance, controllability, and interpretability, while facilitating task-specific fusion in response to complex instructions.", "authors": ["Xurui Li", "Juanjuan Yao"], "categories": ["cs.LG", "cs.AI", "cs.CE"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-12", "url": "https://arxiv.org/abs/2410.10896", "pdf_url": "https://arxiv.org/pdf/2410.10896v2", "arxiv_id": "2410.10896", "doi": "10.48550/arXiv.2410.10896", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "fb00b29d96dac4819e32c7eb6abb66a8ff91ce1286186ccfc7ea367677219a72", "sources": ["arxiv", "semantic_scholar"], "title": "Semi-Supervised Learning of Noisy Mixture of Experts Models", "abstract": "The mixture of experts (MoE) model is a versatile framework for predictive modeling that has gained renewed interest in the age of large language models. A collection of predictive ``experts'' is learned along with a ``gating function'' that controls how much influence each expert is given when a prediction is made. This structure allows relatively simple models to excel in complex, heterogeneous data settings. In many contemporary settings, unlabeled data are widely available while labeled data are difficult to obtain. Semi-supervised learning methods seek to leverage the unlabeled data. We propose a novel method for semi-supervised learning of MoE models. We start from a semi-supervised MoE model that was developed by oceanographers that makes the strong assumption that the latent clustering structure in unlabeled data maps directly to the influence that the gating function should give each expert in the supervised task. We relax this assumption, imagining a noisy connection between the two, and propose an algorithm based on least trimmed squares, which succeeds even in the presence of misaligned data. Our theoretical analysis characterizes the conditions under which our approach yields estimators with a near-parametric rate of convergence. Simulated and real data examples demonstrate the method's efficacy.", "authors": ["Oh-Ran Kwon", "Gourab Mukherjee", "Jacob Bien"], "categories": ["stat.ME"], "fields_of_study": ["Mathematics"], "published_date": "2024-10-11", "url": "https://arxiv.org/abs/2410.09039", "pdf_url": "https://arxiv.org/pdf/2410.09039v1", "arxiv_id": "2410.09039", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "88ca589e2165906e7f3e56e6f2a8f64f7affe7a4d68ffc0d1f069f569a628687", "sources": ["arxiv", "semantic_scholar"], "title": "SLIM: Let LLM Learn More and Forget Less with Soft LoRA and Identity Mixture", "abstract": "Although many efforts have been made, it is still a challenge to balance the training budget, downstream performance, and the general capabilities of the LLMs in many applications. Training the whole model for downstream tasks is expensive, and could easily result in catastrophic forgetting. By introducing parameter-efficient fine-tuning (PEFT), the training cost could be reduced, but it still suffers from forgetting, and limits the learning on the downstream tasks. To efficiently fine-tune the LLMs with less limitation to their downstream performance while mitigating the forgetting of general capabilities, we propose a novel mixture of expert (MoE) framework based on Soft LoRA and Identity Mixture (SLIM), that allows dynamic routing between LoRA adapters and skipping connection, enables the suppression of forgetting. We adopt weight-yielding with sliding clustering for better out-of-domain distinguish to enhance the routing. We also propose to convert the mixture of low-rank adapters to the model merging formulation and introduce fast dynamic merging of LoRA adapters to keep the general capabilities of the base model. Extensive experiments demonstrate that the proposed SLIM is comparable to the state-of-the-art PEFT approaches on the downstream tasks while achieving the leading performance in mitigating catastrophic forgetting.", "authors": ["Jiayi Han", "Liang Du", "Hongwei Du", "Xiangguo Zhou", "Yiwen Wu", "Weibo Zheng", "Donghong Han"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.07739", "pdf_url": "https://arxiv.org/pdf/2410.07739v3", "arxiv_id": "2410.07739", "doi": "10.48550/arXiv.2410.07739", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.294}
{"id": "c5072499d536e3b330b336b1cfce7138e179e4f8aa77866621e8130075246472", "sources": ["arxiv", "semantic_scholar"], "title": "TurboRAG: Accelerating Retrieval-Augmented Generation with Precomputed KV Caches for Chunked Text", "abstract": "Current Retrieval-Augmented Generation (RAG) systems concatenate and process numerous retrieved document chunks for prefill which requires a large volume of computation, therefore leading to significant latency in time-to-first-token (TTFT). To reduce the computation overhead as well as TTFT, we introduce TurboRAG, a novel RAG system that redesigns the inference paradigm of the current RAG system by first pre-computing and storing the key-value (KV) caches of documents offline, and then directly retrieving the saved KV cache for prefill. Hence, online computation of KV caches is eliminated during inference. In addition, we provide a number of insights into the mask matrix and positional embedding mechanisms, plus fine-tune a pretrained language model to maintain model accuracy of TurboRAG. Our approach is applicable to most existing large language models and their applications without any requirement in modification of models and inference systems. Experimental results across a suite of RAG benchmarks demonstrate that TurboRAG reduces TTFT by up to 9.4x compared to the conventional RAG systems (on an average of 8.6x), but reserving comparable performance to the standard RAG systems.", "authors": ["Songshuo Lu", "Hua Wang", "Yutian Rong", "Zhi Chen", "Yaohua Tang"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.07590", "pdf_url": "https://arxiv.org/pdf/2410.07590v1", "arxiv_id": "2410.07590", "doi": "10.48550/arXiv.2410.07590", "citation_count": 55, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4515}
{"id": "dd0f9fa87403ffa3b0b3f4288ab13d66dbb3d288b416d42e6079aeec93062d3c", "sources": ["arxiv", "semantic_scholar"], "title": "Reducing the Cost of Dropout in Flash-Attention by Hiding RNG with GEMM", "abstract": "Dropout, a network operator, when enabled is likely to dramatically impact the performance of Flash-Attention, which in turn increases the end-to-end training time of Large-Language-Models (LLMs). The main contributor to such performance degradation is the Random Number Generation (RNG) phase. The state-of-the-art optimization is to fuse RNG into the Flash-Attention kernel. However, while RNG and Attention do not compete on compute or memory resources, they are bounded on the same lower-level architecture bottlenecks. Fusion can hardly hide RNG latency within the Attention kernel. We propose overlapping RNG with previous GEMM layers in the network to hide RNG latency and improve end-to-end performance. RNG and GEMM have distinct resource requirements and hardware bottlenecks, so they can run together without compromising each other's performance. We propose a fine-grained analytical performance model that analyzes low-level architecture resource utilization to evaluate RNG-GEMM overlapping performance benefits. This model, cross-validated by silicon results, shows 1.26x speedup for overlapping RNG and GEMM layers over a sequential implementation on one Transformer Block (one LLM layer including multi-head attention and feed-forward layers), and 1.22x over state-of-the-art fusion implementation, for Llama3 on GH100 GPUs with FP8 precision. Because the kernel patterns are regular, the findings of the shared bottlenecks, as well as the achievable performance benefits, can be generalized to different model architectures, software implementations and hardware configurations.", "authors": ["Haiyue Ma", "Jian Liu", "Ronny Krashinsky"], "categories": ["cs.AR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.07531", "pdf_url": "https://arxiv.org/pdf/2410.07531v2", "arxiv_id": "2410.07531", "doi": "10.48550/arXiv.2410.07531", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "46783fee3ad70823d1dd6657432247741bfc16a8ed69a57d290d2d728981bc07", "sources": ["arxiv", "semantic_scholar"], "title": "Upcycling Large Language Models into Mixture of Experts", "abstract": "Upcycling pre-trained dense language models into sparse mixture-of-experts (MoE) models is an efficient approach to increase the model capacity of already trained models. However, optimal techniques for upcycling at scale remain unclear. In this work, we conduct an extensive study of upcycling methods and hyperparameters for billion-parameter scale language models. We propose a novel \"virtual group\" initialization scheme and weight scaling approach to enable upcycling into fine-grained MoE architectures. Through ablations, we find that upcycling outperforms continued dense model training. In addition, we show that softmax-then-topK expert routing improves over topK-then-softmax approach and higher granularity MoEs can help improve accuracy. Finally, we upcycled Nemotron-4 15B on 1T tokens and compared it to a continuously trained version of the same model on the same 1T tokens: the continuous trained model achieved 65.3% MMLU, whereas the upcycled model achieved 67.6%. Our results offer insights and best practices to effectively leverage upcycling for building MoE language models. Code is available.", "authors": ["Ethan He", "Abhinav Khattar", "Ryan Prenger", "Vijay Korthikanti", "Zijie Yan", "Tong Liu", "Shiqing Fan", "Ashwath Aithal", "Mohammad Shoeybi", "Bryan Catanzaro"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.07524", "pdf_url": "https://arxiv.org/pdf/2410.07524v2", "arxiv_id": "2410.07524", "doi": "10.48550/arXiv.2410.07524", "citation_count": 42, "influential_citation_count": 6, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4225}
{"id": "47f49a283e0d763c2165b06df565b31aa1d00f0e6b95de267833bcf442a17c85", "sources": ["arxiv", "semantic_scholar"], "title": "QuAILoRA: Quantization-Aware Initialization for LoRA", "abstract": "QLoRA reduces the memory-cost of fine-tuning a large language model (LLM) with LoRA by quantizing the base LLM. However, quantization introduces quantization errors that negatively impact model performance after fine-tuning. In this paper we introduce QuAILoRA, a quantization-aware initialization for LoRA that mitigates this negative impact by decreasing quantization errors at initialization. Our method spends a small amount of computational overhead to compute this quantization-aware initialization, without increasing the memory-cost of fine-tuning. We evaluate our method on several causal language modeling and downstream evaluation tasks using several different model sizes and families. We observe that almost all LLMs fined-tuned with QuAILoRA achieve better validation perplexity. When evaluated on downstream tasks, we find that QuAILoRA yields improvements proportional to the negative effect of quantization error. On average, applying QuAILoRA to 4-bit QLoRA models yields 75% of the validation perplexity decrease and 86% of the downstream task accuracy increase as doubling the quantization precision to 8-bit, without increasing GPU memory utilization during fine-tuning.", "authors": ["Neal Lawton", "Aishwarya Padmakumar", "Judith Gaspers", "Jack FitzGerald", "Anoop Kumar", "Greg Ver Steeg", "Aram Galstyan"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-09", "url": "https://arxiv.org/abs/2410.14713", "pdf_url": "https://arxiv.org/pdf/2410.14713v1", "arxiv_id": "2410.14713", "doi": "10.48550/arXiv.2410.14713", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1193}
{"id": "bf782496a4bb6aa1984b811a33479eca72026c9f1ef605286183adffcfdd30a0", "sources": ["arxiv", "semantic_scholar"], "title": "MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts", "abstract": "In this work, we aim to simultaneously enhance the effectiveness and efficiency of Mixture-of-Experts (MoE) methods. To achieve this, we propose MoE++, a general and heterogeneous MoE framework that integrates both Feed-Forward Network~(FFN) and zero-computation experts. Specifically, we introduce three types of zero-computation experts: the zero expert, copy expert, and constant expert, which correspond to discard, skip, and replace operations, respectively. This design offers three key advantages: (i) Low Computing Overhead: Unlike the uniform mixing mechanism for all tokens within vanilla MoE, MoE++ allows each token to engage with a dynamic number of FFNs, be adjusted by constant vectors, or even skip the MoE layer entirely. (ii) High Performance: By enabling simple tokens to utilize fewer FFN experts, MoE++ allows more experts to focus on challenging tokens, thereby unlocking greater performance potential than vanilla MoE. (iii) Deployment Friendly: Given that zero-computation experts have negligible parameters, we can deploy all zero-computation experts on each GPU, eliminating the significant communication overhead and expert load imbalance associated with FFN experts distributed across different GPUs. Moreover, we leverage gating residuals, enabling each token to consider the pathway taken in the previous layer when selecting the appropriate experts. Extensive experimental results demonstrate that MoE++ achieves better performance while delivering 1.1-2.1x expert forward throughput compared to a vanilla MoE model of the same size, which lays a solid foundation for developing advanced and efficient MoE-related models.", "authors": ["Peng Jin", "Bo Zhu", "Li Yuan", "Shuicheng Yan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-09", "url": "https://arxiv.org/abs/2410.07348", "pdf_url": "https://arxiv.org/pdf/2410.07348v1", "arxiv_id": "2410.07348", "doi": "10.48550/arXiv.2410.07348", "citation_count": 52, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/SkyworkAI/MoE-plus-plus", "venue": "International Conference on Learning Representations", "quality_score": 0.4515}
{"id": "58903eae536555116f1205c7a99c2f9af76b9117a595fc11c8904245d484baf1", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture Compressor for Mixture-of-Experts LLMs Gains More", "abstract": "Mixture-of-Experts large language models (MoE-LLMs) marks a significant step forward of language models, however, they encounter two critical challenges in practice: 1) expert parameters lead to considerable memory consumption and loading latency; and 2) the current activated experts are redundant, as many tokens may only require a single expert. Motivated by these issues, we investigate the MoE-LLMs and make two key observations: a) different experts exhibit varying behaviors on activation reconstruction error, routing scores, and activated frequencies, highlighting their differing importance, and b) not all tokens are equally important -- only a small subset is critical. Building on these insights, we propose MC, a training-free Mixture-Compressor for MoE-LLMs, which leverages the significance of both experts and tokens to achieve an extreme compression. First, to mitigate storage and loading overheads, we introduce Pre-Loading Mixed-Precision Quantization, which formulates the adaptive bit-width allocation as a Linear Programming problem, where the objective function balances multi-factors reflecting the importance of each expert. Additionally, we develop Online Dynamic Pruning, which identifies important tokens to retain and dynamically select activated experts for other tokens during inference to optimize efficiency while maintaining performance. Our MC integrates static quantization and dynamic pruning to collaboratively achieve extreme compression for MoE-LLMs with less accuracy loss, ensuring an optimal trade-off between performance and efficiency. Extensive experiments confirm the effectiveness of our approach. For instance, at 2.54 bits, MC compresses 76.6% of the model, with only a 3.8% average accuracy loss. During dynamic inference, we further reduce activated parameters by 15%, with a performance drop of less than 0.6%.", "authors": ["Wei Huang", "Yue Liao", "Jianhui Liu", "Ruifei He", "Haoru Tan", "Shiming Zhang", "Hongsheng Li", "Si Liu", "Xiaojuan Qi"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-08", "url": "https://arxiv.org/abs/2410.06270", "pdf_url": "https://arxiv.org/pdf/2410.06270v2", "arxiv_id": "2410.06270", "doi": "10.48550/arXiv.2410.06270", "citation_count": 39, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4771}
{"id": "dc66f01630872879d5638e3193a639fe9f610e36a597c1d856b2899c9b689c39", "sources": ["arxiv", "semantic_scholar"], "title": "PrefixQuant: Eliminating Outliers by Prefixed Tokens for Large Language Models Quantization", "abstract": "Existing weight-activation quantization methods for Large Language Models (LLMs) primarily address channel-wise outliers but often neglect token-wise outliers, which limits the accuracy of quantized models. In this work, we propose PrefixQuant, a novel quantization method that achieves state-of-the-art performance across various precision levels (W4A4KV4 and W4A8KV4) and granularities (dynamic and static quantization) by effectively isolating token-wise outliers. First, PrefixQuant eliminates token-wise outliers by prefixing outlier tokens in the KV cache, a process that is training-free and highly efficient (e.g., 1 minutes for Llama-3-70B). Second, PrefixQuant introduces new trainable parameters for block-wise training to compensate for quantization error. Our experiments show that PrefixQuant significantly outperforms existing dynamic quantization methods, even under coarser static quantization settings. For instance, PrefixQuant achieves an average accuracy improvement of +3.08 and +2.85 points over SpinQuant (dynamic quantization) on five zero-shot reasoning tasks under dynamic and static quantization settings, respectively, on W4A4KV4 Llama-3-8B. Additionally, we demonstrate up to 2.74x prefilling speedup and 2.16x decoding speedup for LLMs using W4A4 PrefixQuant. Our code is available at https://github.com/ChenMnZ/PrefixQuant.", "authors": ["Mengzhao Chen", "Yi Liu", "Jiahao Wang", "Yi Bin", "Wenqi Shao", "Ping Luo"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-07", "url": "https://arxiv.org/abs/2410.05265", "pdf_url": "https://arxiv.org/pdf/2410.05265v2", "arxiv_id": "2410.05265", "doi": null, "citation_count": 11, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/ChenMnZ/PrefixQuant", "venue": null, "quality_score": 0.2698}
{"id": "0a74129afdb790f4a7031b2c7a435f07a1cd171b17c818976493beb49d4afb7a", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring the Limitations of Mamba in COPY and CoT Reasoning", "abstract": "Transformers have become the backbone of modern Large Language Models (LLMs); however, their inference overhead grows linearly with the sequence length, posing challenges for modeling long sequences. In light of this, Mamba has attracted attention for maintaining a constant inference size, with empirical evidence demonstrating that it can match Transformer performance in sequence modeling while significantly reducing computational costs. However, an open question remains: can Mamba always bring savings while achieving performance comparable to Transformers? In this paper, we focus on analyzing the expressive ability of Mamba to perform our defined COPY operation and Chain of Thought (CoT) reasoning. First, inspired by the connection between Mamba and linear attention, we show that constant-sized Mamba may struggle to perform COPY operations while Transformers can handle them more easily. However, when the size of Mamba grows linearly with the input sequence length, it can accurately perform COPY, but in this case, Mamba no longer provides overhead savings. Based on this observation, we further analyze Mamba's ability to tackle CoT tasks, which can be described by the Dynamic Programming (DP) problems. Our findings suggest that to solve arbitrary DP problems, the total cost of Mamba is still comparable to standard Transformers. However, similar to efficient Transformers, when facing DP problems with favorable properties such as locality, Mamba can provide savings in overhead. Our experiments on the copy and CoT tasks further demonstrate Mamba's limitations compared to Transformers in learning these tasks.", "authors": ["Ruifeng Ren", "Zhicong Li", "Yong Liu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-04", "url": "https://arxiv.org/abs/2410.03810", "pdf_url": "https://arxiv.org/pdf/2410.03810v3", "arxiv_id": "2410.03810", "doi": "10.18653/v1/2025.emnlp-main.634", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2113}
{"id": "59917b84a540419e78abf7ab2bca6042ff81ff3d9cd7bfffb49c90917763202c", "sources": ["arxiv", "semantic_scholar"], "title": "Compute Or Load KV Cache? Why Not Both?", "abstract": "Large Language Models (LLMs) are increasingly deployed in large-scale online services, enabling sophisticated applications. However, the computational overhead of generating key-value (KV) caches in the prefill stage presents a major bottleneck, particularly for long-context inputs. Prefix caching mitigates this issue by storing KV caches for reuse, reducing redundant computation. Despite its advantages, prefix caching suffers from high latency due to the limited I/O bandwidth of storage devices, constraining inference efficiency. To address this challenge, we introduce Cake, a novel KV cache loading system that optimally utilizes both computational and I/O resources in parallel. Cake employs a bidirectional scheduling strategy that dynamically balances KV cache computation and loading, ensuring efficient resource utilization. Additionally, Cake incorporates an adaptive scheduling mechanism that seamlessly integrates with non-prefix caching requests, improving system throughput and adapting to fluctuating resource availabilty. Through extensive evaluations across various hardware configurations, datasets, and storage conditions, Cake achieves on average 2.6x reduction in Time to First Token (TTFT) compared to compute-only and I/O-only methods. Our findings highlight Cake as an effective and practical solution for optimizing long-context LLM inference, bridging the gap between computation and I/O efficiency in large-scale AI deployments.", "authors": ["Shuowei Jin", "Xueshen Liu", "Qingzhao Zhang", "Z. Morley Mao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-04", "url": "https://arxiv.org/abs/2410.03065", "pdf_url": "https://arxiv.org/pdf/2410.03065v2", "arxiv_id": "2410.03065", "doi": "10.48550/arXiv.2410.03065", "citation_count": 32, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3891}
{"id": "e0913764fd6842cbe6b59e431da5b029f0b4af841b3d7e4273535aa1a73727b5", "sources": ["arxiv", "semantic_scholar"], "title": "LoRC: Low-Rank Compression for LLMs KV Cache with a Progressive Compression Strategy", "abstract": "The Key-Value (KV) cache is a crucial component in serving transformer-based autoregressive large language models (LLMs), enabling faster inference by storing previously computed KV vectors. However, its memory consumption scales linearly with sequence length and batch size, posing a significant bottleneck in LLM deployment. Existing approaches to mitigate this issue include: (1) efficient attention variants integrated in upcycling stages, which requires extensive parameter tuning thus unsuitable for pre-trained LLMs; (2) KV cache compression at test time, primarily through token eviction policies, which often overlook inter-layer dependencies and can be task-specific. This paper introduces an orthogonal approach to KV cache compression. We propose a low-rank approximation of KV weight matrices, allowing for plug-in integration with existing transformer-based LLMs without model retraining. To effectively compress KV cache at the weight level, we adjust for layerwise sensitivity and introduce a progressive compression strategy, which is supported by our theoretical analysis on how compression errors accumulate in deep networks. Our method is designed to function without model tuning in upcycling stages or task-specific profiling in test stages. Extensive experiments with LLaMA models ranging from 8B to 70B parameters across various tasks show that our approach significantly reduces the GPU memory footprint while maintaining performance.", "authors": ["Rongzhi Zhang", "Kuang Wang", "Liyuan Liu", "Shuohang Wang", "Hao Cheng", "Chao Zhang", "Yelong Shen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-04", "url": "https://arxiv.org/abs/2410.03111", "pdf_url": "https://arxiv.org/pdf/2410.03111v1", "arxiv_id": "2410.03111", "doi": "10.48550/arXiv.2410.03111", "citation_count": 43, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4109}
{"id": "e8fe853bf9a8472b2756eeeab5192dd168a14b87545d4988b6b453a9cdff04da", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba in Vision: A Comprehensive Survey of Techniques and Applications", "abstract": "Mamba is emerging as a novel approach to overcome the challenges faced by Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) in computer vision. While CNNs excel at extracting local features, they often struggle to capture long-range dependencies without complex architectural modifications. In contrast, ViTs effectively model global relationships but suffer from high computational costs due to the quadratic complexity of their self-attention mechanisms. Mamba addresses these limitations by leveraging Selective Structured State Space Models to effectively capture long-range dependencies with linear computational complexity. This survey analyzes the unique contributions, computational benefits, and applications of Mamba models while also identifying challenges and potential future research directions. We provide a foundational resource for advancing the understanding and growth of Mamba models in computer vision. An overview of this work is available at https://github.com/maklachur/Mamba-in-Computer-Vision.", "authors": ["Md Maklachur Rahman", "Abdullah Aman Tutul", "Ankur Nath", "Lamyanba Laishram", "Soon Ki Jung", "Tracy Hammond"], "categories": ["cs.CV", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-04", "url": "https://arxiv.org/abs/2410.03105", "pdf_url": "https://arxiv.org/pdf/2410.03105v1", "arxiv_id": "2410.03105", "doi": "10.48550/arXiv.2410.03105", "citation_count": 45, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/maklachur/Mamba-in-Computer-Vision", "venue": "arXiv.org", "quality_score": 0.4157}
{"id": "ec8ab8a3f059338f223048185166a5255adb1941d3ebc86582b732104686bba2", "sources": ["arxiv", "semantic_scholar"], "title": "MIGA: Mixture-of-Experts with Group Aggregation for Stock Market Prediction", "abstract": "Stock market prediction has remained an extremely challenging problem for many decades owing to its inherent high volatility and low information noisy ratio. Existing solutions based on machine learning or deep learning demonstrate superior performance by employing a single model trained on the entire stock dataset to generate predictions across all types of stocks. However, due to the significant variations in stock styles and market trends, a single end-to-end model struggles to fully capture the differences in these stylized stock features, leading to relatively inaccurate predictions for all types of stocks. In this paper, we present MIGA, a novel Mixture of Expert with Group Aggregation framework designed to generate specialized predictions for stocks with different styles by dynamically switching between distinct style experts. To promote collaboration among different experts in MIGA, we propose a novel inner group attention architecture, enabling experts within the same group to share information and thereby enhancing the overall performance of all experts. As a result, MIGA significantly outperforms other end-to-end models on three Chinese Stock Index benchmarks including CSI300, CSI500, and CSI1000. Notably, MIGA-Conv reaches 24 % excess annual return on CSI300 benchmark, surpassing the previous state-of-the-art model by 8% absolute. Furthermore, we conduct a comprehensive analysis of mixture of experts for stock market prediction, providing valuable insights for future research.", "authors": ["Zhaojian Yu", "Yinghao Wu", "Genesis Wang", "Heming Weng"], "categories": ["cs.CE"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-03", "url": "https://arxiv.org/abs/2410.02241", "pdf_url": "https://arxiv.org/pdf/2410.02241v1", "arxiv_id": "2410.02241", "doi": "10.48550/arXiv.2410.02241", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "016a08cf4eef8aa04e72e018f46ba466764e33e5ceb900a449319a03b90b2959", "sources": ["arxiv", "semantic_scholar"], "title": "On Expert Estimation in Hierarchical Mixture of Experts: Beyond Softmax Gating Functions", "abstract": "With the growing prominence of the Mixture of Experts (MoE) architecture in developing large-scale foundation models, we investigate the Hierarchical Mixture of Experts (HMoE), a specialized variant of MoE that excels in handling complex inputs and improving performance on targeted tasks. Our analysis highlights the advantages of using the Laplace gating function over the traditional Softmax gating within the HMoE frameworks. We theoretically demonstrate that applying the Laplace gating function at both levels of the HMoE model helps eliminate undesirable parameter interactions caused by the Softmax gating and, therefore, accelerates the expert convergence as well as enhances the expert specialization. Empirical validation across diverse scenarios supports these theoretical claims. This includes large-scale multimodal tasks, image classification, and latent domain discovery and prediction tasks, where our modified HMoE models show great performance improvements compared to the conventional HMoE models.", "authors": ["Huy Nguyen", "Xing Han", "Carl Harris", "Suchi Saria", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2024-10-03", "url": "https://arxiv.org/abs/2410.02935", "pdf_url": "https://arxiv.org/pdf/2410.02935v2", "arxiv_id": "2410.02935", "doi": "10.48550/arXiv.2410.02935", "citation_count": 15, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.301}
{"id": "e66beee41674e17bca8e8f5eb2175b4cef8c4368bbf5eed8d36178d731ea0a41", "sources": ["arxiv", "semantic_scholar"], "title": "SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration", "abstract": "The transformer architecture predominates across various models. As the heart of the transformer, attention has a computational complexity of $O(N^2)$, compared to $O(N)$ for linear transformations. When handling large sequence lengths, attention becomes the primary time-consuming component. Although quantization has proven to be an effective method for accelerating model inference, existing quantization methods primarily focus on optimizing the linear layer. In response, we first analyze the feasibility of quantization in attention detailedly. Following that, we propose SageAttention, a highly efficient and accurate quantization method for attention. The OPS (operations per second) of our approach outperforms FlashAttention2 and xformers by about 2.1 times and 2.7 times, respectively. SageAttention also achieves superior accuracy performance over FlashAttention3. Comprehensive experiments confirm that our approach incurs almost no end-to-end metrics loss across diverse models, including those for large language processing, image generation, and video generation. The codes are available at https://github.com/thu-ml/SageAttention.", "authors": ["Jintao Zhang", "Jia Wei", "Haofeng Huang", "Pengle Zhang", "Jun Zhu", "Jianfei Chen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-03", "url": "https://arxiv.org/abs/2410.02367", "pdf_url": "https://arxiv.org/pdf/2410.02367v9", "arxiv_id": "2410.02367", "doi": "10.48550/arXiv.2410.02367", "citation_count": 140, "influential_citation_count": 8, "has_code": true, "code_url": "https://github.com/thu-ml/SageAttention", "venue": "International Conference on Learning Representations", "quality_score": 0.5373}
{"id": "d86011f35441c13d021b3c6b4faa2b9e6805b609c3cc2a701e4d2ed645db61a3", "sources": ["arxiv", "semantic_scholar"], "title": "Upcycling Instruction Tuning from Dense to Mixture-of-Experts via Parameter Merging", "abstract": "Mixture-of-Experts (MoE) shines brightly in large language models (LLMs) and demonstrates outstanding performance in plentiful natural language processing tasks. However, existing methods transforming LLMs from dense to MoE face significant data requirements and typically rely on large-scale post-training. In this paper, we propose Upcycling Instruction Tuning (UpIT), a data-efficient approach for tuning a dense pre-trained model into a MoE instruction model. Specifically, we first point out that intermediate checkpoints during instruction tuning of the dense model are naturally suitable for specialized experts, and then propose an expert expansion stage to flexibly achieve models with flexible numbers of experts, where genetic algorithm and parameter merging are introduced to ensure sufficient diversity of new extended experts. To ensure that each specialized expert in the MoE model works as expected, we select a small amount of seed data that each expert excels to pre-optimize the router. Extensive experiments with various data scales and upcycling settings demonstrate the outstanding performance and data efficiency of UpIT, as well as stable improvement in expert or data scaling. Further analysis reveals the importance of ensuring expert diversity in upcycling.", "authors": ["Tingfeng Hui", "Zhenyu Zhang", "Shuohuan Wang", "Yu Sun", "Hua Wu", "Sen Su"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-02", "url": "https://arxiv.org/abs/2410.01610", "pdf_url": "https://arxiv.org/pdf/2410.01610v1", "arxiv_id": "2410.01610", "doi": "10.48550/arXiv.2410.01610", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1945}
{"id": "788f18cb805bbf97cc0e9c0bf7979f2c1d927e7ddbc34540f6f0c447a6ab9c3f", "sources": ["arxiv", "semantic_scholar"], "title": "LayerKV: Optimizing Large Language Model Serving with Layer-wise KV Cache Management", "abstract": "The expanding context windows in large language models (LLMs) have greatly enhanced their capabilities in various applications, but they also introduce significant challenges in maintaining low latency, particularly in Time to First Token (TTFT). This paper identifies that the sharp rise in TTFT as context length increases is predominantly driven by queuing delays, which are caused by the growing demands for GPU Key-Value (KV) cache allocation clashing with the limited availability of KV cache blocks. To address this issue, we propose LayerKV, a simple yet effective plug-in method that effectively reduces TTFT without requiring additional hardware or compromising output performance, while seamlessly integrating with existing parallelism strategies and scheduling techniques. Specifically, LayerKV introduces layer-wise KV block allocation, management, and offloading for fine-grained control over system memory, coupled with an SLO-aware scheduler to optimize overall Service Level Objectives (SLOs). Comprehensive evaluations on representative models, ranging from 7B to 70B parameters, across various GPU configurations, demonstrate that LayerKV improves TTFT latency up to 69x and reduces SLO violation rates by 28.7%, significantly enhancing the user experience.", "authors": ["Yi Xiong", "Hao Wu", "Changxu Shao", "Ziqing Wang", "Rui Zhang", "Yuhong Guo", "Junping Zhao", "Ke Zhang", "Zhenxuan Pan"], "categories": ["cs.DC", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-01", "url": "https://arxiv.org/abs/2410.00428", "pdf_url": "https://arxiv.org/pdf/2410.00428v3", "arxiv_id": "2410.00428", "doi": "10.48550/arXiv.2410.00428", "citation_count": 19, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3253}
{"id": "b8c6a06d974a82341affcfa86a7079fa9e6dac59742274538dae42ef316760e3", "sources": ["arxiv", "semantic_scholar"], "title": "KV-Compress: Paged KV-Cache Compression with Variable Compression Rates per Attention Head", "abstract": "Context lengths of Large Language Models (LLMs) have exploded in recent years, with 128k-token context becoming a standard and million-token context becoming a reality. Efficiently supporting long-context inference remains challenging as the memory that must be allocated in key-value (KV) cache for a generation scales with its context length, limiting the number of long-context requests that can be served concurrently under a given memory budget. KV cache compression can mitigate this issue by removing under-utilized KVs from each attention head's cache and reducing its memory footprint. Higher theoretical compression rates can be achieved when the number of removed KVs varies across attention heads, but application of such a strategy within existing inference frameworks adds fragmentation and cannot realize the theoretical compression rates in physical memory. We introduce KV-Compress, a novel compression method that evicts contiguous KV blocks within a PagedAttention framework, reducing the memory footprint of the KV cache proportionally to this theoretical compression rate. Our method achieves state-of-the-art performance on LongBench for both Mistral-7B-Instruct-v0.2 and Llama-3.1-8B-Instruct while lowering the total number of compressed KVs by 4x compared with prior methods. Evaluations on Llama-3.1-8B-Instruct and Llama-3.1-70B-Instruct-FP8 achieve compression rates up to 8x with negligible impact on performance, and up to 64x while retaining over 90% of full-cache performance for all but three of the suite's subsets. We benchmark an integration of our method with vLLM that increases total throughput by up to 5.18x by enabling larger decoding batches.", "authors": ["Isaac Rehg"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-30", "url": "https://arxiv.org/abs/2410.00161", "pdf_url": "https://arxiv.org/pdf/2410.00161v2", "arxiv_id": "2410.00161", "doi": "10.48550/arXiv.2410.00161", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "754ee48fb9d7dec5582e9c8127e5c069d03c0ca4c6413bf67697cb11fba0f813", "sources": ["arxiv", "semantic_scholar"], "title": "HDMoLE: Mixture of LoRA Experts with Hierarchical Routing and Dynamic Thresholds for Fine-Tuning LLM-based ASR Models", "abstract": "Recent advancements in integrating Large Language Models (LLM) with automatic speech recognition (ASR) have performed remarkably in general domains. While supervised fine-tuning (SFT) of all model parameters is often employed to adapt pre-trained LLM-based ASR models to specific domains, it imposes high computational costs and notably reduces their performance in general domains. In this paper, we propose a novel parameter-efficient multi-domain fine-tuning method for adapting pre-trained LLM-based ASR models to multi-accent domains without catastrophic forgetting named \\textit{HDMoLE}, which leverages hierarchical routing and dynamic thresholds based on combining low-rank adaptation (LoRA) with the mixer of experts (MoE) and can be generalized to any linear layer. Hierarchical routing establishes a clear correspondence between LoRA experts and accent domains, improving cross-domain collaboration among the LoRA experts. Unlike the static Top-K strategy for activating LoRA experts, dynamic thresholds can adaptively activate varying numbers of LoRA experts at each MoE layer. Experiments on the multi-accent and standard Mandarin datasets demonstrate the efficacy of HDMoLE. Applying HDMoLE to an LLM-based ASR model projector module achieves similar performance to full fine-tuning in the target multi-accent domains while using only 9.6% of the trainable parameters required for full fine-tuning and minimal degradation in the source general domain.", "authors": ["Bingshen Mu", "Kun Wei", "Qijie Shao", "Yong Xu", "Lei Xie"], "categories": ["cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-09-30", "url": "https://arxiv.org/abs/2409.19878", "pdf_url": "https://arxiv.org/pdf/2409.19878v3", "arxiv_id": "2409.19878", "doi": "10.1109/ICASSP49660.2025.10888133", "citation_count": 20, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.3306}
{"id": "eff464272f3186094dcadcd67402561f839ab8ed03871b6ad424dab42f258fcb", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Attentional Mixture of LoRAs for Language Model Continual Learning", "abstract": "Fine-tuning large language models (LLMs) with Low-Rank adaption (LoRA) is widely acknowledged as an effective approach for continual learning for new tasks. However, it often suffers from catastrophic forgetting when dealing with multiple tasks sequentially. To this end, we propose Attentional Mixture of LoRAs (AM-LoRA), a continual learning approach tailored for LLMs. Specifically, AM-LoRA learns a sequence of LoRAs for a series of tasks to continually learn knowledge from different tasks. The key of our approach is that we devise an attention mechanism as a knowledge mixture module to adaptively integrate information from each LoRA. With the attention mechanism, AM-LoRA can efficiently leverage the distinctive contributions of each LoRA, while mitigating the risk of mutually negative interactions among them that may lead to catastrophic forgetting. Moreover, we further introduce $L1$ norm in the learning process to make the attention vector more sparse. The sparse constraints can enable the model to lean towards selecting a few highly relevant LoRAs, rather than aggregating and weighting all LoRAs collectively, which can further reduce the impact stemming from mutual interference. Experimental results on continual learning benchmarks indicate the superiority of our proposed method.", "authors": ["Jialin Liu", "Jianhua Wu", "Jie Liu", "Yutai Duan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-29", "url": "https://arxiv.org/abs/2409.19611", "pdf_url": "https://arxiv.org/pdf/2409.19611v1", "arxiv_id": "2409.19611", "doi": "10.48550/arXiv.2409.19611", "citation_count": 13, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2865}
{"id": "1cbdbddeb9c2f8f80136a838fc1c6112d87ecf99ebae8feb1a921c30b2b7fba5", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Multicenter Experts in Multimodal AI for Debiased Radiotherapy Target Delineation", "abstract": "Clinical decision-making reflects diverse strategies shaped by regional patient populations and institutional protocols. However, most existing medical artificial intelligence (AI) models are trained on highly prevalent data patterns, which reinforces biases and fails to capture the breadth of clinical expertise. Inspired by the recent advances in Mixture of Experts (MoE), we propose a Mixture of Multicenter Experts (MoME) framework to address AI bias in the medical domain without requiring data sharing across institutions. MoME integrates specialized expertise from diverse clinical strategies to enhance model generalizability and adaptability across medical centers. We validate this framework using a multimodal target volume delineation model for prostate cancer radiotherapy. With few-shot training that combines imaging and clinical notes from each center, the model outperformed baselines, particularly in settings with high inter-center variability or limited data availability. Furthermore, MoME enables model customization to local clinical preferences without cross-institutional data exchange, making it especially suitable for resource-constrained settings while promoting broadly generalizable medical AI.", "authors": ["Yujin Oh", "Sangjoon Park", "Xiang Li", "Pengfei Jin", "Yi Wang", "Jonathan Paly", "Jason Efstathiou", "Annie Chan", "Jun Won Kim", "Hwa Kyung Byun", "Ik Jae Lee", "Jaeho Cho", "Chan Woo Wee", "Peng Shu", "Peilong Wang", "Nathan Yu", "Jason Holmes", "Jong Chul Ye", "Quanzheng Li", "Wei Liu", "Woong Sub Koom", "Jin Sung Kim", "Kyungsang Kim"], "categories": ["eess.IV", "cs.CV", "cs.LG"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2024-09-27", "url": "https://arxiv.org/abs/2410.00046", "pdf_url": "https://arxiv.org/pdf/2410.00046v3", "arxiv_id": "2410.00046", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "83c8a040d141fbbebdeb87e286f3ad58e2c3972f50e33a994323041a00b26262", "sources": ["arxiv", "semantic_scholar"], "title": "Benign Overfitting in Token Selection of Attention Mechanism", "abstract": "Attention mechanism is a fundamental component of the transformer model and plays a significant role in its success. However, the theoretical understanding of how attention learns to select tokens is still an emerging area of research. In this work, we study the training dynamics and generalization ability of the attention mechanism under classification problems with label noise. We show that, with the characterization of signal-to-noise ratio (SNR), the token selection of attention mechanism achieves benign overfitting, i.e., maintaining high generalization performance despite fitting label noise. Our work also demonstrates an interesting delayed acquisition of generalization after an initial phase of overfitting. Finally, we provide experiments to support our theoretical analysis using both synthetic and real-world datasets.", "authors": ["Keitaro Sakamoto", "Issei Sato"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-26", "url": "https://arxiv.org/abs/2409.17625", "pdf_url": "https://arxiv.org/pdf/2409.17625v3", "arxiv_id": "2409.17625", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/keitaroskmt/benign-attention", "venue": null, "quality_score": 0.0753}
{"id": "6c3b5141f3e3270a7c385e7bb511574d3c79162afe5a4e6c1ee94e3bbfce1bcc", "sources": ["arxiv", "semantic_scholar"], "title": "A Time Series is Worth Five Experts: Heterogeneous Mixture of Experts for Traffic Flow Prediction", "abstract": "Accurate traffic prediction faces significant challenges, necessitating a deep understanding of both temporal and spatial cues and their complex interactions across multiple variables. Recent advancements in traffic prediction systems are primarily due to the development of complex sequence-centric models. However, existing approaches often embed multiple variables and spatial relationships at each time step, which may hinder effective variable-centric learning, ultimately leading to performance degradation in traditional traffic prediction tasks. To overcome these limitations, we introduce variable-centric and prior knowledge-centric modeling techniques. Specifically, we propose a Heterogeneous Mixture of Experts (TITAN) model for traffic flow prediction. TITAN initially consists of three experts focused on sequence-centric modeling. Then, designed a low-rank adaptive method, TITAN simultaneously enables variable-centric modeling. Furthermore, we supervise the gating process using a prior knowledge-centric modeling strategy to ensure accurate routing. Experiments on two public traffic network datasets, METR-LA and PEMS-BAY, demonstrate that TITAN effectively captures variable-centric dependencies while ensuring accurate routing. Consequently, it achieves improvements in all evaluation metrics, ranging from approximately 4.37\\% to 11.53\\%, compared to previous state-of-the-art (SOTA) models. The code is open at \\href{https://github.com/sqlcow/TITAN}{https://github.com/sqlcow/TITAN}.", "authors": ["Guangyu Wang", "Yujie Chen", "Ming Gao", "Zhiqiao Wu", "Jiafu Tang", "Jiabi Zhao"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-26", "url": "https://arxiv.org/abs/2409.17440", "pdf_url": "https://arxiv.org/pdf/2409.17440v1", "arxiv_id": "2409.17440", "doi": "10.48550/arXiv.2409.17440", "citation_count": 9, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/sqlcow/TITAN}{https://github.com/sqlcow/TITAN}", "venue": "arXiv.org", "quality_score": 0.25}
{"id": "e79aadc64132569955c1eed813baff57881ab487d9ec6bc52691cfbefc6de88f", "sources": ["arxiv", "semantic_scholar"], "title": "AlignedKV: Reducing Memory Access of KV-Cache with Precision-Aligned Quantization", "abstract": "Model quantization has become a crucial technique to address the issues of large memory consumption and long inference times associated with LLMs. Mixed-precision quantization, which distinguishes between important and unimportant parameters, stands out among numerous quantization schemes as it achieves a balance between precision and compression rate. However, existing approaches can only identify important parameters through qualitative analysis and manual experiments without quantitatively analyzing how their importance is determined. We propose a new criterion, so-called 'precision alignment', to build a quantitative framework to holistically evaluate the importance of parameters in mixed-precision quantization. Our observations on floating point addition under various real-world scenarios suggest that two addends should have identical precision, otherwise the information in the higher-precision number will be wasted. Such an observation offers an essential principle to determine the precision of each parameter in matrix multiplication operation. As the first step towards applying the above discovery to large model inference, we develop a dynamic KV-Cache quantization technique to effectively reduce memory access latency. Different from existing quantization approaches that focus on memory saving, this work directly aims to accelerate LLM inference through quantifying floating numbers. The proposed technique attains a 25% saving of memory access and delivers up to 1.3x speedup in the computation of attention in the decoding phase of LLM, with almost no loss of precision.", "authors": ["Yifan Tan", "Haoze Wang", "Chao Yan", "Yangdong Deng"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-25", "url": "https://arxiv.org/abs/2409.16546", "pdf_url": "https://arxiv.org/pdf/2409.16546v2", "arxiv_id": "2409.16546", "doi": "10.48550/arXiv.2409.16546", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "1260691c32814d655c417f8e2d16f5b45e2bf5d01e31a8dfdc8e0da4525b3378", "sources": ["arxiv", "semantic_scholar"], "title": "INT-FlashAttention: Enabling Flash Attention for INT8 Quantization", "abstract": "As the foundation of large language models (LLMs), self-attention module faces the challenge of quadratic time and memory complexity with respect to sequence length. FlashAttention accelerates attention computation and reduces its memory usage by leveraging the GPU memory hierarchy. A promising research direction is to integrate FlashAttention with quantization methods. This paper introduces INT-FlashAttention, the first INT8 quantization architecture compatible with the forward workflow of FlashAttention, which significantly improves the inference speed of FlashAttention on Ampere GPUs. We implement our INT-FlashAttention prototype with fully INT8 activations and general matrix-multiplication (GEMM) kernels, making it the first attention operator with fully INT8 input. As a general token-level post-training quantization framework, INT-FlashAttention is also compatible with other data formats like INT4, etc. Experimental results show INT-FlashAttention achieves 72% faster inference speed and 82% smaller quantization error compared to standard FlashAttention with FP16 and FP8 data format.", "authors": ["Shimao Chen", "Zirui Liu", "Zhiying Wu", "Ce Zheng", "Peizhuang Cong", "Zihan Jiang", "Yuhan Wu", "Lei Su", "Tong Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-25", "url": "https://arxiv.org/abs/2409.16997", "pdf_url": "https://arxiv.org/pdf/2409.16997v2", "arxiv_id": "2409.16997", "doi": "10.48550/arXiv.2409.16997", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "7b873e037165b3735415ac4ff10756758f6f17d3875724a72a5891ce21d856e2", "sources": ["arxiv", "semantic_scholar"], "title": "Boosting Code-Switching ASR with Mixture of Experts Enhanced Speech-Conditioned LLM", "abstract": "In this paper, we introduce a speech-conditioned Large Language Model (LLM) integrated with a Mixture of Experts (MoE) based connector to address the challenge of Code-Switching (CS) in Automatic Speech Recognition (ASR). Specifically, we propose an Insertion and Deletion of Interruption Token (IDIT) mechanism for better transfer text generation ability of LLM to speech recognition task. We also present a connecter with MoE architecture that manages multiple languages efficiently. To further enhance the collaboration of multiple experts and leverage the understanding capabilities of LLM, we propose a two-stage progressive training strategy: 1) The connector is unfrozen and trained with language-specialized experts to map speech representations to the text space. 2) The connector and LLM LoRA adaptor are trained with the proposed IDIT mechanism and all experts are activated to learn general representations. Experimental results demonstrate that our method significantly outperforms state-of-the-art models, including end-to-end and large-scale audio-language models.", "authors": ["Fengrun Zhang", "Wang Geng", "Hukai Huang", "Yahui Shan", "Cheng Yi", "He Qu"], "categories": ["cs.SD", "cs.AI", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-09-24", "url": "https://arxiv.org/abs/2409.15905", "pdf_url": "https://arxiv.org/pdf/2409.15905v2", "arxiv_id": "2409.15905", "doi": "10.1109/ICASSP49660.2025.10890030", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.2785}
{"id": "3be7478021ff395fac3fea608bd8b9fc878345cc32cd659d651454df86f4944b", "sources": ["arxiv", "semantic_scholar"], "title": "Leveraging Mixture of Experts for Improved Speech Deepfake Detection", "abstract": "Speech deepfakes pose a significant threat to personal security and content authenticity. Several detectors have been proposed in the literature, and one of the primary challenges these systems have to face is the generalization over unseen data to identify fake signals across a wide range of datasets. In this paper, we introduce a novel approach for enhancing speech deepfake detection performance using a Mixture of Experts architecture. The Mixture of Experts framework is well-suited for the speech deepfake detection task due to its ability to specialize in different input types and handle data variability efficiently. This approach offers superior generalization and adaptability to unseen data compared to traditional single models or ensemble methods. Additionally, its modular structure supports scalable updates, making it more flexible in managing the evolving complexity of deepfake techniques while maintaining high detection accuracy. We propose an efficient, lightweight gating mechanism to dynamically assign expert weights for each input, optimizing detection performance. Experimental results across multiple datasets demonstrate the effectiveness and potential of our proposed approach.", "authors": ["Viola Negroni", "Davide Salvi", "Alessandro Ilic Mezza", "Paolo Bestagini", "Stefano Tubaro"], "categories": ["cs.SD", "cs.AI", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-09-24", "url": "https://arxiv.org/abs/2409.16077", "pdf_url": "https://arxiv.org/pdf/2409.16077v1", "arxiv_id": "2409.16077", "doi": "10.1109/ICASSP49660.2025.10890398", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.3076}
{"id": "29a85013151f4be851e2e63e5f66577cbd9c247fa19114a515f0081a620280a1", "sources": ["arxiv", "semantic_scholar"], "title": "Merging LoRAs like Playing LEGO: Pushing the Modularity of LoRA to Extremes Through Rank-Wise Clustering", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a popular technique for fine-tuning large language models (LLMs) to various domains due to its modular design and widespread availability on platforms like Huggingface. This modularity has sparked interest in combining multiple LoRAs to enhance LLM capabilities. However, existing methods for LoRA composition primarily focus on task-specific adaptations that require additional training, and current model merging techniques often fail to fully leverage LoRA's modular nature, leading to parameter interference and performance degradation. In this paper, we investigate the feasibility of disassembling and reassembling multiple LoRAs at a finer granularity, analogous to assembling LEGO blocks. We introduce the concept of Minimal Semantic Units (MSUs), where the parameters corresponding to each rank in LoRA function as independent units. These MSUs demonstrate permutation invariance and concatenation-summation equivalence properties, enabling flexible combinations to create new LoRAs. Building on these insights, we propose the LoRA-LEGO framework. This framework conducts rank-wise parameter clustering by grouping MSUs from different LoRAs into $k$ clusters. The centroid of each cluster serves as a representative MSU, enabling the assembly of a merged LoRA with an adjusted rank of $k$. Additionally, we apply a dual reweighting strategy to optimize the scale of the merged LoRA. Experiments across various benchmarks demonstrate that our method outperforms existing approaches in LoRA merging.", "authors": ["Ziyu Zhao", "Tao Shen", "Didi Zhu", "Zexi Li", "Jing Su", "Xuwu Wang", "Kun Kuang", "Fei Wu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-24", "url": "https://arxiv.org/abs/2409.16167", "pdf_url": "https://arxiv.org/pdf/2409.16167v3", "arxiv_id": "2409.16167", "doi": "10.48550/arXiv.2409.16167", "citation_count": 48, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4225}
{"id": "d1b7932d9bf0591b22211865839c1a20170353aa32a4c709fd20d96bc308d142", "sources": ["arxiv", "semantic_scholar"], "title": "Efficiently Dispatching Flash Attention For Partially Filled Attention Masks", "abstract": "Transformers are widely used across various applications, many of which yield sparse or partially filled attention matrices. Examples include attention masks designed to reduce the quadratic complexity of attention, sequence packing techniques, and recent innovations like tree masking for fast validation in MEDUSA. Despite the inherent sparsity in these matrices, the state-of-the-art algorithm Flash Attention still processes them with quadratic complexity as though they were dense. In this paper, we introduce Binary Block Masking, a highly efficient modification that enhances Flash Attention by making it mask-aware. We further propose two optimizations: one tailored for masks with contiguous non-zero patterns and another for extremely sparse masks. Our experiments on attention masks derived from real-world scenarios demonstrate up to a 9x runtime improvement. The implementation will be publicly released to foster further research and application.", "authors": ["Agniv Sharma", "Jonas Geiping"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-23", "url": "https://arxiv.org/abs/2409.15097", "pdf_url": "https://arxiv.org/pdf/2409.15097v2", "arxiv_id": "2409.15097", "doi": "10.48550/arXiv.2409.15097", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "5ce0122c4880e0e972dfdd95deff55c13f0503cd468bb4dea9b1a7f16633b67d", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Efficient Diffusion Experts Through Automatic Interval and Sub-Network Selection", "abstract": "Diffusion probabilistic models can generate high-quality samples. Yet, their sampling process requires numerous denoising steps, making it slow and computationally intensive. We propose to reduce the sampling cost by pruning a pretrained diffusion model into a mixture of efficient experts. First, we study the similarities between pairs of denoising timesteps, observing a natural clustering, even across different datasets. This suggests that rather than having a single model for all time steps, separate models can serve as ``experts'' for their respective time intervals. As such, we separately fine-tune the pretrained model on each interval, with elastic dimensions in depth and width, to obtain experts specialized in their corresponding denoising interval. To optimize the resource usage between experts, we introduce our Expert Routing Agent, which learns to select a set of proper network configurations. By doing so, our method can allocate the computing budget between the experts in an end-to-end manner without requiring manual heuristics. Finally, with a selected configuration, we fine-tune our pruned experts to obtain our mixture of efficient experts. We demonstrate the effectiveness of our method, DiffPruning, across several datasets, LSUN-Church, LSUN-Beds, FFHQ, and ImageNet, on the Latent Diffusion Model architecture.", "authors": ["Alireza Ganjdanesh", "Yan Kang", "Yuchen Liu", "Richard Zhang", "Zhe Lin", "Heng Huang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-23", "url": "https://arxiv.org/abs/2409.15557", "pdf_url": "https://arxiv.org/pdf/2409.15557v1", "arxiv_id": "2409.15557", "doi": "10.48550/arXiv.2409.15557", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "European Conference on Computer Vision", "quality_score": 0.294}
{"id": "1567fa15736ca7a01d3b3222394f788724b64594470a7ea765f3f5452d37a2f8", "sources": ["arxiv", "semantic_scholar"], "title": "A Gated Residual Kolmogorov-Arnold Networks for Mixtures of Experts", "abstract": "This paper introduces KAMoE, a novel Mixture of Experts (MoE) framework based on Gated Residual Kolmogorov-Arnold Networks (GRKAN). We propose GRKAN as an alternative to the traditional gating function, aiming to enhance efficiency and interpretability in MoE modeling. Through extensive experiments on digital asset markets and real estate valuation, we demonstrate that KAMoE consistently outperforms traditional MoE architectures across various tasks and model types. Our results show that GRKAN exhibits superior performance compared to standard Gating Residual Networks, particularly in LSTM-based models for sequential tasks. We also provide insights into the trade-offs between model complexity and performance gains in MoE and KAMoE architectures.", "authors": ["Hugo Inzirillo", "Remi Genet"], "categories": ["cs.LG", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-23", "url": "https://arxiv.org/abs/2409.15161", "pdf_url": "https://arxiv.org/pdf/2409.15161v2", "arxiv_id": "2409.15161", "doi": "10.48550/arXiv.2409.15161", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "859ff02ceb53b35c556875bca0a0ef19f627e8351114cf2a909127f03326056e", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Performance and Scalability of Large-Scale Recommendation Systems with Jagged Flash Attention", "abstract": "The integration of hardware accelerators has significantly advanced the capabilities of modern recommendation systems, enabling the exploration of complex ranking paradigms previously deemed impractical. However, the GPU-based computational costs present substantial challenges. In this paper, we demonstrate our development of an efficiency-driven approach to explore these paradigms, moving beyond traditional reliance on native PyTorch modules. We address the specific challenges posed by ranking models' dependence on categorical features, which vary in length and complicate GPU utilization. We introduce Jagged Feature Interaction Kernels, a novel method designed to extract fine-grained insights from long categorical features through efficient handling of dynamically sized tensors. We further enhance the performance of attention mechanisms by integrating Jagged tensors with Flash Attention. Our novel Jagged Flash Attention achieves up to 9x speedup and 22x memory reduction compared to dense attention. Notably, it also outperforms dense flash attention, with up to 3x speedup and 53% more memory efficiency. In production models, we observe 10% QPS improvement and 18% memory savings, enabling us to scale our recommendation systems with longer features and more complex architectures.", "authors": ["Rengan Xu", "Junjie Yang", "Yifan Xu", "Hong Li", "Xing Liu", "Devashish Shankar", "Haoci Zhang", "Meng Liu", "Boyang Li", "Yuxi Hu", "Mingwei Tang", "Zehua Zhang", "Tunhou Zhang", "Dai Li", "Sijia Chen", "Gian-Paolo Musumeci", "Jiaqi Zhai", "Bill Zhu", "Hong Yan", "Srihari Reddy"], "categories": ["cs.LG", "cs.AI", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-19", "url": "https://arxiv.org/abs/2409.15373", "pdf_url": "https://arxiv.org/pdf/2409.15373v1", "arxiv_id": "2409.15373", "doi": "10.1145/3640457.3688040", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ACM Conference on Recommender Systems", "quality_score": 0.1945}
{"id": "fe9b3af4afd924f9594fb075e2849ed4c128eb64747ee8b1d16c5019a494250c", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Diverse Size Experts", "abstract": "The Sparsely-Activated Mixture-of-Experts (MoE) has gained increasing popularity for scaling up large language models (LLMs) without exploding computational costs. Despite its success, the current design faces a challenge where all experts have the same size, limiting the ability of tokens to choose the experts with the most appropriate size for generating the next token. In this paper, we propose the Mixture of Diverse Size Experts (MoDSE), a new MoE architecture with layers designed to have experts of different sizes. Our analysis of difficult token generation tasks shows that experts of various sizes achieve better predictions, and the routing path of the experts tends to be stable after a training period. However, having experts of diverse sizes can lead to uneven workload distribution. To tackle this limitation, we introduce an expert-pair allocation strategy to evenly distribute the workload across multiple GPUs. Comprehensive evaluations across multiple benchmarks demonstrate the effectiveness of MoDSE, as it outperforms existing MoEs by allocating the parameter budget to experts adaptively while maintaining the same total parameter size and the number of experts.", "authors": ["Manxi Sun", "Wei Liu", "Jian Luan", "Pengzhi Gao", "Bin Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-18", "url": "https://arxiv.org/abs/2409.12210", "pdf_url": "https://arxiv.org/pdf/2409.12210v1", "arxiv_id": "2409.12210", "doi": "10.48550/arXiv.2409.12210", "citation_count": 5, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2386}
{"id": "391927aa8f8164c995d7cbf239eb037cb14912729b6d17d7f202b671d7aaa764", "sources": ["arxiv", "semantic_scholar"], "title": "LPT++: Efficient Training on Mixture of Long-tailed Experts", "abstract": "We introduce LPT++, a comprehensive framework for long-tailed classification that combines parameter-efficient fine-tuning (PEFT) with a learnable model ensemble. LPT++ enhances frozen Vision Transformers (ViTs) through the integration of three core components. The first is a universal long-tailed adaptation module, which aggregates long-tailed prompts and visual adapters to adapt the pretrained model to the target domain, meanwhile improving its discriminative ability. The second is the mixture of long-tailed experts framework with a mixture-of-experts (MoE) scorer, which adaptively calculates reweighting coefficients for confidence scores from both visual-only and visual-language (VL) model experts to generate more accurate predictions. Finally, LPT++ employs a three-phase training framework, wherein each critical module is learned separately, resulting in a stable and effective long-tailed classification training paradigm. Besides, we also propose the simple version of LPT++ namely LPT, which only integrates visual-only pretrained ViT and long-tailed prompts to formulate a single model method. LPT can clearly illustrate how long-tailed prompts works meanwhile achieving comparable performance without VL pretrained models. Experiments show that, with only ~1% extra trainable parameters, LPT++ achieves comparable accuracy against all the counterparts.", "authors": ["Bowen Dong", "Pan Zhou", "Wangmeng Zuo"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-17", "url": "https://arxiv.org/abs/2409.11323", "pdf_url": "https://arxiv.org/pdf/2409.11323v1", "arxiv_id": "2409.11323", "doi": "10.48550/arXiv.2409.11323", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "c515036bac91fb7dc543e7c188733efe80dc1aeee63f79ea1f5dd182c4f8087f", "sources": ["arxiv", "semantic_scholar"], "title": "KVPruner: Structural Pruning for Faster and Memory-Efficient Large Language Models", "abstract": "The bottleneck associated with the key-value(KV) cache presents a significant challenge during the inference processes of large language models. While depth pruning accelerates inference, it requires extensive recovery training, which can take up to two weeks. On the other hand, width pruning retains much of the performance but offers slight speed gains. To tackle these challenges, we propose KVPruner to improve model efficiency while maintaining performance. Our method uses global perplexity-based analysis to determine the importance ratio for each block and provides multiple strategies to prune non-essential KV channels within blocks. Compared to the original model, KVPruner reduces runtime memory usage by 50% and boosts throughput by over 35%. Additionally, our method requires only two hours of LoRA fine-tuning on small datasets to recover most of the performance.", "authors": ["Bo Lv", "Quan Zhou", "Xuanang Ding", "Yan Wang", "Zeming Ma"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-17", "url": "https://arxiv.org/abs/2409.11057", "pdf_url": "https://arxiv.org/pdf/2409.11057v1", "arxiv_id": "2409.11057", "doi": "10.1109/ICASSP49660.2025.10889000", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.2386}
{"id": "bbaebe4f0af0914b77dd4a3920144bbcba764edcbdb4f36b5d84c5c0bf136d37", "sources": ["arxiv", "semantic_scholar"], "title": "CSKV: Training-Efficient Channel Shrinking for KV Cache in Long-Context Scenarios", "abstract": "Large Language Models (LLMs) have been widely adopted to process long-context tasks. However, the large memory overhead of the key-value (KV) cache poses significant challenges in long-context scenarios. Existing training-free KV cache compression methods typically focus on quantization and token pruning, which have compression limits, and excessive sparsity can lead to severe performance degradation. Other methods design new architectures with less KV overhead but require significant training overhead. To address the above two drawbacks, we further explore the redundancy in the channel dimension and apply an architecture-level design with minor training costs. Therefore, we introduce CSKV, a training-efficient Channel Shrinking technique for KV cache compression: (1) We first analyze the singular value distribution of the KV cache, revealing significant redundancy and compression potential along the channel dimension. Based on this observation, we propose using low-rank decomposition for key and value layers and storing the low-dimension features. (2) To preserve model performance, we introduce a bi-branch KV cache, including a window-based full-precision KV cache and a low-precision compressed KV cache. (3) To reduce the training costs, we minimize the layer-wise reconstruction loss for the compressed KV cache instead of retraining the entire LLMs. Extensive experiments show that CSKV can reduce the memory overhead of the KV cache by 80% while maintaining the model's long-context capability. Moreover, we show that our method can be seamlessly combined with quantization to further reduce the memory overhead, achieving a compression ratio of up to 95%. Code is available at https://github.com/wln20/CSKV.", "authors": ["Luning Wang", "Shiyao Li", "Xuefei Ning", "Zhihang Yuan", "Shengen Yan", "Guohao Dai", "Yu Wang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-16", "url": "https://arxiv.org/abs/2409.10593", "pdf_url": "https://arxiv.org/pdf/2409.10593v3", "arxiv_id": "2409.10593", "doi": "10.48550/arXiv.2409.10593", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/wln20/CSKV", "venue": null, "quality_score": 0.1193}
{"id": "8efc0ff16ae766ca9d6ddf94a558b3e06839f30fb3a7a44c18e6f6ec459244e2", "sources": ["arxiv", "semantic_scholar"], "title": "Flash STU: Fast Spectral Transform Units", "abstract": "Recent advances in state-space model architectures have shown great promise for efficient sequence modeling, but challenges remain in balancing computational efficiency with model expressiveness. We propose the Flash STU architecture, a hybrid model that interleaves spectral state space model layers with sliding window attention, enabling scalability to billions of parameters for language modeling while maintaining a near-linear time complexity. We evaluate the Flash STU and its variants on diverse sequence prediction tasks, including linear dynamical systems, robotics control, and language modeling. We find that, given a fixed parameter budget, the Flash STU architecture consistently outperforms the Transformer and other leading state-space models such as S4 and Mamba-2.", "authors": ["Y. Isabel Liu", "Windsor Nguyen", "Yagiz Devre", "Evan Dogariu", "Anirudha Majumdar", "Elad Hazan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-16", "url": "https://arxiv.org/abs/2409.10489", "pdf_url": "https://arxiv.org/pdf/2409.10489v6", "arxiv_id": "2409.10489", "doi": "10.1109/CDC57313.2025.11312729", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Conference on Decision and Control", "quality_score": 0.1945}
{"id": "74f5d524a1422429e46920f2f0727a810a4ba39d5ac04d14bf1a0b74b2240b62", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba for Scalable and Efficient Personalized Recommendations", "abstract": "In this effort, we propose using the Mamba for handling tabular data in personalized recommendation systems. We present the \\textit{FT-Mamba} (Feature Tokenizer\\,$+$\\,Mamba), a novel hybrid model that replaces Transformer layers with Mamba layers within the FT-Transformer architecture, for handling tabular data in personalized recommendation systems. The \\textit{Mamba model} offers an efficient alternative to Transformers, reducing computational complexity from quadratic to linear by enhancing the capabilities of State Space Models (SSMs). FT-Mamba is designed to improve the scalability and efficiency of recommendation systems while maintaining performance. We evaluate FT-Mamba in comparison to a traditional Transformer-based model within a Two-Tower architecture on three datasets: Spotify music recommendation, H\\&M fashion recommendation, and vaccine messaging recommendation. Each model is trained on 160,000 user-action pairs, and performance is measured using precision (P), recall (R), Mean Reciprocal Rank (MRR), and Hit Ratio (HR) at several truncation values. Our results demonstrate that FT-Mamba outperforms the Transformer-based model in terms of computational efficiency while maintaining or exceeding performance across key recommendation metrics. By leveraging Mamba layers, FT-Mamba provides a scalable and effective solution for large-scale personalized recommendation systems, showcasing the potential of the Mamba architecture to enhance both efficiency and accuracy.", "authors": ["Andrew Starnes", "Clayton Webster"], "categories": ["cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-11", "url": "https://arxiv.org/abs/2409.17165", "pdf_url": "https://arxiv.org/pdf/2409.17165v1", "arxiv_id": "2409.17165", "doi": "10.1109/ICDMW65004.2024.00018", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "473ed090d07a54918ac90c943098284a2508c25aa0fe0589c44126c536e0380c", "sources": ["arxiv", "semantic_scholar"], "title": "DA-MoE: Towards Dynamic Expert Allocation for Mixture-of-Experts Models", "abstract": "Transformer-based Mixture-of-Experts (MoE) models have been driving several recent technological advancements in Natural Language Processing (NLP). These MoE models adopt a router mechanism to determine which experts to activate for routing input tokens. However, existing router mechanisms allocate a fixed number of experts to each token, which neglects the varying importance of different input tokens. In this study, we propose a novel dynamic router mechanism that Dynamically Allocates a variable number of experts for Mixture-of-Experts (DA-MoE) models based on an effective token importance measure. First, we show that the Transformer attention mechanism provides a natural and effective way of calculating token importance. Second, we propose a dynamic router mechanism that effectively decides the optimal number of experts (K) and allocates the top-K experts for each input token. Third, comprehensive experiments on several benchmark datasets demonstrate that our DA-MoE approach consistently outperforms the state-of-the-art Transformer based MoE model on the popular GLUE benchmark.", "authors": ["Maryam Akhavan Aghdam", "Hongpeng Jin", "Yanzhao Wu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-10", "url": "https://arxiv.org/abs/2409.06669", "pdf_url": "https://arxiv.org/pdf/2409.06669v1", "arxiv_id": "2409.06669", "doi": "10.48550/arXiv.2409.06669", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3076}
{"id": "2410db988c6e12c8ffbe18e44ee29050755d4c17f4c2e0ca3a13ed1c2040ec9f", "sources": ["arxiv", "semantic_scholar"], "title": "STUN: Structured-Then-Unstructured Pruning for Scalable MoE Pruning", "abstract": "Mixture-of-experts (MoEs) have been adopted for reducing inference costs by sparsely activating experts in Large language models (LLMs). Despite this reduction, the massive number of experts in MoEs still makes them expensive to serve. In this paper, we study how to address this, by pruning MoEs. Among pruning methodologies, unstructured pruning has been known to achieve the highest performance for a given pruning ratio, compared to structured pruning, since the latter imposes constraints on the sparsification structure. This is intuitive, as the solution space of unstructured pruning subsumes that of structured pruning. However, our counterintuitive finding reveals that expert pruning, a form of structured pruning, can actually precede unstructured pruning to outperform unstructured-only pruning. As existing expert pruning, requiring $O(\\frac{k^n}{\\sqrt{n}})$ forward passes for $n$ experts, cannot scale for recent MoEs, we propose a scalable alternative with $O(1)$ complexity, yet outperforming the more expensive methods. The key idea is leveraging a latent structure between experts, based on behavior similarity, such that the greedy decision of whether to prune closely captures the joint pruning effect. Ours is highly effective -- for Snowflake Arctic, a 480B-sized MoE with 128 experts, our method needs only one H100 and two hours to achieve nearly no loss in performance with 40% sparsity, even in generative tasks such as GSM8K, where state-of-the-art unstructured pruning fails to. The code will be made publicly available.", "authors": ["Jaeseong Lee", "seung-won hwang", "Aurick Qiao", "Daniel F Campos", "Zhewei Yao", "Yuxiong He"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-10", "url": "https://arxiv.org/abs/2409.06211", "pdf_url": "https://arxiv.org/pdf/2409.06211v2", "arxiv_id": "2409.06211", "doi": "10.48550/arXiv.2409.06211", "citation_count": 23, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3451}
{"id": "e8ce648218a49ba17f6ee10c58b85434d27958fd1354b7d43e8ea8ea5e6c50c4", "sources": ["arxiv", "semantic_scholar"], "title": "ChartMoE: Mixture of Diversely Aligned Expert Connector for Chart Understanding", "abstract": "Automatic chart understanding is crucial for content comprehension and document parsing. Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in chart understanding through domain-specific alignment and fine-tuning. However, current MLLMs still struggle to provide faithful data and reliable analysis only based on charts. To address it, we propose ChartMoE, which employs the Mixture of Expert (MoE) architecture to replace the traditional linear projector to bridge the modality gap. Specifically, we train several linear connectors through distinct alignment tasks, which are utilized as the foundational initialization parameters for different experts. Additionally, we introduce ChartMoE-Align, a dataset with nearly 1 million chart-table-JSON-code quadruples to conduct three alignment tasks (chart-table/JSON/code). Combined with the vanilla connector, we initialize different experts diversely and adopt high-quality knowledge learning to further refine the MoE connector and LLM parameters. Extensive experiments demonstrate the effectiveness of the MoE connector and our initialization strategy, e.g., ChartMoE improves the accuracy of the previous state-of-the-art from 80.48\\% to 84.64\\% on the ChartQA benchmark.", "authors": ["Zhengzhuo Xu", "Bowen Qu", "Yiyan Qi", "Sinan Du", "Chengjin Xu", "Chun Yuan", "Jian Guo"], "categories": ["cs.AI", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-05", "url": "https://arxiv.org/abs/2409.03277", "pdf_url": "https://arxiv.org/pdf/2409.03277v3", "arxiv_id": "2409.03277", "doi": null, "citation_count": 45, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4157}
{"id": "90fdd04167f4883e16e2aa25e4c1cd8a378c5f922b6ca76dac56cd31994678ac", "sources": ["arxiv", "semantic_scholar"], "title": "RouterRetriever: Routing over a Mixture of Expert Embedding Models", "abstract": "Information retrieval methods often rely on a single embedding model trained on large, general-domain datasets like MSMARCO. While this approach can produce a retriever with reasonable overall performance, they often underperform models trained on domain-specific data when testing on their respective domains. Prior work in information retrieval has tackled this through multi-task training, but the idea of routing over a mixture of domain-specific expert retrievers remains unexplored despite the popularity of such ideas in language model generation research. In this work, we introduce RouterRetriever, a retrieval model that leverages a mixture of domain-specific experts by using a routing mechanism to select the most appropriate expert for each query. RouterRetriever is lightweight and allows easy addition or removal of experts without additional training. Evaluation on the BEIR benchmark demonstrates that RouterRetriever outperforms both models trained on MSMARCO (+2.1 absolute nDCG@10) and multi-task models (+3.2). This is achieved by employing our routing mechanism, which surpasses other routing techniques (+1.8 on average) commonly used in language modeling. Furthermore, the benefit generalizes well to other datasets, even in the absence of a specific expert on the dataset. RouterRetriever is the first work to demonstrate the advantages of routing over a mixture of domain-specific expert embedding models as an alternative to a single, general-purpose embedding model, especially when retrieving from diverse, specialized domains.", "authors": ["Hyunji Lee", "Luca Soldaini", "Arman Cohan", "Minjoon Seo", "Kyle Lo"], "categories": ["cs.IR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-04", "url": "https://arxiv.org/abs/2409.02685", "pdf_url": "https://arxiv.org/pdf/2409.02685v2", "arxiv_id": "2409.02685", "doi": "10.1609/aaai.v39i11.33306", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2603}
{"id": "836ac44689f3e3e4ff686ad56af24f65814574fe418e4e59f21248b8b8d59d71", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Code-Switching Speech Recognition with LID-Based Collaborative Mixture of Experts Model", "abstract": "Due to the inherent difficulty in modeling phonetic similarities across different languages, code-switching speech recognition presents a formidable challenge. This study proposes a Collaborative-MoE, a Mixture of Experts (MoE) model that leverages a collaborative mechanism among expert groups. Initially, a preceding routing network explicitly learns Language Identification (LID) tasks and selects experts based on acquired LID weights. This process ensures robust routing information to the MoE layer, mitigating interference from diverse language domains on expert network parameter updates. The LID weights are also employed to facilitate inter-group collaboration, enabling the integration of language-specific representations. Furthermore, within each language expert group, a gating network operates unsupervised to foster collaboration on attributes beyond language. Extensive experiments demonstrate the efficacy of our approach, achieving significant performance enhancements compared to alternative methods. Importantly, our method preserves the efficient inference capabilities characteristic of MoE models without necessitating additional pre-training.", "authors": ["Hukai Huang", "Jiayan Lin", "Kaidi Wang", "Yishuang Li", "Wenhao Guan", "Lin Li", "Qingyang Hong"], "categories": ["cs.CL", "cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-09-03", "url": "https://arxiv.org/abs/2409.02050", "pdf_url": "https://arxiv.org/pdf/2409.02050v2", "arxiv_id": "2409.02050", "doi": "10.1109/SLT61566.2024.10832265", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Spoken Language Technology Workshop", "quality_score": 0.1505}
{"id": "6b6567695f703faddf7d2cbb05cb65ace5407815aa9ce3c3eee264cfbe5dd7d2", "sources": ["arxiv", "semantic_scholar"], "title": "Revisiting SMoE Language Models by Evaluating Inefficiencies with Task Specific Expert Pruning", "abstract": "Sparse Mixture of Expert (SMoE) models have emerged as a scalable alternative to dense models in language modeling. These models use conditionally activated feedforward subnetworks in transformer blocks, allowing for a separation between total model parameters and per-example computation. However, large token-routed SMoE models face a significant challenge: during inference, the entire model must be used for a sequence or a batch, resulting in high latencies in a distributed setting that offsets the advantages of per-token sparse activation. Our research explores task-specific model pruning to inform decisions about designing SMoE architectures, mainly modulating the choice of expert counts in pretraining. We investigate whether such pruned models offer advantages over smaller SMoE models trained from scratch, when evaluating and comparing them individually on tasks. To that end, we introduce an adaptive task-aware pruning technique UNCURL to reduce the number of experts per MoE layer in an offline manner post-training. Our findings reveal a threshold pruning factor for the reduction that depends on the number of experts used in pretraining, above which, the reduction starts to degrade model performance. These insights contribute to our understanding of model design choices when pretraining with SMoE architectures, particularly useful when considering task-specific inference optimization for later stages.", "authors": ["Soumajyoti Sarkar", "Leonard Lausen", "Volkan Cevher", "Sheng Zha", "Thomas Brox", "George Karypis"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-02", "url": "https://arxiv.org/abs/2409.01483", "pdf_url": "https://arxiv.org/pdf/2409.01483v1", "arxiv_id": "2409.01483", "doi": "10.48550/arXiv.2409.01483", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.25}
{"id": "6e5cae3a6548803001ac5e14a8ea6155c60669d5ee686ec1bd4194636c6b55e4", "sources": ["arxiv", "semantic_scholar"], "title": "Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batching", "abstract": "Large language models (LLMs) have emerged due to their capability to generate high-quality content across diverse contexts. To reduce their explosively increasing demands for computing resources, a mixture of experts (MoE) has emerged. The MoE layer enables exploiting a huge number of parameters with less computation. Applying state-of-the-art continuous batching increases throughput; however, it leads to frequent DRAM access in the MoE and attention layers. We observe that conventional computing devices have limitations when processing the MoE and attention layers, which dominate the total execution time and exhibit low arithmetic intensity (Op/B). Processing MoE layers only with devices targeting low-Op/B such as processing-in-memory (PIM) architectures is challenging due to the fluctuating Op/B in the MoE layer caused by continuous batching. To address these challenges, we propose Duplex, which comprises xPU tailored for high-Op/B and Logic-PIM to effectively perform low-Op/B operation within a single device. Duplex selects the most suitable processor based on the Op/B of each layer within LLMs. As the Op/B of the MoE layer is at least 1 and that of the attention layer has a value of 4-8 for grouped query attention, prior PIM architectures are not efficient, which place processing units inside DRAM dies and only target extremely low-Op/B (under one) operations. Based on recent trends, Logic-PIM adds more through-silicon vias (TSVs) to enable high-bandwidth communication between the DRAM die and the logic die and place powerful processing units on the logic die, which is best suited for handling low-Op/B operations ranging from few to a few dozens. To maximally utilize the xPU and Logic-PIM, we propose expert and attention co-processing.", "authors": ["Sungmin Yun", "Kwanhee Kyung", "Juhwan Cho", "Jaewan Choi", "Jongmin Kim", "Byeongho Kim", "Sukhan Lee", "Kyomin Sohn", "Jung Ho Ahn"], "categories": ["cs.AR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-02", "url": "https://arxiv.org/abs/2409.01141", "pdf_url": "https://arxiv.org/pdf/2409.01141v1", "arxiv_id": "2409.01141", "doi": "10.1109/MICRO61859.2024.00105", "citation_count": 55, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "Micro", "quality_score": 0.4771}
{"id": "ec7a0d00d0f2bef0e995c6ee17ef4456d5d3c348fc54c0174d56562b0bba4d64", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Parameter Count: Implicit Bias in Soft Mixture of Experts", "abstract": "The traditional viewpoint on Sparse Mixture of Experts (MoE) models is that instead of training a single large expert, which is computationally expensive, we can train many small experts. The hope is that if the total parameter count of the small experts equals that of the singular large expert, then we retain the representation power of the large expert while gaining computational tractability and promoting expert specialization. The recently introduced Soft MoE replaces the Sparse MoE's discrete routing mechanism with a differentiable gating function that smoothly mixes tokens. While this smooth gating function successfully mitigates the various training instabilities associated with Sparse MoE, it is unclear whether it induces implicit biases that affect Soft MoE's representation power or potential for expert specialization. We prove that Soft MoE with a single arbitrarily powerful expert cannot represent simple convex functions. This justifies that Soft MoE's success cannot be explained by the traditional viewpoint of many small experts collectively mimicking the representation power of a single large expert, and that multiple experts are actually necessary to achieve good representation power (even for a fixed total parameter count). Continuing along this line of investigation, we introduce a notion of expert specialization for Soft MoE, and while varying the number of experts yet fixing the total parameter count, we consider the following (computationally intractable) task. Given any input, how can we discover the expert subset that is specialized to predict this input's label? We empirically show that when there are many small experts, the architecture is implicitly biased in a fashion that allows us to efficiently approximate the specialized expert subset. Our method can be easily implemented to potentially reduce computation during inference.", "authors": ["Youngseog Chung", "Dhruv Malik", "Jeff Schneider", "Yuanzhi Li", "Aarti Singh"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-02", "url": "https://arxiv.org/abs/2409.00879", "pdf_url": "https://arxiv.org/pdf/2409.00879v1", "arxiv_id": "2409.00879", "doi": "10.48550/arXiv.2409.00879", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "a4bf5e6b20be096be6d110f609d6c145202206e16a0c5484b52943666a76727d", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Mamba: Introducing Controllability, Observability, And Stability To Structural State Space Models", "abstract": "Structured state space models' (SSMs) development in recent studies, such as Mamba and Mamba2, outperformed and solved the computational inefficiency of transformers and large language models at small to medium scale. In this work, we introduce the concept of controllability and observability to the original Mamba SSM's architecture in our Sparse-Mamba (S-Mamba) for natural language processing (NLP) applications. Moreover, we reinforce stability on the $nxn$ $A$ matrix on Mmaba2. The Mamba SSMs architecture drops the need for attention layers or multilayer perception blocks in transformers. However, current Mamba models lack reinforcement of controllability in state-space equations for computing the $A$, $B$, $C$, and $D$ matrices at each time step, leading to increased complexity and computational costs. Furthermore, the $A$ matrix in Mamba2 is not always stable. We demonstrate a reduction of parameters compared to the first published Mamba and Mamba2. We showcase an improvement in perplexity by 5\\% and a decrease in training time by 3\\% after reinforcing controllability and observability on the original Mamba architecture in our proposed S-Mamba. We further enforce stability on the $A$ matrix in Mamba2 to improve the loss and perplexity of the model. The controllable and stable $n \\times n$ state matrix $A$ is sparse, and it has only $n$ free parameters. Our novel approach will ensure controllable/observable and stable SSMs, which will be the gate key for Mamba3.", "authors": ["Emadeldeen Hamdan", "Hongyi Pan", "Ahmet Enis Cetin"], "categories": ["cs.LG", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-08-31", "url": "https://arxiv.org/abs/2409.00563", "pdf_url": "https://arxiv.org/pdf/2409.00563v3", "arxiv_id": "2409.00563", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "a7c930fc76fcc57268cce443d0ef3b8588b1ca0283dc2b33d623f588452d79b3", "sources": ["arxiv", "semantic_scholar"], "title": "Flexible and Effective Mixing of Large Language Models into a Mixture of Domain Experts", "abstract": "We present a toolkit for creating low-cost Mixture-of-Domain-Experts (MOE) from trained models. The toolkit can be used for creating a mixture from models or from adapters. We perform extensive tests and offer guidance on defining the architecture of the resulting MOE using the toolkit. A public repository is available.", "authors": ["Rhui Dih Lee", "Laura Wynter", "Raghu Kiran Ganti"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-30", "url": "https://arxiv.org/abs/2408.17280", "pdf_url": "https://arxiv.org/pdf/2408.17280v2", "arxiv_id": "2408.17280", "doi": "10.48550/arXiv.2408.17280", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "31ecf871871b035ceb62296316f0978d26b3f8ce3fae628bfc9fa6921db9821d", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba or Transformer for Time Series Forecasting? Mixture of Universals (MoU) Is All You Need", "abstract": "Time series forecasting requires balancing short-term and long-term dependencies for accurate predictions. Existing methods mainly focus on long-term dependency modeling, neglecting the complexities of short-term dynamics, which may hinder performance. Transformers are superior in modeling long-term dependencies but are criticized for their quadratic computational cost. Mamba provides a near-linear alternative but is reported less effective in time series longterm forecasting due to potential information loss. Current architectures fall short in offering both high efficiency and strong performance for long-term dependency modeling. To address these challenges, we introduce Mixture of Universals (MoU), a versatile model to capture both short-term and long-term dependencies for enhancing performance in time series forecasting. MoU is composed of two novel designs: Mixture of Feature Extractors (MoF), an adaptive method designed to improve time series patch representations for short-term dependency, and Mixture of Architectures (MoA), which hierarchically integrates Mamba, FeedForward, Convolution, and Self-Attention architectures in a specialized order to model long-term dependency from a hybrid perspective. The proposed approach achieves state-of-the-art performance while maintaining relatively low computational costs. Extensive experiments on seven real-world datasets demonstrate the superiority of MoU. Code is available at https://github.com/lunaaa95/mou/.", "authors": ["Sijia Peng", "Yun Xiong", "Yangyong Zhu", "Zhiqiang Shen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-28", "url": "https://arxiv.org/abs/2408.15997", "pdf_url": "https://arxiv.org/pdf/2408.15997v1", "arxiv_id": "2408.15997", "doi": "10.48550/arXiv.2408.15997", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/lunaaa95/mou/", "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "2791ad937298baa4174761d88aedde0a90573c1bc6953173c392b7039e61bb3c", "sources": ["arxiv", "semantic_scholar"], "title": "Nexus: Specialization meets Adaptability for Efficiently Training Mixture of Experts", "abstract": "Efficiency, specialization, and adaptability to new data distributions are qualities that are hard to combine in current Large Language Models. The Mixture of Experts (MoE) architecture has been the focus of significant research because its inherent conditional computation enables such desirable properties. In this work, we focus on \"upcycling\" dense expert models into an MoE, aiming to improve specialization while also adding the ability to adapt to new tasks easily. We introduce Nexus, an enhanced MoE architecture with adaptive routing where the model learns to project expert embeddings from domain representations. This approach allows Nexus to flexibly add new experts after the initial upcycling through separately trained dense models, without requiring large-scale MoE training for unseen data domains. Our experiments show that Nexus achieves a relative gain of up to 2.1% over the baseline for initial upcycling, and a 18.8% relative gain for extending the MoE with a new expert by using limited finetuning data. This flexibility of Nexus is crucial to enable an open-source ecosystem where every user continuously assembles their own MoE-mix according to their needs.", "authors": ["Nikolas Gritsch", "Qizhen Zhang", "Acyr Locatelli", "Sara Hooker", "Ahmet Üstün"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-28", "url": "https://arxiv.org/abs/2408.15901", "pdf_url": "https://arxiv.org/pdf/2408.15901v1", "arxiv_id": "2408.15901", "doi": "10.48550/arXiv.2408.15901", "citation_count": 8, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "1f22ba0ee53481efdf91eead6a612f1f7df5e12173747592725a9bcd8ca0b70b", "sources": ["arxiv", "semantic_scholar"], "title": "PAT: Pruning-Aware Tuning for Large Language Models", "abstract": "Large language models (LLMs) excel in language tasks, especially with supervised fine-tuning after pre-training. However, their substantial memory and computational requirements hinder practical applications. Structural pruning, which reduces less significant weight dimensions, is one solution. Yet, traditional post-hoc pruning often leads to significant performance loss, with limited recovery from further fine-tuning due to reduced capacity. Since the model fine-tuning refines the general and chaotic knowledge in pre-trained models, we aim to incorporate structural pruning with the fine-tuning, and propose the Pruning-Aware Tuning (PAT) paradigm to eliminate model redundancy while preserving the model performance to the maximum extend. Specifically, we insert the innovative Hybrid Sparsification Modules (HSMs) between the Attention and FFN components to accordingly sparsify the upstream and downstream linear modules. The HSM comprises a lightweight operator and a globally shared trainable mask. The lightweight operator maintains a training overhead comparable to that of LoRA, while the trainable mask unifies the channels to be sparsified, ensuring structural pruning. Additionally, we propose the Identity Loss which decouples the transformation and scaling properties of the HSMs to enhance training robustness. Extensive experiments demonstrate that PAT excels in both performance and efficiency. For example, our Llama2-7b model with a 25\\% pruning ratio achieves 1.33$\\times$ speedup while outperforming the LoRA-finetuned model by up to 1.26\\% in accuracy with a similar training cost. Code: https://github.com/kriskrisliu/PAT_Pruning-Aware-Tuning", "authors": ["Yijiang Liu", "Huanrui Yang", "Youxin Chen", "Rongyu Zhang", "Miao Wang", "Yuan Du", "Li Du"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-27", "url": "https://arxiv.org/abs/2408.14721", "pdf_url": "https://arxiv.org/pdf/2408.14721v2", "arxiv_id": "2408.14721", "doi": "10.48550/arXiv.2408.14721", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/kriskrisliu/PAT_Pruning-Aware-Tuning", "venue": "arXiv.org", "quality_score": 0.0}
{"id": "205dfe1411606cb893f513a06b5c6f8246220c60644d1d9ee538199e59a5b72c", "sources": ["arxiv", "semantic_scholar"], "title": "HMoE: Heterogeneous Mixture of Experts for Language Modeling", "abstract": "Mixture of Experts (MoE) offers remarkable performance and computational efficiency by selectively activating subsets of model parameters. Traditionally, MoE models use homogeneous experts, each with identical capacity. However, varying complexity in input data necessitates experts with diverse capabilities, while homogeneous MoE hinders effective expert specialization and efficient parameter utilization. In this study, we propose a novel Heterogeneous Mixture of Experts (HMoE), where experts differ in size and thus possess diverse capacities. This heterogeneity allows for more specialized experts to handle varying token complexities more effectively. To address the imbalance in expert activation, we propose a novel training objective that encourages the frequent activation of smaller experts, enhancing computational efficiency and parameter utilization. Extensive experiments demonstrate that HMoE achieves lower loss with fewer activated parameters and outperforms conventional homogeneous MoE models on various pre-training evaluation benchmarks. Codes will be released upon acceptance.", "authors": ["An Wang", "Xingwu Sun", "Ruobing Xie", "Shuaipeng Li", "Jiaqi Zhu", "Zhen Yang", "Pinxue Zhao", "J. N. Han", "Zhanhui Kang", "Di Wang", "Naoaki Okazaki", "Cheng-zhong Xu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-20", "url": "https://arxiv.org/abs/2408.10681", "pdf_url": "https://arxiv.org/pdf/2408.10681v1", "arxiv_id": "2408.10681", "doi": "10.48550/arXiv.2408.10681", "citation_count": 40, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4225}
{"id": "73989ea0ceff853659c0304ac96ea3e5027fb501227aee7cdd5986250934338a", "sources": ["arxiv", "semantic_scholar"], "title": "Customizing Language Models with Instance-wise LoRA for Sequential Recommendation", "abstract": "Sequential recommendation systems predict the next interaction item based on users' past interactions, aligning recommendations with individual preferences. Leveraging the strengths of Large Language Models (LLMs) in knowledge comprehension and reasoning, recent approaches are eager to apply LLMs to sequential recommendation. A common paradigm is converting user behavior sequences into instruction data, and fine-tuning the LLM with parameter-efficient fine-tuning (PEFT) methods like Low-Rank Adaption (LoRA). However, the uniform application of LoRA across diverse user behaviors is insufficient to capture individual variability, resulting in negative transfer between disparate sequences. To address these challenges, we propose Instance-wise LoRA (iLoRA). We innovatively treat the sequential recommendation task as a form of multi-task learning, integrating LoRA with the Mixture of Experts (MoE) framework. This approach encourages different experts to capture various aspects of user behavior. Additionally, we introduce a sequence representation guided gate function that generates customized expert participation weights for each user sequence, which allows dynamic parameter adjustment for instance-wise recommendations. In sequential recommendation, iLoRA achieves an average relative improvement of 11.4\\% over basic LoRA in the hit ratio metric, with less than a 1\\% relative increase in trainable parameters. Extensive experiments on three benchmark datasets demonstrate the effectiveness of iLoRA, highlighting its superior performance compared to existing methods in mitigating negative transfer and improving recommendation accuracy. Our data and code are available at https://github.com/AkaliKong/iLoRA.", "authors": ["Xiaoyu Kong", "Jiancan Wu", "An Zhang", "Leheng Sheng", "Hui Lin", "Xiang Wang", "Xiangnan He"], "categories": ["cs.IR", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-19", "url": "https://arxiv.org/abs/2408.10159", "pdf_url": "https://arxiv.org/pdf/2408.10159v4", "arxiv_id": "2408.10159", "doi": "10.48550/arXiv.2408.10159", "citation_count": 57, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/AkaliKong/iLoRA", "venue": "Neural Information Processing Systems", "quality_score": 0.4409}
{"id": "101f83a2eb36231686bbaff32d98d7d7780786295c41808561aaa288fa7888a1", "sources": ["arxiv", "semantic_scholar"], "title": "SMILE: Zero-Shot Sparse Mixture of Low-Rank Experts Construction From Pre-Trained Foundation Models", "abstract": "Deep model training on extensive datasets is increasingly becoming cost-prohibitive, prompting the widespread adoption of deep model fusion techniques to leverage knowledge from pre-existing models. From simple weight averaging to more sophisticated methods like AdaMerging, model fusion effectively improves model performance and accelerates the development of new models. However, potential interference between parameters of individual models and the lack of interpretability in the fusion progress remain significant challenges. Existing methods often try to resolve the parameter interference issue by evaluating attributes of parameters, such as their magnitude or sign, or by parameter pruning. In this study, we begin by examining the fine-tuning of linear layers through the lens of subspace analysis and explicitly define parameter interference as an optimization problem to shed light on this subject. Subsequently, we introduce an innovative approach to model fusion called zero-shot Sparse MIxture of Low-rank Experts (SMILE) construction, which allows for the upscaling of source models into an MoE model without extra data or further training. Our approach relies on the observation that fine-tuning mostly keeps the important parts from the pre-training, but it uses less significant or unused areas to adapt to new tasks. Also, the issue of parameter interference, which is intrinsically intractable in the original parameter space, can be managed by expanding the dimensions. We conduct extensive experiments across diverse scenarios, such as image classification and text generation tasks, using full fine-tuning and LoRA fine-tuning, and we apply our method to large language models (CLIP models, Flan-T5 models, and Mistral-7B models), highlighting the adaptability and scalability of SMILE. Code is available at https://github.com/tanganke/fusion_bench", "authors": ["Anke Tang", "Li Shen", "Yong Luo", "Shuai Xie", "Han Hu", "Lefei Zhang", "Bo Du", "Dacheng Tao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-19", "url": "https://arxiv.org/abs/2408.10174", "pdf_url": "https://arxiv.org/pdf/2408.10174v2", "arxiv_id": "2408.10174", "doi": "10.48550/arXiv.2408.10174", "citation_count": 11, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/tanganke/fusion_bench", "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "83676369a333933123b2439de17a56ea46f6c7d2dbffd85cf3295bd9782094de", "sources": ["arxiv", "semantic_scholar"], "title": "Galaxy cluster profiles: A Gaussian mixture model approach to halo miscentering", "abstract": "Measurements of the galaxy density and weak-lensing profiles of galaxy clusters typically rely on an assumed cluster center, which is taken to be the brightest cluster galaxy or other proxies for the true halo center. Departure of the assumed cluster center from the true halo center bias the resultant profile measurements, an effect known as miscentering bias. Currently, miscentering is typically modeled in stacked profiles of clusters with a two parameter model. We use an alternate approach in which the profiles of individual clusters are used with the corresponding likelihood computed using a Gaussian mixture model. We test the approach using halos and the corresponding subhalo profiles from the IllustrisTNG hydrodynamic simulations. We obtain significantly improved estimates of the miscentering parameters for both 3D and projected 2D profiles relevant for imaging surveys. We discuss applications to upcoming cosmological surveys.", "authors": ["Matthew Currie", "Kyle Miller", "Tae-Hyeon Shin", "Eric Baxter", "Bhuvnesh Jain"], "categories": ["astro-ph.IM", "astro-ph.CO"], "fields_of_study": ["Physics"], "published_date": "2024-08-19", "url": "https://arxiv.org/abs/2408.10371", "pdf_url": "https://arxiv.org/pdf/2408.10371v1", "arxiv_id": "2408.10371", "doi": "10.1088/1475-7516/2025/07/072", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/KyleMiller1/Halo-Miscentering-Mixture-Model", "venue": "Journal of Cosmology and Astroparticle Physics", "quality_score": 0.0753}
{"id": "eae208ebbb03a1f1982bba1a6d1f4773c2eb2242d707223e21e45c7749e86550", "sources": ["arxiv", "semantic_scholar"], "title": "BAM! Just Like That: Simple and Efficient Parameter Upcycling for Mixture of Experts", "abstract": "The Mixture of Experts (MoE) framework has become a popular architecture for large language models due to its superior performance over dense models. However, training MoEs from scratch in a large-scale regime is prohibitively expensive. Existing methods mitigate this by pre-training multiple dense expert models independently and using them to initialize an MoE. This is done by using experts' feed-forward network (FFN) to initialize the MoE's experts while merging other parameters. However, this method limits the reuse of dense model parameters to only the FFN layers, thereby constraining the advantages when \"upcycling\" these models into MoEs. We propose BAM (Branch-Attend-Mix), a simple yet effective method that addresses this shortcoming. BAM makes full use of specialized dense models by not only using their FFN to initialize the MoE layers but also leveraging experts' attention parameters fully by initializing them into a soft-variant of Mixture of Attention (MoA) layers. We explore two methods for upcycling attention parameters: 1) initializing separate attention experts from dense models including all attention parameters for the best model performance; and 2) sharing key and value parameters across all experts to facilitate for better inference efficiency. To further improve efficiency, we adopt a parallel attention transformer architecture to MoEs, which allows the attention experts and FFN experts to be computed concurrently. Our experiments on seed models ranging from 590 million to 2 billion parameters demonstrate that BAM surpasses baselines in both perplexity and downstream task performance, within the same computational and data constraints.", "authors": ["Qizhen Zhang", "Nikolas Gritsch", "Dwaraknath Gnaneshwar", "Simon Guo", "David Cairuz", "Bharat Venkitesh", "Jakob Foerster", "Phil Blunsom", "Sebastian Ruder", "Ahmet Ustun", "Acyr Locatelli"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-15", "url": "https://arxiv.org/abs/2408.08274", "pdf_url": "https://arxiv.org/pdf/2408.08274v2", "arxiv_id": "2408.08274", "doi": "10.48550/arXiv.2408.08274", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3076}
