Daily Papers

Low-rank adaptation (LoRA) enables parameter-efficient specialization of foundation models, but the proliferation of task-specific adapters fragments capabilities across many adapters, complicating reuse and deployment. We study the problem of merging T LoRAs into a single rank-r LoRA, thereby preserving the benefits of low-rank structure. Existing Merge-then-Compress pipelines treat the rank constraint as an afterthought: they merge adapters in the full parameter space, then compress the merged result to rank r via truncated SVD. However, full-parameter merging may destroy the low-rank structure, making it difficult for subsequent compression to recover an effective rank-r LoRA. We propose Compress-then-Merge (CtM), a reversed pipeline that enforces the rank-r bottleneck before merging: CtM computes shared r-dimensional subspaces using only the LoRA weights to capture cross-adapter common structure, projects each adapter into the shared subspaces to obtain rtimes r coordinates, and then applies standard merging rules in this reduced space. CtM guarantees a rank-r LoRA by construction, avoiding post-hoc truncation, and enables efficient computation in the core space spanned by concatenated LoRA factors. Experiments across multiple models and tasks show that CtM consistently outperforms existing single-LoRA-output baselines while narrowing the performance gap to full-parameter merging methods.

Sparsely activated Mixture-of-Experts (SMoE) has shown promise to scale up the learning capacity of neural networks, however, they have issues like (a) High Memory Usage, due to duplication of the network layers into multiple copies as experts; and (b) Redundancy in Experts, as common learning-based routing policies suffer from representational collapse. Therefore, vanilla SMoE models are memory inefficient and non-scalable, especially for resource-constrained downstream scenarios. In this paper, we ask: Can we craft a compact SMoE model by consolidating expert information? What is the best recipe to merge multiple experts into fewer but more knowledgeable experts? Our pilot investigation reveals that conventional model merging methods fail to be effective in such expert merging for SMoE. The potential reasons are: (1) redundant information overshadows critical experts; (2) appropriate neuron permutation for each expert is missing to bring all of them in alignment. To address this, we propose M-SMoE, which leverages routing statistics to guide expert merging. Specifically, it starts with neuron permutation alignment for experts; then, dominant experts and their "group members" are formed; lastly, every expert group is merged into a single expert by utilizing each expert's activation frequency as their weight for merging, thus diminishing the impact of insignificant experts. Moreover, we observed that our proposed merging promotes a low dimensionality in the merged expert's weight space, naturally paving the way for additional compression. Hence, our final method, MC-SMoE (i.e., Merge, then Compress SMoE), further decomposes the merged experts into low-rank and structural sparse alternatives. Extensive experiments across 8 benchmarks validate the effectiveness of MC-SMoE. For instance, our MC-SMoE achieves up to 80% memory and a 20% FLOPs reduction, with virtually no loss in performance.

Long-context large language models (LC LLMs) combined with retrieval-augmented generation (RAG) hold strong potential for complex multi-hop and large-document tasks. However, existing RAG systems often suffer from imprecise retrieval, incomplete context coverage under constrained windows, and fragmented information from suboptimal context construction. We introduce Multi-scale Adaptive Context RAG (MacRAG), a hierarchical RAG framework that compresses and partitions documents into coarse-to-fine granularities, then adaptively merges relevant contexts through real-time chunk- and document-level expansions. By initiating with finest-level retrieval and progressively incorporating broader, higher-level context, MacRAG constructs effective query-specific long contexts, optimizing both precision and coverage. Evaluations on challenging LongBench expansions of HotpotQA, 2WikiMultihopQA, and Musique confirm MacRAG consistently surpasses baseline RAG pipelines in single- and multi-step generation using Llama-3.1-8B, Gemini-1.5-pro, and GPT-4o. Our results establish MacRAG as an efficient, scalable solution for real-world long-context, multi-hop reasoning. Our code is available at https://github.com/Leezekun/MacRAG.

Summarizing book-length documents (>100K tokens) that exceed the context window size of large language models (LLMs) requires first breaking the input document into smaller chunks and then prompting an LLM to merge, update, and compress chunk-level summaries. Despite the complexity and importance of this task, it has yet to be meaningfully studied due to the challenges of evaluation: existing book-length summarization datasets (e.g., BookSum) are in the pretraining data of most public LLMs, and existing evaluation methods struggle to capture errors made by modern LLM summarizers. In this paper, we present the first study of the coherence of LLM-based book-length summarizers implemented via two prompting workflows: (1) hierarchically merging chunk-level summaries, and (2) incrementally updating a running summary. We obtain 1193 fine-grained human annotations on GPT-4 generated summaries of 100 recently-published books and identify eight common types of coherence errors made by LLMs. Because human evaluation is expensive and time-consuming, we develop an automatic metric, BooookScore, that measures the proportion of sentences in a summary that do not contain any of the identified error types. BooookScore has high agreement with human annotations and allows us to systematically evaluate the impact of many other critical parameters (e.g., chunk size, base LLM) while saving $15K USD and 500 hours in human evaluation costs. We find that closed-source LLMs such as GPT-4 and Claude 2 produce summaries with higher BooookScore than those generated by open-source models. While LLaMA 2 falls behind other models, Mixtral achieves performance on par with GPT-3.5-Turbo. Incremental updating yields lower BooookScore but higher level of detail than hierarchical merging, a trade-off sometimes preferred by annotators.