{"id": "8b8629ee72eb54e693ed080856b5215313e24db110dce9725a832332c4dd82a3", "sources": ["arxiv", "semantic_scholar"], "title": "PQV-Mobile: A Combined Pruning and Quantization Toolkit to Optimize Vision Transformers for Mobile Applications", "abstract": "While Vision Transformers (ViTs) are extremely effective at computer vision tasks and are replacing convolutional neural networks as the new state-of-the-art, they are complex and memory-intensive models. In order to effectively run these models on resource-constrained mobile/edge systems, there is a need to not only compress these models but also to optimize them and convert them into deployment-friendly formats. To this end, this paper presents a combined pruning and quantization tool, called PQV-Mobile, to optimize vision transformers for mobile applications. The tool is able to support different types of structured pruning based on magnitude importance, Taylor importance, and Hessian importance. It also supports quantization from FP32 to FP16 and int8, targeting different mobile hardware backends. We demonstrate the capabilities of our tool and show important latency-memory-accuracy trade-offs for different amounts of pruning and int8 quantization with Facebook Data Efficient Image Transformer (DeiT) models. Our results show that even pruning a DeiT model by 9.375% and quantizing it to int8 from FP32 followed by optimizing for mobile applications, we find a latency reduction by 7.18X with a small accuracy loss of 2.24%. The tool is open source.", "authors": ["Kshitij Bhardwaj"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-15", "url": "https://arxiv.org/abs/2408.08437", "pdf_url": "https://arxiv.org/pdf/2408.08437v1", "arxiv_id": "2408.08437", "doi": "10.48550/arXiv.2408.08437", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "1dfa030847098504d35ace529d464769a9de2ecc3b19c34bd38196e096bf672a", "sources": ["arxiv", "semantic_scholar"], "title": "FactorLLM: Factorizing Knowledge via Mixture of Experts for Large Language Models", "abstract": "Recent research has demonstrated that Feed-Forward Networks (FFNs) in Large Language Models (LLMs) play a pivotal role in storing diverse linguistic and factual knowledge. Conventional methods frequently face challenges due to knowledge confusion stemming from their monolithic and redundant architectures, which calls for more efficient solutions with minimal computational overhead, particularly for LLMs. In this paper, we explore the FFN computation paradigm in LLMs and introduce FactorLLM, a novel approach that decomposes well-trained dense FFNs into sparse sub-networks without requiring any further modifications, while maintaining the same level of performance. Furthermore, we embed a router from the Mixture-of-Experts (MoE), combined with our devised Prior-Approximate (PA) loss term that facilitates the dynamic activation of experts and knowledge adaptation, thereby accelerating computational processes and enhancing performance using minimal training data and fine-tuning steps. FactorLLM thus enables efficient knowledge factorization and activates select groups of experts specifically tailored to designated tasks, emulating the interactive functional segmentation of the human brain. Extensive experiments across various benchmarks demonstrate the effectiveness of our proposed FactorLLM which achieves comparable performance to the source model securing up to 85% model performance while obtaining over a 30% increase in inference speed. Code: https://github.com/zhenwuweihe/FactorLLM.", "authors": ["Zhongyu Zhao", "Menghang Dong", "Rongyu Zhang", "Wenzhao Zheng", "Yunpeng Zhang", "Huanrui Yang", "Dalong Du", "Kurt Keutzer", "Shanghang Zhang"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-15", "url": "https://arxiv.org/abs/2408.11855", "pdf_url": "https://arxiv.org/pdf/2408.11855v1", "arxiv_id": "2408.11855", "doi": "10.48550/arXiv.2408.11855", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/zhenwuweihe/FactorLLM", "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "cb1991065baec153171fbe923a0c3ffa79efdff2694056a4808889e20263df44", "sources": ["arxiv", "semantic_scholar"], "title": "Eigen Attention: Attention in Low-Rank Space for KV Cache Compression", "abstract": "Large language models (LLMs) represent a groundbreaking advancement in the domain of natural language processing due to their impressive reasoning abilities. Recently, there has been considerable interest in increasing the context lengths for these models to enhance their applicability to complex tasks. However, at long context lengths and large batch sizes, the key-value (KV) cache, which stores the attention keys and values, emerges as the new bottleneck in memory usage during inference. To address this, we propose Eigen Attention, which performs the attention operation in a low-rank space, thereby reducing the KV cache memory overhead. Our proposed approach is orthogonal to existing KV cache compression techniques and can be used synergistically with them. Through extensive experiments over OPT, MPT, and Llama model families, we demonstrate that Eigen Attention results in up to 40% reduction in KV cache sizes and up to 60% reduction in attention operation latency with minimal drop in performance. Code is available at https://github.com/UtkarshSaxena1/EigenAttn.", "authors": ["Utkarsh Saxena", "Gobinda Saha", "Sakshi Choudhary", "Kaushik Roy"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-10", "url": "https://arxiv.org/abs/2408.05646", "pdf_url": "https://arxiv.org/pdf/2408.05646v2", "arxiv_id": "2408.05646", "doi": "10.48550/arXiv.2408.05646", "citation_count": 55, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/UtkarshSaxena1/EigenAttn", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4515}
{"id": "b8b68f847c89737ff45066a16c6bea368339294b6e329c443e74da928cd4f503", "sources": ["arxiv", "semantic_scholar"], "title": "NACL: A General and Effective KV Cache Eviction Framework for LLMs at Inference Time", "abstract": "Large Language Models (LLMs) have ignited an innovative surge of AI applications, marking a new era of exciting possibilities equipped with extended context windows. However, hosting these models is cost-prohibitive mainly due to the extensive memory consumption of KV Cache involving long-context modeling. Despite several works proposing to evict unnecessary tokens from the KV Cache, most of them rely on the biased local statistics of accumulated attention scores and report performance using unconvincing metric like perplexity on inadequate short-text evaluation. In this paper, we propose NACL, a general framework for long-context KV cache eviction that achieves more optimal and efficient eviction in a single operation during the encoding phase. Due to NACL's efficiency, we combine more accurate attention score statistics in PROXY TOKENS EVICTION with the diversified random eviction strategy of RANDOM EVICTION, aiming to alleviate the issue of attention bias and enhance the robustness in maintaining pivotal tokens for long-context modeling tasks. Notably, our method significantly improves the performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 50% with over 95% performance maintenance. The code is available at https://github.com/PaddlePaddle/Research/tree/master/NLP/ACL2024-NACL.", "authors": ["Yilong Chen", "Guoxia Wang", "Junyuan Shang", "Shiyao Cui", "Zhenyu Zhang", "Tingwen Liu", "Shuohuan Wang", "Yu Sun", "Dianhai Yu", "Hua Wu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-07", "url": "https://arxiv.org/abs/2408.03675", "pdf_url": "https://arxiv.org/pdf/2408.03675v2", "arxiv_id": "2408.03675", "doi": "10.48550/arXiv.2408.03675", "citation_count": 40, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/PaddlePaddle/Research/tree/master/NLP/ACL2024-NACL", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4032}
{"id": "10e3cd1c7a515211dfe89673e056948a0ea0a917d8a5c6f748308ba90b4d702a", "sources": ["arxiv", "semantic_scholar"], "title": "Attention is all you need for an improved CNN-based flash flood susceptibility modeling. The case of the ungauged Rheraya watershed, Morocco", "abstract": "Effective flood hazard management requires evaluating and predicting flash flood susceptibility. Convolutional neural networks (CNNs) are commonly used for this task but face issues like gradient explosion and overfitting. This study explores the use of an attention mechanism, specifically the convolutional block attention module (CBAM), to enhance CNN models for flash flood susceptibility in the ungauged Rheraya watershed, a flood prone region. We used ResNet18, DenseNet121, and Xception as backbone architectures, integrating CBAM at different locations. Our dataset included 16 conditioning factors and 522 flash flood inventory points. Performance was evaluated using accuracy, precision, recall, F1-score, and the area under the curve (AUC) of the receiver operating characteristic (ROC). Results showed that CBAM significantly improved model performance, with DenseNet121 incorporating CBAM in each convolutional block achieving the best results (accuracy = 0.95, AUC = 0.98). Distance to river and drainage density were identified as key factors. These findings demonstrate the effectiveness of the attention mechanism in improving flash flood susceptibility modeling and offer valuable insights for disaster management.", "authors": ["Akram Elghouat", "Ahmed Algouti", "Abdellah Algouti", "Soukaina Baid"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-03", "url": "https://arxiv.org/abs/2408.02692", "pdf_url": "https://arxiv.org/pdf/2408.02692v1", "arxiv_id": "2408.02692", "doi": "10.48550/arXiv.2408.02692", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "aef280a4cec17a974d4aff68de21d3c7ae7890278662102e99afe3f79924e088", "sources": ["arxiv", "semantic_scholar"], "title": "MoDE: Effective Multi-task Parameter Efficient Fine-Tuning with a Mixture of Dyadic Experts", "abstract": "Parameter-efficient fine-tuning techniques like Low-Rank Adaptation (LoRA) have revolutionized the adaptation of large language models (LLMs) to diverse tasks. Recent efforts have explored mixtures of LoRA modules for multi-task settings. However, our analysis reveals redundancy in the down-projection matrices of these architectures. This observation motivates our proposed method, Mixture of Dyadic Experts (MoDE), which introduces a novel design for efficient multi-task adaptation. This is done by sharing the down-projection matrix across tasks and employing atomic rank-one adapters, coupled with routers that allow more sophisticated task-level specialization. Our design allows for more fine-grained mixing, thereby increasing the model's ability to jointly handle multiple tasks. We evaluate MoDE on the Supernatural Instructions (SNI) benchmark consisting of a diverse set of 700+ tasks and demonstrate that it outperforms state-of-the-art multi-task parameter-efficient fine-tuning (PEFT) methods, without introducing additional parameters. Our findings contribute to a deeper understanding of parameter efficiency in multi-task LLM adaptation and provide a practical solution for deploying high-performing, lightweight models.", "authors": ["Lin Ning", "Harsh Lara", "Meiqi Guo", "Abhinav Rastogi"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-02", "url": "https://arxiv.org/abs/2408.01505", "pdf_url": "https://arxiv.org/pdf/2408.01505v1", "arxiv_id": "2408.01505", "doi": "10.48550/arXiv.2408.01505", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.2113}
{"id": "2d304f1a313da1604999471aa52da2e9ec8a579c92406f719faeb29eb7ecdc20", "sources": ["arxiv", "semantic_scholar"], "title": "A Survey of Mamba", "abstract": "As one of the most representative DL techniques, Transformer architecture has empowered numerous advanced models, especially the large language models (LLMs) that comprise billions of parameters, becoming a cornerstone in deep learning. Despite the impressive achievements, Transformers still face inherent limitations, particularly the time-consuming inference resulting from the quadratic computation complexity of attention calculation. Recently, a novel architecture named Mamba, drawing inspiration from classical state space models (SSMs), has emerged as a promising alternative for building foundation models, delivering comparable modeling abilities to Transformers while preserving near-linear scalability concerning sequence length. This has sparked an increasing number of studies actively exploring Mamba's potential to achieve impressive performance across diverse domains. Given such rapid evolution, there is a critical need for a systematic review that consolidates existing Mamba-empowered models, offering a comprehensive understanding of this emerging model architecture. In this survey, we therefore conduct an in-depth investigation of recent Mamba-associated studies, covering three main aspects: the advancements of Mamba-based models, the techniques of adapting Mamba to diverse data, and the applications where Mamba can excel. Specifically, we first review the foundational knowledge of various representative deep learning models and the details of Mamba-1&2 as preliminaries. Then, to showcase the significance of Mamba for AI, we comprehensively review the related studies focusing on Mamba models' architecture design, data adaptability, and applications. Finally, we present a discussion of current limitations and explore various promising research directions to provide deeper insights for future investigations.", "authors": ["Haohao Qu", "Liangbo Ning", "Rui An", "Wenqi Fan", "Tyler Derr", "Hui Liu", "Xin Xu", "Qing Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-02", "url": "https://arxiv.org/abs/2408.01129", "pdf_url": "https://arxiv.org/pdf/2408.01129v8", "arxiv_id": "2408.01129", "doi": "10.48550/arXiv.2408.01129", "citation_count": 94, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4944}
{"id": "e3b8dfaf27876a1db8e87531194c0851f3453368543950c9ef04e1c6f643b7f4", "sources": ["arxiv", "semantic_scholar"], "title": "PMoE: Progressive Mixture of Experts with Asymmetric Transformer for Continual Learning", "abstract": "Large Language Models (LLMs) encounter significant challenges in continual learning due to catastrophic forgetting, where new information overwrites previously acquired knowledge. This limitation leads to substantial environmental and economic waste. In this study, we introduce the PMoE, Progressive Mixture of Experts with Asymmetric Transformer, which aims to minimize forgetting by utilizing an asymmetric design with shallow layers dedicated to general knowledge and deep layers for new knowledge. PMoE incorporates progressively added experts in deep layers and a router that allocates new knowledge to the appropriate experts efficiently. The router, positioned adjacent to the deep layers, utilizes deep features aggregating consolidated information. This enables the router to perform efficiently, allocating new knowledge to the appropriate experts, which progressively increase in the deep layers. Extensive experiments on TRACE datasets and general language understanding datasets demonstrate that the proposed PMoE outperforms previous state-of-the-art approaches.", "authors": ["Min Jae Jung", "JooHee Kim"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-31", "url": "https://arxiv.org/abs/2407.21571", "pdf_url": "https://arxiv.org/pdf/2407.21571v1", "arxiv_id": "2407.21571", "doi": "10.48550/arXiv.2407.21571", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "d9e21c6e4f9b334b0fd4728da8ba67a883602bdb3a83cc83e948cf0a5efe78aa", "sources": ["arxiv", "semantic_scholar"], "title": "MoMa: Efficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts", "abstract": "We introduce MoMa, a novel modality-aware mixture-of-experts (MoE) architecture designed for pre-training mixed-modal, early-fusion language models. MoMa processes images and text in arbitrary sequences by dividing expert modules into modality-specific groups. These groups exclusively process designated tokens while employing learned routing within each group to maintain semantically informed adaptivity. Our empirical results reveal substantial pre-training efficiency gains through this modality-specific parameter allocation. Under a 1-trillion-token training budget, the MoMa 1.4B model, featuring 4 text experts and 4 image experts, achieves impressive FLOPs savings: 3.7x overall, with 2.6x for text and 5.2x for image processing compared to a compute-equivalent dense baseline, measured by pre-training loss. This outperforms the standard expert-choice MoE with 8 mixed-modal experts, which achieves 3x overall FLOPs savings (3x for text, 2.8x for image). Combining MoMa with mixture-of-depths (MoD) further improves pre-training FLOPs savings to 4.2x overall (text: 3.4x, image: 5.3x), although this combination hurts performance in causal inference due to increased sensitivity to router accuracy. These results demonstrate MoMa's potential to significantly advance the efficiency of mixed-modal, early-fusion language model pre-training, paving the way for more resource-efficient and capable multimodal AI systems.", "authors": ["Xi Victoria Lin", "Akshat Shrivastava", "Liang Luo", "Srinivasan Iyer", "Mike Lewis", "Gargi Ghosh", "Luke Zettlemoyer", "Armen Aghajanyan"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-31", "url": "https://arxiv.org/abs/2407.21770", "pdf_url": "https://arxiv.org/pdf/2407.21770v3", "arxiv_id": "2407.21770", "doi": "10.48550/arXiv.2407.21770", "citation_count": 71, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4643}
{"id": "1678294e1071d6d64dfed1ab2303c6e22cacbeb8eadeb030a3eff93a83b34612", "sources": ["arxiv", "semantic_scholar"], "title": "Palu: Compressing KV-Cache with Low-Rank Projection", "abstract": "Post-training KV-Cache compression methods typically either sample a subset of effectual tokens or quantize the data into lower numerical bit width. However, these methods cannot exploit redundancy in the hidden dimension of the KV tensors. This paper presents a hidden dimension compression approach called Palu, a KV-Cache compression framework that utilizes low-rank projection to reduce inference-time LLM memory usage. Palu decomposes the linear layers into low-rank matrices, caches compressed intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, compression rate, and efficiency, Palu further encompasses (1) a medium-grained low-rank decomposition scheme, (2) an efficient rank search algorithm, (3) low-rank-aware quantization compatibility enhancements, and (4) optimized GPU kernels with operators fusion. Extensive experiments with popular LLMs show that Palu compresses KV-Cache by 50% while maintaining strong accuracy and delivering up to 1.89x on the RoPE-based attention module. When combined with quantization, Palu's inherent quantization-friendly design yields small to negligible extra accuracy degradation while saving additional memory than quantization-only methods and achieving up to 2.91x speedup for the RoPE-based attention. Moreover, it maintains comparable or even better accuracy (up to 1.19 lower perplexity) compared to quantization-only methods. These results demonstrate Palu's superior capability to effectively address the efficiency and memory challenges of LLM inference posed by KV-Cache. Our code is publicly available at: https://github.com/shadowpa0327/Palu", "authors": ["Chi-Chih Chang", "Wei-Cheng Lin", "Chien-Yu Lin", "Chong-Yan Chen", "Yu-Fang Hu", "Pei-Shuo Wang", "Ning-Chi Huang", "Luis Ceze", "Mohamed S. Abdelfattah", "Kai-Chiang Wu"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-30", "url": "https://arxiv.org/abs/2407.21118", "pdf_url": "https://arxiv.org/pdf/2407.21118v2", "arxiv_id": "2407.21118", "doi": null, "citation_count": 32, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/shadowpa0327/Palu", "venue": "International Conference on Learning Representations", "quality_score": 0.3796}
{"id": "f782af87a021d28cf0c0c6a2caa029df75361d710acbaf4d45835371f2c8e8ca", "sources": ["arxiv", "semantic_scholar"], "title": "ThinK: Thinner Key Cache by Query-Driven Pruning", "abstract": "Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications. However, their increased computational and memory demands present significant challenges, especially when handling long sequences. This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference. Unlike existing approaches that optimize the memory based on the sequence length, we identify substantial redundancy in the channel dimension of the KV cache, as indicated by an uneven magnitude distribution and a low-rank structure in the attention weights. In response, we propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels. Our approach not only maintains or enhances model accuracy but also achieves a reduction in KV cache memory costs by over 20% compared with vanilla KV cache eviction and quantization methods. For instance, ThinK integrated with KIVI can achieve a 2.8x reduction in peak memory usage while maintaining nearly the same quality, enabling up to a 5x increase in batch size when using a single GPU. Extensive evaluations on the LLaMA and Mistral models across various long-sequence datasets verified the efficiency of ThinK, establishing a new baseline algorithm for efficient LLM deployment without compromising performance. Our code has been made available at https://github.com/SalesforceAIResearch/ThinK.", "authors": ["Yuhui Xu", "Zhanming Jie", "Hanze Dong", "Lei Wang", "Xudong Lu", "Aojun Zhou", "Amrita Saha", "Caiming Xiong", "Doyen Sahoo"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-30", "url": "https://arxiv.org/abs/2407.21018", "pdf_url": "https://arxiv.org/pdf/2407.21018v3", "arxiv_id": "2407.21018", "doi": "10.48550/arXiv.2407.21018", "citation_count": 56, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/SalesforceAIResearch/ThinK", "venue": "International Conference on Learning Representations", "quality_score": 0.4515}
{"id": "f391aba987d7aa8c9730ef826322a55638b4f9f44328d90f6addc36672681136", "sources": ["arxiv", "semantic_scholar"], "title": "A2SF: Accumulative Attention Scoring with Forgetting Factor for Token Pruning in Transformer Decoder", "abstract": "Recently, large language models (LLM) based on transformers are facing memory bottleneck issues due to KV cache, especially in long sequence handling. Previous researches proposed KV cache compression techniques that identify insignificant tokens based on Accumulative Attention Scores and removes their items from KV cache, noting that only few tokens play an important role in attention operations. However, we have observed that the existing Accumulative Attention Score is not suitable for the transformer decoder structure. In the decoder model, the number of times the Attention Score accumulates varies depending on the order of token appearance due to the effect of masking, causing an uneven comparison between tokens. To solve this, we propose Accumulative Attention Score with Forgetting Factor (A2SF) technique, which introduces a Forgetting Factor in the Attention Score accumulation process. A2SF applies a penalty to the past Attention Score generated from old tokens by repeatedly multiplying the Forgetting Factor to the Attention Score over time. Therefore, older tokens receive a larger penalty, providing fairness among different ages of tokens. Through the fair comparison among tokens, we can more effectively select important tokens. We have verified the accuracy improvement through A2SF in the OPT and LLaMA models and A2SF improves the accuracy of LLaMA 2 by up to 7.8% and 5.1% on 1-shot and 0-shot.", "authors": ["Hyun-rae Jo", "Dongkun Shin"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-30", "url": "https://arxiv.org/abs/2407.20485", "pdf_url": "https://arxiv.org/pdf/2407.20485v2", "arxiv_id": "2407.20485", "doi": "10.48550/arXiv.2407.20485", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "e50a8cefec4dee1272ddec94887556093d3033fd4ab087580762c7604e4bb56d", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Nested Experts: Adaptive Processing of Visual Tokens", "abstract": "The visual medium (images and videos) naturally contains a large amount of information redundancy, thereby providing a great opportunity for leveraging efficiency in processing. While Vision Transformer (ViT) based models scale effectively to large data regimes, they fail to capitalize on this inherent redundancy, leading to higher computational costs. Mixture of Experts (MoE) networks demonstrate scalability while maintaining same inference-time costs, but they come with a larger parameter footprint. We present Mixture of Nested Experts (MoNE), which utilizes a nested structure for experts, wherein individual experts fall on an increasing compute-accuracy curve. Given a compute budget, MoNE learns to dynamically choose tokens in a priority order, and thus redundant tokens are processed through cheaper nested experts. Using this framework, we achieve equivalent performance as the baseline models, while reducing inference time compute by over two-fold. We validate our approach on standard image and video datasets - ImageNet-21K, Kinetics400, and Something-Something-v2. We further highlight MoNE$'$s adaptability by showcasing its ability to maintain strong performance across different inference-time compute budgets on videos, using only a single trained model.", "authors": ["Gagan Jain", "Nidhi Hegde", "Aditya Kusupati", "Arsha Nagrani", "Shyamal Buch", "Prateek Jain", "Anurag Arnab", "Sujoy Paul"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-29", "url": "https://arxiv.org/abs/2407.19985", "pdf_url": "https://arxiv.org/pdf/2407.19985v2", "arxiv_id": "2407.19985", "doi": "10.48550/arXiv.2407.19985", "citation_count": 24, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3495}
{"id": "7f8a199ebabcae3309ea5e9bf9003386b7aa73225250dc8b60b92bf1f8f2b4fd", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Modular Experts: Distilling Knowledge from a Multilingual Teacher into Specialized Modular Language Models", "abstract": "This research combines Knowledge Distillation (KD) and Mixture of Experts (MoE) to develop modular, efficient multilingual language models. Key objectives include evaluating adaptive versus fixed alpha methods in KD and comparing modular MoE architectures for handling multi-domain inputs and preventing catastrophic forgetting. KD compresses large language models (LLMs) into smaller, efficient models, while MoE enhances modularity with specialized tasks. Experiments showed similar performance for both KD methods, with marginal improvements from adaptive alpha. A combined loss approach provided more stable learning. The router, trained to classify input sequences into English, French, German, or Python, achieved 99.95% precision, recall, and F1 score, with Logistic Regression being the most effective classifier. Evaluations of modular MoE architectures revealed that Pre-trained Language Experts (PLE) and Joint Expert Embedding Training (JEET) performed similarly, while the MoE with Common Expert (MoE-CE) setup showed slightly lower performance. Including a common expert in MoE-CE improved its performance. Studies on catastrophic forgetting indicated that sequential training led to significant forgetting, while single-session training with balanced batches and the MoE approach mitigated this issue. The MoE architecture preserved knowledge across multiple languages effectively. The research contributes open-sourced resources including the dataset (https://zenodo.org/doi/10.5281/zenodo.12677631), a balanced dataset creation tool (https://github.com/padas-lab-de/multi-language-dataset-creator), and the research codebase (https://github.com/ModMaamari/mixture-modular-experts).", "authors": ["Mohammed Al-Maamari", "Mehdi Ben Amor", "Michael Granitzer"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-28", "url": "https://arxiv.org/abs/2407.19610", "pdf_url": "https://arxiv.org/pdf/2407.19610v1", "arxiv_id": "2407.19610", "doi": "10.48550/arXiv.2407.19610", "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/padas-lab-de/multi-language-dataset-creator", "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "c348e65d58b9a46b4e387b55a7cd1d030e8e1caef9b55ee2d00c7bbef70ae23e", "sources": ["arxiv", "semantic_scholar"], "title": "Keep the Cost Down: A Review on Methods to Optimize LLM' s KV-Cache Consumption", "abstract": "Large Language Models (LLMs), epitomized by ChatGPT's release in late 2022, have revolutionized various industries with their advanced language comprehension. However, their efficiency is challenged by the Transformer architecture's struggle with handling long texts. KV Cache has emerged as a pivotal solution to this issue, converting the time complexity of token generation from quadratic to linear, albeit with increased GPU memory overhead proportional to conversation length. With the development of the LLM community and academia, various KV Cache compression methods have been proposed. In this review, we dissect the various properties of KV Cache and elaborate on various methods currently used to optimize the KV Cache space usage of LLMs. These methods span the pre-training phase, deployment phase, and inference phase, and we summarize the commonalities and differences among these methods. Additionally, we list some metrics for evaluating the long-text capabilities of large language models, from both efficiency and capability perspectives. Our review thus sheds light on the evolving landscape of LLM optimization, offering insights into future advancements in this dynamic field. Links to the papers mentioned in this review can be found in our Github Repo https://github.com/zcli-charlie/Awesome-KV-Cache.", "authors": ["Luohe Shi", "Hongyi Zhang", "Yao Yao", "Zuchao Li", "Hai Zhao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-25", "url": "https://arxiv.org/abs/2407.18003", "pdf_url": "https://arxiv.org/pdf/2407.18003v4", "arxiv_id": "2407.18003", "doi": "10.48550/arXiv.2407.18003", "citation_count": 119, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/zcli-charlie/Awesome-KV-Cache", "venue": "arXiv.org", "quality_score": 0.5198}
{"id": "bf02b402cfb44e746e91e4b198ae4f20d5495fbfc7ca26b905f6e90622ce2574", "sources": ["arxiv", "semantic_scholar"], "title": "M4: Multi-Proxy Multi-Gate Mixture of Experts Network for Multiple Instance Learning in Histopathology Image Analysis", "abstract": "Multiple instance learning (MIL) has been successfully applied for whole slide images (WSIs) analysis in computational pathology, enabling a wide range of prediction tasks from tumor subtyping to inferring genetic mutations and multi-omics biomarkers. However, existing MIL methods predominantly focus on single-task learning, resulting in not only overall low efficiency but also the overlook of inter-task relatedness. To address these issues, we proposed an adapted architecture of Multi-gate Mixture-of-experts with Multi-proxy for Multiple instance learning (M4), and applied this framework for simultaneous prediction of multiple genetic mutations from WSIs. The proposed M4 model has two main innovations: (1) utilizing a mixture of experts with multiple gating strategies for multi-genetic mutation prediction on a single pathological slide; (2) constructing multi-proxy expert network and gate network for comprehensive and effective modeling of pathological image information. Our model achieved significant improvements across five tested TCGA datasets in comparison to current state-of-the-art single-task methods. The code is available at:https://github.com/Bigyehahaha/M4.", "authors": ["Junyu Li", "Ye Zhang", "Wen Shu", "Xiaobing Feng", "Yingchun Wang", "Pengju Yan", "Xiaolin Li", "Chulin Sha", "Min He"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2024-07-24", "url": "https://arxiv.org/abs/2407.17267", "pdf_url": "https://arxiv.org/pdf/2407.17267v1", "arxiv_id": "2407.17267", "doi": "10.48550/arXiv.2407.17267", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Bigyehahaha/M4", "venue": null, "quality_score": 0.2113}
{"id": "df2689e1ed53305dae4079ec6f895d8866e09211e9822afc4da55fa8e64721e7", "sources": ["arxiv", "semantic_scholar"], "title": "Compensate Quantization Errors+: Quantized Models Are Inquisitive Learners", "abstract": "The quantization of large language models (LLMs) has been a prominent research area aimed at enabling their lightweight deployment in practice. Existing research about LLM's quantization has mainly explored the interplay between weights and activations, or employing auxiliary components while neglecting the necessity of adjusting weights during quantization. Consequently, original weight distributions frequently fail to yield desired results after round-to-nearest (RTN) quantization. Even though incorporating techniques such as mixed precision and low-rank error approximation in LLM's quantization can yield improved results, they inevitably introduce additional computational overhead. On the other hand, traditional techniques for weight quantization, such as Generative Post-Training Quantization, rely on manually tweaking weight distributions to minimize local errors, but they fall short of achieving globally optimal outcomes. Although the recently proposed Learnable Singular-value Increment improves global weight quantization by modifying weight distributions, it disrupts the original distribution considerably. This introduces pronounced bias toward the training data and can degrade downstream task performance. In this paper, we introduce Singular-value Diagonal Expansion, a more nuanced approach to refining weight distributions to achieve better quantization alignment. Furthermore, we introduce Cross-layer Learning that improves overall quantization outcomes by distributing errors more evenly across layers. Our plug-and-play weight-quantization methods demonstrate substantial performance improvements over state-of-the-art approaches, including OmniQuant, DuQuant, and PrefixQuant.", "authors": ["Yifei Gao", "Jie Ou", "Lei Wang", "Jun Cheng", "Mengchu Zhou"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-22", "url": "https://arxiv.org/abs/2407.15508", "pdf_url": "https://arxiv.org/pdf/2407.15508v3", "arxiv_id": "2407.15508", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "ea6bcad76576699ef4e44fc4a7f49f5ec44dd95a7f9744e065bd6b47f0d2f106", "sources": ["arxiv", "semantic_scholar"], "title": "RazorAttention: Efficient KV Cache Compression Through Retrieval Heads", "abstract": "The memory and computational demands of Key-Value (KV) cache present significant challenges for deploying long-context language models. Previous approaches attempt to mitigate this issue by selectively dropping tokens, which irreversibly erases critical information that might be needed for future queries. In this paper, we propose a novel compression technique for KV cache that preserves all token information. Our investigation reveals that: i) Most attention heads primarily focus on the local context; ii) Only a few heads, denoted as retrieval heads, can essentially pay attention to all input tokens. These key observations motivate us to use separate caching strategy for attention heads. Therefore, we propose RazorAttention, a training-free KV cache compression algorithm, which maintains a full cache for these crucial retrieval heads and discards the remote tokens in non-retrieval heads. Furthermore, we introduce a novel mechanism involving a \"compensation token\" to further recover the information in the dropped tokens. Extensive evaluations across a diverse set of large language models (LLMs) demonstrate that RazorAttention achieves a reduction in KV cache size by over 70% without noticeable impacts on performance. Additionally, RazorAttention is compatible with FlashAttention, rendering it an efficient and plug-and-play solution that enhances LLM inference efficiency without overhead or retraining of the original model.", "authors": ["Hanlin Tang", "Yang Lin", "Jing Lin", "Qingsen Han", "Shikuan Hong", "Yiwu Yao", "Gongyi Wang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-22", "url": "https://arxiv.org/abs/2407.15891", "pdf_url": "https://arxiv.org/pdf/2407.15891v1", "arxiv_id": "2407.15891", "doi": "10.48550/arXiv.2407.15891", "citation_count": 88, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4873}
{"id": "590f5d0464b49d74299d45cbb7f8fdd1d1efe0df0abb5ca803fff98cfe314c84", "sources": ["arxiv", "semantic_scholar"], "title": "FMamba: Mamba based on Fast-attention for Multivariate Time-series Forecasting", "abstract": "In multivariate time-series forecasting (MTSF), extracting the temporal correlations of the input sequences is crucial. While popular Transformer-based predictive models can perform well, their quadratic computational complexity results in inefficiency and high overhead. The recently emerged Mamba, a selective state space model, has shown promising results in many fields due to its strong temporal feature extraction capabilities and linear computational complexity. However, due to the unilateral nature of Mamba, channel-independent predictive models based on Mamba cannot attend to the relationships among all variables in the manner of Transformer-based models. To address this issue, we combine fast-attention with Mamba to introduce a novel framework named FMamba for MTSF. Technically, we first extract the temporal features of the input variables through an embedding layer, then compute the dependencies among input variables via the fast-attention module. Subsequently, we use Mamba to selectively deal with the input features and further extract the temporal dependencies of the variables through the multi-layer perceptron block (MLP-block). Finally, FMamba obtains the predictive results through the projector, a linear layer. Experimental results on eight public datasets demonstrate that FMamba can achieve state-of-the-art performance while maintaining low computational overhead.", "authors": ["Shusen Ma", "Yu Kang", "Peng Bai", "Yun-Bo Zhao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-20", "url": "https://arxiv.org/abs/2407.14814", "pdf_url": "https://arxiv.org/pdf/2407.14814v1", "arxiv_id": "2407.14814", "doi": "10.48550/arXiv.2407.14814", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "56ececd97103353a7cb789852a8c8b5368e50df25c88f9f91a741bef8f1f0647", "sources": ["arxiv", "semantic_scholar"], "title": "EEGMamba: Bidirectional State Space Model with Mixture of Experts for EEG Multi-task Classification", "abstract": "In recent years, with the development of deep learning, electroencephalogram (EEG) classification networks have achieved certain progress. Transformer-based models can perform well in capturing long-term dependencies in EEG signals. However, their quadratic computational complexity poses a substantial computational challenge. Moreover, most EEG classification models are only suitable for single tasks and struggle with generalization across different tasks, particularly when faced with variations in signal length and channel count. In this paper, we introduce EEGMamba, the first universal EEG classification network to truly implement multi-task learning for EEG applications. EEGMamba seamlessly integrates the Spatio-Temporal-Adaptive (ST-Adaptive) module, bidirectional Mamba, and Mixture of Experts (MoE) into a unified framework. The proposed ST-Adaptive module performs unified feature extraction on EEG signals of different lengths and channel counts through spatial-adaptive convolution and incorporates a class token to achieve temporal-adaptability. Moreover, we design a bidirectional Mamba particularly suitable for EEG signals for further feature extraction, balancing high accuracy, fast inference speed, and efficient memory-usage in processing long EEG signals. To enhance the processing of EEG data across multiple tasks, we introduce task-aware MoE with a universal expert, effectively capturing both differences and commonalities among EEG data from different tasks. We evaluate our model on eight publicly available EEG datasets, and the experimental results demonstrate its superior performance in four types of tasks: seizure detection, emotion recognition, sleep stage classification, and motor imagery. The code is set to be released soon.", "authors": ["Yiyu Gui", "MingZhi Chen", "Yuqi Su", "Guibo Luo", "Yuchao Yang"], "categories": ["eess.SP", "cs.LG"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2024-07-20", "url": "https://arxiv.org/abs/2407.20254", "pdf_url": "https://arxiv.org/pdf/2407.20254v2", "arxiv_id": "2407.20254", "doi": null, "citation_count": 23, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3451}
{"id": "8b6e620e093f70854a435643d8ba3b96c8977865e16554888900ff1e4382df9e", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts with Mixture of Precisions for Tuning Quality of Service", "abstract": "The increasing demand for deploying large Mixture-of-Experts (MoE) models in resource-constrained environments necessitates efficient approaches to address their high memory and computational requirements challenges. Moreover, given that tasks come in different user-defined constraints and the available resources change over time in multi-tenant environments, it is necessary to design an approach which provides a flexible configuration space. This paper presents an adaptive serving approach for the efficient deployment of MoE models, capitalizing on partial quantization of the experts. By dynamically determining the number of quantized experts and their distribution across CPU and GPU, our approach explores the Pareto frontier and offers a fine-grained range of configurations for tuning throughput and model quality. Our evaluation on an NVIDIA A100 GPU using a Mixtral 8x7B MoE model for three language modelling benchmarks demonstrates that the throughput of token generation can be adjusted from 0.63 to 13.00 token per second. This enhancement comes with a marginal perplexity increase of 3.81 to 4.00, 13.59 to 14.17, and 7.24 to 7.40 for WikiText2, PTB, and C4 datasets respectively under maximum quantization. These results highlight the practical applicability of our approach in dynamic and accuracy-sensitive applications where both memory usage and output quality are important.", "authors": ["HamidReza Imani", "Abdolah Amirany", "Tarek El-Ghazawi"], "categories": ["cs.DC", "cs.AI", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-19", "url": "https://arxiv.org/abs/2407.14417", "pdf_url": "https://arxiv.org/pdf/2407.14417v2", "arxiv_id": "2407.14417", "doi": "10.1109/ICRC64395.2024.10937027", "citation_count": 21, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Rebooting Computing", "quality_score": 0.3356}
{"id": "8340e233dacd0cad9179b29eec157240edb00cf1977df1c5a0b5c6d7ac997e38", "sources": ["arxiv", "semantic_scholar"], "title": "Investigating the Indirect Object Identification circuit in Mamba", "abstract": "How well will current interpretability techniques generalize to future models? A relevant case study is Mamba, a recent recurrent architecture with scaling comparable to Transformers. We adapt pre-Mamba techniques to Mamba and partially reverse-engineer the circuit responsible for the Indirect Object Identification (IOI) task. Our techniques provide evidence that 1) Layer 39 is a key bottleneck, 2) Convolutions in layer 39 shift names one position forward, and 3) The name entities are stored linearly in Layer 39's SSM. Finally, we adapt an automatic circuit discovery tool, positional Edge Attribution Patching, to identify a Mamba IOI circuit. Our contributions provide initial evidence that circuit-based mechanistic interpretability tools work well for the Mamba architecture.", "authors": ["Danielle Ensign", "Adrià Garriga-Alonso"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-19", "url": "https://arxiv.org/abs/2407.14008", "pdf_url": "https://arxiv.org/pdf/2407.14008v2", "arxiv_id": "2407.14008", "doi": "10.48550/arXiv.2407.14008", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "97dcd326f540970adb9f30c15fe4ae9af60e0e8d54e4b8605c997b95418fbef9", "sources": ["arxiv", "semantic_scholar"], "title": "Attention in SRAM on Tenstorrent Grayskull", "abstract": "When implementations of the Transformer's self-attention layer utilize SRAM instead of DRAM, they can achieve significant speedups. The Tenstorrent Grayskull architecture provides a large SRAM, distributed across a grid of cores. This work presents a fused kernel for Grayskull, that exclusively utilizes its large SRAM by combining matrix multiplication, attention score scaling and Softmax operations. Additionally, a dedicated Softmax kernel utilizing the SRAM and a CPU implementation serving as a baseline are presented. The Softmax operation consumes most of the runtime in the computation of attention weights from queries and keys on Grayskull. The speedup of the dedicated Softmax kernel compared to the CPU implementation is up to $10 \\times$, and the Softmax implementation inside the fused kernel is approximately $1.8 \\times$ faster than the dedicated Softmax kernel. The time and memory complexity of all implementations is quadratic in sequence length. Currently, the Grayskull e150 is approximately $30 \\times$ cheaper for the general public than an Nvidia H100 PCIe (a state-of-the-art GPU) and offers approximately $1.5 \\times$ more SRAM.", "authors": ["Moritz Thüning"], "categories": ["cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-18", "url": "https://arxiv.org/abs/2407.13885", "pdf_url": "https://arxiv.org/pdf/2407.13885v1", "arxiv_id": "2407.13885", "doi": "10.48550/arXiv.2407.13885", "citation_count": 9, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/moritztng/grayskull-attention", "venue": "arXiv.org", "quality_score": 0.25}
{"id": "5dd7caed123ba68c82496fef166ac67c29a4675930897aac70d3def088c18151", "sources": ["arxiv", "semantic_scholar"], "title": "MoME: Mixture of Multimodal Experts for Generalist Multimodal Large Language Models", "abstract": "Multimodal large language models (MLLMs) have demonstrated impressive capabilities across various vision-language tasks. However, a generalist MLLM typically underperforms compared with a specialist MLLM on most VL tasks, which can be attributed to task interference. In this paper, we propose a mixture of multimodal experts (MoME) to mitigate task interference and obtain a generalist MLLM. Our MoME is composed of two key components, a mixture of vision experts (MoVE) and a mixture of language experts (MoLE). MoVE can adaptively modulate the features transformed from various vision encoders, and has a strong compatibility in transformation architecture. MoLE incorporates sparsely gated experts into LLMs to achieve painless improvements with roughly unchanged inference costs. In response to task interference, our MoME specializes in both vision and language modality to adapt to task discrepancies. Extensive experiments show that MoME significantly improves the performance of generalist MLLMs across various VL tasks. The source code is released at https://github.com/JiuTian-VL/MoME", "authors": ["Leyang Shen", "Gongwei Chen", "Rui Shao", "Weili Guan", "Liqiang Nie"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-17", "url": "https://arxiv.org/abs/2407.12709", "pdf_url": "https://arxiv.org/pdf/2407.12709v1", "arxiv_id": "2407.12709", "doi": "10.48550/arXiv.2407.12709", "citation_count": 47, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/JiuTian-VL/MoME", "venue": "Neural Information Processing Systems", "quality_score": 0.4203}
{"id": "ba2c7979838621cd0a994959a0c0a6b91f64838646337ee0c1109ab0878d1df7", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba-PTQ: Outlier Channels in Recurrent Large Language Models", "abstract": "Modern recurrent layers are emerging as a promising path toward edge deployment of foundation models, especially in the context of large language models (LLMs). Compressing the whole input sequence in a finite-dimensional representation enables recurrent layers to model long-range dependencies while maintaining a constant inference cost for each token and a fixed memory requirement. However, the practical deployment of LLMs in resource-limited environments often requires further model compression, such as quantization and pruning. While these techniques are well-established for attention-based models, their effects on recurrent layers remain underexplored. In this preliminary work, we focus on post-training quantization for recurrent LLMs and show that Mamba models exhibit the same pattern of outlier channels observed in attention-based LLMs. We show that the reason for the difficulty of quantizing SSMs is caused by activation outliers, similar to those observed in transformer-based LLMs. We report baseline results for post-training quantization of Mamba that do not take into account the activation outliers and suggest first steps for outlier-aware quantization.", "authors": ["Alessandro Pierro", "Steven Abreu"], "categories": ["cs.LG", "cs.AI", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-17", "url": "https://arxiv.org/abs/2407.12397", "pdf_url": "https://arxiv.org/pdf/2407.12397v1", "arxiv_id": "2407.12397", "doi": "10.48550/arXiv.2407.12397", "citation_count": 15, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.301}
{"id": "071e94324303159e2c758efff15529ab47300e7c67178401a3aa44f42c2d1efa", "sources": ["arxiv", "semantic_scholar"], "title": "Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference", "abstract": "Large Language Models have excelled in various domains but face efficiency challenges due to the growing Key-Value (KV) cache required for long-sequence inference. Recent efforts aim to reduce KV cache size by evicting vast non-critical cache elements during runtime while preserving generation quality. However, these methods typically allocate compression budgets uniformly across all attention heads, ignoring the unique attention patterns of each head. In this paper, we establish a theoretical loss upper bound between pre- and post-eviction attention output, explaining the optimization target of prior cache eviction methods, while guiding the optimization of adaptive budget allocation. Base on this, we propose {\\it Ada-KV}, the first head-wise adaptive budget allocation strategy. It offers plug-and-play benefits, enabling seamless integration with prior cache eviction methods. Extensive evaluations on 13 datasets from Ruler and 16 datasets from LongBench, all conducted under both question-aware and question-agnostic scenarios, demonstrate substantial quality improvements over existing methods. Our code is available at https://github.com/FFY0/AdaKV.", "authors": ["Yuan Feng", "Junlin Lv", "Yukun Cao", "Xike Xie", "S. Kevin Zhou"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-16", "url": "https://arxiv.org/abs/2407.11550", "pdf_url": "https://arxiv.org/pdf/2407.11550v5", "arxiv_id": "2407.11550", "doi": "10.48550/arXiv.2407.11550", "citation_count": 170, "influential_citation_count": 31, "has_code": true, "code_url": "https://github.com/FFY0/AdaKV", "venue": "arXiv.org", "quality_score": 0.7526}
{"id": "680daf7f84ebad7a7f153e4be6f792db48b099e5f0891453ffbb583549ff9a2b", "sources": ["arxiv", "semantic_scholar"], "title": "GoldFinch: High Performance RWKV/Transformer Hybrid with Linear Pre-Fill and Extreme KV-Cache Compression", "abstract": "We introduce GoldFinch, a hybrid Linear Attention/Transformer sequence model that uses a new technique to efficiently generate a highly compressed and reusable KV-Cache in linear time and space with respect to sequence length. GoldFinch stacks our new GOLD transformer on top of an enhanced version of the Finch (RWKV-6) architecture. We train up to 1.5B parameter class models of the Finch, Llama, and GoldFinch architectures, and find dramatically improved modeling performance relative to both Finch and Llama. Our cache size savings increase linearly with model layer count, ranging from 756-2550 times smaller than the traditional transformer cache for common sizes, enabling inference of extremely large context lengths even on limited hardware. Although autoregressive generation has O(n) time complexity per token because of attention, pre-fill computation of the entire initial cache state for a submitted context costs only O(1) time per token due to the use of a recurrent neural network (RNN) to generate this cache. We release our trained weights and training code under the Apache 2.0 license for community use.", "authors": ["Daniel Goldstein", "Fares Obeid", "Eric Alcaide", "Guangyu Song", "Eugene Cheah"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-16", "url": "https://arxiv.org/abs/2407.12077", "pdf_url": "https://arxiv.org/pdf/2407.12077v1", "arxiv_id": "2407.12077", "doi": "10.48550/arXiv.2407.12077", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "310d23d6f7047599f9b1cf1bee6704a191d4e211994654d34a74c09511244e5e", "sources": ["arxiv", "semantic_scholar"], "title": "Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering", "abstract": "Since the introduction of NeRFs, considerable attention has been focused on improving their training and inference times, leading to the development of Fast-NeRFs models. Despite demonstrating impressive rendering speed and quality, the rapid convergence of such models poses challenges for further improving reconstruction quality. Common strategies to improve rendering quality involves augmenting model parameters or increasing the number of sampled points. However, these computationally intensive approaches encounter limitations in achieving significant quality enhancements. This study introduces a model-agnostic framework inspired by Sparsely-Gated Mixture of Experts to enhance rendering quality without escalating computational complexity. Our approach enables specialization in rendering different scene components by employing a mixture of experts with varying resolutions. We present a novel gate formulation designed to maximize expert capabilities and propose a resolution-based routing technique to effectively induce sparsity and decompose scenes. Our work significantly improves reconstruction quality while maintaining competitive performance.", "authors": ["Francesco Di Sario", "Riccardo Renzulli", "Enzo Tartaglione", "Marco Grangetto"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-15", "url": "https://arxiv.org/abs/2407.10389", "pdf_url": "https://arxiv.org/pdf/2407.10389v3", "arxiv_id": "2407.10389", "doi": "10.48550/arXiv.2407.10389", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "European Conference on Computer Vision", "quality_score": 0.2113}
{"id": "4a0167c133c8561ea01824cd6edb995ff5285ea4ee5f815b879f3a5e6cec8a10", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond KV Caching: Shared Attention for Efficient LLMs", "abstract": "The efficiency of large language models (LLMs) remains a critical challenge, particularly in contexts where computational resources are limited. Traditional attention mechanisms in these models, while powerful, require significant computational and memory resources due to the necessity of recalculating and storing attention weights across different layers. This paper introduces a novel Shared Attention (SA) mechanism, designed to enhance the efficiency of LLMs by directly sharing computed attention weights across multiple layers. Unlike previous methods that focus on sharing intermediate Key-Value (KV) caches, our approach utilizes the isotropic tendencies of attention distributions observed in advanced LLMs post-pretraining to reduce both the computational flops and the size of the KV cache required during inference. We empirically demonstrate that implementing SA across various LLMs results in minimal accuracy loss on standard benchmarks. Our findings suggest that SA not only conserves computational resources but also maintains robust model performance, thereby facilitating the deployment of more efficient LLMs in resource-constrained environments.", "authors": ["Bingli Liao", "Danilo Vasconcellos Vargas"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-13", "url": "https://arxiv.org/abs/2407.12866", "pdf_url": "https://arxiv.org/pdf/2407.12866v1", "arxiv_id": "2407.12866", "doi": "10.48550/arXiv.2407.12866", "citation_count": 12, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Neurocomputing", "quality_score": 0.2785}
{"id": "83068f8f9555fdc2dac24264c8bd893992fca72bbe51837883701ca50c22cc74", "sources": ["arxiv", "semantic_scholar"], "title": "Speech Slytherin: Examining the Performance and Efficiency of Mamba for Speech Separation, Recognition, and Synthesis", "abstract": "It is too early to conclude that Mamba is a better alternative to transformers for speech before comparing Mamba with transformers in terms of both performance and efficiency in multiple speech-related tasks. To reach this conclusion, we propose and evaluate three models for three tasks: Mamba-TasNet for speech separation, ConMamba for speech recognition, and VALL-M for speech synthesis. We compare them with transformers of similar sizes in performance, memory, and speed. Our Mamba or Mamba-transformer hybrid models show comparable or higher performance than their transformer counterparts: Sepformer, Conformer, and VALL-E. They are more efficient than transformers in memory and speed for speech longer than a threshold duration, inversely related to the resolution of a speech token. Mamba for separation is the most efficient, and Mamba for recognition is the least. Further, we show that Mamba is not more efficient than transformer for speech shorter than the threshold duration and performs worse in models that require joint modeling of text and speech, such as cross or masked attention of two inputs. Therefore, we argue that the superiority of Mamba or transformer depends on particular problems and models. Code available at https://github.com/xi-j/Mamba-TasNet and https://github.com/xi-j/Mamba-ASR.", "authors": ["Xilin Jiang", "Yinghao Aaron Li", "Adrian Nicolas Florea", "Cong Han", "Nima Mesgarani"], "categories": ["eess.AS", "cs.LG", "cs.SD"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-07-13", "url": "https://arxiv.org/abs/2407.09732", "pdf_url": "https://arxiv.org/pdf/2407.09732v1", "arxiv_id": "2407.09732", "doi": "10.1109/ICASSP49660.2025.10889391", "citation_count": 40, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/xi-j/Mamba-TasNet", "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.4032}
{"id": "424709b475eee829ffe9d1602fd1306829c0475f77de3e5c572b35a0ac77bc79", "sources": ["arxiv", "semantic_scholar"], "title": "Diversifying the Expert Knowledge for Task-Agnostic Pruning in Sparse Mixture-of-Experts", "abstract": "By increasing model parameters but activating them sparsely when performing a task, the use of Mixture-of-Experts (MoE) architecture significantly improves the performance of Large Language Models (LLMs) without increasing the inference cost. However, the memory consumption due to the growing number of experts presents a challenge to the deployment of these models in many real world settings. Our empirical study reveals that some experts encode redundant knowledge during pre-training. We thus propose a method of grouping and pruning similar experts to improve the model's parameter efficiency. We validate the effectiveness of our method by pruning three state-of-the-art MoE architectures, including Mixtral, Deepseek-MoE, and Qwen. The evaluation shows that our method outperforms other model pruning methods on a range of natural language tasks. We will release our code to facilitate future research.", "authors": ["Zeliang Zhang", "Xiaodong Liu", "Hao Cheng", "Chenliang Xu", "Jianfeng Gao"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-12", "url": "https://arxiv.org/abs/2407.09590", "pdf_url": "https://arxiv.org/pdf/2407.09590v4", "arxiv_id": "2407.09590", "doi": "10.48550/arXiv.2407.09590", "citation_count": 35, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3891}
{"id": "27ff2a8e575fe0cdd4db3cc9aa2634b64ac7d2a91a1a9e536938fea6037b6ecd", "sources": ["arxiv", "semantic_scholar"], "title": "Model Tells You Where to Merge: Adaptive KV Cache Merging for LLMs on Long-Context Tasks", "abstract": "How to efficiently serve Large Language Models (LLMs) has become a pressing issue because of their huge computational cost in their autoregressive generation process. To mitigate computational costs, LLMs often employ the KV Cache technique to improve the generation speed. While improving the computational efficiency, the storage requirements of the KV cache are substantial, particularly in long-context scenarios, leading to significant memory consumption. Existing KV cache eviction methods often degrade the performance of LLMs in long-context scenarios due to the information loss introduced by eviction. In this paper, we propose a novel KV cache merging approach, called KVMerger, to achieve adaptive KV cache compression for long-context tasks without significant performance degradation under constrained memory budgets. Our approach is inspired by the intriguing observation that key states exhibit high similarity at the token level within a single sequence. To facilitate merging, we develop an effective yet straightforward merging set identification algorithm to identify suitable KV states for merging. Our merging set identification algorithm stimulates the second observation that KV cache sparsity, from similarity perspective, is independent of the dataset and remains persistent at the model level. Subsequently, we propose a Gaussian kernel weighted merging algorithm to selectively merge all states within each merging set. We conduct extensive experiments to demonstrate the effectiveness of KVMerger for long-context tasks under constrained memory budgets, applying it to models including Llama2-7B-chat and Llama2-13B-chat. Using the LongBench and ZeroScroll benchmarks, we compare our method with other KV cache compression techniques, including H2O and CaM, showing that our method achieves superior performance across tasks with both 50% and 35% KV cache budgets.", "authors": ["Zheng Wang", "Boxiao Jin", "Zhongzhi Yu", "Minjia Zhang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-11", "url": "https://arxiv.org/abs/2407.08454", "pdf_url": "https://arxiv.org/pdf/2407.08454v2", "arxiv_id": "2407.08454", "doi": "10.48550/arXiv.2407.08454", "citation_count": 79, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4758}
{"id": "76f475c3166914f56c84addf5def2fcf1c0a03b4aec698eee711bf8cb505e3e5", "sources": ["arxiv", "semantic_scholar"], "title": "Distributed Deep Reinforcement Learning Based Gradient Quantization for Federated Learning Enabled Vehicle Edge Computing", "abstract": "Federated Learning (FL) can protect the privacy of the vehicles in vehicle edge computing (VEC) to a certain extent through sharing the gradients of vehicles' local models instead of local data. The gradients of vehicles' local models are usually large for the vehicular artificial intelligence (AI) applications, thus transmitting such large gradients would cause large per-round latency. Gradient quantization has been proposed as one effective approach to reduce the per-round latency in FL enabled VEC through compressing gradients and reducing the number of bits, i.e., the quantization level, to transmit gradients. The selection of quantization level and thresholds determines the quantization error, which further affects the model accuracy and training time. To do so, the total training time and quantization error (QE) become two key metrics for the FL enabled VEC. It is critical to jointly optimize the total training time and QE for the FL enabled VEC. However, the time-varying channel condition causes more challenges to solve this problem. In this paper, we propose a distributed deep reinforcement learning (DRL)-based quantization level allocation scheme to optimize the long-term reward in terms of the total training time and QE. Extensive simulations identify the optimal weighted factors between the total training time and QE, and demonstrate the feasibility and effectiveness of the proposed scheme.", "authors": ["Cui Zhang", "Wenjun Zhang", "Qiong Wu", "Pingyi Fan", "Qiang Fan", "Jiangzhou Wang", "Khaled B. Letaief"], "categories": ["cs.LG", "cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-11", "url": "https://arxiv.org/abs/2407.08462", "pdf_url": "https://arxiv.org/pdf/2407.08462v2", "arxiv_id": "2407.08462", "doi": "10.1109/JIOT.2024.3447036", "citation_count": 91, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/qiongwu86/Distributed-Deep-Reinforcement-Learning-Based-Gradient", "venue": "IEEE Internet of Things Journal", "quality_score": 0.4909}
{"id": "ba48bd8dd0d9d251ffa3861c9b8ed69775544a3213f88ab377a04b773bdcd234", "sources": ["arxiv", "semantic_scholar"], "title": "MoVEInt: Mixture of Variational Experts for Learning Human-Robot Interactions from Demonstrations", "abstract": "Shared dynamics models are important for capturing the complexity and variability inherent in Human-Robot Interaction (HRI). Therefore, learning such shared dynamics models can enhance coordination and adaptability to enable successful reactive interactions with a human partner. In this work, we propose a novel approach for learning a shared latent space representation for HRIs from demonstrations in a Mixture of Experts fashion for reactively generating robot actions from human observations. We train a Variational Autoencoder (VAE) to learn robot motions regularized using an informative latent space prior that captures the multimodality of the human observations via a Mixture Density Network (MDN). We show how our formulation derives from a Gaussian Mixture Regression formulation that is typically used approaches for learning HRI from demonstrations such as using an HMM/GMM for learning a joint distribution over the actions of the human and the robot. We further incorporate an additional regularization to prevent \"mode collapse\", a common phenomenon when using latent space mixture models with VAEs. We find that our approach of using an informative MDN prior from human observations for a VAE generates more accurate robot motions compared to previous HMM-based or recurrent approaches of learning shared latent representations, which we validate on various HRI datasets involving interactions such as handshakes, fistbumps, waving, and handovers. Further experiments in a real-world human-to-robot handover scenario show the efficacy of our approach for generating successful interactions with four different human interaction partners.", "authors": ["Vignesh Prasad", "Alap Kshirsagar", "Dorothea Koert", "Ruth Stock-Homburg", "Jan Peters", "Georgia Chalvatzaki"], "categories": ["cs.RO", "cs.HC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-10", "url": "https://arxiv.org/abs/2407.07636", "pdf_url": "https://arxiv.org/pdf/2407.07636v2", "arxiv_id": "2407.07636", "doi": "10.1109/LRA.2024.3396074", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Robotics and Automation Letters", "quality_score": 0.2785}
{"id": "9004e1eac2c57a421743319c9fd057b4b01cca11a7642347796e745fe2a6e3a3", "sources": ["arxiv", "semantic_scholar"], "title": "A Simple Architecture for Enterprise Large Language Model Applications based on Role based security and Clearance Levels using Retrieval-Augmented Generation or Mixture of Experts", "abstract": "This study proposes a simple architecture for Enterprise application for Large Language Models (LLMs) for role based security and NATO clearance levels. Our proposal aims to address the limitations of current LLMs in handling security and information access. The proposed architecture could be used while utilizing Retrieval-Augmented Generation (RAG) and fine tuning of Mixture of experts models (MoE). It could be used only with RAG, or only with MoE or with both of them. Using roles and security clearance level of the user, documents in RAG and experts in MoE are filtered. This way information leakage is prevented.", "authors": ["Atilla Özgür", "Yılmaz Uygun"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-09", "url": "https://arxiv.org/abs/2407.06718", "pdf_url": "https://arxiv.org/pdf/2407.06718v1", "arxiv_id": "2407.06718", "doi": "10.48550/arXiv.2407.06718", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "e7b276b3a25131684ec72ea93fa6e8c68172768dd37df405d6737a16f9fef767", "sources": ["arxiv", "semantic_scholar"], "title": "A Survey on LoRA of Large Language Models", "abstract": "Low-Rank Adaptation~(LoRA), which updates the dense neural network layers with pluggable low-rank matrices, is one of the best performed parameter efficient fine-tuning paradigms. Furthermore, it has significant advantages in cross-task generalization and privacy-preserving. Hence, LoRA has gained much attention recently, and the number of related literature demonstrates exponential growth. It is necessary to conduct a comprehensive overview of the current progress on LoRA. This survey categorizes and reviews the progress from the perspectives of (1) downstream adaptation improving variants that improve LoRA's performance on downstream tasks; (2) cross-task generalization methods that mix multiple LoRA plugins to achieve cross-task generalization; (3) efficiency-improving methods that boost the computation-efficiency of LoRA; (4) data privacy-preserving methods that use LoRA in federated learning; (5) application. Besides, this survey also discusses the future directions in this field. At last, we provide a Github page~\\footnote{\\href{https://github.com/ZJU-LLMs/Awesome-LoRAs.git}{https://github.com/ZJU-LLMs/Awesome-LoRAs.git}} for readers to check the updates and initiate discussions on this survey paper.", "authors": ["Yuren Mao", "Yuhang Ge", "Yijiang Fan", "Wenyi Xu", "Yu Mi", "Zhonghao Hu", "Yunjun Gao"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-08", "url": "https://arxiv.org/abs/2407.11046", "pdf_url": "https://arxiv.org/pdf/2407.11046v4", "arxiv_id": "2407.11046", "doi": "10.1007/s11704-024-40663-9", "citation_count": 148, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/ZJU-LLMs/Awesome-LoRAs.git}{https://github.com/ZJU-LLMs/Awesome-LoRAs.git}}", "venue": null, "quality_score": 0.5433}
{"id": "d904876197274fb419bd589288fe7b052a82e299b5687bc8131147cf31bf6454", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning Large Language Models to Intra-module Low-rank Architecture with Transitional Activations", "abstract": "Structured pruning fundamentally reduces computational and memory overheads of large language models (LLMs) and offers a feasible solution for end-side LLM deployment. Structurally pruned models remain dense and high-precision, highly compatible with further tuning and compression. However, as the coarse-grained structured pruning poses large damage to the highly interconnected model, achieving a high compression ratio for scaled-up LLMs remains a challenge. In this paper, we introduce a task-agnostic structured pruning approach coupled with a compact Transformer architecture design. The proposed approach, named TransAct, reduces transitional activations inside multi-head attention (MHA) and multi-layer perceptron (MLP) modules, while preserving the inter-module activations that are sensitive to perturbations. Hence, the LLM is pruned into an intra-module low-rank architecture, significantly reducing weights, KV Cache and attention computation. TransAct is implemented on the LLaMA model and evaluated on downstream benchmarks. Results verify the optimality of our approach at high compression with respect to both efficiency and performance. Further, ablation studies reveal the strength of activation-guided iterative pruning and provide experimental analysis on the redundancy of MHA and MLP modules.", "authors": ["Bowen Shen", "Zheng Lin", "Daren Zha", "Wei Liu", "Jian Luan", "Bin Wang", "Weiping Wang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-08", "url": "https://arxiv.org/abs/2407.05690", "pdf_url": "https://arxiv.org/pdf/2407.05690v1", "arxiv_id": "2407.05690", "doi": "10.48550/arXiv.2407.05690", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1747}
{"id": "4aa3a417d2cb6a5956842d25699c50413bc2732c502887097cff7c13e4470423", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba Hawkes Process", "abstract": "Irregular and asynchronous event sequences are prevalent in many domains, such as social media, finance, and healthcare. Traditional temporal point processes (TPPs), like Hawkes processes, often struggle to model mutual inhibition and nonlinearity effectively. While recent neural network models, including RNNs and Transformers, address some of these issues, they still face challenges with long-term dependencies and computational efficiency. In this paper, we introduce the Mamba Hawkes Process (MHP), which leverages the Mamba state space architecture to capture long-range dependencies and dynamic event interactions. Our results show that MHP outperforms existing models across various datasets. Additionally, we propose the Mamba Hawkes Process Extension (MHP-E), which combines Mamba and Transformer models to enhance predictive capabilities. We present the novel application of the Mamba architecture to Hawkes processes, a flexible and extensible model structure, and a theoretical analysis of the synergy between state space models and Hawkes processes. Experimental results demonstrate the superior performance of both MHP and MHP-E, advancing the field of temporal point process modeling.", "authors": ["Anningzhe Gao", "Shan Dai", "Yan Hu"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-07-07", "url": "https://arxiv.org/abs/2407.05302", "pdf_url": "https://arxiv.org/pdf/2407.05302v1", "arxiv_id": "2407.05302", "doi": "10.48550/arXiv.2407.05302", "citation_count": 5, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "809025225eaf1e0b98f7df32ebb3bce9e3d2f7f28f729edfe3bf1d8345350c80", "sources": ["arxiv", "semantic_scholar"], "title": "The Impact of Quantization and Pruning on Deep Reinforcement Learning Models", "abstract": "Deep reinforcement learning (DRL) has achieved remarkable success across various domains, such as video games, robotics, and, recently, large language models. However, the computational costs and memory requirements of DRL models often limit their deployment in resource-constrained environments. The challenge underscores the urgent need to explore neural network compression methods to make RDL models more practical and broadly applicable. Our study investigates the impact of two prominent compression methods, quantization and pruning on DRL models. We examine how these techniques influence four performance factors: average return, memory, inference time, and battery utilization across various DRL algorithms and environments. Despite the decrease in model size, we identify that these compression techniques generally do not improve the energy efficiency of DRL models, but the model size decreases. We provide insights into the trade-offs between model compression and DRL performance, offering guidelines for deploying efficient DRL models in resource-constrained settings.", "authors": ["Heng Lu", "Mehdi Alemi", "Reza Rawassizadeh"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-05", "url": "https://arxiv.org/abs/2407.04803", "pdf_url": "https://arxiv.org/pdf/2407.04803v1", "arxiv_id": "2407.04803", "doi": "10.48550/arXiv.2407.04803", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "0343b5e22dd5e3d4ea2b94257f6520cd131df17beabae7f0f105b123b7f89115", "sources": ["arxiv", "semantic_scholar"], "title": "QET: Enhancing Quantized LLM Parameters and KV cache Compression through Element Substitution and Residual Clustering", "abstract": "The matrix quantization entails representing matrix elements in a more space-efficient form to reduce storage usage, with dequantization restoring the original matrix for use. We formulate the Quantization Error Minimization (QEM) problem as minimizing the distance between a matrix before and after quantization, under the condition that the quantized matrix occupies the same memory space. Matrix quantization is crucial in various applications, including Large Language Models (LLMs) weight quantization, vector databases, KV cache quantization, graph compression, and image compression. Recent advancements in LLMs, such as GPT-4 and BERT, have highlighted the importance of matrix compression due to the large size of parameters and KV cache, which are stored as matrices. We propose Quantum Entanglement Trees (QET) to address the QEM problem by leveraging the local orderliness of matrix elements, involving iterative element swapping to form a locally ordered matrix. This matrix is then grouped and quantized by columns. To enhance QET, we introduce two optimizations: further quantizing residuals to reduce MSE, and using masking and batch processing to accelerate the algorithm. Experimental results demonstrate that QET can effectively reduce MSE to 5.05%, 13.33%, and 11.89% of the current best method on the LLM dataset, K cache, and V cache, respectively. Our contributions include the abstraction of the QEM problem, the design of the QET algorithm, and the proposal of two optimizations to improve accuracy and speed.", "authors": ["Yanshu Wang", "Wang Li", "Zhaoqian Yao", "Tong Yang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-04", "url": "https://arxiv.org/abs/2407.03637", "pdf_url": "https://arxiv.org/pdf/2407.03637v4", "arxiv_id": "2407.03637", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "44a53ceae24cba9b717c5bf8062d45514bb975ab6f280bceefbe12df1df541d0", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of A Million Experts", "abstract": "The feedforward (FFW) layers in standard transformer architectures incur a linear increase in computational costs and activation memory as the hidden layer width grows. Sparse mixture-of-experts (MoE) architectures have emerged as a viable approach to address this issue by decoupling model size from computational cost. The recent discovery of the fine-grained MoE scaling law shows that higher granularity leads to better performance. However, existing MoE models are limited to a small number of experts due to computational and optimization challenges. This paper introduces PEER (parameter efficient expert retrieval), a novel layer design that utilizes the product key technique for sparse retrieval from a vast pool of tiny experts (over a million). Experiments on language modeling tasks demonstrate that PEER layers outperform dense FFWs and coarse-grained MoEs in terms of performance-compute trade-off. By enabling efficient utilization of a massive number of experts, PEER unlocks the potential for further scaling of transformer models while maintaining computational efficiency.", "authors": ["Xu Owen He"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-04", "url": "https://arxiv.org/abs/2407.04153", "pdf_url": "https://arxiv.org/pdf/2407.04153v1", "arxiv_id": "2407.04153", "doi": "10.48550/arXiv.2407.04153", "citation_count": 65, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4549}
{"id": "64570b8952b0ad4fce5840e30e45f6a9241bf645534e55b754d1da07d8ebd141", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Expert Pruning for Sparse Mixture-of-Experts Language Models: Enhancing Performance and Reducing Inference Costs", "abstract": "The rapid advancement of large language models (LLMs) has led to architectures with billions to trillions of parameters, posing significant deployment challenges due to their substantial demands on memory, processing power, and energy consumption. Sparse Mixture-of-Experts (SMoE) architectures have emerged as a solution, activating only a subset of parameters per token, thereby achieving faster inference while maintaining performance. However, SMoE models still face limitations in broader deployment due to their large parameter counts and significant GPU memory requirements. In this work, we introduce a gradient-free evolutionary strategy named EEP (Efficient Expert P}runing) to enhance the pruning of experts in SMoE models. EEP relies solely on model inference (i.e., no gradient computation) and achieves greater sparsity while maintaining or even improving performance on downstream tasks. EEP can be used to reduce both the total number of experts (thus saving GPU memory) and the number of active experts (thus accelerating inference). For example, we demonstrate that pruning up to 75% of experts in Mixtral $8\\times7$B-Instruct results in a substantial reduction in parameters with minimal performance loss. Remarkably, we observe improved performance on certain tasks, such as a significant increase in accuracy on the SQuAD dataset (from 53.4% to 75.4%), when pruning half of the experts. With these results, EEP not only lowers the barrier to deploying SMoE models,but also challenges the conventional understanding of model pruning by showing that fewer experts can lead to better task-specific performance without any fine-tuning. Code is available at https://github.com/imagination-research/EEP.", "authors": ["Enshu Liu", "Junyi Zhu", "Zinan Lin", "Xuefei Ning", "Matthew B. Blaschko", "Shengen Yan", "Guohao Dai", "Huazhong Yang", "Yu Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-01", "url": "https://arxiv.org/abs/2407.00945", "pdf_url": "https://arxiv.org/pdf/2407.00945v1", "arxiv_id": "2407.00945", "doi": "10.48550/arXiv.2407.00945", "citation_count": 43, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/imagination-research/EEP", "venue": "arXiv.org", "quality_score": 0.4109}
{"id": "2104a6befdcec421b448efe2d30b47632ca836c3a3abf1d9628140e34924dd4e", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache Compression, But What Must We Give in Return? A Comprehensive Benchmark of Long Context Capable Approaches", "abstract": "Long context capability is a crucial competency for large language models (LLMs) as it mitigates the human struggle to digest long-form texts. This capability enables complex task-solving scenarios such as book summarization, code assistance, and many more tasks that are traditionally manpower-intensive. However, transformer-based LLMs face significant challenges with long context input due to the growing size of the KV cache and the intrinsic complexity of attending to extended inputs; where multiple schools of efficiency-driven approaches - such as KV cache quantization, token dropping, prompt compression, linear-time sequence models, and hybrid architectures - have been proposed to produce efficient yet long context-capable models. Despite these advancements, no existing work has comprehensively benchmarked these methods in a reasonably aligned environment. In this work, we fill this gap by providing a taxonomy of current methods and evaluating 10+ state-of-the-art approaches across seven categories of long context tasks. Our work reveals numerous previously unknown phenomena and offers insights - as well as a friendly workbench - for the future development of long context-capable LLMs. The source code is available at https://github.com/henryzhongsc/longctx_bench.", "authors": ["Jiayi Yuan", "Hongyi Liu", "Shaochen Zhong", "Yu-Neng Chuang", "Songchen Li", "Guanchu Wang", "Duy Le", "Hongye Jin", "Vipin Chaudhary", "Zhaozhuo Xu", "Zirui Liu", "Xia Hu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-01", "url": "https://arxiv.org/abs/2407.01527", "pdf_url": "https://arxiv.org/pdf/2407.01527v2", "arxiv_id": "2407.01527", "doi": "10.48550/arXiv.2407.01527", "citation_count": 45, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/henryzhongsc/longctx_bench", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4157}
{"id": "fba569d6051fa8eebcf16140703494a853d209d39170f08002899d5b2277aa78", "sources": ["arxiv", "semantic_scholar"], "title": "Joint Pruning and Channel-wise Mixed-Precision Quantization for Efficient Deep Neural Networks", "abstract": "The resource requirements of deep neural networks (DNNs) pose significant challenges to their deployment on edge devices. Common approaches to address this issue are pruning and mixed-precision quantization, which lead to latency and memory occupation improvements. These optimization techniques are usually applied independently. We propose a novel methodology to apply them jointly via a lightweight gradient-based search, and in a hardware-aware manner, greatly reducing the time required to generate Pareto-optimal DNNs in terms of accuracy versus cost (i.e., latency or memory). We test our approach on three edge-relevant benchmarks, namely CIFAR-10, Google Speech Commands, and Tiny ImageNet. When targeting the optimization of the memory footprint, we are able to achieve a size reduction of 47.50% and 69.54% at iso-accuracy with the baseline networks with all weights quantized at 8 and 2-bit, respectively. Our method surpasses a previous state-of-the-art approach with up to 56.17% size reduction at iso-accuracy. With respect to the sequential application of state-of-the-art pruning and mixed-precision optimizations, we obtain comparable or superior results, but with a significantly lowered training time. In addition, we show how well-tailored cost models can improve the cost versus accuracy trade-offs when targeting specific hardware for deployment.", "authors": ["Beatrice Alessandra Motetti", "Matteo Risso", "Alessio Burrello", "Enrico Macii", "Massimo Poncino", "Daniele Jahier Pagliari"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-01", "url": "https://arxiv.org/abs/2407.01054", "pdf_url": "https://arxiv.org/pdf/2407.01054v2", "arxiv_id": "2407.01054", "doi": "10.1109/TC.2024.3449084", "citation_count": 15, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE transactions on computers", "quality_score": 0.301}
{"id": "afb119d21924f4a027fbfc0ee0785364d9f673215631d49e306a27245e23f6c1", "sources": ["arxiv", "semantic_scholar"], "title": "SAML: Speaker Adaptive Mixture of LoRA Experts for End-to-End ASR", "abstract": "Mixture-of-experts (MoE) models have achieved excellent results in many tasks. However, conventional MoE models are often very large, making them challenging to deploy on resource-constrained edge devices. In this paper, we propose a novel speaker adaptive mixture of LoRA experts (SAML) approach, which uses low-rank adaptation (LoRA) modules as experts to reduce the number of trainable parameters in MoE. Specifically, SAML is applied to the quantised and personalised end-to-end automatic speech recognition models, which combines test-time speaker adaptation to improve the performance of heavily compressed models in speaker-specific scenarios. Experiments have been performed on the LibriSpeech and the TED-LIUM 3 corpora. Remarkably, with a 7x reduction in model size, 29.1% and 31.1% relative word error rate reductions were achieved on the quantised Whisper model and Conformer-based attention-based encoder-decoder ASR model respectively, comparing to the original full precision models.", "authors": ["Qiuming Zhao", "Guangzhi Sun", "Chao Zhang", "Mingxing Xu", "Thomas Fang Zheng"], "categories": ["cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-06-28", "url": "https://arxiv.org/abs/2406.19706", "pdf_url": "https://arxiv.org/pdf/2406.19706v1", "arxiv_id": "2406.19706", "doi": "10.48550/arXiv.2406.19706", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.2113}
{"id": "c0dc204b2fb9bea998104fe6065f4388406817d40f835df23df6e8623ace1477", "sources": ["arxiv", "semantic_scholar"], "title": "One Prompt is not Enough: Automated Construction of a Mixture-of-Expert Prompts", "abstract": "Large Language Models (LLMs) exhibit strong generalization capabilities to novel tasks when prompted with language instructions and in-context demos. Since this ability sensitively depends on the quality of prompts, various methods have been explored to automate the instruction design. While these methods demonstrated promising results, they also restricted the searched prompt to one instruction. Such simplification significantly limits their capacity, as a single demo-free instruction might not be able to cover the entire complex problem space of the targeted task. To alleviate this issue, we adopt the Mixture-of-Expert paradigm and divide the problem space into a set of sub-regions; Each sub-region is governed by a specialized expert, equipped with both an instruction and a set of demos. A two-phase process is developed to construct the specialized expert for each region: (1) demo assignment: Inspired by the theoretical connection between in-context learning and kernel regression, we group demos into experts based on their semantic similarity; (2) instruction assignment: A region-based joint search of an instruction per expert complements the demos assigned to it, yielding a synergistic effect. The resulting method, codenamed Mixture-of-Prompts (MoP), achieves an average win rate of 81% against prior arts across several major benchmarks.", "authors": ["Ruochen Wang", "Sohyun An", "Minhao Cheng", "Tianyi Zhou", "Sung Ju Hwang", "Cho-Jui Hsieh"], "categories": ["cs.AI", "cs.CL", "cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-28", "url": "https://arxiv.org/abs/2407.00256", "pdf_url": "https://arxiv.org/pdf/2407.00256v1", "arxiv_id": "2407.00256", "doi": "10.48550/arXiv.2407.00256", "citation_count": 24, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ruocwang/mixture-of-prompts", "venue": "International Conference on Machine Learning", "quality_score": 0.3495}
{"id": "104da362f5dfd9aa95b339524e4ff99acd4ab545a19379c61377e5f68685a42d", "sources": ["arxiv", "semantic_scholar"], "title": "InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management", "abstract": "Transformer-based large language models (LLMs) demonstrate impressive performance across various natural language processing tasks. Serving LLM inference for generating long contents, however, poses a challenge due to the enormous memory footprint of the transient state, known as the key-value (KV) cache, which scales with the sequence length and batch size. In this paper, we present InfiniGen, a novel KV cache management framework tailored for long-text generation, which synergistically works with modern offloading-based inference systems. InfiniGen leverages the key insight that a few important tokens that are essential for computing the subsequent attention layer in the Transformer can be speculated by performing a minimal rehearsal with the inputs of the current layer and part of the query weight and key cache of the subsequent layer. This allows us to prefetch only the essential KV cache entries (without fetching them all), thereby mitigating the fetch overhead from the host memory in offloading-based LLM serving systems. Our evaluation on several representative LLMs shows that InfiniGen improves the overall performance of a modern offloading-based system by up to 3.00x compared to prior KV cache management methods while offering substantially better model accuracy.", "authors": ["Wonbeom Lee", "Jungi Lee", "Junghwan Seo", "Jaewoong Sim"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-28", "url": "https://arxiv.org/abs/2406.19707", "pdf_url": "https://arxiv.org/pdf/2406.19707v1", "arxiv_id": "2406.19707", "doi": "10.48550/arXiv.2406.19707", "citation_count": 258, "influential_citation_count": 28, "has_code": false, "code_url": null, "venue": "USENIX Symposium on Operating Systems Design and Implementation", "quality_score": 0.7312}
{"id": "30c8d904e2f40e80763dc30a22497c1590e0b68b1367d3a49ed58a2b14a0e155", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of In-Context Experts Enhance LLMs' Long Context Awareness", "abstract": "Many studies have revealed that large language models (LLMs) exhibit uneven awareness of different contextual positions. Their limited context awareness can lead to overlooking critical information and subsequent task failures. While several approaches have been proposed to enhance LLMs' context awareness, achieving both effectiveness and efficiency remains challenging. In this paper, for LLMs utilizing RoPE as position embeddings, we introduce a novel method called \"Mixture of In-Context Experts\" (MoICE) to address this challenge. MoICE comprises two key components: a router integrated into each attention head within LLMs and a lightweight router-only training optimization strategy: (1) MoICE views each RoPE angle as an `in-context' expert, demonstrated to be capable of directing the attention of a head to specific contextual positions. Consequently, each attention head flexibly processes tokens using multiple RoPE angles dynamically selected by the router to attend to the needed positions. This approach mitigates the risk of overlooking essential contextual information. (2) The router-only training strategy entails freezing LLM parameters and exclusively updating routers for only a few steps. When applied to open-source LLMs including Llama and Mistral, MoICE surpasses prior methods across multiple tasks on long context understanding and generation, all while maintaining commendable inference efficiency.", "authors": ["Hongzhan Lin", "Ang Lv", "Yuhan Chen", "Chen Zhu", "Yang Song", "Hengshu Zhu", "Rui Yan"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-28", "url": "https://arxiv.org/abs/2406.19598", "pdf_url": "https://arxiv.org/pdf/2406.19598v2", "arxiv_id": "2406.19598", "doi": "10.48550/arXiv.2406.19598", "citation_count": 23, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3451}
{"id": "8f3808127e6e287172aed2974193556343d86749a23dc88b5fdd45101800d4c6", "sources": ["arxiv", "semantic_scholar"], "title": "Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model", "abstract": "The Mixture-of-Experts (MoE) has gained increasing attention in studying Large Vision-Language Models (LVLMs). It uses a sparse model to replace the dense model, achieving comparable performance while activating fewer parameters during inference, thus significantly reducing the inference cost. Existing MoE methods in LVLM encourage different experts to specialize in different tokens, and they usually employ a router to predict the routing of each token. However, the router is not optimized concerning distinct parameter optimization directions generated from tokens within an expert. This may lead to severe interference between tokens within an expert. To address this problem, we propose to use the token-level gradient analysis to Solving Token Gradient Conflict (STGC) in this paper. Specifically, we first use token-level gradients to identify conflicting tokens in experts. After that, we add a regularization loss tailored to encourage conflicting tokens routing from their current experts to other experts, for reducing interference between tokens within an expert. Our method can serve as a plug-in for diverse LVLM methods, and extensive experimental results demonstrate its effectiveness. The code will be publicly available at https://github.com/longrongyang/STGC.", "authors": ["Longrong Yang", "Dong Shen", "Chaoxiang Cai", "Fan Yang", "Tingting Gao", "Di Zhang", "Xi Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-28", "url": "https://arxiv.org/abs/2406.19905", "pdf_url": "https://arxiv.org/pdf/2406.19905v3", "arxiv_id": "2406.19905", "doi": "10.48550/arXiv.2406.19905", "citation_count": 11, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/longrongyang/STGC", "venue": "International Conference on Learning Representations", "quality_score": 0.2698}
{"id": "8cc55905b4ae81c8e7d8db923fb1f5c9b76ae107422a9b6a696bd17599c92d47", "sources": ["arxiv", "semantic_scholar"], "title": "LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference", "abstract": "Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs' KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge. In this work, we introduce LOOK-M, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. LOOK-M demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by 80% in some cases, it not only achieves up to 1.5x faster decoding but also maintains or even enhances performance across a variety of long context multimodal tasks.", "authors": ["Zhongwei Wan", "Ziang Wu", "Che Liu", "Jinfa Huang", "Zhihong Zhu", "Peng Jin", "Longyue Wang", "Li Yuan"], "categories": ["cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-26", "url": "https://arxiv.org/abs/2406.18139", "pdf_url": "https://arxiv.org/pdf/2406.18139v1", "arxiv_id": "2406.18139", "doi": "10.48550/arXiv.2406.18139", "citation_count": 98, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4989}
{"id": "edbcea7aad07e43aea079497303aae1e574fdcb60eff57d226b6c53c3cb7c002", "sources": ["arxiv", "semantic_scholar"], "title": "A Survey on Mixture of Experts in Large Language Models", "abstract": "Large language models (LLMs) have garnered unprecedented advancements across diverse fields, ranging from natural language processing to computer vision and beyond. The prowess of LLMs is underpinned by their substantial model size, extensive and diverse datasets, and the vast computational power harnessed during training, all of which contribute to the emergent abilities of LLMs (e.g., in-context learning) that are not present in small models. Within this context, the mixture of experts (MoE) has emerged as an effective method for substantially scaling up model capacity with minimal computation overhead, gaining significant attention from academia and industry. Despite its growing prevalence, there lacks a systematic and comprehensive review of the literature on MoE. This survey seeks to bridge that gap, serving as an essential resource for researchers delving into the intricacies of MoE. We first briefly introduce the structure of the MoE layer, followed by proposing a new taxonomy of MoE. Next, we overview the core designs for various MoE models including both algorithmic and systemic aspects, alongside collections of available open-source implementations, hyperparameter configurations and empirical evaluations. Furthermore, we delineate the multifaceted applications of MoE in practice, and outline some potential directions for future research. To facilitate ongoing updates and the sharing of cutting-edge advances in MoE research, we have established a resource repository at https://github.com/withinmiaov/A-Survey-on-Mixture-of-Experts-in-LLMs.", "authors": ["Weilin Cai", "Juyong Jiang", "Fan Wang", "Jing Tang", "Sunghun Kim", "Jiayi Huang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-26", "url": "https://arxiv.org/abs/2407.06204", "pdf_url": "https://arxiv.org/pdf/2407.06204v3", "arxiv_id": "2407.06204", "doi": "10.1109/TKDE.2025.3554028", "citation_count": 318, "influential_citation_count": 19, "has_code": true, "code_url": "https://github.com/withinmiaov/A-Survey-on-Mixture-of-Experts-in-LLMs", "venue": "IEEE Transactions on Knowledge and Data Engineering", "quality_score": 0.6505}
{"id": "ac3eb6b14947c2dedc6da780345f8878e979d127a32690f8e689ebfbfe7fe937", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts in a Mixture of RL settings", "abstract": "Mixtures of Experts (MoEs) have gained prominence in (self-)supervised learning due to their enhanced inference efficiency, adaptability to distributed training, and modularity. Previous research has illustrated that MoEs can significantly boost Deep Reinforcement Learning (DRL) performance by expanding the network's parameter count while reducing dormant neurons, thereby enhancing the model's learning capacity and ability to deal with non-stationarity. In this work, we shed more light on MoEs' ability to deal with non-stationarity and investigate MoEs in DRL settings with \"amplified\" non-stationarity via multi-task training, providing further evidence that MoEs improve learning capacity. In contrast to previous work, our multi-task results allow us to better understand the underlying causes for the beneficial effect of MoE in DRL training, the impact of the various MoE components, and insights into how best to incorporate them in actor-critic-based DRL networks. Finally, we also confirm results from previous work.", "authors": ["Timon Willi", "Johan Obando-Ceron", "Jakob Foerster", "Karolina Dziugaite", "Pablo Samuel Castro"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-26", "url": "https://arxiv.org/abs/2406.18420", "pdf_url": "https://arxiv.org/pdf/2406.18420v1", "arxiv_id": "2406.18420", "doi": "10.48550/arXiv.2406.18420", "citation_count": 21, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3356}
{"id": "3ab03e1ae17f46cf307c4e93fa09c516f4637f53f8c88ff31c8bf754531f83e6", "sources": ["arxiv", "semantic_scholar"], "title": "A Closer Look into Mixture-of-Experts in Large Language Models", "abstract": "Mixture-of-experts (MoE) is gaining increasing attention due to its unique properties and remarkable performance, especially for language tasks. By sparsely activating a subset of parameters for each token, MoE architecture could increase the model size without sacrificing computational efficiency, achieving a better trade-off between performance and training costs. However, the underlying mechanism of MoE still lacks further exploration, and its modularization degree remains questionable. In this paper, we make an initial attempt to understand the inner workings of MoE-based large language models. Concretely, we comprehensively study the parametric and behavioral features of three popular MoE-based models and reveal some intriguing observations, including 1) Neurons act like fine-grained experts; 2) The router of MoE usually selects experts with larger output norms; 3) The expert diversity increases as the layer increases, while the last layer is an outlier, which is further validated by an initial experiment. Based on the observations, we also provide suggestions for a broad spectrum of MoE practitioners, such as router design and expert allocation. We hope this work could shed light on future research on the MoE framework and other modular architectures. Code is available at https://github.com/kamanphoebe/Look-into-MoEs.", "authors": ["Ka Man Lo", "Zeyu Huang", "Zihan Qiu", "Zili Wang", "Jie Fu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-26", "url": "https://arxiv.org/abs/2406.18219", "pdf_url": "https://arxiv.org/pdf/2406.18219v3", "arxiv_id": "2406.18219", "doi": "10.48550/arXiv.2406.18219", "citation_count": 46, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/kamanphoebe/Look-into-MoEs", "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.418}
{"id": "82cb56d68d1cd70ce96072248e639a34c655eca5eb01c8271a3a0b1aee5fa87c", "sources": ["arxiv", "semantic_scholar"], "title": "MoESD: Mixture of Experts Stable Diffusion to Mitigate Gender Bias", "abstract": "Text-to-image models are known to propagate social biases. For example, when prompted to generate images of people in certain professions, these models tend to systematically generate specific genders or ethnicities. In this paper, we show that this bias is already present in the text encoder of the model and introduce a Mixture-of-Experts approach by identifying text-encoded bias in the latent space and then creating a Bias-Identification Gate mechanism. More specifically, we propose MoESD (Mixture of Experts Stable Diffusion) with BiAs (Bias Adapters) to mitigate gender bias in text-to-image models. We also demonstrate that introducing an arbitrary special token to the prompt is essential during the mitigation process. With experiments focusing on gender bias, we show that our approach successfully mitigates gender bias while maintaining image quality.", "authors": ["Guorun Wang", "Lucia Specia"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-25", "url": "https://arxiv.org/abs/2407.11002", "pdf_url": "https://arxiv.org/pdf/2407.11002v2", "arxiv_id": "2407.11002", "doi": "10.48550/arXiv.2407.11002", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "c526a36817e15201c8db51e47af1d67ec710d401cdd2f29d61044ea605003487", "sources": ["arxiv", "semantic_scholar"], "title": "Retrieval-Augmented Mixture of LoRA Experts for Uploadable Machine Learning", "abstract": "Low-Rank Adaptation (LoRA) offers an efficient way to fine-tune large language models (LLMs). Its modular and plug-and-play nature allows the integration of various domain-specific LoRAs, enhancing LLM capabilities. Open-source platforms like Huggingface and Modelscope have introduced a new computational paradigm, Uploadable Machine Learning (UML). In UML, contributors use decentralized data to train specialized adapters, which are then uploaded to a central platform to improve LLMs. This platform uses these domain-specific adapters to handle mixed-task requests requiring personalized service. Previous research on LoRA composition either focuses on specific tasks or fixes the LoRA selection during training. However, in UML, the pool of LoRAs is dynamically updated with new uploads, requiring a generalizable selection mechanism for unseen LoRAs. Additionally, the mixed-task nature of downstream requests necessitates personalized services. To address these challenges, we propose Retrieval-Augmented Mixture of LoRA Experts (RAMoLE), a framework that adaptively retrieves and composes multiple LoRAs based on input prompts. RAMoLE has three main components: LoraRetriever for identifying and retrieving relevant LoRAs, an on-the-fly MoLE mechanism for coordinating the retrieved LoRAs, and efficient batch inference for handling heterogeneous requests. Experimental results show that RAMoLE consistently outperforms baselines, highlighting its effectiveness and scalability.", "authors": ["Ziyu Zhao", "Leilei Gan", "Guoyin Wang", "Yuwei Hu", "Tao Shen", "Hongxia Yang", "Kun Kuang", "Fei Wu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-24", "url": "https://arxiv.org/abs/2406.16989", "pdf_url": "https://arxiv.org/pdf/2406.16989v2", "arxiv_id": "2406.16989", "doi": "10.48550/arXiv.2406.16989", "citation_count": 17, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3138}
{"id": "d4839f9e2af992f2af6df0614e1fa5f315ad7c3107edb7c36a244cce3d18e0e9", "sources": ["arxiv", "semantic_scholar"], "title": "Finding Transformer Circuits with Edge Pruning", "abstract": "The path to interpreting a language model often proceeds via analysis of circuits -- sparse computational subgraphs of the model that capture specific aspects of its behavior. Recent work has automated the task of discovering circuits. Yet, these methods have practical limitations, as they rely either on inefficient search algorithms or inaccurate approximations. In this paper, we frame automated circuit discovery as an optimization problem and propose *Edge Pruning* as an effective and scalable solution. Edge Pruning leverages gradient-based pruning techniques, but instead of removing neurons or components, it prunes the \\emph{edges} between components. Our method finds circuits in GPT-2 that use less than half the number of edges compared to circuits found by previous methods while being equally faithful to the full model predictions on standard circuit-finding tasks. Edge Pruning is efficient even with as many as 100K examples, outperforming previous methods in speed and producing substantially better circuits. It also perfectly recovers the ground-truth circuits in two models compiled with Tracr. Thanks to its efficiency, we scale Edge Pruning to CodeLlama-13B, a model over 100x the scale that prior methods operate on. We use this setting for a case study comparing the mechanisms behind instruction prompting and in-context learning. We find two circuits with more than 99.96% sparsity that match the performance of the full model and reveal that the mechanisms in the two settings overlap substantially. Our case study shows that Edge Pruning is a practical and scalable tool for interpretability and sheds light on behaviors that only emerge in large models.", "authors": ["Adithya Bhaskar", "Alexander Wettig", "Dan Friedman", "Danqi Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-24", "url": "https://arxiv.org/abs/2406.16778", "pdf_url": "https://arxiv.org/pdf/2406.16778v3", "arxiv_id": "2406.16778", "doi": "10.48550/arXiv.2406.16778", "citation_count": 49, "influential_citation_count": 10, "has_code": true, "code_url": "https://github.com/princeton-nlp/Edge-Pruning", "venue": "Neural Information Processing Systems", "quality_score": 0.5207}
{"id": "32b159f5c624caad98c7e7a64c6b583f36e2b1647c4faa77eb30d4518bdc4067", "sources": ["arxiv", "semantic_scholar"], "title": "Compensate Quantization Errors: Make Weights Hierarchical to Compensate Each Other", "abstract": "Emergent Large Language Models (LLMs) use their extraordinary performance and powerful deduction capacity to discern from traditional language models. However, the expenses of computational resources and storage for these LLMs are stunning, quantization then arises as a trending conversation. To address accuracy decay caused by quantization, two streams of works in post-training quantization methods stand out. One uses other weights to compensate existing quantization error, while the other transfers the quantization difficulty to other parts in the model. Combining both merits, we introduce Learnable Singular value Increment (LSI) as an advanced solution. LSI uses Singular Value Decomposition to extract singular values of the weights and make them learnable to help weights compensate each other conditioned on activation. Incorporating LSI with existing techniques, we achieve state-of-the-art performance in diverse quantization settings, no matter in weight-only, weight-activation or extremely low bit scenarios. By unleashing the potential of LSI, efficient finetuning on quantized model is no longer a prohibitive problem.", "authors": ["Yifei Gao", "Jie Ou", "Lei Wang", "Yuting Xiao", "Zhiyuan Xiang", "Ruiting Dai", "Jun Cheng"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-24", "url": "https://arxiv.org/abs/2406.16299", "pdf_url": "https://arxiv.org/pdf/2406.16299v1", "arxiv_id": "2406.16299", "doi": "10.48550/arXiv.2406.16299", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "24552bab1a8452c3e4cf8f8278d7271727d0de740358c299b215214f3eafa614", "sources": ["arxiv", "semantic_scholar"], "title": "Peirce in the Machine: How Mixture of Experts Models Perform Hypothesis Construction", "abstract": "Mixture of experts is a prediction aggregation method in machine learning that aggregates the predictions of specialized experts. This method often outperforms Bayesian methods despite the Bayesian having stronger inductive guarantees. We argue that this is due to the greater functional capacity of mixture of experts. We prove that in a limiting case of mixture of experts will have greater capacity than equivalent Bayesian methods, which we vouchsafe through experiments on non-limiting cases. Finally, we conclude that mixture of experts is a type of abductive reasoning in the Peircian sense of hypothesis construction.", "authors": ["Bruce Rushing"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-24", "url": "https://arxiv.org/abs/2406.17150", "pdf_url": "https://arxiv.org/pdf/2406.17150v1", "arxiv_id": "2406.17150", "doi": "10.1017/psa.2025.10159", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Philosophia Scientiæ", "quality_score": 0.1193}
{"id": "40771f2fabd855de2c1fda0819457331c4dc5ea4a890dfc6e2722c7f6374f1fd", "sources": ["arxiv", "semantic_scholar"], "title": "Training-Free Exponential Context Extension via Cascading KV Cache", "abstract": "The transformer's context window is vital for tasks such as few-shot learning and conditional generation as it preserves previous tokens for active memory. However, as the context lengths increase, the computational costs grow quadratically, hindering the deployment of large language models (LLMs) in real-world, long sequence scenarios. Although some recent key-value caching (KV Cache) methods offer linear inference complexity, they naively manage the stored context, prematurely evicting tokens and losing valuable information. Moreover, they lack an optimized prefill/prompt stage strategy, resulting in higher latency than even quadratic attention for realistic context sizes. In response, we introduce a novel mechanism that leverages cascading sub-cache buffers to selectively retain the most relevant tokens, enabling the model to maintain longer context histories without increasing the cache size. Our approach outperforms linear caching baselines across key benchmarks, including streaming perplexity, question answering, book summarization, and passkey retrieval, where it retains better retrieval accuracy at 1M tokens after four doublings of the cache size of 65K. Additionally, our method reduces prefill stage latency by a factor of 6.8 when compared to flash attention on 1M tokens. These innovations not only enhance the computational efficiency of LLMs but also pave the way for their effective deployment in resource-constrained environments, enabling large-scale, real-time applications with significantly reduced latency.", "authors": ["Jeffrey Willette", "Heejun Lee", "Youngwan Lee", "Myeongjae Jeon", "Sung Ju Hwang"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-24", "url": "https://arxiv.org/abs/2406.17808", "pdf_url": "https://arxiv.org/pdf/2406.17808v4", "arxiv_id": "2406.17808", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.1747}
{"id": "7a21c27d27477dafac89b3fc37e67ec46b30a9ed2744915829b12105f630108d", "sources": ["arxiv", "semantic_scholar"], "title": "Theory on Mixture-of-Experts in Continual Learning", "abstract": "Continual learning (CL) has garnered significant attention because of its ability to adapt to new tasks that arrive over time. Catastrophic forgetting (of old tasks) has been identified as a major issue in CL, as the model adapts to new tasks. The Mixture-of-Experts (MoE) model has recently been shown to effectively mitigate catastrophic forgetting in CL, by employing a gating network to sparsify and distribute diverse tasks among multiple experts. However, there is a lack of theoretical analysis of MoE and its impact on the learning performance in CL. This paper provides the first theoretical results to characterize the impact of MoE in CL via the lens of overparameterized linear regression tasks. We establish the benefit of MoE over a single expert by proving that the MoE model can diversify its experts to specialize in different tasks, while its router learns to select the right expert for each task and balance the loads across all experts. Our study further suggests an intriguing fact that the MoE in CL needs to terminate the update of the gating network after sufficient training rounds to attain system convergence, which is not needed in the existing MoE studies that do not consider the continual task arrival. Furthermore, we provide explicit expressions for the expected forgetting and overall generalization error to characterize the benefit of MoE in the learning performance in CL. Interestingly, adding more experts requires additional rounds before convergence, which may not enhance the learning performance. Finally, we conduct experiments on both synthetic and real datasets to extend these insights from linear models to deep neural networks (DNNs), which also shed light on the practical algorithm design for MoE in CL.", "authors": ["Hongbo Li", "Sen Lin", "Lingjie Duan", "Yingbin Liang", "Ness B. Shroff"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-24", "url": "https://arxiv.org/abs/2406.16437", "pdf_url": "https://arxiv.org/pdf/2406.16437v3", "arxiv_id": "2406.16437", "doi": "10.48550/arXiv.2406.16437", "citation_count": 55, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.437}
{"id": "28ac9e06cc98f78b807f875a709fa2d89b54a7ae19bc826b3c190659a0516e29", "sources": ["arxiv", "semantic_scholar"], "title": "OTCE: Hybrid SSM and Attention with Cross Domain Mixture of Experts to construct Observer-Thinker-Conceiver-Expresser", "abstract": "Recent research has shown that combining Mamba with Transformer architecture, which has selective state space and quadratic self-attention mechanism, outperforms using Mamba or Transformer architecture alone in language modeling tasks. The quadratic self-attention mechanism effectively alleviates the shortcomings of selective state space in handling long-term dependencies of any element in the sequence. We propose a position information injection method that connects the selective state space model with the quadratic attention, and integrates these two architectures with hybrid experts with cross-sharing domains, so that we can enjoy the advantages of both. We design a new architecture with a more biomimetic idea: Observer-Thinker-Conceiver-Expresser (OTCE), which can compete with well-known medium-scale open-source language models on a small scale in language modeling tasks.", "authors": ["Jingze Shi", "Ting Xie", "Bingheng Wu", "Chunjun Zheng", "Kai Wang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-24", "url": "https://arxiv.org/abs/2406.16495", "pdf_url": "https://arxiv.org/pdf/2406.16495v3", "arxiv_id": "2406.16495", "doi": "10.48550/arXiv.2406.16495", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "1557a998d68c3da3a8b3467ab3c7db47c68d5f0aa3f9e19a9b1e026744f227b6", "sources": ["arxiv", "semantic_scholar"], "title": "Elliptical Attention", "abstract": "Pairwise dot-product self-attention is key to the success of transformers that achieve state-of-the-art performance across a variety of applications in language and vision. This dot-product self-attention computes attention weights among the input tokens using Euclidean distance, which makes the model prone to representation collapse and vulnerable to contaminated samples. In this paper, we propose using a Mahalanobis distance metric for computing the attention weights to stretch the underlying feature space in directions of high contextual relevance. In particular, we define a hyper-ellipsoidal neighborhood around each query to increase the attention weights of the tokens lying in the contextually important directions. We term this novel class of attention Elliptical Attention. Our Elliptical Attention provides two benefits: 1) reducing representation collapse and 2) enhancing the model's robustness as Elliptical Attention pays more attention to contextually relevant information rather than focusing on some small subset of informative features. We empirically demonstrate the advantages of Elliptical Attention over the baseline dot-product attention and state-of-the-art attention methods on various practical tasks, including object classification, image segmentation, and language modeling across different data modalities.", "authors": ["Stefan K. Nielsen", "Laziz U. Abdullaev", "Rachel S. Y. Teo", "Tan M. Nguyen"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-19", "url": "https://arxiv.org/abs/2406.13770", "pdf_url": "https://arxiv.org/pdf/2406.13770v2", "arxiv_id": "2406.13770", "doi": "10.48550/arXiv.2406.13770", "citation_count": 10, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/stefvk/Elliptical-Attention", "venue": "Neural Information Processing Systems", "quality_score": 0.2603}
{"id": "f29ec0e5eefa07eddb35d026cea6bfbee0b3f1e040819b8b2277cb73f116de18", "sources": ["arxiv", "semantic_scholar"], "title": "AdaMoE: Token-Adaptive Routing with Null Experts for Mixture-of-Experts Language Models", "abstract": "Mixture of experts (MoE) has become the standard for constructing production-level large language models (LLMs) due to its promise to boost model capacity without causing significant overheads. Nevertheless, existing MoE methods usually enforce a constant top-k routing for all tokens, which is arguably restrictive because various tokens (e.g., \"<EOS>\" vs. \"apple\") may require various numbers of experts for feature abstraction. Lifting such a constraint can help make the most of limited resources and unleash the potential of the model for downstream tasks. In this sense, we introduce AdaMoE to realize token-adaptive routing for MoE, where different tokens are permitted to select a various number of experts. AdaMoE makes minimal modifications to the vanilla MoE with top-k routing -- it simply introduces a fixed number of null experts, which do not consume any FLOPs, to the expert set and increases the value of k. AdaMoE does not force each token to occupy a fixed number of null experts but ensures the average usage of the null experts with a load-balancing loss, leading to an adaptive number of null/true experts used by each token. AdaMoE exhibits a strong resemblance to MoEs with expert choice routing while allowing for trivial auto-regressive modeling. AdaMoE is easy to implement and can be effectively applied to pre-trained (MoE-)LLMs. Extensive studies show that AdaMoE can reduce average expert load (FLOPs) while achieving superior performance. For example, on the ARC-C dataset, applying our method to fine-tuning Mixtral-8x7B can reduce FLOPs by 14.5% while increasing accuracy by 1.69%.", "authors": ["Zihao Zeng", "Yibo Miao", "Hongcheng Gao", "Hao Zhang", "Zhijie Deng"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-19", "url": "https://arxiv.org/abs/2406.13233", "pdf_url": "https://arxiv.org/pdf/2406.13233v2", "arxiv_id": "2406.13233", "doi": "10.48550/arXiv.2406.13233", "citation_count": 41, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4225}
{"id": "9a6b0a116aec34f26818ac2468f51565f4f3df4108f9adf664d55979e25d6011", "sources": ["arxiv", "semantic_scholar"], "title": "BoA: Attention-aware Post-training Quantization without Backpropagation", "abstract": "Post-training quantization (PTQ) is a promising solution for deploying large language models (LLMs) on resource-constrained devices. Early methods developed for small-scale networks, such as ResNet, rely on gradient-based optimization, which becomes impractical for hyper-scale LLMs with billions of parameters. While recently proposed backpropagation-free or transformation-based methods alleviate this issue, they ignore inter-layer interactions or use the naive nearest-rounding-based quantized weight assignment to save the heavy computational cost of weight optimization. In this paper, we introduce a novel backpropagation-free PTQ algorithm that optimizes quantized weights by considering inter-layer dependencies. The key innovation is the development of attention-aware Hessian matrices that capture inter-layer interactions within the attention module. Extensive experiments demonstrate that our approach not only outperforms existing weight quantization methods but also shows good synergy with conventional methods to suppress activation outliers, leading to state-of-the-art weight-activation quantization performance. The code will be available at https://github.com/SamsungLabs/BoA.", "authors": ["Junhan Kim", "Ho-young Kim", "Eulrang Cho", "Chungman Lee", "Joonyoung Kim", "Yongkweon Jeon"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-19", "url": "https://arxiv.org/abs/2406.13474", "pdf_url": "https://arxiv.org/pdf/2406.13474v3", "arxiv_id": "2406.13474", "doi": "10.48550/arXiv.2406.13474", "citation_count": 9, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/SamsungLabs/BoA", "venue": "International Conference on Machine Learning", "quality_score": 0.25}
{"id": "502d3528493b87bcaf26ea7755b4da94e67ec684f3726cda05c2c97f76149da9", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Score is not All You Need for Token Importance Indicator in KV Cache Reduction: Value Also Matters", "abstract": "Scaling the context size of large language models (LLMs) enables them to perform various new tasks, e.g., book summarization. However, the memory cost of the Key and Value (KV) cache in attention significantly limits the practical applications of LLMs. Recent works have explored token pruning for KV cache reduction in LLMs, relying solely on attention scores as a token importance indicator. However, our investigation into value vector norms revealed a notably non-uniform pattern questioning their reliance only on attention scores. Inspired by this, we propose a new method: Value-Aware Token Pruning (VATP) which uses both attention scores and the $ \\ell_{1} $ norm of value vectors to evaluate token importance. Extensive experiments on LLaMA2-7B-chat and Vicuna-v1.5-7B across 16 LongBench tasks demonstrate that VATP outperforms attention-score-only baselines in over 12 tasks, confirming the effectiveness of incorporating value vector norms into token importance evaluation of LLMs.", "authors": ["Zhiyu Guo", "Hidetaka Kamigaito", "Taro Watanabe"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-18", "url": "https://arxiv.org/abs/2406.12335", "pdf_url": "https://arxiv.org/pdf/2406.12335v2", "arxiv_id": "2406.12335", "doi": "10.48550/arXiv.2406.12335", "citation_count": 70, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4628}
{"id": "e64172c0ecd29f450f5eedb25a193c350bffa279bab4bf6001062fddbdbcac19", "sources": ["arxiv", "semantic_scholar"], "title": "A Comprehensive Study of Structural Pruning for Vision Models", "abstract": "Structural pruning has emerged as a promising approach for producing more efficient models. Nevertheless, the community suffers from a lack of standardized benchmarks and metrics, leaving the progress in this area not fully comprehended. To fill this gap, we present the first comprehensive benchmark, termed PruningBench, for structural pruning. PruningBench showcases the following three characteristics: 1) PruningBench employs a unified and consistent framework for evaluating the effectiveness of diverse structural pruning techniques; 2) PruningBench systematically evaluates 16 existing pruning methods, encompassing a wide array of models (e.g., CNNs and ViTs) and tasks (e.g., classification and detection); 3) PruningBench provides easily implementable interfaces to facilitate the implementation of future pruning methods, and enables the subsequent researchers to incorporate their work into our leaderboards. We provide an online pruning platform for customizing pruning tasks and reproducing all results in this paper. Leaderboard results can also be available.", "authors": ["Changhao Li", "Haoling Li", "Mengqi Xue", "Gongfan Fang", "Sheng Zhou", "Zunlei Feng", "Huiqiong Wang", "Mingli Song", "Jie Song"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-18", "url": "https://arxiv.org/abs/2406.12315", "pdf_url": "https://arxiv.org/pdf/2406.12315v7", "arxiv_id": "2406.12315", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "862388236a2b0ae3754fb3ec73322fa9fd11ec8046adae655f9da7dc76ad125f", "sources": ["arxiv", "semantic_scholar"], "title": "Bayesian-LoRA: LoRA based Parameter Efficient Fine-Tuning using Optimal Quantization levels and Rank Values trough Differentiable Bayesian Gates", "abstract": "It is a common practice in natural language processing to pre-train a single model on a general domain and then fine-tune it for downstream tasks. However, when it comes to Large Language Models, fine-tuning the entire model can be computationally expensive, resulting in very intensive energy consumption. As a result, several Parameter Efficient Fine-Tuning (PEFT) approaches were recently proposed. One of the most popular approaches is low-rank adaptation (LoRA), where the key insight is decomposing the update weights of the pre-trained model into two low-rank matrices. However, the proposed approaches either use the same rank value across all different weight matrices, which has been shown to be a sub-optimal choice, or do not use any quantization technique, one of the most important factors when it comes to a model's energy consumption. In this work, we propose Bayesian-LoRA which approaches low-rank adaptation and quantization from a Bayesian perspective by employing a prior distribution on both quantization levels and rank values. As a result, B-LoRA is able to fine-tune a pre-trained model on a specific downstream task, finding the optimal rank values and quantization levels for every low-rank matrix. We validate the proposed model by fine-tuning a pre-trained DeBERTaV3 on the GLUE benchmark. Moreover, we compare it to relevant baselines and present both qualitative and quantitative results, showing how the proposed approach is able to learn optimal-rank quantized matrices. B-LoRA performs on par with or better than the baselines while reducing the total number of bit operations by roughly 70% compared to the baseline methods.", "authors": ["Cristian Meo", "Ksenia Sycheva", "Anirudh Goyal", "Justin Dauwels"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-18", "url": "https://arxiv.org/abs/2406.13046", "pdf_url": "https://arxiv.org/pdf/2406.13046v3", "arxiv_id": "2406.13046", "doi": "10.48550/arXiv.2406.13046", "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2865}
{"id": "bd29c2619357dd58bf719dc63b6aec3b816a2415cc236d8adf8583bc4c30b7f2", "sources": ["arxiv", "semantic_scholar"], "title": "P-React: Synthesizing Topic-Adaptive Reactions of Personality Traits via Mixture of Specialized LoRA Experts", "abstract": "Personalized large language models (LLMs) have attracted great attention in many applications, such as emotional support and role-playing. However, existing works primarily focus on modeling explicit character profiles, while ignoring the underlying personality traits that truly shape behaviors and decision-making, hampering the development of more anthropomorphic and psychologically-grounded AI systems. In this paper, we explore the modeling of Big Five personality traits, which is the most widely used trait theory in psychology, and propose P-React, a mixture of experts (MoE)-based personalized LLM. Particularly, we integrate a Personality Specialization Loss (PSL) to better capture individual trait expressions, providing a more nuanced and psychologically grounded personality simulacrum. To facilitate research in this field, we curate OCEAN-Chat, a high-quality, human-verified dataset designed to train LLMs in expressing personality traits across diverse topics. Extensive experiments demonstrate the effectiveness of P-React in maintaining consistent and real personality.", "authors": ["Yuhao Dan", "Jie Zhou", "Qin Chen", "Junfeng Tian", "Liang He"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-18", "url": "https://arxiv.org/abs/2406.12548", "pdf_url": "https://arxiv.org/pdf/2406.12548v3", "arxiv_id": "2406.12548", "doi": "10.18653/v1/2025.findings-acl.328", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2113}
{"id": "5bb168dc0f9400d9fa3b39093b2f42482bac12067f73f4a8a9bf6754da5a0bfb", "sources": ["arxiv", "semantic_scholar"], "title": "Variational Distillation of Diffusion Policies into Mixture of Experts", "abstract": "This work introduces Variational Diffusion Distillation (VDD), a novel method that distills denoising diffusion policies into Mixtures of Experts (MoE) through variational inference. Diffusion Models are the current state-of-the-art in generative modeling due to their exceptional ability to accurately learn and represent complex, multi-modal distributions. This ability allows Diffusion Models to replicate the inherent diversity in human behavior, making them the preferred models in behavior learning such as Learning from Human Demonstrations (LfD). However, diffusion models come with some drawbacks, including the intractability of likelihoods and long inference times due to their iterative sampling process. The inference times, in particular, pose a significant challenge to real-time applications such as robot control. In contrast, MoEs effectively address the aforementioned issues while retaining the ability to represent complex distributions but are notoriously difficult to train. VDD is the first method that distills pre-trained diffusion models into MoE models, and hence, combines the expressiveness of Diffusion Models with the benefits of Mixture Models. Specifically, VDD leverages a decompositional upper bound of the variational objective that allows the training of each expert separately, resulting in a robust optimization scheme for MoEs. VDD demonstrates across nine complex behavior learning tasks, that it is able to: i) accurately distill complex distributions learned by the diffusion model, ii) outperform existing state-of-the-art distillation methods, and iii) surpass conventional methods for training MoE.", "authors": ["Hongyi Zhou", "Denis Blessing", "Ge Li", "Onur Celik", "Xiaogang Jia", "Gerhard Neumann", "Rudolf Lioutikov"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-18", "url": "https://arxiv.org/abs/2406.12538", "pdf_url": "https://arxiv.org/pdf/2406.12538v2", "arxiv_id": "2406.12538", "doi": "10.48550/arXiv.2406.12538", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.2785}
{"id": "840482196a57d92dd23e7144aeb6992f7200e19dbcaf887a1cec129c0acac604", "sources": ["arxiv", "semantic_scholar"], "title": "A Simple and Effective $L_2$ Norm-Based Strategy for KV Cache Compression", "abstract": "The deployment of large language models (LLMs) is often hindered by the extensive memory requirements of the Key-Value (KV) cache, especially as context lengths increase. Existing approaches to reduce the KV cache size involve either fine-tuning the model to learn a compression strategy or leveraging attention scores to reduce the sequence length. We analyse the attention distributions in decoder-only Transformers-based models and observe that attention allocation patterns stay consistent across most layers. Surprisingly, we find a clear correlation between the $L_2$ and the attention scores over cached KV pairs, where a low $L_2$ of a key embedding usually leads to a high attention score during decoding. This finding indicates that the influence of a KV pair is potentially determined by the key embedding itself before being queried. Based on this observation, we compress the KV cache based on the $L_2$ of key embeddings. Our experimental results show that this simple strategy can reduce the KV cache size by 50% on language modelling and needle-in-a-haystack tasks and 90% on passkey retrieval tasks without losing accuracy. Moreover, without relying on the attention scores, this approach remains compatible with FlashAttention, enabling broader applicability.", "authors": ["Alessio Devoto", "Yu Zhao", "Simone Scardapane", "Pasquale Minervini"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-17", "url": "https://arxiv.org/abs/2406.11430", "pdf_url": "https://arxiv.org/pdf/2406.11430v4", "arxiv_id": "2406.11430", "doi": "10.48550/arXiv.2406.11430", "citation_count": 76, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4716}
{"id": "879c6d3f38f6609173b36dfec9138e80ebc44b304708e9538f37877b4b6a61b5", "sources": ["arxiv", "semantic_scholar"], "title": "Graph Knowledge Distillation to Mixture of Experts", "abstract": "In terms of accuracy, Graph Neural Networks (GNNs) are the best architectural choice for the node classification task. Their drawback in real-world deployment is the latency that emerges from the neighbourhood processing operation. One solution to the latency issue is to perform knowledge distillation from a trained GNN to a Multi-Layer Perceptron (MLP), where the MLP processes only the features of the node being classified (and possibly some pre-computed structural information). However, the performance of such MLPs in both transductive and inductive settings remains inconsistent for existing knowledge distillation techniques. We propose to address the performance concerns by using a specially-designed student model instead of an MLP. Our model, named Routing-by-Memory (RbM), is a form of Mixture-of-Experts (MoE), with a design that enforces expert specialization. By encouraging each expert to specialize on a certain region on the hidden representation space, we demonstrate experimentally that it is possible to derive considerably more consistent performance across multiple datasets. Code available at https://github.com/Rufaim/routing-by-memory.", "authors": ["Pavel Rumiantsev", "Mark Coates"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-17", "url": "https://arxiv.org/abs/2406.11919", "pdf_url": "https://arxiv.org/pdf/2406.11919v2", "arxiv_id": "2406.11919", "doi": "10.48550/arXiv.2406.11919", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Rufaim/routing-by-memory", "venue": null, "quality_score": 0.1193}
{"id": "4d4adaf0fea459330849e892166531c86068e8ac44fb8dcf59fbd077e99b2577", "sources": ["arxiv", "semantic_scholar"], "title": "Promoting Data and Model Privacy in Federated Learning through Quantized LoRA", "abstract": "Conventional federated learning primarily aims to secure the privacy of data distributed across multiple edge devices, with the global model dispatched to edge devices for parameter updates during the learning process. However, the development of large language models (LLMs) requires substantial data and computational resources, rendering them valuable intellectual properties for their developers and owners. To establish a mechanism that protects both data and model privacy in a federated learning context, we introduce a method that just needs to distribute a quantized version of the model's parameters during training. This method enables accurate gradient estimations for parameter updates while preventing clients from accessing a model whose performance is comparable to the centrally hosted one. Moreover, we combine this quantization strategy with LoRA, a popular and parameter-efficient fine-tuning method, to significantly reduce communication costs in federated learning. The proposed framework, named \\textsc{FedLPP}, successfully ensures both data and model privacy in the federated learning context. Additionally, the learned central model exhibits good generalization and can be trained in a resource-efficient manner.", "authors": ["JianHao Zhu", "Changze Lv", "Xiaohua Wang", "Muling Wu", "Wenhao Liu", "Tianlong Li", "Zixuan Ling", "Cenyuan Zhang", "Xiaoqing Zheng", "Xuanjing Huang"], "categories": ["cs.LG", "cs.CL", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-16", "url": "https://arxiv.org/abs/2406.10976", "pdf_url": "https://arxiv.org/pdf/2406.10976v1", "arxiv_id": "2406.10976", "doi": "10.48550/arXiv.2406.10976", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.294}
{"id": "3c30a0da647c02cb6a7518fcb0034695a8018c7339cdbe3cc947d7ce02cca8ee", "sources": ["arxiv", "semantic_scholar"], "title": "A Training-free Sub-quadratic Cost Transformer Model Serving Framework With Hierarchically Pruned Attention", "abstract": "In modern large language models (LLMs), increasing the context length is crucial for improving comprehension and coherence in long-context, multi-modal, and retrieval-augmented language generation. While many recent transformer models attempt to extend their context length over a million tokens, they remain impractical due to the quadratic time and space complexities. Although recent works on linear and sparse attention mechanisms can achieve this goal, their real-world applicability is often limited by the need to re-train from scratch and significantly worse performance. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which reduces the time complexity of the attention mechanism to $O(T \\log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length. We notice a pattern in the attention scores of pretrained LLMs where tokens close together tend to have similar scores, which we call ``attention locality''. Based on this observation, we utilize a novel tree-search-like algorithm that estimates the top-$k$ key tokens for a given query on the fly, which is mathematically guaranteed to have better performance than random attention pruning. In addition to improving the time complexity of the attention mechanism, we further optimize GPU memory usage by implementing KV cache offloading, which stores only $O(\\log T)$ tokens on the GPU while maintaining similar decoding throughput. Experiments on benchmarks show that HiP, with its training-free nature, significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation. HiP enables pretrained LLMs to scale up to millions of tokens on commodity GPUs, potentially unlocking long-context LLM applications previously deemed infeasible.", "authors": ["Heejun Lee", "Geon Park", "Youngwan Lee", "Jaduk Suh", "Jina Kim", "Wonyoung Jeong", "Bumsik Kim", "Hyemin Lee", "Myeongjae Jeon", "Sung Ju Hwang"], "categories": ["cs.CL", "cs.CV", "cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-14", "url": "https://arxiv.org/abs/2406.09827", "pdf_url": "https://arxiv.org/pdf/2406.09827v3", "arxiv_id": "2406.09827", "doi": null, "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.25}
{"id": "260fdc5dd1a769ab3fc36ecf72cc9d17735fa2d72ccdd086a72a8e810c02ad40", "sources": ["arxiv", "semantic_scholar"], "title": "MoME: Mixture of Multimodal Experts for Cancer Survival Prediction", "abstract": "Survival analysis, as a challenging task, requires integrating Whole Slide Images (WSIs) and genomic data for comprehensive decision-making. There are two main challenges in this task: significant heterogeneity and complex inter- and intra-modal interactions between the two modalities. Previous approaches utilize co-attention methods, which fuse features from both modalities only once after separate encoding. However, these approaches are insufficient for modeling the complex task due to the heterogeneous nature between the modalities. To address these issues, we propose a Biased Progressive Encoding (BPE) paradigm, performing encoding and fusion simultaneously. This paradigm uses one modality as a reference when encoding the other. It enables deep fusion of the modalities through multiple alternating iterations, progressively reducing the cross-modal disparities and facilitating complementary interactions. Besides modality heterogeneity, survival analysis involves various biomarkers from WSIs, genomics, and their combinations. The critical biomarkers may exist in different modalities under individual variations, necessitating flexible adaptation of the models to specific scenarios. Therefore, we further propose a Mixture of Multimodal Experts (MoME) layer to dynamically selects tailored experts in each stage of the BPE paradigm. Experts incorporate reference information from another modality to varying degrees, enabling a balanced or biased focus on different modalities during the encoding process. Extensive experimental results demonstrate the superior performance of our method on various datasets, including TCGA-BLCA, TCGA-UCEC and TCGA-LUAD. Codes are available at https://github.com/BearCleverProud/MoME.", "authors": ["Conghao Xiong", "Hao Chen", "Hao Zheng", "Dong Wei", "Yefeng Zheng", "Joseph J. Y. Sung", "Irwin King"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-06-14", "url": "https://arxiv.org/abs/2406.09696", "pdf_url": "https://arxiv.org/pdf/2406.09696v1", "arxiv_id": "2406.09696", "doi": "10.48550/arXiv.2406.09696", "citation_count": 37, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/BearCleverProud/MoME", "venue": "International Conference on Medical Image Computing and Computer-Assisted Intervention", "quality_score": 0.3949}
{"id": "4d959c5d6811d8233e45d77369d57c0b717b4c5d62e3449ea29fd60283066c90", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Cascading Mixture-of-Experts for Urban Traffic Congestion Prediction", "abstract": "Rapid urbanization has significantly escalated traffic congestion, underscoring the need for advanced congestion prediction services to bolster intelligent transportation systems. As one of the world's largest ride-hailing platforms, DiDi places great emphasis on the accuracy of congestion prediction to enhance the effectiveness and reliability of their real-time services, such as travel time estimation and route planning. Despite numerous efforts have been made on congestion prediction, most of them fall short in handling heterogeneous and dynamic spatio-temporal dependencies (e.g., periodic and non-periodic congestions), particularly in the presence of noisy and incomplete traffic data. In this paper, we introduce a Congestion Prediction Mixture-of-Experts, CP-MoE, to address the above challenges. We first propose a sparsely-gated Mixture of Adaptive Graph Learners (MAGLs) with congestion-aware inductive biases to improve the model capacity for efficiently capturing complex spatio-temporal dependencies in varying traffic scenarios. Then, we devise two specialized experts to help identify stable trends and periodic patterns within the traffic data, respectively. By cascading these experts with MAGLs, CP-MoE delivers congestion predictions in a more robust and interpretable manner. Furthermore, an ordinal regression strategy is adopted to facilitate effective collaboration among diverse experts. Extensive experiments on real-world datasets demonstrate the superiority of our proposed method compared with state-of-the-art spatio-temporal prediction models. More importantly, CP-MoE has been deployed in DiDi to improve the accuracy and reliability of the travel time estimation system.", "authors": ["Wenzhao Jiang", "Jindong Han", "Hao Liu", "Tao Tao", "Naiqiang Tan", "Hui Xiong"], "categories": ["cs.LG", "cs.MA"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-14", "url": "https://arxiv.org/abs/2406.12923", "pdf_url": "https://arxiv.org/pdf/2406.12923v1", "arxiv_id": "2406.12923", "doi": "10.1145/3637528.3671507", "citation_count": 21, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Knowledge Discovery and Data Mining", "quality_score": 0.3356}
{"id": "0b6ea7b16a21ead3c53583c915bedd30c36e6dbba760a0203b05fbfaf3d5a2a4", "sources": ["arxiv", "semantic_scholar"], "title": "QuantMoE-Bench: Examining Post-Training Quantization for Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) is a promising way to scale up the learning capacity of large language models. It increases the number of parameters while keeping FLOPs nearly constant during inference through sparse activation. Yet, it still suffers from significant memory overheads due to the vast parameter size, necessitating model compression techniques. Post-training quantization offers a powerful approach for model compression. Existing methods adopt a fixed quantization precision for the entire MoE model. This rigid setup can lead to suboptimal performance, without considering the inherent sparse structure. For example, MoE's sparse routing mechanism leads to different activation patterns, where shared experts are accessed by all tokens while token-conditioned experts are selectively activated. This activation disparity suggests different quantization requirements, with consistently activated shared experts potentially needing higher precision to maintain model quality. In this paper, we study a fine-grained precision setup for MoE quantization. We explore MoE structure-aware quantization heuristics, ranging from coarse (e.g., MoE layers) to fine granularity (e.g., linear layers). Our investigations reveal critical principles, where different MoE structures require varying numbers of bits for effective quantization. Conclusions are supported by extensive benchmarking across two representative MoE models and six tasks including commonsense reasoning and natural language understanding. We further show that an MoE quantized in a fined-grained mixed precision achieved state-of-the-art 65.35% performance on average compared to the baseline 64.30% (i.e., GPTQ). Moreover, based on the findings, we introduce novel data-driven techniques for optimizing bit allocation in MoE quantization, including the outlier-aware linear layer scorer and MoE block importance predictor.", "authors": ["Pingzhi Li", "Xiaolong Jin", "Zhen Tan", "Yu Cheng", "Tianlong Chen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-12", "url": "https://arxiv.org/abs/2406.08155", "pdf_url": "https://arxiv.org/pdf/2406.08155v2", "arxiv_id": "2406.08155", "doi": null, "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/UNITES-Lab/moe-quantization", "venue": null, "quality_score": 0.2113}
{"id": "fcfb153c5a2b0ac9b109768a806247676e27938e966492cd61f5e1cb3da7e7cc", "sources": ["arxiv", "semantic_scholar"], "title": "MaIL: Improving Imitation Learning with Mamba", "abstract": "This work presents Mamba Imitation Learning (MaIL), a novel imitation learning (IL) architecture that provides an alternative to state-of-the-art (SoTA) Transformer-based policies. MaIL leverages Mamba, a state-space model designed to selectively focus on key features of the data. While Transformers are highly effective in data-rich environments due to their dense attention mechanisms, they can struggle with smaller datasets, often leading to overfitting or suboptimal representation learning. In contrast, Mamba's architecture enhances representation learning efficiency by focusing on key features and reducing model complexity. This approach mitigates overfitting and enhances generalization, even when working with limited data. Extensive evaluations on the LIBERO benchmark demonstrate that MaIL consistently outperforms Transformers on all LIBERO tasks with limited data and matches their performance when the full dataset is available. Additionally, MaIL's effectiveness is validated through its superior performance in three real robot experiments. Our code is available at https://github.com/ALRhub/MaIL.", "authors": ["Xiaogang Jia", "Qian Wang", "Atalay Donat", "Bowen Xing", "Ge Li", "Hongyi Zhou", "Onur Celik", "Denis Blessing", "Rudolf Lioutikov", "Gerhard Neumann"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-12", "url": "https://arxiv.org/abs/2406.08234", "pdf_url": "https://arxiv.org/pdf/2406.08234v2", "arxiv_id": "2406.08234", "doi": "10.48550/arXiv.2406.08234", "citation_count": 25, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/ALRhub/MaIL", "venue": "arXiv.org", "quality_score": 0.3537}
{"id": "e76bb9b622f74a93060f4da4a777c99cd1f7fd032f07ab11a00f22c8d7b6ff2d", "sources": ["arxiv", "semantic_scholar"], "title": "Effectively Compress KV Heads for LLM", "abstract": "The advent of pre-trained large language models (LLMs) has revolutionized various natural language processing tasks. These models predominantly employ an auto-regressive decoding mechanism that utilizes Key-Value (KV) caches to eliminate redundant calculations for previous tokens. Nevertheless, as context lengths and batch sizes increase, the linear expansion in memory footprint of KV caches becomes a key bottleneck of LLM deployment, which decreases generation speeds significantly. To mitigate this issue, previous techniques like multi-query attention (MQA) and grouped-query attention (GQA) have been developed, in order to reduce KV heads to accelerate inference with comparable accuracy to multi-head attention (MHA). Despite their effectiveness, existing strategies for compressing MHA often overlook the intrinsic properties of the KV caches. In this work, we explore the low-rank characteristics of the KV caches and propose a novel approach for compressing KV heads. In particular, we carefully optimize the MHA-to-GQA transformation to minimize compression error, and to remain compatible with rotary position embeddings (RoPE), we also introduce specialized strategies for key caches with RoPE. We demonstrate that our method can compress half or even three-quarters of KV heads while maintaining performance comparable to the original LLMs, which presents a promising direction for more efficient LLM deployment in resource-constrained environments.", "authors": ["Hao Yu", "Zelan Yang", "Shen Li", "Yong Li", "Jianxin Wu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-11", "url": "https://arxiv.org/abs/2406.07056", "pdf_url": "https://arxiv.org/pdf/2406.07056v1", "arxiv_id": "2406.07056", "doi": "10.48550/arXiv.2406.07056", "citation_count": 31, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3763}
{"id": "a858ac87b732420f5bb8ffa9b0af3a0ce25ff114db944d101c46b58dd7f488c9", "sources": ["arxiv", "semantic_scholar"], "title": "Attention as a Hypernetwork", "abstract": "Transformers can under some circumstances generalize to novel problem instances whose constituent parts might have been encountered during training, but whose compositions have not. What mechanisms underlie this ability for compositional generalization? By reformulating multi-head attention as a hypernetwork, we reveal that a composable, low-dimensional latent code specifies key-query specific operations. We find empirically that this latent code is predictive of the subtasks the network performs on unseen task compositions, revealing that latent codes acquired during training are reused to solve unseen problem instances. To further examine the hypothesis that the intrinsic hypernetwork of multi-head attention supports compositional generalization, we ablate whether making the hypernetwork-generated linear value network nonlinear strengthens compositionality. We find that this modification improves compositional generalization on abstract reasoning tasks. In particular, we introduce a symbolic version of the Raven's Progressive Matrices human intelligence test, which gives us precise control over the problem compositions encountered during training and evaluation. We demonstrate on this task how scaling model size and data enables compositional generalization in transformers and gives rise to a functionally structured latent space.", "authors": ["Simon Schug", "Seijin Kobayashi", "Yassir Akram", "João Sacramento", "Razvan Pascanu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-09", "url": "https://arxiv.org/abs/2406.05816", "pdf_url": "https://arxiv.org/pdf/2406.05816v4", "arxiv_id": "2406.05816", "doi": "10.48550/arXiv.2406.05816", "citation_count": 13, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/smonsays/hypernetwork-attention", "venue": "International Conference on Learning Representations", "quality_score": 0.2865}
{"id": "ed61cae21a5bd5de47212b9e15b18100577eb968636129cf4910cf3e55ffae74", "sources": ["arxiv", "semantic_scholar"], "title": "QCQA: Quality and Capacity-aware grouped Query Attention", "abstract": "Excessive memory requirements of key and value features (KV-cache) present significant challenges in the autoregressive inference of large language models (LLMs), restricting both the speed and length of text generation. Approaches such as Multi-Query Attention (MQA) and Grouped Query Attention (GQA) mitigate these challenges by grouping query heads and consequently reducing the number of corresponding key and value heads. However, MQA and GQA decrease the KV-cache size requirements at the expense of LLM accuracy (quality of text generation). These methods do not ensure an optimal tradeoff between KV-cache size and text generation quality due to the absence of quality-aware grouping of query heads. To address this issue, we propose Quality and Capacity-Aware Grouped Query Attention (QCQA), which identifies optimal query head groupings using an evolutionary algorithm with a computationally efficient and inexpensive fitness function. We demonstrate that QCQA achieves a significantly better tradeoff between KV-cache capacity and LLM accuracy compared to GQA. For the Llama2 $7\\,$B model, QCQA achieves $\\mathbf{20}$\\% higher accuracy than GQA with similar KV-cache size requirements in the absence of fine-tuning. After fine-tuning both QCQA and GQA, for a similar KV-cache size, QCQA provides $\\mathbf{10.55}\\,$\\% higher accuracy than GQA. Furthermore, QCQA requires $40\\,$\\% less KV-cache size than GQA to attain similar accuracy. The proposed quality and capacity-aware grouping of query heads can serve as a new paradigm for KV-cache optimization in autoregressive LLM inference.", "authors": ["Vinay Joshi", "Prashant Laddha", "Shambhavi Sinha", "Om Ji Omer", "Sreenivas Subramoney"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-08", "url": "https://arxiv.org/abs/2406.10247", "pdf_url": "https://arxiv.org/pdf/2406.10247v1", "arxiv_id": "2406.10247", "doi": "10.48550/arXiv.2406.10247", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "dfa907496f976d2f57c4d92d33418c76402648a4884dc1056cdb491065c73d4c", "sources": ["arxiv", "semantic_scholar"], "title": "Federated LoRA with Sparse Communication", "abstract": "Low-rank adaptation (LoRA) is a natural method for finetuning in communication-constrained machine learning settings such as cross-device federated learning. Prior work that has studied LoRA in the context of federated learning has focused on improving LoRA's robustness to heterogeneity and privacy. In this work, we instead consider techniques for further improving communication-efficiency in federated LoRA. Unfortunately, we show that centralized ML methods that improve the efficiency of LoRA through unstructured pruning do not transfer well to federated settings. We instead study a simple approach, \\textbf{FLASC}, that applies sparsity to LoRA during communication while allowing clients to locally fine-tune the entire LoRA module. Across four common federated learning tasks, we demonstrate that this method matches the performance of dense LoRA with up to $10\\times$ less communication. Additionally, despite being designed primarily to target communication, we find that this approach has benefits in terms of heterogeneity and privacy relative to existing approaches tailored to these specific concerns. Overall, our work highlights the importance of considering system-specific constraints when developing communication-efficient finetuning approaches, and serves as a simple and competitive baseline for future work in federated finetuning.", "authors": ["Kevin Kuo", "Arian Raje", "Kousik Rajesh", "Virginia Smith"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-07", "url": "https://arxiv.org/abs/2406.05233", "pdf_url": "https://arxiv.org/pdf/2406.05233v1", "arxiv_id": "2406.05233", "doi": "10.48550/arXiv.2406.05233", "citation_count": 31, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3763}
{"id": "9c3de1db38c56ec795ab031d0c0b21ef055486f9fc50d13d759e0ad4aa4274db", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient 3D Shape Generation via Diffusion Mamba with Bidirectional SSMs", "abstract": "Recent advancements in sequence modeling have led to the development of the Mamba architecture, noted for its selective state space approach, offering a promising avenue for efficient long sequence handling. However, its application in 3D shape generation, particularly at high resolutions, remains underexplored. Traditional diffusion transformers (DiT) with self-attention mechanisms, despite their potential, face scalability challenges due to the cubic complexity of attention operations as input length increases. This complexity becomes a significant hurdle when dealing with high-resolution voxel sizes. To address this challenge, we introduce a novel diffusion architecture tailored for 3D point clouds generation-Diffusion Mamba (DiM-3D). This architecture forgoes traditional attention mechanisms, instead utilizing the inherent efficiency of the Mamba architecture to maintain linear complexity with respect to sequence length. DiM-3D is characterized by fast inference times and substantially lower computational demands, quantified in reduced Gflops, thereby addressing the key scalability issues of prior models. Our empirical results on the ShapeNet benchmark demonstrate that DiM-3D achieves state-of-the-art performance in generating high-fidelity and diverse 3D shapes. Additionally, DiM-3D shows superior capabilities in tasks like 3D point cloud completion. This not only proves the model's scalability but also underscores its efficiency in generating detailed, high-resolution voxels necessary for advanced 3D shape modeling, particularly excelling in environments requiring high-resolution voxel sizes. Through these findings, we illustrate the exceptional scalability and efficiency of the Diffusion Mamba framework in 3D shape generation, setting a new standard for the field and paving the way for future explorations in high-resolution 3D modeling technologies.", "authors": ["Shentong Mo"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-07", "url": "https://arxiv.org/abs/2406.05038", "pdf_url": "https://arxiv.org/pdf/2406.05038v1", "arxiv_id": "2406.05038", "doi": "10.48550/arXiv.2406.05038", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "cc2fb4375eae90e79e7dccca67b7010f9320345b42281a55bbdcf96bb4b40074", "sources": ["arxiv", "semantic_scholar"], "title": "QJL: 1-Bit Quantized JL Transform for KV Cache Quantization with Zero Overhead", "abstract": "Serving LLMs requires substantial memory due to the storage requirements of Key-Value (KV) embeddings in the KV cache, which grows with sequence length. An effective approach to compress KV cache is quantization. However, traditional quantization methods face significant memory overhead due to the need to store quantization constants (at least a zero point and a scale) in full precision per data block. Depending on the block size, this overhead can add 1 or 2 bits per quantized number. We introduce QJL, a new quantization approach that consists of a Johnson-Lindenstrauss (JL) transform followed by sign-bit quantization. In contrast to existing methods, QJL eliminates memory overheads by removing the need for storing quantization constants. We propose an asymmetric estimator for the inner product of two vectors and demonstrate that applying QJL to one vector and a standard JL transform without quantization to the other provides an unbiased estimator with minimal distortion. We have developed an efficient implementation of the QJL sketch and its corresponding inner product estimator, incorporating a lightweight CUDA kernel for optimized computation. When applied across various LLMs and NLP tasks to quantize the KV cache to only 3 bits, QJL demonstrates a more than fivefold reduction in KV cache memory usage without compromising accuracy, all while achieving faster runtime. Codes are available at \\url{https://github.com/amirzandieh/QJL}.", "authors": ["Amir Zandieh", "Majid Daliri", "Insu Han"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-05", "url": "https://arxiv.org/abs/2406.03482", "pdf_url": "https://arxiv.org/pdf/2406.03482v2", "arxiv_id": "2406.03482", "doi": "10.48550/arXiv.2406.03482", "citation_count": 41, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/amirzandieh/QJL}", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4058}
{"id": "1aabaf102753528b5aca0963535f8b37b4cd9d042a1a5d59cbd381591d6301f6", "sources": ["arxiv", "semantic_scholar"], "title": "Continual Traffic Forecasting via Mixture of Experts", "abstract": "The real-world traffic networks undergo expansion through the installation of new sensors, implying that the traffic patterns continually evolve over time. Incrementally training a model on the newly added sensors would make the model forget the past knowledge, i.e., catastrophic forgetting, while retraining the model on the entire network to capture these changes is highly inefficient. To address these challenges, we propose a novel Traffic Forecasting Mixture of Experts (TFMoE) for traffic forecasting under evolving networks. The main idea is to segment the traffic flow into multiple homogeneous groups, and assign an expert model responsible for a specific group. This allows each expert model to concentrate on learning and adapting to a specific set of patterns, while minimizing interference between the experts during training, thereby preventing the dilution or replacement of prior knowledge, which is a major cause of catastrophic forgetting. Through extensive experiments on a real-world long-term streaming network dataset, PEMSD3-Stream, we demonstrate the effectiveness and efficiency of TFMoE. Our results showcase superior performance and resilience in the face of catastrophic forgetting, underscoring the effectiveness of our approach in dealing with continual learning for traffic flow forecasting in long-term streaming networks.", "authors": ["Sanghyun Lee", "Chanyoung Park"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-05", "url": "https://arxiv.org/abs/2406.03140", "pdf_url": "https://arxiv.org/pdf/2406.03140v1", "arxiv_id": "2406.03140", "doi": "10.48550/arXiv.2406.03140", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "436000ec8c54f7d2743eadc6a1eeace5adbf2a100ec0c2165b6fb201606be76f", "sources": ["arxiv", "semantic_scholar"], "title": "Style Mixture of Experts for Expressive Text-To-Speech Synthesis", "abstract": "Recent advances in style transfer text-to-speech (TTS) have improved the expressiveness of synthesized speech. However, encoding stylistic information (e.g., timbre, emotion, and prosody) from diverse and unseen reference speech remains a challenge. This paper introduces StyleMoE, an approach that addresses the issue of learning averaged style representations in the style encoder by creating style experts that learn from subsets of data. The proposed method replaces the style encoder in a TTS framework with a Mixture of Experts (MoE) layer. The style experts specialize by learning from subsets of reference speech routed to them by the gating network, enabling them to handle different aspects of the style space. As a result, StyleMoE improves the style coverage of the style encoder for style transfer TTS. Our experiments, both objective and subjective, demonstrate improved style transfer for diverse and unseen reference speech. The proposed method enhances the performance of existing state-of-the-art style transfer TTS models and represents the first study of style MoE in TTS.", "authors": ["Ahad Jawaid", "Shreeram Suresh Chandra", "Junchen Lu", "Berrak Sisman"], "categories": ["eess.AS", "cs.CL", "cs.LG", "cs.SD"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-06-05", "url": "https://arxiv.org/abs/2406.03637", "pdf_url": "https://arxiv.org/pdf/2406.03637v2", "arxiv_id": "2406.03637", "doi": "10.48550/arXiv.2406.03637", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "ac05d8379c6fd5637e8966934bb15f12831118785315c94811d6f56637e13ba3", "sources": ["arxiv", "semantic_scholar"], "title": "Node-wise Filtering in Graph Neural Networks: A Mixture of Experts Approach", "abstract": "Graph Neural Networks (GNNs) have proven to be highly effective for node classification tasks across diverse graph structural patterns. Traditionally, GNNs employ a uniform global filter, typically a low-pass filter for homophilic graphs and a high-pass filter for heterophilic graphs. However, real-world graphs often exhibit a complex mix of homophilic and heterophilic patterns, rendering a single global filter approach suboptimal. In this work, we theoretically demonstrate that a global filter optimized for one pattern can adversely affect performance on nodes with differing patterns. To address this, we introduce a novel GNN framework Node-MoE that utilizes a mixture of experts to adaptively select the appropriate filters for different nodes. Extensive experiments demonstrate the effectiveness of Node-MoE on both homophilic and heterophilic graphs.", "authors": ["Haoyu Han", "Juanhui Li", "Wei Huang", "Xianfeng Tang", "Hanqing Lu", "Chen Luo", "Hui Liu", "Jiliang Tang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-05", "url": "https://arxiv.org/abs/2406.03464", "pdf_url": "https://arxiv.org/pdf/2406.03464v1", "arxiv_id": "2406.03464", "doi": "10.48550/arXiv.2406.03464", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3197}
{"id": "f2690dc814da5b22fb6f6a69fa69c949646361cd9d4d0d04a8888b43f241bcdb", "sources": ["arxiv", "semantic_scholar"], "title": "PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling", "abstract": "In this study, we investigate whether attention-based information flow inside large language models (LLMs) is aggregated through noticeable patterns for long context processing. Our observations reveal that LLMs aggregate information through Pyramidal Information Funneling where attention is scattering widely in lower layers, progressively consolidating within specific contexts, and ultimately focusing on critical tokens (a.k.a massive activation or attention sink) in higher layers. Motivated by these insights, we developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers, allocating more cache in lower layers and less in higher ones, diverging from traditional methods that maintain a uniform KV cache size. Our experimental evaluations, utilizing the LongBench benchmark, show that PyramidKV matches the performance of models with a full KV cache while retaining only 12% of the KV cache, thus significantly reducing memory usage. In scenarios emphasizing memory efficiency, where only 0.7% of the KV cache is maintained, PyramidKV surpasses other KV cache compression techniques, achieving up to a 20.5 absolute accuracy improvement on TREC dataset. In the Needle-in-a-Haystack experiment, PyramidKV outperforms competing methods in maintaining long-context comprehension in LLMs; notably, retaining just 128 KV cache entries enables the LLAMA-3-70B model to achieve 100.0 Acc. performance.", "authors": ["Zefan Cai", "Yichi Zhang", "Bofei Gao", "Yuliang Liu", "Yucheng Li", "Tianyu Liu", "Keming Lu", "Wayne Xiong", "Yue Dong", "Junjie Hu", "Wen Xiao"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-04", "url": "https://arxiv.org/abs/2406.02069", "pdf_url": "https://arxiv.org/pdf/2406.02069v4", "arxiv_id": "2406.02069", "doi": "10.48550/arXiv.2406.02069", "citation_count": 319, "influential_citation_count": 64, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.9065}
{"id": "a03d85f18d68dc51be151367cde9f375f84147f18aadcfe7792759a169d6b581", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Efficient Mixture of Experts: A Holistic Study of Compression Techniques", "abstract": "Scaling large language models has driven remarkable advancements across various domains, yet the continual increase in model size presents significant challenges for real-world deployment. The Mixture of Experts (MoE) architecture offers a promising solution by dynamically selecting and activating only a subset of experts during inference, thus substantially reducing computational costs while preserving high performance. Despite these benefits, MoE introduces new inefficiencies, such as excessive parameters and communication overhead. In this work, we present a holistic study of compression techniques for Mixture of Experts to enhance both efficiency and scalability. While recent efforts have focused on Expert Trimming, which reduces the number of experts, these approaches still suffer from considerable communication and computational costs. To address this, we propose more aggressive strategies, such as Layer Drop, which removes entire MoE layers, and Block Drop, which eliminates transformer blocks. Surprisingly, these aggressive pruning techniques not only preserve model performance but also substantially improve computation and memory efficiency. Furthermore, beyond Expert Trimming, we also introduce Expert Slimming, which compresses individual experts to further boost performance and can be seamlessly integrated with Expert Trimming. Extensive experimental results demonstrate the effectiveness of our proposed methods-Layer Drop and Block Drop-along with the comprehensive recipe that integrates Expert Slimming and Expert Trimming, achieving a 6.05x speedup with 77.1% reduced memory usage while maintaining over 92% of performance on Mixtral-8x7B. Our code is released at https://github.com/CASE-Lab-UMD/Unified-MoE-Compression.", "authors": ["Shwai He", "Daize Dong", "Liang Ding", "Ang Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-04", "url": "https://arxiv.org/abs/2406.02500", "pdf_url": "https://arxiv.org/pdf/2406.02500v3", "arxiv_id": "2406.02500", "doi": null, "citation_count": 22, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/CASE-Lab-UMD/Unified-MoE-Compression", "venue": null, "quality_score": 0.3404}
{"id": "cd6ed72c04db6c1dbec85d698d89a14e7421f2395411dc544f673f0722e2fa96", "sources": ["arxiv", "semantic_scholar"], "title": "Focus on the Core: Efficient Attention via Pruned Token Compression for Document Classification", "abstract": "Transformer-based models have achieved dominant performance in numerous NLP tasks. Despite their remarkable successes, pre-trained transformers such as BERT suffer from a computationally expensive self-attention mechanism that interacts with all tokens, including the ones unfavorable to classification performance. To overcome these challenges, we propose integrating two strategies: token pruning and token combining. Token pruning eliminates less important tokens in the attention mechanism's key and value as they pass through the layers. Additionally, we adopt fuzzy logic to handle uncertainty and alleviate potential mispruning risks arising from an imbalanced distribution of each token's importance. Token combining, on the other hand, condenses input sequences into smaller sizes in order to further compress the model. By integrating these two approaches, we not only improve the model's performance but also reduce its computational demands. Experiments with various datasets demonstrate superior performance compared to baseline models, especially with the best improvement over the existing BERT model, achieving +5%p in accuracy and +5.6%p in F1 score. Additionally, memory cost is reduced to 0.61x, and a speedup of 1.64x is achieved.", "authors": ["Jungmin Yun", "Mihyeon Kim", "Youngbin Kim"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-03", "url": "https://arxiv.org/abs/2406.01283", "pdf_url": "https://arxiv.org/pdf/2406.01283v1", "arxiv_id": "2406.01283", "doi": "10.18653/v1/2023.findings-emnlp.909", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2785}
{"id": "bbc8b9839f85b2a85cf2831ead0b4cb7585f637f8711526c8b005dc0e881530d", "sources": ["arxiv", "semantic_scholar"], "title": "Automatic Channel Pruning for Multi-Head Attention", "abstract": "Despite the strong performance of Transformers, their quadratic computation complexity presents challenges in applying them to vision tasks. Automatic pruning is one of effective methods for reducing computation complexity without heuristic approaches. However, directly applying it to multi-head attention is not straightforward due to channel misalignment. In this paper, we propose an automatic channel pruning method to take into account the multi-head attention mechanism. First, we incorporate channel similarity-based weights into the pruning indicator to preserve more informative channels in each head. Then, we adjust pruning indicator to enforce removal of channels in equal proportions across all heads, preventing the channel misalignment. We also add a reweight module to compensate for information loss resulting from channel removal, and an effective initialization step for pruning indicator based on difference of attention between original structure and each channel. Our proposed method can be used to not only original attention, but also linear attention, which is more efficient as linear complexity with respect to the number of tokens. On ImageNet-1K, applying our pruning method to the FLattenTransformer, which includes both attention mechanisms, shows outperformed accuracy for several MACs compared with previous state-of-the-art efficient models and pruned methods. Code will be available soon.", "authors": ["Eunho Lee", "Youngbae Hwang"], "categories": ["cs.CV", "cs.AI", "cs.CC"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-31", "url": "https://arxiv.org/abs/2405.20867", "pdf_url": "https://arxiv.org/pdf/2405.20867v1", "arxiv_id": "2405.20867", "doi": "10.48550/arXiv.2405.20867", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "9a32bbd1d37d6c41b66ab90503745137dd0dd7412cf9e80c8629d793f8347c62", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba State-Space Models Are Lyapunov-Stable Learners", "abstract": "Mamba state-space models (SSMs) have recently outperformed state-of-the-art (SOTA) Transformer large language models (LLMs) in various tasks and been widely adapted. However, a major concern for stable learning in recurrent-based deep models (such as SSMs) is the sensitivity of their recurrent dynamics. Despite widespread adaptation, the sensitivity of Mamba's recurrent dynamics under common fine-tuning methods-e.g., mixed-precision fine-tuning (MPFT) and parameter-efficient fine-tuning (PEFT)-remains unexplored. Empirically, we show that Mamba LLMs are extremely stable to changes introduced by combinations of MPFT and PEFT, in stark contrast to Transformer LLMs, which we demonstrate may drastically diverge from their respective full-precision counterparts under different combinations of MPFT and PEFT (despite the near-ubiquitous adaptation of these fine-tuning frameworks for attention-based models). The demonstrated robustness of Mamba LLMs are due to their recurrent dynamics, which we prove are guaranteed to be stable using dynamical systems theory (in particular, Lyapunov stability). We conclude by using MPFT and PEFT to novelly study Mamba LLMs' in-context learning (ICL) abilities on natural language tasks, thus supplementing other recent work.", "authors": ["John T. Halloran", "Manbir Gulati", "Paul F. Roysdon"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-31", "url": "https://arxiv.org/abs/2406.00209", "pdf_url": "https://arxiv.org/pdf/2406.00209v3", "arxiv_id": "2406.00209", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "57f379353c5a4dd83ab44f6b46e6b00164a79ea36d2e6a9e377b93706c7cdaba", "sources": ["arxiv", "semantic_scholar"], "title": "MLAE: Masked LoRA Experts for Visual Parameter-Efficient Fine-Tuning", "abstract": "In response to the challenges posed by the extensive parameter updates required for full fine-tuning of large-scale pre-trained models, parameter-efficient fine-tuning (PEFT) methods, exemplified by Low-Rank Adaptation (LoRA), have emerged. LoRA simplifies the fine-tuning process but may still struggle with a certain level of redundancy in low-rank matrices and limited effectiveness from merely increasing their rank. To address these issues, a natural idea is to enhance the independence and diversity of the learning process for the low-rank matrices. Therefore, we propose Masked LoRA Experts (MLAE), an innovative approach that applies the concept of masking to visual PEFT. Our method incorporates a cellular decomposition strategy that transforms a low-rank matrix into independent rank-1 submatrices, or \"experts\", thus enhancing independence. Additionally, we introduce a binary mask matrix that selectively activates these experts during training to promote more diverse and anisotropic learning, based on expert-level dropout strategies. Our investigations reveal that this selective activation not only enhances performance but also fosters a more diverse acquisition of knowledge with a marked decrease in parameter similarity among MLAE, significantly boosting the quality of the model. Remarkably, MLAE achieves new state-of-the-art (SOTA) performance with an average accuracy score of 78.8% on the VTAB-1k benchmark and 90.9% on the FGVC benchmark, surpassing the previous SOTA result by an average of 0.8% on both benchmarks with approximately half parameters. Our code is available at https://github.com/jie040109/MLAE.", "authors": ["Junjie Wang", "Guangjing Yang", "Wentao Chen", "Huahui Yi", "Xiaohu Wu", "Zhouchen Lin", "Qicheng Lao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-29", "url": "https://arxiv.org/abs/2405.18897", "pdf_url": "https://arxiv.org/pdf/2405.18897v2", "arxiv_id": "2405.18897", "doi": null, "citation_count": 8, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/jie040109/MLAE", "venue": null, "quality_score": 0.2386}
{"id": "d2455165de97ad49098297f3306a4e462e3dc04e2b1bafa6c1172867e7f19df8", "sources": ["arxiv", "semantic_scholar"], "title": "MoNDE: Mixture of Near-Data Experts for Large-Scale Sparse Models", "abstract": "Mixture-of-Experts (MoE) large language models (LLM) have memory requirements that often exceed the GPU memory capacity, requiring costly parameter movement from secondary memories to the GPU for expert computation. In this work, we present Mixture of Near-Data Experts (MoNDE), a near-data computing solution that efficiently enables MoE LLM inference. MoNDE reduces the volume of MoE parameter movement by transferring only the $\\textit{hot}$ experts to the GPU, while computing the remaining $\\textit{cold}$ experts inside the host memory device. By replacing the transfers of massive expert parameters with the ones of small activations, MoNDE enables far more communication-efficient MoE inference, thereby resulting in substantial speedups over the existing parameter offloading frameworks for both encoder and decoder operations.", "authors": ["Taehyun Kim", "Kwanseok Choi", "Youngmock Cho", "Jaehoon Cho", "Hyuk-Jae Lee", "Jaewoong Sim"], "categories": ["cs.LG", "cs.AI", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-29", "url": "https://arxiv.org/abs/2405.18832", "pdf_url": "https://arxiv.org/pdf/2405.18832v1", "arxiv_id": "2405.18832", "doi": "10.1145/3649329.3655951", "citation_count": 14, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Design Automation Conference", "quality_score": 0.294}
{"id": "1fde663289f2242784dc1fed3ea3ad7ec151c113005e71fdba2784b73562a300", "sources": ["arxiv", "semantic_scholar"], "title": "MEMoE: Enhancing Model Editing with Mixture of Experts Adaptors", "abstract": "Model editing aims to efficiently alter the behavior of Large Language Models (LLMs) within a desired scope, while ensuring no adverse impact on other inputs. Recent years have witnessed various model editing methods been proposed. However, these methods either exhibit poor overall performance or struggle to strike a balance between generalization and locality. We propose MEMoE, a model editing adapter utilizing a Mixture of Experts (MoE) architecture with a knowledge anchor routing strategy. MEMoE updates knowledge using a bypass MoE structure, keeping the original parameters unchanged to preserve the general ability of LLMs. And, the knowledge anchor routing ensures that inputs requiring similar knowledge are routed to the same expert, thereby enhancing the generalization of the updated knowledge. Experimental results show the superiority of our approach over both batch editing and sequential batch editing tasks, exhibiting exceptional overall performance alongside outstanding balance between generalization and locality. Our code will be available.", "authors": ["Renzhi Wang", "Piji Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-29", "url": "https://arxiv.org/abs/2405.19086", "pdf_url": "https://arxiv.org/pdf/2405.19086v2", "arxiv_id": "2405.19086", "doi": "10.48550/arXiv.2405.19086", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "5172b9e00d8be99f86654b71d47c9585ee52d3e4adc227f6e26b739ec372f785", "sources": ["arxiv", "semantic_scholar"], "title": "Yuan 2.0-M32: Mixture of Experts with Attention Router", "abstract": "Yuan 2.0-M32, with a similar base architecture as Yuan-2.0 2B, uses a mixture-of-experts architecture with 32 experts of which 2 experts are active. A new router network, Attention Router, is proposed and adopted for a more efficient selection of experts, which improves the accuracy compared to the model with classical router network. Yuan 2.0-M32 is trained with 2000B tokens from scratch, and the training computation consumption is only 9.25% of a dense model at the same parameter scale. Yuan 2.0-M32 demonstrates competitive capability on coding, math, and various domains of expertise, with only 3.7B active parameters of 40B in total, and 7.4 GFlops forward computation per token, both of which are only 1/19 of Llama3-70B. Yuan 2.0-M32 surpass Llama3-70B on MATH and ARC-Challenge benchmark, with accuracy of 55.89 and 95.8 respectively. The models and source codes of Yuan 2.0-M32 are released at Github1.", "authors": ["Shaohua Wu", "Jiangang Luo", "Xi Chen", "Lingjun Li", "Xudong Zhao", "Tong Yu", "Chao Wang", "Yue Wang", "Fei Wang", "Weixu Qiao", "Houbo He", "Zeru Zhang", "Zeyu Sun", "Junxiong Mao", "Chong Shen"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-28", "url": "https://arxiv.org/abs/2405.17976", "pdf_url": "https://arxiv.org/pdf/2405.17976v2", "arxiv_id": "2405.17976", "doi": "10.48550/arXiv.2405.17976", "citation_count": 21, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4225}
{"id": "e3a67557019bbc178472ce8cfc673e6220a16594b07ea860c532b12f521055a7", "sources": ["arxiv", "semantic_scholar"], "title": "Mixtures of Unsupervised Lexicon Classification", "abstract": "This paper presents a mixture version of the method-of-moment unsupervised lexicon classification by an incorporation of a Dirichlet process.", "authors": ["Peratham Wiriyathammabhum"], "categories": ["cs.CL", "cs.AI", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-27", "url": "https://arxiv.org/abs/2405.17116", "pdf_url": "https://arxiv.org/pdf/2405.17116v1", "arxiv_id": "2405.17116", "doi": "10.48550/arXiv.2405.17116", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "d078db03e882815318b477ea555a893b60b866a56fc1053ec7de91d147a824a7", "sources": ["arxiv", "semantic_scholar"], "title": "Multiple Heads are Better than One: Mixture of Modality Knowledge Experts for Entity Representation Learning", "abstract": "Learning high-quality multi-modal entity representations is an important goal of multi-modal knowledge graph (MMKG) representation learning, which can enhance reasoning tasks within the MMKGs, such as MMKG completion (MMKGC). The main challenge is to collaboratively model the structural information concealed in massive triples and the multi-modal features of the entities. Existing methods focus on crafting elegant entity-wise multi-modal fusion strategies, yet they overlook the utilization of multi-perspective features concealed within the modalities under diverse relational contexts. To address this issue, we introduce a novel framework with Mixture of Modality Knowledge experts (MoMoK for short) to learn adaptive multi-modal entity representations for better MMKGC. We design relation-guided modality knowledge experts to acquire relation-aware modality embeddings and integrate the predictions from multi-modalities to achieve joint decisions. Additionally, we disentangle the experts by minimizing their mutual information. Experiments on four public MMKG benchmarks demonstrate the outstanding performance of MoMoK under complex scenarios.", "authors": ["Yichi Zhang", "Zhuo Chen", "Lingbing Guo", "Yajing Xu", "Binbin Hu", "Ziqi Liu", "Wen Zhang", "Huajun Chen"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-27", "url": "https://arxiv.org/abs/2405.16869", "pdf_url": "https://arxiv.org/pdf/2405.16869v4", "arxiv_id": "2405.16869", "doi": null, "citation_count": 23, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3451}
{"id": "7ca1fd470586f8873d86418af3b2ee803c4293914f583687167c103466f1fd00", "sources": ["arxiv", "semantic_scholar"], "title": "A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts", "abstract": "The sparsely gated mixture of experts (MoE) architecture sends different inputs to different subnetworks, i.e., experts, through trainable routers. MoE reduces the training computation significantly for large models, but its deployment can be still memory or computation expensive for some downstream tasks. Model pruning is a popular approach to reduce inference computation, but its application in MoE architecture is largely unexplored. To the best of our knowledge, this paper provides the first provably efficient technique for pruning experts in finetuned MoE models. We theoretically prove that prioritizing the pruning of the experts with a smaller change of the routers l2 norm from the pretrained model guarantees the preservation of test accuracy, while significantly reducing the model size and the computational requirements. Although our theoretical analysis is centered on binary classification tasks on simplified MoE architecture, our expert pruning method is verified on large vision MoE models such as VMoE and E3MoE finetuned on benchmark datasets such as CIFAR10, CIFAR100, and ImageNet.", "authors": ["Mohammed Nowaz Rabbani Chowdhury", "Meng Wang", "Kaoutar El Maghraoui", "Naigang Wang", "Pin-Yu Chen", "Christopher Carothers"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-26", "url": "https://arxiv.org/abs/2405.16646", "pdf_url": "https://arxiv.org/pdf/2405.16646v3", "arxiv_id": "2405.16646", "doi": "10.48550/arXiv.2405.16646", "citation_count": 21, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3356}
{"id": "f16129f480a2b9ebc192084f55659a0f8c4052466109ec16a64d3d5ebfd8d365", "sources": ["arxiv", "semantic_scholar"], "title": "CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion", "abstract": "Large language models (LLMs) often incorporate multiple text chunks in their inputs to provide the necessary contexts. To speed up the prefill of the long LLM inputs, one can pre-compute the KV cache of a text and re-use the KV cache when the context is reused as the prefix of another LLM input. However, the reused text chunks are not always the input prefix, which makes precomputed KV caches not directly usable since they ignore the text's cross-attention with the preceding texts. Thus, the benefits of reusing KV caches remain largely unrealized. This paper tackles just one challenge: when an LLM input contains multiple text chunks, how to quickly combine their precomputed KV caches in order to achieve the same generation quality as the expensive full prefill (i.e., without reusing KV cache)? This challenge naturally arises in retrieval-augmented generation (RAG) where the input is supplemented with multiple retrieved texts as the context. We present CacheBlend, a scheme that reuses the precomputed KV caches, regardless prefix or not, and selectively recomputes the KV values of a small subset of tokens to partially update each reused KV cache. In the meantime, the small extra delay for recomputing some tokens can be pipelined with the retrieval of KV caches within the same job, allowing CacheBlend to store KV caches in slower devices with more storage capacity while retrieving them without increasing the inference delay. By comparing CacheBlend with the state-of-the-art KV cache reusing schemes on three open-source LLMs of various sizes and four popular benchmark datasets of different tasks, we show that CacheBlend reduces time-to-first-token (TTFT) by 2.2-3.3x and increases the inference throughput by 2.8-5x from full KV recompute without compromising generation quality. The code is available at https://github.com/LMCache/LMCache.", "authors": ["Jiayi Yao", "Hanchen Li", "Yuhan Liu", "Siddhant Ray", "Yihua Cheng", "Qizheng Zhang", "Kuntai Du", "Shan Lu", "Junchen Jiang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-26", "url": "https://arxiv.org/abs/2405.16444", "pdf_url": "https://arxiv.org/pdf/2405.16444v3", "arxiv_id": "2405.16444", "doi": "10.1145/3790254", "citation_count": 169, "influential_citation_count": 22, "has_code": true, "code_url": "https://github.com/LMCache/LMCache", "venue": "ACM Transactions on Computer Systems", "quality_score": 0.6809}
{"id": "379a5fce3c8759798c10646a76d68d15f4a8313efb723ca5bfc8f15bca0c2502", "sources": ["arxiv", "semantic_scholar"], "title": "Clustering Survival Data using a Mixture of Non-parametric Experts", "abstract": "Survival analysis aims to predict the timing of future events across various fields, from medical outcomes to customer churn. However, the integration of clustering into survival analysis, particularly for precision medicine, remains underexplored. This study introduces SurvMixClust, a novel algorithm for survival analysis that integrates clustering with survival function prediction within a unified framework. SurvMixClust learns latent representations for clustering while also predicting individual survival functions using a mixture of non-parametric experts. Our evaluations on five public datasets show that SurvMixClust creates balanced clusters with distinct survival curves, outperforms clustering baselines, and competes with non-clustering survival models in predictive accuracy, as measured by the time-dependent c-index and log-rank metrics.", "authors": ["Gabriel Buginga", "Edmundo de Souza e Silva"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-24", "url": "https://arxiv.org/abs/2405.15934", "pdf_url": "https://arxiv.org/pdf/2405.15934v1", "arxiv_id": "2405.15934", "doi": "10.48550/arXiv.2405.15934", "citation_count": 4, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "2703bac725307796cd1236e2579637588e9fb084144f6573c4cc01b7cbf7c6bc", "sources": ["arxiv", "semantic_scholar"], "title": "ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification", "abstract": "KV cache stores key and value states from previous tokens to avoid re-computation, yet it demands substantial storage space, especially for long sequences. Adaptive KV cache compression seeks to discern the saliency of tokens, preserving vital information while aggressively compressing those of less importance. However, previous methods of this approach exhibit significant performance degradation at high compression ratios due to inaccuracies in identifying salient tokens. In this paper, we present ZipCache, an accurate and efficient KV cache quantization method for LLMs. First, we construct a strong baseline for quantizing KV cache. Through the proposed channel-separable tokenwise quantization scheme, the memory overhead of quantization parameters are substantially reduced compared to fine-grained groupwise quantization. To enhance the compression ratio, we propose normalized attention score as an effective metric for identifying salient tokens by considering the lower triangle characteristics of the attention matrix. Moreover, we develop an efficient approximation method that decouples the saliency metric from full attention scores, enabling compatibility with fast attention implementations like FlashAttention. Extensive experiments demonstrate that ZipCache achieves superior compression ratios, fast generation speed and minimal performance losses compared with previous KV cache compression methods. For instance, when evaluating Mistral-7B model on GSM8k dataset, ZipCache is capable of compressing the KV cache by $4.98\\times$, with only a $0.38\\%$ drop in accuracy. In terms of efficiency, ZipCache also showcases a $37.3\\%$ reduction in prefill-phase latency, a $56.9\\%$ reduction in decoding-phase latency, and a $19.8\\%$ reduction in GPU memory usage when evaluating LLaMA3-8B model with a input length of $4096$.", "authors": ["Yefei He", "Luoming Zhang", "Weijia Wu", "Jing Liu", "Hong Zhou", "Bohan Zhuang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14256", "pdf_url": "https://arxiv.org/pdf/2405.14256v1", "arxiv_id": "2405.14256", "doi": "10.48550/arXiv.2405.14256", "citation_count": 75, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4702}
{"id": "3c4de940db30f87750f235bcf6c1bc9a6a6958549d31e1dad87d468d5aa9b200", "sources": ["arxiv", "semantic_scholar"], "title": "MiniCache: KV Cache Compression in Depth Dimension for Large Language Models", "abstract": "A critical approach for efficiently deploying computationally demanding large language models (LLMs) is Key-Value (KV) caching. The KV cache stores key-value states of previously generated tokens, significantly reducing the need for repetitive computations and thereby lowering latency in autoregressive generation. However, the size of the KV cache grows linearly with sequence length, posing challenges for applications requiring long context input and extensive sequence generation. In this paper, we present a simple yet effective approach, called MiniCache, to compress the KV cache across layers from a novel depth perspective, significantly reducing the memory footprint for LLM inference. Our approach is based on the observation that KV cache states exhibit high similarity between the adjacent layers in the middle-to-deep portion of LLMs. To facilitate merging, we propose disentangling the states into the magnitude and direction components, interpolating the directions of the state vectors while preserving their lengths unchanged. Furthermore, we introduce a token retention strategy to keep highly distinct state pairs unmerged, thus preserving the information with minimal additional storage overhead. Our MiniCache is training-free and general, complementing existing KV cache compression strategies, such as quantization and sparsity. We conduct a comprehensive evaluation of MiniCache utilizing various models including LLaMA-2, LLaMA-3, Phi-3, Mistral, and Mixtral across multiple benchmarks, demonstrating its exceptional performance in achieving superior compression ratios and high throughput. On the ShareGPT dataset, LLaMA-2-7B with 4-bit MiniCache achieves a remarkable compression ratio of up to 5.02x, enhances inference throughput by approximately 5x, and reduces the memory footprint by 41% compared to the FP16 full cache baseline, all while maintaining near-lossless performance.", "authors": ["Akide Liu", "Jing Liu", "Zizheng Pan", "Yefei He", "Gholamreza Haffari", "Bohan Zhuang"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14366", "pdf_url": "https://arxiv.org/pdf/2405.14366v2", "arxiv_id": "2405.14366", "doi": "10.48550/arXiv.2405.14366", "citation_count": 126, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.526}
{"id": "274ddbcc76e383beca69e8ab9132e6e13252544505e40f11278e98739d21786c", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models", "abstract": "The Sparse Mixture of Experts (SMoE) has been widely employed to enhance the efficiency of training and inference for Transformer-based foundational models, yielding promising results.However, the performance of SMoE heavily depends on the choice of hyper-parameters, such as the number of experts and the number of experts to be activated (referred to as top-k), resulting in significant computational overhead due to the extensive model training by searching over various hyper-parameter configurations. As a remedy, we introduce the Dynamic Mixture of Experts (DynMoE) technique. DynMoE incorporates (1) a novel gating method that enables each token to automatically determine the number of experts to activate. (2) An adaptive process automatically adjusts the number of experts during training. Extensive numerical results across Vision, Language, and Vision-Language tasks demonstrate the effectiveness of our approach to achieve competitive performance compared to GMoE for vision and language tasks, and MoE-LLaVA for vision-language tasks, while maintaining efficiency by activating fewer parameters. Our code is available at https://github.com/LINs-lab/DynMoE.", "authors": ["Yongxin Guo", "Zhenglin Cheng", "Xiaoying Tang", "Zhaopeng Tu", "Tao Lin"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14297", "pdf_url": "https://arxiv.org/pdf/2405.14297v4", "arxiv_id": "2405.14297", "doi": "10.48550/arXiv.2405.14297", "citation_count": 51, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/LINs-lab/DynMoE", "venue": "International Conference on Learning Representations", "quality_score": 0.429}
{"id": "f484e49a772dc4bca0e29bfcf940d53eff1bb599be338c796c531ea4d7926e3b", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts Meets Prompt-Based Continual Learning", "abstract": "Exploiting the power of pre-trained models, prompt-based approaches stand out compared to other continual learning solutions in effectively preventing catastrophic forgetting, even with very few learnable parameters and without the need for a memory buffer. While existing prompt-based continual learning methods excel in leveraging prompts for state-of-the-art performance, they often lack a theoretical explanation for the effectiveness of prompting. This paper conducts a theoretical analysis to unravel how prompts bestow such advantages in continual learning, thus offering a new perspective on prompt design. We first show that the attention block of pre-trained models like Vision Transformers inherently encodes a special mixture of experts architecture, characterized by linear experts and quadratic gating score functions. This realization drives us to provide a novel view on prefix tuning, reframing it as the addition of new task-specific experts, thereby inspiring the design of a novel gating mechanism termed Non-linear Residual Gates (NoRGa). Through the incorporation of non-linear activation and residual connection, NoRGa enhances continual learning performance while preserving parameter efficiency. The effectiveness of NoRGa is substantiated both theoretically and empirically across diverse benchmarks and pretraining paradigms. Our code is publicly available at https://github.com/Minhchuyentoancbn/MoE_PromptCL", "authors": ["Minh Le", "An Nguyen", "Huy Nguyen", "Trang Nguyen", "Trang Pham", "Linh Van Ngo", "Nhat Ho"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14124", "pdf_url": "https://arxiv.org/pdf/2405.14124v4", "arxiv_id": "2405.14124", "doi": "10.48550/arXiv.2405.14124", "citation_count": 51, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/Minhchuyentoancbn/MoE_PromptCL", "venue": "Neural Information Processing Systems", "quality_score": 0.429}
{"id": "5926b3d2d98afa3b52047c37bf69248b5a0012278734857b776bbb09dc8a8a64", "sources": ["arxiv", "semantic_scholar"], "title": "Statistical Advantages of Perturbing Cosine Router in Mixture of Experts", "abstract": "The cosine router in Mixture of Experts (MoE) has recently emerged as an attractive alternative to the conventional linear router. Indeed, the cosine router demonstrates favorable performance in image and language tasks and exhibits better ability to mitigate the representation collapse issue, which often leads to parameter redundancy and limited representation potentials. Despite its empirical success, a comprehensive analysis of the cosine router in MoE has been lacking. Considering the least square estimation of the cosine routing MoE, we demonstrate that due to the intrinsic interaction of the model parameters in the cosine router via some partial differential equations, regardless of the structures of the experts, the estimation rates of experts and model parameters can be as slow as $\\mathcal{O}(1/\\log^τ(n))$ where $τ> 0$ is some constant and $n$ is the sample size. Surprisingly, these pessimistic non-polynomial convergence rates can be circumvented by the widely used technique in practice to stabilize the cosine router -- simply adding noises to the $\\ell^2$-norms in the cosine router, which we refer to as \\textit{perturbed cosine router}. Under the strongly identifiable settings of the expert functions, we prove that the estimation rates for both the experts and model parameters under the perturbed cosine routing MoE are significantly improved to polynomial rates. Finally, we conduct extensive simulation studies in both synthetic and real data settings to empirically validate our theoretical results.", "authors": ["Huy Nguyen", "Pedram Akbarian", "Trang Pham", "Trang Nguyen", "Shujian Zhang", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14131", "pdf_url": "https://arxiv.org/pdf/2405.14131v3", "arxiv_id": "2405.14131", "doi": null, "citation_count": 15, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.301}
{"id": "92486a38ccae7ca739ad5487b2656d8b30656440b544d7e735420d392f069c1d", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-Ensemble: Efficient Uncertainty Modelling for Self-Attention Networks", "abstract": "Numerous real-world decisions rely on machine learning algorithms and require calibrated uncertainty estimates. However, modern methods often yield overconfident, uncalibrated predictions. The dominant approach to quantifying the uncertainty inherent in the model is to train an ensemble of separate predictors and measure their empirical variance. In an explicit implementation, the ensemble has a high computational cost and memory footprint, especially if the base model itself is already large, like modern transformers. This motivates efforts to develop implicit ensemble methods that emulate the ensemble without explicitly instantiating all its members. We introduce LoRA-Ensemble, a parameter-efficient ensembling method for self-attention networks. It is based on Low-Rank Adaptation (LoRA), originally developed for efficient LLM fine-tuning, and extends it into an implicit ensembling scheme, where all ensemble members share the same, pre-trained self-attention network, but have individual low-rank matrices for the attention projections. The resulting method not only outperforms state-of-the-art implicit techniques like BatchEnsemble, but even matches or exceeds the accuracy of an Explicit Ensemble, while at the same time achieving superior calibration.", "authors": ["Dominik J. Mühlematter", "Michelle Halbheer", "Alexander Becker", "Dominik Narnhofer", "Helge Aasen", "Konrad Schindler", "Mehmet Ozgur Turkoglu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14438", "pdf_url": "https://arxiv.org/pdf/2405.14438v5", "arxiv_id": "2405.14438", "doi": "10.48550/arXiv.2405.14438", "citation_count": 18, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Transactions on Machine Learning Research, 2026", "quality_score": 0.3197}
{"id": "b32b7c66f43c4d8c0e9a486310c813e648c8288b8eba787b77894d94c6e34ef1", "sources": ["arxiv", "semantic_scholar"], "title": "Sigmoid Gating is More Sample Efficient than Softmax Gating in Mixture of Experts", "abstract": "The softmax gating function is arguably the most popular choice in mixture of experts modeling. Despite its widespread use in practice, the softmax gating may lead to unnecessary competition among experts, potentially causing the undesirable phenomenon of representation collapse due to its inherent structure. In response, the sigmoid gating function has been recently proposed as an alternative and has been demonstrated empirically to achieve superior performance. However, a rigorous examination of the sigmoid gating function is lacking in current literature. In this paper, we verify theoretically that the sigmoid gating, in fact, enjoys a higher sample efficiency than the softmax gating for the statistical task of expert estimation. Towards that goal, we consider a regression framework in which the unknown regression function is modeled as a mixture of experts, and study the rates of convergence of the least squares estimator under the over-specified case in which the number of fitted experts is larger than the true value. We show that two gating regimes naturally arise and, in each of them, we formulate an identifiability condition for the expert functions and derive the corresponding convergence rates. In both cases, we find that experts formulated as feed-forward networks with commonly used activation such as ReLU and GELU enjoy faster convergence rates under the sigmoid gating than those under softmax gating. Furthermore, given the same choice of experts, we demonstrate that the sigmoid gating function requires a smaller sample size than its softmax counterpart to attain the same error of expert estimation and, therefore, is more sample efficient.", "authors": ["Huy Nguyen", "Nhat Ho", "Alessandro Rinaldo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2024-05-22", "url": "https://arxiv.org/abs/2405.13997", "pdf_url": "https://arxiv.org/pdf/2405.13997v3", "arxiv_id": "2405.13997", "doi": "10.48550/arXiv.2405.13997", "citation_count": 28, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3656}
{"id": "24fee02c28a5cdd5753c1e654b98db07e8f18f1655e0d4a4ef2df8aa3094a6ec", "sources": ["arxiv", "semantic_scholar"], "title": "Unlocking Data-free Low-bit Quantization with Matrix Decomposition for KV Cache Compression", "abstract": "Key-value~(KV) caching is an important technique to accelerate the inference of large language models~(LLMs), but incurs significant memory overhead. To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment. In this paper, we introduce \\textbf{DecoQuant}, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache. Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors. Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range. Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor. Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant. Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a $\\sim$75\\% reduction in memory footprint while maintaining comparable generation quality.", "authors": ["Peiyu Liu", "Ze-Feng Gao", "Wayne Xin Zhao", "Yipeng Ma", "Tao Wang", "Ji-Rong Wen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-21", "url": "https://arxiv.org/abs/2405.12591", "pdf_url": "https://arxiv.org/pdf/2405.12591v1", "arxiv_id": "2405.12591", "doi": "10.48550/arXiv.2405.12591", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2603}
{"id": "c8e67550cce9ebecee9c3f098117c6c65441547299930624499f8534963f2452", "sources": ["arxiv", "semantic_scholar"], "title": "PyramidInfer: Pyramid KV Cache Compression for High-throughput LLM Inference", "abstract": "Large Language Models (LLMs) have shown remarkable comprehension abilities but face challenges in GPU memory usage during inference, hindering their scalability for real-time applications like chatbots. To accelerate inference, we store computed keys and values (KV cache) in the GPU memory. Existing methods study the KV cache compression to reduce memory by pruning the pre-computed KV cache. However, they neglect the inter-layer dependency between layers and huge memory consumption in pre-computation. To explore these deficiencies, we find that the number of crucial keys and values that influence future generations decreases layer by layer and we can extract them by the consistency in attention weights. Based on the findings, we propose PyramidInfer, a method that compresses the KV cache by layer-wise retaining crucial context. PyramidInfer saves significant memory by computing fewer keys and values without sacrificing performance. Experimental results show PyramidInfer improves 2.2x throughput compared to Accelerate with over 54% GPU memory reduction in KV cache.", "authors": ["Dongjie Yang", "XiaoDong Han", "Yan Gao", "Yao Hu", "Shilin Zhang", "Hai Zhao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-21", "url": "https://arxiv.org/abs/2405.12532", "pdf_url": "https://arxiv.org/pdf/2405.12532v2", "arxiv_id": "2405.12532", "doi": "10.48550/arXiv.2405.12532", "citation_count": 150, "influential_citation_count": 17, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.6276}
{"id": "e68bc5abba5f51e5b66e345aac7e94641f2b423e5680a4ca5782b2133041439d", "sources": ["arxiv", "semantic_scholar"], "title": "Reducing Transformer Key-Value Cache Size with Cross-Layer Attention", "abstract": "Key-value (KV) caching plays an essential role in accelerating decoding for transformer-based autoregressive large language models (LLMs). However, the amount of memory required to store the KV cache can become prohibitive at long sequence lengths and large batch sizes. Since the invention of the transformer, two of the most effective interventions discovered for reducing the size of the KV cache have been Multi-Query Attention (MQA) and its generalization, Grouped-Query Attention (GQA). MQA and GQA both modify the design of the attention block so that multiple query heads can share a single key/value head, reducing the number of distinct key/value heads by a large factor while only minimally degrading accuracy. In this paper, we show that it is possible to take Multi-Query Attention a step further by also sharing key and value heads between adjacent layers, yielding a new attention design we call Cross-Layer Attention (CLA). With CLA, we find that it is possible to reduce the size of the KV cache by another 2x while maintaining nearly the same accuracy as unmodified MQA. In experiments training 1B- and 3B-parameter models from scratch, we demonstrate that CLA provides a Pareto improvement over the memory/accuracy tradeoffs which are possible with traditional MQA, enabling inference with longer sequence lengths and larger batch sizes than would otherwise be possible", "authors": ["William Brandon", "Mayank Mishra", "Aniruddha Nrusimha", "Rameswar Panda", "Jonathan Ragan Kelly"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-21", "url": "https://arxiv.org/abs/2405.12981", "pdf_url": "https://arxiv.org/pdf/2405.12981v1", "arxiv_id": "2405.12981", "doi": "10.48550/arXiv.2405.12981", "citation_count": 121, "influential_citation_count": 10, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5216}
{"id": "f77d803b9a58fb4bac52e4e60d14b60d13b902308ffa4c4ff5d8dc1a1fe6ce00", "sources": ["arxiv", "semantic_scholar"], "title": "Learning More Generalized Experts by Merging Experts in Mixture-of-Experts", "abstract": "We observe that incorporating a shared layer in a mixture-of-experts can lead to performance degradation. This leads us to hypothesize that learning shared features poses challenges in deep learning, potentially caused by the same feature being learned as various different features. To address this issue, we track each expert's usage frequency and merge the two most frequently selected experts. We then update the least frequently selected expert using the combination of experts. This approach, combined with the subsequent learning of the router's expert selection, allows the model to determine if the most frequently selected experts have learned the same feature differently. If they have, the combined expert can be further trained to learn a more general feature. Consequently, our algorithm enhances transfer learning and mitigates catastrophic forgetting when applied to multi-domain task incremental learning.", "authors": ["Sejik Park"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-19", "url": "https://arxiv.org/abs/2405.11530", "pdf_url": "https://arxiv.org/pdf/2405.11530v1", "arxiv_id": "2405.11530", "doi": "10.48550/arXiv.2405.11530", "citation_count": 11, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "ff725a30566bec0f560967074fe10c1bc8cc5ec2220556579c05b0c89ed16d8c", "sources": ["arxiv", "semantic_scholar"], "title": "MeteoRA: Multiple-tasks Embedded LoRA for Large Language Models", "abstract": "The pretrain+fine-tune paradigm is foundational for deploying large language models (LLMs) across various downstream applications. Within this framework, Low-Rank Adaptation (LoRA) stands out for its parameter-efficient fine-tuning (PEFT), producing numerous reusable task-specific LoRA adapters. However, this approach requires explicit task intention selection, posing challenges for autonomous task sensing and switching during inference with multiple existing LoRA adapters embedded in a single LLM. In this work, we introduce MeteoRA (Multiple-tasks embedded LoRA), a scalable and efficient framework that reuses multiple task-specific LoRA adapters into the base LLM via a full-mode Mixture-of-Experts (MoE) architecture. This framework also includes novel MoE forward acceleration strategies to address the efficiency challenges of traditional MoE implementations. Our evaluation, using the LlaMA2-13B and LlaMA3-8B base models equipped with 28 existing LoRA adapters through MeteoRA, demonstrates equivalent performance with the traditional PEFT method. Moreover, the LLM equipped with MeteoRA achieves superior performance in handling composite tasks, effectively solving ten sequential problems in a single inference pass, thereby demonstrating the framework's enhanced capability for timely adapter switching.", "authors": ["Jingwei Xu", "Junyu Lai", "Yunpeng Huang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-19", "url": "https://arxiv.org/abs/2405.13053", "pdf_url": "https://arxiv.org/pdf/2405.13053v3", "arxiv_id": "2405.13053", "doi": "10.48550/arXiv.2405.13053", "citation_count": 15, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.301}
{"id": "b428310811cfe6eb56efcb8a86ee2338fa30579a149640f548fc3c9d21d213bf", "sources": ["arxiv", "semantic_scholar"], "title": "Uni-MoE: Scaling Unified Multimodal LLMs with Mixture of Experts", "abstract": "Recent advancements in Multimodal Large Language Models (MLLMs) underscore the significance of scalable models and data to boost performance, yet this often incurs substantial computational costs. Although the Mixture of Experts (MoE) architecture has been employed to efficiently scale large language and image-text models, these efforts typically involve fewer experts and limited modalities. To address this, our work presents the pioneering attempt to develop a unified MLLM with the MoE architecture, named Uni-MoE that can handle a wide array of modalities. Specifically, it features modality-specific encoders with connectors for a unified multimodal representation. We also implement a sparse MoE architecture within the LLMs to enable efficient training and inference through modality-level data parallelism and expert-level model parallelism. To enhance the multi-expert collaboration and generalization, we present a progressive training strategy: 1) Cross-modality alignment using various connectors with different cross-modality data, 2) Training modality-specific experts with cross-modality instruction data to activate experts' preferences, and 3) Tuning the Uni-MoE framework utilizing Low-Rank Adaptation (LoRA) on mixed multimodal instruction data. We evaluate the instruction-tuned Uni-MoE on a comprehensive set of multimodal datasets. The extensive experimental results demonstrate Uni-MoE's principal advantage of significantly reducing performance bias in handling mixed multimodal datasets, alongside improved multi-expert collaboration and generalization. Our findings highlight the substantial potential of MoE frameworks in advancing MLLMs and the code is available at https://github.com/HITsz-TMG/UMOE-Scaling-Unified-Multimodal-LLMs.", "authors": ["Yunxin Li", "Shenyuan Jiang", "Baotian Hu", "Longyue Wang", "Wanqi Zhong", "Wenhan Luo", "Lin Ma", "Min Zhang"], "categories": ["cs.AI", "cs.CL", "cs.CV", "cs.MM"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2024-05-18", "url": "https://arxiv.org/abs/2405.11273", "pdf_url": "https://arxiv.org/pdf/2405.11273v1", "arxiv_id": "2405.11273", "doi": "10.1109/TPAMI.2025.3532688", "citation_count": 144, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/HITsz-TMG/UMOE-Scaling-Unified-Multimodal-LLMs", "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.5403}
{"id": "f54bb6b6122628bf6b925096cb0a3cf67f663d038057952f6657ae7c1c66c540", "sources": ["arxiv", "semantic_scholar"], "title": "Layer-Condensed KV Cache for Efficient Inference of Large Language Models", "abstract": "Huge memory consumption has been a major bottleneck for deploying high-throughput large language models in real-world applications. In addition to the large number of parameters, the key-value (KV) cache for the attention mechanism in the transformer architecture consumes a significant amount of memory, especially when the number of layers is large for deep language models. In this paper, we propose a novel method that only computes and caches the KVs of a small number of layers, thus significantly saving memory consumption and improving inference throughput. Our experiments on large language models show that our method achieves up to 26$\\times$ higher throughput than standard transformers and competitive performance in language modeling and downstream tasks. In addition, our method is orthogonal to existing transformer memory-saving techniques, so it is straightforward to integrate them with our model, achieving further improvement in inference efficiency. Our code is available at https://github.com/whyNLP/LCKV.", "authors": ["Haoyi Wu", "Kewei Tu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-17", "url": "https://arxiv.org/abs/2405.10637", "pdf_url": "https://arxiv.org/pdf/2405.10637v2", "arxiv_id": "2405.10637", "doi": "10.48550/arXiv.2405.10637", "citation_count": 59, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/whyNLP/LCKV", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4445}