Fine-tuning large language models (LLMs) with low-rank adapters (LoRAs) has become common practice, often yielding numerous copies of the same LLM differing only in their LoRA updates. This paradigm presents challenges for systems that serve real-time responses to queries that each involve a different LoRA. Prior works optimize the design of such systems but still require continuous loading and offloading of LoRAs, as it is infeasible to store thousands of LoRAs in GPU memory. To mitigate this issue, we investigate the efficacy of compression when serving LoRA adapters. We consider compressing adapters individually via SVD and propose a method for joint compression of LoRAs into a shared basis paired with LoRA-specific scaling matrices. Our experiments with up to 500 LoRAs demonstrate that compressed LoRAs preserve performance while offering major throughput gains in realistic serving scenarios with over a thousand LoRAs, maintaining 75% of the throughput of serving a single LoRA.

Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy is to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.

Reasoning models produce long chain-of-thought traces that are costly to distill and encourage verbose student outputs. We study post-hoc compression of such traces before knowledge distillation. Two teachers, Qwen3.5-397B-A17B and gpt-oss-120B, generate about 283k correct traces each; two instruction-tuned models then compress them to 8.6-21.0% of their original character length. Across a 48-run main grid plus seven Qwen-teacher truncation ablations, compressed traces reduce training tokens to 12-30% of raw, speed up training by 2.0-7.6x, and shorten inference outputs by 3-19x with smaller reductions under the shorter gpt-oss teacher. However, raw traces retain the highest downstream accuracy at every scale and for both teachers. A length-matched raw-trace truncation ablation shows that compression is not merely benefiting from a smaller token budget: model-compressed traces usually beat or match naive truncation, especially for smaller students, while maintaining shorter inference outputs. Overall, reasoning-trace compression offers an accuracy-efficiency trade-off rather than a free improvement: students retain up to 96% of raw-trace accuracy while gaining up to 18x higher per-token efficiency, and at the 0.8B scale under LoRA compressed traces narrow the raw-vs-compressed gap but do not exceed raw.

Autonomous driving demands high-quality LiDAR data, yet the cost of physical LiDAR sensors presents a significant scaling-up challenge. While recent efforts have explored deep generative models to address this issue, they often consume substantial computational resources with slow generation speeds while suffering from a lack of realism. To address these limitations, we introduce RangeLDM, a novel approach for rapidly generating high-quality range-view LiDAR point clouds via latent diffusion models. We achieve this by correcting range-view data distribution for accurate projection from point clouds to range images via Hough voting, which has a critical impact on generative learning. We then compress the range images into a latent space with a variational autoencoder, and leverage a diffusion model to enhance expressivity. Additionally, we instruct the model to preserve 3D structural fidelity by devising a range-guided discriminator. Experimental results on KITTI-360 and nuScenes datasets demonstrate both the robust expressiveness and fast speed of our LiDAR point cloud generation.

Visual language models (VLMs) have made significant advances in accuracy in recent years. However, their efficiency has received much less attention. This paper introduces NVILA, a family of open VLMs designed to optimize both efficiency and accuracy. Building on top of VILA, we improve its model architecture by first scaling up the spatial and temporal resolutions, and then compressing visual tokens. This "scale-then-compress" approach enables NVILA to efficiently process high-resolution images and long videos. We also conduct a systematic investigation to enhance the efficiency of NVILA throughout its entire lifecycle, from training and fine-tuning to deployment. NVILA matches or surpasses the accuracy of many leading open and proprietary VLMs across a wide range of image and video benchmarks. At the same time, it reduces training costs by 4.5X, fine-tuning memory usage by 3.4X, pre-filling latency by 1.6-2.2X, and decoding latency by 1.2-2.8X. We will soon make our code and models available to facilitate reproducibility.