{"id": "6b65b067d7509192003d2e09885e2961a6494276fbb5f3086ed7e68a04a87cfb", "sources": ["arxiv", "semantic_scholar"], "title": "A Foundation Model for Brain Lesion Segmentation with Mixture of Modality Experts", "abstract": "Brain lesion segmentation plays an essential role in neurological research and diagnosis. As brain lesions can be caused by various pathological alterations, different types of brain lesions tend to manifest with different characteristics on different imaging modalities. Due to this complexity, brain lesion segmentation methods are often developed in a task-specific manner. A specific segmentation model is developed for a particular lesion type and imaging modality. However, the use of task-specific models requires predetermination of the lesion type and imaging modality, which complicates their deployment in real-world scenarios. In this work, we propose a universal foundation model for 3D brain lesion segmentation, which can automatically segment different types of brain lesions for input data of various imaging modalities. We formulate a novel Mixture of Modality Experts (MoME) framework with multiple expert networks attending to different imaging modalities. A hierarchical gating network combines the expert predictions and fosters expertise collaboration. Furthermore, we introduce a curriculum learning strategy during training to avoid the degeneration of each expert network and preserve their specialization. We evaluated the proposed method on nine brain lesion datasets, encompassing five imaging modalities and eight lesion types. The results show that our model outperforms state-of-the-art universal models and provides promising generalization to unseen datasets.", "authors": ["Xinru Zhang", "Ni Ou", "Berke Doga Basaran", "Marco Visentin", "Mengyun Qiao", "Renyang Gu", "Cheng Ouyang", "Yaou Liu", "Paul M. Matthew", "Chuyang Ye", "Wenjia Bai"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-05-16", "url": "https://arxiv.org/abs/2405.10246", "pdf_url": "https://arxiv.org/pdf/2405.10246v2", "arxiv_id": "2405.10246", "doi": "10.48550/arXiv.2405.10246", "citation_count": 24, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Medical Image Computing and Computer-Assisted Intervention", "quality_score": 0.3495}
{"id": "ac987478de395c1c2d9cd99fb1fdc913eb54c0d682a0d5d98616a885c1516433", "sources": ["arxiv", "semantic_scholar"], "title": "M$^4$oE: A Foundation Model for Medical Multimodal Image Segmentation with Mixture of Experts", "abstract": "Medical imaging data is inherently heterogeneous across different modalities and clinical centers, posing unique challenges for developing generalizable foundation models. Conventional entails training distinct models per dataset or using a shared encoder with modality-specific decoders. However, these approaches incur heavy computational overheads and suffer from poor scalability. To address these limitations, we propose the Medical Multimodal Mixture of Experts (M$^4$oE) framework, leveraging the SwinUNet architecture. Specifically, M$^4$oE comprises modality-specific experts; each separately initialized to learn features encoding domain knowledge. Subsequently, a gating network is integrated during fine-tuning to modulate each expert's contribution to the collective predictions dynamically. This enhances model interpretability and generalization ability while retaining expertise specialization. Simultaneously, the M$^4$oE architecture amplifies the model's parallel processing capabilities, and it also ensures the model's adaptation to new modalities with ease. Experiments across three modalities reveal that M$^4$oE can achieve 3.45% over STU-Net-L, 5.11% over MED3D, and 11.93% over SAM-Med2D across the MICCAI FLARE22, AMOS2022, and ATLAS2023 datasets. Moreover, M$^4$oE showcases a significant reduction in training duration with 7 hours less while maintaining a parameter count that is only 30% of its compared methods. The code is available at https://github.com/JefferyJiang-YF/M4oE.", "authors": ["Yufeng Jiang", "Yiqing Shen"], "categories": ["eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-05-15", "url": "https://arxiv.org/abs/2405.09446", "pdf_url": "https://arxiv.org/pdf/2405.09446v1", "arxiv_id": "2405.09446", "doi": "10.1007/978-3-031-72390-2_58", "citation_count": 41, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/JefferyJiang-YF/M4oE", "venue": "International Conference on Medical Image Computing and Computer-Assisted Intervention", "quality_score": 0.4058}
{"id": "a8d79809b3575cfbd23134e3feb39ffd4f77ba493d928dda72f9429e232d2809", "sources": ["arxiv", "semantic_scholar"], "title": "Decision Mamba Architectures", "abstract": "Recent advancements in imitation learning have been largely fueled by the integration of sequence models, which provide a structured flow of information to effectively mimic task behaviours. Currently, Decision Transformer (DT) and subsequently, the Hierarchical Decision Transformer (HDT), presented Transformer-based approaches to learn task policies. Recently, the Mamba architecture has shown to outperform Transformers across various task domains. In this work, we introduce two novel methods, Decision Mamba (DM) and Hierarchical Decision Mamba (HDM), aimed at enhancing the performance of the Transformer models. Through extensive experimentation across diverse environments such as OpenAI Gym and D4RL, leveraging varying demonstration data sets, we demonstrate the superiority of Mamba models over their Transformer counterparts in a majority of tasks. Results show that DM outperforms other methods in most settings. The code can be found at https://github.com/meowatthemoon/DecisionMamba.", "authors": ["André Correia", "Luís A. Alexandre"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-13", "url": "https://arxiv.org/abs/2405.07943", "pdf_url": "https://arxiv.org/pdf/2405.07943v2", "arxiv_id": "2405.07943", "doi": null, "citation_count": 1, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/meowatthemoon/DecisionMamba", "venue": null, "quality_score": 0.1505}
{"id": "9f70a6ab2dcf0492fbafac78a89abdcdb893c2ef76a7c68bcfd262d288e8489c", "sources": ["arxiv", "semantic_scholar"], "title": "SKVQ: Sliding-window Key and Value Cache Quantization for Large Language Models", "abstract": "Large language models (LLMs) can now handle longer sequences of tokens, enabling complex tasks like book understanding and generating lengthy novels. However, the key-value (KV) cache required for LLMs consumes substantial memory as context length increasing, becoming the bottleneck for deployment. In this paper, we present a strategy called SKVQ, which stands for sliding-window KV cache quantization, to address the issue of extremely low bitwidth KV cache quantization. To achieve this, SKVQ rearranges the channels of the KV cache in order to improve the similarity of channels in quantization groups, and applies clipped dynamic quantization at the group level. Additionally, SKVQ ensures that the most recent window tokens in the KV cache are preserved with high precision. This helps maintain the accuracy of a small but important portion of the KV cache.SKVQ achieves high compression ratios while maintaining accuracy. Our evaluation on LLMs demonstrates that SKVQ surpasses previous quantization approaches, allowing for quantization of the KV cache to 2-bit keys and 1.5-bit values with minimal loss of accuracy. With SKVQ, it is possible to process context lengths of up to 1M on an 80GB memory GPU for a 7b model and up to 7 times faster decoding.", "authors": ["Haojie Duanmu", "Zhihang Yuan", "Xiuhong Li", "Jiangfei Duan", "Xingcheng Zhang", "Dahua Lin"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-10", "url": "https://arxiv.org/abs/2405.06219", "pdf_url": "https://arxiv.org/pdf/2405.06219v3", "arxiv_id": "2405.06219", "doi": "10.48550/arXiv.2405.06219", "citation_count": 36, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5}
{"id": "54538da3a105f0a698144d5d098676de884a7f15020e150acad5da868698f1a1", "sources": ["arxiv", "semantic_scholar"], "title": "A Mixture of Experts Approach to 3D Human Motion Prediction", "abstract": "This project addresses the challenge of human motion prediction, a critical area for applications such as au- tonomous vehicle movement detection. Previous works have emphasized the need for low inference times to provide real time performance for applications like these. Our primary objective is to critically evaluate existing model ar- chitectures, identifying their advantages and opportunities for improvement by replicating the state-of-the-art (SOTA) Spatio-Temporal Transformer model as best as possible given computational con- straints. These models have surpassed the limitations of RNN-based models and have demonstrated the ability to generate plausible motion sequences over both short and long term horizons through the use of spatio-temporal rep- resentations. We also propose a novel architecture to ad- dress challenges of real time inference speed by incorpo- rating a Mixture of Experts (MoE) block within the Spatial- Temporal (ST) attention layer. The particular variation that is used is Soft MoE, a fully-differentiable sparse Transformer that has shown promising ability to enable larger model capacity at lower inference cost. We make out code publicly available at https://github.com/edshieh/motionprediction", "authors": ["Edmund Shieh", "Joshua Lee Franco", "Kang Min Bae", "Tej Lalvani"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-09", "url": "https://arxiv.org/abs/2405.06088", "pdf_url": "https://arxiv.org/pdf/2405.06088v1", "arxiv_id": "2405.06088", "doi": "10.48550/arXiv.2405.06088", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/edshieh/motionprediction", "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "6690108cebd450fb052233400d921cbff5a810300889641bb4ee87dd01380d0b", "sources": ["arxiv", "semantic_scholar"], "title": "A Mixture-of-Experts Approach to Few-Shot Task Transfer in Open-Ended Text Worlds", "abstract": "Open-ended worlds are those in which there are no pre-specified goals or environmental reward signal. As a consequence, an agent must know how to perform a multitude of tasks. However, when a new task is presented to an agent, we expect it to be able to reuse some of what it knows from previous tasks to rapidly learn that new task. We introduce a novel technique whereby policies for different a priori known tasks are combined into a Mixture-of-Experts model with an attention mechanism across a mix of frozen and unfrozen experts. The model learns when to attend to frozen task-specific experts when appropriate and learns new experts to handle novel situations. We work in an open-ended text-based environment in which the agent is tasked with behaving like different types of character roles and must rapidly learn behaviors associated with new character role types. We show that our agent both obtains more rewards in the zero-shot setting, and discovers these rewards with greater sample efficiency in the few-shot learning settings.", "authors": ["Christopher Z. Cui", "Xiangyu Peng", "Mark O. Riedl"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-09", "url": "https://arxiv.org/abs/2405.06059", "pdf_url": "https://arxiv.org/pdf/2405.06059v1", "arxiv_id": "2405.06059", "doi": "10.48550/arXiv.2405.06059", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "e22766ac69ae224c0e07a0b413c04216e96a96da1f00bcdf0000d93ba6eab542", "sources": ["arxiv", "semantic_scholar"], "title": "CuMo: Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts", "abstract": "Recent advancements in Multimodal Large Language Models (LLMs) have focused primarily on scaling by increasing text-image pair data and enhancing LLMs to improve performance on multimodal tasks. However, these scaling approaches are computationally expensive and overlook the significance of improving model capabilities from the vision side. Inspired by the successful applications of Mixture-of-Experts (MoE) in LLMs, which improves model scalability during training while keeping inference costs similar to those of smaller models, we propose CuMo. CuMo incorporates Co-upcycled Top-K sparsely-gated Mixture-of-experts blocks into both the vision encoder and the MLP connector, thereby enhancing the multimodal LLMs with minimal additional activated parameters during inference. CuMo first pre-trains the MLP blocks and then initializes each expert in the MoE block from the pre-trained MLP block during the visual instruction tuning stage. Auxiliary losses are used to ensure a balanced loading of experts. CuMo outperforms state-of-the-art multimodal LLMs across various VQA and visual-instruction-following benchmarks using models within each model size group, all while training exclusively on open-sourced datasets. The code and model weights for CuMo are open-sourced at https://github.com/SHI-Labs/CuMo.", "authors": ["Jiachen Li", "Xinyao Wang", "Sijie Zhu", "Chia-Wen Kuo", "Lu Xu", "Fan Chen", "Jitesh Jain", "Humphrey Shi", "Longyin Wen"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-09", "url": "https://arxiv.org/abs/2405.05949", "pdf_url": "https://arxiv.org/pdf/2405.05949v1", "arxiv_id": "2405.05949", "doi": "10.48550/arXiv.2405.05949", "citation_count": 71, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/SHI-Labs/CuMo", "venue": "Neural Information Processing Systems", "quality_score": 0.4643}
{"id": "8f6acd0228ad0a6fa9b43d948f60d6427a30814984b56c72f5ff029b724dfa81", "sources": ["arxiv", "semantic_scholar"], "title": "KV-Runahead: Scalable Causal LLM Inference by Parallel Key-Value Cache Generation", "abstract": "Large Language Model or LLM inference has two phases, the prompt (or prefill) phase to output the first token and the extension (or decoding) phase to the generate subsequent tokens. In this work, we propose an efficient parallelization scheme, KV-Runahead to accelerate the prompt phase. The key observation is that the extension phase generates tokens faster than the prompt phase because of key-value cache (KV-cache). Hence, KV-Runahead parallelizes the prompt phase by orchestrating multiple processes to populate the KV-cache and minimizes the time-to-first-token (TTFT). Dual-purposing the KV-cache scheme has two main benefits. First, since KV-cache is designed to leverage the causal attention map, we minimize computation and computation automatically. Second, since it already exists for the extension phase, KV-Runahead is easy to implement. We further propose context-level load-balancing to handle uneven KV-cache generation (due to the causal attention) and to optimize TTFT. Compared with an existing parallelization scheme such as tensor or sequential parallelization where keys and values are locally generated and exchanged via all-gather collectives, our experimental results demonstrate that KV-Runahead can offer over 1.4x and 1.6x speedups for Llama 7B and Falcon 7B respectively.", "authors": ["Minsik Cho", "Mohammad Rastegari", "Devang Naik"], "categories": ["cs.DC", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-08", "url": "https://arxiv.org/abs/2405.05329", "pdf_url": "https://arxiv.org/pdf/2405.05329v2", "arxiv_id": "2405.05329", "doi": "10.48550/arXiv.2405.05329", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2785}
{"id": "fd4fbe415b2efcd80a15f270f3d6c59516294ae7742fb5bb89d153abc641d657", "sources": ["arxiv", "semantic_scholar"], "title": "Vision Mamba: A Comprehensive Survey and Taxonomy", "abstract": "State Space Model (SSM) is a mathematical model used to describe and analyze the behavior of dynamic systems. This model has witnessed numerous applications in several fields, including control theory, signal processing, economics and machine learning. In the field of deep learning, state space models are used to process sequence data, such as time series analysis, natural language processing (NLP) and video understanding. By mapping sequence data to state space, long-term dependencies in the data can be better captured. In particular, modern SSMs have shown strong representational capabilities in NLP, especially in long sequence modeling, while maintaining linear time complexity. Notably, based on the latest state-space models, Mamba merges time-varying parameters into SSMs and formulates a hardware-aware algorithm for efficient training and inference. Given its impressive efficiency and strong long-range dependency modeling capability, Mamba is expected to become a new AI architecture that may outperform Transformer. Recently, a number of works have attempted to study the potential of Mamba in various fields, such as general vision, multi-modal, medical image analysis and remote sensing image analysis, by extending Mamba from natural language domain to visual domain. To fully understand Mamba in the visual domain, we conduct a comprehensive survey and present a taxonomy study. This survey focuses on Mamba's application to a variety of visual tasks and data types, and discusses its predecessors, recent advances and far-reaching impact on a wide range of domains. Since Mamba is now on an upward trend, please actively notice us if you have new findings, and new progress on Mamba will be included in this survey in a timely manner and updated on the Mamba project at https://github.com/lx6c78/Vision-Mamba-A-Comprehensive-Survey-and-Taxonomy.", "authors": ["Xiao Liu", "Chenxu Zhang", "Lei Zhang"], "categories": ["cs.CV", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-07", "url": "https://arxiv.org/abs/2405.04404", "pdf_url": "https://arxiv.org/pdf/2405.04404v1", "arxiv_id": "2405.04404", "doi": "10.48550/arXiv.2405.04404", "citation_count": 139, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/lx6c78/Vision-Mamba-A-Comprehensive-Survey-and-Taxonomy", "venue": "arXiv.org", "quality_score": 0.5365}
{"id": "c7ca818515634a4a3da4e6e1a8f65b66f1674f0489a3e97799c9c4db6d05a953", "sources": ["arxiv", "semantic_scholar"], "title": "KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization", "abstract": "Efficient deployment of Large Language Models (LLMs) requires batching multiple requests together to improve throughput. As the batch size, context length, or model size increases, the size of the key and value (KV) cache can quickly become the main contributor to GPU memory usage and the bottleneck of inference latency. Quantization has emerged as an effective technique for KV cache compression, but existing methods still fail at very low bit widths. We observe that distinct channels of a key/value activation embedding are highly inter-dependent, and the joint entropy of multiple channels grows at a slower rate than the sum of their marginal entropies. Based on this insight, we propose Coupled Quantization (CQ), which couples multiple key/value channels together to exploit their inter-dependency and encode the activations in a more information-efficient manner. Extensive experiments reveal that CQ outperforms or is competitive with existing baselines in preserving model quality. Furthermore, we demonstrate that CQ can preserve model quality with KV cache quantized down to 1-bit.", "authors": ["Tianyi Zhang", "Jonah Yi", "Zhaozhuo Xu", "Anshumali Shrivastava"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-07", "url": "https://arxiv.org/abs/2405.03917", "pdf_url": "https://arxiv.org/pdf/2405.03917v1", "arxiv_id": "2405.03917", "doi": "10.48550/arXiv.2405.03917", "citation_count": 81, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4785}
{"id": "437219ad4747a1ea29efb5a8d530e6df1d39c16319b8e4429078e09fc4a54513", "sources": ["arxiv", "semantic_scholar"], "title": "QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving", "abstract": "Quantization can accelerate large language model (LLM) inference. Going beyond INT8 quantization, the research community is actively exploring even lower precision, such as INT4. Nonetheless, state-of-the-art INT4 quantization techniques only accelerate low-batch, edge LLM inference, failing to deliver performance gains in large-batch, cloud-based LLM serving. We uncover a critical issue: existing INT4 quantization methods suffer from significant runtime overhead (20-90%) when dequantizing either weights or partial sums on GPUs. To address this challenge, we introduce QoQ, a W4A8KV4 quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache. QoQ stands for quattuor-octo-quattuor, which represents 4-8-4 in Latin. QoQ is implemented by the QServe inference library that achieves measured speedup. The key insight driving QServe is that the efficiency of LLM serving on GPUs is critically influenced by operations on low-throughput CUDA cores. Building upon this insight, in QoQ algorithm, we introduce progressive quantization that can allow low dequantization overhead in W4A8 GEMM. Additionally, we develop SmoothAttention to effectively mitigate the accuracy degradation incurred by 4-bit KV quantization. In the QServe system, we perform compute-aware weight reordering and take advantage of register-level parallelism to reduce dequantization latency. We also make fused attention memory-bound, harnessing the performance gain brought by KV4 quantization. As a result, QServe improves the maximum achievable serving throughput of Llama-3-8B by 1.2x on A100, 1.4x on L40S; and Qwen1.5-72B by 2.4x on A100, 3.5x on L40S, compared to TensorRT-LLM. Remarkably, QServe on L40S GPU can achieve even higher throughput than TensorRT-LLM on A100. Thus, QServe effectively reduces the dollar cost of LLM serving by 3x. Code is available at https://github.com/mit-han-lab/omniserve.", "authors": ["Yujun Lin", "Haotian Tang", "Shang Yang", "Zhekai Zhang", "Guangxuan Xiao", "Chuang Gan", "Song Han"], "categories": ["cs.CL", "cs.AI", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-07", "url": "https://arxiv.org/abs/2405.04532", "pdf_url": "https://arxiv.org/pdf/2405.04532v3", "arxiv_id": "2405.04532", "doi": "10.48550/arXiv.2405.04532", "citation_count": 216, "influential_citation_count": 38, "has_code": true, "code_url": "https://github.com/mit-han-lab/omniserve", "venue": "Conference on Machine Learning and Systems", "quality_score": 0.7955}
{"id": "7c571637f7ccda3eeb5d810b3236556abf10a6128d58ce0e0680d0bdf0767e08", "sources": ["arxiv", "semantic_scholar"], "title": "Simple Drop-in LoRA Conditioning on Attention Layers Will Improve Your Diffusion Model", "abstract": "Current state-of-the-art diffusion models employ U-Net architectures containing convolutional and (qkv) self-attention layers. The U-Net processes images while being conditioned on the time embedding input for each sampling step and the class or caption embedding input corresponding to the desired conditional generation. Such conditioning involves scale-and-shift operations to the convolutional layers but does not directly affect the attention layers. While these standard architectural choices are certainly effective, not conditioning the attention layers feels arbitrary and potentially suboptimal. In this work, we show that simply adding LoRA conditioning to the attention layers without changing or tuning the other parts of the U-Net architecture improves the image generation quality. For example, a drop-in addition of LoRA conditioning to EDM diffusion model yields FID scores of 1.91/1.75 for unconditional and class-conditional CIFAR-10 generation, improving upon the baseline of 1.97/1.79.", "authors": ["Joo Young Choi", "Jaesung R. Park", "Inkyu Park", "Jaewoong Cho", "Albert No", "Ernest K. Ryu"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-07", "url": "https://arxiv.org/abs/2405.03958", "pdf_url": "https://arxiv.org/pdf/2405.03958v3", "arxiv_id": "2405.03958", "doi": "10.48550/arXiv.2405.03958", "citation_count": 11, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2698}
{"id": "a08df7911fb1faca8950b4af478c3a4c5b121f16d9713fc2b631a79b77d46cfe", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Physical Layer Communication Security through Generative AI with Mixture of Experts", "abstract": "AI technologies have become more widely adopted in wireless communications. As an emerging type of AI technologies, the generative artificial intelligence (GAI) gains lots of attention in communication security. Due to its powerful learning ability, GAI models have demonstrated superiority over conventional AI methods. However, GAI still has several limitations, including high computational complexity and limited adaptability. Mixture of Experts (MoE), which uses multiple expert models for prediction through a gate mechanism, proposes possible solutions. Firstly, we review GAI model's applications in physical layer communication security, discuss limitations, and explore how MoE can help GAI overcome these limitations. Furthermore, we propose an MoE-enabled GAI framework for network optimization problems for communication security. To demonstrate the framework's effectiveness, we provide a case study in a cooperative friendly jamming scenario. The experimental results show that the MoE-enabled framework effectively assists the GAI algorithm, solves its limitations, and enhances communication security.", "authors": ["Changyuan Zhao", "Hongyang Du", "Dusit Niyato", "Jiawen Kang", "Zehui Xiong", "Dong In Kim", " Xuemin", " Shen", "Khaled B. Letaief"], "categories": ["cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-07", "url": "https://arxiv.org/abs/2405.04198", "pdf_url": "https://arxiv.org/pdf/2405.04198v1", "arxiv_id": "2405.04198", "doi": "10.1109/MWC.001.2400150", "citation_count": 25, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE wireless communications", "quality_score": 0.3537}
{"id": "c30468920e35510aa1997ca49b27cdd09773fadf23447b9d14ef4f248f9a3e99", "sources": ["arxiv", "semantic_scholar"], "title": "DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model", "abstract": "We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.", "authors": [" DeepSeek-AI", "Aixin Liu", "Bei Feng", "Bin Wang", "Bingxuan Wang", "Bo Liu", "Chenggang Zhao", "Chengqi Dengr", "Chong Ruan", "Damai Dai", "Daya Guo", "Dejian Yang", "Deli Chen", "Dongjie Ji", "Erhang Li", "Fangyun Lin", "Fuli Luo", "Guangbo Hao", "Guanting Chen", "Guowei Li", "H. Zhang", "Hanwei Xu", "Hao Yang", "Haowei Zhang", "Honghui Ding", "Huajian Xin", "Huazuo Gao", "Hui Li", "Hui Qu", "J. L. Cai", "Jian Liang", "Jianzhong Guo", "Jiaqi Ni", "Jiashi Li", "Jin Chen", "Jingyang Yuan", "Junjie Qiu", "Junxiao Song", "Kai Dong", "Kaige Gao", "Kang Guan", "Lean Wang", "Lecong Zhang", "Lei Xu", "Leyi Xia", "Liang Zhao", "Liyue Zhang", "Meng Li", "Miaojun Wang", "Mingchuan Zhang", "Minghua Zhang", "Minghui Tang", "Mingming Li", "Ning Tian", "Panpan Huang", "Peiyi Wang", "Peng Zhang", "Qihao Zhu", "Qinyu Chen", "Qiushi Du", "R. J. Chen", "R. L. Jin", "Ruiqi Ge", "Ruizhe Pan", "Runxin Xu", "Ruyi Chen", "S. S. Li", "Shanghao Lu", "Shangyan Zhou", "Shanhuang Chen", "Shaoqing Wu", "Shengfeng Ye", "Shirong Ma", "Shiyu Wang", "Shuang Zhou", "Shuiping Yu", "Shunfeng Zhou", "Size Zheng", "T. Wang", "Tian Pei", "Tian Yuan", "Tianyu Sun", "W. L. Xiao", "Wangding Zeng", "Wei An", "Wen Liu", "Wenfeng Liang", "Wenjun Gao", "Wentao Zhang", "X. Q. Li", "Xiangyue Jin", "Xianzu Wang", "Xiao Bi", "Xiaodong Liu", "Xiaohan Wang", "Xiaojin Shen", "Xiaokang Chen", "Xiaosha Chen", "Xiaotao Nie", "Xiaowen Sun", "Xiaoxiang Wang", "Xin Liu", "Xin Xie", "Xingkai Yu", "Xinnan Song", "Xinyi Zhou", "Xinyu Yang", "Xuan Lu", "Xuecheng Su", "Y. Wu", "Y. K. Li", "Y. X. Wei", "Y. X. Zhu", "Yanhong Xu", "Yanping Huang", "Yao Li", "Yao Zhao", "Yaofeng Sun", "Yaohui Li", "Yaohui Wang", "Yi Zheng", "Yichao Zhang", "Yiliang Xiong", "Yilong Zhao", "Ying He", "Ying Tang", "Yishi Piao", "Yixin Dong", "Yixuan Tan", "Yiyuan Liu", "Yongji Wang", "Yongqiang Guo", "Yuchen Zhu", "Yuduan Wang", "Yuheng Zou", "Yukun Zha", "Yunxian Ma", "Yuting Yan", "Yuxiang You", "Yuxuan Liu", "Z. Z. Ren", "Zehui Ren", "Zhangli Sha", "Zhe Fu", "Zhen Huang", "Zhen Zhang", "Zhenda Xie", "Zhewen Hao", "Zhihong Shao", "Zhiniu Wen", "Zhipeng Xu", "Zhongyu Zhang", "Zhuoshu Li", "Zihan Wang", "Zihui Gu", "Zilin Li", "Ziwei Xie"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-07", "url": "https://arxiv.org/abs/2405.04434", "pdf_url": "https://arxiv.org/pdf/2405.04434v5", "arxiv_id": "2405.04434", "doi": "10.48550/arXiv.2405.04434", "citation_count": 1241, "influential_citation_count": 123, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 1.0}
{"id": "f5823d06263087791a7263c92a4a9fffaa8a16d7fd4088fa0955b70d8ce4175a", "sources": ["arxiv", "semantic_scholar"], "title": "MEET: Mixture of Experts Extra Tree-Based sEMG Hand Gesture Identification", "abstract": "Artificial intelligence (AI) has made significant advances in recent years and opened up new possibilities in exploring applications in various fields such as biomedical, robotics, education, industry, etc. Among these fields, human hand gesture recognition is a subject of study that has recently emerged as a research interest in robotic hand control using electromyography (EMG). Surface electromyography (sEMG) is a primary technique used in EMG, which is popular due to its non-invasive nature and is used to capture gesture movements using signal acquisition devices placed on the surface of the forearm. Moreover, these signals are pre-processed to extract significant handcrafted features through time and frequency domain analysis. These are helpful and act as input to machine learning (ML) models to identify hand gestures. However, handling multiple classes and biases are major limitations that can affect the performance of an ML model. Therefore, to address this issue, a new mixture of experts extra tree (MEET) model is proposed to identify more accurate and effective hand gesture movements. This model combines individual ML models referred to as experts, each focusing on a minimal class of two. Moreover, a fully trained model known as the gate is employed to weigh the output of individual expert models. This amalgamation of the expert models with the gate model is known as a mixture of experts extra tree (MEET) model. In this study, four subjects with six hand gesture movements have been considered and their identification is evaluated among eleven models, including the MEET classifier. Results elucidate that the MEET classifier performed best among other algorithms and identified hand gesture movement accurately.", "authors": ["Naveen Gehlot", "Ashutosh Jena", "Rajesh Kumar", "Mahipal Bukya"], "categories": ["eess.SP", "cs.LG"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-05-06", "url": "https://arxiv.org/abs/2405.09562", "pdf_url": "https://arxiv.org/pdf/2405.09562v1", "arxiv_id": "2405.09562", "doi": "10.48550/arXiv.2405.09562", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "c5168b110f21e8fd34aadfbf700052b1389cea6d861eeba58d993edd424161f5", "sources": ["arxiv", "semantic_scholar"], "title": "WDMoE: Wireless Distributed Large Language Models with Mixture of Experts", "abstract": "Large Language Models (LLMs) have achieved significant success in various natural language processing tasks, but how wireless communications can support LLMs has not been extensively studied. In this paper, we propose a wireless distributed LLMs paradigm based on Mixture of Experts (MoE), named WDMoE, deploying LLMs collaboratively across edge servers of base station (BS) and mobile devices in the wireless communications system. Specifically, we decompose the MoE layer in LLMs by deploying the gating network and the preceding neural network layer at BS, while distributing the expert networks across the devices. This arrangement leverages the parallel capabilities of expert networks on distributed devices. Moreover, to overcome the instability of wireless communications, we design an expert selection policy by taking into account both the performance of the model and the end-to-end latency, which includes both transmission delay and inference delay. Evaluations conducted across various LLMs and multiple datasets demonstrate that WDMoE not only outperforms existing models, such as Llama 2 with 70 billion parameters, but also significantly reduces end-to-end latency.", "authors": ["Nan Xue", "Yaping Sun", "Zhiyong Chen", "Meixia Tao", "Xiaodong Xu", "Liang Qian", "Shuguang Cui", "Ping Zhang"], "categories": ["cs.IT", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-06", "url": "https://arxiv.org/abs/2405.03131", "pdf_url": "https://arxiv.org/pdf/2405.03131v1", "arxiv_id": "2405.03131", "doi": "10.1109/GLOBECOM52923.2024.10901084", "citation_count": 25, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Global Communications Conference", "quality_score": 0.3537}
{"id": "0facaab7794865af13d95d8a7ac6e19c19adb2e294d59bbfa19526db5898d4fd", "sources": ["arxiv", "semantic_scholar"], "title": "Is Flash Attention Stable?", "abstract": "Training large-scale machine learning models poses distinct system challenges, given both the size and complexity of today's workloads. Recently, many organizations training state-of-the-art Generative AI models have reported cases of instability during training, often taking the form of loss spikes. Numeric deviation has emerged as a potential cause of this training instability, although quantifying this is especially challenging given the costly nature of training runs. In this work, we develop a principled approach to understanding the effects of numeric deviation, and construct proxies to put observations into context when downstream effects are difficult to quantify. As a case study, we apply this framework to analyze the widely-adopted Flash Attention optimization. We find that Flash Attention sees roughly an order of magnitude more numeric deviation as compared to Baseline Attention at BF16 when measured during an isolated forward pass. We then use a data-driven analysis based on the Wasserstein Distance to provide upper bounds on how this numeric deviation impacts model weights during training, finding that the numerical deviation present in Flash Attention is 2-5 times less significant than low-precision training.", "authors": ["Alicia Golden", "Samuel Hsia", "Fei Sun", "Bilge Acun", "Basil Hosmer", "Yejin Lee", "Zachary DeVito", "Jeff Johnson", "Gu-Yeon Wei", "David Brooks", "Carole-Jean Wu"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-05", "url": "https://arxiv.org/abs/2405.02803", "pdf_url": "https://arxiv.org/pdf/2405.02803v1", "arxiv_id": "2405.02803", "doi": "10.48550/arXiv.2405.02803", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "ccd8bc98dc355db535fc37a27ff6a9cf99f4e12dd15643ad5a42573d2d082908", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of partially linear experts", "abstract": "In the mixture of experts model, a common assumption is the linearity between a response variable and covariates. While this assumption has theoretical and computational benefits, it may lead to suboptimal estimates by overlooking potential nonlinear relationships among the variables. To address this limitation, we propose a partially linear structure that incorporates unspecified functions to capture nonlinear relationships. We establish the identifiability of the proposed model under mild conditions and introduce a practical estimation algorithm. We present the performance of our approach through numerical studies, including simulations and real data analysis.", "authors": ["Yeongsan Hwang", "Byungtae Seo", "Sangkon Oh"], "categories": ["stat.ME", "stat.ML"], "fields_of_study": ["Mathematics"], "published_date": "2024-05-05", "url": "https://arxiv.org/abs/2405.02905", "pdf_url": "https://arxiv.org/pdf/2405.02905v1", "arxiv_id": "2405.02905", "doi": "10.1002/sta4.70062", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Stat, Volume 14, Issue 2, e70062, 2025", "quality_score": 0.0753}
{"id": "7bed596e5fd899daa5e4d171e9b14934286ef115687dd7be5a3b362b3bebe693", "sources": ["arxiv", "semantic_scholar"], "title": "Estimating Complier Average Causal Effects with Mixtures of Experts", "abstract": "Treatment non-compliance, where individuals deviate from their assigned experimental conditions, frequently complicates the estimation of causal effects. To address this, we introduce a novel learning framework based on a mixture of experts architecture to estimate the Complier Average Causal Effect (CACE). Our framework provides a flexible alternative to classical instrumental variable methods by relaxing their strict monotonicity and exclusion restriction assumptions. We develop a principled, two-step procedure where each step is optimized with a dedicated Expectation-Maximization (EM) algorithm. Crucially, we provide formal proofs that the model's components are identifiable, ensuring the learning procedure is well-posed. The resulting CACE estimators are proven to be consistent and asymptotically normal. Extensive simulations demonstrate that our method achieves a substantially lower root mean squared error than traditional instrumental variable approaches when their assumptions fail, an advantage that persists even when our own mixture of experts are misspecified. We illustrate the framework's practical utility on data from a large-scale randomized trial.", "authors": ["François Grolleau", "Céline Béji", "Raphaël Porcher", "François Petit"], "categories": ["stat.ME"], "fields_of_study": ["Mathematics"], "published_date": "2024-05-05", "url": "https://arxiv.org/abs/2405.02779", "pdf_url": "https://arxiv.org/pdf/2405.02779v2", "arxiv_id": "2405.02779", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "c757f129491357b28785740601d608b2b1b4e0e9a2481c53890b545fefafd427", "sources": ["arxiv", "semantic_scholar"], "title": "AdaMoLE: Fine-Tuning Large Language Models with Adaptive Mixture of Low-Rank Adaptation Experts", "abstract": "We introduce AdaMoLE, a novel method for fine-tuning large language models (LLMs) through an Adaptive Mixture of Low-Rank Adaptation (LoRA) Experts. Moving beyond conventional methods that employ a static top-k strategy for activating experts, AdaMoLE dynamically adjusts the activation threshold using a dedicated threshold network, adaptively responding to the varying complexities of different tasks. By replacing a single LoRA in a layer with multiple LoRA experts and integrating a gating function with the threshold mechanism, AdaMoLE effectively selects and activates the most appropriate experts based on the input context. Our extensive evaluations across a variety of commonsense reasoning and natural language processing tasks show that AdaMoLE exceeds baseline performance. This enhancement highlights the advantages of AdaMoLE's adaptive selection of LoRA experts, improving model effectiveness without a corresponding increase in the expert count. The experimental validation not only confirms AdaMoLE as a robust approach for enhancing LLMs but also suggests valuable directions for future research in adaptive expert selection mechanisms, potentially broadening the scope for optimizing model performance across diverse language processing tasks.", "authors": ["Zefang Liu", "Jiahua Luo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-01", "url": "https://arxiv.org/abs/2405.00361", "pdf_url": "https://arxiv.org/pdf/2405.00361v2", "arxiv_id": "2405.00361", "doi": "10.48550/arXiv.2405.00361", "citation_count": 35, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3891}
{"id": "249a5a4822bf1c83f6e3d35822b9296d0f5d6f5f82ca9eda481e940af348e43c", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of insighTful Experts (MoTE): The Synergy of Thought Chains and Expert Mixtures in Self-Alignment", "abstract": "As the capabilities of large language models (LLMs) continue to expand, aligning these models with human values remains a significant challenge. Recent studies show that reasoning abilities contribute significantly to model safety, while integrating Mixture-of-Experts (MoE) architectures can further enhance alignment. In this work, we address a fundamental question: How to effectively incorporate reasoning abilities and MoE architectures into self-alignment process in LLMs? We propose Mixture of insighTful Experts (MoTE), a novel framework that synergistically combines reasoning chains and expert mixtures to improve self-alignments. From a data perspective, MoTE employs a structured reasoning chain comprising four key stages: Question Analysis, Answer Guidance, Safe Answer, and Safety Checking. This approach enhances safety through multi-step reasoning and proves effective even for smaller and less powerful LLMs (e.g., 7B models). From an architectural perspective, MoTE adopts a multi-LoRA framework with step-level routing, where each expert is dedicated to a specific reasoning step. This design eliminates the need for balance losses, ensures stable training, and supports adaptive inference lengths. Experimental results demonstrate that MoTE significantly improves model safety, jailbreak resistance, and over-refusal capabilities, achieving performance comparable to OpenAI's state-of-the-art o1 model.", "authors": ["Zhili Liu", "Yunhao Gou", "Kai Chen", "Lanqing Hong", "Jiahui Gao", "Fei Mi", "Yu Zhang", "Zhenguo Li", "Xin Jiang", "Qun Liu", "James T. Kwok"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-01", "url": "https://arxiv.org/abs/2405.00557", "pdf_url": "https://arxiv.org/pdf/2405.00557v5", "arxiv_id": "2405.00557", "doi": "10.18653/v1/2025.acl-long.151", "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2865}
{"id": "9216f520c52c7fcecbe1460109dadd270449fcaba5ecfeb87f26f35affdbceeb", "sources": ["arxiv", "semantic_scholar"], "title": "Bi-Mamba+: Bidirectional Mamba for Time Series Forecasting", "abstract": "Long-term time series forecasting (LTSF) provides longer insights into future trends and patterns. Over the past few years, deep learning models especially Transformers have achieved advanced performance in LTSF tasks. However, LTSF faces inherent challenges such as long-term dependencies capturing and sparse semantic characteristics. Recently, a new state space model (SSM) named Mamba is proposed. With the selective capability on input data and the hardware-aware parallel computing algorithm, Mamba has shown great potential in balancing predicting performance and computational efficiency compared to Transformers. To enhance Mamba's ability to preserve historical information in a longer range, we design a novel Mamba+ block by adding a forget gate inside Mamba to selectively combine the new features with the historical features in a complementary manner. Furthermore, we apply Mamba+ both forward and backward and propose Bi-Mamba+, aiming to promote the model's ability to capture interactions among time series elements. Additionally, multivariate time series data in different scenarios may exhibit varying emphasis on intra- or inter-series dependencies. Therefore, we propose a series-relation-aware decider that controls the utilization of channel-independent or channel-mixing tokenization strategy for specific datasets. Extensive experiments on 8 real-world datasets show that our model achieves more accurate predictions compared with state-of-the-art methods.", "authors": ["Aobo Liang", "Xingguo Jiang", "Yan Sun", "Xiaohou Shi", "Ke Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-24", "url": "https://arxiv.org/abs/2404.15772", "pdf_url": "https://arxiv.org/pdf/2404.15772v3", "arxiv_id": "2404.15772", "doi": null, "citation_count": 29, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3693}
{"id": "d436f61df850c5fd192b04fe4bcd6ed2b3ec824eb5f92aae104681c4edbcd696", "sources": ["arxiv", "semantic_scholar"], "title": "SST: Multi-Scale Hybrid Mamba-Transformer Experts for Time Series Forecasting", "abstract": "Time series forecasting has made significant advances, including with Transformer-based models. The attention mechanism in Transformer effectively captures temporal dependencies by attending to all past inputs simultaneously. However, its quadratic complexity with respect to sequence length limits the scalability for long-range modeling. Recent state space models (SSMs) such as Mamba offer a promising alternative by achieving linear complexity without attention. Yet, Mamba compresses historical information into a fixed-size latent state, potentially causing information loss and limiting representational effectiveness. This raises a key research question: Can we design a hybrid Mamba-Transformer architecture that is both effective and efficient for time series forecasting? To address it, we adapt a hybrid Mamba-Transformer architecture Mambaformer, originally proposed for language modeling, to the time series domain. Preliminary experiments reveal that naively stacking Mamba and Transformer layers in Mambaformer is suboptimal for time series forecasting, due to an information interference problem. To mitigate this issue, we introduce a new time series decomposition strategy that separates time series into long-range patterns and short-range variations. Then we show that Mamba excels at capturing long-term structures, while Transformer is more effective at modeling short-term dynamics. Building on this insight, we propose State Space Transformer (SST), a multi-scale hybrid model with expert modules: a Mamba expert for long-range patterns and a Transformer expert for short-term variations. SST also employs a multi-scale patching mechanism to adaptively adjust time series resolution: low resolution for long-term patterns and high resolution for short-term variations. Experiments show that SST obtains SOTA performance with linear scalability. The code is at https://github.com/XiongxiaoXu/SST.", "authors": ["Xiongxiao Xu", "Canyu Chen", "Yueqing Liang", "Baixiang Huang", "Guangji Bai", "Liang Zhao", "Kai Shu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-23", "url": "https://arxiv.org/abs/2404.14757", "pdf_url": "https://arxiv.org/pdf/2404.14757v3", "arxiv_id": "2404.14757", "doi": "10.1145/3746252.3761394", "citation_count": 32, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/XiongxiaoXu/SST", "venue": "International Conference on Information and Knowledge Management", "quality_score": 0.3796}
{"id": "8eff6d2475432dc57ed606361cbf9a4c46c34c0434a86e9d3cf63f038407ffd2", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Head Mixture-of-Experts", "abstract": "Sparse Mixtures of Experts (SMoE) scales model capacity without significant increases in training and inference costs, but exhibits the following two issues: (1) Low expert activation, where only a small subset of experts are activated for optimization. (2) Lacking fine-grained analytical capabilities for multiple semantic concepts within individual tokens. We propose Multi-Head Mixture-of-Experts (MH-MoE), which employs a multi-head mechanism to split each token into multiple sub-tokens. These sub-tokens are then assigned to and processed by a diverse set of experts in parallel, and seamlessly reintegrated into the original token form. The multi-head mechanism enables the model to collectively attend to information from various representation spaces within different experts, while significantly enhances expert activation, thus deepens context understanding and alleviate overfitting. Moreover, our MH-MoE is straightforward to implement and decouples from other SMoE optimization methods, making it easy to integrate with other SMoE models for enhanced performance. Extensive experimental results across three tasks: English-focused language modeling, Multi-lingual language modeling and Masked multi-modality modeling tasks, demonstrate the effectiveness of MH-MoE.", "authors": ["Xun Wu", "Shaohan Huang", "Wenhui Wang", "Furu Wei"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-23", "url": "https://arxiv.org/abs/2404.15045", "pdf_url": "https://arxiv.org/pdf/2404.15045v1", "arxiv_id": "2404.15045", "doi": "10.48550/arXiv.2404.15045", "citation_count": 38, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3978}
{"id": "4252c8359138eb725c8d8a2d37afcd87faa1ecd6d45c3aae6dee275b40a81650", "sources": ["arxiv", "semantic_scholar"], "title": "MixLoRA: Enhancing Large Language Models Fine-Tuning with LoRA-based Mixture of Experts", "abstract": "Fine-tuning Large Language Models (LLMs) is a common practice to adapt pre-trained models for specific applications. While methods like LoRA have effectively addressed GPU memory constraints during fine-tuning, their performance often falls short, especially in multi-task scenarios. In contrast, Mixture-of-Expert (MoE) models, such as Mixtral 8x7B, demonstrate remarkable performance in multi-task learning scenarios while maintaining a reduced parameter count. However, the resource requirements of these MoEs remain challenging, particularly for consumer-grade GPUs with less than 24GB memory. To tackle these challenges, we propose MixLoRA, an approach to construct a resource-efficient sparse MoE model based on LoRA. MixLoRA inserts multiple LoRA-based experts within the feed-forward network block of a frozen pre-trained dense model and employs a commonly used top-k router. Unlike other LoRA-based MoE methods, MixLoRA enhances model performance by utilizing independent attention-layer LoRA adapters. Additionally, an auxiliary load balance loss is employed to address the imbalance problem of the router. Our evaluations show that MixLoRA improves about 9% accuracy compared to state-of-the-art PEFT methods in multi-task learning scenarios. We also propose a new high-throughput framework to alleviate the computation and memory bottlenecks during the training and inference of MOE models. This framework reduces GPU memory consumption by 40% and token computation latency by 30% during both training and inference.", "authors": ["Dengchun Li", "Yingzi Ma", "Naizheng Wang", "Zhengmao Ye", "Zhiyuan Cheng", "Yinghao Tang", "Yan Zhang", "Lei Duan", "Jie Zuo", "Cal Yang", "Mingjie Tang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-22", "url": "https://arxiv.org/abs/2404.15159", "pdf_url": "https://arxiv.org/pdf/2404.15159v3", "arxiv_id": "2404.15159", "doi": "10.48550/arXiv.2404.15159", "citation_count": 147, "influential_citation_count": 27, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7236}
{"id": "70aa412994c31282df2ceae126c9d71e694a7b9933e6a5dc06334099f6ad605e", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of LoRA Experts", "abstract": "LoRA has gained widespread acceptance in the fine-tuning of large pre-trained models to cater to a diverse array of downstream tasks, showcasing notable effectiveness and efficiency, thereby solidifying its position as one of the most prevalent fine-tuning techniques. Due to the modular nature of LoRA's plug-and-play plugins, researchers have delved into the amalgamation of multiple LoRAs to empower models to excel across various downstream tasks. Nonetheless, extant approaches for LoRA fusion grapple with inherent challenges. Direct arithmetic merging may result in the loss of the original pre-trained model's generative capabilities or the distinct identity of LoRAs, thereby yielding suboptimal outcomes. On the other hand, Reference tuning-based fusion exhibits limitations concerning the requisite flexibility for the effective combination of multiple LoRAs. In response to these challenges, this paper introduces the Mixture of LoRA Experts (MoLE) approach, which harnesses hierarchical control and unfettered branch selection. The MoLE approach not only achieves superior LoRA fusion performance in comparison to direct arithmetic merging but also retains the crucial flexibility for combining LoRAs effectively. Extensive experimental evaluations conducted in both the Natural Language Processing (NLP) and Vision & Language (V&L) domains substantiate the efficacy of MoLE.", "authors": ["Xun Wu", "Shaohan Huang", "Furu Wei"], "categories": ["cs.CL", "cs.LG", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-21", "url": "https://arxiv.org/abs/2404.13628", "pdf_url": "https://arxiv.org/pdf/2404.13628v1", "arxiv_id": "2404.13628", "doi": "10.48550/arXiv.2404.13628", "citation_count": 194, "influential_citation_count": 19, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.6505}
{"id": "c431a04a34cd37f7db52238a2a735239b101b9f4b8495ccf9af94e80b6f0b817", "sources": ["arxiv", "semantic_scholar"], "title": "decoupleQ: Towards 2-bit Post-Training Uniform Quantization via decoupling Parameters into Integer and Floating Points", "abstract": "Quantization emerges as one of the most promising compression technologies for deploying efficient large models for various real time application in recent years. Considering that the storage and IO of weights take up the vast majority of the overhead inside a large model, weight only quantization can lead to large gains. However, existing quantization schemes suffer from significant accuracy degradation at very low bits, or require some additional computational overhead when deployed, making it difficult to be applied to large-scale applications in industry. In this paper, we propose decoupleQ, achieving a substantial increase in model accuracy, especially at very low bits. decoupleQ abandons the traditional heuristic quantization paradigm and decouples the model parameters into integer and floating-point parts, thus transforming the quantization problem into a traditional mathematical optimization problem with constraints, which is then solved alternatively by off-the-shelf optimization methods. Quantization via decoupleQ is linear and uniform, making it hardware-friendlier than non-uniform counterpart, and enabling the idea to be migrated to high-bit quantization to enhance its robustness. Our method has achieved well on-line accuracy near fp16/bf16 on the 2-bit quantization of large speech models in ByteDance. The code is available at https://github.com/bytedance/decoupleQ", "authors": ["Yi Guo", "Fanliu Kong", "Xiaoyang Li", "Hui Li", "Wei Chen", "Xiaogang Tian", "Jinping Cai", "Yang Zhang", "Shouda Liu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-19", "url": "https://arxiv.org/abs/2404.12759", "pdf_url": "https://arxiv.org/pdf/2404.12759v1", "arxiv_id": "2404.12759", "doi": "10.48550/arXiv.2404.12759", "citation_count": 8, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/bytedance/decoupleQ", "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "3f3695b09ce67a4a159fc750477756789e0f1b2a5761995df6391f4d61b66c99", "sources": ["arxiv", "semantic_scholar"], "title": "MoVA: Adapting Mixture of Vision Experts to Multimodal Context", "abstract": "As the key component in multimodal large language models (MLLMs), the ability of the visual encoder greatly affects MLLM's understanding on diverse image content. Although some large-scale pretrained vision encoders such as vision encoders in CLIP and DINOv2 have brought promising performance, we found that there is still no single vision encoder that can dominate various image content understanding, e.g., the CLIP vision encoder leads to outstanding results on general image understanding but poor performance on document or chart content. To alleviate the bias of CLIP vision encoder, we first delve into the inherent behavior of different pre-trained vision encoders and then propose the MoVA, a powerful and novel MLLM, adaptively routing and fusing task-specific vision experts with a coarse-to-fine mechanism. In the coarse-grained stage, we design a context-aware expert routing strategy to dynamically select the most suitable vision experts according to the user instruction, input image, and expertise of vision experts. This benefits from the powerful model function understanding ability of the large language model (LLM). In the fine-grained stage, we elaborately conduct the mixture-of-vision-expert adapter (MoV-Adapter) to extract and fuse task-specific knowledge from various experts. This coarse-to-fine paradigm effectively leverages representations from experts based on multimodal context and model expertise, further enhancing the generalization ability. We conduct extensive experiments to evaluate the effectiveness of the proposed approach. Without any bells and whistles, MoVA can achieve significant performance gains over current state-of-the-art methods in a wide range of challenging multimodal benchmarks.", "authors": ["Zhuofan Zong", "Bingqi Ma", "Dazhong Shen", "Guanglu Song", "Hao Shao", "Dongzhi Jiang", "Hongsheng Li", "Yu Liu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-19", "url": "https://arxiv.org/abs/2404.13046", "pdf_url": "https://arxiv.org/pdf/2404.13046v2", "arxiv_id": "2404.13046", "doi": "10.48550/arXiv.2404.13046", "citation_count": 101, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5022}
{"id": "3fa15e570353563f0bcfe34e8f864a35b88ccc0dc2ce2131bca6ba8f559cc26a", "sources": ["arxiv", "semantic_scholar"], "title": "SNP: Structured Neuron-level Pruning to Preserve Attention Scores", "abstract": "Multi-head self-attention (MSA) is a key component of Vision Transformers (ViTs), which have achieved great success in various vision tasks. However, their high computational cost and memory footprint hinder their deployment on resource-constrained devices. Conventional pruning approaches can only compress and accelerate the MSA module using head pruning, although the head is not an atomic unit. To address this issue, we propose a novel graph-aware neuron-level pruning method, Structured Neuron-level Pruning (SNP). SNP prunes neurons with less informative attention scores and eliminates redundancy among heads. Specifically, it prunes graphically connected query and key layers having the least informative attention scores while preserving the overall attention scores. Value layers, which can be pruned independently, are pruned to eliminate inter-head redundancy. Our proposed method effectively compresses and accelerates Transformer-based models for both edge devices and server processors. For instance, the DeiT-Small with SNP runs 3.1$\\times$ faster than the original model and achieves performance that is 21.94\\% faster and 1.12\\% higher than the DeiT-Tiny. Additionally, SNP combine successfully with conventional head or block pruning approaches. SNP with head pruning could compress the DeiT-Base by 80\\% of the parameters and computational costs and achieve 3.85$\\times$ faster inference speed on RTX3090 and 4.93$\\times$ on Jetson Nano.", "authors": ["Kyunghwan Shim", "Jaewoong Yun", "Shinkook Choi"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-18", "url": "https://arxiv.org/abs/2404.11630", "pdf_url": "https://arxiv.org/pdf/2404.11630v1", "arxiv_id": "2404.11630", "doi": "10.1007/978-3-031-73007-8_6", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "European Conference on Computer Vision", "quality_score": 0.2258}
{"id": "0aa3ecb2c47e6690721b62e65a3ce1af58a1f29aa66cf59c1da48adc807f8492", "sources": ["arxiv", "semantic_scholar"], "title": "MoA: Mixture-of-Attention for Subject-Context Disentanglement in Personalized Image Generation", "abstract": "We introduce a new architecture for personalization of text-to-image diffusion models, coined Mixture-of-Attention (MoA). Inspired by the Mixture-of-Experts mechanism utilized in large language models (LLMs), MoA distributes the generation workload between two attention pathways: a personalized branch and a non-personalized prior branch. MoA is designed to retain the original model's prior by fixing its attention layers in the prior branch, while minimally intervening in the generation process with the personalized branch that learns to embed subjects in the layout and context generated by the prior branch. A novel routing mechanism manages the distribution of pixels in each layer across these branches to optimize the blend of personalized and generic content creation. Once trained, MoA facilitates the creation of high-quality, personalized images featuring multiple subjects with compositions and interactions as diverse as those generated by the original model. Crucially, MoA enhances the distinction between the model's pre-existing capability and the newly augmented personalized intervention, thereby offering a more disentangled subject-context control that was previously unattainable. Project page: https://snap-research.github.io/mixture-of-attention", "authors": ["Kuan-Chieh Wang", "Daniil Ostashev", "Yuwei Fang", "Sergey Tulyakov", "Kfir Aberman"], "categories": ["cs.CV", "cs.AI", "cs.GR"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-17", "url": "https://arxiv.org/abs/2404.11565", "pdf_url": "https://arxiv.org/pdf/2404.11565v2", "arxiv_id": "2404.11565", "doi": "10.1145/3680528.3687662", "citation_count": 53, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia", "quality_score": 0.4331}
{"id": "2808705db603c03ec54fd23fc6da4b3b507dfa9e4a364f4bd02e514c76b21b1c", "sources": ["arxiv", "semantic_scholar"], "title": "T-REX: Mixture-of-Rank-One-Experts with Semantic-aware Intuition for Multi-task Large Language Model Finetuning", "abstract": "Large language models (LLMs) encounter significant adaptation challenges in diverse multitask finetuning. Mixture-of-experts (MoE) provides a promising solution with a dynamic architecture, enabling effective task decoupling. However, scaling up the number of MoE experts incurs substantial parameter and computational overheads and suffers from limited performance gain due to naive routing mechanisms. In this paper, we design a novel framework, mix\\underline{\\textbf{T}}ure\\underline{\\textbf{-}}of-\\underline{\\textbf{R}}ank-on\\underline{\\textbf{E}}-e\\underline{\\textbf{X}}perts (\\texttt{T-REX}), which leverages the combination of ultra-low rank experts to construct LoRA weights on pretrained LLMs. The rank-1 experts enable a mix-and-match mechanism to quadratically expand the vector subspace of experts with linear parameter overheads, achieving approximate error reduction with optimal efficiency. In addition, T-REX offers implicit guidance to the router, leveraging the inherent semantic clustering of training embeddings as prior knowledge, enabling optimized feature allocation across experts for a smoother convergence. Extensive theoretical and empirical results demonstrate that T-REX achieves superior efficiency and generalizability across diverse tasks. Compared with other LoRA-based methods, T-REX achieves up to 1.78\\% mean accuracy improvement with around 30\\%-40\\% less trainable parameters across 14 public datasets. \\href{https://github.com/RoyZry98/T-REX-Pytorch}{Code} is available.", "authors": ["Rongyu Zhang", "Yijiang Liu", "Huanrui Yang", "Shenli Zheng", "Dan Wang", "Yuan Du", "Li Du", "Shanghang Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-13", "url": "https://arxiv.org/abs/2404.08985", "pdf_url": "https://arxiv.org/pdf/2404.08985v2", "arxiv_id": "2404.08985", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/RoyZry98/T-REX-Pytorch}{Code}", "venue": null, "quality_score": 0.1945}
{"id": "7135596a2c7c369ef72fb065ff5f14dc71047786987438fb4743066218f1aaa5", "sources": ["arxiv", "semantic_scholar"], "title": "MoPE: Mixture of Prefix Experts for Zero-Shot Dialogue State Tracking", "abstract": "Zero-shot dialogue state tracking (DST) transfers knowledge to unseen domains, reducing the cost of annotating new datasets. Previous zero-shot DST models mainly suffer from domain transferring and partial prediction problems. To address these challenges, we propose Mixture of Prefix Experts (MoPE) to establish connections between similar slots in different domains, which strengthens the model transfer performance in unseen domains. Empirical results demonstrate that MoPE-DST achieves the joint goal accuracy of 57.13% on MultiWOZ2.1 and 55.40% on SGD.", "authors": ["Tianwen Tang", "Tong Zhu", "Haodong Liu", "Yin Bai", "Jia Cheng", "Wenliang Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-12", "url": "https://arxiv.org/abs/2404.08559", "pdf_url": "https://arxiv.org/pdf/2404.08559v1", "arxiv_id": "2404.08559", "doi": "10.48550/arXiv.2404.08559", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Language Resources and Evaluation", "quality_score": 0.0753}
{"id": "c2836682c1cea45bb1d3fa93d741ccfe22a30e0855eeef674fafddfb0cf040bd", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-FFD: Mixture of Experts for Generalized and Parameter-Efficient Face Forgery Detection", "abstract": "Deepfakes have recently raised significant trust issues and security concerns among the public. Compared to CNN face forgery detectors, ViT-based methods take advantage of the expressivity of transformers, achieving superior detection performance. However, these approaches still exhibit the following limitations: (1) Fully fine-tuning ViT-based models from ImageNet weights demands substantial computational and storage resources; (2) ViT-based methods struggle to capture local forgery clues, leading to model bias; (3) These methods limit their scope on only one or few face forgery features, resulting in limited generalizability. To tackle these challenges, this work introduces Mixture-of-Experts modules for Face Forgery Detection (MoE-FFD), a generalized yet parameter-efficient ViT-based approach. MoE-FFD only updates lightweight Low-Rank Adaptation (LoRA) and Adapter layers while keeping the ViT backbone frozen, thereby achieving parameter-efficient training. Moreover, MoE-FFD leverages the expressivity of transformers and local priors of CNNs to simultaneously extract global and local forgery clues. Additionally, novel MoE modules are designed to scale the model's capacity and smartly select optimal forgery experts, further enhancing forgery detection performance. Our proposed learning scheme can be seamlessly adapted to various transformer backbones in a plug-and-play manner. Extensive experimental results demonstrate that the proposed method achieves state-of-the-art face forgery detection performance with significantly reduced parameter overhead. The code is released at: https://github.com/LoveSiameseCat/MoE-FFD.", "authors": ["Chenqi Kong", "Anwei Luo", "Peijun Bao", "Yi Yu", "Haoliang Li", "Zengwei Zheng", "Shiqi Wang", "Alex C. Kot"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-12", "url": "https://arxiv.org/abs/2404.08452", "pdf_url": "https://arxiv.org/pdf/2404.08452v3", "arxiv_id": "2404.08452", "doi": "10.1109/TDSC.2025.3604443", "citation_count": 51, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/LoveSiameseCat/MoE-FFD", "venue": "IEEE Transactions on Dependable and Secure Computing", "quality_score": 0.429}
{"id": "06a9f27bcd1334676d3e017b731c4497b22e1836f9bd6da0108fc22c6d632ced", "sources": ["arxiv", "semantic_scholar"], "title": "SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget", "abstract": "Optimizing the Key-Value (KV) cache of the Large Language Model (LLM) has been considered critical to saving the cost of inference. Most of the existing KV-cache compression algorithms attempted to sparsify the sequence of tokens by taking advantage of the different importance of tokens. However, most of these methods treat all layers equally, allocating the same KV budget to each layer. This approach is suboptimal, as some layers may be less sensitive to input tokens yet still receive the same budget as others. In this work, we found that by identifying the importance of attention layers, we could optimize the KV-cache jointly from two dimensions, i.e., sequence-wise and layer-wise. Based on our observations regarding layer-wise importance in inference, we propose SqueezeAttention to precisely optimize the allocation of KV-cache budget among layers on-the-fly and then incorporate three representative sequence-wise algorithms to compress the KV-cache for each layer with its very own budget. Specifically, we first measure each layer's importance by calculating the cosine similarity of the input prompt differences before and after the self-attention layers. Based on this similarity, we then categorize the layers into two groups and adjust their KV budgets accordingly. By optimizing the KV-cache from both sequence's and layer's dimensions, SqueezeAttention achieves around 30% to 70% of the memory reductions and up to 2.2 times of throughput improvements in a wide range of LLMs and benchmarks. The code is available at https://github.com/hetailang/SqueezeAttention.", "authors": ["Zihao Wang", "Bin Cui", "Shaoduo Gan"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-07", "url": "https://arxiv.org/abs/2404.04793", "pdf_url": "https://arxiv.org/pdf/2404.04793v2", "arxiv_id": "2404.04793", "doi": "10.48550/arXiv.2404.04793", "citation_count": 18, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/hetailang/SqueezeAttention", "venue": "International Conference on Learning Representations", "quality_score": 0.3197}
{"id": "07ed2a685b37ac9d04adb203ec78277150b1e46b57c91e0f5cc585c8fac02004", "sources": ["arxiv", "semantic_scholar"], "title": "SEER-MoE: Sparse Expert Efficiency through Regularization for Mixture-of-Experts", "abstract": "The advancement of deep learning has led to the emergence of Mixture-of-Experts (MoEs) models, known for their dynamic allocation of computational resources based on input. Despite their promise, MoEs face challenges, particularly in terms of memory requirements. To address this, our work introduces SEER-MoE, a novel two-stage framework for reducing both the memory footprint and compute requirements of pre-trained MoE models. The first stage involves pruning the total number of experts using a heavy-hitters counting guidance, while the second stage employs a regularization-based fine-tuning strategy to recover accuracy loss and reduce the number of activated experts during inference. Our empirical studies demonstrate the effectiveness of our method, resulting in a sparse MoEs model optimized for inference efficiency with minimal accuracy trade-offs.", "authors": ["Alexandre Muzio", "Alex Sun", "Churan He"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-07", "url": "https://arxiv.org/abs/2404.05089", "pdf_url": "https://arxiv.org/pdf/2404.05089v1", "arxiv_id": "2404.05089", "doi": "10.48550/arXiv.2404.05089", "citation_count": 36, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3921}
{"id": "97a59d64e1b76ff39d4d94bf5aa3440eddfbfaa0fe85c5e3e17bc7690507c7cf", "sources": ["arxiv", "semantic_scholar"], "title": "A Note on LoRA", "abstract": "LoRA (Low-Rank Adaptation) has emerged as a preferred method for efficiently adapting Large Language Models (LLMs) with remarkable simplicity and efficacy. This note extends the original LoRA paper by offering new perspectives that were not initially discussed and presents a series of insights for deploying LoRA at scale. Without introducing new experiments, we aim to improve the understanding and application of LoRA.", "authors": ["Vlad Fomenko", "Han Yu", "Jongho Lee", "Stanley Hsieh", "Weizhu Chen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-07", "url": "https://arxiv.org/abs/2404.05086", "pdf_url": "https://arxiv.org/pdf/2404.05086v1", "arxiv_id": "2404.05086", "doi": "10.48550/arXiv.2404.05086", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3076}
{"id": "95c36a384d682624fbebd458febdaf1583ac5e8abf266159e736218d882b22af", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Low-rank Experts for Transferable AI-Generated Image Detection", "abstract": "Generative models have shown a giant leap in synthesizing photo-realistic images with minimal expertise, sparking concerns about the authenticity of online information. This study aims to develop a universal AI-generated image detector capable of identifying images from diverse sources. Existing methods struggle to generalize across unseen generative models when provided with limited sample sources. Inspired by the zero-shot transferability of pre-trained vision-language models, we seek to harness the nontrivial visual-world knowledge and descriptive proficiency of CLIP-ViT to generalize over unknown domains. This paper presents a novel parameter-efficient fine-tuning approach, mixture of low-rank experts, to fully exploit CLIP-ViT's potential while preserving knowledge and expanding capacity for transferable detection. We adapt only the MLP layers of deeper ViT blocks via an integration of shared and separate LoRAs within an MoE-based structure. Extensive experiments on public benchmarks show that our method achieves superiority over state-of-the-art approaches in cross-generator generalization and robustness to perturbations. Remarkably, our best-performing ViT-L/14 variant requires training only 0.08% of its parameters to surpass the leading baseline by +3.64% mAP and +12.72% avg.Acc across unseen diffusion and autoregressive models. This even outperforms the baseline with just 0.28% of the training data. Our code and pre-trained models will be available at https://github.com/zhliuworks/CLIPMoLE.", "authors": ["Zihan Liu", "Hanyi Wang", "Yaoyu Kang", "Shilin Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-07", "url": "https://arxiv.org/abs/2404.04883", "pdf_url": "https://arxiv.org/pdf/2404.04883v1", "arxiv_id": "2404.04883", "doi": "10.48550/arXiv.2404.04883", "citation_count": 30, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/zhliuworks/CLIPMoLE", "venue": "arXiv.org", "quality_score": 0.3728}
{"id": "eacb7a23e570117e6950fea83ec82d6833c8b9b4166447442c33aa62cd69e3aa", "sources": ["arxiv", "semantic_scholar"], "title": "Fusion of Mixture of Experts and Generative Artificial Intelligence in Mobile Edge Metaverse", "abstract": "In the digital transformation era, Metaverse offers a fusion of virtual reality (VR), augmented reality (AR), and web technologies to create immersive digital experiences. However, the evolution of the Metaverse is slowed down by the challenges of content creation, scalability, and dynamic user interaction. Our study investigates an integration of Mixture of Experts (MoE) models with Generative Artificial Intelligence (GAI) for mobile edge computing to revolutionize content creation and interaction in the Metaverse. Specifically, we harness an MoE model's ability to efficiently manage complex data and complex tasks by dynamically selecting the most relevant experts running various sub-models to enhance the capabilities of GAI. We then present a novel framework that improves video content generation quality and consistency, and demonstrate its application through case studies. Our findings underscore the efficacy of MoE and GAI integration to redefine virtual experiences by offering a scalable, efficient pathway to harvest the Metaverse's full potential.", "authors": ["Guangyuan Liu", "Hongyang Du", "Dusit Niyato", "Jiawen Kang", "Zehui Xiong", "Abbas Jamalipour", "Shiwen Mao", "Dong In Kim"], "categories": ["cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-04", "url": "https://arxiv.org/abs/2404.03321", "pdf_url": "https://arxiv.org/pdf/2404.03321v1", "arxiv_id": "2404.03321", "doi": "10.48550/arXiv.2404.03321", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Network", "quality_score": 0.2603}
{"id": "1756a8bd38b2706db999761bb378272c66261dd3d9563211744d08407284fa98", "sources": ["arxiv", "semantic_scholar"], "title": "CATP: Cross-Attention Token Pruning for Accuracy Preserved Multimodal Model Inference", "abstract": "In response to the rising interest in large multimodal models, we introduce Cross-Attention Token Pruning (CATP), a precision-focused token pruning method. Our approach leverages cross-attention layers in multimodal models, exemplified by BLIP-2, to extract valuable information for token importance determination. CATP employs a refined voting strategy across model heads and layers. In evaluations, CATP achieves up to 12.1X higher accuracy compared to existing token pruning methods, addressing the trade-off between computational efficiency and model precision.", "authors": ["Ruqi Liao", "Chuqing Zhao", "Jin Li", "Weiqi Feng", "Yi Lyu", "Bingxian Chen", "Haochen Yang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-02", "url": "https://arxiv.org/abs/2404.08567", "pdf_url": "https://arxiv.org/pdf/2404.08567v2", "arxiv_id": "2404.08567", "doi": "10.1109/CAI64502.2025.00191", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Algebraic Informatics", "quality_score": 0.2603}
{"id": "b0e014a46b5ee6e40ba56180a410ddb821c9b41f5eeb39921eb6ecb4583342b8", "sources": ["arxiv", "semantic_scholar"], "title": "DeFT: Decoding with Flash Tree-attention for Efficient Tree-structured LLM Inference", "abstract": "Large language models (LLMs) are increasingly employed for complex tasks that process multiple generation calls in a tree structure with shared prefixes of tokens, including few-shot prompting, multi-step reasoning, speculative decoding, etc. However, existing inference systems for tree-based applications are inefficient due to improper partitioning of queries and KV cache during attention calculation. This leads to two main issues: (1) a lack of memory access (IO) reuse for KV cache of shared prefixes, and (2) poor load balancing.As a result, there is redundant KV cache IO between GPU global memory and shared memory, along with low GPU utilization. To address these challenges, we propose DeFT(Decoding with Flash Tree-Attention), a hardware-efficient attention algorithm with prefix-aware and load-balanced KV cache partitions. DeFT reduces the number of read/write operations of KV cache during attention calculation through KV-Guided Grouping, a method that avoids repeatedly loading KV cache of shared prefixes in attention computation. Additionally, we propose Flattened Tree KV Splitting, a mechanism that ensures even distribution of the KV cache across partitions with little computation redundancy, enhancing GPU utilization during attention computations. By reducing 73-99% KV cache IO and nearly 100% IO for partial results during attention calculation, DeFT achieves up to 2.23/3.59x speedup in the end-to-end/attention latency across three practical tree-based workloads compared to state-of-the-art attention algorithms. Our code is available at https://github.com/LINs-lab/DeFT.", "authors": ["Jinwei Yao", "Kaiqi Chen", "Kexun Zhang", "Jiaxuan You", "Binhang Yuan", "Zeke Wang", "Tao Lin"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-30", "url": "https://arxiv.org/abs/2404.00242", "pdf_url": "https://arxiv.org/pdf/2404.00242v4", "arxiv_id": "2404.00242", "doi": null, "citation_count": 14, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/LINs-lab/DeFT", "venue": "International Conference on Learning Representations", "quality_score": 0.294}
{"id": "fdaeb81c02a2cc7d6ae7678c3f881afaf38a55754ec7171dbc5f84e426307edb", "sources": ["arxiv", "semantic_scholar"], "title": "Jamba: A Hybrid Transformer-Mamba Language Model", "abstract": "We present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of both model families. MoE is added in some of these layers to increase model capacity while keeping active parameter usage manageable. This flexible architecture allows resource- and objective-specific configurations. In the particular configuration we have implemented, we end up with a powerful model that fits in a single 80GB GPU. Built at large scale, Jamba provides high throughput and small memory footprint compared to vanilla Transformers, and at the same time state-of-the-art performance on standard language model benchmarks and long-context evaluations. Remarkably, the model presents strong results for up to 256K tokens context length. We study various architectural decisions, such as how to combine Transformer and Mamba layers, and how to mix experts, and show that some of them are crucial in large scale modeling. We also describe several interesting properties of these architectures which the training and evaluation of Jamba have revealed, and plan to release checkpoints from various ablation runs, to encourage further exploration of this novel architecture. We make the weights of our implementation of Jamba publicly available under a permissive license.", "authors": ["Opher Lieber", "Barak Lenz", "Hofit Bata", "Gal Cohen", "Jhonathan Osin", "Itay Dalmedigos", "Erez Safahi", "Shaked Meirom", "Yonatan Belinkov", "Shai Shalev-Shwartz", "Omri Abend", "Raz Alon", "Tomer Asida", "Amir Bergman", "Roman Glozman", "Michael Gokhman", "Avashalom Manevich", "Nir Ratner", "Noam Rozen", "Erez Shwartz", "Mor Zusman", "Yoav Shoham"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-28", "url": "https://arxiv.org/abs/2403.19887", "pdf_url": "https://arxiv.org/pdf/2403.19887v2", "arxiv_id": "2403.19887", "doi": "10.48550/arXiv.2403.19887", "citation_count": 439, "influential_citation_count": 33, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7657}
{"id": "47029a25cc76aa7e699ef6bef1404e56da54b6bf0a1c3ccd7bb26dc92561f997", "sources": ["arxiv", "semantic_scholar"], "title": "Contrastive Test-Time Composition of Multiple LoRA Models for Image Generation", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a powerful and popular technique for personalization, enabling efficient adaptation of pre-trained image generation models for specific tasks without comprehensive retraining. While employing individual pre-trained LoRA models excels at representing single concepts, such as those representing a specific dog or a cat, utilizing multiple LoRA models to capture a variety of concepts in a single image still poses a significant challenge. Existing methods often fall short, primarily because the attention mechanisms within different LoRA models overlap, leading to scenarios where one concept may be completely ignored (e.g., omitting the dog) or where concepts are incorrectly combined (e.g., producing an image of two cats instead of one cat and one dog). We introduce CLoRA, a training-free approach that addresses these limitations by updating the attention maps of multiple LoRA models at test-time, and leveraging the attention maps to create semantic masks for fusing latent representations. This enables the generation of composite images that accurately reflect the characteristics of each LoRA. Our comprehensive qualitative and quantitative evaluations demonstrate that CLoRA significantly outperforms existing methods in multi-concept image generation using LoRAs.", "authors": ["Tuna Han Salih Meral", "Enis Simsar", "Federico Tombari", "Pinar Yanardag"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-28", "url": "https://arxiv.org/abs/2403.19776", "pdf_url": "https://arxiv.org/pdf/2403.19776v2", "arxiv_id": "2403.19776", "doi": "10.1109/ICCV51701.2025.01681", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.2386}
{"id": "1cc9a5820d12ffec95b0aec463aab354bfea4be27ca79560aae3a1aa8efa02dd", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Task Dense Prediction via Mixture of Low-Rank Experts", "abstract": "Previous multi-task dense prediction methods based on the Mixture of Experts (MoE) have received great performance but they neglect the importance of explicitly modeling the global relations among all tasks. In this paper, we present a novel decoder-focused method for multi-task dense prediction, called Mixture-of-Low-Rank-Experts (MLoRE). To model the global task relationships, MLoRE adds a generic convolution path to the original MoE structure, where each task feature can go through this path for explicit parameter sharing. Furthermore, to control the parameters and computational cost brought by the increase in the number of experts, we take inspiration from LoRA and propose to leverage the low-rank format of a vanilla convolution in the expert network. Since the low-rank experts have fewer parameters and can be dynamically parameterized into the generic convolution, the parameters and computational cost do not change much with the increase of experts. Benefiting from this design, we increase the number of experts and its reception field to enlarge the representation capacity, facilitating multiple dense tasks learning in a unified network. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks show that our MLoRE achieves superior performance compared to previous state-of-the-art methods on all metrics. Our code is available at https://github.com/YuqiYang213/MLoRE.", "authors": ["Yuqi Yang", "Peng-Tao Jiang", "Qibin Hou", "Hao Zhang", "Jinwei Chen", "Bo Li"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-26", "url": "https://arxiv.org/abs/2403.17749", "pdf_url": "https://arxiv.org/pdf/2403.17749v2", "arxiv_id": "2403.17749", "doi": "10.1109/CVPR52733.2024.02638", "citation_count": 77, "influential_citation_count": 11, "has_code": true, "code_url": "https://github.com/YuqiYang213/MLoRE", "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.5396}
{"id": "43d2d2b5f11f690cc2ff63b3a96ffb66350c96a5700ed4c5ad427ef907d9f529", "sources": ["arxiv", "semantic_scholar"], "title": "ALISA: Accelerating Large Language Model Inference via Sparsity-Aware KV Caching", "abstract": "The Transformer architecture has significantly advanced natural language processing (NLP) and has been foundational in developing large language models (LLMs) such as LLaMA and OPT, which have come to dominate a broad range of NLP tasks. Despite their superior accuracy, LLMs present unique challenges in practical inference, concerning the compute and memory-intensive nature. Thanks to the autoregressive characteristic of LLM inference, KV caching for the attention layers in Transformers can effectively accelerate LLM inference by substituting quadratic-complexity computation with linear-complexity memory accesses. Yet, this approach requires increasing memory as demand grows for processing longer sequences. The overhead leads to reduced throughput due to I/O bottlenecks and even out-of-memory errors, particularly on resource-constrained systems like a single commodity GPU. In this paper, we propose ALISA, a novel algorithm-system co-design solution to address the challenges imposed by KV caching. On the algorithm level, ALISA prioritizes tokens that are most important in generating a new token via a Sparse Window Attention (SWA) algorithm. SWA introduces high sparsity in attention layers and reduces the memory footprint of KV caching at negligible accuracy loss. On the system level, ALISA employs three-phase token-level dynamical scheduling and optimizes the trade-off between caching and recomputation, thus maximizing the overall performance in resource-constrained systems. In a single GPU-CPU system, we demonstrate that under varying workloads, ALISA improves the throughput of baseline systems such as FlexGen and vLLM by up to 3X and 1.9X, respectively.", "authors": ["Youpeng Zhao", "Di Wu", "Jun Wang"], "categories": ["cs.AI", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-26", "url": "https://arxiv.org/abs/2403.17312", "pdf_url": "https://arxiv.org/pdf/2403.17312v1", "arxiv_id": "2403.17312", "doi": "10.1109/ISCA59077.2024.00077", "citation_count": 65, "influential_citation_count": 11, "has_code": false, "code_url": null, "venue": "International Symposium on Computer Architecture", "quality_score": 0.5396}
{"id": "e8281e850740ea260ad1494faff1a9c7d76749c1d20bb89eaa39d5d25df0153a", "sources": ["arxiv", "semantic_scholar"], "title": "Generalization Error Analysis for Sparse Mixture-of-Experts: A Preliminary Study", "abstract": "Mixture-of-Experts (MoE) represents an ensemble methodology that amalgamates predictions from several specialized sub-models (referred to as experts). This fusion is accomplished through a router mechanism, dynamically assigning weights to each expert's contribution based on the input data. Conventional MoE mechanisms select all available experts, incurring substantial computational costs. In contrast, Sparse Mixture-of-Experts (Sparse MoE) selectively engages only a limited number, or even just one expert, significantly reducing computation overhead while empirically preserving, and sometimes even enhancing, performance. Despite its wide-ranging applications and these advantageous characteristics, MoE's theoretical underpinnings have remained elusive. In this paper, we embark on an exploration of Sparse MoE's generalization error concerning various critical factors. Specifically, we investigate the impact of the number of data samples, the total number of experts, the sparsity in expert selection, the complexity of the routing mechanism, and the complexity of individual experts. Our analysis sheds light on \\textit{how \\textbf{sparsity} contributes to the MoE's generalization}, offering insights from the perspective of classical learning theory.", "authors": ["Jinze Zhao", "Peihao Wang", "Zhangyang Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-26", "url": "https://arxiv.org/abs/2403.17404", "pdf_url": "https://arxiv.org/pdf/2403.17404v1", "arxiv_id": "2403.17404", "doi": "10.48550/arXiv.2403.17404", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "146f126c734cb55492707c515e7ca93754117b6a51c9411d192f8e87dc4c7d4c", "sources": ["arxiv", "semantic_scholar"], "title": "Be Yourself: Bounded Attention for Multi-Subject Text-to-Image Generation", "abstract": "Text-to-image diffusion models have an unprecedented ability to generate diverse and high-quality images. However, they often struggle to faithfully capture the intended semantics of complex input prompts that include multiple subjects. Recently, numerous layout-to-image extensions have been introduced to improve user control, aiming to localize subjects represented by specific tokens. Yet, these methods often produce semantically inaccurate images, especially when dealing with multiple semantically or visually similar subjects. In this work, we study and analyze the causes of these limitations. Our exploration reveals that the primary issue stems from inadvertent semantic leakage between subjects in the denoising process. This leakage is attributed to the diffusion model's attention layers, which tend to blend the visual features of different subjects. To address these issues, we introduce Bounded Attention, a training-free method for bounding the information flow in the sampling process. Bounded Attention prevents detrimental leakage among subjects and enables guiding the generation to promote each subject's individuality, even with complex multi-subject conditioning. Through extensive experimentation, we demonstrate that our method empowers the generation of multiple subjects that better align with given prompts and layouts.", "authors": ["Omer Dahary", "Or Patashnik", "Kfir Aberman", "Daniel Cohen-Or"], "categories": ["cs.CV", "cs.AI", "cs.GR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-25", "url": "https://arxiv.org/abs/2403.16990", "pdf_url": "https://arxiv.org/pdf/2403.16990v1", "arxiv_id": "2403.16990", "doi": "10.48550/arXiv.2403.16990", "citation_count": 69, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "European Conference on Computer Vision", "quality_score": 0.4613}
{"id": "e8e8d2c12c4c66d27273267a90b931f938844eafe778258767b9de87c841a72b", "sources": ["arxiv", "semantic_scholar"], "title": "Implicit Style-Content Separation using B-LoRA", "abstract": "Image stylization involves manipulating the visual appearance and texture (style) of an image while preserving its underlying objects, structures, and concepts (content). The separation of style and content is essential for manipulating the image's style independently from its content, ensuring a harmonious and visually pleasing result. Achieving this separation requires a deep understanding of both the visual and semantic characteristics of images, often necessitating the training of specialized models or employing heavy optimization. In this paper, we introduce B-LoRA, a method that leverages LoRA (Low-Rank Adaptation) to implicitly separate the style and content components of a single image, facilitating various image stylization tasks. By analyzing the architecture of SDXL combined with LoRA, we find that jointly learning the LoRA weights of two specific blocks (referred to as B-LoRAs) achieves style-content separation that cannot be achieved by training each B-LoRA independently. Consolidating the training into only two blocks and separating style and content allows for significantly improving style manipulation and overcoming overfitting issues often associated with model fine-tuning. Once trained, the two B-LoRAs can be used as independent components to allow various image stylization tasks, including image style transfer, text-based image stylization, consistent style generation, and style-content mixing.", "authors": ["Yarden Frenkel", "Yael Vinker", "Ariel Shamir", "Daniel Cohen-Or"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-21", "url": "https://arxiv.org/abs/2403.14572", "pdf_url": "https://arxiv.org/pdf/2403.14572v2", "arxiv_id": "2403.14572", "doi": "10.48550/arXiv.2403.14572", "citation_count": 154, "influential_citation_count": 22, "has_code": false, "code_url": null, "venue": "European Conference on Computer Vision", "quality_score": 0.6809}
{"id": "32aa9b565ca69618d719b7cf84b3869063ec93c4e0d7ffbc6f26917356bccf8a", "sources": ["arxiv", "semantic_scholar"], "title": "OrthCaps: An Orthogonal CapsNet with Sparse Attention Routing and Pruning", "abstract": "Redundancy is a persistent challenge in Capsule Networks (CapsNet),leading to high computational costs and parameter counts. Although previous works have introduced pruning after the initial capsule layer, dynamic routing's fully connected nature and non-orthogonal weight matrices reintroduce redundancy in deeper layers. Besides, dynamic routing requires iterating to converge, further increasing computational demands. In this paper, we propose an Orthogonal Capsule Network (OrthCaps) to reduce redundancy, improve routing performance and decrease parameter counts. Firstly, an efficient pruned capsule layer is introduced to discard redundant capsules. Secondly, dynamic routing is replaced with orthogonal sparse attention routing, eliminating the need for iterations and fully connected structures. Lastly, weight matrices during routing are orthogonalized to sustain low capsule similarity, which is the first approach to introduce orthogonality into CapsNet as far as we know. Our experiments on baseline datasets affirm the efficiency and robustness of OrthCaps in classification tasks, in which ablation studies validate the criticality of each component. Remarkably, OrthCaps-Shallow outperforms other Capsule Network benchmarks on four datasets, utilizing only 110k parameters, which is a mere 1.25% of a standard Capsule Network's total. To the best of our knowledge, it achieves the smallest parameter count among existing Capsule Networks. Similarly, OrthCaps-Deep demonstrates competitive performance across four datasets, utilizing only 1.2% of the parameters required by its counterparts.", "authors": ["Xinyu Geng", "Jiaming Wang", "Jiawei Gong", "Yuerong Xue", "Jun Xu", "Fanglin Chen", "Xiaolin Huang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-20", "url": "https://arxiv.org/abs/2403.13351", "pdf_url": "https://arxiv.org/pdf/2403.13351v1", "arxiv_id": "2403.13351", "doi": "10.1109/CVPR52733.2024.00577", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.2785}
{"id": "146239c06b15b57a62a2a1da36233ec01ea3f7bcc35222cca79a23138684b34d", "sources": ["arxiv", "semantic_scholar"], "title": "Improving LoRA in Privacy-preserving Federated Learning", "abstract": "Low-rank adaptation (LoRA) is one of the most popular task-specific parameter-efficient fine-tuning (PEFT) methods on pre-trained language models for its good performance and computational efficiency. LoRA injects a product of two trainable rank decomposition matrices over the top of each frozen pre-trained model module. However, when applied in the setting of privacy-preserving federated learning (FL), LoRA may become unstable due to the following facts: 1) the effects of data heterogeneity and multi-step local updates are non-negligible, 2) additive noise enforced on updating gradients to guarantee differential privacy (DP) can be amplified and 3) the final performance is susceptible to hyper-parameters. A key factor leading to these phenomena is the discordance between jointly optimizing the two low-rank matrices by local clients and separately aggregating them by the central server. Thus, this paper proposes an efficient and effective version of LoRA, Federated Freeze A LoRA (FFA-LoRA), to alleviate these challenges and further halve the communication cost of federated fine-tuning LLMs. The core idea of FFA-LoRA is to fix the randomly initialized non-zero matrices and only fine-tune the zero-initialized matrices. Compared to LoRA, FFA-LoRA is motivated by practical and theoretical benefits in privacy-preserved FL. Our experiments demonstrate that FFA-LoRA provides more consistent performance with better computational efficiency over vanilla LoRA in various FL tasks.", "authors": ["Youbang Sun", "Zitao Li", "Yaliang Li", "Bolin Ding"], "categories": ["cs.LG", "cs.CR", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-18", "url": "https://arxiv.org/abs/2403.12313", "pdf_url": "https://arxiv.org/pdf/2403.12313v1", "arxiv_id": "2403.12313", "doi": "10.48550/arXiv.2403.12313", "citation_count": 191, "influential_citation_count": 30, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.7457}
{"id": "162974f92df1d6217f89cc5e250b83b0eebc4053f25e38c355f75f6de7d62711", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Prompt-Experts for Multi-modal Semantic Understanding", "abstract": "Deep multimodal semantic understanding that goes beyond the mere superficial content relation mining has received increasing attention in the realm of artificial intelligence. The challenges of collecting and annotating high-quality multi-modal data have underscored the significance of few-shot learning. In this paper, we focus on two critical tasks under this context: few-shot multi-modal sarcasm detection (MSD) and multi-modal sentiment analysis (MSA). To address them, we propose Mixture-of-Prompt-Experts with Block-Aware Prompt Fusion (MoPE-BAF), a novel multi-modal soft prompt framework based on the unified vision-language model (VLM). Specifically, we design three experts of soft prompts: a text prompt and an image prompt that extract modality-specific features to enrich the single-modal representation, and a unified prompt to assist multi-modal interaction. Additionally, we reorganize Transformer layers into several blocks and introduce cross-modal prompt attention between adjacent blocks, which smoothens the transition from single-modal representation to multi-modal fusion. On both MSD and MSA datasets in few-shot setting, our proposed model not only surpasses the 8.2B model InstructBLIP with merely 2% parameters (150M), but also significantly outperforms other widely-used prompt methods on VLMs or task-specific methods.", "authors": ["Zichen Wu", "Hsiu-Yuan Huang", "Fanyi Qu", "Yunfang Wu"], "categories": ["cs.CL", "cs.MM"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-17", "url": "https://arxiv.org/abs/2403.11311", "pdf_url": "https://arxiv.org/pdf/2403.11311v2", "arxiv_id": "2403.11311", "doi": "10.48550/arXiv.2403.11311", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Language Resources and Evaluation", "quality_score": 0.2603}
{"id": "9c4680904d6612966ef4d2d09eb62ceefe174faf61c4d7563643fb4b725f6b6e", "sources": ["arxiv", "semantic_scholar"], "title": "DEFA: Efficient Deformable Attention Acceleration via Pruning-Assisted Grid-Sampling and Multi-Scale Parallel Processing", "abstract": "Multi-scale deformable attention (MSDeformAttn) has emerged as a key mechanism in various vision tasks, demonstrating explicit superiority attributed to multi-scale grid-sampling. However, this newly introduced operator incurs irregular data access and enormous memory requirement, leading to severe PE underutilization. Meanwhile, existing approaches for attention acceleration cannot be directly applied to MSDeformAttn due to lack of support for this distinct procedure. Therefore, we propose a dedicated algorithm-architecture co-design dubbed DEFA, the first-of-its-kind method for MSDeformAttn acceleration. At the algorithm level, DEFA adopts frequency-weighted pruning and probability-aware pruning for feature maps and sampling points respectively, alleviating the memory footprint by over 80%. At the architecture level, it explores the multi-scale parallelism to boost the throughput significantly and further reduces the memory access via fine-grained layer fusion and feature map reusing. Extensively evaluated on representative benchmarks, DEFA achieves 10.1-31.9x speedup and 20.3-37.7x energy efficiency boost compared to powerful GPUs. It also rivals the related accelerators by 2.2-3.7x energy efficiency improvement while providing pioneering support for MSDeformAttn.", "authors": ["Yansong Xu", "Dongxu Lyu", "Zhenyu Li", "Zilong Wang", "Yuzhou Chen", "Gang Wang", "Zhican Wang", "Haomin Li", "Guanghui He"], "categories": ["cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-16", "url": "https://arxiv.org/abs/2403.10913", "pdf_url": "https://arxiv.org/pdf/2403.10913v1", "arxiv_id": "2403.10913", "doi": "10.1145/3649329.3657328", "citation_count": 9, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Design Automation Conference", "quality_score": 0.301}
{"id": "c0f4acc197aa5cc4c13e98eac620656222616c3a725b52d9de8a1a40d81980fe", "sources": ["arxiv", "semantic_scholar"], "title": "Keyformer: KV Cache Reduction through Key Tokens Selection for Efficient Generative Inference", "abstract": "Transformers have emerged as the underpinning architecture for Large Language Models (LLMs). In generative language models, the inference process involves two primary phases: prompt processing and token generation. Token generation, which constitutes the majority of the computational workload, primarily entails vector-matrix multiplications and interactions with the Key-Value (KV) Cache. This phase is constrained by memory bandwidth due to the overhead of transferring weights and KV cache values from the memory system to the computing units. This memory bottleneck becomes particularly pronounced in applications that require long-context and extensive text generation, both of which are increasingly crucial for LLMs. This paper introduces \"Keyformer\", an innovative inference-time approach, to mitigate the challenges associated with KV cache size and memory bandwidth utilization. Keyformer leverages the observation that approximately 90% of the attention weight in generative inference focuses on a specific subset of tokens, referred to as \"key\" tokens. Keyformer retains only the key tokens in the KV cache by identifying these crucial tokens using a novel score function. This approach effectively reduces both the KV cache size and memory bandwidth usage without compromising model accuracy. We evaluate Keyformer's performance across three foundational models: GPT-J, Cerebras-GPT, and MPT, which employ various positional embedding algorithms. Our assessment encompasses a variety of tasks, with a particular emphasis on summarization and conversation tasks involving extended contexts. Keyformer's reduction of KV cache reduces inference latency by 2.1x and improves token generation throughput by 2.4x, while preserving the model's accuracy.", "authors": ["Muhammad Adnan", "Akhil Arunkumar", "Gaurav Jain", "Prashant J. Nair", "Ilya Soloveychik", "Purushotham Kamath"], "categories": ["cs.LG", "cs.AI", "cs.AR", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-14", "url": "https://arxiv.org/abs/2403.09054", "pdf_url": "https://arxiv.org/pdf/2403.09054v2", "arxiv_id": "2403.09054", "doi": "10.48550/arXiv.2403.09054", "citation_count": 141, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Conference on Machine Learning and Systems", "quality_score": 0.5381}
{"id": "7f584cc9824081c231bb67f4d3a3666cad7a9ba232231fab8e14b7ac479d7b44", "sources": ["arxiv", "semantic_scholar"], "title": "SAM-Lightening: A Lightweight Segment Anything Model with Dilated Flash Attention to Achieve 30 times Acceleration", "abstract": "Segment Anything Model (SAM) has garnered significant attention in segmentation tasks due to their zero-shot generalization ability. However, a broader application of SAMs to real-world practice has been restricted by their low inference speed and high computational memory demands, which mainly stem from the attention mechanism. Existing work concentrated on optimizing the encoder, yet has not adequately addressed the inefficiency of the attention mechanism itself, even when distilled to a smaller model, which thus leaves space for further improvement. In response, we introduce SAM-Lightening, a variant of SAM, that features a re-engineered attention mechanism, termed Dilated Flash Attention. It not only facilitates higher parallelism, enhancing processing efficiency but also retains compatibility with the existing FlashAttention. Correspondingly, we propose a progressive distillation to enable an efficient knowledge transfer from the vanilla SAM without costly training from scratch. Experiments on COCO and LVIS reveal that SAM-Lightening significantly outperforms the state-of-the-art methods in both run-time efficiency and segmentation accuracy. Specifically, it can achieve an inference speed of 7 milliseconds (ms) per image, for images of size 1024*1024 pixels, which is 30.1 times faster than the vanilla SAM and 2.1 times than the state-of-the-art. Moreover, it takes only 244MB memory, which is 3.5\\% of the vanilla SAM. The code and weights are available at https://anonymous.4open.science/r/SAM-LIGHTENING-BC25/.", "authors": ["Yanfei Song", "Bangzheng Pu", "Peng Wang", "Hongxu Jiang", "Dong Dong", "Yongxiang Cao", "Yiqing Shen"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-14", "url": "https://arxiv.org/abs/2403.09195", "pdf_url": "https://arxiv.org/pdf/2403.09195v2", "arxiv_id": "2403.09195", "doi": "10.48550/arXiv.2403.09195", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3197}
{"id": "04e7032e5f874c265354e6f57955cac915f058f7c1f1a9d997dc6c2c00236492", "sources": ["arxiv", "semantic_scholar"], "title": "MoPE: Mixture of Prompt Experts for Parameter-Efficient and Scalable Multimodal Fusion", "abstract": "Despite the demonstrated parameter efficiency of prompt-based fusion, its limited adaptivity and expressiveness hinder its effectiveness for multimodal applications at scale. In this paper, we present the first comprehensive study addressing these limitations. Our key motivation is to ``divide and conquer'' the vanilla prompt, traditionally shared across all instances, by generating instance-specific prompts. Specifically, we propose the Mixture of Prompt Experts (MoPE), a framework that significantly enhances prompt adaptivity and expressiveness by dynamically generating instance-specific prompts. MoPE leverages multimodal pairings as additional evidence, allowing the model to adaptively select optimal prompts tailored to each individual instance. Unlike traditional prompt-fusion methods, which encounter scalability bottlenecks when optimizing long unified prompts, MoPE maintains fixed prompt length while effectively scaling the number of specialized experts. Moreover, we investigate regularization terms to encourage expert specialization, resulting in highly adaptive and interpretable prompting. MoPE fundamentally changes the scaling dynamic, unlocking greater expressiveness and adaptability to complex multimodal relationships, enabling the model to selectively attend to task-relevant sub-sequences based on instance-specific multimodal input. Extensive experiments across six multimodal datasets spanning four modalities demonstrate state-of-the-art performance for multimodal fusion, matching or surpassing the performance of fine-tuning while requiring only 0.8% of the trainable parameters. Code is available: https://github.com/songrise/MoPE.", "authors": ["Ruixiang Jiang", "Lingbo Liu", "Changwen Chen"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-14", "url": "https://arxiv.org/abs/2403.10568", "pdf_url": "https://arxiv.org/pdf/2403.10568v4", "arxiv_id": "2403.10568", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/songrise/MoPE", "venue": null, "quality_score": 0.1747}
{"id": "6e018d55684c9af56125d4a82ef1a7da06e74a293098a0e0735f37b0973f92d2", "sources": ["arxiv", "semantic_scholar"], "title": "Scattered Mixture-of-Experts Implementation", "abstract": "We present ScatterMoE, an implementation of Sparse Mixture-of-Experts (SMoE) on GPUs. ScatterMoE builds upon existing implementations, and overcoming some of the limitations to improve inference and training speed, and memory footprint. This implementation achieves this by avoiding padding and making excessive copies of the input. We introduce ParallelLinear, the main component we use to build our implementation and the various kernels used to speed up the operation. We benchmark our implementation against Megablocks, and show that it enables a higher throughput and lower memory footprint. We also show how ParallelLinear enables extension of the Mixture-of-Experts concept by demonstrating with an implementation of Mixture of Attention.", "authors": ["Shawn Tan", "Yikang Shen", "Rameswar Panda", "Aaron Courville"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-13", "url": "https://arxiv.org/abs/2403.08245", "pdf_url": "https://arxiv.org/pdf/2403.08245v2", "arxiv_id": "2403.08245", "doi": "10.48550/arXiv.2403.08245", "citation_count": 14, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.301}
{"id": "e08f34e22377259558561a7d1968c7e2540af6f7b9655277e6daf2ba0b4f19c1", "sources": ["arxiv", "semantic_scholar"], "title": "Branch-Train-MiX: Mixing Expert LLMs into a Mixture-of-Experts LLM", "abstract": "We investigate efficient methods for training Large Language Models (LLMs) to possess capabilities in multiple specialized domains, such as coding, math reasoning and world knowledge. Our method, named Branch-Train-MiX (BTX), starts from a seed model, which is branched to train experts in embarrassingly parallel fashion with high throughput and reduced communication cost. After individual experts are asynchronously trained, BTX brings together their feedforward parameters as experts in Mixture-of-Expert (MoE) layers and averages the remaining parameters, followed by an MoE-finetuning stage to learn token-level routing. BTX generalizes two special cases, the Branch-Train-Merge method, which does not have the MoE finetuning stage to learn routing, and sparse upcycling, which omits the stage of training experts asynchronously. Compared to alternative approaches, BTX achieves the best accuracy-efficiency tradeoff.", "authors": ["Sainbayar Sukhbaatar", "Olga Golovneva", "Vasu Sharma", "Hu Xu", "Xi Victoria Lin", "Baptiste Rozière", "Jacob Kahn", "Daniel Li", "Wen-tau Yih", "Jason Weston", "Xian Li"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-12", "url": "https://arxiv.org/abs/2403.07816", "pdf_url": "https://arxiv.org/pdf/2403.07816v1", "arxiv_id": "2403.07816", "doi": "10.48550/arXiv.2403.07816", "citation_count": 109, "influential_citation_count": 14, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.588}
{"id": "0c23d298bb26b8f667bd7a165c97e6e886a1b5d5f7b9b14ab930ef8b79c6cfc3", "sources": ["arxiv", "semantic_scholar"], "title": "Acquiring Diverse Skills using Curriculum Reinforcement Learning with Mixture of Experts", "abstract": "Reinforcement learning (RL) is a powerful approach for acquiring a good-performing policy. However, learning diverse skills is challenging in RL due to the commonly used Gaussian policy parameterization. We propose \\textbf{Di}verse \\textbf{Skil}l \\textbf{L}earning (Di-SkilL\\footnote{Videos and code are available on the project webpage: \\url{https://alrhub.github.io/di-skill-website/}}), an RL method for learning diverse skills using Mixture of Experts, where each expert formalizes a skill as a contextual motion primitive. Di-SkilL optimizes each expert and its associate context distribution to a maximum entropy objective that incentivizes learning diverse skills in similar contexts. The per-expert context distribution enables automatic curricula learning, allowing each expert to focus on its best-performing sub-region of the context space. To overcome hard discontinuities and multi-modalities without any prior knowledge of the environment's unknown context probability space, we leverage energy-based models to represent the per-expert context distributions and demonstrate how we can efficiently train them using the standard policy gradient objective. We show on challenging robot simulation tasks that Di-SkilL can learn diverse and performant skills.", "authors": ["Onur Celik", "Aleksandar Taranovic", "Gerhard Neumann"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-11", "url": "https://arxiv.org/abs/2403.06966", "pdf_url": "https://arxiv.org/pdf/2403.06966v2", "arxiv_id": "2403.06966", "doi": "10.48550/arXiv.2403.06966", "citation_count": 20, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3306}
{"id": "d6a6872a6f345f4d8a1ebadd274e22886b3249172995e7ffcb66ee71110398dc", "sources": ["arxiv", "semantic_scholar"], "title": "GEAR: An Efficient KV Cache Compression Recipe for Near-Lossless Generative Inference of LLM", "abstract": "Key-value (KV) caching has become the de-facto to accelerate generation speed for large language models (LLMs) inference. However, the growing cache demand with increasing sequence length has transformed LLM inference to be a memory bound problem, significantly constraining the system throughput. Existing methods rely on dropping unimportant tokens or quantizing all entries uniformly. Such methods, however, often incur high approximation errors to represent the compressed matrices. The autoregressive decoding process further compounds the error of each step, resulting in critical deviation in model generation and deterioration of performance. To tackle this challenge, we propose GEAR, an efficient KV cache compression framework that achieves near-lossless high-ratio compression. GEAR first applies quantization to majority of entries of similar magnitudes to ultra-low precision. It then employs a low rank matrix to approximate the quantization error, and a sparse matrix to remedy individual errors from outlier entries. By adeptly integrating three techniques, GEAR is able to fully exploit their synergistic potentials. Our experiments demonstrate that compared to alternatives, GEAR achieves near-lossless 4-bit KV cache compression with up to 2.38x throughput improvement, while reducing peak-memory size up to 2.29x. Our code is publicly available at https://github.com/HaoKang-Timmy/GEAR.", "authors": ["Hao Kang", "Qingru Zhang", "Souvik Kundu", "Geonhwa Jeong", "Zaoxing Liu", "Tushar Krishna", "Tuo Zhao"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-08", "url": "https://arxiv.org/abs/2403.05527", "pdf_url": "https://arxiv.org/pdf/2403.05527v4", "arxiv_id": "2403.05527", "doi": "10.48550/arXiv.2403.05527", "citation_count": 162, "influential_citation_count": 8, "has_code": true, "code_url": "https://github.com/HaoKang-Timmy/GEAR", "venue": "arXiv.org", "quality_score": 0.553}
{"id": "805e844d692e2e87f16216ed7618a20f5c0c721a5f77d2f38ae107550d2a7016", "sources": ["arxiv", "semantic_scholar"], "title": "QAQ: Quality Adaptive Quantization for LLM KV Cache", "abstract": "The emergence of LLMs has ignited a fresh surge of breakthroughs in NLP applications, particularly in domains such as question-answering systems and text generation. As the need for longer context grows, a significant bottleneck in model deployment emerges due to the linear expansion of the Key-Value (KV) cache with the context length. Existing methods primarily rely on various hypotheses, such as sorting the KV cache based on attention scores for replacement or eviction, to compress the KV cache and improve model throughput. However, heuristics used by these strategies may wrongly evict essential KV cache, which can significantly degrade model performance. In this paper, we propose QAQ, a Quality Adaptive Quantization scheme for the KV cache. We theoretically demonstrate that key cache and value cache exhibit distinct sensitivities to quantization, leading to the formulation of separate quantization strategies for their non-uniform quantization. Through the integration of dedicated outlier handling, as well as an improved attention-aware approach, QAQ achieves up to 10x the compression ratio of the KV cache size with a neglectable impact on model performance. QAQ significantly reduces the practical hurdles of deploying LLMs, opening up new possibilities for longer-context applications. The code is available at github.com/ClubieDong/KVCacheQuantization.", "authors": ["Shichen Dong", "Wen Cheng", "Jiayu Qin", "Wei Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-07", "url": "https://arxiv.org/abs/2403.04643", "pdf_url": "https://arxiv.org/pdf/2403.04643v2", "arxiv_id": "2403.04643", "doi": "10.1109/ICCVW69036.2025.00267", "citation_count": 75, "influential_citation_count": 4, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.4702}
{"id": "2a3ca41ceac43dfb3a6db863089e102e45b6968f556c58a1754d6e543dd2fbf0", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-LoRAs: An Efficient Multitask Tuning for Large Language Models", "abstract": "Instruction Tuning has the potential to stimulate or enhance specific capabilities of large language models (LLMs). However, achieving the right balance of data is crucial to prevent catastrophic forgetting and interference between tasks. To address these limitations and enhance training flexibility, we propose the Mixture-of-LoRAs (MoA) architecture which is a novel and parameter-efficient tuning method designed for multi-task learning with LLMs. In this paper, we start by individually training multiple domain-specific LoRA modules using corresponding supervised corpus data. These LoRA modules can be aligned with the expert design principles observed in Mixture-of-Experts (MoE). Subsequently, we combine the multiple LoRAs using an explicit routing strategy and introduce domain labels to facilitate multi-task learning, which help prevent interference between tasks and ultimately enhances the performance of each individual task. Furthermore, each LoRA model can be iteratively adapted to a new domain, allowing for quick domain-specific adaptation. Experiments on diverse tasks demonstrate superior and robust performance, which can further promote the wide application of domain-specific LLMs.", "authors": ["Wenfeng Feng", "Chuzhan Hao", "Yuewei Zhang", "Yu Han", "Hao Wang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-06", "url": "https://arxiv.org/abs/2403.03432", "pdf_url": "https://arxiv.org/pdf/2403.03432v1", "arxiv_id": "2403.03432", "doi": "10.48550/arXiv.2403.03432", "citation_count": 36, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Language Resources and Evaluation", "quality_score": 0.3921}
{"id": "d18f8774c918ca823c13e83395854647d6123fc17c62a0c093317c9d7160a07b", "sources": ["arxiv", "semantic_scholar"], "title": "Video Relationship Detection Using Mixture of Experts", "abstract": "Machine comprehension of visual information from images and videos by neural networks faces two primary challenges. Firstly, there exists a computational and inference gap in connecting vision and language, making it difficult to accurately determine which object a given agent acts on and represent it through language. Secondly, classifiers trained by a single, monolithic neural network often lack stability and generalization. To overcome these challenges, we introduce MoE-VRD, a novel approach to visual relationship detection utilizing a mixture of experts. MoE-VRD identifies language triplets in the form of < subject, predicate, object> tuples to extract relationships from visual processing. Leveraging recent advancements in visual relationship detection, MoE-VRD addresses the requirement for action recognition in establishing relationships between subjects (acting) and objects (being acted upon). In contrast to single monolithic networks, MoE-VRD employs multiple small models as experts, whose outputs are aggregated. Each expert in MoE-VRD specializes in visual relationship learning and object tagging. By utilizing a sparsely-gated mixture of experts, MoE-VRD enables conditional computation and significantly enhances neural network capacity without increasing computational complexity. Our experimental results demonstrate that the conditional computation capabilities and scalability of the mixture-of-experts approach lead to superior performance in visual relationship detection compared to state-of-the-art methods.", "authors": ["Ala Shaabana", "Zahra Gharaee", "Paul Fieguth"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-06", "url": "https://arxiv.org/abs/2403.03994", "pdf_url": "https://arxiv.org/pdf/2403.03994v1", "arxiv_id": "2403.03994", "doi": "10.1109/ACCESS.2023.3257280", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Access", "quality_score": 0.2113}
{"id": "979bbc8f8c4d381ae5c7d0d7effbeba61b288f1b0121c60523df13e3ffa76f2f", "sources": ["arxiv", "semantic_scholar"], "title": "TESTAM: A Time-Enhanced Spatio-Temporal Attention Model with Mixture of Experts", "abstract": "Accurate traffic forecasting is challenging due to the complex dependency on road networks, various types of roads, and the abrupt speed change due to the events. Recent works mainly focus on dynamic spatial modeling with adaptive graph embedding or graph attention having less consideration for temporal characteristics and in-situ modeling. In this paper, we propose a novel deep learning model named TESTAM, which individually models recurring and non-recurring traffic patterns by a mixture-of-experts model with three experts on temporal modeling, spatio-temporal modeling with static graph, and dynamic spatio-temporal dependency modeling with dynamic graph. By introducing different experts and properly routing them, TESTAM could better model various circumstances, including spatially isolated nodes, highly related nodes, and recurring and non-recurring events. For the proper routing, we reformulate a gating problem into a classification problem with pseudo labels. Experimental results on three public traffic network datasets, METR-LA, PEMS-BAY, and EXPY-TKY, demonstrate that TESTAM achieves a better indication and modeling of recurring and non-recurring traffic. We published the official code at https://github.com/HyunWookL/TESTAM", "authors": ["Hyunwook Lee", "Sungahn Ko"], "categories": ["cs.LG", "cs.SI"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-05", "url": "https://arxiv.org/abs/2403.02600", "pdf_url": "https://arxiv.org/pdf/2403.02600v1", "arxiv_id": "2403.02600", "doi": "10.48550/arXiv.2403.02600", "citation_count": 47, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/HyunWookL/TESTAM", "venue": "International Conference on Learning Representations", "quality_score": 0.4203}
{"id": "786e7b64da1005ecf119d366d83f0540899b917ab8def5ecb802e19a42fb4423", "sources": ["arxiv", "semantic_scholar"], "title": "DéjàVu: KV-cache Streaming for Fast, Fault-tolerant Generative LLM Serving", "abstract": "Distributed LLM serving is costly and often underutilizes hardware accelerators due to three key challenges: bubbles in pipeline-parallel deployments caused by the bimodal latency of prompt and token processing, GPU memory overprovisioning, and long recovery times in case of failures. In this paper, we propose DéjàVu, a system to address all these challenges using a versatile and efficient KV cache streaming library (DéjàVuLib). Using DéjàVuLib, we propose and implement efficient prompt-token disaggregation to reduce pipeline bubbles, microbatch swapping for efficient GPU memory management, and state replication for fault-tolerance. We highlight the efficacy of these solutions on a range of large models across cloud deployments.", "authors": ["Foteini Strati", "Sara Mcallister", "Amar Phanishayee", "Jakub Tarnawski", "Ana Klimovic"], "categories": ["cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-04", "url": "https://arxiv.org/abs/2403.01876", "pdf_url": "https://arxiv.org/pdf/2403.01876v1", "arxiv_id": "2403.01876", "doi": "10.48550/arXiv.2403.01876", "citation_count": 63, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4515}
{"id": "03bc65d10f40fdb0de717236242c2dff5141165191c7061a4b293e0682f9fa64", "sources": ["arxiv", "semantic_scholar"], "title": "The Hidden Attention of Mamba Models", "abstract": "The Mamba layer offers an efficient selective state space model (SSM) that is highly effective in modeling multiple domains, including NLP, long-range sequence processing, and computer vision. Selective SSMs are viewed as dual models, in which one trains in parallel on the entire sequence via an IO-aware parallel scan, and deploys in an autoregressive manner. We add a third view and show that such models can be viewed as attention-driven models. This new perspective enables us to empirically and theoretically compare the underlying mechanisms to that of the self-attention layers in transformers and allows us to peer inside the inner workings of the Mamba model with explainability methods. Our code is publicly available.", "authors": ["Ameen Ali", "Itamar Zimerman", "Lior Wolf"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-03", "url": "https://arxiv.org/abs/2403.01590", "pdf_url": "https://arxiv.org/pdf/2403.01590v2", "arxiv_id": "2403.01590", "doi": "10.48550/arXiv.2403.01590", "citation_count": 113, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.5142}
{"id": "5d7eea29f53c798fb163f0cfeb427e512e65c8abb7e6d69744c470fd89638dc0", "sources": ["arxiv", "semantic_scholar"], "title": "No Token Left Behind: Reliable KV Cache Compression via Importance-Aware Mixed Precision Quantization", "abstract": "Key-Value (KV) Caching has become an essential technique for accelerating the inference speed and throughput of generative Large Language Models~(LLMs). However, the memory footprint of the KV cache poses a critical bottleneck in LLM deployment as the cache size grows with batch size and sequence length, often surpassing even the size of the model itself. Although recent methods were proposed to select and evict unimportant KV pairs from the cache to reduce memory consumption, the potential ramifications of eviction on the generative process are yet to be thoroughly examined. In this paper, we examine the detrimental impact of cache eviction and observe that unforeseen risks arise as the information contained in the KV pairs is exhaustively discarded, resulting in safety breaches, hallucinations, and context loss. Surprisingly, we find that preserving even a small amount of information contained in the evicted KV pairs via reduced precision quantization substantially recovers the incurred degradation. On the other hand, we observe that the important KV pairs must be kept at a relatively higher precision to safeguard the generation quality. Motivated by these observations, we propose \\textit{Mixed-precision KV cache}~(MiKV), a reliable cache compression method that simultaneously preserves the context details by retaining the evicted KV pairs in low-precision and ensure generation quality by keeping the important KV pairs in high-precision. Experiments on diverse benchmarks and LLM backbones show that our proposed method offers a state-of-the-art trade-off between compression ratio and performance, compared to other baselines.", "authors": ["June Yong Yang", "Byeongwook Kim", "Jeongin Bae", "Beomseok Kwon", "Gunho Park", "Eunho Yang", "Se Jung Kwon", "Dongsoo Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-28", "url": "https://arxiv.org/abs/2402.18096", "pdf_url": "https://arxiv.org/pdf/2402.18096v1", "arxiv_id": "2402.18096", "doi": "10.48550/arXiv.2402.18096", "citation_count": 90, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4898}
{"id": "5c1c750ae40c159308e3e4165e527ee63842d6a898a029589cda207c9a5b27c6", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-LoRA Composition for Image Generation", "abstract": "Low-Rank Adaptation (LoRA) is extensively utilized in text-to-image models for the accurate rendition of specific elements like distinct characters or unique styles in generated images. Nonetheless, existing methods face challenges in effectively composing multiple LoRAs, especially as the number of LoRAs to be integrated grows, thus hindering the creation of complex imagery. In this paper, we study multi-LoRA composition through a decoding-centric perspective. We present two training-free methods: LoRA Switch, which alternates between different LoRAs at each denoising step, and LoRA Composite, which simultaneously incorporates all LoRAs to guide more cohesive image synthesis. To evaluate the proposed approaches, we establish ComposLoRA, a new comprehensive testbed as part of this research. It features a diverse range of LoRA categories with 480 composition sets. Utilizing an evaluation framework based on GPT-4V, our findings demonstrate a clear improvement in performance with our methods over the prevalent baseline, particularly evident when increasing the number of LoRAs in a composition. The code, benchmarks, LoRA weights, and all evaluation details are available on our project website: https://maszhongming.github.io/Multi-LoRA-Composition.", "authors": ["Ming Zhong", "Yelong Shen", "Shuohang Wang", "Yadong Lu", "Yizhu Jiao", "Siru Ouyang", "Donghan Yu", "Jiawei Han", "Weizhu Chen"], "categories": ["cs.CV", "cs.AI", "cs.CL", "cs.GR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-26", "url": "https://arxiv.org/abs/2402.16843", "pdf_url": "https://arxiv.org/pdf/2402.16843v2", "arxiv_id": "2402.16843", "doi": "10.48550/arXiv.2402.16843", "citation_count": 96, "influential_citation_count": 12, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.557}
{"id": "74f30123ad539002b2724d60f92593c5b67212e3b8fb411f99b1b78d32553861", "sources": ["arxiv", "semantic_scholar"], "title": "Multimodal Instruction Tuning with Conditional Mixture of LoRA", "abstract": "Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in diverse tasks across different domains, with an increasing focus on improving their zero-shot generalization capabilities for unseen multimodal tasks. Multimodal instruction tuning has emerged as a successful strategy for achieving zero-shot generalization by fine-tuning pre-trained models on diverse multimodal tasks through instructions. As MLLMs grow in complexity and size, the need for parameter-efficient fine-tuning methods like Low-Rank Adaption (LoRA), which fine-tunes with a minimal set of parameters, becomes essential. However, applying LoRA in multimodal instruction tuning presents the challenge of task interference, which leads to performance degradation, especially when dealing with a broad array of multimodal tasks. To address this, this paper introduces a novel approach that integrates multimodal instruction tuning with Conditional Mixture-of-LoRA (MixLoRA). It innovates upon LoRA by dynamically constructing low-rank adaptation matrices tailored to the unique demands of each input instance, aiming to mitigate task interference. Experimental results on various multimodal evaluation datasets indicate that MixLoRA not only outperforms the conventional LoRA with the same or even higher ranks, demonstrating its efficacy and adaptability in diverse multimodal tasks.", "authors": ["Ying Shen", "Zhiyang Xu", "Qifan Wang", "Yu Cheng", "Wenpeng Yin", "Lifu Huang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-24", "url": "https://arxiv.org/abs/2402.15896", "pdf_url": "https://arxiv.org/pdf/2402.15896v2", "arxiv_id": "2402.15896", "doi": "10.48550/arXiv.2402.15896", "citation_count": 41, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4058}
{"id": "a244686f5a8050fd390385679e3696c14f6b92420ee70a9184d1bf6f37065450", "sources": ["arxiv", "semantic_scholar"], "title": "ChunkAttention: Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase Partition", "abstract": "Self-attention is an essential component of large language models (LLM) but a significant source of inference latency for long sequences. In multi-tenant LLM serving scenarios, the compute and memory operation cost of self-attention can be optimized by using the probability that multiple LLM requests have shared system prompts in prefixes. In this paper, we introduce ChunkAttention, a prefix-aware self-attention module that can detect matching prompt prefixes across multiple requests and share their key/value tensors in memory at runtime to improve the memory utilization of KV cache. This is achieved by breaking monolithic key/value tensors into smaller chunks and structuring them into the auxiliary prefix tree. Consequently, on top of the prefix-tree based KV cache, we design an efficient self-attention kernel, where a two-phase partition algorithm is implemented to improve the data locality during self-attention computation in the presence of shared system prompts. Experiments show that ChunkAttention can speed up the self-attention kernel by 3.2-4.8$\\times$ compared to the state-of-the-art implementation, with the length of the system prompt ranging from 1024 to 4096.", "authors": ["Lu Ye", "Ze Tao", "Yong Huang", "Yang Li"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-23", "url": "https://arxiv.org/abs/2402.15220", "pdf_url": "https://arxiv.org/pdf/2402.15220v4", "arxiv_id": "2402.15220", "doi": "10.48550/arXiv.2402.15220", "citation_count": 92, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4921}
{"id": "4db19b656ad2afffb9b54b9fecf50e80ad7426ead3a98a899e9ecd22abe2ffc4", "sources": ["arxiv", "semantic_scholar"], "title": "Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models", "abstract": "A pivotal advancement in the progress of large language models (LLMs) is the emergence of the Mixture-of-Experts (MoE) LLMs. Compared to traditional LLMs, MoE LLMs can achieve higher performance with fewer parameters, but it is still hard to deploy them due to their immense parameter sizes. Different from previous weight pruning methods that rely on specifically designed hardware, this paper mainly aims to enhance the deployment efficiency of MoE LLMs by introducing plug-and-play expert-level sparsification techniques. Specifically, we propose, for the first time to our best knowledge, post-training approaches for task-agnostic and task-specific expert pruning and skipping of MoE LLMs, tailored to improve deployment efficiency while maintaining model performance across a wide range of tasks. Extensive experiments show that our proposed methods can simultaneously reduce model sizes and increase the inference speed, while maintaining satisfactory performance. Data and code will be available at https://github.com/Lucky-Lance/Expert_Sparsity.", "authors": ["Xudong Lu", "Qi Liu", "Yuhui Xu", "Aojun Zhou", "Siyuan Huang", "Bo Zhang", "Junchi Yan", "Hongsheng Li"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-22", "url": "https://arxiv.org/abs/2402.14800", "pdf_url": "https://arxiv.org/pdf/2402.14800v2", "arxiv_id": "2402.14800", "doi": "10.48550/arXiv.2402.14800", "citation_count": 119, "influential_citation_count": 24, "has_code": true, "code_url": "https://github.com/Lucky-Lance/Expert_Sparsity", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.699}
{"id": "b95e1e4f0764d04c328d09924217651b2df92b3dcdb5c309d0cbc3a8cbb110c6", "sources": ["arxiv", "semantic_scholar"], "title": "HyperMoE: Towards Better Mixture of Experts via Transferring Among Experts", "abstract": "The Mixture of Experts (MoE) for language models has been proven effective in augmenting the capacity of models by dynamically routing each input token to a specific subset of experts for processing. Despite the success, most existing methods face a challenge for balance between sparsity and the availability of expert knowledge: enhancing performance through increased use of expert knowledge often results in diminishing sparsity during expert selection. To mitigate this contradiction, we propose HyperMoE, a novel MoE framework built upon Hypernetworks. This framework integrates the computational processes of MoE with the concept of knowledge transferring in multi-task learning. Specific modules generated based on the information of unselected experts serve as supplementary information, which allows the knowledge of experts not selected to be used while maintaining selection sparsity. Our comprehensive empirical evaluations across multiple datasets and backbones establish that HyperMoE significantly outperforms existing MoE methods under identical conditions concerning the number of experts.", "authors": ["Hao Zhao", "Zihan Qiu", "Huijia Wu", "Zili Wang", "Zhaofeng He", "Jie Fu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-20", "url": "https://arxiv.org/abs/2402.12656", "pdf_url": "https://arxiv.org/pdf/2402.12656v4", "arxiv_id": "2402.12656", "doi": "10.18653/v1/2024.acl-long.571", "citation_count": 28, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3656}
{"id": "b5645099cb61ead3b7ed71c42e95157eb93845a50d18b1aa55570079f1a86d2d", "sources": ["arxiv", "semantic_scholar"], "title": "MoELoRA: Contrastive Learning Guided Mixture of Experts on Parameter-Efficient Fine-Tuning for Large Language Models", "abstract": "Fine-tuning is often necessary to enhance the adaptability of Large Language Models (LLM) to downstream tasks. Nonetheless, the process of updating billions of parameters demands significant computational resources and training time, which poses a substantial obstacle to the widespread application of large-scale models in various scenarios. To address this issue, Parameter-Efficient Fine-Tuning (PEFT) has emerged as a prominent paradigm in recent research. However, current PEFT approaches that employ a limited set of global parameters (such as LoRA, which adds low-rank approximation matrices to all weights) face challenges in flexibly combining different computational modules in downstream tasks. In this work, we introduce a novel PEFT method: MoELoRA. We consider LoRA as Mixture of Experts (MoE), and to mitigate the random routing phenomenon observed in MoE, we propose the utilization of contrastive learning to encourage experts to learn distinct features. We conducted experiments on 11 tasks in math reasoning and common-sense reasoning benchmarks. With the same number of parameters, our approach outperforms LoRA significantly. In math reasoning, MoELoRA achieved an average performance that was 4.2% higher than LoRA, and demonstrated competitive performance compared to the 175B GPT-3.5 on several benchmarks.", "authors": ["Tongxu Luo", "Jiahe Lei", "Fangyu Lei", "Weihao Liu", "Shizhu He", "Jun Zhao", "Kang Liu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-20", "url": "https://arxiv.org/abs/2402.12851", "pdf_url": "https://arxiv.org/pdf/2402.12851v1", "arxiv_id": "2402.12851", "doi": "10.48550/arXiv.2402.12851", "citation_count": 59, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4445}
{"id": "a1e3c7ec9b0520560ab9bd60709056519850bb92b6d4a63f65984a68f51b4726", "sources": ["arxiv", "semantic_scholar"], "title": "WKVQuant: Quantizing Weight and Key/Value Cache for Large Language Models Gains More", "abstract": "Large Language Models (LLMs) face significant deployment challenges due to their substantial memory requirements and the computational demands of auto-regressive text generation process. This paper addresses these challenges by focusing on the quantization of LLMs, a technique that reduces memory consumption by converting model parameters and activations into low-bit integers. We critically analyze the existing quantization approaches, identifying their limitations in balancing the accuracy and efficiency of the quantized LLMs. To advance beyond these limitations, we propose WKVQuant, a PTQ framework especially designed for quantizing weights and the key/value (KV) cache of LLMs. Specifically, we incorporates past-only quantization to improve the computation of attention. Additionally, we introduce two-dimensional quantization strategy to handle the distribution of KV cache, along with a cross-block reconstruction regularization for parameter optimization. Experiments show that WKVQuant achieves almost comparable memory savings to weight-activation quantization, while also approaching the performance of weight-only quantization.", "authors": ["Yuxuan Yue", "Zhihang Yuan", "Haojie Duanmu", "Sifan Zhou", "Jianlong Wu", "Liqiang Nie"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-19", "url": "https://arxiv.org/abs/2402.12065", "pdf_url": "https://arxiv.org/pdf/2402.12065v2", "arxiv_id": "2402.12065", "doi": "10.48550/arXiv.2402.12065", "citation_count": 85, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4836}
{"id": "8bf3cfeebdf26f6f1a0cd2563ab131292a6bd4889d9f964de107cc2539ad3c62", "sources": ["arxiv", "semantic_scholar"], "title": "Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization", "abstract": "The Mixture of Experts (MoE) paradigm provides a powerful way to decompose dense layers into smaller, modular computations often more amenable to human interpretation, debugging, and editability. However, a major challenge lies in the computational cost of scaling the number of experts high enough to achieve fine-grained specialization. In this paper, we propose the Multilinear Mixture of Experts ($μ$MoE) layer to address this, focusing on vision models. $μ$MoE layers enable scalable expert specialization by performing an implicit computation on prohibitively large weight tensors entirely in factorized form. Consequently, $μ$MoEs (1) avoid the restrictively high inference-time costs of dense MoEs, yet (2) do not inherit the training issues of the popular sparse MoEs' discrete (non-differentiable) expert routing. We present both qualitative and quantitative evidence that scaling $μ$MoE layers when fine-tuning foundation models for vision tasks leads to more specialized experts at the class-level, further enabling manual bias correction in CelebA attribute classification. Finally, we show qualitative results demonstrating the expert specialism achieved when pre-training large GPT2 and MLP-Mixer models with parameter-matched $μ$MoE blocks at every layer, maintaining comparable accuracy. Our code is available at: https://github.com/james-oldfield/muMoE.", "authors": ["James Oldfield", "Markos Georgopoulos", "Grigorios G. Chrysos", "Christos Tzelepis", "Yannis Panagakis", "Mihalis A. Nicolaou", "Jiankang Deng", "Ioannis Patras"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-19", "url": "https://arxiv.org/abs/2402.12550", "pdf_url": "https://arxiv.org/pdf/2402.12550v4", "arxiv_id": "2402.12550", "doi": "10.48550/arXiv.2402.12550", "citation_count": 24, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/james-oldfield/muMoE", "venue": "Neural Information Processing Systems", "quality_score": 0.3495}
{"id": "9b171b9991973c6676eb1c58506f1d7f43d15df6fbc850c7d805add15c0c6e47", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-Flow: Dynamic LoRA Fusion for Large Language Models in Generative Tasks", "abstract": "LoRA employs lightweight modules to customize large language models (LLMs) for each downstream task or domain, where different learned additional modules represent diverse skills. Combining existing LoRAs to address new tasks can enhance the reusability of learned LoRAs, particularly beneficial for tasks with limited annotated data. Most prior works on LoRA combination primarily rely on task-level weights for each involved LoRA, making different examples and tokens share the same LoRA weights. However, in generative tasks, different tokens may necessitate diverse skills to manage. Taking the Chinese math task as an example, understanding the problem description may depend more on the Chinese LoRA, while the calculation part may rely more on the math LoRA. To this end, we propose LoRA-Flow, which utilizes dynamic weights to adjust the impact of different LoRAs. The weights at each step are determined by a fusion gate with extremely few parameters, which can be learned with only 200 training examples. Experiments across six generative tasks demonstrate that our method consistently outperforms baselines with task-level fusion weights. This underscores the necessity of introducing dynamic fusion weights for LoRA combination.", "authors": ["Hanqing Wang", "Bowen Ping", "Shuo Wang", "Xu Han", "Yun Chen", "Zhiyuan Liu", "Maosong Sun"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-18", "url": "https://arxiv.org/abs/2402.11455", "pdf_url": "https://arxiv.org/pdf/2402.11455v1", "arxiv_id": "2402.11455", "doi": "10.48550/arXiv.2402.11455", "citation_count": 42, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4084}
{"id": "ec02e12d5d3c73f4042d3c767540c43af52b23902b2c995b7566e06baf6f556d", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts for Network Optimization: A Large Language Model-enabled Approach", "abstract": "Optimizing various wireless user tasks poses a significant challenge for networking systems because of the expanding range of user requirements. Despite advancements in Deep Reinforcement Learning (DRL), the need for customized optimization tasks for individual users complicates developing and applying numerous DRL models, leading to substantial computation resource and energy consumption and can lead to inconsistent outcomes. To address this issue, we propose a novel approach utilizing a Mixture of Experts (MoE) framework, augmented with Large Language Models (LLMs), to analyze user objectives and constraints effectively, select specialized DRL experts, and weigh each decision from the participating experts. Specifically, we develop a gate network to oversee the expert models, allowing a collective of experts to tackle a wide array of new tasks. Furthermore, we innovatively substitute the traditional gate network with an LLM, leveraging its advanced reasoning capabilities to manage expert model selection for joint decisions. Our proposed method reduces the need to train new DRL models for each unique optimization problem, decreasing energy consumption and AI model implementation costs. The LLM-enabled MoE approach is validated through a general maze navigation task and a specific network service provider utility maximization task, demonstrating its effectiveness and practical applicability in optimizing complex networking systems.", "authors": ["Hongyang Du", "Guangyuan Liu", "Yijing Lin", "Dusit Niyato", "Jiawen Kang", "Zehui Xiong", "Dong In Kim"], "categories": ["cs.NI", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-02-15", "url": "https://arxiv.org/abs/2402.09756", "pdf_url": "https://arxiv.org/pdf/2402.09756v1", "arxiv_id": "2402.09756", "doi": "10.48550/arXiv.2402.09756", "citation_count": 15, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.301}
{"id": "7377a4cb1821a06c317791ddb7eb6cdb5e01078dfc9b2aa65fbb1f89267e6f8b", "sources": ["arxiv", "semantic_scholar"], "title": "Get More with LESS: Synthesizing Recurrence with KV Cache Compression for Efficient LLM Inference", "abstract": "Many computational factors limit broader deployment of large language models. In this paper, we focus on a memory bottleneck imposed by the key-value (KV) cache, a computational shortcut that requires storing previous KV pairs during decoding. While existing KV cache methods approach this problem by pruning or evicting large swaths of relatively less important KV pairs to dramatically reduce the memory footprint of the cache, they can have limited success in tasks that require recollecting a majority of previous tokens. To alleviate this issue, we propose LESS, a simple integration of a (nearly free) constant sized cache with eviction-based cache methods, such that all tokens can be queried at later decoding steps. Its ability to retain information throughout time shows merit on a variety of tasks where we demonstrate LESS can help reduce the performance gap from caching everything, sometimes even matching it, all while being efficient. Relevant code can be found at https://github.com/hdong920/LESS.", "authors": ["Harry Dong", "Xinyu Yang", "Zhenyu Zhang", "Zhangyang Wang", "Yuejie Chi", "Beidi Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-14", "url": "https://arxiv.org/abs/2402.09398", "pdf_url": "https://arxiv.org/pdf/2402.09398v2", "arxiv_id": "2402.09398", "doi": "10.48550/arXiv.2402.09398", "citation_count": 86, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/hdong920/LESS", "venue": "International Conference on Machine Learning", "quality_score": 0.4849}
{"id": "b6e7799b01d82aa8a46acd7a4a8d9369114000298ce4fa7abdb94bdb51678565", "sources": ["arxiv", "semantic_scholar"], "title": "Higher Layers Need More LoRA Experts", "abstract": "Parameter-efficient tuning (PEFT) techniques like low-rank adaptation (LoRA) offer training efficiency on Large Language Models, but their impact on model performance remains limited. Recent efforts integrate LoRA and Mixture-of-Experts (MoE) to improve the performance of PEFT methods. Despite promising results, research on improving the efficiency of LoRA with MoE is still in its early stages. Recent studies have shown that experts in the MoE architecture have different strengths and also exhibit some redundancy. Does this statement also apply to parameter-efficient MoE? In this paper, we introduce a novel parameter-efficient MoE method, \\textit{\\textbf{M}oE-L\\textbf{o}RA with \\textbf{L}ayer-wise Expert \\textbf{A}llocation (MoLA)} for Transformer-based models, where each model layer has the flexibility to employ a varying number of LoRA experts. We investigate several architectures with varying layer-wise expert configurations. Experiments on six well-known NLP and commonsense QA benchmarks demonstrate that MoLA achieves equal or superior performance compared to all baselines. We find that allocating more LoRA experts to higher layers further enhances the effectiveness of models with a certain number of experts in total. With much fewer parameters, this allocation strategy outperforms the setting with the same number of experts in every layer. This work can be widely used as a plug-and-play parameter-efficient tuning approach for various applications. The code is available at https://github.com/GCYZSL/MoLA.", "authors": ["Chongyang Gao", "Kezhen Chen", "Jinmeng Rao", "Baochen Sun", "Ruibo Liu", "Daiyi Peng", "Yawen Zhang", "Xiaoyuan Guo", "Jie Yang", "VS Subrahmanian"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-13", "url": "https://arxiv.org/abs/2402.08562", "pdf_url": "https://arxiv.org/pdf/2402.08562v1", "arxiv_id": "2402.08562", "doi": "10.48550/arXiv.2402.08562", "citation_count": 108, "influential_citation_count": 11, "has_code": true, "code_url": "https://github.com/GCYZSL/MoLA", "venue": "arXiv.org", "quality_score": 0.5396}
{"id": "22a1de6e34df578ba824839af150566cb43b3abd75d7b1d5d1d7f41f1773223f", "sources": ["arxiv", "semantic_scholar"], "title": "AMEND: A Mixture of Experts Framework for Long-tailed Trajectory Prediction", "abstract": "Accurate prediction of pedestrians' future motions is critical for intelligent driving systems. Developing models for this task requires rich datasets containing diverse sets of samples. However, the existing naturalistic trajectory prediction datasets are generally imbalanced in favor of simpler samples and lack challenging scenarios. Such a long-tail effect causes prediction models to underperform on the tail portion of the data distribution containing safety-critical scenarios. Previous methods tackle the long-tail problem using methods such as contrastive learning and class-conditioned hypernetworks. These approaches, however, are not modular and cannot be applied to many machine learning architectures. In this work, we propose a modular model-agnostic framework for trajectory prediction that leverages a specialized mixture of experts. In our approach, each expert is trained with a specialized skill with respect to a particular part of the data. To produce predictions, we utilise a router network that selects the best expert by generating relative confidence scores. We conduct experimentation on common pedestrian trajectory prediction datasets and show that our method improves performance on long-tail scenarios. We further conduct ablation studies to highlight the contribution of different proposed components.", "authors": ["Ray Coden Mercurius", "Ehsan Ahmadi", "Soheil Mohamad Alizadeh Shabestary", "Amir Rasouli"], "categories": ["cs.CV", "cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-13", "url": "https://arxiv.org/abs/2402.08698", "pdf_url": "https://arxiv.org/pdf/2402.08698v2", "arxiv_id": "2402.08698", "doi": "10.48550/arXiv.2402.08698", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "ece96f6e46ef63fedc21e4fb399b6841e476b36061ba0f327acef81a4b47c681", "sources": ["arxiv", "semantic_scholar"], "title": "LoRA-drop: Efficient LoRA Parameter Pruning based on Output Evaluation", "abstract": "Low-Rank Adaptation (LoRA) is currently the most commonly used Parameter-efficient fine-tuning (PEFT) method, it introduces auxiliary parameters for each layer to fine-tune the pre-trained model under limited computing resources. However, it still faces resource consumption challenges during training when scaling up to larger models. Most previous studies have tackled this issue by using pruning techniques, which involve removing LoRA parameters deemed unimportant. Nonetheless, these efforts only analyze LoRA parameter features to evaluate their importance, such as parameter count, size, and gradient. In fact, the output of LoRA (product of LoRA parameter and hidden state), directly impacts the final results. Preliminary experiments indicate that a fraction of LoRA elements possesses significantly high output values, substantially influencing the layer output. Motivated by the observation, we propose LoRA-drop. Concretely, LoRA-drop evaluates the importance of LoRA based on the LoRA output. Then we retain LoRA for important layers and the other layers share the same LoRA. We conduct abundant experiments with models of different scales on NLU and NLG tasks. Results demonstrate that LoRA-drop can achieve performance comparable to full fine-tuning and LoRA, while retaining 50\\% of the LoRA parameters on average.", "authors": ["Hongyun Zhou", "Xiangyu Lu", "Wang Xu", "Conghui Zhu", "Tiejun Zhao", "Muyun Yang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-12", "url": "https://arxiv.org/abs/2402.07721", "pdf_url": "https://arxiv.org/pdf/2402.07721v2", "arxiv_id": "2402.07721", "doi": "10.48550/arXiv.2402.07721", "citation_count": 43, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4109}
{"id": "f613f8983fc81035458cd6aa31af86de0a84094c1521b358c7dc40e14aabb2e8", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling Laws for Fine-Grained Mixture of Experts", "abstract": "Mixture of Experts (MoE) models have emerged as a primary solution for reducing the computational cost of Large Language Models. In this work, we analyze their scaling properties, incorporating an expanded range of variables. Specifically, we introduce a new hyperparameter, granularity, whose adjustment enables precise control over the size of the experts. Building on this, we establish scaling laws for fine-grained MoE, taking into account the number of training tokens, model size, and granularity. Leveraging these laws, we derive the optimal training configuration for a given computational budget. Our findings not only show that MoE models consistently outperform dense Transformers but also highlight that the efficiency gap between dense and MoE models widens as we scale up the model size and training budget. Furthermore, we demonstrate that the common practice of setting the size of experts in MoE to mirror the feed-forward layer is not optimal at almost any computational budget.", "authors": ["Jakub Krajewski", "Jan Ludziejewski", "Kamil Adamczewski", "Maciej Pióro", "Michał Krutul", "Szymon Antoniak", "Kamil Ciebiera", "Krystian Król", "Tomasz Odrzygóźdź", "Piotr Sankowski", "Marek Cygan", "Sebastian Jaszczur"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-12", "url": "https://arxiv.org/abs/2402.07871", "pdf_url": "https://arxiv.org/pdf/2402.07871v1", "arxiv_id": "2402.07871", "doi": "10.48550/arXiv.2402.07871", "citation_count": 162, "influential_citation_count": 10, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.553}
{"id": "a4d05d6d7fa04c8f551b6e7015b8894144a07e894a3d4fa58469030ed72d5354", "sources": ["arxiv", "semantic_scholar"], "title": "X-LoRA: Mixture of Low-Rank Adapter Experts, a Flexible Framework for Large Language Models with Applications in Protein Mechanics and Molecular Design", "abstract": "We report a mixture of expert strategy to create fine-tuned large language models using a deep layer-wise token-level approach based on low-rank adaptation (LoRA). Starting with a set of pre-trained LoRA adapters, our gating strategy uses the hidden states to dynamically mix adapted layers, allowing the resulting X-LoRA model to draw upon different capabilities and create never-before-used deep layer-wise combinations to solve tasks. The design is inspired by the biological principles of universality and diversity, where neural network building blocks are reused in different hierarchical manifestations. Hence, the X-LoRA model can be easily implemented for any existing large language model (LLM) without a need for modifications of the underlying structure. We develop a tailored X-LoRA model that offers scientific capabilities including forward/inverse analysis tasks and enhanced reasoning capability, focused on biomaterial analysis, protein mechanics and design. The impact of this work include access to readily expandable and adaptable models with strong domain knowledge and the capability to integrate across areas of knowledge. Featuring experts in biology, mathematics, reasoning, bio-inspired materials, mechanics and materials, chemistry, protein biophysics, mechanics and quantum-mechanics based molecular properties, we conduct a series of physics-focused case studies. We examine knowledge recall, protein mechanics forward/inverse tasks, protein design, adversarial agentic modeling including ontological knowledge graph construction, as well as molecular design. The model is capable not only of making quantitative predictions of nanomechanical properties of proteins or quantum mechanical molecular properties, but also reasons over the results and correctly predicts likely mechanisms that explain distinct molecular behaviors.", "authors": ["Eric L. Buehler", "Markus J. Buehler"], "categories": ["cond-mat.soft", "cond-mat.dis-nn", "cs.AI", "cs.CL", "cs.LG", "q-bio.QM"], "fields_of_study": ["Computer Science", "Physics", "Biology"], "published_date": "2024-02-11", "url": "https://arxiv.org/abs/2402.07148", "pdf_url": "https://arxiv.org/pdf/2402.07148v2", "arxiv_id": "2402.07148", "doi": "10.48550/arXiv.2402.07148", "citation_count": 62, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "APL Machine Learning", "quality_score": 0.4498}
{"id": "1d370911e80b90af12397f93eb1a52a3fcc3d037b21b65c3caf8b6d1b2a9ff90", "sources": ["arxiv", "semantic_scholar"], "title": "Differentially Private Training of Mixture of Experts Models", "abstract": "This position paper investigates the integration of Differential Privacy (DP) in the training of Mixture of Experts (MoE) models within the field of natural language processing. As Large Language Models (LLMs) scale to billions of parameters, leveraging expansive datasets, they exhibit enhanced linguistic capabilities and emergent abilities. However, this growth raises significant computational and privacy concerns. Our study addresses these issues by exploring the potential of MoE models, known for their computational efficiency, and the application of DP, a standard for privacy preservation. We present the first known attempt to train MoE models under the constraints of DP, addressing the unique challenges posed by their architecture and the complexities of DP integration. Our initial experimental studies demonstrate that MoE models can be effectively trained with DP, achieving performance that is competitive with their non-private counterparts. This initial study aims to provide valuable insights and ignite further research in the domain of privacy-preserving MoE models, softly laying the groundwork for prospective developments in this evolving field.", "authors": ["Pierre Tholoniat", "Huseyin A. Inan", "Janardhan Kulkarni", "Robert Sim"], "categories": ["cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-11", "url": "https://arxiv.org/abs/2402.07334", "pdf_url": "https://arxiv.org/pdf/2402.07334v1", "arxiv_id": "2402.07334", "doi": "10.48550/arXiv.2402.07334", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "85f690e45ced5bbf8d8864ad0258c32043bbe47f5fc146cff3429655323d3aa0", "sources": ["arxiv", "semantic_scholar"], "title": "Toward Scalable Generative AI via Mixture of Experts in Mobile Edge Networks", "abstract": "The advancement of generative artificial intelligence (GAI) has driven revolutionary applications like ChatGPT. The widespread of these applications relies on the mixture of experts (MoE), which contains multiple experts and selectively engages them for each task to lower operation costs while maintaining performance. Despite MoE, GAI faces challenges in resource consumption when deployed on user devices. This paper proposes mobile edge networks supported MoE-based GAI. We first review the MoE from traditional AI and GAI perspectives, including structure, principles, and applications. We then propose a framework that transfers subtasks to devices in mobile edge networks, aiding GAI model operation on user devices. We discuss challenges in this process and introduce a deep reinforcement learning based algorithm to select edge devices for subtask execution. Experimental results will show that our framework not only facilitates GAI's deployment on resource-limited devices but also generates higher-quality content compared to methods without edge network support.", "authors": ["Jiacheng Wang", "Hongyang Du", "Dusit Niyato", "Jiawen Kang", "Zehui Xiong", "Dong In Kim", "Khaled B. Letaief"], "categories": ["cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-10", "url": "https://arxiv.org/abs/2402.06942", "pdf_url": "https://arxiv.org/pdf/2402.06942v1", "arxiv_id": "2402.06942", "doi": "10.1109/MWC.003.2400046", "citation_count": 31, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE wireless communications", "quality_score": 0.3763}
{"id": "6a34ff5446b01af9505262260c77e83e827f5dcf3e6a23b86f1589c383375aa6", "sources": ["arxiv", "semantic_scholar"], "title": "Accurate LoRA-Finetuning Quantization of LLMs via Information Retention", "abstract": "The LoRA-finetuning quantization of LLMs has been extensively studied to obtain accurate yet compact LLMs for deployment on resource-constrained hardware. However, existing methods cause the quantized LLM to severely degrade and even fail to benefit from the finetuning of LoRA. This paper proposes a novel IR-QLoRA for pushing quantized LLMs with LoRA to be highly accurate through information retention. The proposed IR-QLoRA mainly relies on two technologies derived from the perspective of unified information: (1) statistics-based Information Calibration Quantization allows the quantized parameters of LLM to retain original information accurately; (2) finetuning-based Information Elastic Connection makes LoRA utilizes elastic representation transformation with diverse information. Comprehensive experiments show that IR-QLoRA can significantly improve accuracy across LLaMA and LLaMA2 families under 2-4 bit-widths, e.g., 4- bit LLaMA-7B achieves 1.4% improvement on MMLU compared with the state-of-the-art methods. The significant performance gain requires only a tiny 0.31% additional time consumption, revealing the satisfactory efficiency of our IR-QLoRA. We highlight that IR-QLoRA enjoys excellent versatility, compatible with various frameworks (e.g., NormalFloat and Integer quantization) and brings general accuracy gains. The code is available at https://github.com/htqin/ir-qlora.", "authors": ["Haotong Qin", "Xudong Ma", "Xingyu Zheng", "Xiaoyang Li", "Yang Zhang", "Shouda Liu", "Jie Luo", "Xianglong Liu", "Michele Magno"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-08", "url": "https://arxiv.org/abs/2402.05445", "pdf_url": "https://arxiv.org/pdf/2402.05445v2", "arxiv_id": "2402.05445", "doi": "10.48550/arXiv.2402.05445", "citation_count": 86, "influential_citation_count": 9, "has_code": true, "code_url": "https://github.com/htqin/ir-qlora", "venue": "International Conference on Machine Learning", "quality_score": 0.5}
{"id": "5af02393c8be4a6b9ce9fa23e226443fbfb43b18d8702ae5f3c3d35a21df25c0", "sources": ["arxiv", "semantic_scholar"], "title": "Buffer Overflow in Mixture of Experts", "abstract": "Mixture of Experts (MoE) has become a key ingredient for scaling large foundation models while keeping inference costs steady. We show that expert routing strategies that have cross-batch dependencies are vulnerable to attacks. Malicious queries can be sent to a model and can affect a model's output on other benign queries if they are grouped in the same batch. We demonstrate this via a proof-of-concept attack in a toy experimental setting.", "authors": ["Jamie Hayes", "Ilia Shumailov", "Itay Yona"], "categories": ["cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-08", "url": "https://arxiv.org/abs/2402.05526", "pdf_url": "https://arxiv.org/pdf/2402.05526v1", "arxiv_id": "2402.05526", "doi": "10.48550/arXiv.2402.05526", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.294}
{"id": "0f085685aa78046a213c04263eb842e94d95603a45012cf123be2c6b95a8d0b2", "sources": ["arxiv", "semantic_scholar"], "title": "Task-customized Masked AutoEncoder via Mixture of Cluster-conditional Experts", "abstract": "Masked Autoencoder~(MAE) is a prevailing self-supervised learning method that achieves promising results in model pre-training. However, when the various downstream tasks have data distributions different from the pre-training data, the semantically irrelevant pre-training information might result in negative transfer, impeding MAE's scalability. To address this issue, we propose a novel MAE-based pre-training paradigm, Mixture of Cluster-conditional Experts (MoCE), which can be trained once but provides customized pre-training models for diverse downstream tasks. Different from the mixture of experts (MoE), our MoCE trains each expert only with semantically relevant images by using cluster-conditional gates. Thus, each downstream task can be allocated to its customized model pre-trained with data most similar to the downstream data. Experiments on a collection of 11 downstream tasks show that MoCE outperforms the vanilla MAE by 2.45\\% on average. It also obtains new state-of-the-art self-supervised learning results on detection and segmentation.", "authors": ["Zhili Liu", "Kai Chen", "Jianhua Han", "Lanqing Hong", "Hang Xu", "Zhenguo Li", "James T. Kwok"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-08", "url": "https://arxiv.org/abs/2402.05382", "pdf_url": "https://arxiv.org/pdf/2402.05382v1", "arxiv_id": "2402.05382", "doi": "10.48550/arXiv.2402.05382", "citation_count": 33, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3829}
{"id": "96d69c789741a58ecc5512b6c6ec2d53ca82860ed8f69705c125c528df1e44aa", "sources": ["arxiv", "semantic_scholar"], "title": "On Parameter Estimation in Deviated Gaussian Mixture of Experts", "abstract": "We consider the parameter estimation problem in the deviated Gaussian mixture of experts in which the data are generated from $(1 - λ^{\\ast}) g_0(Y| X)+ λ^{\\ast} \\sum_{i = 1}^{k_{\\ast}} p_{i}^{\\ast} f(Y|(a_{i}^{\\ast})^{\\top}X+b_i^{\\ast},σ_{i}^{\\ast})$, where $X, Y$ are respectively a covariate vector and a response variable, $g_{0}(Y|X)$ is a known function, $λ^{\\ast} \\in [0, 1]$ is true but unknown mixing proportion, and $(p_{i}^{\\ast}, a_{i}^{\\ast}, b_{i}^{\\ast}, σ_{i}^{\\ast})$ for $1 \\leq i \\leq k^{\\ast}$ are unknown parameters of the Gaussian mixture of experts. This problem arises from the goodness-of-fit test when we would like to test whether the data are generated from $g_{0}(Y|X)$ (null hypothesis) or they are generated from the whole mixture (alternative hypothesis). Based on the algebraic structure of the expert functions and the distinguishability between $g_0$ and the mixture part, we construct novel Voronoi-based loss functions to capture the convergence rates of maximum likelihood estimation (MLE) for our models. We further demonstrate that our proposed loss functions characterize the local convergence rates of parameter estimation more accurately than the generalized Wasserstein, a loss function being commonly used for estimating parameters in the Gaussian mixture of experts.", "authors": ["Huy Nguyen", "Khai Nguyen", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-07", "url": "https://arxiv.org/abs/2402.05220", "pdf_url": "https://arxiv.org/pdf/2402.05220v2", "arxiv_id": "2402.05220", "doi": "10.48550/arXiv.2402.05220", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.1193}
{"id": "05f0b5167f4b6fe00633bf32ae5249aa17fc86e0ee0743c6d368832968bb3224", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Guided CAM: Visual Explanations of Vision Transformer Guided by Self-Attention", "abstract": "Vision Transformer(ViT) is one of the most widely used models in the computer vision field with its great performance on various tasks. In order to fully utilize the ViT-based architecture in various applications, proper visualization methods with a decent localization performance are necessary, but these methods employed in CNN-based models are still not available in ViT due to its unique structure. In this work, we propose an attention-guided visualization method applied to ViT that provides a high-level semantic explanation for its decision. Our method selectively aggregates the gradients directly propagated from the classification output to each self-attention, collecting the contribution of image features extracted from each location of the input image. These gradients are additionally guided by the normalized self-attention scores, which are the pairwise patch correlation scores. They are used to supplement the gradients on the patch-level context information efficiently detected by the self-attention mechanism. This approach of our method provides elaborate high-level semantic explanations with great localization performance only with the class labels. As a result, our method outperforms the previous leading explainability methods of ViT in the weakly-supervised localization task and presents great capability in capturing the full instances of the target class object. Meanwhile, our method provides a visualization that faithfully explains the model, which is demonstrated in the perturbation comparison test.", "authors": ["Saebom Leem", "Hyunseok Seo"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-07", "url": "https://arxiv.org/abs/2402.04563", "pdf_url": "https://arxiv.org/pdf/2402.04563v1", "arxiv_id": "2402.04563", "doi": "10.48550/arXiv.2402.04563", "citation_count": 44, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/LeemSaebom/Attention-Guided-CAM-Visual-Explanations-of-Vision-Transformer-Guided-by-Self-Attention.git", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4133}
{"id": "75217fd65d176d54d6c931f2a33c316c34b50e01c2b92a7b11b4ea820388a842", "sources": ["arxiv", "semantic_scholar"], "title": "KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache", "abstract": "Efficiently serving large language models (LLMs) requires batching of many requests to reduce the cost per request. Yet, with larger batch sizes and longer context lengths, the key-value (KV) cache, which stores attention keys and values to avoid re-computations, significantly increases memory demands and becomes the new bottleneck in speed and memory usage. Additionally, the loading of the KV cache causes the computational core to be idle, which limits the inference speed. A straightforward and effective solution to reduce KV cache size is quantization, which decreases the total bytes taken by KV cache. However, there is a lack of in-depth studies that explore the element distribution of KV cache to understand the hardness and limitation of KV cache quantization. To fill the gap, we conducted a comprehensive study on the element distribution in KV cache of popular LLMs. Our findings indicate that the key cache should be quantized per-channel, i.e., group elements along the channel dimension and quantize them together. In contrast, the value cache should be quantized per-token. From this analysis, we developed a tuning-free 2bit KV cache quantization algorithm named KIVI. With hardware-friendly implementation, KIVI can enable Llama, Falcon, and Mistral models to maintain almost the same quality while using $\\mathbf{2.6\\times}$ less peak memory (including model weight). This reduction in memory usage enables up to $\\mathbf{4\\times}$ larger batch size, bringing $\\mathbf{2.35\\times \\sim 3.47\\times}$ throughput on real LLM inference workload. The source code is available at https://github.com/jy-yuan/KIVI.", "authors": ["Zirui Liu", "Jiayi Yuan", "Hongye Jin", "Shaochen Zhong", "Zhaozhuo Xu", "Vladimir Braverman", "Beidi Chen", "Xia Hu"], "categories": ["cs.CL", "cs.LG", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-05", "url": "https://arxiv.org/abs/2402.02750", "pdf_url": "https://arxiv.org/pdf/2402.02750v2", "arxiv_id": "2402.02750", "doi": "10.13140/RG.2.2.28167.37282", "citation_count": 507, "influential_citation_count": 85, "has_code": true, "code_url": "https://github.com/jy-yuan/KIVI", "venue": "International Conference on Machine Learning", "quality_score": 0.9672}
{"id": "e991b109d3288f222a403efda787055ff0611dde2d1f0c05473de8343f579002", "sources": ["arxiv", "semantic_scholar"], "title": "Shortened LLaMA: Depth Pruning for Large Language Models with Comparison of Retraining Methods", "abstract": "Structured pruning of modern large language models (LLMs) has emerged as a way of decreasing their high computational needs. Width pruning reduces the size of projection weight matrices (e.g., by removing attention heads) while maintaining the number of layers. Depth pruning, in contrast, removes entire layers or blocks, while keeping the size of the remaining weights unchanged. Most current research focuses on either width-only or a blend of width and depth pruning, with little comparative analysis between the two units (width vs. depth) concerning their impact on LLM inference efficiency. In this work, we show that simple depth pruning can effectively compress LLMs while achieving comparable or superior performance to recent width pruning studies. Our pruning method boosts inference speeds, especially under memory-constrained conditions that require limited batch sizes for running LLMs, where width pruning is ineffective. In retraining pruned models for quality recovery, continued pretraining on a large corpus markedly outperforms LoRA-based tuning, particularly at severe pruning ratios. We hope this work can help build compact yet capable LLMs. Code and models can be found at: https://github.com/Nota-NetsPresso/shortened-llm", "authors": ["Bo-Kyeong Kim", "Geonmin Kim", "Tae-Ho Kim", "Thibault Castells", "Shinkook Choi", "Junho Shin", "Hyoung-Kyu Song"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-05", "url": "https://arxiv.org/abs/2402.02834", "pdf_url": "https://arxiv.org/pdf/2402.02834v2", "arxiv_id": "2402.02834", "doi": "10.48550/arXiv.2402.02834", "citation_count": 81, "influential_citation_count": 24, "has_code": true, "code_url": "https://github.com/Nota-NetsPresso/shortened-llm", "venue": "arXiv.org", "quality_score": 0.699}
{"id": "0f6ef6342815be2ccf67f2177ac64ecd607a53fc34aa8d153cbb9a6265792409", "sources": ["arxiv", "semantic_scholar"], "title": "On Least Square Estimation in Softmax Gating Mixture of Experts", "abstract": "Mixture of experts (MoE) model is a statistical machine learning design that aggregates multiple expert networks using a softmax gating function in order to form a more intricate and expressive model. Despite being commonly used in several applications owing to their scalability, the mathematical and statistical properties of MoE models are complex and difficult to analyze. As a result, previous theoretical works have primarily focused on probabilistic MoE models by imposing the impractical assumption that the data are generated from a Gaussian MoE model. In this work, we investigate the performance of the least squares estimators (LSE) under a deterministic MoE model where the data are sampled according to a regression model, a setting that has remained largely unexplored. We establish a condition called strong identifiability to characterize the convergence behavior of various types of expert functions. We demonstrate that the rates for estimating strongly identifiable experts, namely the widely used feed-forward networks with activation functions $\\mathrm{sigmoid}(\\cdot)$ and $\\tanh(\\cdot)$, are substantially faster than those of polynomial experts, which we show to exhibit a surprising slow estimation rate. Our findings have important practical implications for expert selection.", "authors": ["Huy Nguyen", "Nhat Ho", "Alessandro Rinaldo"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2024-02-05", "url": "https://arxiv.org/abs/2402.02952", "pdf_url": "https://arxiv.org/pdf/2402.02952v2", "arxiv_id": "2402.02952", "doi": null, "citation_count": 26, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3578}
{"id": "50325cdf0492ef38b2fbe7f3113cc479d7bd565598de7645ff51b6bd1544193b", "sources": ["arxiv", "semantic_scholar"], "title": "Approximation Rates and VC-Dimension Bounds for (P)ReLU MLP Mixture of Experts", "abstract": "Mixture-of-Experts (MoEs) can scale up beyond traditional deep learning models by employing a routing strategy in which each input is processed by a single \"expert\" deep learning model. This strategy allows us to scale up the number of parameters defining the MoE while maintaining sparse activation, i.e., MoEs only load a small number of their total parameters into GPU VRAM for the forward pass depending on the input. In this paper, we provide an approximation and learning-theoretic analysis of mixtures of expert MLPs with (P)ReLU activation functions. We first prove that for every error level $\\varepsilon>0$ and every Lipschitz function $f:[0,1]^n\\to \\mathbb{R}$, one can construct a MoMLP model (a Mixture-of-Experts comprising of (P)ReLU MLPs) which uniformly approximates $f$ to $\\varepsilon$ accuracy over $[0,1]^n$, while only requiring networks of $\\mathcal{O}(\\varepsilon^{-1})$ parameters to be loaded in memory. Additionally, we show that MoMLPs can generalize since the entire MoMLP model has a (finite) VC dimension of $\\tilde{O}(L\\max\\{nL,JW\\})$, if there are $L$ experts and each expert has a depth and width of $J$ and $W$, respectively.", "authors": ["Anastasis Kratsios", "Haitz Sáez de Ocáriz Borde", "Takashi Furuya", "Marc T. Law"], "categories": ["stat.ML", "cs.LG", "cs.NE", "math.CO", "math.NA"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2024-02-05", "url": "https://arxiv.org/abs/2402.03460", "pdf_url": "https://arxiv.org/pdf/2402.03460v2", "arxiv_id": "2402.03460", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "441fa6865111ed6a9b36bad70a65ba426f34cb14a07ab8100c91e2091c27dc16", "sources": ["arxiv", "semantic_scholar"], "title": "Is Mamba Capable of In-Context Learning?", "abstract": "State of the art foundation models such as GPT-4 perform surprisingly well at in-context learning (ICL), a variant of meta-learning concerning the learned ability to solve tasks during a neural network forward pass, exploiting contextual information provided as input to the model. This useful ability emerges as a side product of the foundation model's massive pretraining. While transformer models are currently the state of the art in ICL, this work provides empirical evidence that Mamba, a newly proposed state space model which scales better than transformers w.r.t. the input sequence length, has similar ICL capabilities. We evaluated Mamba on tasks involving simple function approximation as well as more complex natural language processing problems. Our results demonstrate that, across both categories of tasks, Mamba closely matches the performance of transformer models for ICL. Further analysis reveals that, like transformers, Mamba appears to solve ICL problems by incrementally optimizing its internal representations. Overall, our work suggests that Mamba can be an efficient alternative to transformers for ICL tasks involving long input sequences. This is an exciting finding in meta-learning and may enable generalizations of in-context learned AutoML algorithms (like TabPFN or Optformer) to long input sequences.", "authors": ["Riccardo Grazzi", "Julien Siems", "Simon Schrodi", "Thomas Brox", "Frank Hutter"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-05", "url": "https://arxiv.org/abs/2402.03170", "pdf_url": "https://arxiv.org/pdf/2402.03170v2", "arxiv_id": "2402.03170", "doi": "10.48550/arXiv.2402.03170", "citation_count": 65, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.4549}
{"id": "5743f4db6d90ea691ed873ad6999b7d2f6a2e34ae57b65103b6341aaf52670a2", "sources": ["arxiv", "semantic_scholar"], "title": "CompeteSMoE -- Effective Training of Sparse Mixture of Experts via Competition", "abstract": "Sparse mixture of experts (SMoE) offers an appealing solution to scale up the model complexity beyond the mean of increasing the network's depth or width. However, effective training of SMoE has proven to be challenging due to the representation collapse issue, which causes parameter redundancy and limited representation potentials. In this work, we propose a competition mechanism to address this fundamental challenge of representation collapse. By routing inputs only to experts with the highest neural response, we show that, under mild assumptions, competition enjoys the same convergence rate as the optimal estimator. We further propose CompeteSMoE, an effective and efficient algorithm to train large language models by deploying a simple router that predicts the competition outcomes. Consequently, CompeteSMoE enjoys strong performance gains from the competition routing policy while having low computation overheads. Our extensive empirical evaluations on two transformer architectures and a wide range of tasks demonstrate the efficacy, robustness, and scalability of CompeteSMoE compared to state-of-the-art SMoE strategies.", "authors": ["Quang Pham", "Giang Do", "Huy Nguyen", "TrungTin Nguyen", "Chenghao Liu", "Mina Sartipi", "Binh T. Nguyen", "Savitha Ramasamy", "Xiaoli Li", "Steven Hoi", "Nhat Ho"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-04", "url": "https://arxiv.org/abs/2402.02526", "pdf_url": "https://arxiv.org/pdf/2402.02526v1", "arxiv_id": "2402.02526", "doi": "10.48550/arXiv.2402.02526", "citation_count": 24, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "844fe0d810378533321db6338fac2da29ecd6928fce44cf8f814f86990ef1ae6", "sources": ["arxiv", "semantic_scholar"], "title": "Attention-Mamba: A Mamba-Enhanced Multi-Scale Parallel Inference Network for Medical Image Segmentation", "abstract": "U-shaped architectures have long dominated the field of medical image segmentation, while Transformers are widely employed for modeling long-range dependencies. The former typically handles scale variations implicitly by aggregating multi-level features, whereas the efficiency of the latter is constrained by its quadratic computational and memory complexity. In this work, we propose an effective alternative to traditional U-shaped architectures by constructing parallel branches at different levels to obtain multi-scale features and corresponding predictions. Furthermore, we enhance our network by integrating Mamba, a state space model that captures long-range dependencies with linear complexity. First, a dual-path architecture with lateral connections aggregates high-level semantic information and low-level spatial details at each branch. Then, we introduce a Recursive Alignment Module (RAM) that restores spatial details in low-resolution features through stepwise alignment, optimizing them for subsequent global feature learning and multi-scale fusion. We further build parallel Mamba branches upon aligned features to establish hierarchical global representations. Finally, we propose a Mamba-based attention mechanism for adaptive multi-scale prediction fusion; this mechanism utilizes Mamba to enhance information exchange across scales along both the channel and spatial dimensions. Experiments across three imaging modalities (MRI, CT, and dermoscopy) underscore the superior generalization of the proposed network. Compared to state-of-the-art 2D CNN, Transformer, and Mamba-based networks, our model achieves the highest segmentation performance on the Synapse, ACDC, ISIC-2018, and PH2 datasets while maintaining high efficiency, featuring the second-smallest parameters (14.05 M) and moderate computational complexity (8.94 GFLOPs).", "authors": ["Yanhua Zhang", "Ke Zhang", "Jingyu Wang", "Gabriella Balestra", "Samanta Rosati", "Yulin Wu", "Wuwei Wang", "Valentina Giannini"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-03", "url": "https://arxiv.org/abs/2402.02286", "pdf_url": "https://arxiv.org/pdf/2402.02286v4", "arxiv_id": "2402.02286", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "9456f6e168a25f32aa4c9e2ce5ab7f3cc765a6169e79001bb51cae76cf5e7f8c", "sources": ["arxiv", "semantic_scholar"], "title": "Self-Attention through Kernel-Eigen Pair Sparse Variational Gaussian Processes", "abstract": "While the great capability of Transformers significantly boosts prediction accuracy, it could also yield overconfident predictions and require calibrated uncertainty estimation, which can be commonly tackled by Gaussian processes (GPs). Existing works apply GPs with symmetric kernels under variational inference to the attention kernel; however, omitting the fact that attention kernels are in essence asymmetric. Moreover, the complexity of deriving the GP posteriors remains high for large-scale data. In this work, we propose Kernel-Eigen Pair Sparse Variational Gaussian Processes (KEP-SVGP) for building uncertainty-aware self-attention where the asymmetry of attention kernels is tackled by Kernel SVD (KSVD) and a reduced complexity is acquired. Through KEP-SVGP, i) the SVGP pair induced by the two sets of singular vectors from KSVD w.r.t. the attention kernel fully characterizes the asymmetry; ii) using only a small set of adjoint eigenfunctions from KSVD, the derivation of SVGP posteriors can be based on the inversion of a diagonal matrix containing singular values, contributing to a reduction in time complexity; iii) an evidence lower bound is derived so that variational parameters and network weights can be optimized with it. Experiments verify our excellent performances and efficiency on in-distribution, distribution-shift and out-of-distribution benchmarks.", "authors": ["Yingyi Chen", "Qinghua Tao", "Francesco Tonin", "Johan A. K. Suykens"], "categories": ["cs.LG", "cs.AI", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-02", "url": "https://arxiv.org/abs/2402.01476", "pdf_url": "https://arxiv.org/pdf/2402.01476v2", "arxiv_id": "2402.01476", "doi": "10.48550/arXiv.2402.01476", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1945}
{"id": "a48d264e80415c0ba43066fcbf161cc5094af0953b56918009c90072cbe67774", "sources": ["arxiv", "semantic_scholar"], "title": "BlackMamba: Mixture of Experts for State-Space Models", "abstract": "State-space models (SSMs) have recently demonstrated competitive performance to transformers at large-scale language modeling benchmarks while achieving linear time and memory complexity as a function of sequence length. Mamba, a recently released SSM model, shows impressive performance in both language modeling and long sequence processing tasks. Simultaneously, mixture-of-expert (MoE) models have shown remarkable performance while significantly reducing the compute and latency costs of inference at the expense of a larger memory footprint. In this paper, we present BlackMamba, a novel architecture that combines the Mamba SSM with MoE to obtain the benefits of both. We demonstrate that BlackMamba performs competitively against both Mamba and transformer baselines, and outperforms in inference and training FLOPs. We fully train and open-source 340M/1.5B and 630M/2.8B BlackMamba models on 300B tokens of a custom dataset. We show that BlackMamba inherits and combines both of the benefits of SSM and MoE architectures, combining linear-complexity generation from SSM with cheap and fast inference from MoE. We release all weights, checkpoints, and inference code open-source. Inference code at: https://github.com/Zyphra/BlackMamba", "authors": ["Quentin Anthony", "Yury Tokpanov", "Paolo Glorioso", "Beren Millidge"], "categories": ["cs.CL", "cs.AI", "cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-01", "url": "https://arxiv.org/abs/2402.01771", "pdf_url": "https://arxiv.org/pdf/2402.01771v1", "arxiv_id": "2402.01771", "doi": "10.48550/arXiv.2402.01771", "citation_count": 40, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/Zyphra/BlackMamba", "venue": "arXiv.org", "quality_score": 0.4032}
{"id": "a38639bf35efec39b9c73338b81673c7ff5edb1c138b1b3464f3e6bdc04de3a2", "sources": ["arxiv", "semantic_scholar"], "title": "KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization", "abstract": "LLMs are seeing growing use for applications which require large context windows, and with these large context windows KV cache activations surface as the dominant contributor to memory consumption during inference. Quantization is a promising approach for compressing KV cache activations; however, existing solutions fail to represent activations accurately in sub-4-bit precision. Our work, KVQuant, facilitates low precision KV cache quantization by incorporating several novel methods: (i) Per-Channel Key Quantization, where we adjust the dimension along which we quantize the Key activations to better match the distribution; (ii) Pre-RoPE Key Quantization, where we quantize Key activations before the rotary positional embedding to mitigate its impact on quantization; (iii) Non-Uniform KV Cache Quantization, where we derive per-layer sensitivity-weighted non-uniform datatypes that better represent the distributions; and (iv) Per-Vector Dense-and-Sparse Quantization, where we isolate outliers separately for each vector to minimize skews in quantization ranges. By applying our method to the LLaMA, Llama-2, Llama-3, and Mistral models, we achieve < 0.1 perplexity degradation with 3-bit quantization on both Wikitext-2 and C4, outperforming existing approaches. Our method enables serving LLaMA-7B with a context length of up to 1 million on a single A100-80GB GPU and up to 10 million on an 8-GPU system. We develop custom CUDA kernels for KVQuant, showing that we can achieve up to ~1.7x speedups, compared to baseline fp16 matrix-vector multiplications, for the LLaMA-7B model.", "authors": ["Coleman Hooper", "Sehoon Kim", "Hiva Mohammadzadeh", "Michael W. Mahoney", "Yakun Sophia Shao", "Kurt Keutzer", "Amir Gholami"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-31", "url": "https://arxiv.org/abs/2401.18079", "pdf_url": "https://arxiv.org/pdf/2401.18079v6", "arxiv_id": "2401.18079", "doi": "10.48550/arXiv.2401.18079", "citation_count": 547, "influential_citation_count": 58, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.8854}
{"id": "a99b5a50e13d8a4db9632cbb45e7b5ff79f3afd643ec6cfa4a99530087d7d4a1", "sources": ["arxiv", "semantic_scholar"], "title": "MoDE: A Mixture-of-Experts Model with Mutual Distillation among the Experts", "abstract": "The application of mixture-of-experts (MoE) is gaining popularity due to its ability to improve model's performance. In an MoE structure, the gate layer plays a significant role in distinguishing and routing input features to different experts. This enables each expert to specialize in processing their corresponding sub-tasks. However, the gate's routing mechanism also gives rise to narrow vision: the individual MoE's expert fails to use more samples in learning the allocated sub-task, which in turn limits the MoE to further improve its generalization ability. To effectively address this, we propose a method called Mixture-of-Distilled-Expert (MoDE), which applies moderate mutual distillation among experts to enable each expert to pick up more features learned by other experts and gain more accurate perceptions on their original allocated sub-tasks. We conduct plenty experiments including tabular, NLP and CV datasets, which shows MoDE's effectiveness, universality and robustness. Furthermore, we develop a parallel study through innovatively constructing \"expert probing\", to experimentally prove why MoDE works: moderate distilling knowledge can improve each individual expert's test performances on their assigned tasks, leading to MoE's overall performance improvement.", "authors": ["Zhitian Xie", "Yinger Zhang", "Chenyi Zhuang", "Qitao Shi", "Zhining Liu", "Jinjie Gu", "Guannan Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-31", "url": "https://arxiv.org/abs/2402.00893", "pdf_url": "https://arxiv.org/pdf/2402.00893v1", "arxiv_id": "2402.00893", "doi": "10.48550/arXiv.2402.00893", "citation_count": 23, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3451}
{"id": "bb0e4b0f8cd968c7cc2526c5df3301bef928e173bd78344a47bd59f25d9fa23b", "sources": ["arxiv", "semantic_scholar"], "title": "LLaVA-MoLE: Sparse Mixture of LoRA Experts for Mitigating Data Conflicts in Instruction Finetuning MLLMs", "abstract": "Instruction finetuning on a variety of image-text instruction data is the key to obtaining a versatile Multimodal Large Language Model (MLLM), and different configurations of the instruction data can lead to finetuned models with different capabilities. However, we have discovered that data conflicts are inevitable when mixing instruction data from distinct domains, which can result in performance drops for tasks of a specific domain. To address this issue, we propose to apply an efficient Mixture of Experts (MoE) design, which is a sparse Mixture of LoRA Experts (MoLE) for instruction finetuning MLLMs. Within the Transformer layers, we extend the popular Low-Rank Adaption (LoRA) method by creating a set of LoRA experts specifically for the MLP layer, and route each token to the top-1 expert based on a routing function, allowing adaptive choices for tokens from different domains. Since the LoRA experts are sparsely activated, the training and inference cost are kept roughly constant compared to the original LoRA method. By replacing the plain-LoRA of LLaVA-1.5 with our MoE design, our final model is named LLaVA-MoLE. Extensive experiments proved that LLaVA-MoLE effectively mitigates the data conflict issue when mixing multiple distinct instruction datasets with various configurations, and achieves consistent performance gains over the strong plain-LoRA baselines. Most importantly, on the mixed datasets, LLaVA-MoLE can even outperform the plain-LoRA baseline trained with twice the samples.", "authors": ["Shaoxiang Chen", "Zequn Jie", "Lin Ma"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-29", "url": "https://arxiv.org/abs/2401.16160", "pdf_url": "https://arxiv.org/pdf/2401.16160v2", "arxiv_id": "2401.16160", "doi": "10.48550/arXiv.2401.16160", "citation_count": 101, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5022}
{"id": "acc03edc62e4e3ab3d6201aae2d831a0fcf0a46b8b9cc96db8f35527091982ec", "sources": ["arxiv", "semantic_scholar"], "title": "Routers in Vision Mixture of Experts: An Empirical Study", "abstract": "Mixture-of-Experts (MoE) models are a promising way to scale up model capacity without significantly increasing computational cost. A key component of MoEs is the router, which decides which subset of parameters (experts) process which feature embeddings (tokens). In this paper, we present a comprehensive study of routers in MoEs for computer vision tasks. We introduce a unified MoE formulation that subsumes different MoEs with two parametric routing tensors. This formulation covers both sparse MoE, which uses a binary or hard assignment between experts and tokens, and soft MoE, which uses a soft assignment between experts and weighted combinations of tokens. Routers for sparse MoEs can be further grouped into two variants: Token Choice, which matches experts to each token, and Expert Choice, which matches tokens to each expert. We conduct head-to-head experiments with 6 different routers, including existing routers from prior work and new ones we introduce. We show that (i) many routers originally developed for language modeling can be adapted to perform strongly in vision tasks, (ii) in sparse MoE, Expert Choice routers generally outperform Token Choice routers, and (iii) soft MoEs generally outperform sparse MoEs with a fixed compute budget. These results provide new insights regarding the crucial role of routers in vision MoE models.", "authors": ["Tianlin Liu", "Mathieu Blondel", "Carlos Riquelme", "Joan Puigcerver"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-29", "url": "https://arxiv.org/abs/2401.15969", "pdf_url": "https://arxiv.org/pdf/2401.15969v2", "arxiv_id": "2401.15969", "doi": "10.48550/arXiv.2401.15969", "citation_count": 20, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3306}
{"id": "4fca1256200b13e5b72caaac9a5182e9f71e649e5ce0fd25684d7f9719c54f40", "sources": ["arxiv", "semantic_scholar"], "title": "Contrastive Learning and Mixture of Experts Enables Precise Vector Embeddings", "abstract": "The advancement of transformer neural networks has significantly elevated the capabilities of sentence similarity models, but they still struggle with highly discriminative tasks and may produce sub-optimal representations of important documents like scientific literature. With the increased reliance on retrieval augmentation and search, representing diverse documents as concise and descriptive vectors is crucial. This paper improves upon the vectors embeddings of scientific text by assembling niche datasets using co-citations as a similarity metric, focusing on biomedical domains. We apply a novel Mixture of Experts (MoE) extension pipeline to pretrained BERT models, where every multi-layer perceptron section is enlarged and copied into multiple distinct experts. Our MoE variants perform well over $N$ scientific domains with $N$ dedicated experts, whereas standard BERT models excel in only one domain at a time. Notably, extending just a single transformer block to MoE captures 85% of the benefit seen from full MoE extension at every layer. This holds promise for versatile and efficient One-Size-Fits-All transformer networks for numerically representing diverse inputs. Our methodology marks advancements in representation learning and holds promise for enhancing vector database search and compilation.", "authors": ["Logan Hallee", "Rohan Kapur", "Arjun Patel", "Jason P. Gleghorn", "Bohdan Khomtchouk"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-28", "url": "https://arxiv.org/abs/2401.15713", "pdf_url": "https://arxiv.org/pdf/2401.15713v3", "arxiv_id": "2401.15713", "doi": "10.48550/arXiv.2401.15713", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "e29183fbae54a78080d5e81f37bbab99a6d33b4eb279c1d4081e723f2ab10c81", "sources": ["arxiv", "semantic_scholar"], "title": "Is Temperature Sample Efficient for Softmax Gaussian Mixture of Experts?", "abstract": "Dense-to-sparse gating mixture of experts (MoE) has recently become an effective alternative to a well-known sparse MoE. Rather than fixing the number of activated experts as in the latter model, which could limit the investigation of potential experts, the former model utilizes the temperature to control the softmax weight distribution and the sparsity of the MoE during training in order to stabilize the expert specialization. Nevertheless, while there are previous attempts to theoretically comprehend the sparse MoE, a comprehensive analysis of the dense-to-sparse gating MoE has remained elusive. Therefore, we aim to explore the impacts of the dense-to-sparse gate on the maximum likelihood estimation under the Gaussian MoE in this paper. We demonstrate that due to interactions between the temperature and other model parameters via some partial differential equations, the convergence rates of parameter estimations are slower than any polynomial rates, and could be as slow as $\\mathcal{O}(1/\\log(n))$, where $n$ denotes the sample size. To address this issue, we propose using a novel activation dense-to-sparse gate, which routes the output of a linear layer to an activation function before delivering them to the softmax function. By imposing linearly independence conditions on the activation function and its derivatives, we show that the parameter estimation rates are significantly improved to polynomial rates. Finally, we conduct a simulation study to empirically validate our theoretical results.", "authors": ["Huy Nguyen", "Pedram Akbarian", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-01-25", "url": "https://arxiv.org/abs/2401.13875", "pdf_url": "https://arxiv.org/pdf/2401.13875v2", "arxiv_id": "2401.13875", "doi": "10.48550/arXiv.2401.13875", "citation_count": 20, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3306}
{"id": "4ecf6c33e475bb09acdc573e614d3fa6d5e376f013084e57016ba7e0e5975336", "sources": ["arxiv", "semantic_scholar"], "title": "Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model", "abstract": "Recently the state space models (SSMs) with efficient hardware-aware designs, i.e., the Mamba deep learning model, have shown great potential for long sequence modeling. Meanwhile building efficient and generic vision backbones purely upon SSMs is an appealing direction. However, representing visual data is challenging for SSMs due to the position-sensitivity of visual data and the requirement of global context for visual understanding. In this paper, we show that the reliance on self-attention for visual representation learning is not necessary and propose a new generic vision backbone with bidirectional Mamba blocks (Vim), which marks the image sequences with position embeddings and compresses the visual representation with bidirectional state space models. On ImageNet classification, COCO object detection, and ADE20k semantic segmentation tasks, Vim achieves higher performance compared to well-established vision transformers like DeiT, while also demonstrating significantly improved computation & memory efficiency. For example, Vim is 2.8$\\times$ faster than DeiT and saves 86.8% GPU memory when performing batch inference to extract features on images with a resolution of 1248$\\times$1248. The results demonstrate that Vim is capable of overcoming the computation & memory constraints on performing Transformer-style understanding for high-resolution images and it has great potential to be the next-generation backbone for vision foundation models. Code is available at https://github.com/hustvl/Vim.", "authors": ["Lianghui Zhu", "Bencheng Liao", "Qian Zhang", "Xinlong Wang", "Wenyu Liu", "Xinggang Wang"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-17", "url": "https://arxiv.org/abs/2401.09417", "pdf_url": "https://arxiv.org/pdf/2401.09417v3", "arxiv_id": "2401.09417", "doi": "10.48550/arXiv.2401.09417", "citation_count": 1834, "influential_citation_count": 190, "has_code": true, "code_url": "https://github.com/hustvl/Vim", "venue": "International Conference on Machine Learning", "quality_score": 1.0}
{"id": "9cf88d2e677dc5703a2b3e4ae68cd63cd81c016e73d4abf896f4244132e1989d", "sources": ["arxiv", "semantic_scholar"], "title": "Exploiting Inter-Layer Expert Affinity for Accelerating Mixture-of-Experts Model Inference", "abstract": "In large language models like the Generative Pre-trained Transformer, the Mixture of Experts paradigm has emerged as a powerful technique for enhancing model expressiveness and accuracy. However, deploying GPT MoE models for parallel inference on distributed systems presents significant challenges, primarily due to the extensive Alltoall communication required for expert routing and aggregation. This communication bottleneck exacerbates the already complex computational landscape, hindering the efficient utilization of high-performance computing resources. In this paper, we propose a lightweight optimization technique called ExFlow, to largely accelerate the inference of these MoE models. We take a new perspective on alleviating the communication overhead by exploiting the inter-layer expert affinity. Unlike previous methods, our solution can be directly applied to pre-trained MoE models without any fine-tuning or accuracy degradation. By proposing a context-coherent expert parallelism on distributed systems, our design only uses one Alltoall communication to deliver the same functionality while previous methods all require two Alltoalls. By carefully examining the conditional probability in tokens' routing across multiple layers, we proved that pre-trained GPT MoE models implicitly exhibit a strong inter-layer expert affinity. We then design an efficient integer programming model to capture such features and show that by properly placing the experts on corresponding GPUs, we can reduce up to 67% cross-GPU routing latency. Our solution beats the cutting-edge MoE implementations with experts from 8 to 64, with up to 2.2x improvement in inference throughput. We further provide a detailed study of how the model implicitly acquires this expert affinity at the very early training stage and how this affinity evolves and stabilizes during training.", "authors": ["Jinghan Yao", "Quentin Anthony", "Aamir Shafi", "Hari Subramoni", "Dhabaleswar K.", " Panda"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-16", "url": "https://arxiv.org/abs/2401.08383", "pdf_url": "https://arxiv.org/pdf/2401.08383v2", "arxiv_id": "2401.08383", "doi": "10.1109/IPDPS57955.2024.00086", "citation_count": 38, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "IEEE International Parallel and Distributed Processing Symposium", "quality_score": 0.4515}
{"id": "a0a81b47c9720442ca8fb146acaa65ef6b143a5cb96f337f7c63cc05b44fb384", "sources": ["arxiv", "semantic_scholar"], "title": "DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models", "abstract": "In the era of large language models, Mixture-of-Experts (MoE) is a promising architecture for managing computational costs when scaling up model parameters. However, conventional MoE architectures like GShard, which activate the top-$K$ out of $N$ experts, face challenges in ensuring expert specialization, i.e. each expert acquires non-overlapping and focused knowledge. In response, we propose the DeepSeekMoE architecture towards ultimate expert specialization. It involves two principal strategies: (1) finely segmenting the experts into $mN$ ones and activating $mK$ from them, allowing for a more flexible combination of activated experts; (2) isolating $K_s$ experts as shared ones, aiming at capturing common knowledge and mitigating redundancy in routed experts. Starting from a modest scale with 2B parameters, we demonstrate that DeepSeekMoE 2B achieves comparable performance with GShard 2.9B, which has 1.5 times the expert parameters and computation. In addition, DeepSeekMoE 2B nearly approaches the performance of its dense counterpart with the same number of total parameters, which set the upper bound of MoE models. Subsequently, we scale up DeepSeekMoE to 16B parameters and show that it achieves comparable performance with LLaMA2 7B, with only about 40% of computations. Further, our preliminary efforts to scale up DeepSeekMoE to 145B parameters consistently validate its substantial advantages over the GShard architecture, and show its performance comparable with DeepSeek 67B, using only 28.5% (maybe even 18.2%) of computations.", "authors": ["Damai Dai", "Chengqi Deng", "Chenggang Zhao", "R. X. Xu", "Huazuo Gao", "Deli Chen", "Jiashi Li", "Wangding Zeng", "Xingkai Yu", "Y. Wu", "Zhenda Xie", "Y. K. Li", "Panpan Huang", "Fuli Luo", "Chong Ruan", "Zhifang Sui", "Wenfeng Liang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-11", "url": "https://arxiv.org/abs/2401.06066", "pdf_url": "https://arxiv.org/pdf/2401.06066v1", "arxiv_id": "2401.06066", "doi": "10.48550/arXiv.2401.06066", "citation_count": 892, "influential_citation_count": 105, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 1.0}
{"id": "4a4571ad5847b8df440723d86fd0cb9114b9ac17cbd0b69c1ece7a6e28eb0248", "sources": ["arxiv", "semantic_scholar"], "title": "Lightning Attention-2: A Free Lunch for Handling Unlimited Sequence Lengths in Large Language Models", "abstract": "Linear attention is an efficient attention mechanism that has recently emerged as a promising alternative to conventional softmax attention. With its ability to process tokens in linear computational complexities, linear attention, in theory, can handle sequences of unlimited length without sacrificing speed, i.e., maintaining a constant training speed for various sequence lengths with a fixed memory consumption. However, due to the issue with cumulative summation (cumsum), current linear attention algorithms cannot demonstrate their theoretical advantage in a causal setting. In this paper, we present Lightning Attention-2, the first linear attention implementation that enables linear attention to realize its theoretical computational benefits. To achieve this, we leverage the thought of tiling, separately handling the intra-block and inter-block components in linear attention calculation. Specifically, we utilize the conventional attention computation mechanism for the intra-blocks and apply linear attention kernel tricks for the inter-blocks. A tiling technique is adopted through both forward and backward procedures to take full advantage of the GPU hardware. We implement our algorithm in Triton to make it IO-aware and hardware-friendly. Various experiments are conducted on different model sizes and sequence lengths. Lightning Attention-2 retains consistent training and inference speed regardless of input sequence length and is significantly faster than other attention mechanisms. The source code is available at https://github.com/OpenNLPLab/lightning-attention.", "authors": ["Zhen Qin", "Weigao Sun", "Dong Li", "Xuyang Shen", "Weixuan Sun", "Yiran Zhong"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-09", "url": "https://arxiv.org/abs/2401.04658", "pdf_url": "https://arxiv.org/pdf/2401.04658v2", "arxiv_id": "2401.04658", "doi": "10.48550/arXiv.2401.04658", "citation_count": 55, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/OpenNLPLab/lightning-attention", "venue": "arXiv.org", "quality_score": 0.437}
{"id": "1f701b3e1bae9ceb6e22d474cecdc14f9db106aeef4a020a446439ab73050647", "sources": ["arxiv", "semantic_scholar"], "title": "MoE-Mamba: Efficient Selective State Space Models with Mixture of Experts", "abstract": "State Space Models (SSMs) have become serious contenders in the field of sequential modeling, challenging the dominance of Transformers. At the same time, Mixture of Experts (MoE) has significantly improved Transformer-based Large Language Models, including recent state-of-the-art open models. We propose that to unlock the potential of SSMs for scaling, they should be combined with MoE. We showcase this on Mamba, a recent SSM-based model that achieves remarkable performance. Our model, MoE-Mamba, outperforms both Mamba and baseline Transformer-MoE. In particular, MoE-Mamba reaches the same performance as Mamba in $2.35\\times$ fewer training steps while preserving the inference performance gains of Mamba against Transformer.", "authors": ["Maciej Pióro", "Kamil Ciebiera", "Krystian Król", "Jan Ludziejewski", "Michał Krutul", "Jakub Krajewski", "Szymon Antoniak", "Piotr Miłoś", "Marek Cygan", "Sebastian Jaszczur"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-08", "url": "https://arxiv.org/abs/2401.04081", "pdf_url": "https://arxiv.org/pdf/2401.04081v2", "arxiv_id": "2401.04081", "doi": "10.48550/arXiv.2401.04081", "citation_count": 97, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4978}
{"id": "10590fa10041f5b9dfcd5170653fdfa75fb8d9849d644369b769da63bb7e58d2", "sources": ["arxiv", "semantic_scholar"], "title": "Mixtral of Experts", "abstract": "We introduce Mixtral 8x7B, a Sparse Mixture of Experts (SMoE) language model. Mixtral has the same architecture as Mistral 7B, with the difference that each layer is composed of 8 feedforward blocks (i.e. experts). For every token, at each layer, a router network selects two experts to process the current state and combine their outputs. Even though each token only sees two experts, the selected experts can be different at each timestep. As a result, each token has access to 47B parameters, but only uses 13B active parameters during inference. Mixtral was trained with a context size of 32k tokens and it outperforms or matches Llama 2 70B and GPT-3.5 across all evaluated benchmarks. In particular, Mixtral vastly outperforms Llama 2 70B on mathematics, code generation, and multilingual benchmarks. We also provide a model fine-tuned to follow instructions, Mixtral 8x7B - Instruct, that surpasses GPT-3.5 Turbo, Claude-2.1, Gemini Pro, and Llama 2 70B - chat model on human benchmarks. Both the base and instruct models are released under the Apache 2.0 license.", "authors": ["Albert Q. Jiang", "Alexandre Sablayrolles", "Antoine Roux", "Arthur Mensch", "Blanche Savary", "Chris Bamford", "Devendra Singh Chaplot", "Diego de las Casas", "Emma Bou Hanna", "Florian Bressand", "Gianna Lengyel", "Guillaume Bour", "Guillaume Lample", "Lélio Renard Lavaud", "Lucile Saulnier", "Marie-Anne Lachaux", "Pierre Stock", "Sandeep Subramanian", "Sophia Yang", "Szymon Antoniak", "Teven Le Scao", "Théophile Gervet", "Thibaut Lavril", "Thomas Wang", "Timothée Lacroix", "William El Sayed"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-08", "url": "https://arxiv.org/abs/2401.04088", "pdf_url": "https://arxiv.org/pdf/2401.04088v1", "arxiv_id": "2401.04088", "doi": "10.48550/arXiv.2401.04088", "citation_count": 1942, "influential_citation_count": 188, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 1.0}
{"id": "20982f547081ceb69493e7bc0b1ec371d121695aee01d71dea081735f5b418b8", "sources": ["arxiv", "semantic_scholar"], "title": "Frequency-Adaptive Pan-Sharpening with Mixture of Experts", "abstract": "Pan-sharpening involves reconstructing missing high-frequency information in multi-spectral images with low spatial resolution, using a higher-resolution panchromatic image as guidance. Although the inborn connection with frequency domain, existing pan-sharpening research has not almost investigated the potential solution upon frequency domain. To this end, we propose a novel Frequency Adaptive Mixture of Experts (FAME) learning framework for pan-sharpening, which consists of three key components: the Adaptive Frequency Separation Prediction Module, the Sub-Frequency Learning Expert Module, and the Expert Mixture Module. In detail, the first leverages the discrete cosine transform to perform frequency separation by predicting the frequency mask. On the basis of generated mask, the second with low-frequency MOE and high-frequency MOE takes account for enabling the effective low-frequency and high-frequency information reconstruction. Followed by, the final fusion module dynamically weights high-frequency and low-frequency MOE knowledge to adapt to remote sensing images with significant content variations. Quantitative and qualitative experiments over multiple datasets demonstrate that our method performs the best against other state-of-the-art ones and comprises a strong generalization ability for real-world scenes. Code will be made publicly at \\url{https://github.com/alexhe101/FAME-Net}.", "authors": ["Xuanhua He", "Keyu Yan", "Rui Li", "Chengjun Xie", "Jie Zhang", "Man Zhou"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-04", "url": "https://arxiv.org/abs/2401.02151", "pdf_url": "https://arxiv.org/pdf/2401.02151v1", "arxiv_id": "2401.02151", "doi": "10.48550/arXiv.2401.02151", "citation_count": 44, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/alexhe101/FAME-Net}", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4133}
{"id": "c11eebe3531aeb14e311298b31ec0c3d94a5bedd349e454c9c492c91625178a5", "sources": ["arxiv", "semantic_scholar"], "title": "Fast Inference of Mixture-of-Experts Language Models with Offloading", "abstract": "With the widespread adoption of Large Language Models (LLMs), many deep learning practitioners are looking for strategies of running these models more efficiently. One such strategy is to use sparse Mixture-of-Experts (MoE) - a type of model architectures where only a fraction of model layers are active for any given input. This property allows MoE-based language models to generate tokens faster than their dense counterparts, but it also increases model size due to having multiple experts. Unfortunately, this makes state-of-the-art MoE language models difficult to run without high-end GPUs. In this work, we study the problem of running large MoE language models on consumer hardware with limited accelerator memory. We build upon parameter offloading algorithms and propose a novel strategy that accelerates offloading by taking advantage of innate properties of MoE LLMs. Using this strategy, we build can run Mixtral-8x7B with mixed quantization on desktop hardware and free-tier Google Colab instances.", "authors": ["Artyom Eliseev", "Denis Mazur"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-28", "url": "https://arxiv.org/abs/2312.17238", "pdf_url": "https://arxiv.org/pdf/2312.17238v1", "arxiv_id": "2312.17238", "doi": "10.48550/arXiv.2312.17238", "citation_count": 90, "influential_citation_count": 30, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7457}
{"id": "12be47a6c6cc6b23639c2474264ea9087a522874f3ae3f8d3048b76daeae18d4", "sources": ["arxiv", "semantic_scholar"], "title": "Crossed Products, Conditional Expectations and Constraint Quantization", "abstract": "Recent work has highlighted the importance of crossed products in correctly elucidating the operator algebraic approach to quantum field theories. In the gravitational context, the crossed product simultaneously promotes von Neumann algebras associated with subregions in diffeomorphism covariant quantum field theories from type III to type II, and provides the necessary ingredients to gravitationally dress operators, thereby enforcing the constraints of the theory. In this note we enhance the crossed product construction to the context of general gauge theories with arbitrary combinations of internal and spacetime local symmetries. This is done by leveraging the correspondence between the crossed product and the extended phase space. We then undertake a detailed study of constraint quantization from the perspective of a generic crossed product algebra. We study and compare four distinct approaches to constraint quantization from this point of view: refined algebraic quantization, BRST quantization, path integral quantization, and the commutation theorem for crossed products. Far from simply reproducing existing analyses, the operator algebraic viewpoint sheds new light on old problems by reformulating the dressing of operators in terms of conditional expectations and other closely related projection maps. We conclude by applying our approach to the constraint quantization of three distinct gauge theories including a discussion of gravity on null hypersurfaces.", "authors": ["Marc S. Klinger", "Robert G. Leigh"], "categories": ["hep-th"], "fields_of_study": ["Physics"], "published_date": "2023-12-27", "url": "https://arxiv.org/abs/2312.16678", "pdf_url": "https://arxiv.org/pdf/2312.16678v3", "arxiv_id": "2312.16678", "doi": "10.1016/j.nuclphysb.2024.116622", "citation_count": 22, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Nuclear Physics B", "quality_score": 0.3404}