Extractive prompt compression promises to cut LLM inference costs by removing low-information tokens, and learned compressors such as LLMLingua-2 report strong results on English benchmarks. Most other languages already pay a token premium: the same content costs 1.3-1.8x more tokens than in English. We ask whether compression closes or widens this gap. Using fully parallel data in ten languages spanning five scripts, with controls budget-matched in the target model's tokenizer, we audit four learned compressors against four deterministic baselines, on eleven target models from ten vendors (over 250,000 evaluation calls). Three of the compressors are trained with English supervision (LLMLingua-2 XLM-R/mBERT; Kompress-v2 from the production Headroom stack); the fourth, XProvence, is trained multilingually. First, the transfer gap is real, replicates across target models and compressor backbones, and is strongly rate-dependent: at a 0.33 keep-rate English retains 57-62% of normalized context utilization while Lithuanian retains 10-24% and Chinese essentially none, despite Chinese having the smallest token premium. Second, the gap tracks compression supervision data, not architecture. All three English-trained compressors show it, deterministic methods show no comparable gap, and the multilingually trained XProvence v1 shows none. Its v2 release, retrained on translated data, empties 92% of Chinese contexts at its aggressive threshold without any warning. Third, in a harder long-context setting, aggressive learned compression drives compressed contexts to or below no-context utility in three of five non-English languages. A translate-then-compress pipeline matches or beats native compression at roughly half the token cost in three of five tested languages. We release all code, compressions, and model outputs. Safe compression budgets are much smaller outside English.

The rapid progress of large language models (LLMs) has laid the foundation for multimodal models. However, visual language models (VLMs) still face heavy computational costs when extended from images to videos due to high frame rates and long durations. Token compression is a promising solution, yet most existing training-free methods cause information loss and performance degradation. To overcome this, we propose Memory-Augmented Reinforcement Learning-based Token Compression (MARC), which integrates structured retrieval and RL-based distillation. MARC adopts a retrieve-then-compress strategy using a Visual Memory Retriever (VMR) to select key clips and a Compression Group Relative Policy Optimization (C-GRPO) framework to distil reasoning ability from a teacher to a student model. Experiments on six video benchmarks show that MARC achieves near-baseline accuracy using only one frame's tokens -- reducing visual tokens by 95\%, GPU memory by 72\%, and latency by 23.9\%. This demonstrates its potential for efficient, real-time video understanding in resource-constrained settings such as video QA, surveillance, and autonomous driving.

Knowledge Distillation (KD) is a predominant approach for BERT compression. Previous KD-based methods focus on designing extra alignment losses for the student model to mimic the behavior of the teacher model. These methods transfer the knowledge in an indirect way. In this paper, we propose a novel Weight-Inherited Distillation (WID), which directly transfers knowledge from the teacher. WID does not require any additional alignment loss and trains a compact student by inheriting the weights, showing a new perspective of knowledge distillation. Specifically, we design the row compactors and column compactors as mappings and then compress the weights via structural re-parameterization. Experimental results on the GLUE and SQuAD benchmarks show that WID outperforms previous state-of-the-art KD-based baselines. Further analysis indicates that WID can also learn the attention patterns from the teacher model without any alignment loss on attention distributions. The code is available at https://github.com/wutaiqiang/WID-NAACL2024.

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.

Reinforcement learning with verifiable rewards (RLVR) broadcasts a single response-level reward to every token, while on-policy distillation (OPD) scores each token against a stronger teacher for a dense advantage but caps performance at teacher quality and discourages exploration beyond it. Their complementarity makes combining RLVR and OPD promising, but we find that fusing the two advantages with a fixed coefficient triggers entropy collapse from two miscalibrations: a magnitude mismatch, where token-level OPD advantages can spike far beyond the bounded RLVR advantage and erase its signal, and a temporal mismatch, where sustained full-strength OPD keeps pulling the student toward the teacher and limits exploration needed to surpass it. We propose SAF, a Stable Advantage Fusion framework that resolves both issues via a lightweight, four-stage pipeline applied only to the OPD advantage: a sparsify-then-compress mechanism for magnitude control paired with a warm-up-then-anneal mechanism for temporal control, with each stage independently switchable and adding negligible overhead. Instantiating RLVR with GRPO, we evaluate SAF across seven mathematical reasoning and code generation benchmarks with Qwen3-1.7B/4B/8B: SAF avoids entropy collapse and consistently outperforms fixed-coefficient GRPO+OPD fusion, improving the aggregate score by 0.51-2.70% across all six model-domain settings while achieving more stable training.

Federated Learning is a machine learning setting where the goal is to train a high-quality centralized model while training data remains distributed over a large number of clients each with unreliable and relatively slow network connections. We consider learning algorithms for this setting where on each round, each client independently computes an update to the current model based on its local data, and communicates this update to a central server, where the client-side updates are aggregated to compute a new global model. The typical clients in this setting are mobile phones, and communication efficiency is of the utmost importance. In this paper, we propose two ways to reduce the uplink communication costs: structured updates, where we directly learn an update from a restricted space parametrized using a smaller number of variables, e.g. either low-rank or a random mask; and sketched updates, where we learn a full model update and then compress it using a combination of quantization, random rotations, and subsampling before sending it to the server. Experiments on both convolutional and recurrent networks show that the proposed methods can reduce the communication cost by two orders of magnitude.