{"id": "0ee409539185bc3c6051aca994e863c099e75fec6f750dfdbbe6f9fe08c6432a", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Deweather Mixture-of-Experts with Uncertainty-aware Feature-wise Linear Modulation", "abstract": "The Mixture-of-Experts (MoE) approach has demonstrated outstanding scalability in multi-task learning including low-level upstream tasks such as concurrent removal of multiple adverse weather effects. However, the conventional MoE architecture with parallel Feed Forward Network (FFN) experts leads to significant parameter and computational overheads that hinder its efficient deployment. In addition, the naive MoE linear router is suboptimal in assigning task-specific features to multiple experts which limits its further scalability. In this work, we propose an efficient MoE architecture with weight sharing across the experts. Inspired by the idea of linear feature modulation (FM), our architecture implicitly instantiates multiple experts via learnable activation modulations on a single shared expert block. The proposed Feature Modulated Expert (FME) serves as a building block for the novel Mixture-of-Feature-Modulation-Experts (MoFME) architecture, which can scale up the number of experts with low overhead. We further propose an Uncertainty-aware Router (UaR) to assign task-specific features to different FM modules with well-calibrated weights. This enables MoFME to effectively learn diverse expert functions for multiple tasks. The conducted experiments on the multi-deweather task show that our MoFME outperforms the baselines in the image restoration quality by 0.1-0.2 dB and achieves SOTA-compatible performance while saving more than 72% of parameters and 39% inference time over the conventional MoE counterpart. Experiments on the downstream segmentation and classification tasks further demonstrate the generalizability of MoFME to real open-world applications.", "authors": ["Rongyu Zhang", "Yulin Luo", "Jiaming Liu", "Huanrui Yang", "Zhen Dong", "Denis Gudovskiy", "Tomoyuki Okuno", "Yohei Nakata", "Kurt Keutzer", "Yuan Du", "Shanghang Zhang"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-27", "url": "https://arxiv.org/abs/2312.16610", "pdf_url": "https://arxiv.org/pdf/2312.16610v1", "arxiv_id": "2312.16610", "doi": "10.48550/arXiv.2312.16610", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "4b58442696e521d030069cddb0edd2de25ba0388c2227de82e92b05f6f01b131", "sources": ["arxiv", "semantic_scholar"], "title": "Hardware-Aware DNN Compression via Diverse Pruning and Mixed-Precision Quantization", "abstract": "Deep Neural Networks (DNNs) have shown significant advantages in a wide variety of domains. However, DNNs are becoming computationally intensive and energy hungry at an exponential pace, while at the same time, there is a vast demand for running sophisticated DNN-based services on resource constrained embedded devices. In this paper, we target energy-efficient inference on embedded DNN accelerators. To that end, we propose an automated framework to compress DNNs in a hardware-aware manner by jointly employing pruning and quantization. We explore, for the first time, per-layer fine- and coarse-grained pruning, in the same DNN architecture, in addition to low bit-width mixed-precision quantization for weights and activations. Reinforcement Learning (RL) is used to explore the associated design space and identify the pruning-quantization configuration so that the energy consumption is minimized whilst the prediction accuracy loss is retained at acceptable levels. Using our novel composite RL agent we are able to extract energy-efficient solutions without requiring retraining and/or fine tuning. Our extensive experimental evaluation over widely used DNNs and the CIFAR-10/100 and ImageNet datasets demonstrates that our framework achieves $39\\%$ average energy reduction for $1.7\\%$ average accuracy loss and outperforms significantly the state-of-the-art approaches.", "authors": ["Konstantinos Balaskas", "Andreas Karatzas", "Christos Sad", "Kostas Siozios", "Iraklis Anagnostopoulos", "Georgios Zervakis", "Jörg Henkel"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-23", "url": "https://arxiv.org/abs/2312.15322", "pdf_url": "https://arxiv.org/pdf/2312.15322v1", "arxiv_id": "2312.15322", "doi": "10.1109/TETC.2023.3346944", "citation_count": 33, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Emerging Topics in Computing", "quality_score": 0.3829}
{"id": "f4599c832ac1f3954e22bc39ef8f7c15bd47063d4b953b873c8aaea990f3935e", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Cluster-conditional LoRA Experts for Vision-language Instruction Tuning", "abstract": "Instruction tuning of Large Vision-language Models (LVLMs) has revolutionized the development of versatile models with zero-shot generalization across a wide range of downstream vision-language tasks. However, the diversity of training tasks of different sources and formats would lead to inevitable task conflicts, where different tasks conflict for the same set of model parameters, resulting in sub-optimal instruction-following abilities. To address that, we propose the Mixture of Cluster-conditional LoRA Experts (MoCLE), a novel Mixture of Experts (MoE) architecture designed to activate the task-customized model parameters based on the instruction clusters. A separate universal expert is further incorporated to improve generalization capabilities of MoCLE for novel instructions. Extensive experiments on InstructBLIP and LLaVA demonstrate the effectiveness of MoCLE.", "authors": ["Yunhao Gou", "Zhili Liu", "Kai Chen", "Lanqing Hong", "Hang Xu", "Aoxue Li", "Dit-Yan Yeung", "James T. Kwok", "Yu Zhang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2023-12-19", "url": "https://arxiv.org/abs/2312.12379", "pdf_url": "https://arxiv.org/pdf/2312.12379v5", "arxiv_id": "2312.12379", "doi": "10.1109/TIP.2026.3678087", "citation_count": 117, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Image Processing", "quality_score": 0.518}
{"id": "e73a9eada3b9340071d864c7b2b77c2f61bcddf83e66e25276094612d9f83c4f", "sources": ["arxiv", "semantic_scholar"], "title": "SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention", "abstract": "Despite many recent works on Mixture of Experts (MoEs) for resource-efficient Transformer language models, existing methods mostly focus on MoEs for feedforward layers. Previous attempts at extending MoE to the self-attention layer fail to match the performance of the parameter-matched baseline. Our novel SwitchHead is an effective MoE method for the attention layer that successfully reduces both the compute and memory requirements, achieving wall-clock speedup, while matching the language modeling performance of the baseline Transformer. Our novel MoE mechanism allows SwitchHead to compute up to 8 times fewer attention matrices than the standard Transformer. SwitchHead can also be combined with MoE feedforward layers, resulting in fully-MoE \"SwitchAll\" Transformers. For our 262M parameter model trained on C4, SwitchHead matches the perplexity of standard models with only 44% compute and 27% memory usage. Zero-shot experiments on downstream tasks confirm the performance of SwitchHead, e.g., achieving more than 3.5% absolute improvements on BliMP compared to the baseline with an equal compute resource.", "authors": ["Róbert Csordás", "Piotr Piękos", "Kazuki Irie", "Jürgen Schmidhuber"], "categories": ["cs.LG", "cs.CL", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-13", "url": "https://arxiv.org/abs/2312.07987", "pdf_url": "https://arxiv.org/pdf/2312.07987v3", "arxiv_id": "2312.07987", "doi": "10.48550/arXiv.2312.07987", "citation_count": 38, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3978}
{"id": "f6fe8ddc8cfd882d45dd9dc360503a81864ff97ce73acd50ab95906091f96dec", "sources": ["arxiv", "semantic_scholar"], "title": "Training of Neural Networks with Uncertain Data: A Mixture of Experts Approach", "abstract": "This paper introduces the \"Uncertainty-aware Mixture of Experts\" (uMoE), a novel solution aimed at addressing aleatoric uncertainty within Neural Network (NN) based predictive models. While existing methodologies primarily concentrate on managing uncertainty during inference, uMoE uniquely embeds uncertainty into the training phase. Employing a \"Divide and Conquer\" strategy, uMoE strategically partitions the uncertain input space into more manageable subspaces. It comprises Expert components, individually trained on their respective subspace uncertainties. Overarching the Experts, a Gating Unit, leveraging additional information regarding the distribution of uncertain in-puts across these subspaces, dynamically adjusts the weighting to minimize deviations from ground truth. Our findings demonstrate the superior performance of uMoE over baseline methods in effectively managing data uncertainty. Furthermore, through a comprehensive robustness analysis, we showcase its adaptability to varying uncertainty levels and propose optimal threshold parameters. This innovative approach boasts broad applicability across diverse da-ta-driven domains, including but not limited to biomedical signal processing, autonomous driving, and production quality control.", "authors": ["Lucas Luttner"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-12-13", "url": "https://arxiv.org/abs/2312.08083", "pdf_url": "https://arxiv.org/pdf/2312.08083v4", "arxiv_id": "2312.08083", "doi": "10.5281/zenodo.10050097", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "7f0bcbde324fa4f59cc67328e56cee937f1be0c45c253fbab0ae5652a743c429", "sources": ["arxiv", "semantic_scholar"], "title": "HyperRouter: Towards Efficient Training and Inference of Sparse Mixture of Experts", "abstract": "By routing input tokens to only a few split experts, Sparse Mixture-of-Experts has enabled efficient training of large language models. Recent findings suggest that fixing the routers can achieve competitive performance by alleviating the collapsing problem, where all experts eventually learn similar representations. However, this strategy has two key limitations: (i) the policy derived from random routers might be sub-optimal, and (ii) it requires extensive resources during training and evaluation, leading to limited efficiency gains. This work introduces \\HyperRout, which dynamically generates the router's parameters through a fixed hypernetwork and trainable embeddings to achieve a balance between training the routers and freezing them to learn an improved routing policy. Extensive experiments across a wide range of tasks demonstrate the superior performance and efficiency gains of \\HyperRouter compared to existing routing methods. Our implementation is publicly available at {\\url{https://github.com/giangdip2410/HyperRouter}}.", "authors": ["Giang Do", "Khiem Le", "Quang Pham", "TrungTin Nguyen", "Thanh-Nam Doan", "Bint T. Nguyen", "Chenghao Liu", "Savitha Ramasamy", "Xiaoli Li", "Steven Hoi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-12", "url": "https://arxiv.org/abs/2312.07035", "pdf_url": "https://arxiv.org/pdf/2312.07035v1", "arxiv_id": "2312.07035", "doi": "10.18653/v1/2023.emnlp-main.351", "citation_count": 26, "influential_citation_count": 6, "has_code": true, "code_url": "https://github.com/giangdip2410/HyperRouter}}", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4225}
{"id": "7d3533296b5f48311ed92bbaabffce5ffde0dd64045a4ea13632c84020775ae3", "sources": ["arxiv", "semantic_scholar"], "title": "Stoichiometry preservation and generalization of Bilger mixture fraction for non-premixed combustion with differential molecular diffusion", "abstract": "The Bilger mixture fraction is a widely used parameter in non-premixed combustion when considering differential molecular diffusion, a prevalent phenomenon in hydrogen or hydrogen-blended fuel combustion. The property of stoichiometry preservation of mixture fractions is investigated. Two different Bilger mixture fraction formulations are clarified. It is found that they belong to a class of one-parameter generalized mixture fraction definitions discovered in this work. Specific definitions from the class of mixture fractions are compared for hydrocarbon fuels. The comparison shows that the difference can be significant. An optimal mixture fraction definition is sought from the general definitions by minimizing its deviation from the desired properties. The obtained optimal mixture fractions show overall better preservation of stoichiometry than Bilger's definitions. The extension of the generalized mixture fraction to other fuels that contain nitrogen (like ammonia $\\mathrm{NH_3}$) or sulfur (like hydrogen sulfide $\\mathrm{H_2S}$) is also demonstrated.", "authors": ["Haifeng Wang", "Tianfang Xie"], "categories": ["physics.flu-dyn"], "fields_of_study": ["Physics"], "published_date": "2023-12-08", "url": "https://arxiv.org/abs/2312.05204", "pdf_url": "https://arxiv.org/pdf/2312.05204v2", "arxiv_id": "2312.05204", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "daa8a223eaef66cc9a9ceb2c06f422c4d722122e137e9e284f4b52ee0c9278e4", "sources": ["arxiv", "semantic_scholar"], "title": "GraphMETRO: Mitigating Complex Graph Distribution Shifts via Mixture of Aligned Experts", "abstract": "Graph data are inherently complex and heterogeneous, leading to a high natural diversity of distributional shifts. However, it remains unclear how to build machine learning architectures that generalize to the complex distributional shifts naturally occurring in the real world. Here, we develop GraphMETRO, a Graph Neural Network architecture that models natural diversity and captures complex distributional shifts. GraphMETRO employs a Mixture-of-Experts (MoE) architecture with a gating model and multiple expert models, where each expert model targets a specific distributional shift to produce a referential representation w.r.t. a reference model, and the gating model identifies shift components. Additionally, we design a novel objective that aligns the representations from different expert models to ensure reliable optimization. GraphMETRO achieves state-of-the-art results on four datasets from the GOOD benchmark, which is comprised of complex and natural real-world distribution shifts, improving by 67% and 4.2% on the WebKB and Twitch datasets. Code and data are available at https://github.com/Wuyxin/GraphMETRO.", "authors": ["Shirley Wu", "Kaidi Cao", "Bruno Ribeiro", "James Zou", "Jure Leskovec"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-07", "url": "https://arxiv.org/abs/2312.04693", "pdf_url": "https://arxiv.org/pdf/2312.04693v3", "arxiv_id": "2312.04693", "doi": "10.52202/079017-0297", "citation_count": 23, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/Wuyxin/GraphMETRO", "venue": "Neural Information Processing Systems", "quality_score": 0.3451}
{"id": "506ab863492261fd3cf56248f46c58950548b27cb506260fdf598efac1c09842", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Molecular Property Prediction via Mixture of Collaborative Experts", "abstract": "Molecular Property Prediction (MPP) task involves predicting biochemical properties based on molecular features, such as molecular graph structures, contributing to the discovery of lead compounds in drug development. To address data scarcity and imbalance in MPP, some studies have adopted Graph Neural Networks (GNN) as an encoder to extract commonalities from molecular graphs. However, these approaches often use a separate predictor for each task, neglecting the shared characteristics among predictors corresponding to different tasks. In response to this limitation, we introduce the GNN-MoCE architecture. It employs the Mixture of Collaborative Experts (MoCE) as predictors, exploiting task commonalities while confronting the homogeneity issue in the expert pool and the decision dominance dilemma within the expert group. To enhance expert diversity for collaboration among all experts, the Expert-Specific Projection method is proposed to assign a unique projection perspective to each expert. To balance decision-making influence for collaboration within the expert group, the Expert-Specific Loss is presented to integrate individual expert loss into the weighted decision loss of the group for more equitable training. Benefiting from the enhancements of MoCE in expert creation, dynamic expert group formation, and experts' collaboration, our model demonstrates superior performance over traditional methods on 24 MPP datasets, especially in tasks with limited data or high imbalance.", "authors": ["Xu Yao", "Shuang Liang", "Songqiao Han", "Hailiang Huang"], "categories": ["cs.LG", "cs.MA", "q-bio.QM"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2023-12-06", "url": "https://arxiv.org/abs/2312.03292", "pdf_url": "https://arxiv.org/pdf/2312.03292v1", "arxiv_id": "2312.03292", "doi": "10.48550/arXiv.2312.03292", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "06cba2c0250ab8d01408bfc4e1f813169c26a7810757433f9d2acb618ba92acf", "sources": ["arxiv", "semantic_scholar"], "title": "Run LoRA Run: Faster and Lighter LoRA Implementations", "abstract": "LoRA is a technique that reduces the number of trainable parameters in a neural network by introducing low-rank adapters to linear layers. This technique is used both for fine-tuning and full training of large language models. This paper presents the RunLoRA framework for efficient implementations of LoRA that significantly improves the speed of neural network training and fine-tuning using low-rank adapters. The proposed implementation optimizes the computation of LoRA operations based on dimensions of corresponding linear layer, layer input dimensions and lora rank by choosing best forward and backward computation graph based on FLOPs and time estimations, resulting in faster training without sacrificing accuracy. The experimental results show up to 28\\% speedup on language modeling networks.", "authors": ["Daria Cherniuk", "Aleksandr Mikhalev", "Ivan Oseledets"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-06", "url": "https://arxiv.org/abs/2312.03415", "pdf_url": "https://arxiv.org/pdf/2312.03415v2", "arxiv_id": "2312.03415", "doi": "10.48550/arXiv.2312.03415", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1193}
{"id": "fdd0293fb2399f48f7b3124199464b57fcb91a99e7361353a2ebd506eef6c315", "sources": ["arxiv", "semantic_scholar"], "title": "MoEC: Mixture of Experts Implicit Neural Compression", "abstract": "Emerging Implicit Neural Representation (INR) is a promising data compression technique, which represents the data using the parameters of a Deep Neural Network (DNN). Existing methods manually partition a complex scene into local regions and overfit the INRs into those regions. However, manually designing the partition scheme for a complex scene is very challenging and fails to jointly learn the partition and INRs. To solve the problem, we propose MoEC, a novel implicit neural compression method based on the theory of mixture of experts. Specifically, we use a gating network to automatically assign a specific INR to a 3D point in the scene. The gating network is trained jointly with the INRs of different local regions. Compared with block-wise and tree-structured partitions, our learnable partition can adaptively find the optimal partition in an end-to-end manner. We conduct detailed experiments on massive and diverse biomedical data to demonstrate the advantages of MoEC against existing approaches. In most of experiment settings, we have achieved state-of-the-art results. Especially in cases of extreme compression ratios, such as 6000x, we are able to uphold the PSNR of 48.16.", "authors": ["Jianchen Zhao", "Cheng-Ching Tseng", "Ming Lu", "Ruichuan An", "Xiaobao Wei", "He Sun", "Shanghang Zhang"], "categories": ["cs.CV", "cs.LG", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-12-03", "url": "https://arxiv.org/abs/2312.01361", "pdf_url": "https://arxiv.org/pdf/2312.01361v1", "arxiv_id": "2312.01361", "doi": "10.48550/arXiv.2312.01361", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "c18fc9a9a621e5c0e7d93555fb3cdf211146b5f4ab61846ae06d3527216c3319", "sources": ["arxiv", "semantic_scholar"], "title": "Physics Inspired Criterion for Pruning-Quantization Joint Learning", "abstract": "Pruning-quantization joint learning always facilitates the deployment of deep neural networks (DNNs) on resource-constrained edge devices. However, most existing methods do not jointly learn a global criterion for pruning and quantization in an interpretable way. In this paper, we propose a novel physics inspired criterion for pruning-quantization joint learning (PIC-PQ), which is explored from an analogy we first draw between elasticity dynamics (ED) and model compression (MC). Specifically, derived from Hooke's law in ED, we establish a linear relationship between the filters' importance distribution and the filter property (FP) by a learnable deformation scale in the physics inspired criterion (PIC). Furthermore, we extend PIC with a relative shift variable for a global view. To ensure feasibility and flexibility, available maximum bitwidth and penalty factor are introduced in quantization bitwidth assignment. Experiments on benchmarks of image classification demonstrate that PIC-PQ yields a good trade-off between accuracy and bit-operations (BOPs) compression ratio e.g., 54.96X BOPs compression ratio in ResNet56 on CIFAR10 with 0.10% accuracy drop and 53.24X in ResNet18 on ImageNet with 0.61% accuracy drop). The code will be available at https://github.com/fanxxxxyi/PIC-PQ.", "authors": ["Weiying Xie", "Xiaoyi Fan", "Xin Zhang", "Yunsong Li", "Jie Lei", "Leyuan Fang"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-01", "url": "https://arxiv.org/abs/2312.00851", "pdf_url": "https://arxiv.org/pdf/2312.00851v2", "arxiv_id": "2312.00851", "doi": "10.48550/arXiv.2312.00851", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/fanxxxxyi/PIC-PQ", "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "3d4ec33029b8cf20b56a268f47a0dc89e119429fc9c9cb56f0f012007eefe785", "sources": ["arxiv", "semantic_scholar"], "title": "Omni-SMoLA: Boosting Generalist Multimodal Models with Soft Mixture of Low-rank Experts", "abstract": "Large multi-modal models (LMMs) exhibit remarkable performance across numerous tasks. However, generalist LMMs often suffer from performance degradation when tuned over a large collection of tasks. Recent research suggests that Mixture of Experts (MoE) architectures are useful for instruction tuning, but for LMMs of parameter size around O(50-100B), the prohibitive cost of replicating and storing the expert models severely limits the number of experts we can use. We propose Omni-SMoLA, an architecture that uses the Soft MoE approach to (softly) mix many multimodal low rank experts, and avoids introducing a significant number of new parameters compared to conventional MoE models. The core intuition here is that the large model provides a foundational backbone, while different lightweight experts residually learn specialized knowledge, either per-modality or multimodally. Extensive experiments demonstrate that the SMoLA approach helps improve the generalist performance across a broad range of generative vision-and-language tasks, achieving new SoTA generalist performance that often matches or outperforms single specialized LMM baselines, as well as new SoTA specialist performance.", "authors": ["Jialin Wu", "Xia Hu", "Yaqing Wang", "Bo Pang", "Radu Soricut"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-01", "url": "https://arxiv.org/abs/2312.00968", "pdf_url": "https://arxiv.org/pdf/2312.00968v2", "arxiv_id": "2312.00968", "doi": "10.1109/CVPR52733.2024.01347", "citation_count": 37, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.3949}
{"id": "7264502b4db06e88186f18dd76311211020afb4fde34a93922b2c7363c3bc73a", "sources": ["arxiv", "semantic_scholar"], "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces", "abstract": "Foundation models, now powering most of the exciting applications in deep learning, are almost universally based on the Transformer architecture and its core attention module. Many subquadratic-time architectures such as linear attention, gated convolution and recurrent models, and structured state space models (SSMs) have been developed to address Transformers' computational inefficiency on long sequences, but they have not performed as well as attention on important modalities such as language. We identify that a key weakness of such models is their inability to perform content-based reasoning, and make several improvements. First, simply letting the SSM parameters be functions of the input addresses their weakness with discrete modalities, allowing the model to selectively propagate or forget information along the sequence length dimension depending on the current token. Second, even though this change prevents the use of efficient convolutions, we design a hardware-aware parallel algorithm in recurrent mode. We integrate these selective SSMs into a simplified end-to-end neural network architecture without attention or even MLP blocks (Mamba). Mamba enjoys fast inference (5$\\times$ higher throughput than Transformers) and linear scaling in sequence length, and its performance improves on real data up to million-length sequences. As a general sequence model backbone, Mamba achieves state-of-the-art performance across several modalities such as language, audio, and genomics. On language modeling, our Mamba-3B model outperforms Transformers of the same size and matches Transformers twice its size, both in pretraining and downstream evaluation.", "authors": ["Albert Gu", "Tri Dao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-01", "url": "https://arxiv.org/abs/2312.00752", "pdf_url": "https://arxiv.org/pdf/2312.00752v2", "arxiv_id": "2312.00752", "doi": null, "citation_count": 7448, "influential_citation_count": 1147, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 1.0}
{"id": "467a68a09dd48184b3d86414e8d5a4b9ef9236fe73055e139e174e441704b499", "sources": ["arxiv", "semantic_scholar"], "title": "LQ-LoRA: Low-rank Plus Quantized Matrix Decomposition for Efficient Language Model Finetuning", "abstract": "We propose a simple approach for memory-efficient adaptation of pretrained language models. Our approach uses an iterative algorithm to decompose each pretrained matrix into a high-precision low-rank component and a memory-efficient quantized component. During finetuning, the quantized component remains fixed and only the low-rank component is updated. We present an integer linear programming formulation of the quantization component which enables dynamic configuration of quantization parameters (e.g., bit-width, block size) for each matrix given an overall target memory budget. We further explore a data-aware version of the algorithm which uses an approximation of the Fisher information matrix to weight the reconstruction objective during matrix decomposition. Experiments on finetuning RoBERTa and LLaMA-2 (7B and 70B) demonstrate that our low-rank plus quantized matrix decomposition approach (LQ-LoRA) outperforms strong QLoRA and GPTQ-LoRA baselines and enables aggressive quantization to sub-3 bits with only minor performance degradations. When finetuned on a language modeling calibration dataset, LQ-LoRA can also be used for model compression; in this setting our 2.75-bit LLaMA-2-70B model (which has 2.85 bits on average when including the low-rank components and requires 27GB of GPU memory) performs respectably compared to the 16-bit baseline.", "authors": ["Han Guo", "Philip Greengard", "Eric P. Xing", "Yoon Kim"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-20", "url": "https://arxiv.org/abs/2311.12023", "pdf_url": "https://arxiv.org/pdf/2311.12023v4", "arxiv_id": "2311.12023", "doi": "10.48550/arXiv.2311.12023", "citation_count": 98, "influential_citation_count": 13, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.5731}
{"id": "3687ce032cbd56b457fb00a400758f427a41c4a134b341da142935b72d23ebeb", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts", "abstract": "Multi-Task Reinforcement Learning (MTRL) tackles the long-standing problem of endowing agents with skills that generalize across a variety of problems. To this end, sharing representations plays a fundamental role in capturing both unique and common characteristics of the tasks. Tasks may exhibit similarities in terms of skills, objects, or physical properties while leveraging their representations eases the achievement of a universal policy. Nevertheless, the pursuit of learning a shared set of diverse representations is still an open challenge. In this paper, we introduce a novel approach for representation learning in MTRL that encapsulates common structures among the tasks using orthogonal representations to promote diversity. Our method, named Mixture Of Orthogonal Experts (MOORE), leverages a Gram-Schmidt process to shape a shared subspace of representations generated by a mixture of experts. When task-specific information is provided, MOORE generates relevant representations from this shared subspace. We assess the effectiveness of our approach on two MTRL benchmarks, namely MiniGrid and MetaWorld, showing that MOORE surpasses related baselines and establishes a new state-of-the-art result on MetaWorld.", "authors": ["Ahmed Hendawy", "Jan Peters", "Carlo D'Eramo"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-19", "url": "https://arxiv.org/abs/2311.11385", "pdf_url": "https://arxiv.org/pdf/2311.11385v2", "arxiv_id": "2311.11385", "doi": "10.48550/arXiv.2311.11385", "citation_count": 50, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4771}
{"id": "7c4ae9f7b35711e9d46e1fec6a3cf3f095b5f50fcb36eaeddc54a595ccc247b9", "sources": ["arxiv", "semantic_scholar"], "title": "Token-Level Adaptation of LoRA Adapters for Downstream Task Generalization", "abstract": "This paper introduces a method for adapting LoRA adapters in smaller-sized language models to arbitrary downstream tasks. Unlike standard mixture-of-expert architectures, our method employs a gradient-free routing function to choose a weighted combination of experts without increasing the compute requirements for training or inference. The results show that token-level adaptation of LoRA adapters outperforms the base Llama-2-7b model across mathematical (GSM8K), scientific (ARC-Challenge), reading comprehension (SQuAD), and coding (CodeAlpaca-20k) tasks. Further evaluations also show that the average performance of token-level adaptation outperforms individual models fine-tuned for each of the tasks with the best performance observed in adaptation of every-other token during inference. The code for this study is made available through a public repository.", "authors": ["Joshua Belofsky"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-17", "url": "https://arxiv.org/abs/2311.10847", "pdf_url": "https://arxiv.org/pdf/2311.10847v2", "arxiv_id": "2311.10847", "doi": "10.1145/3639592.3639615", "citation_count": 18, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Artificial Intelligence and Cloud Computing Conference", "quality_score": 0.3197}
{"id": "46aec8aa3ebe17b26ada335cd57e22ca0a358fb0bc0f8da5cb8654ec9bf6bdf6", "sources": ["arxiv", "semantic_scholar"], "title": "On the Impact of Calibration Data in Post-training Quantization and Pruning", "abstract": "Quantization and pruning form the foundation of compression for neural networks, enabling efficient inference for large language models (LLMs). Recently, various quantization and pruning techniques have demonstrated remarkable performance in a post-training setting. They rely upon calibration data, a small set of unlabeled examples that are used to generate layer activations. However, no prior work has systematically investigated how the calibration data impacts the effectiveness of model compression methods. In this paper, we present the first extensive empirical study on the effect of calibration data upon LLM performance. We trial a variety of quantization and pruning methods, datasets, tasks, and models. Surprisingly, we find substantial variations in downstream task performance, contrasting existing work that suggests a greater level of robustness to the calibration data. Finally, we make a series of recommendations for the effective use of calibration data in LLM quantization and pruning.", "authors": ["Miles Williams", "Nikolaos Aletras"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-16", "url": "https://arxiv.org/abs/2311.09755", "pdf_url": "https://arxiv.org/pdf/2311.09755v2", "arxiv_id": "2311.09755", "doi": "10.18653/v1/2024.acl-long.544", "citation_count": 51, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.429}
{"id": "a0d8f5ca2794a9aeb0b0fe46c00cd1862e641e2cee9c5b8582b17c5e358113d5", "sources": ["arxiv", "semantic_scholar"], "title": "Tied-Lora: Enhancing parameter efficiency of LoRA with weight tying", "abstract": "We introduce Tied-LoRA, a novel paradigm leveraging weight tying and selective training to enhance the parameter efficiency of Low-rank Adaptation (LoRA). Our exploration encompasses different plausible combinations of parameter training and freezing, coupled with weight tying, aimed at identifying the optimal trade-off between performance and the count of trainable parameters. Across $5$ diverse tasks and two foundational language models with different parameter counts, our experiments provide comprehensive insights into the inherent trade-offs between efficiency and performance. Our findings reveal a specific Tied-LoRA configuration that distinguishes itself by showcasing comparable performance to LoRA across multiple tasks while utilizing only a fraction of the parameters employed by the standard LoRA method, particularly at elevated ranks. This underscores the efficacy of Tied-LoRA in achieving impressive results with significantly reduced model complexity.", "authors": ["Adithya Renduchintala", "Tugrul Konuk", "Oleksii Kuchaiev"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-16", "url": "https://arxiv.org/abs/2311.09578", "pdf_url": "https://arxiv.org/pdf/2311.09578v2", "arxiv_id": "2311.09578", "doi": "10.18653/v1/2024.naacl-long.481", "citation_count": 81, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.4785}
{"id": "14e3febb9c49f09a6d7c20f2a9af88e6424d757a3c49a4937d5e84ab58f56e4b", "sources": ["arxiv", "semantic_scholar"], "title": "Memory Augmented Language Models through Mixture of Word Experts", "abstract": "Scaling up the number of parameters of language models has proven to be an effective approach to improve performance. For dense models, increasing model size proportionally increases the model's computation footprint. In this work, we seek to aggressively decouple learning capacity and FLOPs through Mixture-of-Experts (MoE) style models with large knowledge-rich vocabulary based routing functions and experts. Our proposed approach, dubbed Mixture of Word Experts (MoWE), can be seen as a memory augmented model, where a large set of word-specific experts play the role of a sparse memory. We demonstrate that MoWE performs significantly better than the T5 family of models with similar number of FLOPs in a variety of NLP tasks. Additionally, MoWE outperforms regular MoE models on knowledge intensive tasks and has similar performance to more complex memory augmented approaches that often require to invoke custom mechanisms to search the sparse memory.", "authors": ["Cicero Nogueira dos Santos", "James Lee-Thorp", "Isaac Noble", "Chung-Ching Chang", "David Uthus"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-15", "url": "https://arxiv.org/abs/2311.10768", "pdf_url": "https://arxiv.org/pdf/2311.10768v1", "arxiv_id": "2311.10768", "doi": "10.48550/arXiv.2311.10768", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.2785}
{"id": "46b7bafc4912f0312e2298a5fd66bce5cabe42a129a2cf7bcf677768d8580270", "sources": ["arxiv", "semantic_scholar"], "title": "SiRA: Sparse Mixture of Low Rank Adaptation", "abstract": "Parameter Efficient Tuning has been an prominent approach to adapt the Large Language Model to downstream tasks. Most previous works considers adding the dense trainable parameters, where all parameters are used to adapt certain task. We found this less effective empirically using the example of LoRA that introducing more trainable parameters does not help. Motivated by this we investigate the importance of leveraging \"sparse\" computation and propose SiRA: sparse mixture of low rank adaption. SiRA leverages the Sparse Mixture of Expert(SMoE) to boost the performance of LoRA. Specifically it enforces the top $k$ experts routing with a capacity limit restricting the maximum number of tokens each expert can process. We propose a novel and simple expert dropout on top of gating network to reduce the over-fitting issue. Through extensive experiments, we verify SiRA performs better than LoRA and other mixture of expert approaches across different single tasks and multitask settings.", "authors": ["Yun Zhu", "Nevan Wichers", "Chu-Cheng Lin", "Xinyi Wang", "Tianlong Chen", "Lei Shu", "Han Lu", "Canoee Liu", "Liangchen Luo", "Jindong Chen", "Lei Meng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-15", "url": "https://arxiv.org/abs/2311.09179", "pdf_url": "https://arxiv.org/pdf/2311.09179v1", "arxiv_id": "2311.09179", "doi": "10.48550/arXiv.2311.09179", "citation_count": 40, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4032}
{"id": "495c91f8c594c0b739ac46c60471d1c1e98d868434477640333f2a194a2fda4b", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Weak & Strong Experts on Graphs", "abstract": "Realistic graphs contain both (1) rich self-features of nodes and (2) informative structures of neighborhoods, jointly handled by a Graph Neural Network (GNN) in the typical setup. We propose to decouple the two modalities by Mixture of weak and strong experts (Mowst), where the weak expert is a light-weight Multi-layer Perceptron (MLP), and the strong expert is an off-the-shelf GNN. To adapt the experts' collaboration to different target nodes, we propose a \"confidence\" mechanism based on the dispersion of the weak expert's prediction logits. The strong expert is conditionally activated in the low-confidence region when either the node's classification relies on neighborhood information, or the weak expert has low model quality. We reveal interesting training dynamics by analyzing the influence of the confidence function on loss: our training algorithm encourages the specialization of each expert by effectively generating soft splitting of the graph. In addition, our \"confidence\" design imposes a desirable bias toward the strong expert to benefit from GNN's better generalization capability. Mowst is easy to optimize and achieves strong expressive power, with a computation cost comparable to a single GNN. Empirically, Mowst on 4 backbone GNN architectures show significant accuracy improvement on 6 standard node classification benchmarks, including both homophilous and heterophilous graphs (https://github.com/facebookresearch/mowst-gnn).", "authors": ["Hanqing Zeng", "Hanjia Lyu", "Diyi Hu", "Yinglong Xia", "Jiebo Luo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-09", "url": "https://arxiv.org/abs/2311.05185", "pdf_url": "https://arxiv.org/pdf/2311.05185v2", "arxiv_id": "2311.05185", "doi": "10.48550/arXiv.2311.05185", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/facebookresearch/mowst-gnn", "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "005924600d1d561a41e58f127eb55b2a4c77af234b6386ae573eba5a025ee598", "sources": ["arxiv", "semantic_scholar"], "title": "Leveraging Speculative Sampling and KV-Cache Optimizations Together for Generative AI using OpenVINO", "abstract": "Inference optimizations are critical for improving user experience and reducing infrastructure costs and power consumption. In this article, we illustrate a form of dynamic execution known as speculative sampling to reduce the overall latency of text generation and compare it with standard autoregressive sampling. This can be used together with model-based optimizations (e.g. quantization) to provide an optimized solution. Both sampling methods make use of KV caching. A Jupyter notebook and some sample executions are provided.", "authors": ["Haim Barad", "Ekaterina Aidova", "Yury Gorbachev"], "categories": ["cs.LG", "cs.AI", "cs.PF"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-08", "url": "https://arxiv.org/abs/2311.04951", "pdf_url": "https://arxiv.org/pdf/2311.04951v2", "arxiv_id": "2311.04951", "doi": "10.48550/arXiv.2311.04951", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/openvinotoolkit/openvino_notebooks/tree/latest/notebooks/speculative-sampling", "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "6ebdb86eb82c354814b9042a3546b3a9c3e54869fc999e7d035f9b56292d7327", "sources": ["arxiv", "semantic_scholar"], "title": "DAMEX: Dataset-aware Mixture-of-Experts for visual understanding of mixture-of-datasets", "abstract": "Construction of a universal detector poses a crucial question: How can we most effectively train a model on a large mixture of datasets? The answer lies in learning dataset-specific features and ensembling their knowledge but do all this in a single model. Previous methods achieve this by having separate detection heads on a common backbone but that results in a significant increase in parameters. In this work, we present Mixture-of-Experts as a solution, highlighting that MoEs are much more than a scalability tool. We propose Dataset-Aware Mixture-of-Experts, DAMEX where we train the experts to become an `expert' of a dataset by learning to route each dataset tokens to its mapped expert. Experiments on Universal Object-Detection Benchmark show that we outperform the existing state-of-the-art by average +10.2 AP score and improve over our non-MoE baseline by average +2.0 AP score. We also observe consistent gains while mixing datasets with (1) limited availability, (2) disparate domains and (3) divergent label sets. Further, we qualitatively show that DAMEX is robust against expert representation collapse.", "authors": ["Yash Jain", "Harkirat Behl", "Zsolt Kira", "Vibhav Vineet"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-08", "url": "https://arxiv.org/abs/2311.04894", "pdf_url": "https://arxiv.org/pdf/2311.04894v1", "arxiv_id": "2311.04894", "doi": "10.48550/arXiv.2311.04894", "citation_count": 39, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/jinga-lala/DAMEX", "venue": "Neural Information Processing Systems", "quality_score": 0.4005}
{"id": "eef765fe3d687f820f9ebb65824709aa4009d3d80fc1507f319da53eae3d1c67", "sources": ["arxiv", "semantic_scholar"], "title": "S-LoRA: Serving Thousands of Concurrent LoRA Adapters", "abstract": "The \"pretrain-then-finetune\" paradigm is commonly adopted in the deployment of large language models. Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, is often employed to adapt a base model to a multitude of tasks, resulting in a substantial collection of LoRA adapters derived from one base model. We observe that this paradigm presents significant opportunities for batched inference during serving. To capitalize on these opportunities, we present S-LoRA, a system designed for the scalable serving of many LoRA adapters. S-LoRA stores all adapters in the main memory and fetches the adapters used by the currently running queries to the GPU memory. To efficiently use the GPU memory and reduce fragmentation, S-LoRA proposes Unified Paging. Unified Paging uses a unified memory pool to manage dynamic adapter weights with different ranks and KV cache tensors with varying sequence lengths. Additionally, S-LoRA employs a novel tensor parallelism strategy and highly optimized custom CUDA kernels for heterogeneous batching of LoRA computation. Collectively, these features enable S-LoRA to serve thousands of LoRA adapters on a single GPU or across multiple GPUs with a small overhead. Compared to state-of-the-art libraries such as HuggingFace PEFT and vLLM (with naive support of LoRA serving), S-LoRA can improve the throughput by up to 4 times and increase the number of served adapters by several orders of magnitude. As a result, S-LoRA enables scalable serving of many task-specific fine-tuned models and offers the potential for large-scale customized fine-tuning services. The code is available at https://github.com/S-LoRA/S-LoRA", "authors": ["Ying Sheng", "Shiyi Cao", "Dacheng Li", "Coleman Hooper", "Nicholas Lee", "Shuo Yang", "Christopher Chou", "Banghua Zhu", "Lianmin Zheng", "Kurt Keutzer", "Joseph E. Gonzalez", "Ion Stoica"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-06", "url": "https://arxiv.org/abs/2311.03285", "pdf_url": "https://arxiv.org/pdf/2311.03285v3", "arxiv_id": "2311.03285", "doi": "10.48550/arXiv.2311.03285", "citation_count": 189, "influential_citation_count": 18, "has_code": true, "code_url": "https://github.com/S-LoRA/S-LoRA", "venue": "arXiv.org", "quality_score": 0.6394}
{"id": "f859ef87c4aedcb3744d0e41072bd649bde9618871c153f1a38846870ae6ccf2", "sources": ["arxiv", "semantic_scholar"], "title": "Cross-Image Attention for Zero-Shot Appearance Transfer", "abstract": "Recent advancements in text-to-image generative models have demonstrated a remarkable ability to capture a deep semantic understanding of images. In this work, we leverage this semantic knowledge to transfer the visual appearance between objects that share similar semantics but may differ significantly in shape. To achieve this, we build upon the self-attention layers of these generative models and introduce a cross-image attention mechanism that implicitly establishes semantic correspondences across images. Specifically, given a pair of images -- one depicting the target structure and the other specifying the desired appearance -- our cross-image attention combines the queries corresponding to the structure image with the keys and values of the appearance image. This operation, when applied during the denoising process, leverages the established semantic correspondences to generate an image combining the desired structure and appearance. In addition, to improve the output image quality, we harness three mechanisms that either manipulate the noisy latent codes or the model's internal representations throughout the denoising process. Importantly, our approach is zero-shot, requiring no optimization or training. Experiments show that our method is effective across a wide range of object categories and is robust to variations in shape, size, and viewpoint between the two input images.", "authors": ["Yuval Alaluf", "Daniel Garibi", "Or Patashnik", "Hadar Averbuch-Elor", "Daniel Cohen-Or"], "categories": ["cs.CV", "cs.GR"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-06", "url": "https://arxiv.org/abs/2311.03335", "pdf_url": "https://arxiv.org/pdf/2311.03335v1", "arxiv_id": "2311.03335", "doi": "10.1145/3641519.3657423", "citation_count": 166, "influential_citation_count": 34, "has_code": false, "code_url": null, "venue": "International Conference on Computer Graphics and Interactive Techniques", "quality_score": 0.772}
{"id": "15a71b9a64c8cb13ddd0a8c1ebdd578384f585f70083806993fe86f33c54785a", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Linguistic-Experts Adapters for Improving and Interpreting Pre-trained Language Models", "abstract": "In this work, we propose a method that combines two popular research areas by injecting linguistic structures into pre-trained language models in the parameter-efficient fine-tuning (PEFT) setting. In our approach, parallel adapter modules encoding different linguistic structures are combined using a novel Mixture-of-Linguistic-Experts architecture, where Gumbel-Softmax gates are used to determine the importance of these modules at each layer of the model. To reduce the number of parameters, we first train the model for a fixed small number of steps before pruning the experts based on their importance scores. Our experiment results with three different pre-trained models show that our approach can outperform state-of-the-art PEFT methods with a comparable number of parameters. In addition, we provide additional analysis to examine the experts selected by each model at each layer to provide insights for future studies.", "authors": ["Raymond Li", "Gabriel Murray", "Giuseppe Carenini"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-24", "url": "https://arxiv.org/abs/2310.16240", "pdf_url": "https://arxiv.org/pdf/2310.16240v1", "arxiv_id": "2310.16240", "doi": "10.48550/arXiv.2310.16240", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.1945}
{"id": "294e58d38be889534834ce0916634682e2682811146b8db8f190a783119e236c", "sources": ["arxiv", "semantic_scholar"], "title": "A General Theory for Softmax Gating Multinomial Logistic Mixture of Experts", "abstract": "Mixture-of-experts (MoE) model incorporates the power of multiple submodels via gating functions to achieve greater performance in numerous regression and classification applications. From a theoretical perspective, while there have been previous attempts to comprehend the behavior of that model under the regression settings through the convergence analysis of maximum likelihood estimation in the Gaussian MoE model, such analysis under the setting of a classification problem has remained missing in the literature. We close this gap by establishing the convergence rates of density estimation and parameter estimation in the softmax gating multinomial logistic MoE model. Notably, when part of the expert parameters vanish, these rates are shown to be slower than polynomial rates owing to an inherent interaction between the softmax gating and expert functions via partial differential equations. To address this issue, we propose using a novel class of modified softmax gating functions which transform the input before delivering them to the gating functions. As a result, the previous interaction disappears and the parameter estimation rates are significantly improved.", "authors": ["Huy Nguyen", "Pedram Akbarian", "TrungTin Nguyen", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-10-22", "url": "https://arxiv.org/abs/2310.14188", "pdf_url": "https://arxiv.org/pdf/2310.14188v2", "arxiv_id": "2310.14188", "doi": "10.48550/arXiv.2310.14188", "citation_count": 29, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3693}
{"id": "19997a0c2f918ccbf60b93b2835e75572a886b3f8cd6910f19ccf7ceab774fcd", "sources": ["arxiv", "semantic_scholar"], "title": "Merging Experts into One: Improving Computational Efficiency of Mixture of Experts", "abstract": "Scaling the size of language models usually leads to remarkable advancements in NLP tasks. But it often comes with a price of growing computational cost. Although a sparse Mixture of Experts (MoE) can reduce the cost by activating a small subset of parameters (e.g., one expert) for each input, its computation escalates significantly if increasing the number of activated experts, limiting its practical utility. Can we retain the advantages of adding more experts without substantially increasing the computational costs? In this paper, we first demonstrate the superiority of selecting multiple experts and then propose a computation-efficient approach called \\textbf{\\texttt{Merging Experts into One}} (MEO), which reduces the computation cost to that of a single expert. Extensive experiments show that MEO significantly improves computational efficiency, e.g., FLOPS drops from 72.0G of vanilla MoE to 28.6G (MEO). Moreover, we propose a token-level attention block that further enhances the efficiency and performance of token-level MEO, e.g., 83.3\\% (MEO) vs. 82.6\\% (vanilla MoE) average score on the GLUE benchmark. Our code will be released upon acceptance. Code will be released at: \\url{https://github.com/Shwai-He/MEO}.", "authors": ["Shwai He", "Run-Ze Fan", "Liang Ding", "Li Shen", "Tianyi Zhou", "Dacheng Tao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-15", "url": "https://arxiv.org/abs/2310.09832", "pdf_url": "https://arxiv.org/pdf/2310.09832v3", "arxiv_id": "2310.09832", "doi": "10.48550/arXiv.2310.09832", "citation_count": 43, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/Shwai-He/MEO}", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.4109}
{"id": "ba2378cef867c1f825dbac6bb56e5cd149b6aaf87c7244f77e478e48fd03f26e", "sources": ["arxiv", "semantic_scholar"], "title": "Diversifying the Mixture-of-Experts Representation for Language Models with Orthogonal Optimizer", "abstract": "The Mixture of Experts (MoE) has emerged as a highly successful technique in deep learning, based on the principle of divide-and-conquer to maximize model capacity without significant additional computational cost. Even in the era of large-scale language models (LLMs), MoE continues to play a crucial role, as some researchers have indicated that GPT-4 adopts the MoE structure to ensure diverse inference results. However, MoE is susceptible to performance degeneracy, particularly evident in the issues of imbalance and homogeneous representation among experts. While previous studies have extensively addressed the problem of imbalance, the challenge of homogeneous representation remains unresolved. In this study, we shed light on the homogeneous representation problem, wherein experts in the MoE fail to specialize and lack diversity, leading to frustratingly high similarities in their representations (up to 99\\% in a well-performed MoE model). This problem restricts the expressive power of the MoE and, we argue, contradicts its original intention. To tackle this issue, we propose a straightforward yet highly effective solution: OMoE, an orthogonal expert optimizer. Additionally, we introduce an alternating training strategy that encourages each expert to update in a direction orthogonal to the subspace spanned by other experts. Our algorithm facilitates MoE training in two key ways: firstly, it explicitly enhances representation diversity, and secondly, it implicitly fosters interaction between experts during orthogonal weights computation. Through extensive experiments, we demonstrate that our proposed optimization algorithm significantly improves the performance of fine-tuning the MoE model on the GLUE benchmark, SuperGLUE benchmark, question-answering task, and name entity recognition tasks.", "authors": ["Boan Liu", "Liang Ding", "Li Shen", "Keqin Peng", "Yu Cao", "Dazhao Cheng", "Dacheng Tao"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-15", "url": "https://arxiv.org/abs/2310.09762", "pdf_url": "https://arxiv.org/pdf/2310.09762v2", "arxiv_id": "2310.09762", "doi": "10.48550/arXiv.2310.09762", "citation_count": 29, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "European Conference on Artificial Intelligence", "quality_score": 0.3693}
{"id": "e21f6c7a33648a417d385f539c4a063a89f0e99400b874863bcf83cf4372c311", "sources": ["arxiv", "semantic_scholar"], "title": "LoftQ: LoRA-Fine-Tuning-Aware Quantization for Large Language Models", "abstract": "Quantization is an indispensable technique for serving Large Language Models (LLMs) and has recently found its way into LoRA fine-tuning. In this work we focus on the scenario where quantization and LoRA fine-tuning are applied together on a pre-trained model. In such cases it is common to observe a consistent gap in the performance on downstream tasks between full fine-tuning and quantization plus LoRA fine-tuning approach. In response, we propose LoftQ (LoRA-Fine-Tuning-aware Quantization), a novel quantization framework that simultaneously quantizes an LLM and finds a proper low-rank initialization for LoRA fine-tuning. Such an initialization alleviates the discrepancy between the quantized and full-precision model and significantly improves generalization in downstream tasks. We evaluate our method on natural language understanding, question answering, summarization, and natural language generation tasks. Experiments show that our method is highly effective and outperforms existing quantization methods, especially in the challenging 2-bit and 2/4-bit mixed precision regimes. The code is available on https://github.com/yxli2123/LoftQ.", "authors": ["Yixiao Li", "Yifan Yu", "Chen Liang", "Pengcheng He", "Nikos Karampatziakis", "Weizhu Chen", "Tuo Zhao"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-12", "url": "https://arxiv.org/abs/2310.08659", "pdf_url": "https://arxiv.org/pdf/2310.08659v4", "arxiv_id": "2310.08659", "doi": "10.48550/arXiv.2310.08659", "citation_count": 226, "influential_citation_count": 27, "has_code": true, "code_url": "https://github.com/yxli2123/LoftQ", "venue": "International Conference on Learning Representations", "quality_score": 0.7236}
{"id": "f9248d78db8d5c5aeb701b864bf5a6a89a2eab14eab77093fd8c803e088bb74f", "sources": ["arxiv", "semantic_scholar"], "title": "CacheGen: KV Cache Compression and Streaming for Fast Large Language Model Serving", "abstract": "As large language models (LLMs) take on complex tasks, their inputs are supplemented with longer contexts that incorporate domain knowledge. Yet using long contexts is challenging, as nothing can be generated until the whole context is processed by the LLM. While the context-processing delay can be reduced by reusing the KV cache of a context across different inputs, fetching the KV cache, which contains large tensors, over the network can cause high extra network delays. CacheGen is a fast context-loading module for LLM systems. First, CacheGen uses a custom tensor encoder, leveraging KV cache's distributional properties to encode a KV cache into more compact bitstream representations with negligible decoding overhead, to save bandwidth usage. Second, CacheGen adapts the compression level of different parts of a KV cache to cope with changes in available bandwidth, in order to maintain low context-loading delay and high generation quality. % When available bandwidth drops, CacheGen may raise the compression level for a part of the context or recompute its KV cache on the fly. We test CacheGen on popular LLMs and datasets. Compared to the recent systems that reuse the KV cache, CacheGen reduces the KV cache size by 3.5-4.3x and the total delay in fetching and processing contexts by 3.2-3.7x with negligible impact on the LLM response quality. Our code is at: https://github.com/UChi-JCL/CacheGen.", "authors": ["Yuhan Liu", "Hanchen Li", "Yihua Cheng", "Siddhant Ray", "Yuyang Huang", "Qizheng Zhang", "Kuntai Du", "Jiayi Yao", "Shan Lu", "Ganesh Ananthanarayanan", "Michael Maire", "Henry Hoffmann", "Ari Holtzman", "Junchen Jiang"], "categories": ["cs.NI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-11", "url": "https://arxiv.org/abs/2310.07240", "pdf_url": "https://arxiv.org/pdf/2310.07240v6", "arxiv_id": "2310.07240", "doi": "10.1145/3651890.3672274", "citation_count": 229, "influential_citation_count": 13, "has_code": true, "code_url": "https://github.com/UChi-JCL/CacheGen", "venue": "Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication", "quality_score": 0.5904}
{"id": "84d4075c36e64f9d7e3bc67e72c1f98bc9db2ba86ce5e376dcfe74759d9f5295", "sources": ["arxiv", "semantic_scholar"], "title": "Compresso: Structured Pruning with Collaborative Prompting Learns Compact Large Language Models", "abstract": "Despite the remarkable success of Large Language Models (LLMs), the massive size poses significant deployment challenges, particularly on resource-constrained hardware. While existing LLM compression methods focus on quantization, pruning remains relatively unexplored due to the high cost of training-based approaches and data collection challenges. One-shot pruning methods, although cost-effective and data-free, have become dominant in LLM pruning, but lead to performance decline under the structured pruning setting. In this work, we introduce a new paradigm for structurally pruning LLMs, called Compresso. Our approach, through the collaboration of the proposed resource-efficient pruning algorithm and the LLM itself, learns optimal pruning decisions during the training process. Compresso addresses the challenges of expensive training costs and data collection by incorporating Low-Rank Adaptation (LoRA) into the $L_0$ regularization during the instruction tuning process. Then, we further augment the pruning algorithm by introducing a collaborative prompt that fosters collaboration between the LLM and the pruning algorithm, significantly boosting the overall performance. To this end, Compresso prunes LLaMA-7B to 5.4B, maintaining original performance and even surpassing LLaMA-7B in reading comprehension by 2.62%. Extensive experiments demonstrate that Compresso significantly outperforms one-shot pruning baselines across various sparsity ratios, achieving up to 2.21%, 11.43%, 7.04%, and 4.81% higher scores on the commonsense reasoning, reading comprehension, MMLU, and BBH benchmarks, respectively.", "authors": ["Song Guo", "Jiahang Xu", "Li Lyna Zhang", "Mao Yang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-08", "url": "https://arxiv.org/abs/2310.05015", "pdf_url": "https://arxiv.org/pdf/2310.05015v2", "arxiv_id": "2310.05015", "doi": "10.48550/arXiv.2310.05015", "citation_count": 22, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3404}
{"id": "b22d80193d41d831476ae4b097bf0ffa1f5022e0d0a5a9d34960200b9b330025", "sources": ["arxiv", "semantic_scholar"], "title": "Exploiting Activation Sparsity with Dense to Dynamic-k Mixture-of-Experts Conversion", "abstract": "Transformer models can face practical limitations due to their high computational requirements. At the same time, such models exhibit significant activation sparsity, which can be leveraged to reduce the inference cost by converting parts of the network into equivalent Mixture-of-Experts (MoE) layers. Despite the crucial role played by activation sparsity, its impact on this process remains unexplored. We demonstrate that the efficiency of the conversion can be significantly enhanced by a proper regularization of the activation sparsity of the base model. Moreover, motivated by the high variance of the number of activated neurons for different inputs, we introduce a more effective dynamic-$k$ expert selection rule that adjusts the number of executed experts on a per-token basis. To achieve further savings, we extend this approach to multi-head attention projections. Finally, we develop an efficient implementation that translates these computational savings into actual wall-clock speedup. The proposed method, Dense to Dynamic-$k$ Mixture-of-Experts (D2DMoE), outperforms existing approaches on common NLP and vision tasks, reducing inference cost by up to 60% without significantly impacting performance.", "authors": ["Filip Szatkowski", "Bartosz Wójcik", "Mikołaj Piórczyński", "Simone Scardapane"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-06", "url": "https://arxiv.org/abs/2310.04361", "pdf_url": "https://arxiv.org/pdf/2310.04361v4", "arxiv_id": "2310.04361", "doi": "10.52202/079017-1370", "citation_count": 14, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.301}
{"id": "e71b31026e35561e604180aca91fd6d19f4dfb448592fdfa550583c03c945b5b", "sources": ["arxiv", "semantic_scholar"], "title": "BA-MoE: Boundary-Aware Mixture-of-Experts Adapter for Code-Switching Speech Recognition", "abstract": "Mixture-of-experts based models, which use language experts to extract language-specific representations effectively, have been well applied in code-switching automatic speech recognition. However, there is still substantial space to improve as similar pronunciation across languages may result in ineffective multi-language modeling and inaccurate language boundary estimation. To eliminate these drawbacks, we propose a cross-layer language adapter and a boundary-aware training method, namely Boundary-Aware Mixture-of-Experts (BA-MoE). Specifically, we introduce language-specific adapters to separate language-specific representations and a unified gating layer to fuse representations within each encoder layer. Second, we compute language adaptation loss of the mean output of each language-specific adapter to improve the adapter module's language-specific representation learning. Besides, we utilize a boundary-aware predictor to learn boundary representations for dealing with language boundary confusion. Our approach achieves significant performance improvement, reducing the mixture error rate by 16.55\\% compared to the baseline on the ASRU 2019 Mandarin-English code-switching challenge dataset.", "authors": ["Peikun Chen", "Fan Yu", "Yuhao Lian", "Hongfei Xue", "Xucheng Wan", "Naijun Zheng", "Huan Zhou", "Lei Xie"], "categories": ["cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-10-04", "url": "https://arxiv.org/abs/2310.02629", "pdf_url": "https://arxiv.org/pdf/2310.02629v2", "arxiv_id": "2310.02629", "doi": "10.1109/ASRU57964.2023.10389798", "citation_count": 14, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Automatic Speech Recognition & Understanding", "quality_score": 0.294}
{"id": "24f47b840cfae7063b48b86285a5174572970bf1bef43cc0fc730a584d755331", "sources": ["arxiv", "semantic_scholar"], "title": "NOLA: Compressing LoRA using Linear Combination of Random Basis", "abstract": "Fine-tuning Large Language Models (LLMs) and storing them for each downstream task or domain is impractical because of the massive model size (e.g., 350GB in GPT-3). Current literature, such as LoRA, showcases the potential of low-rank modifications to the original weights of an LLM, enabling efficient adaptation and storage for task-specific models. These methods can reduce the number of parameters needed to fine-tune an LLM by several orders of magnitude. Yet, these methods face two primary limitations: (1) the parameter count is lower-bounded by the rank one decomposition, and (2) the extent of reduction is heavily influenced by both the model architecture and the chosen rank. We introduce NOLA, which overcomes the rank one lower bound present in LoRA. It achieves this by re-parameterizing the low-rank matrices in LoRA using linear combinations of randomly generated matrices (basis) and optimizing the linear mixture coefficients only. This approach allows us to decouple the number of trainable parameters from both the choice of rank and the network architecture. We present adaptation results using GPT-2, LLaMA-2, and ViT in natural language and computer vision tasks. NOLA performs as well as LoRA models with much fewer number of parameters compared to LoRA with rank one, the best compression LoRA can archive. Particularly, on LLaMA-2 70B, our method is almost 20 times more compact than the most compressed LoRA without degradation in accuracy. Our code is available here: https://github.com/UCDvision/NOLA", "authors": ["Soroush Abbasi Koohpayegani", "KL Navaneet", "Parsa Nooralinejad", "Soheil Kolouri", "Hamed Pirsiavash"], "categories": ["cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-04", "url": "https://arxiv.org/abs/2310.02556", "pdf_url": "https://arxiv.org/pdf/2310.02556v2", "arxiv_id": "2310.02556", "doi": null, "citation_count": 35, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/UCDvision/NOLA", "venue": "International Conference on Learning Representations", "quality_score": 0.3891}
{"id": "0e4bbe6ee648c0dab4a697fec4bc27943d5ad4b15ca31b3de177462f76b867d7", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Quantized Experts (MoQE): Complementary Effect of Low-bit Quantization and Robustness", "abstract": "Large Mixture of Experts (MoE) models could achieve state-of-the-art quality on various language tasks, including machine translation task, thanks to the efficient model scaling capability with expert parallelism. However, it has brought a fundamental issue of larger memory consumption and increased memory bandwidth bottleneck at deployment time. In this paper, we propose Mixture of Quantized Experts (MoQE) which is a simple weight-only quantization method applying ultra low-bit down to 2-bit quantizations only to expert weights for mitigating the increased memory and latency issues of MoE models. We show that low-bit quantization together with the MoE architecture delivers a reliable model performance while reducing the memory size significantly even without any additional training in most cases. In particular, expert layers in MoE models are much more robust to the quantization than conventional feedforward networks (FFN) layers. In our comprehensive analysis, we show that MoE models with 2-bit expert weights can deliver better model performance than the dense model trained on the same dataset. As a result of low-bit quantization, we show the model size can be reduced by 79.6% of the original half precision floating point (fp16) MoE model. Combined with an optimized GPU runtime implementation, it also achieves 1.24X speed-up on A100 GPUs.", "authors": ["Young Jin Kim", "Raffy Fahim", "Hany Hassan Awadalla"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-03", "url": "https://arxiv.org/abs/2310.02410", "pdf_url": "https://arxiv.org/pdf/2310.02410v1", "arxiv_id": "2310.02410", "doi": "10.48550/arXiv.2310.02410", "citation_count": 37, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3949}
{"id": "f119b5c4357b65084b694a2505c18cd9cd4b07b59984226ade302a182c790786", "sources": ["arxiv", "semantic_scholar"], "title": "Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs", "abstract": "In this study, we introduce adaptive KV cache compression, a plug-and-play method that reduces the memory footprint of generative inference for Large Language Models (LLMs). Different from the conventional KV cache that retains key and value vectors for all context tokens, we conduct targeted profiling to discern the intrinsic structure of attention modules. Based on the recognized structure, we then construct the KV cache in an adaptive manner: evicting long-range contexts on attention heads emphasizing local contexts, discarding non-special tokens on attention heads centered on special tokens, and only employing the standard KV cache for attention heads that broadly attend to all tokens. Moreover, with the lightweight attention profiling used to guide the construction of the adaptive KV cache, FastGen can be deployed without resource-intensive fine-tuning or re-training. In our experiments across various asks, FastGen demonstrates substantial reduction on GPU memory consumption with negligible generation quality loss. We will release our code and the compatible CUDA kernel for reproducibility.", "authors": ["Suyu Ge", "Yunan Zhang", "Liyuan Liu", "Minjia Zhang", "Jiawei Han", "Jianfeng Gao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-03", "url": "https://arxiv.org/abs/2310.01801", "pdf_url": "https://arxiv.org/pdf/2310.01801v4", "arxiv_id": "2310.01801", "doi": "10.48550/arXiv.2310.01801", "citation_count": 473, "influential_citation_count": 43, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.8217}
{"id": "c8ac18eba23bbef0bb2f5b7066a77cd682f997bfbfc88a052c74ba37210c7887", "sources": ["arxiv", "semantic_scholar"], "title": "QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models", "abstract": "Recently years have witnessed a rapid development of large language models (LLMs). Despite the strong ability in many language-understanding tasks, the heavy computational burden largely restricts the application of LLMs especially when one needs to deploy them onto edge devices. In this paper, we propose a quantization-aware low-rank adaptation (QA-LoRA) algorithm. The motivation lies in the imbalanced degrees of freedom of quantization and adaptation, and the solution is to use group-wise operators which increase the degree of freedom of quantization meanwhile decreasing that of adaptation. QA-LoRA is easily implemented with a few lines of code, and it equips the original LoRA with two-fold abilities: (i) during fine-tuning, the LLM's weights are quantized (e.g., into INT4) to reduce time and memory usage; (ii) after fine-tuning, the LLM and auxiliary weights are naturally integrated into a quantized model without loss of accuracy. We apply QA-LoRA to the LLaMA and LLaMA2 model families and validate its effectiveness in different fine-tuning datasets and downstream scenarios. Code will be made available at https://github.com/yuhuixu1993/qa-lora.", "authors": ["Yuhui Xu", "Lingxi Xie", "Xiaotao Gu", "Xin Chen", "Heng Chang", "Hengheng Zhang", "Zhengsu Chen", "Xiaopeng Zhang", "Qi Tian"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-26", "url": "https://arxiv.org/abs/2309.14717", "pdf_url": "https://arxiv.org/pdf/2309.14717v2", "arxiv_id": "2309.14717", "doi": "10.48550/arXiv.2309.14717", "citation_count": 189, "influential_citation_count": 14, "has_code": true, "code_url": "https://github.com/yuhuixu1993/qa-lora", "venue": "International Conference on Learning Representations", "quality_score": 0.588}
{"id": "61da2b562e433140da2c3211c6d61099b0405344a5b67a8006508f5bf7a13f49", "sources": ["arxiv", "semantic_scholar"], "title": "MoCaE: Mixture of Calibrated Experts Significantly Improves Object Detection", "abstract": "Combining the strengths of many existing predictors to obtain a Mixture of Experts which is superior to its individual components is an effective way to improve the performance without having to develop new architectures or train a model from scratch. However, surprisingly, we find that naïvely combining expert object detectors in a similar way to Deep Ensembles, can often lead to degraded performance. We identify that the primary cause of this issue is that the predictions of the experts do not match their performance, a term referred to as miscalibration. Consequently, the most confident detector dominates the final predictions, preventing the mixture from leveraging all the predictions from the experts appropriately. To address this, when constructing the Mixture of Experts, we propose to combine their predictions in a manner which reflects the individual performance of the experts; an objective we achieve by first calibrating the predictions before filtering and refining them. We term this approach the Mixture of Calibrated Experts and demonstrate its effectiveness through extensive experiments on 5 different detection tasks using a variety of detectors, showing that it: (i) improves object detectors on COCO and instance segmentation methods on LVIS by up to $\\sim 2.5$ AP; (ii) reaches state-of-the-art on COCO test-dev with $65.1$ AP and on DOTA with $82.62$ $\\mathrm{AP_{50}}$; (iii) outperforms single models consistently on recent detection tasks such as Open Vocabulary Object Detection.", "authors": ["Kemal Oksuz", "Selim Kuzucu", "Tom Joy", "Puneet K. Dokania"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-26", "url": "https://arxiv.org/abs/2309.14976", "pdf_url": "https://arxiv.org/pdf/2309.14976v4", "arxiv_id": "2309.14976", "doi": "10.48550/arXiv.2309.14976", "citation_count": 19, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3253}
{"id": "2a194205da75cc3bdb3065d80bfc364a0aa5f175987d714b989be473f7dfb97d", "sources": ["arxiv", "semantic_scholar"], "title": "Statistical Perspective of Top-K Sparse Softmax Gating Mixture of Experts", "abstract": "Top-K sparse softmax gating mixture of experts has been widely used for scaling up massive deep-learning architectures without increasing the computational cost. Despite its popularity in real-world applications, the theoretical understanding of that gating function has remained an open problem. The main challenge comes from the structure of the top-K sparse softmax gating function, which partitions the input space into multiple regions with distinct behaviors. By focusing on a Gaussian mixture of experts, we establish theoretical results on the effects of the top-K sparse softmax gating function on both density and parameter estimations. Our results hinge upon defining novel loss functions among parameters to capture different behaviors of the input regions. When the true number of experts $k_{\\ast}$ is known, we demonstrate that the convergence rates of density and parameter estimations are both parametric on the sample size. However, when $k_{\\ast}$ becomes unknown and the true model is over-specified by a Gaussian mixture of $k$ experts where $k > k_{\\ast}$, our findings suggest that the number of experts selected from the top-K sparse softmax gating function must exceed the total cardinality of a certain number of Voronoi cells associated with the true parameters to guarantee the convergence of the density estimation. Moreover, while the density estimation rate remains parametric under this setting, the parameter estimation rates become substantially slow due to an intrinsic interaction between the softmax gating and expert functions.", "authors": ["Huy Nguyen", "Pedram Akbarian", "Fanqi Yan", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-09-25", "url": "https://arxiv.org/abs/2309.13850", "pdf_url": "https://arxiv.org/pdf/2309.13850v2", "arxiv_id": "2309.13850", "doi": "10.48550/arXiv.2309.13850", "citation_count": 31, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3763}
{"id": "17657bc1213f94c0749498e2394f32e20cefa1f8377865105b61a5896e79d93a", "sources": ["arxiv", "semantic_scholar"], "title": "Multiagent Reinforcement Learning with an Attention Mechanism for Improving Energy Efficiency in LoRa Networks", "abstract": "Long Range (LoRa) wireless technology, characterized by low power consumption and a long communication range, is regarded as one of the enabling technologies for the Industrial Internet of Things (IIoT). However, as the network scale increases, the energy efficiency (EE) of LoRa networks decreases sharply due to severe packet collisions. To address this issue, it is essential to appropriately assign transmission parameters such as the spreading factor and transmission power for each end device (ED). However, due to the sporadic traffic and low duty cycle of LoRa networks, evaluating the system EE performance under different parameter settings is time-consuming. Therefore, we first formulate an analytical model to calculate the system EE. On this basis, we propose a transmission parameter allocation algorithm based on multiagent reinforcement learning (MALoRa) with the aim of maximizing the system EE of LoRa networks. Notably, MALoRa employs an attention mechanism to guide each ED to better learn how much ''attention'' should be given to the parameter assignments for relevant EDs when seeking to improve the system EE. Simulation results demonstrate that MALoRa significantly improves the system EE compared with baseline algorithms with an acceptable degradation in packet delivery rate (PDR).", "authors": ["Xu Zhang", "Ziqi Lin", "Shimin Gong", "Bo Gu", "Dusit Niyato"], "categories": ["cs.AI", "cs.LG", "cs.MA"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-16", "url": "https://arxiv.org/abs/2309.08965", "pdf_url": "https://arxiv.org/pdf/2309.08965v1", "arxiv_id": "2309.08965", "doi": "10.1109/GLOBECOM54140.2023.10437219", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Global Communications Conference", "quality_score": 0.1505}
{"id": "3fb7c0d2b9bbbfc69d9f8e72f3b335c2e65add5ed1aa92042172af1473505a65", "sources": ["arxiv", "semantic_scholar"], "title": "Augmented Quantization: Mixture Models for Risk-Oriented Sensitivity Analysis", "abstract": "A central question in risk analysis is to identify the factors that drive the system toward a specific hazardous outcome, such as the exceedance of a given threshold. When relying on numerical simulators, we propose to study the distribution of the inputs, transformed into uniform variables via their cumulative distributions, conditionally on the occurrence of the hazardous event. To represent this multivariate conditional distribution for sensitivity analysis, we introduce an original quantization approach based on estimating a mixture of Dirac and local uniform distributions. For each marginal of this mixture, a Dirac component indicates a strong influence of the corresponding variable, whereas a uniform component with wide support reflects weak influence. A notable advantage of this method is its ability to identify the regions of the input space that most strongly influence the occurrence of the risk event, while also capturing the joint effects of multiple variables. However, learning mixture models typically relies on likelihood-based methods, which are not well suited to mixtures involving singular or Dirac components. To address this, we propose an \\emph{Augmented Quantization} method, a reformulation of the classical quantization problem based on the p-Wasserstein distance, which can be computed in very general distribution spaces. The performance of Augmented Quantization in estimating such mixture models is first demonstrated on analytical toy problems, and then applied to sensitivity analysis of both an analytical function and a practical flooding case study on a section of the Loire River.", "authors": ["Charlie Sire", "Didier Rullière", "Rodolphe Le Riche", "Jérémy Rohmer", "Yann Richet", "Lucie Pheulpin"], "categories": ["stat.AP"], "fields_of_study": ["Mathematics"], "published_date": "2023-09-15", "url": "https://arxiv.org/abs/2309.08389", "pdf_url": "https://arxiv.org/pdf/2309.08389v7", "arxiv_id": "2309.08389", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "7e5ea6cea0adb1b333b6de7af5b718d7c60e1417dfab5795ea336236e7fcd07f", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Point Detection of the Powerful Gamma Ray Burst GRB221009A Propagation through the Heliosphere on October 9, 2022", "abstract": "We present the results of processing the effects of the powerful Gamma Ray Burst GRB221009A captured by the charged particle detectors (electrostatic analyzers and solid-state detectors) onboard spacecraft at different points in the heliosphere on October 9, 2022. To follow the GRB221009A propagation through the heliosphere we used the electron and proton flux measurements from solar missions Solar Orbiter and STEREO-A; Earth magnetosphere and the solar wind missions THEMIS and Wind; meteorological satellites POES15, POES19, MetOp3; and MAVEN - a NASA mission orbiting Mars. GRB221009A had a structure of four bursts: less intense Pulse 1 - the triggering impulse - was detected by gamma-ray observatories at 131659 UT (near the Earth); the most intense Pulses 2 and 3 were detected on board all the spacecraft from the list, and Pulse 4 detected in more than 500 s after Pulse 1. Due to their different scientific objectives, the spacecraft, which data was used in this study, were separated by more than 1 AU (Solar Orbiter and MAVEN). This enabled tracking GRB221009A as it was propagating across the heliosphere. STEREO-A was the first to register Pulse 2 and 3 of the GRB, almost 100 seconds before their detection by spacecraft in the vicinity of Earth. MAVEN detected GRB221009A Pulses 2, 3, and 4 at the orbit of Mars about 237 seconds after their detection near Earth. By processing the time delays observed we show that the source location of the GRB221009A was at RA 288.5 degrees, Dec 18.5 degrees (J2000) with an error cone of 2 degrees", "authors": ["Andrii Voshchepynets", "Oleksiy Agapitov", "Lynn Wilson", "Vassilis Angelopoulos", "Samer T. Alnussirat", "Michael Balikhin", "Myroslava Hlebena", "Ihor Korol", "Davin Larson", "David Mitchell", "Christopher Owen", "Ali Rahmati", "Department of System Analysis", "Optimization Theory", "Uzhhorod National University", " Uzhhorod", " Ukraine", "Space Sciences Laboratory", " :", "University of California Berkeley Berkeley", "CA 94720", " Astronomy", "Space Physics Department", "National Taras Shevchenko University of Kyiv", " Kyiv", " Ukraine", "Goddard Space Flight Center", "National Aeronautics", "Space Administration", " Greenbelt", " MD", "Department of Earth", " Planetary", "Space Sciences", "University of California Los Angeles", "Los Angeles", " CA}", "University of Sheffield", " Sheffield", " UK", "Department of Algebra", "Differential Equation", "Uzhhorod National University", " Uzhhorod", " Ukraine", "Department of Mathematical Analysis", "The John Paul II Catholic University of Lublin", " Lublin", " Poland}"], "categories": ["astro-ph.HE", "astro-ph.IM", "astro-ph.SR"], "fields_of_study": ["Physics"], "published_date": "2023-09-11", "url": "https://arxiv.org/abs/2309.05856", "pdf_url": "https://arxiv.org/pdf/2309.05856v1", "arxiv_id": "2309.05856", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "6a460675cb8309e557ef631d6fbdc4e5fb14de7b7f297cb38229d853e5e72748", "sources": ["arxiv", "semantic_scholar"], "title": "Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning", "abstract": "The Mixture of Experts (MoE) is a widely known neural architecture where an ensemble of specialized sub-models optimizes overall performance with a constant computational cost. However, conventional MoEs pose challenges at scale due to the need to store all experts in memory. In this paper, we push MoE to the limit. We propose extremely parameter-efficient MoE by uniquely combining MoE architecture with lightweight experts.Our MoE architecture outperforms standard parameter-efficient fine-tuning (PEFT) methods and is on par with full fine-tuning by only updating the lightweight experts -- less than 1% of an 11B parameters model. Furthermore, our method generalizes to unseen tasks as it does not depend on any prior task knowledge. Our research underscores the versatility of the mixture of experts architecture, showcasing its ability to deliver robust performance even when subjected to rigorous parameter constraints. Our code used in all the experiments is publicly available here: https://github.com/for-ai/parameter-efficient-moe.", "authors": ["Ted Zadouri", "Ahmet Üstün", "Arash Ahmadian", "Beyza Ermiş", "Acyr Locatelli", "Sara Hooker"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-11", "url": "https://arxiv.org/abs/2309.05444", "pdf_url": "https://arxiv.org/pdf/2309.05444v1", "arxiv_id": "2309.05444", "doi": "10.48550/arXiv.2309.05444", "citation_count": 176, "influential_citation_count": 19, "has_code": true, "code_url": "https://github.com/for-ai/parameter-efficient-moe", "venue": "International Conference on Learning Representations", "quality_score": 0.6505}
{"id": "dcf865a6afb9fd4132ef5aec1d9c76ca7b278afbc755444c618242a7468e0f04", "sources": ["arxiv", "semantic_scholar"], "title": "Liu-type Shrinkage Estimators for Mixture of Poisson Regressions with Experts: A Heart Disease Study", "abstract": "Count data play a critical role in medical research, such as heart disease. The Poisson regression model is a common technique for evaluating the impact of a set of covariates on the count responses. The mixture of Poisson regression models with experts is a practical tool to exploit the covariates, not only to handle the heterogeneity in the Poisson regressions but also to learn the mixing structure of the population. Multicollinearity is one of the most common challenges with regression models, leading to ill-conditioned design matrices of Poisson regression components and expert classes. The maximum likelihood method produces unreliable and misleading estimates for the effects of the covariates in multicollinearity. In this research, we develop Ridge and Liu-type methods as two shrinkage approaches to cope with the ill-conditioned design matrices of the mixture of Poisson regression models with experts. Through various numerical studies, we demonstrate that the shrinkage methods offer more reliable estimates for the coefficients of the mixture model in multicollinearity while maintaining the classification performance of the ML method. The shrinkage methods are finally applied to a heart study to analyze the heart disease rate stages.", "authors": ["Elsayed Ghanem", "Moein Yoosefi", "Armin Hatefi"], "categories": ["stat.ME", "stat.CO", "stat.ML"], "fields_of_study": ["Mathematics"], "published_date": "2023-09-11", "url": "https://arxiv.org/abs/2309.05838", "pdf_url": "https://arxiv.org/pdf/2309.05838v1", "arxiv_id": "2309.05838", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "2c33c7d6a8868309bd7125ef6ba431574be7a6895c0dc50725b018ed29d3a2cf", "sources": ["arxiv", "semantic_scholar"], "title": "DeepTriNet: A Tri-Level Attention Based DeepLabv3+ Architecture for Semantic Segmentation of Satellite Images", "abstract": "The segmentation of satellite images is crucial in remote sensing applications. Existing methods face challenges in recognizing small-scale objects in satellite images for semantic segmentation primarily due to ignoring the low-level characteristics of the underlying network and due to containing distinct amounts of information by different feature maps. Thus, in this research, a tri-level attention-based DeepLabv3+ architecture (DeepTriNet) is proposed for the semantic segmentation of satellite images. The proposed hybrid method combines squeeze-and-excitation networks (SENets) and tri-level attention units (TAUs) with the vanilla DeepLabv3+ architecture, where the TAUs are used to bridge the semantic feature gap among encoders output and the SENets used to put more weight on relevant features. The proposed DeepTriNet finds which features are the more relevant and more generalized way by its self-supervision rather we annotate them. The study showed that the proposed DeepTriNet performs better than many conventional techniques with an accuracy of 98% and 77%, IoU 80% and 58%, precision 88% and 68%, and recall of 79% and 55% on the 4-class Land-Cover.ai dataset and the 15-class GID-2 dataset respectively. The proposed method will greatly contribute to natural resource management and change detection in rural and urban regions through efficient and semantic satellite image segmentation", "authors": ["Tareque Bashar Ovi", "Shakil Mosharrof", "Nomaiya Bashree", "Md Shofiqul Islam", "Muhammad Nazrul Islam"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-05", "url": "https://arxiv.org/abs/2310.06848", "pdf_url": "https://arxiv.org/pdf/2310.06848v1", "arxiv_id": "2310.06848", "doi": "10.48550/arXiv.2310.06848", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "66f48ccc9c435cfb30a5b0682ba152a15c16a65838a4d24475e1d7f15caf212d", "sources": ["arxiv", "semantic_scholar"], "title": "Mask-Attention-Free Transformer for 3D Instance Segmentation", "abstract": "Recently, transformer-based methods have dominated 3D instance segmentation, where mask attention is commonly involved. Specifically, object queries are guided by the initial instance masks in the first cross-attention, and then iteratively refine themselves in a similar manner. However, we observe that the mask-attention pipeline usually leads to slow convergence due to low-recall initial instance masks. Therefore, we abandon the mask attention design and resort to an auxiliary center regression task instead. Through center regression, we effectively overcome the low-recall issue and perform cross-attention by imposing positional prior. To reach this goal, we develop a series of position-aware designs. First, we learn a spatial distribution of 3D locations as the initial position queries. They spread over the 3D space densely, and thus can easily capture the objects in a scene with a high recall. Moreover, we present relative position encoding for the cross-attention and iterative refinement for more accurate position queries. Experiments show that our approach converges 4x faster than existing work, sets a new state of the art on ScanNetv2 3D instance segmentation benchmark, and also demonstrates superior performance across various datasets. Code and models are available at https://github.com/dvlab-research/Mask-Attention-Free-Transformer.", "authors": ["Xin Lai", "Yuhui Yuan", "Ruihang Chu", "Yukang Chen", "Han Hu", "Jiaya Jia"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-04", "url": "https://arxiv.org/abs/2309.01692", "pdf_url": "https://arxiv.org/pdf/2309.01692v1", "arxiv_id": "2309.01692", "doi": "10.1109/ICCV51070.2023.00342", "citation_count": 57, "influential_citation_count": 14, "has_code": true, "code_url": "https://github.com/dvlab-research/Mask-Attention-Free-Transformer", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.588}
{"id": "7a9805135472d8b859ed93c1e601a997342b355d563c438f1f61c60105f688c5", "sources": ["arxiv", "semantic_scholar"], "title": "Implementation and Evaluation of Physical Layer Key Generation on SDR based LoRa Platform", "abstract": "Physical layer key generation technology which leverages channel randomness to generate secret keys has attracted extensive attentions in long range (LoRa)-based networks recently. We in this paper develop a software-defined radio (SDR) based LoRa communications platform using GNU Radio on universal software radio peripheral (USRP) to implement and evaluate typical physical layer key generation schemes. Thanks to the flexibility and configurability of GNU Radio to extract LoRa packets, we are able to obtain the fine-grained channel frequency response (CFR) through LoRa preamble based channel estimation for key generation. Besides, we propose a lowcomplexity preprocessing method to enhance the randomness of quantization while reducing the secret key disagreement ratio. The results indicate that we can achieve 367 key bits with a high level of randomness through just a single effective channel probing in an indoor environment at a distance of 2 meters under the circumstance of a spreading factor (SF) of 7, a preamble length of 8, a signal bandwidth of 250 kHz, and a sampling rate of 1 MHz.", "authors": ["Yingying Hu", "Dongyang Xu", "Tiantian Zhang"], "categories": ["eess.SP"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2023-08-30", "url": "https://arxiv.org/abs/2308.15696", "pdf_url": "https://arxiv.org/pdf/2308.15696v1", "arxiv_id": "2308.15696", "doi": "10.1109/VTC2023-Fall60731.2023.10333718", "citation_count": 8, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "IEEE Vehicular Technology Conference", "quality_score": 0.2386}
{"id": "b340bd748db8c5d1dcf0509257c49ca7caaa516b8c5a4e593f175031f6936ed5", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Mobile Face Anti-Spoofing: A Robust Framework for Diverse Attack Types under Screen Flash", "abstract": "Face anti-spoofing (FAS) is crucial for securing face recognition systems. However, existing FAS methods with handcrafted binary or pixel-wise labels have limitations due to diverse presentation attacks (PAs). In this paper, we propose an attack type robust face anti-spoofing framework under light flash, called ATR-FAS. Due to imaging differences caused by various attack types, traditional FAS methods based on single binary classification network may result in excessive intra-class distance of spoof faces, leading to a challenge of decision boundary learning. Therefore, we employed multiple networks to reconstruct multi-frame depth maps as auxiliary supervision, and each network experts in one type of attack. A dual gate module (DGM) consisting of a type gate and a frame-attention gate is introduced, which perform attack type recognition and multi-frame attention generation, respectively. The outputs of DGM are utilized as weight to mix the result of multiple expert networks. The multi-experts mixture enables ATR-FAS to generate spoof-differentiated depth maps, and stably detects spoof faces without being affected by different types of PAs. Moreover, we design a differential normalization procedure to convert original flash frames into differential frames. This simple but effective processing enhances the details in flash frames, aiding in the generation of depth maps. To verify the effectiveness of our framework, we collected a large-scale dataset containing 12,660 live and spoof videos with diverse PAs under dynamic flash from the smartphone screen. Extensive experiments illustrate that the proposed ATR-FAS significantly outperforms existing state-of-the-art methods. The code and dataset will be available at https://github.com/Chaochao-Lin/ATR-FAS.", "authors": ["Weihua Liu", "Chaochao Lin", "Yu Yan"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-29", "url": "https://arxiv.org/abs/2308.15346", "pdf_url": "https://arxiv.org/pdf/2308.15346v1", "arxiv_id": "2308.15346", "doi": "10.48550/arXiv.2308.15346", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Chaochao-Lin/ATR-FAS", "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "7d23e41b9371d888ca7aa086c14767bbd2abfe7e8beb061fa8b3129407c5b05b", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning Self-Attention for Zero-Shot Multi-Speaker Text-to-Speech", "abstract": "For personalized speech generation, a neural text-to-speech (TTS) model must be successfully implemented with limited data from a target speaker. To this end, the baseline TTS model needs to be amply generalized to out-of-domain data (i.e., target speaker's speech). However, approaches to address this out-of-domain generalization problem in TTS have yet to be thoroughly studied. In this work, we propose an effective pruning method for a transformer known as sparse attention, to improve the TTS model's generalization abilities. In particular, we prune off redundant connections from self-attention layers whose attention weights are below the threshold. To flexibly determine the pruning strength for searching optimal degree of generalization, we also propose a new differentiable pruning method that allows the model to automatically learn the thresholds. Evaluations on zero-shot multi-speaker TTS verify the effectiveness of our method in terms of voice quality and speaker similarity.", "authors": ["Hyungchan Yoon", "Changhwan Kim", "Eunwoo Song", "Hyun-Wook Yoon", "Hong-Goo Kang"], "categories": ["cs.SD", "cs.LG", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-08-28", "url": "https://arxiv.org/abs/2308.14909", "pdf_url": "https://arxiv.org/pdf/2308.14909v1", "arxiv_id": "2308.14909", "doi": "10.21437/Interspeech.2023-1301", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.1193}
{"id": "8ed166a338e02f54d656ce0149c868042ce5bf6dda4238f135f25c006915d1bb", "sources": ["arxiv", "semantic_scholar"], "title": "Partition-and-Debias: Agnostic Biases Mitigation via A Mixture of Biases-Specific Experts", "abstract": "Bias mitigation in image classification has been widely researched, and existing methods have yielded notable results. However, most of these methods implicitly assume that a given image contains only one type of known or unknown bias, failing to consider the complexities of real-world biases. We introduce a more challenging scenario, agnostic biases mitigation, aiming at bias removal regardless of whether the type of bias or the number of types is unknown in the datasets. To address this difficult task, we present the Partition-and-Debias (PnD) method that uses a mixture of biases-specific experts to implicitly divide the bias space into multiple subspaces and a gating module to find a consensus among experts to achieve debiased classification. Experiments on both public and constructed benchmarks demonstrated the efficacy of the PnD. Code is available at: https://github.com/Jiaxuan-Li/PnD.", "authors": ["Jiaxuan Li", "Duc Minh Vo", "Hideki Nakayama"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-19", "url": "https://arxiv.org/abs/2308.10005", "pdf_url": "https://arxiv.org/pdf/2308.10005v1", "arxiv_id": "2308.10005", "doi": "10.1109/ICCV51070.2023.00454", "citation_count": 7, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Jiaxuan-Li/PnD", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.2258}
{"id": "990be4b6087c3a38e05726ad95248bc81e6dab98987ac136549a8531596e8cb4", "sources": ["arxiv", "semantic_scholar"], "title": "Robust Mixture-of-Expert Training for Convolutional Neural Networks", "abstract": "Sparsely-gated Mixture of Expert (MoE), an emerging deep model architecture, has demonstrated a great promise to enable high-accuracy and ultra-efficient model inference. Despite the growing popularity of MoE, little work investigated its potential to advance convolutional neural networks (CNNs), especially in the plane of adversarial robustness. Since the lack of robustness has become one of the main hurdles for CNNs, in this paper we ask: How to adversarially robustify a CNN-based MoE model? Can we robustly train it like an ordinary CNN model? Our pilot study shows that the conventional adversarial training (AT) mechanism (developed for vanilla CNNs) no longer remains effective to robustify an MoE-CNN. To better understand this phenomenon, we dissect the robustness of an MoE-CNN into two dimensions: Robustness of routers (i.e., gating functions to select data-specific experts) and robustness of experts (i.e., the router-guided pathways defined by the subnetworks of the backbone CNN). Our analyses show that routers and experts are hard to adapt to each other in the vanilla AT. Thus, we propose a new router-expert alternating Adversarial training framework for MoE, termed AdvMoE. The effectiveness of our proposal is justified across 4 commonly-used CNN model architectures over 4 benchmark datasets. We find that AdvMoE achieves 1% ~ 4% adversarial robustness improvement over the original dense CNN, and enjoys the efficiency merit of sparsity-gated MoE, leading to more than 50% inference cost reduction. Codes are available at https://github.com/OPTML-Group/Robust-MoE-CNN.", "authors": ["Yihua Zhang", "Ruisi Cai", "Tianlong Chen", "Guanhua Zhang", "Huan Zhang", "Pin-Yu Chen", "Shiyu Chang", "Zhangyang Wang", "Sijia Liu"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-19", "url": "https://arxiv.org/abs/2308.10110", "pdf_url": "https://arxiv.org/pdf/2308.10110v1", "arxiv_id": "2308.10110", "doi": "10.1109/ICCV51070.2023.00015", "citation_count": 44, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/OPTML-Group/Robust-MoE-CNN", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.4133}
{"id": "a6aa6cbe90d4bf34061b0ae41a312e95b518b3ed773bf00f343f5f88d00fc9ca", "sources": ["arxiv", "semantic_scholar"], "title": "Unified Data-Free Compression: Pruning and Quantization without Fine-Tuning", "abstract": "Structured pruning and quantization are promising approaches for reducing the inference time and memory footprint of neural networks. However, most existing methods require the original training dataset to fine-tune the model. This not only brings heavy resource consumption but also is not possible for applications with sensitive or proprietary data due to privacy and security concerns. Therefore, a few data-free methods are proposed to address this problem, but they perform data-free pruning and quantization separately, which does not explore the complementarity of pruning and quantization. In this paper, we propose a novel framework named Unified Data-Free Compression(UDFC), which performs pruning and quantization simultaneously without any data and fine-tuning process. Specifically, UDFC starts with the assumption that the partial information of a damaged(e.g., pruned or quantized) channel can be preserved by a linear combination of other channels, and then derives the reconstruction form from the assumption to restore the information loss due to compression. Finally, we formulate the reconstruction error between the original network and its compressed network, and theoretically deduce the closed-form solution. We evaluate the UDFC on the large-scale image classification task and obtain significant improvements over various network architectures and compression methods. For example, we achieve a 20.54% accuracy improvement on ImageNet dataset compared to SOTA method with 30% pruning ratio and 6-bit quantization on ResNet-34.", "authors": ["Shipeng Bai", "Jun Chen", "Xintian Shen", "Yixuan Qian", "Yong Liu"], "categories": ["cs.LG", "cs.CV", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-08-14", "url": "https://arxiv.org/abs/2308.07209", "pdf_url": "https://arxiv.org/pdf/2308.07209v1", "arxiv_id": "2308.07209", "doi": "10.1109/ICCV51070.2023.00540", "citation_count": 32, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.3796}
{"id": "0988fa127abfcefd919f5baf9f3941aa8abde4913dafa9d7498e5f2c3b3c42bb", "sources": ["arxiv", "semantic_scholar"], "title": "From Sparse to Soft Mixtures of Experts", "abstract": "Sparse mixture of expert architectures (MoEs) scale model capacity without significant increases in training or inference costs. Despite their success, MoEs suffer from a number of issues: training instability, token dropping, inability to scale the number of experts, or ineffective finetuning. In this work, we propose Soft MoE, a fully-differentiable sparse Transformer that addresses these challenges, while maintaining the benefits of MoEs. Soft MoE performs an implicit soft assignment by passing different weighted combinations of all input tokens to each expert. As in other MoEs, experts in Soft MoE only process a subset of the (combined) tokens, enabling larger model capacity (and performance) at lower inference cost. In the context of visual recognition, Soft MoE greatly outperforms dense Transformers (ViTs) and popular MoEs (Tokens Choice and Experts Choice). Furthermore, Soft MoE scales well: Soft MoE Huge/14 with 128 experts in 16 MoE layers has over 40x more parameters than ViT Huge/14, with only 2% increased inference time, and substantially better quality.", "authors": ["Joan Puigcerver", "Carlos Riquelme", "Basil Mustafa", "Neil Houlsby"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-02", "url": "https://arxiv.org/abs/2308.00951", "pdf_url": "https://arxiv.org/pdf/2308.00951v2", "arxiv_id": "2308.00951", "doi": "10.48550/arXiv.2308.00951", "citation_count": 281, "influential_citation_count": 22, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.6809}
{"id": "e8a9a8ae6064cbb933a87c02dbf2697ba75955a9cee47cbcef9035aac8eb954c", "sources": ["arxiv", "semantic_scholar"], "title": "Pay Attention to What You Need", "abstract": "Although large language models (LLMs) have achieved significant success in natural language processing, they still struggle with long-context comprehension. Traditional approaches to mitigating this issue typically rely on fine-tuning or retraining, which is both resource-intensive and challenging to deploy in lightweight industrial settings. In this paper, we investigate the potential to accomplish this without any additional resources. Through an in-depth study of the attention mechanism in LLMs, we propose a method called Scaled ReAttention (SRA) to strengthen LLMs' ability to interpret and retrieve information by strategically manipulating their attention scores during inference. Through extensive experiments, we demonstrate that integrating SRA significantly boosts LLMs' performance on a variety of downstream tasks, highlighting its practical potential for enhancing language understanding without incurring the overhead of traditional training.", "authors": ["Yifei Gao", "Shaohong Chen", "Lei Wang", "Ruiting Dai", "Ziyun Zhang", "Kerui Ren", "Jiaji Wu", "Jun Cheng"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-25", "url": "https://arxiv.org/abs/2307.13365", "pdf_url": "https://arxiv.org/pdf/2307.13365v3", "arxiv_id": "2307.13365", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "b6528947169a189fa5c1dbb31b229655d8e3d5153f608d2d4e6dd03debf1be35", "sources": ["arxiv", "semantic_scholar"], "title": "Language-Routing Mixture of Experts for Multilingual and Code-Switching Speech Recognition", "abstract": "Multilingual speech recognition for both monolingual and code-switching speech is a challenging task. Recently, based on the Mixture of Experts (MoE), many works have made good progress in multilingual and code-switching ASR, but present huge computational complexity with the increase of supported languages. In this work, we propose a computation-efficient network named Language-Routing Mixture of Experts (LR-MoE) for multilingual and code-switching ASR. LR-MoE extracts language-specific representations through the Mixture of Language Experts (MLE), which is guided to learn by a frame-wise language routing mechanism. The weight-shared frame-level language identification (LID) network is jointly trained as the shared pre-router of each MoE layer. Experiments show that the proposed method significantly improves multilingual and code-switching speech recognition performances over baseline with comparable computational efficiency.", "authors": ["Wenxuan Wang", "Guodong Ma", "Yuke Li", "Binbin Du"], "categories": ["cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-07-12", "url": "https://arxiv.org/abs/2307.05956", "pdf_url": "https://arxiv.org/pdf/2307.05956v2", "arxiv_id": "2307.05956", "doi": "10.48550/arXiv.2307.05956", "citation_count": 46, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.418}
{"id": "fba9696034f2005b69856d800bace686cd5212f075385390ae446a7e1c3e6cd2", "sources": ["arxiv", "semantic_scholar"], "title": "Bidirectional Attention as a Mixture of Continuous Word Experts", "abstract": "Bidirectional attention $\\unicode{x2013}$ composed of self-attention with positional encodings and the masked language model (MLM) objective $\\unicode{x2013}$ has emerged as a key component of modern large language models (LLMs). Despite its empirical success, few studies have examined its statistical underpinnings: What statistical model is bidirectional attention implicitly fitting? What sets it apart from its non-attention predecessors? We explore these questions in this paper. The key observation is that fitting a single-layer single-head bidirectional attention, upon reparameterization, is equivalent to fitting a continuous bag of words (CBOW) model with mixture-of-experts (MoE) weights. Further, bidirectional attention with multiple heads and multiple layers is equivalent to stacked MoEs and a mixture of MoEs, respectively. This statistical viewpoint reveals the distinct use of MoE in bidirectional attention, which aligns with its practical effectiveness in handling heterogeneous data. It also suggests an immediate extension to categorical tabular data, if we view each word location in a sentence as a tabular feature. Across empirical studies, we find that this extension outperforms existing tabular extensions of transformers in out-of-distribution (OOD) generalization. Finally, this statistical perspective of bidirectional attention enables us to theoretically characterize when linear word analogies are present in its word embeddings. These analyses show that bidirectional attention can require much stronger assumptions to exhibit linear word analogies than its non-attention predecessors.", "authors": ["Kevin Christian Wibisono", "Yixin Wang"], "categories": ["cs.CL", "cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-07-08", "url": "https://arxiv.org/abs/2307.04057", "pdf_url": "https://arxiv.org/pdf/2307.04057v2", "arxiv_id": "2307.04057", "doi": "10.48550/arXiv.2307.04057", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Uncertainty in Artificial Intelligence", "quality_score": 0.0753}
{"id": "f71b41cd4d64fd5ab4b7401231d1a60f128aab9576f67f92ea393545c61a045f", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning vs Quantization: Which is Better?", "abstract": "Neural network pruning and quantization techniques are almost as old as neural networks themselves. However, to date only ad-hoc comparisons between the two have been published. In this paper, we set out to answer the question on which is better: neural network quantization or pruning? By answering this question, we hope to inform design decisions made on neural network hardware going forward. We provide an extensive comparison between the two techniques for compressing deep neural networks. First, we give an analytical comparison of expected quantization and pruning error for general data distributions. Then, we provide lower bounds for the per-layer pruning and quantization error in trained networks, and compare these to empirical error after optimization. Finally, we provide an extensive experimental comparison for training 8 large-scale models on 3 tasks. Our results show that in most cases quantization outperforms pruning. Only in some scenarios with very high compression ratio, pruning might be beneficial from an accuracy standpoint.", "authors": ["Andrey Kuzmin", "Markus Nagel", "Mart van Baalen", "Arash Behboodi", "Tijmen Blankevoort"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-06", "url": "https://arxiv.org/abs/2307.02973", "pdf_url": "https://arxiv.org/pdf/2307.02973v2", "arxiv_id": "2307.02973", "doi": "10.48550/arXiv.2307.02973", "citation_count": 130, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5293}
{"id": "128bdba5bce6734067467201ea5b15a02b784f3b2318bc8be777514925cb12e9", "sources": ["arxiv", "semantic_scholar"], "title": "Quantization Variation: A New Perspective on Training Transformers with Low-Bit Precision", "abstract": "Despite the outstanding performance of transformers in both language and vision tasks, the expanding computation and model size have increased the demand for efficient deployment. To address the heavy computation and parameter drawbacks, quantization is frequently studied in the community as a representative model compression technique and has seen extensive use on ConvNets. However, due to the unique properties of transformers, the low-bit quantization applications are still limited and underexplored. In this paper, we identify the difficulty of transformer low-bit quantization-aware training on its unique variation behaviors, which significantly differ from ConvNets. Based on comprehensive quantitative analysis, we observe variation in three hierarchies: various module quantization sensitivities, outliers in static weight and activation distribution, and oscillation in dynamic parameter fluctuations. These variations of transformers bring instability to the quantization-aware training (QAT) and negatively influence the performance. We explore the best practices to alleviate the variation's influence during low-bit transformer QAT and propose a variation-aware quantization scheme for both vision and language transformers. We extensively verify and show our scheme can alleviate the variation and improve the performance of transformers across various models and tasks. Our solution substantially improves the 2-bit Swin-T and binary BERT-base, achieving a 3.35% and 1.4% accuracy improvement over previous state-of-the-art methods on ImageNet-1K and GLUE. Codes and models are available at https://github.com/HuangOwen/Quantization-Variation.", "authors": ["Xijie Huang", "Zhiqiang Shen", "Pingcheng Dong", "Kwang-Ting Cheng"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-01", "url": "https://arxiv.org/abs/2307.00331", "pdf_url": "https://arxiv.org/pdf/2307.00331v2", "arxiv_id": "2307.00331", "doi": null, "citation_count": 24, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/HuangOwen/Quantization-Variation", "venue": null, "quality_score": 0.3495}
{"id": "e15f0eb2e5d615f85138224f87f88d5add5aefa1aac98964f1fe5ebfbb0c5323", "sources": ["arxiv", "semantic_scholar"], "title": "Computational Flash Photography through Intrinsics", "abstract": "Flash is an essential tool as it often serves as the sole controllable light source in everyday photography. However, the use of flash is a binary decision at the time a photograph is captured with limited control over its characteristics such as strength or color. In this work, we study the computational control of the flash light in photographs taken with or without flash. We present a physically motivated intrinsic formulation for flash photograph formation and develop flash decomposition and generation methods for flash and no-flash photographs, respectively. We demonstrate that our intrinsic formulation outperforms alternatives in the literature and allows us to computationally control flash in in-the-wild images.", "authors": ["Sepideh Sarajian Maralan", "Chris Careaga", "Yağız Aksoy"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-09", "url": "https://arxiv.org/abs/2306.06089", "pdf_url": "https://arxiv.org/pdf/2306.06089v1", "arxiv_id": "2306.06089", "doi": "10.1109/CVPR52729.2023.01598", "citation_count": 5, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.2386}
{"id": "8c4779bf05f4e8c1115df71d2f190b2b0dab8e71c2f4c48ec66e36343c2fbd5b", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Weighted Mixture of Experts with Contrastive Learning for Personalized Ranking in E-commerce", "abstract": "Ranking model plays an essential role in e-commerce search and recommendation. An effective ranking model should give a personalized ranking list for each user according to the user preference. Existing algorithms usually extract a user representation vector from the user behavior sequence, then feed the vector into a feed-forward network (FFN) together with other features for feature interactions, and finally produce a personalized ranking score. Despite tremendous progress in the past, there is still room for improvement. Firstly, the personalized patterns of feature interactions for different users are not explicitly modeled. Secondly, most of existing algorithms have poor personalized ranking results for long-tail users with few historical behaviors due to the data sparsity. To overcome the two challenges, we propose Attention Weighted Mixture of Experts (AW-MoE) with contrastive learning for personalized ranking. Firstly, AW-MoE leverages the MoE framework to capture personalized feature interactions for different users. To model the user preference, the user behavior sequence is simultaneously fed into expert networks and the gate network. Within the gate network, one gate unit and one activation unit are designed to adaptively learn the fine-grained activation vector for experts using an attention mechanism. Secondly, a random masking strategy is applied to the user behavior sequence to simulate long-tail users, and an auxiliary contrastive loss is imposed to the output of the gate network to improve the model generalization for these users. This is validated by a higher performance gain on the long-tail user test set. Experiment results on a JD real production dataset and a public dataset demonstrate the effectiveness of AW-MoE, which significantly outperforms state-of-art methods. Notably, AW-MoE has been successfully deployed in the JD e-commerce search engine, ...", "authors": ["Juan Gong", "Zhenlin Chen", "Chaoyi Ma", "Zhuojian Xiao", "Haonan Wang", "Guoyu Tang", "Lin Liu", "Sulong Xu", "Bo Long", "Yunjiang Jiang"], "categories": ["cs.IR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-08", "url": "https://arxiv.org/abs/2306.05011", "pdf_url": "https://arxiv.org/pdf/2306.05011v1", "arxiv_id": "2306.05011", "doi": "10.1109/ICDE55515.2023.00247", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Data Engineering", "quality_score": 0.2386}
{"id": "4f0a629230f1eb399741becc0cfd44291d27ac13c82ef10d3cf088c6e545acc9", "sources": ["arxiv", "semantic_scholar"], "title": "Grounded Text-to-Image Synthesis with Attention Refocusing", "abstract": "Driven by the scalable diffusion models trained on large-scale datasets, text-to-image synthesis methods have shown compelling results. However, these models still fail to precisely follow the text prompt involving multiple objects, attributes, or spatial compositions. In this paper, we reveal the potential causes in the diffusion model's cross-attention and self-attention layers. We propose two novel losses to refocus attention maps according to a given spatial layout during sampling. Creating the layouts manually requires additional effort and can be tedious. Therefore, we explore using large language models (LLM) to produce these layouts for our method. We conduct extensive experiments on the DrawBench, HRS, and TIFA benchmarks to evaluate our proposed method. We show that our proposed attention refocusing effectively improves the controllability of existing approaches.", "authors": ["Quynh Phung", "Songwei Ge", "Jia-Bin Huang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-08", "url": "https://arxiv.org/abs/2306.05427", "pdf_url": "https://arxiv.org/pdf/2306.05427v2", "arxiv_id": "2306.05427", "doi": "10.1109/CVPR52733.2024.00758", "citation_count": 177, "influential_citation_count": 24, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.699}
{"id": "0269ffbc21678ccffd1c913ec5503c1dfa96ee09286c83310088aef77ff00bc5", "sources": ["arxiv", "semantic_scholar"], "title": "Magnitude Attention-based Dynamic Pruning", "abstract": "Existing pruning methods utilize the importance of each weight based on specified criteria only when searching for a sparse structure but do not utilize it during training. In this work, we propose a novel approach - \\textbf{M}agnitude \\textbf{A}ttention-based Dynamic \\textbf{P}runing (MAP) method, which applies the importance of weights throughout both the forward and backward paths to explore sparse model structures dynamically. Magnitude attention is defined based on the magnitude of weights as continuous real-valued numbers enabling a seamless transition from a redundant to an effective sparse network by promoting efficient exploration. Additionally, the attention mechanism ensures more effective updates for important layers within the sparse network. In later stages of training, our approach shifts from exploration to exploitation, exclusively updating the sparse model composed of crucial weights based on the explored structure, resulting in pruned models that not only achieve performance comparable to dense models but also outperform previous pruning methods on CIFAR-10/100 and ImageNet.", "authors": ["Jihye Back", "Namhyuk Ahn", "Jangho Kim"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-08", "url": "https://arxiv.org/abs/2306.05056", "pdf_url": "https://arxiv.org/pdf/2306.05056v1", "arxiv_id": "2306.05056", "doi": "10.48550/arXiv.2306.05056", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Expert systems with applications", "quality_score": 0.2698}
{"id": "debd00061c4c37dfc9ae7e7b29bc39ab351d3000d88144028baccbf03cc6c5c7", "sources": ["arxiv", "semantic_scholar"], "title": "ModuleFormer: Modularity Emerges from Mixture-of-Experts", "abstract": "Large Language Models (LLMs) have achieved remarkable results. However, existing models are expensive to train and deploy, and it is also difficult to expand their knowledge beyond pre-training data without forgetting previous knowledge. This paper proposes a new neural network architecture, ModuleFormer, that leverages modularity to improve the efficiency and flexibility of large language models. ModuleFormer is based on the Sparse Mixture of Experts (SMoE). Unlike the previous SMoE-based modular language model, which requires domain-labeled data to learn domain-specific experts, ModuleFormer can induce modularity from uncurated data with its new load balancing and concentration losses. ModuleFormer is a modular architecture that includes two different types of modules: new stick-breaking attention heads and feedforward experts. Different modules are sparsely activated conditions on the input token during training and inference. In our experiment, we found that the modular architecture enables three important abilities for large pre-trained language models: 1) Efficiency, since ModuleFormer only activates a subset of its modules for each input token, thus it could achieve the same performance as dense LLMs with more than two times throughput; 2) Extendability, ModuleFormer is more immune to catastrophic forgetting than dense LLMs and can be easily extended with new modules to learn new knowledge that is not included in the training data; 3) Specialisation, finetuning ModuleFormer could specialize a subset of modules to the finetuning task and the task-unrelated modules could be easily pruned for a lightweight deployment.", "authors": ["Yikang Shen", "Zheyu Zhang", "Tianyou Cao", "Shawn Tan", "Zhenfang Chen", "Chuang Gan"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-07", "url": "https://arxiv.org/abs/2306.04640", "pdf_url": "https://arxiv.org/pdf/2306.04640v2", "arxiv_id": "2306.04640", "doi": null, "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3076}
{"id": "f9e702d107f50de36a9af4dbbab53596e249b084e5b68c6463e9455797bea33b", "sources": ["arxiv", "semantic_scholar"], "title": "Patch-level Routing in Mixture-of-Experts is Provably Sample-efficient for Convolutional Neural Networks", "abstract": "In deep learning, mixture-of-experts (MoE) activates one or few experts (sub-networks) on a per-sample or per-token basis, resulting in significant computation reduction. The recently proposed \\underline{p}atch-level routing in \\underline{MoE} (pMoE) divides each input into $n$ patches (or tokens) and sends $l$ patches ($l\\ll n$) to each expert through prioritized routing. pMoE has demonstrated great empirical success in reducing training and inference costs while maintaining test accuracy. However, the theoretical explanation of pMoE and the general MoE remains elusive. Focusing on a supervised classification task using a mixture of two-layer convolutional neural networks (CNNs), we show for the first time that pMoE provably reduces the required number of training samples to achieve desirable generalization (referred to as the sample complexity) by a factor in the polynomial order of $n/l$, and outperforms its single-expert counterpart of the same or even larger capacity. The advantage results from the discriminative routing property, which is justified in both theory and practice that pMoE routers can filter label-irrelevant patches and route similar class-discriminative patches to the same expert. Our experimental results on MNIST, CIFAR-10, and CelebA support our theoretical findings on pMoE's generalization and show that pMoE can avoid learning spurious correlations.", "authors": ["Mohammed Nowaz Rabbani Chowdhury", "Shuai Zhang", "Meng Wang", "Sijia Liu", "Pin-Yu Chen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-07", "url": "https://arxiv.org/abs/2306.04073", "pdf_url": "https://arxiv.org/pdf/2306.04073v1", "arxiv_id": "2306.04073", "doi": "10.48550/arXiv.2306.04073", "citation_count": 47, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4203}
{"id": "a4f4dc20221a6012c3d649f41ceac3ee811d7661bfb1035209291391fe5f57f7", "sources": ["arxiv", "semantic_scholar"], "title": "LoRa Backscatter Communications: Temporal, Spectral, and Error Performance Analysis", "abstract": "LoRa backscatter (LB) communication systems can be considered as a potential candidate for ultra low power wide area networks (LPWAN) because of their low cost and low power consumption. In this paper, we comprehensively analyze LB modulation from various aspects, i.e., temporal, spectral, and error performance characteristics. First, we propose a signal model for LB signals that accounts for the limited number of loads in the tag. Then, we investigate the spectral properties of LB signals, obtaining a closed-form expression for the power spectrum. Finally, we derived the symbol error rate (SER) of LB with two decoders, i.e., the maximum likelihood (ML) and fast Fourier transform (FFT) decoders, in both additive white Gaussian noise (AWGN) and double Nakagami-m fading channels. The spectral analysis shows that out-of-band emissions for LB satisfy the European Telecommunications Standards Institute (ETSI) regulation only when considering a relatively large number of loads. For the error performance, unlike conventional LoRa, the FFT decoder is not optimal. Nevertheless, the ML decoder can achieve a performance similar to conventional LoRa with a moderate number of loads.", "authors": ["Ganghui Lin", "Ahmed Elzanaty", "Mohamed-Slim Alouini"], "categories": ["cs.IT", "eess.SP"], "fields_of_study": ["Computer Science", "Engineering", "Mathematics"], "published_date": "2023-06-04", "url": "https://arxiv.org/abs/2306.02323", "pdf_url": "https://arxiv.org/pdf/2306.02323v2", "arxiv_id": "2306.02323", "doi": "10.1109/JIOT.2023.3268113", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SlinGovie/LoRa-Backscatter-Performance-Analysis", "venue": "IEEE Internet of Things Journal", "quality_score": 0.2698}
{"id": "a601753cffa435606ae97bcf3221dcb1e671375a595c0f2c61b9c99852da13e2", "sources": ["arxiv", "semantic_scholar"], "title": "Group channel pruning and spatial attention distilling for object detection", "abstract": "Due to the over-parameterization of neural networks, many model compression methods based on pruning and quantization have emerged. They are remarkable in reducing the size, parameter number, and computational complexity of the model. However, most of the models compressed by such methods need the support of special hardware and software, which increases the deployment cost. Moreover, these methods are mainly used in classification tasks, and rarely directly used in detection tasks. To address these issues, for the object detection network we introduce a three-stage model compression method: dynamic sparse training, group channel pruning, and spatial attention distilling. Firstly, to select out the unimportant channels in the network and maintain a good balance between sparsity and accuracy, we put forward a dynamic sparse training method, which introduces a variable sparse rate, and the sparse rate will change with the training process of the network. Secondly, to reduce the effect of pruning on network accuracy, we propose a novel pruning method called group channel pruning. In particular, we divide the network into multiple groups according to the scales of the feature layer and the similarity of module structure in the network, and then we use different pruning thresholds to prune the channels in each group. Finally, to recover the accuracy of the pruned network, we use an improved knowledge distillation method for the pruned network. Especially, we extract spatial attention information from the feature maps of specific scales in each group as knowledge for distillation. In the experiments, we use YOLOv4 as the object detection network and PASCAL VOC as the training dataset. Our method reduces the parameters of the model by 64.7 % and the calculation by 34.9%.", "authors": ["Yun Chu", "Pu Li", "Yong Bai", "Zhuhua Hu", "Yongqing Chen", "Jiafeng Lu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-02", "url": "https://arxiv.org/abs/2306.01526", "pdf_url": "https://arxiv.org/pdf/2306.01526v1", "arxiv_id": "2306.01526", "doi": "10.1007/s10489-022-03293-x", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "[J]. Applied Intelligence, 2022: 1-19", "quality_score": 0.3076}
{"id": "cf66082b7ecb5f353925c0d65792d25d1f3a2bf168eae1d478b78ea5d5e17dca", "sources": ["arxiv", "semantic_scholar"], "title": "Faster Causal Attention Over Large Sequences Through Sparse Flash Attention", "abstract": "Transformer-based language models have found many diverse applications requiring them to process sequences of increasing length. For these applications, the causal self-attention -- which is the only component scaling quadratically w.r.t. the sequence length -- becomes a central concern. While many works have proposed schemes to sparsify the attention patterns and reduce the computational overhead of self-attention, those are often limited by implementations concerns and end up imposing a simple and static structure over the attention matrix. Conversely, implementing more dynamic sparse attentions often results in runtimes significantly slower than computing the full attention using the Flash implementation from Dao et al. (2022). We extend FlashAttention to accommodate a large class of attention sparsity patterns that, in particular, encompass key/query dropping and hashing-based attention. This leads to implementations with no computational complexity overhead and a multi-fold runtime speedup on top of FlashAttention. Even with relatively low degrees of sparsity, our method improves visibly upon FlashAttention as the sequence length increases. Without sacrificing perplexity, we increase the training speed of a transformer language model by $2.0\\times$ and $3.3\\times$ for sequences of respectively $8k$ and $16k$ tokens.", "authors": ["Matteo Pagliardini", "Daniele Paliotta", "Martin Jaggi", "François Fleuret"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-01", "url": "https://arxiv.org/abs/2306.01160", "pdf_url": "https://arxiv.org/pdf/2306.01160v1", "arxiv_id": "2306.01160", "doi": "10.48550/arXiv.2306.01160", "citation_count": 30, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3728}
{"id": "9d74eca9014d9e960dd992b90fd4022d95755914497ed4c8ae77cba42573c20b", "sources": ["arxiv", "semantic_scholar"], "title": "Divide, Conquer, and Combine: Mixture of Semantic-Independent Experts for Zero-Shot Dialogue State Tracking", "abstract": "Zero-shot transfer learning for Dialogue State Tracking (DST) helps to handle a variety of task-oriented dialogue domains without the cost of collecting in-domain data. Existing works mainly study common data- or model-level augmentation methods to enhance the generalization but fail to effectively decouple the semantics of samples, limiting the zero-shot performance of DST. In this paper, we present a simple and effective \"divide, conquer and combine\" solution, which explicitly disentangles the semantics of seen data, and leverages the performance and robustness with the mixture-of-experts mechanism. Specifically, we divide the seen data into semantically independent subsets and train corresponding experts, the newly unseen samples are mapped and inferred with mixture-of-experts with our designed ensemble inference. Extensive experiments on MultiWOZ2.1 upon the T5-Adapter show our schema significantly and consistently improves the zero-shot performance, achieving the SOTA on settings without external knowledge, with only 10M trainable parameters1.", "authors": ["Qingyue Wang", "Liang Ding", "Yanan Cao", "Yibing Zhan", "Zheng Lin", "Shi Wang", "Dacheng Tao", "Li Guo"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-01", "url": "https://arxiv.org/abs/2306.00434", "pdf_url": "https://arxiv.org/pdf/2306.00434v1", "arxiv_id": "2306.00434", "doi": "10.48550/arXiv.2306.00434", "citation_count": 22, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3404}
{"id": "770470d55548aee9f8de810aac5f868ddc32db99ad6b5ea35af3b61fbfb19460", "sources": ["arxiv", "semantic_scholar"], "title": "Primal-Attention: Self-attention through Asymmetric Kernel SVD in Primal Representation", "abstract": "Recently, a new line of works has emerged to understand and improve self-attention in Transformers by treating it as a kernel machine. However, existing works apply the methods for symmetric kernels to the asymmetric self-attention, resulting in a nontrivial gap between the analytical understanding and numerical implementation. In this paper, we provide a new perspective to represent and optimize self-attention through asymmetric Kernel Singular Value Decomposition (KSVD), which is also motivated by the low-rank property of self-attention normally observed in deep layers. Through asymmetric KSVD, $i$) a primal-dual representation of self-attention is formulated, where the optimization objective is cast to maximize the projection variances in the attention outputs; $ii$) a novel attention mechanism, i.e., Primal-Attention, is proposed via the primal representation of KSVD, avoiding explicit computation of the kernel matrix in the dual; $iii$) with KKT conditions, we prove that the stationary solution to the KSVD optimization in Primal-Attention yields a zero-value objective. In this manner, KSVD optimization can be implemented by simply minimizing a regularization loss, so that low-rank property is promoted without extra decomposition. Numerical experiments show state-of-the-art performance of our Primal-Attention with improved efficiency. Moreover, we demonstrate that the deployed KSVD optimization regularizes Primal-Attention with a sharper singular value decay than that of the canonical self-attention, further verifying the great potential of our method. To the best of our knowledge, this is the first work that provides a primal-dual representation for the asymmetric kernel in self-attention and successfully applies it to modeling and optimization.", "authors": ["Yingyi Chen", "Qinghua Tao", "Francesco Tonin", "Johan A. K. Suykens"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-31", "url": "https://arxiv.org/abs/2305.19798", "pdf_url": "https://arxiv.org/pdf/2305.19798v2", "arxiv_id": "2305.19798", "doi": "10.48550/arXiv.2305.19798", "citation_count": 37, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3949}
{"id": "eb76aaad1c86f12bee72e3b24b4fc348a9091ad9251b8730f9bdce4f949c1695", "sources": ["arxiv", "semantic_scholar"], "title": "LoRAPrune: Structured Pruning Meets Low-Rank Parameter-Efficient Fine-Tuning", "abstract": "Large Language Models (LLMs), such as LLaMA and T5, have shown exceptional performance across various tasks through fine-tuning. Although low-rank adaption (LoRA) has emerged to cheaply fine-tune these LLMs on downstream tasks, their deployment is still hindered by the vast model scale and computational costs. Post-training model pruning offers a way to compress LLMs. However, the current pruning methods designed for LLMs are not compatible with LoRA. This is due to their utilization of unstructured pruning on LLMs, impeding the merging of LoRA weights, or their dependence on the gradients of pre-trained weights to guide pruning, which can impose significant memory overhead. To this end, we propose LoRAPrune, a new framework that delivers an accurate structured pruned model in a highly memory-efficient manner. Specifically, we first design a LoRA-guided pruning criterion, which uses the weights and gradients of LoRA, rather than the gradients of pre-trained weights for importance estimation. We subsequently integrate this criterion into an iterative pruning process, effectively removing redundant channels and heads. Extensive experimental results demonstrate the superior performance of our LoRAPrune over existing approaches on the LLaMA series models. At a 50\\% compression rate, LoRAPrune demonstrates superior performance over LLM-Pruner, achieving a reduction in perplexity by 4.81 on WikiText2 and 3.46 on PTB, while also decreasing memory usage by 52.6%. Besides, LoRAPrune also matches semi-structural pruning across multiple LLMs, proving its wide applicability. The code is available at https://github.com/aim-uofa/LoRAPrune.", "authors": ["Mingyang Zhang", "Hao Chen", "Chunhua Shen", "Zhen Yang", "Linlin Ou", "Xinyi Yu", "Bohan Zhuang"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-28", "url": "https://arxiv.org/abs/2305.18403", "pdf_url": "https://arxiv.org/pdf/2305.18403v5", "arxiv_id": "2305.18403", "doi": "10.18653/v1/2024.findings-acl.178", "citation_count": 88, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/aim-uofa/LoRAPrune", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4873}
{"id": "0feed82ca942ccc503d77e8d7ed0f1eea9931395312b7768ce99526952505bfd", "sources": ["arxiv", "semantic_scholar"], "title": "Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers", "abstract": "Deployment of Transformer models on edge devices is becoming increasingly challenging due to the exponentially growing inference cost that scales quadratically with the number of tokens in the input sequence. Token pruning is an emerging solution to address this challenge due to its ease of deployment on various Transformer backbones. However, most token pruning methods require computationally expensive fine-tuning, which is undesirable in many edge deployment cases. In this work, we propose Zero-TPrune, the first zero-shot method that considers both the importance and similarity of tokens in performing token pruning. It leverages the attention graph of pre-trained Transformer models to produce an importance distribution for tokens via our proposed Weighted Page Rank (WPR) algorithm. This distribution further guides token partitioning for efficient similarity-based pruning. Due to the elimination of the fine-tuning overhead, Zero-TPrune can prune large models at negligible computational cost, switch between different pruning configurations at no computational cost, and perform hyperparameter tuning efficiently. We evaluate the performance of Zero-TPrune on vision tasks by applying it to various vision Transformer backbones and testing them on ImageNet. Without any fine-tuning, Zero-TPrune reduces the FLOPs cost of DeiT-S by 34.7% and improves its throughput by 45.3% with only 0.4% accuracy loss. Compared with state-of-the-art pruning methods that require fine-tuning, Zero-TPrune not only eliminates the need for fine-tuning after pruning but also does so with only 0.1% accuracy loss. Compared with state-of-the-art fine-tuning-free pruning methods, Zero-TPrune reduces accuracy loss by up to 49% with similar FLOPs budgets. Project webpage: https://jha-lab.github.io/zerotprune.", "authors": ["Hongjie Wang", "Bhishma Dedhia", "Niraj K. Jha"], "categories": ["cs.CV", "cs.AI", "cs.LG", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-05-27", "url": "https://arxiv.org/abs/2305.17328", "pdf_url": "https://arxiv.org/pdf/2305.17328v3", "arxiv_id": "2305.17328", "doi": "10.1109/CVPR52733.2024.01521", "citation_count": 74, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.4688}
{"id": "f00882d2b172d6396638b46eb23c4346f84791d24363bbfc5725d46e0846df10", "sources": ["arxiv", "semantic_scholar"], "title": "Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Time", "abstract": "Large language models(LLMs) have sparked a new wave of exciting AI applications. Hosting these models at scale requires significant memory resources. One crucial memory bottleneck for the deployment stems from the context window. It is commonly recognized that model weights are memory hungry; however, the size of key-value embedding stored during the generation process (KV cache) can easily surpass the model size. The enormous size of the KV cache puts constraints on the inference batch size, which is crucial for high throughput inference workload. Inspired by an interesting observation of the attention scores, we hypothesize the persistence of importance: only pivotal tokens, which had a substantial influence at one step, will significantly influence future generations. Based on our empirical verification and theoretical analysis around this hypothesis, we propose Scissorhands, a system that maintains the memory usage of the KV cache at a fixed budget without finetuning the model. In essence, Scissorhands manages the KV cache by storing the pivotal tokens with a higher probability. We validate that Scissorhands reduces the inference memory usage of the KV cache by up to 5X without compromising model quality. We further demonstrate that Scissorhands can be combined with 4-bit quantization, traditionally used to compress model weights, to achieve up to 20X compression.", "authors": ["Zichang Liu", "Aditya Desai", "Fangshuo Liao", "Weitao Wang", "Victor Xie", "Zhaozhuo Xu", "Anastasios Kyrillidis", "Anshumali Shrivastava"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-26", "url": "https://arxiv.org/abs/2305.17118", "pdf_url": "https://arxiv.org/pdf/2305.17118v2", "arxiv_id": "2305.17118", "doi": "10.48550/arXiv.2305.17118", "citation_count": 436, "influential_citation_count": 28, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.7312}
{"id": "9ec08b65787c18cdf632db411d62d743f3c96f0a57383bd5d7ce08ef4b1061da", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Expert Conformer for Streaming Multilingual ASR", "abstract": "End-to-end models with large capacity have significantly improved multilingual automatic speech recognition, but their computation cost poses challenges for on-device applications. We propose a streaming truly multilingual Conformer incorporating mixture-of-expert (MoE) layers that learn to only activate a subset of parameters in training and inference. The MoE layer consists of a softmax gate which chooses the best two experts among many in forward propagation. The proposed MoE layer offers efficient inference by activating a fixed number of parameters as the number of experts increases. We evaluate the proposed model on a set of 12 languages, and achieve an average 11.9% relative improvement in WER over the baseline. Compared to an adapter model using ground truth information, our MoE model achieves similar WER and activates similar number of parameters but without any language information. We further show around 3% relative WER improvement by multilingual shallow fusion.", "authors": ["Ke Hu", "Bo Li", "Tara N. Sainath", "Yu Zhang", "Francoise Beaufays"], "categories": ["cs.CL", "cs.SD", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-05-25", "url": "https://arxiv.org/abs/2305.15663", "pdf_url": "https://arxiv.org/pdf/2305.15663v1", "arxiv_id": "2305.15663", "doi": "10.48550/arXiv.2305.15663", "citation_count": 30, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Interspeech", "quality_score": 0.3728}
{"id": "f4762d9b0573d78cdd7222133a5079e9597011ad5d54030807a106616bfc5d1d", "sources": ["arxiv", "semantic_scholar"], "title": "Modeling Task Relationships in Multi-variate Soft Sensor with Balanced Mixture-of-Experts", "abstract": "Accurate estimation of multiple quality variables is critical for building industrial soft sensor models, which have long been confronted with data efficiency and negative transfer issues. Methods sharing backbone parameters among tasks address the data efficiency issue; however, they still fail to mitigate the negative transfer problem. To address this issue, a balanced Mixture-of-Experts (BMoE) is proposed in this work, which consists of a multi-gate mixture of experts (MMoE) module and a task gradient balancing (TGB) module. The MoE module aims to portray task relationships, while the TGB module balances the gradients among tasks dynamically. Both of them cooperate to mitigate the negative transfer problem. Experiments on the typical sulfur recovery unit demonstrate that BMoE models task relationship and balances the training process effectively, and achieves better performance than baseline models significantly.", "authors": ["Yuxin Huang", "Hao Wang", "Zhaoran Liu", "Licheng Pan", "Haozhe Li", "Xinggao Liu"], "categories": ["cs.LG", "cs.CE", "stat.AP"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-05-25", "url": "https://arxiv.org/abs/2305.16360", "pdf_url": "https://arxiv.org/pdf/2305.16360v1", "arxiv_id": "2305.16360", "doi": "10.1109/TII.2022.3202909", "citation_count": 25, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Industrial Informatics", "quality_score": 0.3537}
{"id": "872e45644b5cfa0114a8bfe7a167021295d49e19554ba738d87c68bf887a6fc5", "sources": ["arxiv", "semantic_scholar"], "title": "SmartTrim: Adaptive Tokens and Attention Pruning for Efficient Vision-Language Models", "abstract": "Despite achieving remarkable performance on various vision-language tasks, Transformer-based Vision-Language Models (VLMs) suffer from redundancy in inputs and parameters, significantly hampering their efficiency in real-world applications. Moreover, the degree of redundancy in token representations and model parameters, such as attention heads, varies significantly for different inputs. In light of the challenges, we propose SmartTrim, an adaptive acceleration framework for VLMs, which adjusts the computational overhead per instance. Specifically, we integrate lightweight modules into the original backbone to identify and prune redundant token representations and attention heads within each layer. Furthermore, we devise a self-distillation strategy to enhance the consistency between the predictions of the pruned model and its fully-capacity counterpart. Experimental results across various vision-language tasks consistently demonstrate that SmartTrim accelerates the original model by 2-3 times with minimal performance degradation, highlighting the effectiveness and efficiency compared to previous approaches. Code will be available at https://github.com/kugwzk/SmartTrim.", "authors": ["Zekun Wang", "Jingchang Chen", "Wangchunshu Zhou", "Haichao Zhu", "Jiafeng Liang", "Liping Shan", "Ming Liu", "Dongliang Xu", "Qing Yang", "Bing Qin"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-24", "url": "https://arxiv.org/abs/2305.15033", "pdf_url": "https://arxiv.org/pdf/2305.15033v2", "arxiv_id": "2305.15033", "doi": "10.63317/23hifgwadz5j", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/kugwzk/SmartTrim", "venue": "International Conference on Language Resources and Evaluation", "quality_score": 0.2785}
{"id": "a943ec9fb2f33a8d02db05b65cc79e1afe462bb564edca110be35bc9ac7eb0b8", "sources": ["arxiv", "semantic_scholar"], "title": "Getting MoRE out of Mixture of Language Model Reasoning Experts", "abstract": "While recent large language models (LLMs) improve on various question answering (QA) datasets, it remains difficult for a single model to generalize across question types that require distinct reasoning abilities. We provide empirical evidence that state-of-the-art LLMs suffer from poor generalizability on reasoning types beyond those seen in the prompt. To remedy this, we propose a Mixture-of-Reasoning-Experts (MoRE) framework that ensembles diverse specialized language models. We specialize the backbone language model with prompts optimized for different reasoning categories, including factual, multihop, mathematical, and commonsense reasoning. Our key insight is to leverage agreement among the specialized experts to select the best answer for each question, or to abstain from answering. This gives MoRE higher accuracy than any single specialized model on a collection of 12 QA datasets from four reasoning types. Beyond generalizability, the interpretable design of MoRE improves selective question answering results compared to baselines without incorporating inter-expert agreement. This framework is also more interpretable and useful to human consumers of QA outputs. Our human study confirms that presenting expert predictions and the answer selection process helps annotators more accurately calibrate when to trust the system's output. We release all code and data to facilitate future work.", "authors": ["Chenglei Si", "Weijia Shi", "Chen Zhao", "Luke Zettlemoyer", "Jordan Boyd-Graber"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-24", "url": "https://arxiv.org/abs/2305.14628", "pdf_url": "https://arxiv.org/pdf/2305.14628v2", "arxiv_id": "2305.14628", "doi": "10.18653/v1/2023.findings-emnlp.552", "citation_count": 51, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.429}
{"id": "35e664ba5a87afdc91141b8b731e340e255d2ff759905e089c3da58f1fe3abe1", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Convergence Rates for Parameter Estimation in Gaussian-gated Mixture of Experts", "abstract": "Originally introduced as a neural network for ensemble learning, mixture of experts (MoE) has recently become a fundamental building block of highly successful modern deep neural networks for heterogeneous data analysis in several applications of machine learning and statistics. Despite its popularity in practice, a satisfactory level of theoretical understanding of the MoE model is far from complete. To shed new light on this problem, we provide a convergence analysis for maximum likelihood estimation (MLE) in the Gaussian-gated MoE model. The main challenge of that analysis comes from the inclusion of covariates in the Gaussian gating functions and expert networks, which leads to their intrinsic interaction via some partial differential equations with respect to their parameters. We tackle these issues by designing novel Voronoi loss functions among parameters to accurately capture the heterogeneity of parameter estimation rates. Our findings reveal that the MLE has distinct behaviors under two complement settings of location parameters of the Gaussian gating functions, namely when all these parameters are non-zero versus when at least one among them vanishes. Notably, these behaviors can be characterized by the solvability of two different systems of polynomial equations. Finally, we conduct a simulation study to empirically verify our theoretical results.", "authors": ["Huy Nguyen", "TrungTin Nguyen", "Khai Nguyen", "Nhat Ho"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-05-12", "url": "https://arxiv.org/abs/2305.07572", "pdf_url": "https://arxiv.org/pdf/2305.07572v2", "arxiv_id": "2305.07572", "doi": "10.48550/arXiv.2305.07572", "citation_count": 25, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.3537}
{"id": "254d38d03a1db483fd66787a187769dcc432dbb553da6bb0ebea6880497eedfa", "sources": ["arxiv", "semantic_scholar"], "title": "Demystifying Softmax Gating Function in Gaussian Mixture of Experts", "abstract": "Understanding the parameter estimation of softmax gating Gaussian mixture of experts has remained a long-standing open problem in the literature. It is mainly due to three fundamental theoretical challenges associated with the softmax gating function: (i) the identifiability only up to the translation of parameters; (ii) the intrinsic interaction via partial differential equations between the softmax gating and the expert functions in the Gaussian density; (iii) the complex dependence between the numerator and denominator of the conditional density of softmax gating Gaussian mixture of experts. We resolve these challenges by proposing novel Voronoi loss functions among parameters and establishing the convergence rates of maximum likelihood estimator (MLE) for solving parameter estimation in these models. When the true number of experts is unknown and over-specified, our findings show a connection between the convergence rate of the MLE and a solvability problem of a system of polynomial equations.", "authors": ["Huy Nguyen", "TrungTin Nguyen", "Nhat Ho"], "categories": ["stat.ML", "cs.LG", "math.ST"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-05-05", "url": "https://arxiv.org/abs/2305.03288", "pdf_url": "https://arxiv.org/pdf/2305.03288v2", "arxiv_id": "2305.03288", "doi": "10.52202/075280-0206", "citation_count": 40, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4032}
{"id": "c0dbe5fe8025998a0fcc6f2fc08ac1c8175c270a49d42ce88e99c1af246035bd", "sources": ["arxiv", "semantic_scholar"], "title": "Mixtures of Gaussian process experts based on kernel stick-breaking processes", "abstract": "Mixtures of Gaussian process experts is a class of models that can simultaneously address two of the key limitations inherent in standard Gaussian processes: scalability and predictive performance. In particular, models that use Dirichlet processes as gating functions permit straightforward interpretation and automatic selection of the number of experts in a mixture. While the existing models are intuitive and capable of capturing non-stationarity, multi-modality and heteroskedasticity, the simplicity of their gating functions may limit the predictive performance when applied to complex data-generating processes. Capitalising on the recent advancement in the dependent Dirichlet processes literature, we propose a new mixture model of Gaussian process experts based on kernel stick-breaking processes. Our model maintains the intuitive appeal yet improve the performance of the existing models. To make it practical, we design a sampler for posterior computation based on the slice sampling. The model behaviour and improved predictive performance are demonstrated in experiments using six datasets.", "authors": ["Yuji Saikai", "Khue-Dung Dang"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-04-26", "url": "https://arxiv.org/abs/2304.13833", "pdf_url": "https://arxiv.org/pdf/2304.13833v2", "arxiv_id": "2304.13833", "doi": "10.48550/arXiv.2304.13833", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "c87caa01963e9cff045f9972305f914aeb5458b878c5238eb310b94c0182cf8a", "sources": ["arxiv", "semantic_scholar"], "title": "Revisiting Single-gated Mixtures of Experts", "abstract": "Mixture of Experts (MoE) are rising in popularity as a means to train extremely large-scale models, yet allowing for a reasonable computational cost at inference time. Recent state-of-the-art approaches usually assume a large number of experts, and require training all experts jointly, which often lead to training instabilities such as the router collapsing In contrast, in this work, we propose to revisit the simple single-gate MoE, which allows for more practical training. Key to our work are (i) a base model branch acting both as an early-exit and an ensembling regularization scheme, (ii) a simple and efficient asynchronous training pipeline without router collapse issues, and finally (iii) a per-sample clustering-based initialization. We show experimentally that the proposed model obtains efficiency-to-accuracy trade-offs comparable with other more complex MoE, and outperforms non-mixture baselines. This showcases the merits of even a simple single-gate MoE, and motivates further exploration in this area.", "authors": ["Amelie Royer", "Ilia Karmanov", "Andrii Skliar", "Babak Ehteshami Bejnordi", "Tijmen Blankevoort"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-11", "url": "https://arxiv.org/abs/2304.05497", "pdf_url": "https://arxiv.org/pdf/2304.05497v1", "arxiv_id": "2304.05497", "doi": "10.48550/arXiv.2304.05497", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "British Machine Vision Conference", "quality_score": 0.2785}
{"id": "5e2c92a916817ef347d2303d4490cb622eea52e6bce93358c048921686ddb2fa", "sources": ["arxiv", "semantic_scholar"], "title": "Graph Mixture of Experts: Learning on Large-Scale Graphs with Explicit Diversity Modeling", "abstract": "Graph neural networks (GNNs) have found extensive applications in learning from graph data. However, real-world graphs often possess diverse structures and comprise nodes and edges of varying types. To bolster the generalization capacity of GNNs, it has become customary to augment training graph structures through techniques like graph augmentations and large-scale pre-training on a wider array of graphs. Balancing this diversity while avoiding increased computational costs and the notorious trainability issues of GNNs is crucial. This study introduces the concept of Mixture-of-Experts (MoE) to GNNs, with the aim of augmenting their capacity to adapt to a diverse range of training graph structures, without incurring explosive computational overhead. The proposed Graph Mixture of Experts (GMoE) model empowers individual nodes in the graph to dynamically and adaptively select more general information aggregation experts. These experts are trained to capture distinct subgroups of graph structures and to incorporate information with varying hop sizes, where those with larger hop sizes specialize in gathering information over longer distances. The effectiveness of GMoE is validated through a series of experiments on a diverse set of tasks, including graph, node, and link prediction, using the OGB benchmark. Notably, it enhances ROC-AUC by $1.81\\%$ in ogbg-molhiv and by $1.40\\%$ in ogbg-molbbbp, when compared to the non-MoE baselines. Our code is publicly available at https://github.com/VITA-Group/Graph-Mixture-of-Experts.", "authors": ["Haotao Wang", "Ziyu Jiang", "Yuning You", "Yan Han", "Gaowen Liu", "Jayanth Srinivasa", "Ramana Rao Kompella", "Zhangyang Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-06", "url": "https://arxiv.org/abs/2304.02806", "pdf_url": "https://arxiv.org/pdf/2304.02806v2", "arxiv_id": "2304.02806", "doi": "10.48550/arXiv.2304.02806", "citation_count": 90, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/VITA-Group/Graph-Mixture-of-Experts", "venue": "Neural Information Processing Systems", "quality_score": 0.4898}
{"id": "e55b428d912978cb163ce4adf560def6311832f5841db916ae15b8ec899f65e8", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Map Guided Transformer Pruning for Edge Device", "abstract": "Due to its significant capability of modeling long-range dependencies, vision transformer (ViT) has achieved promising success in both holistic and occluded person re-identification (Re-ID) tasks. However, the inherent problems of transformers such as the huge computational cost and memory footprint are still two unsolved issues that will block the deployment of ViT based person Re-ID models on resource-limited edge devices. Our goal is to reduce both the inference complexity and model size without sacrificing the comparable accuracy on person Re-ID, especially for tasks with occlusion. To this end, we propose a novel attention map guided (AMG) transformer pruning method, which removes both redundant tokens and heads with the guidance of the attention map in a hardware-friendly way. We first calculate the entropy in the key dimension and sum it up for the whole map, and the corresponding head parameters of maps with high entropy will be removed for model size reduction. Then we combine the similarity and first-order gradients of key tokens along the query dimension for token importance estimation and remove redundant key and value tokens to further reduce the inference complexity. Comprehensive experiments on Occluded DukeMTMC and Market-1501 demonstrate the effectiveness of our proposals. For example, our proposed pruning strategy on ViT-Base enjoys \\textup{\\textbf{29.4\\%}} \\textup{\\textbf{FLOPs}} savings with \\textup{\\textbf{0.2\\%}} drop on Rank-1 and \\textup{\\textbf{0.4\\%}} improvement on mAP, respectively.", "authors": ["Junzhu Mao", "Yazhou Yao", "Zeren Sun", "Xingguo Huang", "Fumin Shen", "Heng-Tao Shen"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-04", "url": "https://arxiv.org/abs/2304.01452", "pdf_url": "https://arxiv.org/pdf/2304.01452v1", "arxiv_id": "2304.01452", "doi": "10.48550/arXiv.2304.01452", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "75b6f4877141858d8a0c501a2426f9ccd32ff93c0319da14270d46cf045b2915", "sources": ["arxiv", "semantic_scholar"], "title": "Dual Cross-Attention for Medical Image Segmentation", "abstract": "We propose Dual Cross-Attention (DCA), a simple yet effective attention module that is able to enhance skip-connections in U-Net-based architectures for medical image segmentation. DCA addresses the semantic gap between encoder and decoder features by sequentially capturing channel and spatial dependencies across multi-scale encoder features. First, the Channel Cross-Attention (CCA) extracts global channel-wise dependencies by utilizing cross-attention across channel tokens of multi-scale encoder features. Then, the Spatial Cross-Attention (SCA) module performs cross-attention to capture spatial dependencies across spatial tokens. Finally, these fine-grained encoder features are up-sampled and connected to their corresponding decoder parts to form the skip-connection scheme. Our proposed DCA module can be integrated into any encoder-decoder architecture with skip-connections such as U-Net and its variants. We test our DCA module by integrating it into six U-Net-based architectures such as U-Net, V-Net, R2Unet, ResUnet++, DoubleUnet and MultiResUnet. Our DCA module shows Dice Score improvements up to 2.05% on GlaS, 2.74% on MoNuSeg, 1.37% on CVC-ClinicDB, 1.12% on Kvasir-Seg and 1.44% on Synapse datasets. Our codes are available at: https://github.com/gorkemcanates/Dual-Cross-Attention", "authors": ["Gorkem Can Ates", "Prasoon Mohan", "Emrah Celik"], "categories": ["cs.CV", "cs.LG", "eess.IV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-03-30", "url": "https://arxiv.org/abs/2303.17696", "pdf_url": "https://arxiv.org/pdf/2303.17696v1", "arxiv_id": "2303.17696", "doi": "10.1016/j.engappai.2023.107139", "citation_count": 175, "influential_citation_count": 3, "has_code": true, "code_url": "https://github.com/gorkemcanates/Dual-Cross-Attention", "venue": "Engineering applications of artificial intelligence", "quality_score": 0.5614}
{"id": "a6a9c556e10cf7a6c11754e0e436633631c1cfa2549b2c62b471016073e8be80", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Second-Order Attentive Context for Efficient Correspondence Pruning", "abstract": "Correspondence pruning aims to search consistent correspondences (inliers) from a set of putative correspondences. It is challenging because of the disorganized spatial distribution of numerous outliers, especially when putative correspondences are largely dominated by outliers. It's more challenging to ensure effectiveness while maintaining efficiency. In this paper, we propose an effective and efficient method for correspondence pruning. Inspired by the success of attentive context in correspondence problems, we first extend the attentive context to the first-order attentive context and then introduce the idea of attention in attention (ANA) to model second-order attentive context for correspondence pruning. Compared with first-order attention that focuses on feature-consistent context, second-order attention dedicates to attention weights itself and provides an additional source to encode consistent context from the attention map. For efficiency, we derive two approximate formulations for the naive implementation of second-order attention to optimize the cubic complexity to linear complexity, such that second-order attention can be used with negligible computational overheads. We further implement our formulations in a second-order context layer and then incorporate the layer in an ANA block. Extensive experiments demonstrate that our method is effective and efficient in pruning outliers, especially in high-outlier-ratio cases. Compared with the state-of-the-art correspondence pruning approach LMCNet, our method runs 14 times faster while maintaining a competitive accuracy.", "authors": ["Xinyi Ye", "Weiyue Zhao", "Hao Lu", "Zhiguo Cao"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-28", "url": "https://arxiv.org/abs/2303.15761", "pdf_url": "https://arxiv.org/pdf/2303.15761v1", "arxiv_id": "2303.15761", "doi": "10.1609/aaai.v37i3.25431", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2785}
{"id": "e205aee5dbc115c7a34d161302047035eaac90bdf6d1fec94e6b8d2c55a13697", "sources": ["arxiv", "semantic_scholar"], "title": "Vision Transformer with Quadrangle Attention", "abstract": "Window-based attention has become a popular choice in vision transformers due to its superior performance, lower computational complexity, and less memory footprint. However, the design of hand-crafted windows, which is data-agnostic, constrains the flexibility of transformers to adapt to objects of varying sizes, shapes, and orientations. To address this issue, we propose a novel quadrangle attention (QA) method that extends the window-based attention to a general quadrangle formulation. Our method employs an end-to-end learnable quadrangle regression module that predicts a transformation matrix to transform default windows into target quadrangles for token sampling and attention calculation, enabling the network to model various targets with different shapes and orientations and capture rich context information. We integrate QA into plain and hierarchical vision transformers to create a new architecture named QFormer, which offers minor code modifications and negligible extra computational cost. Extensive experiments on public benchmarks demonstrate that QFormer outperforms existing representative vision transformers on various vision tasks, including classification, object detection, semantic segmentation, and pose estimation. The code will be made publicly available at \\href{https://github.com/ViTAE-Transformer/QFormer}{QFormer}.", "authors": ["Qiming Zhang", "Jing Zhang", "Yufei Xu", "Dacheng Tao"], "categories": ["cs.CV"], "fields_of_study": ["Medicine", "Computer Science"], "published_date": "2023-03-27", "url": "https://arxiv.org/abs/2303.15105", "pdf_url": "https://arxiv.org/pdf/2303.15105v1", "arxiv_id": "2303.15105", "doi": "10.1109/TPAMI.2023.3347693", "citation_count": 81, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/ViTAE-Transformer/QFormer}{QFormer}", "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.4785}
{"id": "ef8b17e7bfc3ab6288e11243b21bb844b7e8626b48eb9a42d091efded70c6a51", "sources": ["arxiv", "semantic_scholar"], "title": "Automatic Attention Pruning: Improving and Automating Model Pruning using Attentions", "abstract": "Pruning is a promising approach to compress deep learning models in order to deploy them on resource-constrained edge devices. However, many existing pruning solutions are based on unstructured pruning, which yields models that cannot efficiently run on commodity hardware; and they often require users to manually explore and tune the pruning process, which is time-consuming and often leads to sub-optimal results. To address these limitations, this paper presents Automatic Attention Pruning (AAP), an adaptive, attention-based, structured pruning approach to automatically generate small, accurate, and hardware-efficient models that meet user objectives. First, it proposes iterative structured pruning using activation-based attention maps to effectively identify and prune unimportant filters. Then, it proposes adaptive pruning policies for automatically meeting the pruning objectives of accuracy-critical, memory-constrained, and latency-sensitive tasks. A comprehensive evaluation shows that AAP substantially outperforms the state-of-the-art structured pruning works for a variety of model architectures. Our code is at: https://github.com/kaiqi123/Automatic-Attention-Pruning.git.", "authors": ["Kaiqi Zhao", "Animesh Jain", "Ming Zhao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-14", "url": "https://arxiv.org/abs/2303.08595", "pdf_url": "https://arxiv.org/pdf/2303.08595v1", "arxiv_id": "2303.08595", "doi": "10.48550/arXiv.2303.08595", "citation_count": 26, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/kaiqi123/Automatic-Attention-Pruning.git", "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.3578}
{"id": "88664b6ec4aee1f109552fba913030581bd317e4f2e6a7f64894d34bd8ec2552", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Expert Specialization in Mixture of Experts", "abstract": "Mixture of experts (MoE), introduced over 20 years ago, is the simplest gated modular neural network architecture. There is renewed interest in MoE because the conditional computation allows only parts of the network to be used during each inference, as was recently demonstrated in large scale natural language processing models. MoE is also of potential interest for continual learning, as experts may be reused for new tasks, and new experts introduced. The gate in the MoE architecture learns task decompositions and individual experts learn simpler functions appropriate to the gate's decomposition. In this paper: (1) we show that the original MoE architecture and its training method do not guarantee intuitive task decompositions and good expert utilization, indeed they can fail spectacularly even for simple data such as MNIST and FashionMNIST; (2) we introduce a novel gating architecture, similar to attention, that improves performance and results in a lower entropy task decomposition; and (3) we introduce a novel data-driven regularization that improves expert specialization. We empirically validate our methods on MNIST, FashionMNIST and CIFAR-100 datasets.", "authors": ["Yamuna Krishnamurthy", "Chris Watkins", "Thomas Gaertner"], "categories": ["cs.LG", "cs.AI", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-28", "url": "https://arxiv.org/abs/2302.14703", "pdf_url": "https://arxiv.org/pdf/2302.14703v1", "arxiv_id": "2302.14703", "doi": "10.48550/arXiv.2302.14703", "citation_count": 22, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3404}
{"id": "278aed9efd869b19b0a673e48c2cb9f1d57aaf0a11ba0a8fbb43a1660885db6c", "sources": ["arxiv", "semantic_scholar"], "title": "MoLE : Mixture of Language Experts for Multi-Lingual Automatic Speech Recognition", "abstract": "Multi-lingual speech recognition aims to distinguish linguistic expressions in different languages and integrate acoustic processing simultaneously. In contrast, current multi-lingual speech recognition research follows a language-aware paradigm, mainly targeted to improve recognition performance rather than discriminate language characteristics. In this paper, we present a multi-lingual speech recognition network named Mixture-of-Language-Expert(MoLE), which digests speech in a variety of languages. Specifically, MoLE analyzes linguistic expression from input speech in arbitrary languages, activating a language-specific expert with a lightweight language tokenizer. The tokenizer not only activates experts, but also estimates the reliability of the activation. Based on the reliability, the activated expert and the language-agnostic expert are aggregated to represent language-conditioned embedding for efficient speech recognition. Our proposed model is evaluated in 5 languages scenario, and the experimental results show that our structure is advantageous on multi-lingual recognition, especially for speech in low-resource language.", "authors": ["Yoohwan Kwon", "Soo-Whan Chung"], "categories": ["eess.AS", "cs.CL", "cs.SD"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-02-27", "url": "https://arxiv.org/abs/2302.13750", "pdf_url": "https://arxiv.org/pdf/2302.13750v1", "arxiv_id": "2302.13750", "doi": "10.1109/ICASSP49357.2023.10096227", "citation_count": 33, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.3829}
{"id": "3d5c81b470827ba001a657bbaec6efaf3002927b87954942b7d3e973856b717b", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Optimal Compression: Joint Pruning and Quantization", "abstract": "Model compression is instrumental in optimizing deep neural network inference on resource-constrained hardware. The prevailing methods for network compression, namely quantization and pruning, have been shown to enhance efficiency at the cost of performance. Determining the most effective quantization and pruning strategies for individual layers and parameters remains a challenging problem, often requiring computationally expensive and ad hoc numerical optimization techniques. This paper introduces FITCompress, a novel method integrating layer-wise mixed-precision quantization and unstructured pruning using a unified heuristic approach. By leveraging the Fisher Information Metric and path planning through compression space, FITCompress optimally selects a combination of pruning mask and mixed-precision quantization configuration for a given pre-trained model and compression constraint. Experiments on computer vision and natural language processing benchmarks demonstrate that our proposed approach achieves a superior compression-performance trade-off compared to existing state-of-the-art methods. FITCompress stands out for its principled derivation, making it versatile across tasks and network architectures, and represents a step towards achieving optimal compression for neural networks.", "authors": ["Ben Zandonati", "Glenn Bucagu", "Adrian Alan Pol", "Maurizio Pierini", "Olya Sirkin", "Tal Kopetz"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-15", "url": "https://arxiv.org/abs/2302.07612", "pdf_url": "https://arxiv.org/pdf/2302.07612v2", "arxiv_id": "2302.07612", "doi": "10.48550/arXiv.2302.07612", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "78161a762519a022e2f1c2a9fb0151c3cece2bd80d50357018ef56aeaa2495cb", "sources": ["arxiv", "semantic_scholar"], "title": "Efficient Attention via Control Variates", "abstract": "Random-feature-based attention (RFA) is an efficient approximation of softmax attention with linear runtime and space complexity. However, the approximation gap between RFA and conventional softmax attention is not well studied. Built upon previous progress of RFA, we characterize this gap through the lens of control variates and show that RFA can be decomposed into a sum of multiple control variate estimators for each element in the sequence. This new framework reveals that exact softmax attention can be recovered from RFA by manipulating each control variate. Besides, it allows us to develop a more flexible form of control variates, resulting in a novel attention mechanism that significantly reduces the approximation gap while maintaining linear complexity. Extensive experiments demonstrate that our model outperforms state-of-the-art efficient attention mechanisms on both vision and language tasks.", "authors": ["Lin Zheng", "Jianbo Yuan", "Chong Wang", "Lingpeng Kong"], "categories": ["cs.LG", "cs.CL", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-09", "url": "https://arxiv.org/abs/2302.04542", "pdf_url": "https://arxiv.org/pdf/2302.04542v1", "arxiv_id": "2302.04542", "doi": "10.48550/arXiv.2302.04542", "citation_count": 27, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/HKUNLP/efficient-attention", "venue": "International Conference on Learning Representations", "quality_score": 0.3618}
{"id": "8959cdf0d86825e798757b3e1ea1e09acedc91aa75f099fa0a9099e3b3faacb9", "sources": ["arxiv", "semantic_scholar"], "title": "Gaussian Process-Gated Hierarchical Mixtures of Experts", "abstract": "In this paper, we propose novel Gaussian process-gated hierarchical mixtures of experts (GPHMEs). Unlike other mixtures of experts with gating models linear in the input, our model employs gating functions built with Gaussian processes (GPs). These processes are based on random features that are non-linear functions of the inputs. Furthermore, the experts in our model are also constructed with GPs. The optimization of the GPHMEs is performed by variational inference. The proposed GPHMEs have several advantages. They outperform tree-based HME benchmarks that partition the data in the input space, and they achieve good performance with reduced complexity. Another advantage is the interpretability they provide for deep GPs, and more generally, for deep Bayesian neural networks. Our GPHMEs demonstrate excellent performance for large-scale data sets, even with quite modest sizes.", "authors": ["Yuhao Liu", "Marzieh Ajirak", "Petar Djuric"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Medicine", "Mathematics"], "published_date": "2023-02-09", "url": "https://arxiv.org/abs/2302.04947", "pdf_url": "https://arxiv.org/pdf/2302.04947v2", "arxiv_id": "2302.04947", "doi": "10.1109/TPAMI.2024.3381936", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.1945}
{"id": "737813f489b08c5a7edaeea1b084273a88bcf7ef1936613109ac3079001623d7", "sources": ["arxiv", "semantic_scholar"], "title": "The Hardware Impact of Quantization and Pruning for Weights in Spiking Neural Networks", "abstract": "Energy efficient implementations and deployments of Spiking neural networks (SNNs) have been of great interest due to the possibility of developing artificial systems that can achieve the computational powers and energy efficiency of the biological brain. Efficient implementations of SNNs on modern digital hardware are also inspired by advances in machine learning and deep neural networks (DNNs). Two techniques widely employed in the efficient deployment of DNNs -- the quantization and pruning of parameters, can both compress the model size, reduce memory footprints, and facilitate low-latency execution. The interaction between quantization and pruning and how they might impact model performance on SNN accelerators is currently unknown. We study various combinations of pruning and quantization in isolation, cumulatively, and simultaneously (jointly) to a state-of-the-art SNN targeting gesture recognition for dynamic vision sensor cameras (DVS). We show that this state-of-the-art model is amenable to aggressive parameter quantization, not suffering from any loss in accuracy down to ternary weights. However, pruning only maintains iso-accuracy up to 80% sparsity, which results in 45% more energy than the best quantization on our architectural model. Applying both pruning and quantization can result in an accuracy loss to offer a favourable trade-off on the energy-accuracy Pareto-frontier for the given hardware configuration.", "authors": ["Clemens JS Schaefer", "Pooria Taheri", "Mark Horeni", "Siddharth Joshi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-08", "url": "https://arxiv.org/abs/2302.04174", "pdf_url": "https://arxiv.org/pdf/2302.04174v1", "arxiv_id": "2302.04174", "doi": "10.1109/TCSII.2023.3260701", "citation_count": 21, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Intelligent-Microsystems-Lab/SNNQuantPrune", "venue": null, "quality_score": 0.3356}
{"id": "889148f13a03f72e7b15a393c4b17eb6ebbfc557e0708ed2760f3eceaf2bd44f", "sources": ["arxiv", "semantic_scholar"], "title": "mixdistreg: An R Package for Fitting Mixture of Experts Distributional Regression with Adaptive First-order Methods", "abstract": "This paper presents a high-level description of the R software package mixdistreg to fit mixture of experts distributional regression models. The proposed framework is implemented in R using the deepregression software template, which is based on TensorFlow and follows the neural structured additive learning principle. The software comprises various approaches as special cases, including mixture density networks and mixture regression approaches. Various code examples are given to demonstrate the package's functionality.", "authors": ["David Rügamer"], "categories": ["stat.CO"], "fields_of_study": ["Mathematics"], "published_date": "2023-02-04", "url": "https://arxiv.org/abs/2302.02043", "pdf_url": "https://arxiv.org/pdf/2302.02043v1", "arxiv_id": "2302.02043", "doi": null, "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "3f2c8f4c747da2508cddd9cfb35ec6fb7b549f08854977799ddb6711c86e153b", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-modal Gated Mixture of Local-to-Global Experts for Dynamic Image Fusion", "abstract": "Infrared and visible image fusion aims to integrate comprehensive information from multiple sources to achieve superior performances on various practical tasks, such as detection, over that of a single modality. However, most existing methods directly combined the texture details and object contrast of different modalities, ignoring the dynamic changes in reality, which diminishes the visible texture in good lighting conditions and the infrared contrast in low lighting conditions. To fill this gap, we propose a dynamic image fusion framework with a multi-modal gated mixture of local-to-global experts, termed MoE-Fusion, to dynamically extract effective and comprehensive information from the respective modalities. Our model consists of a Mixture of Local Experts (MoLE) and a Mixture of Global Experts (MoGE) guided by a multi-modal gate. The MoLE performs specialized learning of multi-modal local features, prompting the fused images to retain the local information in a sample-adaptive manner, while the MoGE focuses on the global information that complements the fused image with overall texture detail and contrast. Extensive experiments show that our MoE-Fusion outperforms state-of-the-art methods in preserving multi-modal image texture and contrast through the local-to-global dynamic learning paradigm, and also achieves superior performance on detection tasks. Our code will be available: https://github.com/SunYM2020/MoE-Fusion.", "authors": ["Yiming Sun", "Bing Cao", "Pengfei Zhu", "Qinghua Hu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-02", "url": "https://arxiv.org/abs/2302.01392", "pdf_url": "https://arxiv.org/pdf/2302.01392v2", "arxiv_id": "2302.01392", "doi": "10.1109/ICCV51070.2023.02153", "citation_count": 116, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/SunYM2020/MoE-Fusion", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.517}
{"id": "ac18c8cd9f3c864e8ebbac79ec7fca01147ba82e1c41b5765e1175c38a26f49d", "sources": ["arxiv", "semantic_scholar"], "title": "ACQ: Improving Generative Data-free Quantization Via Attention Correction", "abstract": "Data-free quantization aims to achieve model quantization without accessing any authentic sample. It is significant in an application-oriented context involving data privacy. Converting noise vectors into synthetic samples through a generator is a popular data-free quantization method, which is called generative data-free quantization. However, there is a difference in attention between synthetic samples and authentic samples. This is always ignored and restricts the quantization performance. First, since synthetic samples of the same class are prone to have homogenous attention, the quantized network can only learn limited modes of attention. Second, synthetic samples in eval mode and training mode exhibit different attention. Hence, the batch-normalization statistics matching tends to be inaccurate. ACQ is proposed in this paper to fix the attention of synthetic samples. An attention center position-condition generator is established regarding the homogenization of intra-class attention. Restricted by the attention center matching loss, the attention center position is treated as the generator's condition input to guide synthetic samples in obtaining diverse attention. Moreover, we design adversarial loss of paired synthetic samples under the same condition to prevent the generator from paying overmuch attention to the condition, which may result in mode collapse. To improve the attention similarity of synthetic samples in different network modes, we introduce a consistency penalty to guarantee accurate BN statistics matching. The experimental results demonstrate that ACQ effectively improves the attention problems of synthetic samples. Under various training settings, ACQ achieves the best quantization performance. For the 4-bit quantization of Resnet18 and Resnet50, ACQ reaches 67.55% and 72.23% accuracy, respectively.", "authors": ["Jixing Li", "Xiaozhou Guo", "Benzhe Dai", "Guoliang Gong", "Min Jin", "Gang Chen", "Wenyu Mao", "Huaxiang Lu"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-18", "url": "https://arxiv.org/abs/2301.07266", "pdf_url": "https://arxiv.org/pdf/2301.07266v2", "arxiv_id": "2301.07266", "doi": "10.48550/arXiv.2301.07266", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Pattern Recognition", "quality_score": 0.2113}
{"id": "7006e0d38d769430bf8170cd724bb1b600a1adcfcb3a7395bd77785d80ec390a", "sources": ["arxiv", "semantic_scholar"], "title": "ChameleMon: Shifting Measurement Attention as Network State Changes", "abstract": "Flow-level network measurement is critical to many network applications. Among various measurement tasks, packet loss detection and heavy-hitter detection are two most important measurement tasks, which we call the two key tasks. In practice, the two key tasks are often required at the same time, but existing works seldom handle both tasks. In this paper, we design ChameleMon to support the two key tasks simultaneously. One key design/novelty of ChameleMon is to shift measurement attention as network state changes, through two dimensions of dynamics: 1) dynamically allocating memory between the two key tasks; 2) dynamically monitoring the flows of importance. To realize the key design, we propose a key technique, leveraging Fermat's little theorem to devise a flexible data structure, namely FermatSketch. FermatSketch is dividable, additive, and subtractive, supporting the two key tasks. We have fully implemented a ChameleMon prototype on a testbed with a Fat-tree topology. We conduct extensive experiments and the results show ChameleMon supports the two key tasks with low memory/bandwidth overhead, and more importantly, it can automatically shift measurement attention as network state changes.", "authors": ["Kaicheng Yang", "Yuhan Wu", "Ruijie Miao", "Tong Yang", "Zirui Liu", "Zicang Xu", "Rui Qiu", "Yikai Zhao", "Hanglong Lv", "Zhigang Ji", "Gaogang Xie"], "categories": ["cs.NI"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-02", "url": "https://arxiv.org/abs/2301.00615", "pdf_url": "https://arxiv.org/pdf/2301.00615v2", "arxiv_id": "2301.00615", "doi": "10.1145/3603269.3604850", "citation_count": 26, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication", "quality_score": 0.3578}
{"id": "86631c736240708f8677988107bbc7212d550b9f14fb7540f71e614b0447c82e", "sources": ["arxiv", "semantic_scholar"], "title": "Memory-efficient NLLB-200: Language-specific Expert Pruning of a Massively Multilingual Machine Translation Model", "abstract": "The recently released NLLB-200 is a set of multilingual Neural Machine Translation models that cover 202 languages. The largest model is based on a Mixture of Experts architecture and achieves SoTA results across many language pairs. It contains 54.5B parameters and requires at least four 32GB GPUs just for inference. In this work, we propose a pruning method that enables the removal of up to 80% of experts without further finetuning and with a negligible loss in translation quality, which makes it feasible to run the model on a single 32GB GPU. Further analysis suggests that our pruning metrics can identify language-specific experts.", "authors": ["Yeskendir Koishekenov", "Alexandre Berard", "Vassilina Nikoulina"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-19", "url": "https://arxiv.org/abs/2212.09811", "pdf_url": "https://arxiv.org/pdf/2212.09811v3", "arxiv_id": "2212.09811", "doi": "10.48550/arXiv.2212.09811", "citation_count": 54, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4351}
{"id": "47391f69e1b1ba7b296a6907eea50b407c427070f470c87e65b608851399ec53", "sources": ["arxiv", "semantic_scholar"], "title": "First De-Trend then Attend: Rethinking Attention for Time-Series Forecasting", "abstract": "Transformer-based models have gained large popularity and demonstrated promising results in long-term time-series forecasting in recent years. In addition to learning attention in time domain, recent works also explore learning attention in frequency domains (e.g., Fourier domain, wavelet domain), given that seasonal patterns can be better captured in these domains. In this work, we seek to understand the relationships between attention models in different time and frequency domains. Theoretically, we show that attention models in different domains are equivalent under linear conditions (i.e., linear kernel to attention scores). Empirically, we analyze how attention models of different domains show different behaviors through various synthetic experiments with seasonality, trend and noise, with emphasis on the role of softmax operation therein. Both these theoretical and empirical analyses motivate us to propose a new method: TDformer (Trend Decomposition Transformer), that first applies seasonal-trend decomposition, and then additively combines an MLP which predicts the trend component with Fourier attention which predicts the seasonal component to obtain the final prediction. Extensive experiments on benchmark time-series forecasting datasets demonstrate that TDformer achieves state-of-the-art performance against existing attention-based models.", "authors": ["Xiyuan Zhang", "Xiaoyong Jin", "Karthick Gopalswamy", "Gaurav Gupta", "Youngsuk Park", "Xingjian Shi", "Hao Wang", "Danielle C. Maddix", "Yuyang Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-15", "url": "https://arxiv.org/abs/2212.08151", "pdf_url": "https://arxiv.org/pdf/2212.08151v1", "arxiv_id": "2212.08151", "doi": "10.48550/arXiv.2212.08151", "citation_count": 46, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.418}
{"id": "f488bc687b72c73ff2247205ea81a1d65a5bb314ecfb41c56f594bb3e270a346", "sources": ["arxiv", "semantic_scholar"], "title": "Gradient-based Intra-attention Pruning on Pre-trained Language Models", "abstract": "Pre-trained language models achieve superior performance but are computationally expensive. Techniques such as pruning and knowledge distillation have been developed to reduce their sizes and latencies. In this work, we propose a structured pruning method GRAIN (Gradient-based Intra-attention pruning), which performs task-specific pruning with knowledge distillation and yields highly effective models. Different from common approaches that prune each attention head as a whole, GRAIN inspects and prunes intra-attention structures, which greatly expands the structure search space and enables more flexible models. We also propose a gradient separation strategy that reduces the interference of distillation on pruning for a better combination of the two approaches. Experiments on GLUE, SQuAD, and CoNLL 2003 show that GRAIN notably outperforms other methods, especially in the high sparsity regime, and achieves $6\\sim7\\times$ speedups while maintaining $93\\%\\sim99\\%$ performance. Under extreme compression where only $3\\%$ transformer weights remain, the pruned model is still competitive compared to larger models.", "authors": ["Ziqing Yang", "Yiming Cui", "Xin Yao", "Shijin Wang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-15", "url": "https://arxiv.org/abs/2212.07634", "pdf_url": "https://arxiv.org/pdf/2212.07634v2", "arxiv_id": "2212.07634", "doi": "10.48550/arXiv.2212.07634", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3076}
{"id": "b99a99c47b601ba9ca1b6e277ebfb55a6a87fddd58918dbc9cd4a2ac52a9ff12", "sources": ["arxiv", "semantic_scholar"], "title": "Mod-Squad: Designing Mixture of Experts As Modular Multi-Task Learners", "abstract": "Optimization in multi-task learning (MTL) is more challenging than single-task learning (STL), as the gradient from different tasks can be contradictory. When tasks are related, it can be beneficial to share some parameters among them (cooperation). However, some tasks require additional parameters with expertise in a specific type of data or discrimination (specialization). To address the MTL challenge, we propose Mod-Squad, a new model that is Modularized into groups of experts (a 'Squad'). This structure allows us to formalize cooperation and specialization as the process of matching experts and tasks. We optimize this matching process during the training of a single model. Specifically, we incorporate mixture of experts (MoE) layers into a transformer model, with a new loss that incorporates the mutual dependence between tasks and experts. As a result, only a small set of experts are activated for each task. This prevents the sharing of the entire backbone model between all tasks, which strengthens the model, especially when the training set size and the number of tasks scale up. More interestingly, for each task, we can extract the small set of experts as a standalone model that maintains the same performance as the large model. Extensive experiments on the Taskonomy dataset with 13 vision tasks and the PASCAL-Context dataset with 5 vision tasks show the superiority of our approach.", "authors": ["Zitian Chen", "Yikang Shen", "Mingyu Ding", "Zhenfang Chen", "Hengshuang Zhao", "Erik Learned-Miller", "Chuang Gan"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-15", "url": "https://arxiv.org/abs/2212.08066", "pdf_url": "https://arxiv.org/pdf/2212.08066v1", "arxiv_id": "2212.08066", "doi": "10.48550/arXiv.2212.08066", "citation_count": 19, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3253}
{"id": "983bae249f439a7d4ba4dfbacfc73a0bc1a303601f662cd5d3a00a862d66fad8", "sources": ["arxiv", "semantic_scholar"], "title": "Modeling Multimodal Aleatoric Uncertainty in Segmentation with Mixture of Stochastic Experts", "abstract": "Equipping predicted segmentation with calibrated uncertainty is essential for safety-critical applications. In this work, we focus on capturing the data-inherent uncertainty (aka aleatoric uncertainty) in segmentation, typically when ambiguities exist in input images. Due to the high-dimensional output space and potential multiple modes in segmenting ambiguous images, it remains challenging to predict well-calibrated uncertainty for segmentation. To tackle this problem, we propose a novel mixture of stochastic experts (MoSE) model, where each expert network estimates a distinct mode of the aleatoric uncertainty and a gating network predicts the probabilities of an input image being segmented in those modes. This yields an efficient two-level uncertainty representation. To learn the model, we develop a Wasserstein-like loss that directly minimizes the distribution distance between the MoSE and ground truth annotations. The loss can easily integrate traditional segmentation quality measures and be efficiently optimized via constraint relaxation. We validate our method on the LIDC-IDRI dataset and a modified multimodal Cityscapes dataset. Results demonstrate that our method achieves the state-of-the-art or competitive performance on all metrics.", "authors": ["Zhitong Gao", "Yucong Chen", "Chuyu Zhang", "Xuming He"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-14", "url": "https://arxiv.org/abs/2212.07328", "pdf_url": "https://arxiv.org/pdf/2212.07328v2", "arxiv_id": "2212.07328", "doi": "10.48550/arXiv.2212.07328", "citation_count": 14, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.294}
{"id": "e93f2f5d9ab1192e16b6a67eef786dbffbb477b46073163026e0e0a1eac6308f", "sources": ["arxiv", "semantic_scholar"], "title": "SMILE: Scaling Mixture-of-Experts with Efficient Bi-level Routing", "abstract": "The mixture of Expert (MoE) parallelism is a recent advancement that scales up the model size with constant computational cost. MoE selects different sets of parameters (i.e., experts) for each incoming token, resulting in a sparsely-activated model. Despite several successful applications of MoE, its training efficiency degrades significantly as the number of experts increases. The routing stage in MoE relies on the efficiency of the All2All communication collective, which suffers from network congestion and has poor scalability. To mitigate these issues, we introduce SMILE, which exploits heterogeneous network bandwidth and splits a single-step routing into bi-level routing. Our experimental results show that the proposed method obtains a 2.5x speedup over Switch Transformer in terms of pretraining throughput on the Colossal Clean Crawled Corpus without losing any convergence speed.", "authors": ["Chaoyang He", "Shuai Zheng", "Aston Zhang", "George Karypis", "Trishul Chilimbi", "Mahdi Soltanolkotabi", "Salman Avestimehr"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-10", "url": "https://arxiv.org/abs/2212.05191", "pdf_url": "https://arxiv.org/pdf/2212.05191v1", "arxiv_id": "2212.05191", "doi": "10.48550/arXiv.2212.05191", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "6076bb45064961401ce0af36a69438b1ffe06edb63c23cebe9ac6b150ff217a1", "sources": ["arxiv", "semantic_scholar"], "title": "Implicit Mixture of Interpretable Experts for Global and Local Interpretability", "abstract": "We investigate the feasibility of using mixtures of interpretable experts (MoIE) to build interpretable image classifiers on MNIST10. MoIE uses a black-box router to assign each input to one of many inherently interpretable experts, thereby providing insight into why a particular classification decision was made. We find that a naively trained MoIE will learn to 'cheat', whereby the black-box router will solve the classification problem by itself, with each expert simply learning a constant function for one particular class. We propose to solve this problem by introducing interpretable routers and training the black-box router's decisions to match the interpretable router. In addition, we propose a novel implicit parameterization scheme that allows us to build mixtures of arbitrary numbers of experts, allowing us to study how classification performance, local and global interpretability vary as the number of experts is increased. Our new model, dubbed Implicit Mixture of Interpretable Experts (IMoIE) can match state-of-the-art classification accuracy on MNIST10 while providing local interpretability, and can provide global interpretability albeit at the cost of reduced classification accuracy.", "authors": ["Nathan Elazar", "Kerry Taylor"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-01", "url": "https://arxiv.org/abs/2212.00471", "pdf_url": "https://arxiv.org/pdf/2212.00471v1", "arxiv_id": "2212.00471", "doi": "10.48550/arXiv.2212.00471", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "ab97bcddbeda8764a1f47b6e93355e48e800e3a889d8d96e47991ae0a5554a7f", "sources": ["arxiv", "semantic_scholar"], "title": "Automatically Extracting Information in Medical Dialogue: Expert System And Attention for Labelling", "abstract": "Medical dialogue information extraction is becoming an increasingly significant problem in modern medical care. It is difficult to extract key information from electronic medical records (EMRs) due to their large numbers. Previously, researchers proposed attention-based models for retrieving features from EMRs, but their limitations were reflected in their inability to recognize different categories in medical dialogues. In this paper, we propose a novel model, Expert System and Attention for Labelling (ESAL). We use mixture of experts and pre-trained BERT to retrieve the semantics of different categories, enabling the model to fuse the differences between them. In our experiment, ESAL was applied to a public dataset and the experimental results indicated that ESAL significantly improved the performance of Medical Information Classification.", "authors": ["Xinshi Wang", "Daniel Tang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-28", "url": "https://arxiv.org/abs/2211.15544", "pdf_url": "https://arxiv.org/pdf/2211.15544v2", "arxiv_id": "2211.15544", "doi": "10.48550/arXiv.2211.15544", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "4c66d9c5f579d78e15c361905ef352e7c6bae249f236190a7f344d231049f7ee", "sources": ["arxiv", "semantic_scholar"], "title": "Spatial Mixture-of-Experts", "abstract": "Many data have an underlying dependence on spatial location; it may be weather on the Earth, a simulation on a mesh, or a registered image. Yet this feature is rarely taken advantage of, and violates common assumptions made by many neural network layers, such as translation equivariance. Further, many works that do incorporate locality fail to capture fine-grained structure. To address this, we introduce the Spatial Mixture-of-Experts (SMoE) layer, a sparsely-gated layer that learns spatial structure in the input domain and routes experts at a fine-grained level to utilize it. We also develop new techniques to train SMoEs, including a self-supervised routing loss and damping expert errors. Finally, we show strong results for SMoEs on numerous tasks, and set new state-of-the-art results for medium-range weather prediction and post-processing ensemble weather forecasts.", "authors": ["Nikoli Dryden", "Torsten Hoefler"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-24", "url": "https://arxiv.org/abs/2211.13491", "pdf_url": "https://arxiv.org/pdf/2211.13491v1", "arxiv_id": "2211.13491", "doi": "10.48550/arXiv.2211.13491", "citation_count": 17, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3138}
{"id": "f11ef3037e8e61e8c46dd9082ec3b3c05ea9918208349bc073c85ba0485b3ddb", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts Distributional Regression: Implementation Using Robust Estimation with Adaptive First-order Methods", "abstract": "In this work, we propose an efficient implementation of mixtures of experts distributional regression models which exploits robust estimation by using stochastic first-order optimization techniques with adaptive learning rate schedulers. We take advantage of the flexibility and scalability of neural network software and implement the proposed framework in mixdistreg, an R software package that allows for the definition of mixtures of many different families, estimation in high-dimensional and large sample size settings and robust optimization based on TensorFlow. Numerical experiments with simulated and real-world data applications show that optimization is as reliable as estimation via classical approaches in many different settings and that results may be obtained for complicated scenarios where classical approaches consistently fail.", "authors": ["David Rügamer", "Florian Pfisterer", "Bernd Bischl", "Bettina Grün"], "categories": ["stat.CO"], "fields_of_study": ["Mathematics"], "published_date": "2022-11-17", "url": "https://arxiv.org/abs/2211.09875", "pdf_url": "https://arxiv.org/pdf/2211.09875v1", "arxiv_id": "2211.09875", "doi": "10.1007/s10182-023-00486-8", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AStA Advances in Statistical Analysis", "quality_score": 0.2113}
{"id": "68ecbd4aa4e6b1451284f2d7425eb87a448265b76ae0ab046356f65fec2ce054", "sources": ["arxiv", "semantic_scholar"], "title": "HMOE: Hypernetwork-based Mixture of Experts for Domain Generalization", "abstract": "Due to domain shifts, machine learning systems typically struggle to generalize well to new domains that differ from those of training data, which is what domain generalization (DG) aims to address. Although a variety of DG methods have been proposed, most of them fall short in interpretability and require domain labels, which are not available in many real-world scenarios. This paper presents a novel DG method, called HMOE: Hypernetwork-based Mixture of Experts (MoE), which does not rely on domain labels and is more interpretable. MoE proves effective in identifying heterogeneous patterns in data. For the DG problem, heterogeneity arises exactly from domain shifts. HMOE employs hypernetworks taking vectors as input to generate the weights of experts, which promotes knowledge sharing among experts and enables the exploration of their similarities in a low-dimensional vector space. We benchmark HMOE against other DG methods under a fair evaluation framework -- DomainBed. Our extensive experiments show that HMOE can effectively separate mixed-domain data into distinct clusters that are surprisingly more consistent with human intuition than original domain labels. Using self-learned domain information, HMOE achieves state-of-the-art results on most datasets and significantly surpasses other DG methods in average accuracy across all datasets.", "authors": ["Jingang Qu", "Thibault Faney", "Ze Wang", "Patrick Gallinari", "Soleiman Yousef", "Jean-Charles de Hemptinne"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-15", "url": "https://arxiv.org/abs/2211.08253", "pdf_url": "https://arxiv.org/pdf/2211.08253v3", "arxiv_id": "2211.08253", "doi": "10.48550/arXiv.2211.08253", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "276254dc41353b475e8a827e17f3f62d7f7867746fb86a22ff2d9e1567b98516", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning Very Deep Neural Network Channels for Efficient Inference", "abstract": "In this paper, we introduce a new channel pruning method to accelerate very deep convolutional neural networks. Given a trained CNN model, we propose an iterative two-step algorithm to effectively prune each layer, by a LASSO regression based channel selection and least square reconstruction. We further generalize this algorithm to multi-layer and multi-branch cases. Our method reduces the accumulated error and enhances the compatibility with various architectures. Our pruned VGG-16 achieves the state-of-the-art results by 5x speed-up along with only 0.3% increase of error. More importantly, our method is able to accelerate modern networks like ResNet, Xception and suffers only 1.4%, 1.0% accuracy loss under 2x speed-up respectively, which is significant. Our code has been made publicly available.", "authors": ["Yihui He"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-14", "url": "https://arxiv.org/abs/2211.08339", "pdf_url": "https://arxiv.org/pdf/2211.08339v1", "arxiv_id": "2211.08339", "doi": "10.48550/arXiv.2211.08339", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "40ae856a80fa7559bb681b77cb641da0aed7d42f8f3628d991d457146fae5d99", "sources": ["arxiv", "semantic_scholar"], "title": "Robust Reflection Removal with Flash-only Cues in the Wild", "abstract": "We propose a simple yet effective reflection-free cue for robust reflection removal from a pair of flash and ambient (no-flash) images. The reflection-free cue exploits a flash-only image obtained by subtracting the ambient image from the corresponding flash image in raw data space. The flash-only image is equivalent to an image taken in a dark environment with only a flash on. This flash-only image is visually reflection-free and thus can provide robust cues to infer the reflection in the ambient image. Since the flash-only image usually has artifacts, we further propose a dedicated model that not only utilizes the reflection-free cue but also avoids introducing artifacts, which helps accurately estimate reflection and transmission. Our experiments on real-world images with various types of reflection demonstrate the effectiveness of our model with reflection-free flash-only cues: our model outperforms state-of-the-art reflection removal approaches by more than 5.23dB in PSNR. We extend our approach to handheld photography to address the misalignment between the flash and no-flash pair. With misaligned training data and the alignment module, our aligned model outperforms our previous version by more than 3.19dB in PSNR on a misaligned dataset. We also study using linear RGB images as training data. Our source code and dataset are publicly available at https://github.com/ChenyangLEI/flash-reflection-removal.", "authors": ["Chenyang Lei", "Xudong Jiang", "Qifeng Chen"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2022-11-05", "url": "https://arxiv.org/abs/2211.02914", "pdf_url": "https://arxiv.org/pdf/2211.02914v2", "arxiv_id": "2211.02914", "doi": "10.1109/TPAMI.2023.3314972", "citation_count": 23, "influential_citation_count": 5, "has_code": true, "code_url": "http://github.com/ChenyangLEI/flash-reflection-removal", "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "quality_score": 0.3891}
{"id": "e6be35d9bf53e6151b15a1b6a68b0256b388d86b25ed06ca85cc300722e0286c", "sources": ["arxiv", "semantic_scholar"], "title": "Contextual Mixture of Experts: Integrating Knowledge into Predictive Modeling", "abstract": "This work proposes a new data-driven model devised to integrate process knowledge into its structure to increase the human-machine synergy in the process industry. The proposed Contextual Mixture of Experts (cMoE) explicitly uses process knowledge along the model learning stage to mold the historical data to represent operators' context related to the process through possibility distributions. This model was evaluated in two real case studies for quality prediction, including a sulfur recovery unit and a polymerization process. The contextual mixture of experts was employed to represent different contexts in both experiments. The results indicate that integrating process knowledge has increased predictive performance while improving interpretability by providing insights into the variables affecting the process's different regimes.", "authors": ["Francisco Souza", "Tim Offermans", "Ruud Barendse", "Geert Postma", "Jeroen Jansen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-01", "url": "https://arxiv.org/abs/2211.00558", "pdf_url": "https://arxiv.org/pdf/2211.00558v1", "arxiv_id": "2211.00558", "doi": "10.1109/TII.2022.3224973", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Industrial Informatics", "quality_score": 0.2603}
{"id": "654cf6d3bcf8bff6f1dcfcc20c71bf784a4398c18a28aee3cf3f2f27bd7ce154", "sources": ["arxiv", "semantic_scholar"], "title": "Prediction Sets for High-Dimensional Mixture of Experts Models", "abstract": "Large datasets make it possible to build predictive models that can capture heterogenous relationships between the response variable and features. The mixture of high-dimensional linear experts model posits that observations come from a mixture of high-dimensional linear regression models, where the mixture weights are themselves feature-dependent. In this paper, we show how to construct valid prediction sets for an $\\ell_1$-penalized mixture of experts model in the high-dimensional setting. We make use of a debiasing procedure to account for the bias induced by the penalization and propose a novel strategy for combining intervals to form a prediction set with coverage guarantees in the mixture setting. Synthetic examples and an application to the prediction of critical temperatures of superconducting materials show our method to have reliable practical performance.", "authors": ["Adel Javanmard", "Simeng Shao", "Jacob Bien"], "categories": ["math.ST", "stat.ME", "stat.ML"], "fields_of_study": ["Mathematics"], "published_date": "2022-10-30", "url": "https://arxiv.org/abs/2210.16710", "pdf_url": "https://arxiv.org/pdf/2210.16710v1", "arxiv_id": "2210.16710", "doi": "10.1093/jrsssb/qkae117", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2386}
{"id": "84d82fdbd5394793e411acdf0f54525c0fb4471c98f20f2f74931ba531e767ef", "sources": ["arxiv", "semantic_scholar"], "title": "Pruning by Active Attention Manipulation", "abstract": "Filter pruning of a CNN is typically achieved by applying discrete masks on the CNN's filter weights or activation maps, post-training. Here, we present a new filter-importance-scoring concept named pruning by active attention manipulation (PAAM), that sparsifies the CNN's set of filters through a particular attention mechanism, during-training. PAAM learns analog filter scores from the filter weights by optimizing a cost function regularized by an additive term in the scores. As the filters are not independent, we use attention to dynamically learn their correlations. Moreover, by training the pruning scores of all layers simultaneously, PAAM can account for layer inter-dependencies, which is essential to finding a performant sparse sub-network. PAAM can also train and generate a pruned network from scratch in a straightforward, one-stage training process without requiring a pre-trained network. Finally, PAAM does not need layer-specific hyperparameters and pre-defined layer budgets, since it can implicitly determine the appropriate number of filters in each layer. Our experimental results on different network architectures suggest that PAAM outperforms state-of-the-art structured-pruning methods (SOTA). On CIFAR-10 dataset, without requiring a pre-trained baseline network, we obtain 1.02% and 1.19% accuracy gain and 52.3% and 54% parameters reduction, on ResNet56 and ResNet110, respectively. Similarly, on the ImageNet dataset, PAAM achieves 1.06% accuracy gain while pruning 51.1% of the parameters on ResNet50. For Cifar-10, this is better than the SOTA with a margin of 9.5% and 6.6%, respectively, and on ImageNet with a margin of 11%.", "authors": ["Zahra Babaiee", "Lucas Liebenwein", "Ramin Hasani", "Daniela Rus", "Radu Grosu"], "categories": ["cs.CV", "cs.AI", "cs.LG", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-20", "url": "https://arxiv.org/abs/2210.11114", "pdf_url": "https://arxiv.org/pdf/2210.11114v1", "arxiv_id": "2210.11114", "doi": "10.48550/arXiv.2210.11114", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "023f0bb174237258d52133d7487f5e278cfb03546ae857b44938a4c88312b9b7", "sources": ["arxiv", "semantic_scholar"], "title": "On the Adversarial Robustness of Mixture of Experts", "abstract": "Adversarial robustness is a key desirable property of neural networks. It has been empirically shown to be affected by their sizes, with larger networks being typically more robust. Recently, Bubeck and Sellke proved a lower bound on the Lipschitz constant of functions that fit the training data in terms of their number of parameters. This raises an interesting open question, do -- and can -- functions with more parameters, but not necessarily more computational cost, have better robustness? We study this question for sparse Mixture of Expert models (MoEs), that make it possible to scale up the model size for a roughly constant computational cost. We theoretically show that under certain conditions on the routing and the structure of the data, MoEs can have significantly smaller Lipschitz constants than their dense counterparts. The robustness of MoEs can suffer when the highest weighted experts for an input implement sufficiently different functions. We next empirically evaluate the robustness of MoEs on ImageNet using adversarial attacks and show they are indeed more robust than dense models with the same computational cost. We make key observations showing the robustness of MoEs to the choice of experts, highlighting the redundancy of experts in models trained in practice.", "authors": ["Joan Puigcerver", "Rodolphe Jenatton", "Carlos Riquelme", "Pranjal Awasthi", "Srinadh Bhojanapalli"], "categories": ["cs.LG", "cs.AI", "cs.CR", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-19", "url": "https://arxiv.org/abs/2210.10253", "pdf_url": "https://arxiv.org/pdf/2210.10253v1", "arxiv_id": "2210.10253", "doi": "10.48550/arXiv.2210.10253", "citation_count": 29, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3693}
{"id": "1cd9927adc90d3b0e02138f58f7c650b442f62fff7e1db8869d2c0ab85a06d9f", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Attention Heads: Selecting Attention Heads Per Token", "abstract": "Mixture-of-Experts (MoE) networks have been proposed as an efficient way to scale up model capacity and implement conditional computing. However, the study of MoE components mostly focused on the feedforward layer in Transformer architecture. This paper proposes the Mixture of Attention Heads (MoA), a new architecture that combines multi-head attention with the MoE mechanism. MoA includes a set of attention heads that each has its own set of parameters. Given an input, a router dynamically selects a subset of $k$ attention heads per token. This conditional computation schema allows MoA to achieve stronger performance than the standard multi-head attention layer. Furthermore, the sparsely gated MoA can easily scale up the number of attention heads and the number of parameters while preserving computational efficiency. In addition to the performance improvements, MoA also automatically differentiates heads' utilities, providing a new perspective to discuss the model's interpretability. We conducted experiments on several important tasks, including Machine Translation and Masked Language Modeling. Experiments have shown promising results on several tasks against strong baselines that involve large and very deep models.", "authors": ["Xiaofeng Zhang", "Yikang Shen", "Zeyu Huang", "Jie Zhou", "Wenge Rong", "Zhang Xiong"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-11", "url": "https://arxiv.org/abs/2210.05144", "pdf_url": "https://arxiv.org/pdf/2210.05144v1", "arxiv_id": "2210.05144", "doi": "10.48550/arXiv.2210.05144", "citation_count": 80, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.5}
{"id": "e3ba94382393222f0ed7fab8936af26a1f8a16b3ae9f2ae5864cde42cef7ad00", "sources": ["arxiv", "semantic_scholar"], "title": "LARF: Two-level Attention-based Random Forests with a Mixture of Contamination Models", "abstract": "New models of the attention-based random forests called LARF (Leaf Attention-based Random Forest) are proposed. The first idea behind the models is to introduce a two-level attention, where one of the levels is the \"leaf\" attention and the attention mechanism is applied to every leaf of trees. The second level is the tree attention depending on the \"leaf\" attention. The second idea is to replace the softmax operation in the attention with the weighted sum of the softmax operations with different parameters. It is implemented by applying a mixture of the Huber's contamination models and can be regarded as an analog of the multi-head attention with \"heads\" defined by selecting a value of the softmax parameter. Attention parameters are simply trained by solving the quadratic optimization problem. To simplify the tuning process of the models, it is proposed to make the tuning contamination parameters to be training and to compute them by solving the quadratic optimization problem. Many numerical experiments with real datasets are performed for studying LARFs. The code of proposed algorithms can be found in https://github.com/andruekonst/leaf-attention-forest.", "authors": ["Andrei V. Konstantinov", "Lev V. Utkin"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-10-11", "url": "https://arxiv.org/abs/2210.05168", "pdf_url": "https://arxiv.org/pdf/2210.05168v1", "arxiv_id": "2210.05168", "doi": "10.48550/arXiv.2210.05168", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/andruekonst/leaf-attention-forest", "venue": "Informatics", "quality_score": 0.0}
{"id": "4199ad341e0691983c15bc2fa71610ceef03f2e0f357b6fec1603c5cf218d942", "sources": ["arxiv", "semantic_scholar"], "title": "FEAMOE: Fair, Explainable and Adaptive Mixture of Experts", "abstract": "Three key properties that are desired of trustworthy machine learning models deployed in high-stakes environments are fairness, explainability, and an ability to account for various kinds of \"drift\". While drifts in model accuracy, for example due to covariate shift, have been widely investigated, drifts in fairness metrics over time remain largely unexplored. In this paper, we propose FEAMOE, a novel \"mixture-of-experts\" inspired framework aimed at learning fairer, more explainable/interpretable models that can also rapidly adjust to drifts in both the accuracy and the fairness of a classifier. We illustrate our framework for three popular fairness measures and demonstrate how drift can be handled with respect to these fairness constraints. Experiments on multiple datasets show that our framework as applied to a mixture of linear experts is able to perform comparably to neural networks in terms of accuracy while producing fairer models. We then use the large-scale HMDA dataset and show that while various models trained on HMDA demonstrate drift with respect to both accuracy and fairness, FEAMOE can ably handle these drifts with respect to all the considered fairness measures and maintain model accuracy as well. We also prove that the proposed framework allows for producing fast Shapley value explanations, which makes computationally efficient feature attribution based explanations of model decisions readily available via FEAMOE.", "authors": ["Shubham Sharma", "Jette Henderson", "Joydeep Ghosh"], "categories": ["cs.LG", "cs.AI", "cs.CY"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-10", "url": "https://arxiv.org/abs/2210.04995", "pdf_url": "https://arxiv.org/pdf/2210.04995v1", "arxiv_id": "2210.04995", "doi": "10.48550/arXiv.2210.04995", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Joint Conference on Artificial Intelligence", "quality_score": 0.2113}
{"id": "a4d4e3deb859bc4a61a949fb9372ecf44b8f7d438ce9b5d394ea8b4d253ad8f7", "sources": ["arxiv", "semantic_scholar"], "title": "The Galactic Chemical Evolution of phosphorus observed with IGRINS", "abstract": "Phosphorus (P) is considered to be one of the key elements for life, making it an important element to look for in the abundance analysis of spectra of stellar systems. Yet, there exists only a handful of spectroscopic studies to estimate the P abundances and investigate its trend across a range of metallicities. We have observed full HK band spectra at a spectral resolving power of R=45,000 with IGRINS instrument. Abundances are determined using SME in combination with 1D MARCS stellar atmosphere models. The investigated sample of stars have reliable stellar parameters estimated using optical FIES spectra (GILD; Jönsson et al. in prep.). In order to determine the P abundances from the 16482.92 Angstrom P line, we take special care of the CO($ν=7-4$) blend. We determine the C, N, O abundances from atomic carbon and a range of non-blended molecular lines (CO, CN, OH) which are aplenty in the H band region of K giant stars, assuring an appropriate modelling of the blending CO($ν=7-4$) line. We present [P/Fe] vs [Fe/H] trend for 38 K giant stars in the metallicity range of -1.2 dex $<$ [Fe/H] $<$ 0.4 dex. We find that our trend matches well with the compiled literature sample of prominently dwarf stars and limited number of giant stars. Our trend is found to be higher by $\\sim$ 0.05 - 0.1 dex compared to the theoretical chemical evolution trend in Cescutti et al. 2012 resulting from core collapse supernova (type II) of massive stars with the P yields from Kobayashi et al. (2006) arbitrarily increased by a factor of 2.75. Thus the enhancement factor might need to be $\\sim$ 0.05 - 0.1 dex higher to match our trend. We also find an empirically determined primary behaviour for phosphorus. Furthermore, the phosphorus abundance is found to be elevated by $\\sim$ 0.6 - 0.9 dex in two metal poor s-enriched stars compared to the theoretical chemical evolution trend.", "authors": ["G. Nandakumar", "N. Ryde", "M. Montelius", "B. Thorsbro", "H. Jönsson", "G. Mace", "Lund Observatory", "Department of Astronomy", "Theoretical Physics", "Lund University", "Box 43", "SE-221 00 Lund", " Sweden", "Kapteyn Astronomical Institute", "University of Groningen", "Landleven 12", "NL-9747 AD Groningen", "the Netherlands", "Department of Astronomy", "School of Science", "The University of Tokyo", "7-3-1 Hongo", " Bunkyo-ku", "Tokyo 113-0033", " Japan", "Materials Science", "Applied Mathematics", "Malmö University", "SE-205 06 Malmö", " Sweden", "Department of Astronomy", "McDonald Observatory", "The University of Texas", " Austin", "TX 78712", " USA"], "categories": ["astro-ph.SR", "astro-ph.GA"], "fields_of_study": ["Physics"], "published_date": "2022-10-10", "url": "https://arxiv.org/abs/2210.04940", "pdf_url": "https://arxiv.org/pdf/2210.04940v1", "arxiv_id": "2210.04940", "doi": "10.1051/0004-6361/202244724", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "A&A 668, A88 (2022)", "quality_score": 0.25}
{"id": "b212e2f29bd7186ae518c9133fba6497558990be77a6fb524965a3cc2787b75d", "sources": ["arxiv", "semantic_scholar"], "title": "Modular Approach to Machine Reading Comprehension: Mixture of Task-Aware Experts", "abstract": "In this work we present a Mixture of Task-Aware Experts Network for Machine Reading Comprehension on a relatively small dataset. We particularly focus on the issue of common-sense learning, enforcing the common ground knowledge by specifically training different expert networks to capture different kinds of relationships between each passage, question and choice triplet. Moreover, we take inspi ration on the recent advancements of multitask and transfer learning by training each network a relevant focused task. By making the mixture-of-networks aware of a specific goal by enforcing a task and a relationship, we achieve state-of-the-art results and reduce over-fitting.", "authors": ["Anirudha Rayasam", "Anusha Kamath", "Gabriel Bayomi Tinoco Kalejaiye"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-04", "url": "https://arxiv.org/abs/2210.01750", "pdf_url": "https://arxiv.org/pdf/2210.01750v1", "arxiv_id": "2210.01750", "doi": "10.48550/arXiv.2210.01750", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "7c99a31a2ccd0e75e053c1073295bdef2ed9f2ec25f746d90aa36b694699f052", "sources": ["arxiv", "semantic_scholar"], "title": "Cooperation in the Latent Space: The Benefits of Adding Mixture Components in Variational Autoencoders", "abstract": "In this paper, we show how the mixture components cooperate when they jointly adapt to maximize the ELBO. We build upon recent advances in the multiple and adaptive importance sampling literature. We then model the mixture components using separate encoder networks and show empirically that the ELBO is monotonically non-decreasing as a function of the number of mixture components. These results hold for a range of different VAE architectures on the MNIST, FashionMNIST, and CIFAR-10 datasets. In this work, we also demonstrate that increasing the number of mixture components improves the latent-representation capabilities of the VAE on both image and single-cell datasets. This cooperative behavior motivates that using Mixture VAEs should be considered a standard approach for obtaining more flexible variational approximations. Finally, Mixture VAEs are here, for the first time, compared and combined with normalizing flows, hierarchical models and/or the VampPrior in an extensive ablation study. Multiple of our Mixture VAEs achieve state-of-the-art log-likelihood results for VAE architectures on the MNIST and FashionMNIST datasets. The experiments are reproducible using our code, provided here: https://github.com/lagergren-lab/mixturevaes.", "authors": ["Oskar Kviman", "Ricky Molén", "Alexandra Hotti", "Semih Kurt", "Víctor Elvira", "Jens Lagergren"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-09-30", "url": "https://arxiv.org/abs/2209.15514", "pdf_url": "https://arxiv.org/pdf/2209.15514v2", "arxiv_id": "2209.15514", "doi": null, "citation_count": 16, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/lagergren-lab/mixturevaes", "venue": "International Conference on Machine Learning", "quality_score": 0.3076}
{"id": "4548a7d2c536ac5b0ecfbbeb8333268c7c770389950adff4cb61fc4f7a389b9a", "sources": ["arxiv", "semantic_scholar"], "title": "Convolutional Neural Networks Quantization with Attention", "abstract": "It has been proven that, compared to using 32-bit floating-point numbers in the training phase, Deep Convolutional Neural Networks (DCNNs) can operate with low precision during inference, thereby saving memory space and power consumption. However, quantizing networks is always accompanied by an accuracy decrease. Here, we propose a method, double-stage Squeeze-and-Threshold (double-stage ST). It uses the attention mechanism to quantize networks and achieve state-of-art results. Using our method, the 3-bit model can achieve accuracy that exceeds the accuracy of the full-precision baseline model. The proposed double-stage ST activation quantization is easy to apply: inserting it before the convolution.", "authors": ["Binyi Wu", "Bernd Waschneck", "Christian Georg Mayr"], "categories": ["cs.AI", "cs.CV"], "fields_of_study": ["Computer Science", "Medicine"], "published_date": "2022-09-30", "url": "https://arxiv.org/abs/2209.15317", "pdf_url": "https://arxiv.org/pdf/2209.15317v1", "arxiv_id": "2209.15317", "doi": "10.1142/S0129065722500514", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Journal of Neural Systems", "quality_score": 0.1505}
{"id": "98228d5491133477a1d500c8f1ea51198fdf4c2ddf04c72f37f0a28c851d7b01", "sources": ["arxiv", "semantic_scholar"], "title": "Dilated Neighborhood Attention Transformer", "abstract": "Transformers are quickly becoming one of the most heavily applied deep learning architectures across modalities, domains, and tasks. In vision, on top of ongoing efforts into plain transformers, hierarchical transformers have also gained significant attention, thanks to their performance and easy integration into existing frameworks. These models typically employ localized attention mechanisms, such as the sliding-window Neighborhood Attention (NA) or Swin Transformer's Shifted Window Self Attention. While effective at reducing self attention's quadratic complexity, local attention weakens two of the most desirable properties of self attention: long range inter-dependency modeling, and global receptive field. In this paper, we introduce Dilated Neighborhood Attention (DiNA), a natural, flexible and efficient extension to NA that can capture more global context and expand receptive fields exponentially at no additional cost. NA's local attention and DiNA's sparse global attention complement each other, and therefore we introduce Dilated Neighborhood Attention Transformer (DiNAT), a new hierarchical vision transformer built upon both. DiNAT variants enjoy significant improvements over strong baselines such as NAT, Swin, and ConvNeXt. Our large model is faster and ahead of its Swin counterpart by 1.6% box AP in COCO object detection, 1.4% mask AP in COCO instance segmentation, and 1.4% mIoU in ADE20K semantic segmentation. Paired with new frameworks, our large variant is the new state of the art panoptic segmentation model on COCO (58.5 PQ) and ADE20K (49.4 PQ), and instance segmentation model on Cityscapes (45.1 AP) and ADE20K (35.4 AP) (no extra data). It also matches the state of the art specialized semantic segmentation models on ADE20K (58.1 mIoU), and ranks second on Cityscapes (84.5 mIoU) (no extra data).", "authors": ["Ali Hassani", "Humphrey Shi"], "categories": ["cs.CV", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-29", "url": "https://arxiv.org/abs/2209.15001", "pdf_url": "https://arxiv.org/pdf/2209.15001v3", "arxiv_id": "2209.15001", "doi": "10.48550/arXiv.2209.15001", "citation_count": 118, "influential_citation_count": 17, "has_code": true, "code_url": "https://github.com/SHI-Labs/Neighborhood-Attention-Transformer", "venue": "arXiv.org", "quality_score": 0.6276}
{"id": "26f0892ad98727aaeeb247206310bd98b6c4c42663cec50609cb52e57af9b800", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Sparse ViT: Towards Learnable Adaptive Token Pruning by Fully Exploiting Self-Attention", "abstract": "Vision transformer has emerged as a new paradigm in computer vision, showing excellent performance while accompanied by expensive computational cost. Image token pruning is one of the main approaches for ViT compression, due to the facts that the complexity is quadratic with respect to the token number, and many tokens containing only background regions do not truly contribute to the final prediction. Existing works either rely on additional modules to score the importance of individual tokens, or implement a fixed ratio pruning strategy for different input instances. In this work, we propose an adaptive sparse token pruning framework with a minimal cost. Specifically, we firstly propose an inexpensive attention head importance weighted class attention scoring mechanism. Then, learnable parameters are inserted as thresholds to distinguish informative tokens from unimportant ones. By comparing token attention scores and thresholds, we can discard useless tokens hierarchically and thus accelerate inference. The learnable thresholds are optimized in budget-aware training to balance accuracy and complexity, performing the corresponding pruning configurations for different input instances. Extensive experiments demonstrate the effectiveness of our approach. Our method improves the throughput of DeiT-S by 50% and brings only 0.2% drop in top-1 accuracy, which achieves a better trade-off between accuracy and latency than the previous methods.", "authors": ["Xiangcheng Liu", "Tianyi Wu", "Guodong Guo"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-28", "url": "https://arxiv.org/abs/2209.13802", "pdf_url": "https://arxiv.org/pdf/2209.13802v2", "arxiv_id": "2209.13802", "doi": "10.48550/arXiv.2209.13802", "citation_count": 50, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/Cydia2018/AS-ViT", "venue": "International Joint Conference on Artificial Intelligence", "quality_score": 0.4269}
{"id": "31b35b10dc78e5dbc47111da9df3d7e6a753af53ff356962f8017abbd6001008", "sources": ["arxiv", "semantic_scholar"], "title": "Quantifying attention via dwell time and engagement in a social media browsing environment", "abstract": "Modern computational systems have an unprecedented ability to detect, leverage and influence human attention. Prior work identified user engagement and dwell time as two key metrics of attention in digital environments, but these metrics have yet to be integrated into a unified model that can advance the theory andpractice of digital attention. We draw on work from cognitive science, digital advertising, and AI to propose a two-stage model of attention for social media environments that disentangles engagement and dwell. In an online experiment, we show that attention operates differently in these two stages and find clear evidence of dissociation: when dwelling on posts (Stage 1), users attend more to sensational than credible content, but when deciding whether to engage with content (Stage 2), users attend more to credible than sensational content. These findings have implications for the design and development of computational systems that measure and model human attention, such as newsfeed algorithms on social media.", "authors": ["Ziv Epstein", "Hause Lin", "Gordon Pennycook", "David Rand"], "categories": ["cs.HC", "cs.CY"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-21", "url": "https://arxiv.org/abs/2209.10464", "pdf_url": "https://arxiv.org/pdf/2209.10464v2", "arxiv_id": "2209.10464", "doi": "10.48550/arXiv.2209.10464", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "a8c248368ba03b8ee0ad52484644bdae6498ec211d5591d34a1c09a2146dd30c", "sources": ["arxiv", "semantic_scholar"], "title": "Flash-Cosmos: In-Flash Bulk Bitwise Operations Using Inherent Computation Capability of NAND Flash Memory", "abstract": "Bulk bitwise operations, i.e., bitwise operations on large bit vectors, are prevalent in a wide range of important application domains, including databases, graph processing, genome analysis, cryptography, and hyper-dimensional computing. In conventional systems, the performance and energy efficiency of bulk bitwise operations are bottlenecked by data movement between the compute units and the memory hierarchy. In-flash processing (i.e., processing data inside NAND flash chips) has a high potential to accelerate bulk bitwise operations by fundamentally reducing data movement through the entire memory hierarchy. We identify two key limitations of the state-of-the-art in-flash processing technique for bulk bitwise operations; (i) it falls short of maximally exploiting the bit-level parallelism of bulk bitwise operations; (ii) it is unreliable because it does not consider the highly error-prone nature of NAND flash memory. We propose Flash-Cosmos (Flash Computation with One-Shot Multi-Operand Sensing), a new in-flash processing technique that significantly increases the performance and energy efficiency of bulk bitwise operations while providing high reliability. Flash-Cosmos introduces two key mechanisms that can be easily supported in modern NAND flash chips: (i) Multi-Wordline Sensing (MWS), which enables bulk bitwise operations on a large number of operands with a single sensing operation, and (ii) Enhanced SLC-mode Programming (ESP), which enables reliable computation inside NAND flash memory. We demonstrate the feasibility of performing bulk bitwise operations with high reliability in Flash-Cosmos by testing 160 real 3D NAND flash chips. Our evaluation shows that Flash-Cosmos improves average performance and energy efficiency by 3.5x/32x and 3.3x/95x, respectively, over the state-of-the-art in-flash/outside-storage processing techniques across three real-world applications.", "authors": ["Jisung Park", "Roknoddin Azizi", "Geraldo F. Oliveira", "Mohammad Sadrosadati", "Rakesh Nadig", "David Novo", "Juan Gómez-Luna", "Myungsuk Kim", "Onur Mutlu"], "categories": ["cs.AR", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-12", "url": "https://arxiv.org/abs/2209.05566", "pdf_url": "https://arxiv.org/pdf/2209.05566v1", "arxiv_id": "2209.05566", "doi": "10.1109/MICRO56248.2022.00069", "citation_count": 54, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Micro", "quality_score": 0.4351}
{"id": "ec1582ed471a9f039442dd481d19f195c32d29954bfdf6e0ceb660b40ba74c6a", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Attention Acceleration with Synergistic In-Memory Pruning and On-Chip Recomputation", "abstract": "As its core computation, a self-attention mechanism gauges pairwise correlations across the entire input sequence. Despite favorable performance, calculating pairwise correlations is prohibitively costly. While recent work has shown the benefits of runtime pruning of elements with low attention scores, the quadratic complexity of self-attention mechanisms and their on-chip memory capacity demands are overlooked. This work addresses these constraints by architecting an accelerator, called SPRINT, which leverages the inherent parallelism of ReRAM crossbar arrays to compute attention scores in an approximate manner. Our design prunes the low attention scores using a lightweight analog thresholding circuitry within ReRAM, enabling SPRINT to fetch only a small subset of relevant data to on-chip memory. To mitigate potential negative repercussions for model accuracy, SPRINT re-computes the attention scores for the few fetched data in digital. The combined in-memory pruning and on-chip recompute of the relevant attention scores enables SPRINT to transform quadratic complexity to a merely linear one. In addition, we identify and leverage a dynamic spatial locality between the adjacent attention operations even after pruning, which eliminates costly yet redundant data fetches. We evaluate our proposed technique on a wide range of state-of-the-art transformer models. On average, SPRINT yields 7.5x speedup and 19.6x energy reduction when total 16KB on-chip memory is used, while virtually on par with iso-accuracy of the baseline models (on average 0.36% degradation).", "authors": ["Amir Yazdanbakhsh", "Ashkan Moradifirouzabadi", "Zheng Li", "Mingu Kang"], "categories": ["cs.LG", "cs.AR"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-01", "url": "https://arxiv.org/abs/2209.00606", "pdf_url": "https://arxiv.org/pdf/2209.00606v1", "arxiv_id": "2209.00606", "doi": "10.1109/MICRO56248.2022.00059", "citation_count": 48, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Micro", "quality_score": 0.4225}
{"id": "cf6b5f341b261f899e799c335f51ea11fc76dfcc101bd80df7df40cba957b418", "sources": ["arxiv", "semantic_scholar"], "title": "Mixtures of Gaussian Process Experts with SMC$^2$", "abstract": "Gaussian processes are a key component of many flexible statistical and machine learning models. However, they exhibit cubic computational complexity and high memory constraints due to the need of inverting and storing a full covariance matrix. To circumvent this, mixtures of Gaussian process experts have been considered where data points are assigned to independent experts, reducing the complexity by allowing inference based on smaller, local covariance matrices. Moreover, mixtures of Gaussian process experts substantially enrich the model's flexibility, allowing for behaviors such as non-stationarity, heteroscedasticity, and discontinuities. In this work, we construct a novel inference approach based on nested sequential Monte Carlo samplers to simultaneously infer both the gating network and Gaussian process expert parameters. This greatly improves inference compared to importance sampling, particularly in settings when a stationary Gaussian process is inappropriate, while still being thoroughly parallelizable.", "authors": ["Teemu Härkönen", "Sara Wade", "Kody Law", "Lassi Roininen"], "categories": ["stat.ML", "cs.LG", "stat.CO"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2022-08-26", "url": "https://arxiv.org/abs/2208.12830", "pdf_url": "https://arxiv.org/pdf/2208.12830v3", "arxiv_id": "2208.12830", "doi": "10.48550/arXiv.2208.12830", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Journal of machine learning research", "quality_score": 0.1747}
{"id": "7017b73181867e8e1fd8228b93e49a7574c8854d32f1ff8586aa34da75903516", "sources": ["arxiv", "semantic_scholar"], "title": "Optimal Brain Compression: A Framework for Accurate Post-Training Quantization and Pruning", "abstract": "We consider the problem of model compression for deep neural networks (DNNs) in the challenging one-shot/post-training setting, in which we are given an accurate trained model, and must compress it without any retraining, based only on a small amount of calibration input data. This problem has become popular in view of the emerging software and hardware support for executing models compressed via pruning and/or quantization with speedup, and well-performing solutions have been proposed independently for both compression approaches. In this paper, we introduce a new compression framework which covers both weight pruning and quantization in a unified setting, is time- and space-efficient, and considerably improves upon the practical performance of existing post-training methods. At the technical level, our approach is based on an exact and efficient realization of the classical Optimal Brain Surgeon (OBS) framework of [LeCun, Denker, and Solla, 1990] extended to also cover weight quantization at the scale of modern DNNs. From the practical perspective, our experimental results show that it can improve significantly upon the compression-accuracy trade-offs of existing post-training methods, and that it can enable the accurate compound application of both pruning and quantization in a post-training setting.", "authors": ["Elias Frantar", "Sidak Pal Singh", "Dan Alistarh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-08-24", "url": "https://arxiv.org/abs/2208.11580", "pdf_url": "https://arxiv.org/pdf/2208.11580v2", "arxiv_id": "2208.11580", "doi": "10.48550/arXiv.2208.11580", "citation_count": 390, "influential_citation_count": 65, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.9098}
{"id": "c6d5008da77423ca0b3d63cad0f17db355a9f27aa0edaaf5a391d19d811b7ee6", "sources": ["arxiv", "semantic_scholar"], "title": "How to Secure Valid Quantizations", "abstract": "Classical mechanics involves position and momentum variables that must be special coordinates chosen to promote to suitable quantum operators. Since classical variables may be broadly chosen, only unique variables should be chosen. We will outline how the favored variables and their suitable quantum operators is guaranteed to assure a truly valid quantization. Invalid quantizations may be mistaken for valid ones, which then leads to incorrect physics. Besides particle examples, there is also a brief run-through for fields and gravity.", "authors": ["John R. Klauder"], "categories": ["physics.gen-ph", "gr-qc"], "fields_of_study": ["Medicine", "Computer Science", "Physics"], "published_date": "2022-08-18", "url": "https://arxiv.org/abs/2209.03180", "pdf_url": "https://arxiv.org/pdf/2209.03180v1", "arxiv_id": "2209.03180", "doi": "10.3390/e24101374", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Entropy", "quality_score": 0.0753}
{"id": "40a1166f098bb8a826105884f135679c114958fce9b730e4c9168e31ced1e28f", "sources": ["arxiv", "semantic_scholar"], "title": "Context-aware Mixture-of-Experts for Unbiased Scene Graph Generation", "abstract": "Scene graph generation (SGG) has gained tremendous progress in recent years. However, its underlying long-tailed distribution of predicate classes is a challenging problem. For extremely unbalanced predicate distributions, existing approaches usually construct complicated context encoders to extract the intrinsic relevance of scene context to predicates and complex networks to improve the learning ability of network models for highly imbalanced predicate distributions. To address the unbiased SGG problem, we introduce a simple yet effective method dubbed Context-Aware Mixture-of-Experts (CAME) to improve model diversity and mitigate biased SGG without complicated design. Specifically, we propose to integrate the mixture of experts with a divide and ensemble strategy to remedy the severely long-tailed distribution of predicate classes, which is applicable to the majority of unbiased scene graph generators. The biased SGG is thereby reduced, and the model tends to anticipate more evenly distributed predicate predictions. To differentiate between various predicate distribution levels, experts with the same weights are not sufficiently diverse. In order to enable the network dynamically exploit the rich scene context and further boost the diversity of model, we simply use the built-in module to create a context encoder. The importance of each expert to scene context and each predicate to each expert is dynamically associated with expert weighting (EW) and predicate weighting (PW) strategy. We have conducted extensive experiments on three tasks using the Visual Genome dataset, showing that CAME outperforms recent methods and achieves state-of-the-art performance. Our code will be available publicly.", "authors": ["Liguang Zhou", "Yuhongze Zhou", "Tin Lun Lam", "Yangsheng Xu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-08-15", "url": "https://arxiv.org/abs/2208.07109", "pdf_url": "https://arxiv.org/pdf/2208.07109v3", "arxiv_id": "2208.07109", "doi": "10.48550/arXiv.2208.07109", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "a4b1d1639c6a61345d5b2f8157bcab5bbe27516fef3175e1a193795f65204cce", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Understanding Mixture of Experts in Deep Learning", "abstract": "The Mixture-of-Experts (MoE) layer, a sparsely-activated model controlled by a router, has achieved great success in deep learning. However, the understanding of such architecture remains elusive. In this paper, we formally study how the MoE layer improves the performance of neural network learning and why the mixture model will not collapse into a single model. Our empirical results suggest that the cluster structure of the underlying problem and the non-linearity of the expert are pivotal to the success of MoE. To further understand this, we consider a challenging classification problem with intrinsic cluster structures, which is hard to learn using a single expert. Yet with the MoE layer, by choosing the experts as two-layer nonlinear convolutional neural networks (CNNs), we show that the problem can be learned successfully. Furthermore, our theory shows that the router can learn the cluster-center features, which helps divide the input complex problem into simpler linear classification sub-problems that individual experts can conquer. To our knowledge, this is the first result towards formally understanding the mechanism of the MoE layer for deep learning.", "authors": ["Zixiang Chen", "Yihe Deng", "Yue Wu", "Quanquan Gu", "Yuanzhi Li"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-08-04", "url": "https://arxiv.org/abs/2208.02813", "pdf_url": "https://arxiv.org/pdf/2208.02813v1", "arxiv_id": "2208.02813", "doi": "10.48550/arXiv.2208.02813", "citation_count": 90, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4898}
{"id": "b5d3bb2a31726d3f9933e4b2611c01b0a09e74a9eab3aa9ecb76572476589da3", "sources": ["arxiv", "semantic_scholar"], "title": "Keyword Spotting System and Evaluation of Pruning and Quantization Methods on Low-power Edge Microcontrollers", "abstract": "Keyword spotting (KWS) is beneficial for voice-based user interactions with low-power devices at the edge. The edge devices are usually always-on, so edge computing brings bandwidth savings and privacy protection. The devices typically have limited memory spaces, computational performances, power and costs, for example, Cortex-M based microcontrollers. The challenge is to meet the high computation and low-latency requirements of deep learning on these devices. This paper firstly shows our small-footprint KWS system running on STM32F7 microcontroller with Cortex-M7 core @216MHz and 512KB static RAM. Our selected convolutional neural network (CNN) architecture has simplified number of operations for KWS to meet the constraint of edge devices. Our baseline system generates classification results for each 37ms including real-time audio feature extraction part. This paper further evaluates the actual performance for different pruning and quantization methods on microcontroller, including different granularity of sparsity, skipping zero weights, weight-prioritized loop order, and SIMD instruction. The result shows that for microcontrollers, there are considerable challenges for accelerate unstructured pruned models, and the structured pruning is more friendly than unstructured pruning. The result also verified that the performance improvement for quantization and SIMD instruction.", "authors": ["Jingyi Wang", "Shengchen Li"], "categories": ["cs.SD", "cs.AI", "cs.AR", "eess.AS"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2022-08-04", "url": "https://arxiv.org/abs/2208.02765", "pdf_url": "https://arxiv.org/pdf/2208.02765v1", "arxiv_id": "2208.02765", "doi": "10.48550/arXiv.2208.02765", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/RoboBachelor/Keyword-Spotting-STM32", "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "db9efd0067962c5ee4a16ed974906e7fa6b0363de35250417a581bbbb485780a", "sources": ["arxiv", "semantic_scholar"], "title": "BotBuster: Multi-platform Bot Detection Using A Mixture of Experts", "abstract": "Despite rapid development, current bot detection models still face challenges in dealing with incomplete data and cross-platform applications. In this paper, we propose BotBuster, a social bot detector built with the concept of a mixture of experts approach. Each expert is trained to analyze a portion of account information, e.g. username, and are combined to estimate the probability that the account is a bot. Experiments on 10 Twitter datasets show that BotBuster outperforms popular bot-detection baselines (avg F1=73.54 vs avg F1=45.12). This is accompanied with F1=60.04 on a Reddit dataset and F1=60.92 on an external evaluation set. Further analysis shows that only 36 posts is required for a stable bot classification. Investigation shows that bot post features have changed across the years and can be difficult to differentiate from human features, making bot detection a difficult and ongoing problem.", "authors": ["Lynnette Hui Xian Ng", "Kathleen M. Carley"], "categories": ["cs.SI"], "fields_of_study": ["Computer Science"], "published_date": "2022-07-27", "url": "https://arxiv.org/abs/2207.13658", "pdf_url": "https://arxiv.org/pdf/2207.13658v1", "arxiv_id": "2207.13658", "doi": "10.48550/arXiv.2207.13658", "citation_count": 60, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Web and Social Media", "quality_score": 0.4463}
{"id": "18eb9e16ff4e8b13d10cebfd496b4f345b1e2c86a5500609a42ed269f4e842f2", "sources": ["arxiv", "semantic_scholar"], "title": "Seeing Far in the Dark with Patterned Flash", "abstract": "Flash illumination is widely used in imaging under low-light environments. However, illumination intensity falls off with propagation distance quadratically, which poses significant challenges for flash imaging at a long distance. We propose a new flash technique, named ``patterned flash'', for flash imaging at a long distance. Patterned flash concentrates optical power into a dot array. Compared with the conventional uniform flash where the signal is overwhelmed by the noise everywhere, patterned flash provides stronger signals at sparsely distributed points across the field of view to ensure the signals at those points stand out from the sensor noise. This enables post-processing to resolve important objects and details. Additionally, the patterned flash projects texture onto the scene, which can be treated as a structured light system for depth perception. Given the novel system, we develop a joint image reconstruction and depth estimation algorithm with a convolutional neural network. We build a hardware prototype and test the proposed flash technique on various scenes. The experimental results demonstrate that our patterned flash has significantly better performance at long distances in low-light environments.", "authors": ["Zhanghao Sun", "Jian Wang", "Yicheng Wu", "Shree Nayar"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2022-07-25", "url": "https://arxiv.org/abs/2207.12570", "pdf_url": "https://arxiv.org/pdf/2207.12570v1", "arxiv_id": "2207.12570", "doi": "10.48550/arXiv.2207.12570", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "European Conference on Computer Vision", "quality_score": 0.1505}
{"id": "5f2c016bb92912a4edbf8e2729e856cdb8587a288c8d02b2abf128049ad94cc7", "sources": ["arxiv", "semantic_scholar"], "title": "Adaptive Mixture of Experts Learning for Generalizable Face Anti-Spoofing", "abstract": "With various face presentation attacks emerging continually, face anti-spoofing (FAS) approaches based on domain generalization (DG) have drawn growing attention. Existing DG-based FAS approaches always capture the domain-invariant features for generalizing on the various unseen domains. However, they neglect individual source domains' discriminative characteristics and diverse domain-specific information of the unseen domains, and the trained model is not sufficient to be adapted to various unseen domains. To address this issue, we propose an Adaptive Mixture of Experts Learning (AMEL) framework, which exploits the domain-specific information to adaptively establish the link among the seen source domains and unseen target domains to further improve the generalization. Concretely, Domain-Specific Experts (DSE) are designed to investigate discriminative and unique domain-specific features as a complement to common domain-invariant features. Moreover, Dynamic Expert Aggregation (DEA) is proposed to adaptively aggregate the complementary information of each source expert based on the domain relevance to the unseen target domain. And combined with meta-learning, these modules work collaboratively to adaptively aggregate meaningful domain-specific information for the various unseen target domains. Extensive experiments and visualizations demonstrate the effectiveness of our method against the state-of-the-art competitors.", "authors": ["Qianyu Zhou", "Ke-Yue Zhang", "Taiping Yao", "Ran Yi", "Shouhong Ding", "Lizhuang Ma"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-07-20", "url": "https://arxiv.org/abs/2207.09868", "pdf_url": "https://arxiv.org/pdf/2207.09868v1", "arxiv_id": "2207.09868", "doi": "10.1145/3503161.3547769", "citation_count": 68, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "ACM Multimedia", "quality_score": 0.4597}
{"id": "f975165a5db65000bdaa39fe2dea19c55931357e72abdf06a2c90d21f68fe909", "sources": ["arxiv", "semantic_scholar"], "title": "MoEC: Mixture of Expert Clusters", "abstract": "Sparsely Mixture of Experts (MoE) has received great interest due to its promising scaling capability with affordable computational overhead. MoE converts dense layers into sparse experts, and utilizes a gated routing network to make experts conditionally activated. However, as the number of experts grows, MoE with outrageous parameters suffers from overfitting and sparse data allocation. Such problems are especially severe on tasks with limited data, thus hindering the progress for MoE models to improve performance by scaling up. In this work, we propose Mixture of Expert Clusters - a general approach to enable expert layers to learn more diverse and appropriate knowledge by imposing variance-based constraints on the routing stage. We further propose a cluster-level expert dropout strategy specifically designed for the expert cluster structure. Our experiments reveal that MoEC could improve performance on machine translation and natural language understanding tasks, and raise the performance upper bound for scaling up experts under limited data. We also verify that MoEC plays a positive role in mitigating overfitting and sparse data allocation.", "authors": ["Yuan Xie", "Shaohan Huang", "Tianyu Chen", "Furu Wei"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-07-19", "url": "https://arxiv.org/abs/2207.09094", "pdf_url": "https://arxiv.org/pdf/2207.09094v1", "arxiv_id": "2207.09094", "doi": "10.48550/arXiv.2207.09094", "citation_count": 27, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3618}
{"id": "8658e076e5a6bb56cdcfcc3f8d88c2227883077690961f77d52722f56c13f61a", "sources": ["arxiv", "semantic_scholar"], "title": "CoCAtt: A Cognitive-Conditioned Driver Attention Dataset (Supplementary Material)", "abstract": "The task of driver attention prediction has drawn considerable interest among researchers in robotics and the autonomous vehicle industry. Driver attention prediction can play an instrumental role in mitigating and preventing high-risk events, like collisions and casualties. However, existing driver attention prediction models neglect the distraction state and intention of the driver, which can significantly influence how they observe their surroundings. To address these issues, we present a new driver attention dataset, CoCAtt (Cognitive-Conditioned Attention). Unlike previous driver attention datasets, CoCAtt includes per-frame annotations that describe the distraction state and intention of the driver. In addition, the attention data in our dataset is captured in both manual and autopilot modes using eye-tracking devices of different resolutions. Our results demonstrate that incorporating the above two driver states into attention modeling can improve the performance of driver attention prediction. To the best of our knowledge, this work is the first to provide autopilot attention data. Furthermore, CoCAtt is currently the largest and the most diverse driver attention dataset in terms of autonomy levels, eye tracker resolutions, and driving scenarios. CoCAtt is available for download at https://cocatt-dataset.github.io.", "authors": ["Yuan Shen", "Niviru Wijayaratne", "Pranav Sriram", "Aamir Hasan", "Peter Du", "Katherine Driggs-Campbell"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-07-08", "url": "https://arxiv.org/abs/2207.04028", "pdf_url": "https://arxiv.org/pdf/2207.04028v1", "arxiv_id": "2207.04028", "doi": "10.48550/arXiv.2207.04028", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "456d58c93c548c067bcebdf0e74e6957a87b2d4ffa81ccc76d11a9f8d2666627", "sources": ["arxiv", "semantic_scholar"], "title": "Attention Round for Post-Training Quantization", "abstract": "At present, the quantification methods of neural network models are mainly divided into post-training quantization (PTQ) and quantization aware training (QAT). Post-training quantization only need a small part of the data to complete the quantification process, but the performance of its quantitative model is not as good as the quantization aware training. This paper presents a novel quantification method called Attention Round. This method gives parameters w the opportunity to be mapped to all possible quantized values, rather than just the two quantized values nearby w in the process of quantization. The probability of being mapped to different quantified values is negatively correlated with the distance between the quantified values and w, and decay with a Gaussian function. In addition, this paper uses the lossy coding length as a measure to assign bit widths to the different layers of the model to solve the problem of mixed precision quantization, which effectively avoids to solve combinatorial optimization problem. This paper also performs quantitative experiments on different models, the results confirm the effectiveness of the proposed method. For ResNet18 and MobileNetV2, the post-training quantization proposed in this paper only require 1,024 training data and 10 minutes to complete the quantization process, which can achieve quantization performance on par with quantization aware training.", "authors": ["Huabin Diao", "Gongyan Li", "Shaoyun Xu", "Yuexing Hao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-07-07", "url": "https://arxiv.org/abs/2207.03088", "pdf_url": "https://arxiv.org/pdf/2207.03088v1", "arxiv_id": "2207.03088", "doi": "10.48550/arXiv.2207.03088", "citation_count": 19, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neurocomputing", "quality_score": 0.3253}
{"id": "ce3ee6f42b634b3e903c02bc591f64bb54895ab7cf250f196ca9910afb45feba", "sources": ["arxiv", "semantic_scholar"], "title": "Multimodal Contrastive Learning with LIMoE: the Language-Image Mixture of Experts", "abstract": "Large sparsely-activated models have obtained excellent performance in multiple domains. However, such models are typically trained on a single modality at a time. We present the Language-Image MoE, LIMoE, a sparse mixture of experts model capable of multimodal learning. LIMoE accepts both images and text simultaneously, while being trained using a contrastive loss. MoEs are a natural fit for a multimodal backbone, since expert layers can learn an appropriate partitioning of modalities. However, new challenges arise; in particular, training stability and balanced expert utilization, for which we propose an entropy-based regularization scheme. Across multiple scales, we demonstrate remarkable performance improvement over dense models of equivalent computational cost. LIMoE-L/16 trained comparably to CLIP-L/14 achieves 78.6% zero-shot ImageNet accuracy (vs. 76.2%), and when further scaled to H/14 (with additional data) it achieves 84.1%, comparable to state-of-the-art methods which use larger custom per-modality backbones and pre-training schemes. We analyse the quantitative and qualitative behavior of LIMoE, and demonstrate phenomena such as differing treatment of the modalities and the organic emergence of modality-specific experts.", "authors": ["Basil Mustafa", "Carlos Riquelme", "Joan Puigcerver", "Rodolphe Jenatton", "Neil Houlsby"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-06-06", "url": "https://arxiv.org/abs/2206.02770", "pdf_url": "https://arxiv.org/pdf/2206.02770v1", "arxiv_id": "2206.02770", "doi": "10.48550/arXiv.2206.02770", "citation_count": 325, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.6283}
{"id": "84829b842d04d04fb3d0a5b640044b503db1edca74d43988b79a6c697b4a4da1", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Mixture of Experts", "abstract": "The need for reliable model explanations is prominent for many machine learning applications, particularly for tabular and time-series data as their use cases often involve high-stakes decision making. Towards this goal, we introduce a novel interpretable modeling framework, Interpretable Mixture of Experts (IME), that yields high accuracy, comparable to `black-box' Deep Neural Networks (DNNs) in many cases, along with useful interpretability capabilities. IME consists of an assignment module and a mixture of experts, with each sample being assigned to a single expert for prediction. We introduce multiple options for IME based on the assignment and experts being interpretable. When the experts are chosen to be interpretable such as linear models, IME yields an inherently-interpretable architecture where the explanations produced by IME are the exact descriptions of how the prediction is computed. In addition to constituting a standalone inherently-interpretable architecture, IME has the premise of being integrated with existing DNNs to offer interpretability to a subset of samples while maintaining the accuracy of the DNNs. Through extensive experiments on 15 tabular and time-series datasets, IME is demonstrated to be more accurate than single interpretable models and perform comparably with existing state-of-the-art DNNs in accuracy. On most datasets, IME even outperforms DNNs, while providing faithful explanations. Lastly, IME's explanations are compared to commonly-used post-hoc explanations methods through a user study -- participants are able to better predict the model behavior when given IME explanations, while finding IME's explanations more faithful and trustworthy.", "authors": ["Aya Abdelsalam Ismail", "Sercan Ö. Arik", "Jinsung Yoon", "Ankur Taly", "Soheil Feizi", "Tomas Pfister"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-06-05", "url": "https://arxiv.org/abs/2206.02107", "pdf_url": "https://arxiv.org/pdf/2206.02107v2", "arxiv_id": "2206.02107", "doi": null, "citation_count": 19, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3253}
{"id": "b6634fbf08a44317bf533b7efa0acf6818e85c2669c252704eed167b81c9b9b9", "sources": ["arxiv", "semantic_scholar"], "title": "A Valid Quantization of a Half-Harmonic Oscillator Field Theory", "abstract": "The usual full- and half-harmonic oscillators are turned into field theories, and that behavior is examined using canonical and affine quantization. The result leads to a valid affine quantization of the half harmonic oscillator field theory, which points toward further valid quantizations of more realistic field theory models.", "authors": ["John R. Klauder"], "categories": ["hep-th"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2022-06-04", "url": "https://arxiv.org/abs/2206.07653", "pdf_url": "https://arxiv.org/pdf/2206.07653v1", "arxiv_id": "2206.07653", "doi": "10.3390/axioms11080360", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Axioms", "quality_score": 0.1193}
{"id": "956ea02f017bd4d60ac9f052529c14fdf73222e1defb618c680402d4176411e6", "sources": ["arxiv", "semantic_scholar"], "title": "Patcher: Patch Transformers with Mixture of Experts for Precise Medical Image Segmentation", "abstract": "We present a new encoder-decoder Vision Transformer architecture, Patcher, for medical image segmentation. Unlike standard Vision Transformers, it employs Patcher blocks that segment an image into large patches, each of which is further divided into small patches. Transformers are applied to the small patches within a large patch, which constrains the receptive field of each pixel. We intentionally make the large patches overlap to enhance intra-patch communication. The encoder employs a cascade of Patcher blocks with increasing receptive fields to extract features from local to global levels. This design allows Patcher to benefit from both the coarse-to-fine feature extraction common in CNNs and the superior spatial relationship modeling of Transformers. We also propose a new mixture-of-experts (MoE) based decoder, which treats the feature maps from the encoder as experts and selects a suitable set of expert features to predict the label for each pixel. The use of MoE enables better specializations of the expert features and reduces interference between them during inference. Extensive experiments demonstrate that Patcher outperforms state-of-the-art Transformer- and CNN-based approaches significantly on stroke lesion segmentation and polyp segmentation. Code for Patcher is released with publication to facilitate future research.", "authors": ["Yanglan Ou", "Ye Yuan", "Xiaolei Huang", "Stephen T. C. Wong", "John Volpi", "James Z. Wang", "Kelvin Wong"], "categories": ["eess.IV", "cs.CV"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2022-06-03", "url": "https://arxiv.org/abs/2206.01741", "pdf_url": "https://arxiv.org/pdf/2206.01741v2", "arxiv_id": "2206.01741", "doi": "10.48550/arXiv.2206.01741", "citation_count": 21, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "International Conference on Medical Image Computing and Computer-Assisted Intervention", "quality_score": 0.3495}
{"id": "c785f0d799b0db5b0286520d8d0b502db5a6d404adf607914853dc2446490cde", "sources": ["arxiv", "semantic_scholar"], "title": "Task-Specific Expert Pruning for Sparse Mixture-of-Experts", "abstract": "The sparse Mixture-of-Experts (MoE) model is powerful for large-scale pre-training and has achieved promising results due to its model capacity. However, with trillions of parameters, MoE is hard to be deployed on cloud or mobile environment. The inference of MoE requires expert parallelism, which is not hardware-friendly and communication expensive. Especially for resource-limited downstream tasks, such sparse structure has to sacrifice a lot of computing efficiency for limited performance gains. In this work, we observe most experts contribute scarcely little to the MoE fine-tuning and inference. We further propose a general method to progressively drop the non-professional experts for the target downstream task, which preserves the benefits of MoE while reducing the MoE model into one single-expert dense model. Our experiments reveal that the fine-tuned single-expert model could preserve 99.3% benefits from MoE across six different types of tasks while enjoying 2x inference speed with free communication cost.", "authors": ["Tianyu Chen", "Shaohan Huang", "Yuan Xie", "Binxing Jiao", "Daxin Jiang", "Haoyi Zhou", "Jianxin Li", "Furu Wei"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-06-01", "url": "https://arxiv.org/abs/2206.00277", "pdf_url": "https://arxiv.org/pdf/2206.00277v2", "arxiv_id": "2206.00277", "doi": "10.48550/arXiv.2206.00277", "citation_count": 80, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4771}
{"id": "3a9bf9579dc8196cf3d745799499aff333d9c226daaba732543098d871099fa4", "sources": ["arxiv", "semantic_scholar"], "title": "A Mixture-of-Expert Approach to RL-based Dialogue Management", "abstract": "Despite recent advancements in language models (LMs), their application to dialogue management (DM) problems and ability to carry on rich conversations remain a challenge. We use reinforcement learning (RL) to develop a dialogue agent that avoids being short-sighted (outputting generic utterances) and maximizes overall user satisfaction. Most existing RL approaches to DM train the agent at the word-level, and thus, have to deal with a combinatorially complex action space even for a medium-size vocabulary. As a result, they struggle to produce a successful and engaging dialogue even if they are warm-started with a pre-trained LM. To address this issue, we develop a RL-based DM using a novel mixture of expert language model (MoE-LM) that consists of (i) a LM capable of learning diverse semantics for conversation histories, (ii) a number of {\\em specialized} LMs (or experts) capable of generating utterances corresponding to a particular attribute or personality, and (iii) a RL-based DM that performs dialogue planning with the utterances generated by the experts. Our MoE approach provides greater flexibility to generate sensible utterances with different intents and allows RL to focus on conversational-level DM. We compare it with SOTA baselines on open-domain dialogues and demonstrate its effectiveness both in terms of the diversity and sensibility of the generated utterances and the overall DM performance.", "authors": ["Yinlam Chow", "Aza Tulepbergenov", "Ofir Nachum", "MoonKyung Ryu", "Mohammad Ghavamzadeh", "Craig Boutilier"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-05-31", "url": "https://arxiv.org/abs/2206.00059", "pdf_url": "https://arxiv.org/pdf/2206.00059v1", "arxiv_id": "2206.00059", "doi": "10.48550/arXiv.2206.00059", "citation_count": 20, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3306}
{"id": "1236ddfb5a99e9db5256ec61d26e7a60563ee8bf3d12ef0d73dc8f80c31c2ccb", "sources": ["arxiv", "semantic_scholar"], "title": "FCN-Pose: A Pruned and Quantized CNN for Robot Pose Estimation for Constrained Devices", "abstract": "IoT devices suffer from resource limitations, such as processor, RAM, and disc storage. These limitations become more evident when handling demanding applications, such as deep learning, well-known for their heavy computational requirements. A case in point is robot pose estimation, an application that predicts the critical points of the desired image object. One way to mitigate processing and storage problems is compressing that deep learning application. This paper proposes a new CNN for the pose estimation while applying the compression techniques of pruning and quantization to reduce his demands and improve the response time. While the pruning process reduces the total number of parameters required for inference, quantization decreases the precision of the floating-point. We run the approach using a pose estimation task for a robotic arm and compare the results in a high-end device and a constrained device. As metrics, we consider the number of Floating-point Operations Per Second(FLOPS), the total of mathematical computations, the calculation of parameters, the inference time, and the number of video frames processed per second. In addition, we undertake a qualitative evaluation where we compare the output image predicted for each pruned network with the corresponding original one. We reduce the originally proposed network to a 70% pruning rate, implying an 88.86% reduction in parameters, 94.45% reduction in FLOPS, and for the disc storage, we reduced the requirement in 70% while increasing error by a mere $1\\%$. With regard input image processing, this metric increases from 11.71 FPS to 41.9 FPS for the Desktop case. When using the constrained device, image processing augmented from 2.86 FPS to 10.04 FPS. The higher processing rate of image frames achieved by the proposed approach allows a much shorter response time.", "authors": ["Marrone Silvério Melo Dantas", "Iago Richard Rodrigues", "Assis Tiago Oliveira Filho", "Gibson Barbosa", "Daniel Bezerra", "Djamel F. H. Sadok", "Judith Kelner", "Maria Marquezini", "Ricardo Silva"], "categories": ["cs.CV", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-05-26", "url": "https://arxiv.org/abs/2205.13272", "pdf_url": "https://arxiv.org/pdf/2205.13272v1", "arxiv_id": "2205.13272", "doi": "10.1007/s10846-025-02336-y", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Journal of Intelligent and Robotic Systems", "quality_score": 0.1193}
{"id": "1d8132a39ccef4b6a3ce7bda737c090cbf4c613c5396ec91e39149e79e620138", "sources": ["arxiv", "semantic_scholar"], "title": "OPQ: Compressing Deep Neural Networks with One-shot Pruning-Quantization", "abstract": "As Deep Neural Networks (DNNs) usually are overparameterized and have millions of weight parameters, it is challenging to deploy these large DNN models on resource-constrained hardware platforms, e.g., smartphones. Numerous network compression methods such as pruning and quantization are proposed to reduce the model size significantly, of which the key is to find suitable compression allocation (e.g., pruning sparsity and quantization codebook) of each layer. Existing solutions obtain the compression allocation in an iterative/manual fashion while finetuning the compressed model, thus suffering from the efficiency issue. Different from the prior art, we propose a novel One-shot Pruning-Quantization (OPQ) in this paper, which analytically solves the compression allocation with pre-trained weight parameters only. During finetuning, the compression module is fixed and only weight parameters are updated. To our knowledge, OPQ is the first work that reveals pre-trained model is sufficient for solving pruning and quantization simultaneously, without any complex iterative/manual optimization at the finetuning stage. Furthermore, we propose a unified channel-wise quantization method that enforces all channels of each layer to share a common codebook, which leads to low bit-rate allocation without introducing extra overhead brought by traditional channel-wise quantization. Comprehensive experiments on ImageNet with AlexNet/MobileNet-V1/ResNet-50 show that our method improves accuracy and training efficiency while obtains significantly higher compression rates compared to the state-of-the-art.", "authors": ["Peng Hu", "Xi Peng", "Hongyuan Zhu", "Mohamed M. Sabry Aly", "Jie Lin"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-05-23", "url": "https://arxiv.org/abs/2205.11141", "pdf_url": "https://arxiv.org/pdf/2205.11141v1", "arxiv_id": "2205.11141", "doi": "10.1609/aaai.v35i9.16950", "citation_count": 81, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4785}
{"id": "84b4924209aeb12ac34e1c7f0cb339092b97cda0a8fec69b3ae4686f1cb504b5", "sources": ["arxiv", "semantic_scholar"], "title": "Optimizing Mixture of Experts using Dynamic Recompilations", "abstract": "The Mixture of Experts architecture allows for outrageously large neural networks by scaling model parameter size independently from computational demand (FLOPs). However, current DNN frameworks cannot effectively support the dynamic data flow in Mixture of Experts, and implementations on top of these frameworks need to use workarounds that introduce significant overheads. To address the limitation of these frameworks, we present DynaMoE, a DNN library that uses dynamic recompilations to optimize and adapt the use of computational resources to the dynamic needs of Mixture of Experts models. Our evaluation shows that DynaMoE achieves a 1.8x speedup and supports 2.3x larger model sizes when compared to existing MoE systems, even when not using recompilations. We then present further optimizations enabled by dynamic recompilations that yield an additional 1.7x speedup while simultaneously reducing memory pressure and improving model quality.", "authors": ["Ferdinand Kossmann", "Zhihao Jia", "Alex Aiken"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2022-05-04", "url": "https://arxiv.org/abs/2205.01848", "pdf_url": "https://arxiv.org/pdf/2205.01848v2", "arxiv_id": "2205.01848", "doi": "10.48550/arXiv.2205.01848", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "e9d70a30a30e7967321d29fa0ef5bf13229c8ebe5d2928c7945fb605d469a781", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Feature Distribution Alignment and Diversity Enhancement for Data-Free Quantization", "abstract": "To obtain lower inference latency and less memory footprint of deep neural networks, model quantization has been widely employed in deep model deployment, by converting the floating points to low-precision integers. However, previous methods (such as quantization aware training and post training quantization) require original data for the fine-tuning or calibration of quantized model, which makes them inapplicable to the cases that original data are not accessed due to privacy or security. This gives birth to the data-free quantization method with synthetic data generation. While current data-free quantization methods still suffer from severe performance degradation when quantizing a model into lower bit, caused by the low inter-class separability of semantic features. To this end, we propose a new and effective data-free quantization method termed ClusterQ, which utilizes the feature distribution alignment for synthetic data generation. To obtain high inter-class separability of semantic features, we cluster and align the feature distribution statistics to imitate the distribution of real data, so that the performance degradation is alleviated. Moreover, we incorporate the diversity enhancement to solve class-wise mode collapse. We also employ the exponential moving average to update the centroid of each cluster for further feature distribution improvement. Extensive experiments based on different deep models (e.g., ResNet-18 and MobileNet-V2) over the ImageNet dataset demonstrate that our proposed ClusterQ model obtains state-of-the-art performance.", "authors": ["Yangcheng Gao", "Zhao Zhang", "Richang Hong", "Haijun Zhang", "Jicong Fan", "Shuicheng Yan"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-30", "url": "https://arxiv.org/abs/2205.00179", "pdf_url": "https://arxiv.org/pdf/2205.00179v2", "arxiv_id": "2205.00179", "doi": "10.1109/ICDM54844.2022.00024", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Industrial Conference on Data Mining", "quality_score": 0.2698}
{"id": "59d8f444dd68c76a769b5daef3d89cd65c606d5e749b1596a45787af2da5603c", "sources": ["arxiv", "semantic_scholar"], "title": "NeurMiPs: Neural Mixture of Planar Experts for View Synthesis", "abstract": "We present Neural Mixtures of Planar Experts (NeurMiPs), a novel planar-based scene representation for modeling geometry and appearance. NeurMiPs leverages a collection of local planar experts in 3D space as the scene representation. Each planar expert consists of the parameters of the local rectangular shape representing geometry and a neural radiance field modeling the color and opacity. We render novel views by calculating ray-plane intersections and composite output colors and densities at intersected points to the image. NeurMiPs blends the efficiency of explicit mesh rendering and flexibility of the neural radiance field. Experiments demonstrate superior performance and speed of our proposed method, compared to other 3D representations in novel view synthesis.", "authors": ["Zhi-Hao Lin", "Wei-Chiu Ma", "Hao-Yu Hsu", "Yu-Chiang Frank Wang", "Shenlong Wang"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-28", "url": "https://arxiv.org/abs/2204.13696", "pdf_url": "https://arxiv.org/pdf/2204.13696v1", "arxiv_id": "2204.13696", "doi": "10.1109/CVPR52688.2022.01525", "citation_count": 35, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.3891}
{"id": "a985d2df37b6521f3d731a94e299f35996059b3ca0930ae8ab2d74691e93663d", "sources": ["arxiv", "semantic_scholar"], "title": "AAU-net: An Adaptive Attention U-net for Breast Lesions Segmentation in Ultrasound Images", "abstract": "Various deep learning methods have been proposed to segment breast lesion from ultrasound images. However, similar intensity distributions, variable tumor morphology and blurred boundaries present challenges for breast lesions segmentation, especially for malignant tumors with irregular shapes. Considering the complexity of ultrasound images, we develop an adaptive attention U-net (AAU-net) to segment breast lesions automatically and stably from ultrasound images. Specifically, we introduce a hybrid adaptive attention module, which mainly consists of a channel self-attention block and a spatial self-attention block, to replace the traditional convolution operation. Compared with the conventional convolution operation, the design of the hybrid adaptive attention module can help us capture more features under different receptive fields. Different from existing attention mechanisms, the hybrid adaptive attention module can guide the network to adaptively select more robust representation in channel and space dimensions to cope with more complex breast lesions segmentation. Extensive experiments with several state-of-the-art deep learning segmentation methods on three public breast ultrasound datasets show that our method has better performance on breast lesion segmentation. Furthermore, robustness analysis and external experiments demonstrate that our proposed AAU-net has better generalization performance on the segmentation of breast lesions. Moreover, the hybrid adaptive attention module can be flexibly applied to existing network frameworks.", "authors": ["Gongping Chen", "Yu Dai", "Jianxun Zhang", "Moi Hoon Yap"], "categories": ["eess.IV", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science", "Engineering", "Medicine"], "published_date": "2022-04-26", "url": "https://arxiv.org/abs/2204.12077", "pdf_url": "https://arxiv.org/pdf/2204.12077v3", "arxiv_id": "2204.12077", "doi": "10.1109/TMI.2022.3226268", "citation_count": 315, "influential_citation_count": 14, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Medical Imaging", "quality_score": 0.6249}
{"id": "0cc7e93cdafc3428daf110ec1400da197576c2f098de61b454e992f57672cc07", "sources": ["arxiv", "semantic_scholar"], "title": "Coarse-to-fine Q-attention with Tree Expansion", "abstract": "Coarse-to-fine Q-attention enables sample-efficient robot manipulation by discretizing the translation space in a coarse-to-fine manner, where the resolution gradually increases at each layer in the hierarchy. Although effective, Q-attention suffers from \"coarse ambiguity\" - when voxelization is significantly coarse, it is not feasible to distinguish similar-looking objects without first inspecting at a finer resolution. To combat this, we propose to envision Q-attention as a tree that can be expanded and used to accumulate value estimates across the top-k voxels at each Q-attention depth. When our extension, Q-attention with Tree Expansion (QTE), replaces standard Q-attention in the Attention-driven Robot Manipulation (ARM) system, we are able to accomplish a larger set of tasks; especially on those that suffer from \"coarse ambiguity\". In addition to evaluating our approach across 12 RLBench tasks, we also show that the improved performance is visible in a real-world task involving small objects.", "authors": ["Stephen James", "Pieter Abbeel"], "categories": ["cs.RO", "cs.AI", "cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-26", "url": "https://arxiv.org/abs/2204.12471", "pdf_url": "https://arxiv.org/pdf/2204.12471v2", "arxiv_id": "2204.12471", "doi": "10.48550/arXiv.2204.12471", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "c914af26731e1c154a1801eccc32dfceee33706f6c12f65361c6babde60f7377", "sources": ["arxiv", "semantic_scholar"], "title": "Sparsely-gated Mixture-of-Expert Layers for CNN Interpretability", "abstract": "Sparsely-gated Mixture of Expert (MoE) layers have been recently successfully applied for scaling large transformers, especially for language modeling tasks. An intriguing side effect of sparse MoE layers is that they convey inherent interpretability to a model via natural expert specialization. In this work, we apply sparse MoE layers to CNNs for computer vision tasks and analyze the resulting effect on model interpretability. To stabilize MoE training, we present both soft and hard constraint-based approaches. With hard constraints, the weights of certain experts are allowed to become zero, while soft constraints balance the contribution of experts with an additional auxiliary loss. As a result, soft constraints handle expert utilization better and support the expert specialization process, while hard constraints maintain more generalized experts and increase overall model performance. Our findings demonstrate that experts can implicitly focus on individual sub-domains of the input space. For example, experts trained for CIFAR-100 image classification specialize in recognizing different domains such as flowers or animals without previous data clustering. Experiments with RetinaNet and the COCO dataset further indicate that object detection experts can also specialize in detecting objects of distinct sizes.", "authors": ["Svetlana Pavlitska", "Christian Hubschneider", "Lukas Struppek", "J. Marius Zöllner"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-22", "url": "https://arxiv.org/abs/2204.10598", "pdf_url": "https://arxiv.org/pdf/2204.10598v3", "arxiv_id": "2204.10598", "doi": "10.1109/IJCNN54540.2023.10191904", "citation_count": 21, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "IEEE International Joint Conference on Neural Network", "quality_score": 0.3356}
{"id": "121e40f91caf82937fb29f7494d54eb4789939a4cc40d349bf445d10481225fd", "sources": ["arxiv", "semantic_scholar"], "title": "On the Representation Collapse of Sparse Mixture of Experts", "abstract": "Sparse mixture of experts provides larger model capacity while requiring a constant computational overhead. It employs the routing mechanism to distribute input tokens to the best-matched experts according to their hidden representations. However, learning such a routing mechanism encourages token clustering around expert centroids, implying a trend toward representation collapse. In this work, we propose to estimate the routing scores between tokens and experts on a low-dimensional hypersphere. We conduct extensive experiments on cross-lingual language model pre-training and fine-tuning on downstream tasks. Experimental results across seven multilingual benchmarks show that our method achieves consistent gains. We also present a comprehensive analysis on the representation and routing behaviors of our models. Our method alleviates the representation collapse issue and achieves more consistent routing than the baseline mixture-of-experts methods.", "authors": ["Zewen Chi", "Li Dong", "Shaohan Huang", "Damai Dai", "Shuming Ma", "Barun Patra", "Saksham Singhal", "Payal Bajaj", "Xia Song", "Xian-Ling Mao", "Heyan Huang", "Furu Wei"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-20", "url": "https://arxiv.org/abs/2204.09179", "pdf_url": "https://arxiv.org/pdf/2204.09179v3", "arxiv_id": "2204.09179", "doi": "10.48550/arXiv.2204.09179", "citation_count": 188, "influential_citation_count": 23, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.6901}
{"id": "963ffb520cf2c7d94f0deab834e4b2e64ca38ae1001fb253f7bd1e06ff2e5326", "sources": ["arxiv", "semantic_scholar"], "title": "Residual Mixture of Experts", "abstract": "Mixture of Experts (MoE) is able to scale up vision transformers effectively. However, it requires prohibiting computation resources to train a large MoE transformer. In this paper, we propose Residual Mixture of Experts (RMoE), an efficient training pipeline for MoE vision transformers on downstream tasks, such as segmentation and detection. RMoE achieves comparable results with the upper-bound MoE training, while only introducing minor additional training cost than the lower-bound non-MoE training pipelines. The efficiency is supported by our key observation: the weights of an MoE transformer can be factored into an input-independent core and an input-dependent residual. Compared with the weight core, the weight residual can be efficiently trained with much less computation resource, e.g., finetuning on the downstream data. We show that, compared with the current MoE training pipeline, we get comparable results while saving over 30% training cost. When compared with state-of-the-art non- MoE transformers, such as Swin-T / CvT-13 / Swin-L, we get +1.1 / 0.9 / 1.0 mIoU gain on ADE20K segmentation and +1.4 / 1.6 / 0.6 AP gain on MS-COCO object detection task with less than 3% additional training cost.", "authors": ["Lemeng Wu", "Mengchen Liu", "Yinpeng Chen", "Dongdong Chen", "Xiyang Dai", "Lu Yuan"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-20", "url": "https://arxiv.org/abs/2204.09636", "pdf_url": "https://arxiv.org/pdf/2204.09636v3", "arxiv_id": "2204.09636", "doi": "10.48550/arXiv.2204.09636", "citation_count": 53, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4331}
{"id": "c5f073d63ac4345648539d7d8a12000b901a3dcd7dcef26c61998717acea9185", "sources": ["arxiv", "semantic_scholar"], "title": "Table-based Fact Verification with Self-adaptive Mixture of Experts", "abstract": "The table-based fact verification task has recently gained widespread attention and yet remains to be a very challenging problem. It inherently requires informative reasoning over natural language together with different numerical and logical reasoning on tables (e.g., count, superlative, comparative). Considering that, we exploit mixture-of-experts and present in this paper a new method: Self-adaptive Mixture-of-Experts Network (SaMoE). Specifically, we have developed a mixture-of-experts neural network to recognize and execute different types of reasoning -- the network is composed of multiple experts, each handling a specific part of the semantics for reasoning, whereas a management module is applied to decide the contribution of each expert network to the verification result. A self-adaptive method is developed to teach the management module combining results of different experts more efficiently without external knowledge. The experimental results illustrate that our framework achieves 85.1% accuracy on the benchmark dataset TabFact, comparable with the previous state-of-the-art models. We hope our framework can serve as a new baseline for table-based verification. Our code is available at https://github.com/THUMLP/SaMoE.", "authors": ["Yuxuan Zhou", "Xien Liu", "Kaiyin Zhou", "Ji Wu"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-19", "url": "https://arxiv.org/abs/2204.08753", "pdf_url": "https://arxiv.org/pdf/2204.08753v1", "arxiv_id": "2204.08753", "doi": "10.48550/arXiv.2204.08753", "citation_count": 20, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/THUMLP/SaMoE", "venue": "Findings", "quality_score": 0.3495}
{"id": "9bcf711e10fdafbeb24c8e0df8ae51ed860dbcc11a31a92b5c705e597e18ae9f", "sources": ["arxiv", "semantic_scholar"], "title": "StableMoE: Stable Routing Strategy for Mixture of Experts", "abstract": "The Mixture-of-Experts (MoE) technique can scale up the model size of Transformers with an affordable computational overhead. We point out that existing learning-to-route MoE methods suffer from the routing fluctuation issue, i.e., the target expert of the same input may change along with training, but only one expert will be activated for the input during inference. The routing fluctuation tends to harm sample efficiency because the same input updates different experts but only one is finally used. In this paper, we propose StableMoE with two training stages to address the routing fluctuation problem. In the first training stage, we learn a balanced and cohesive routing strategy and distill it into a lightweight router decoupled from the backbone model. In the second training stage, we utilize the distilled router to determine the token-to-expert assignment and freeze it for a stable routing strategy. We validate our method on language modeling and multilingual machine translation. The results show that StableMoE outperforms existing MoE methods in terms of both convergence speed and performance.", "authors": ["Damai Dai", "Li Dong", "Shuming Ma", "Bo Zheng", "Zhifang Sui", "Baobao Chang", "Furu Wei"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-18", "url": "https://arxiv.org/abs/2204.08396", "pdf_url": "https://arxiv.org/pdf/2204.08396v1", "arxiv_id": "2204.08396", "doi": "10.48550/arXiv.2204.08396", "citation_count": 121, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.5216}
{"id": "717325a1b7a7d890848626432b01bd84d564f2bdc273db4a9359329cb271ed39", "sources": ["arxiv", "semantic_scholar"], "title": "Sparsely Activated Mixture-of-Experts are Robust Multi-Task Learners", "abstract": "Traditional multi-task learning (MTL) methods use dense networks that use the same set of shared weights across several different tasks. This often creates interference where two or more tasks compete to pull model parameters in different directions. In this work, we study whether sparsely activated Mixture-of-Experts (MoE) improve multi-task learning by specializing some weights for learning shared representations and using the others for learning task-specific information. To this end, we devise task-aware gating functions to route examples from different tasks to specialized experts which share subsets of network weights conditioned on the task. This results in a sparsely activated multi-task model with a large number of parameters, but with the same computational cost as that of a dense model. We demonstrate such sparse networks to improve multi-task learning along three key dimensions: (i) transfer to low-resource tasks from related tasks in the training mixture; (ii) sample-efficient generalization to tasks not seen during training by making use of task-aware routing from seen related tasks; (iii) robustness to the addition of unrelated tasks by avoiding catastrophic forgetting of existing tasks.", "authors": ["Shashank Gupta", "Subhabrata Mukherjee", "Krishan Subudhi", "Eduardo Gonzalez", "Damien Jose", "Ahmed H. Awadallah", "Jianfeng Gao"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-16", "url": "https://arxiv.org/abs/2204.07689", "pdf_url": "https://arxiv.org/pdf/2204.07689v1", "arxiv_id": "2204.07689", "doi": "10.48550/arXiv.2204.07689", "citation_count": 64, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4532}
{"id": "c12259d573f08635e2e84281eb5503d30f9baf10c4b5989acd4c155dc6404731", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Experts for Biomedical Question Answering", "abstract": "Biomedical Question Answering (BQA) has attracted increasing attention in recent years due to its promising application prospect. It is a challenging task because the biomedical questions are professional and usually vary widely. Existing question answering methods answer all questions with a homogeneous model, leading to various types of questions competing for the shared parameters, which will confuse the model decision for each single type of questions. In this paper, in order to alleviate the parameter competition problem, we propose a Mixture-of-Expert (MoE) based question answering method called MoEBQA that decouples the computation for different types of questions by sparse routing. To be specific, we split a pretrained Transformer model into bottom and top blocks. The bottom blocks are shared by all the examples, aiming to capture the general features. The top blocks are extended to an MoE version that consists of a series of independent experts, where each example is assigned to a few experts according to its underlying question type. MoEBQA automatically learns the routing strategy in an end-to-end manner so that each expert tends to deal with the question types it is expert in. We evaluate MoEBQA on three BQA datasets constructed based on real examinations. The results show that our MoE extension significantly boosts the performance of question answering models and achieves new state-of-the-art performance. In addition, we elaborately analyze our MoE modules to reveal how MoEBQA works and find that it can automatically group the questions into human-readable clusters.", "authors": ["Damai Dai", "Wenbin Jiang", "Jiyuan Zhang", "Weihua Peng", "Yajuan Lyu", "Zhifang Sui", "Baobao Chang", "Yong Zhu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-15", "url": "https://arxiv.org/abs/2204.07469", "pdf_url": "https://arxiv.org/pdf/2204.07469v1", "arxiv_id": "2204.07469", "doi": "10.48550/arXiv.2204.07469", "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Natural Language Processing and Chinese Computing", "quality_score": 0.2865}
{"id": "65ee84d2f9e39c86b26f5a1ea3a85422620c18f6a3757e85695ba29d22e0d93f", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-experts VAEs can disregard variation in surjective multimodal data", "abstract": "Machine learning systems are often deployed in domains that entail data from multiple modalities, for example, phenotypic and genotypic characteristics describe patients in healthcare. Previous works have developed multimodal variational autoencoders (VAEs) that generate several modalities. We consider subjective data, where single datapoints from one modality (such as class labels) describe multiple datapoints from another modality (such as images). We theoretically and empirically demonstrate that multimodal VAEs with a mixture of experts posterior can struggle to capture variability in such surjective data.", "authors": ["Jannik Wolff", "Tassilo Klein", "Moin Nabi", "Rahul G. Krishnan", "Shinichi Nakajima"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-04-11", "url": "https://arxiv.org/abs/2204.05229", "pdf_url": "https://arxiv.org/pdf/2204.05229v1", "arxiv_id": "2204.05229", "doi": "10.48550/arXiv.2204.05229", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "ea5dd100853fecbe413c073e3dcb2044bf3892d6ecb43f4739f6fb49513e1433", "sources": ["arxiv", "semantic_scholar"], "title": "Channel Pruning In Quantization-aware Training: An Adaptive Projection-gradient Descent-shrinkage-splitting Method", "abstract": "We propose an adaptive projection-gradient descent-shrinkage-splitting method (APGDSSM) to integrate penalty based channel pruning into quantization-aware training (QAT). APGDSSM concurrently searches weights in both the quantized subspace and the sparse subspace. APGDSSM uses shrinkage operator and a splitting technique to create sparse weights, as well as the Group Lasso penalty to push the weight sparsity into channel sparsity. In addition, we propose a novel complementary transformed l1 penalty to stabilize the training for extreme compression.", "authors": ["Zhijian Li", "Jack Xin"], "categories": ["cs.LG", "cs.AI", "cs.CV", "math.NA"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-04-09", "url": "https://arxiv.org/abs/2204.04375", "pdf_url": "https://arxiv.org/pdf/2204.04375v1", "arxiv_id": "2204.04375", "doi": "10.1109/AI4I54798.2022.00015", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence for Industries", "quality_score": 0.0753}
{"id": "70d0a55d6f4468317e7ba48d87a04e9829871ea30f0f39609b43a2879f976ffa", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerating Attention through Gradient-Based Learned Runtime Pruning", "abstract": "Self-attention is a key enabler of state-of-art accuracy for various transformer-based Natural Language Processing models. This attention mechanism calculates a correlation score for each word with respect to the other words in a sentence. Commonly, only a small subset of words highly correlates with the word under attention, which is only determined at runtime. As such, a significant amount of computation is inconsequential due to low attention scores and can potentially be pruned. The main challenge is finding the threshold for the scores below which subsequent computation will be inconsequential. Although such a threshold is discrete, this paper formulates its search through a soft differentiable regularizer integrated into the loss function of the training. This formulation piggy backs on the back-propagation training to analytically co-optimize the threshold and the weights simultaneously, striking a formally optimal balance between accuracy and computation pruning. To best utilize this mathematical innovation, we devise a bit-serial architecture, dubbed LeOPArd, for transformer language models with bit-level early termination microarchitectural mechanism. We evaluate our design across 43 back-end tasks for MemN2N, BERT, ALBERT, GPT-2, and Vision transformer models. Post-layout results show that, on average, LeOPArd yields 1.9x and 3.9x speedup and energy reduction, respectively, while keeping the average accuracy virtually intact (<0.2% degradation)", "authors": ["Zheng Li", "Soroush Ghodrati", "Amir Yazdanbakhsh", "Hadi Esmaeilzadeh", "Mingu Kang"], "categories": ["cs.CL", "cs.AR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-07", "url": "https://arxiv.org/abs/2204.03227", "pdf_url": "https://arxiv.org/pdf/2204.03227v3", "arxiv_id": "2204.03227", "doi": "10.1145/3470496.3527423", "citation_count": 24, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Symposium on Computer Architecture", "quality_score": 0.3495}
{"id": "033b310f8f5937c97ab355d7fecba5988e790803cb661b6318eab939c96d0631", "sources": ["arxiv", "semantic_scholar"], "title": "Learning to Adapt Clinical Sequences with Residual Mixture of Experts", "abstract": "Clinical event sequences in Electronic Health Records (EHRs) record detailed information about the patient condition and patient care as they occur in time. Recent years have witnessed increased interest of machine learning community in developing machine learning models solving different types of problems defined upon information in EHRs. More recently, neural sequential models, such as RNN and LSTM, became popular and widely applied models for representing patient sequence data and for predicting future events or outcomes based on such data. However, a single neural sequential model may not properly represent complex dynamics of all patients and the differences in their behaviors. In this work, we aim to alleviate this limitation by refining a one-fits-all model using a Mixture-of-Experts (MoE) architecture. The architecture consists of multiple (expert) RNN models covering patient sub-populations and refining the predictions of the base model. That is, instead of training expert RNN models from scratch we define them on the residual signal that attempts to model the differences from the population-wide model. The heterogeneity of various patient sequences is modeled through multiple experts that consist of RNN. Particularly, instead of directly training MoE from scratch, we augment MoE based on the prediction signal from pretrained base GRU model. With this way, the mixture of experts can provide flexible adaptation to the (limited) predictive power of the single base RNN model. We experiment with the newly proposed model on real-world EHRs data and the multivariate clinical event prediction task. We implement RNN using Gated Recurrent Units (GRU). We show 4.1% gain on AUPRC statistics compared to a single RNN prediction.", "authors": ["Jeong Min Lee", "Milos Hauskrecht"], "categories": ["cs.LG", "cs.AI", "cs.CY"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-06", "url": "https://arxiv.org/abs/2204.02687", "pdf_url": "https://arxiv.org/pdf/2204.02687v1", "arxiv_id": "2204.02687", "doi": "10.1007/978-3-031-09342-5_15", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Artificial Intelligence in Medicine in Europe", "quality_score": 0.0}
{"id": "39111b3d02cc2d57e261090bf9ac56661327a26cc6506598c7b4ed0421a0a3a9", "sources": ["arxiv", "semantic_scholar"], "title": "Extended Uncertainty Principle via Dirac Quantization", "abstract": "Unifying quantum theory and gravity remains a fundamental challenge in physics. While most existing literature focuses on the ultraviolet (UV) modifications of quantum theory due to gravity, this work shows that generic infrared (IR) modifications arise when we describe quantum theory in curved spacetime. We explicitly demonstrate that the modifications to the position-momentum algebra are proportional to curvature invariants (such as the Ricci scalar and Kretschmann scalar). Our results, derived through a rigorous application of Dirac's quantization procedure, demonstrate that infrared effects in quantum systems can be axiomatically derived. We study particle dynamics in an arbitrary curved spacetime by embedding them in a higher-dimensional flat geometry. Our approach, which involves embedding particle dynamics in a higher-dimensional flat geometry and utilizing Dirac's quantization procedure, allows us to capture the dynamics of a particle in 4-dimensional curved spacetime through a modified position-momentum algebra. When applied to various spacetimes, this method reveals that the corrections due to the spacetime curvature are universal. We further compare our results with those derived using extended uncertainty principles. Finally, we discuss the implications of our work for black holes and entanglement.", "authors": ["Mytraya Gattu", "S. Shankaranarayanan"], "categories": ["gr-qc", "hep-th", "quant-ph"], "fields_of_study": ["Physics"], "published_date": "2022-04-04", "url": "https://arxiv.org/abs/2204.01780", "pdf_url": "https://arxiv.org/pdf/2204.01780v2", "arxiv_id": "2204.01780", "doi": "10.1007/s10714-024-03292-y", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "General Relativity and Gravitation", "quality_score": 0.0}
{"id": "a0898c4a33d8375f6654e6d7924008211d10197295e70848fa2d92f8f3fdbe30", "sources": ["arxiv", "semantic_scholar"], "title": "Build a Robust QA System with Transformer-based Mixture of Experts", "abstract": "In this paper, we aim to build a robust question answering system that can adapt to out-of-domain datasets. A single network may overfit to the superficial correlation in the training distribution, but with a meaningful number of expert sub-networks, a gating network that selects a sparse combination of experts for each input, and careful balance on the importance of expert sub-networks, the Mixture-of-Experts (MoE) model allows us to train a multi-task learner that can be generalized to out-of-domain datasets. We also explore the possibility of bringing the MoE layers up to the middle of the DistilBERT and replacing the dense feed-forward network with a sparsely-activated switch FFN layers, similar to the Switch Transformer architecture, which simplifies the MoE routing algorithm with reduced communication and computational costs. In addition to model architectures, we explore techniques of data augmentation including Easy Data Augmentation (EDA) and back translation, to create more meaningful variance among the small out-of-domain training data, therefore boosting the performance and robustness of our models. In this paper, we show that our combination of best architecture and data augmentation techniques achieves a 53.477 F1 score in the out-of-domain evaluation, which is a 9.52% performance gain over the baseline. On the final test set, we reported a higher 59.506 F1 and 41.651 EM. We successfully demonstrate the effectiveness of Mixture-of-Expert architecture in a Robust QA task.", "authors": ["Yu Qing Zhou", "Xixuan Julie Liu", "Yuanzhe Dong"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-20", "url": "https://arxiv.org/abs/2204.09598", "pdf_url": "https://arxiv.org/pdf/2204.09598v1", "arxiv_id": "2204.09598", "doi": "10.48550/arXiv.2204.09598", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "0fabf55cdfc9042ee641311375af5bee5fadeadea5ff82fb6a4bd0472472b141", "sources": ["arxiv", "semantic_scholar"], "title": "Diversifying Content Generation for Commonsense Reasoning with Mixture of Knowledge Graph Experts", "abstract": "Generative commonsense reasoning (GCR) in natural language is to reason about the commonsense while generating coherent text. Recent years have seen a surge of interest in improving the generation quality of commonsense reasoning tasks. Nevertheless, these approaches have seldom investigated diversity in the GCR tasks, which aims to generate alternative explanations for a real-world situation or predict all possible outcomes. Diversifying GCR is challenging as it expects to generate multiple outputs that are not only semantically different but also grounded in commonsense knowledge. In this paper, we propose MoKGE, a novel method that diversifies the generative reasoning by a mixture of expert (MoE) strategy on commonsense knowledge graphs (KG). A set of knowledge experts seek diverse reasoning on KG to encourage various generation outputs. Empirical experiments demonstrated that MoKGE can significantly improve the diversity while achieving on par performance on accuracy on two GCR benchmarks, based on both automatic and human evaluations.", "authors": ["Wenhao Yu", "Chenguang Zhu", "Lianhui Qin", "Zhihan Zhang", "Tong Zhao", "Meng Jiang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-14", "url": "https://arxiv.org/abs/2203.07285", "pdf_url": "https://arxiv.org/pdf/2203.07285v1", "arxiv_id": "2203.07285", "doi": "10.48550/arXiv.2203.07285", "citation_count": 37, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/DM2-ND/MoKGE", "venue": "Findings", "quality_score": 0.3949}
{"id": "3d7c530e864c143748578b0740545ead21d29c8dba977b7ec28a7eff35ff5f7b", "sources": ["arxiv", "semantic_scholar"], "title": "Simplicial Attention Neural Networks", "abstract": "The aim of this work is to introduce simplicial attention networks (SANs), i.e., novel neural architectures that operate on data defined on simplicial complexes leveraging masked self-attentional layers. Hinging on formal arguments from topological signal processing, we introduce a proper self-attention mechanism able to process data components at different layers (e.g., nodes, edges, triangles, and so on), while learning how to weight both upper and lower neighborhoods of the given topological domain in a totally task-oriented fashion. The proposed SANs generalize most of the current architectures available for processing data defined on simplicial complexes. The proposed approach compares favorably with other methods when applied to different (inductive and transductive) tasks such as trajectory prediction and missing data imputations in citation complexes.", "authors": ["L. Giusti", "C. Battiloro", "P. Di Lorenzo", "S. Sardellitti", "S. Barbarossa"], "categories": ["cs.LG", "cs.NE"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-14", "url": "https://arxiv.org/abs/2203.07485", "pdf_url": "https://arxiv.org/pdf/2203.07485v2", "arxiv_id": "2203.07485", "doi": "10.48550/arXiv.2203.07485", "citation_count": 37, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3949}
{"id": "52c9e2c947731dc5bdbd310bfce4cb872f433e734ec034c40d5fb6d4e0e4a24c", "sources": ["arxiv", "semantic_scholar"], "title": "Boosting Crowd Counting via Multifaceted Attention", "abstract": "This paper focuses on the challenging crowd counting task. As large-scale variations often exist within crowd images, neither fixed-size convolution kernel of CNN nor fixed-size attention of recent vision transformers can well handle this kind of variation. To address this problem, we propose a Multifaceted Attention Network (MAN) to improve transformer models in local spatial relation encoding. MAN incorporates global attention from a vanilla transformer, learnable local attention, and instance attention into a counting model. Firstly, the local Learnable Region Attention (LRA) is proposed to assign attention exclusively for each feature location dynamically. Secondly, we design the Local Attention Regularization to supervise the training of LRA by minimizing the deviation among the attention for different feature locations. Finally, we provide an Instance Attention mechanism to focus on the most important instances dynamically during training. Extensive experiments on four challenging crowd counting datasets namely ShanghaiTech, UCF-QNRF, JHU++, and NWPU have validated the proposed method. Codes: https://github.com/LoraLinH/Boosting-Crowd-Counting-via-Multifaceted-Attention.", "authors": ["Hui Lin", "Zhiheng Ma", "Rongrong Ji", "Yaowei Wang", "Xiaopeng Hong"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-05", "url": "https://arxiv.org/abs/2203.02636", "pdf_url": "https://arxiv.org/pdf/2203.02636v1", "arxiv_id": "2203.02636", "doi": "10.1109/CVPR52688.2022.01901", "citation_count": 226, "influential_citation_count": 23, "has_code": true, "code_url": "https://github.com/LoraLinH/Boosting-Crowd-Counting-via-Multifaceted-Attention", "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.6901}
{"id": "9220a41e7f38d5c19dc702fc411ba2a0bd380d06b016775e863e006320c4619c", "sources": ["arxiv", "semantic_scholar"], "title": "Parameter-Efficient Mixture-of-Experts Architecture for Pre-trained Language Models", "abstract": "Recently, Mixture-of-Experts (short as MoE) architecture has achieved remarkable success in increasing the model capacity of large-scale language models. However, MoE requires incorporating significantly more parameters than the base model being extended. In this paper, we propose building a parameter-efficient MoE architecture by sharing information among experts. We adopt the matrix product operator (MPO, a tensor decomposition from quantum many-body physics) to reconstruct the parameter matrix in the expert layer and increase model capacity for pre-trained language models by sharing parameters of the central tensor (containing the core information) among different experts while enabling the specificity through the auxiliary tensors (complementing the central tensor) of different experts. To address the unbalanced optimization issue, we further design the gradient mask strategy for the MPO-based MoE architecture. Extensive experiments based on T5 and GPT-2 show improved performance and efficiency of the pre-trained language model (27.2x reduction in total parameters for the superior model performance, compared with the Switch Transformers). Our code is publicly available at https://github.com/RUCAIBox/MPOE.", "authors": ["Ze-Feng Gao", "Peiyu Liu", "Wayne Xin Zhao", "Zhong-Yi Lu", "Ji-Rong Wen"], "categories": ["cs.CL", "cs.AI", "cs.LG", "quant-ph"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2022-03-02", "url": "https://arxiv.org/abs/2203.01104", "pdf_url": "https://arxiv.org/pdf/2203.01104v4", "arxiv_id": "2203.01104", "doi": "10.48550/arXiv.2203.01104", "citation_count": 36, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/RUCAIBox/MPOE", "venue": "International Conference on Computational Linguistics", "quality_score": 0.3921}
{"id": "a429bc000298419f3c71a89d72149ca52e3c20573d1da9b1c0adb93d647ecfb5", "sources": ["arxiv", "semantic_scholar"], "title": "Functional mixture-of-experts for classification", "abstract": "We develop a mixtures-of-experts (ME) approach to the multiclass classification where the predictors are univariate functions. It consists of a ME model in which both the gating network and the experts network are constructed upon multinomial logistic activation functions with functional inputs. We perform a regularized maximum likelihood estimation in which the coefficient functions enjoy interpretable sparsity constraints on targeted derivatives. We develop an EM-Lasso like algorithm to compute the regularized MLE and evaluate the proposed approach on simulated and real data.", "authors": ["Nhat Thien Pham", "Faicel Chamroukhi"], "categories": ["stat.ML", "cs.AI", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2022-02-28", "url": "https://arxiv.org/abs/2202.13934", "pdf_url": "https://arxiv.org/pdf/2202.13934v1", "arxiv_id": "2202.13934", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "4e0395c60fbfbb7fc05bc3fb69ad22b4572460a0b29858223e0268f5ba828781", "sources": ["arxiv", "semantic_scholar"], "title": "High Dimensional Statistical Estimation under Uniformly Dithered One-bit Quantization", "abstract": "In this paper, we propose a uniformly dithered 1-bit quantization scheme for high-dimensional statistical estimation. The scheme contains truncation, dithering, and quantization as typical steps. As canonical examples, the quantization scheme is applied to the estimation problems of sparse covariance matrix estimation, sparse linear regression (i.e., compressed sensing), and matrix completion. We study both sub-Gaussian and heavy-tailed regimes, where the underlying distribution of heavy-tailed data is assumed to have bounded moments of some order. We propose new estimators based on 1-bit quantized data. In sub-Gaussian regime, our estimators achieve near minimax rates, indicating that our quantization scheme costs very little. In heavy-tailed regime, while the rates of our estimators become essentially slower, these results are either the first ones in an 1-bit quantized and heavy-tailed setting, or already improve on existing comparable results from some respect. Under the observations in our setting, the rates are almost tight in compressed sensing and matrix completion. Our 1-bit compressed sensing results feature general sensing vector that is sub-Gaussian or even heavy-tailed. We also first investigate a novel setting where both the covariate and response are quantized. In addition, our approach to 1-bit matrix completion does not rely on likelihood and represent the first method robust to pre-quantization noise with unknown distribution. Experimental results on synthetic data are presented to support our theoretical analysis.", "authors": ["Junren Chen", "Cheng-Long Wang", "Michael K. Ng", "Di Wang"], "categories": ["stat.ML", "cs.LG", "eess.SP"], "fields_of_study": ["Computer Science", "Mathematics", "Engineering"], "published_date": "2022-02-26", "url": "https://arxiv.org/abs/2202.13157", "pdf_url": "https://arxiv.org/pdf/2202.13157v4", "arxiv_id": "2202.13157", "doi": "10.1109/TIT.2023.3266271", "citation_count": 28, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Information Theory", "quality_score": 0.3656}
{"id": "ee42804008fc8ce6bbdba257a1e109877b01a43b9f27b049878c740a8a2421ae", "sources": ["arxiv", "semantic_scholar"], "title": "Static Analysis Deployment Pitfalls", "abstract": "Organizational, political, and configuration mistakes in the deployment of a static source code analysis tool within a software development organization can result in most of the value of the tool being lost, even while apparently meeting management goals. A list of pitfalls encountered as a static analysis consultant is presented, with discussion of techniques for avoiding or mitigating them. This is part of a work in progress, tentatively entitled \"Handbook of Static Analysis Deployment.\"", "authors": ["Flash Sheridan"], "categories": ["cs.SE", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-26", "url": "https://arxiv.org/abs/2202.13026", "pdf_url": "https://arxiv.org/pdf/2202.13026v1", "arxiv_id": "2202.13026", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "c33bfa34eb3a42efdddcdcee04ccdc1c58dd7b9f095e4b90cbde9f500c5c308e", "sources": ["arxiv", "semantic_scholar"], "title": "Support Recovery in Mixture Models with Sparse Parameters", "abstract": "Mixture models are widely used to fit complex and multimodal datasets. In this paper we study mixtures with high dimensional sparse latent parameter vectors and consider the problem of support recovery of those vectors. While parameter learning in mixture models is well-studied, the sparsity constraint remains relatively unexplored. Sparsity of parameter vectors is a natural constraint in variety of settings, and support recovery is a major step towards parameter estimation. We provide efficient algorithms for support recovery that have a logarithmic sample complexity dependence on the dimensionality of the latent space. Our algorithms are quite general, namely they are applicable to 1) mixtures of many different canonical distributions including Uniform, Poisson, Laplace, Gaussians, etc. 2) Mixtures of linear regressions and linear classifiers with Gaussian covariates under different assumptions on the unknown parameters. In most of these settings, our results are the first guarantees on the problem while in the rest, our results provide improvements on existing works.", "authors": ["Arya Mazumdar", "Soumyabrata Pal"], "categories": ["cs.LG", "cs.IT", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-02-24", "url": "https://arxiv.org/abs/2202.11940", "pdf_url": "https://arxiv.org/pdf/2202.11940v2", "arxiv_id": "2202.11940", "doi": "10.1109/TIT.2024.3462937", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Information Theory", "quality_score": 0.0753}
{"id": "ab967767018884a1857f1c23fc22e635e7ab4e041b8f74ca07dcfa6c0479e77b", "sources": ["arxiv", "semantic_scholar"], "title": "Visual Attention Network", "abstract": "While originally designed for natural language processing tasks, the self-attention mechanism has recently taken various computer vision areas by storm. However, the 2D nature of images brings three challenges for applying self-attention in computer vision. (1) Treating images as 1D sequences neglects their 2D structures. (2) The quadratic complexity is too expensive for high-resolution images. (3) It only captures spatial adaptability but ignores channel adaptability. In this paper, we propose a novel linear attention named large kernel attention (LKA) to enable self-adaptive and long-range correlations in self-attention while avoiding its shortcomings. Furthermore, we present a neural network based on LKA, namely Visual Attention Network (VAN). While extremely simple, VAN surpasses similar size vision transformers(ViTs) and convolutional neural networks(CNNs) in various tasks, including image classification, object detection, semantic segmentation, panoptic segmentation, pose estimation, etc. For example, VAN-B6 achieves 87.8% accuracy on ImageNet benchmark and set new state-of-the-art performance (58.2 PQ) for panoptic segmentation. Besides, VAN-B2 surpasses Swin-T 4% mIoU (50.1 vs. 46.1) for semantic segmentation on ADE20K benchmark, 2.6% AP (48.8 vs. 46.2) for object detection on COCO dataset. It provides a novel method and a simple yet strong baseline for the community. Code is available at https://github.com/Visual-Attention-Network.", "authors": ["Meng-Hao Guo", "Cheng-Ze Lu", "Zheng-Ning Liu", "Ming-Ming Cheng", "Shi-Min Hu"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-20", "url": "https://arxiv.org/abs/2202.09741", "pdf_url": "https://arxiv.org/pdf/2202.09741v5", "arxiv_id": "2202.09741", "doi": "10.1007/s41095-023-0364-2", "citation_count": 1019, "influential_citation_count": 65, "has_code": true, "code_url": "https://github.com/Visual-Attention-Network", "venue": "Computational Visual Media", "quality_score": 0.9098}
{"id": "6c4668e251cb0da1c5faf5d4eb06f2762f6a78d302ccef6cf9331f1112cfd821", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture-of-Experts with Expert Choice Routing", "abstract": "Sparsely-activated Mixture-of-experts (MoE) models allow the number of parameters to greatly increase while keeping the amount of computation for a given token or a given sample unchanged. However, a poor expert routing strategy (e.g. one resulting in load imbalance) can cause certain experts to be under-trained, leading to an expert being under or over-specialized. Prior work allocates a fixed number of experts to each token using a top-k function regardless of the relative importance of different tokens. To address this, we propose a heterogeneous mixture-of-experts employing an expert choice method. Instead of letting tokens select the top-k experts, we have experts selecting the top-k tokens. As a result, each token can be routed to a variable number of experts and each expert can have a fixed bucket size. We systematically study pre-training speedups using the same computational resources of the Switch Transformer top-1 and GShard top-2 gating of prior work and find that our method improves training convergence time by more than 2x. For the same computational cost, our method demonstrates higher performance in fine-tuning 11 selected tasks in the GLUE and SuperGLUE benchmarks. For a smaller activation cost, our method outperforms the T5 dense model in 7 out of the 11 tasks.", "authors": ["Yanqi Zhou", "Tao Lei", "Hanxiao Liu", "Nan Du", "Yanping Huang", "Vincent Zhao", "Andrew Dai", "Zhifeng Chen", "Quoc Le", "James Laudon"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-18", "url": "https://arxiv.org/abs/2202.09368", "pdf_url": "https://arxiv.org/pdf/2202.09368v2", "arxiv_id": "2202.09368", "doi": "10.52202/068431-0515", "citation_count": 766, "influential_citation_count": 43, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.8217}
{"id": "615ccc166efd02b35287649b4466ffc6d5f43b75278f591e73ec34c4b7cddfe7", "sources": ["arxiv", "semantic_scholar"], "title": "Mixture of Online and Offline Experts for Non-stationary Time Series", "abstract": "We consider a general and realistic scenario involving non-stationary time series, consisting of several offline intervals with different distributions within a fixed offline time horizon, and an online interval that continuously receives new samples. For non-stationary time series, the data distribution in the current online interval may have appeared in previous offline intervals. We theoretically explore the feasibility of applying knowledge from offline intervals to the current online interval. To this end, we propose the Mixture of Online and Offline Experts (MOOE). MOOE learns static offline experts from offline intervals and maintains a dynamic online expert for the current online interval. It then adaptively combines the offline and online experts using a meta expert to make predictions for the samples received in the online interval. Specifically, we focus on theoretical analysis, deriving parameter convergence, regret bounds, and generalization error bounds to prove the effectiveness of the algorithm.", "authors": ["Zhilin Zhao", "Longbing Cao", "Yuanyu Wan"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-12", "url": "https://arxiv.org/abs/2202.05996", "pdf_url": "https://arxiv.org/pdf/2202.05996v3", "arxiv_id": "2202.05996", "doi": "10.1609/aaai.v39i21.34448", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.0753}
{"id": "613aac8d124f08e23ea5f16711c1f6534bf4e2c8357861ee96c8b964e774f046", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Lightweight Neural Animation : Exploration of Neural Network Pruning in Mixture of Experts-based Animation Models", "abstract": "In the past few years, neural character animation has emerged and offered an automatic method for animating virtual characters. Their motion is synthesized by a neural network. Controlling this movement in real time with a user-defined control signal is also an important task in video games for example. Solutions based on fully-connected layers (MLPs) and Mixture-of-Experts (MoE) have given impressive results in generating and controlling various movements with close-range interactions between the environment and the virtual character. However, a major shortcoming of fully-connected layers is their computational and memory cost which may lead to sub-optimized solution. In this work, we apply pruning algorithms to compress an MLP- MoE neural network in the context of interactive character animation, which reduces its number of parameters and accelerates its computation time with a trade-off between this acceleration and the synthesized motion quality. This work demonstrates that, with the same number of experts and parameters, the pruned model produces less motion artifacts than the dense model and the learned high-level motion features are similar for both", "authors": ["Antoine Maiorca", "Nathan Hubens", "Sohaib Laraba", "Thierry Dutoit"], "categories": ["cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2022-01-11", "url": "https://arxiv.org/abs/2201.04042", "pdf_url": "https://arxiv.org/pdf/2201.04042v2", "arxiv_id": "2201.04042", "doi": "10.5220/0010908700003124", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "VISIGRAPP", "quality_score": 0.1505}