Daily Papers

The memory-bound nature of the decoding stage of large language model (LLM) inference incurs significant latency. Layer-wise training-free network pruning approaches guided by the Hessian have been a prominent solution to this problem, as pruning reduces the number of nonzero parameters read from memory during decoding. Nevertheless, typical methods in this line compute the Hessian using pre-collected natural sequences, whereas the model is fed self-generated tokens during decoding, creating a distribution shift between the two sequences. The Hessian calculated on the natural sequence is different from that calculated on the generated sequence. We observe that this discrepancy causes the activation distribution during generation to deviate from that used for pruning, further hurting the pruned model performance. Moreover, most existing LLM pruning methods that bring actual speedup primarily target the sparse matrix-matrix (SpMM) multiplication, providing limited support for the sparse matrix-vector (SpMV) operations, which dominate decoding. To solve these problems, we introduce SparseDecoding, a principled decoding-aware pruning framework tailored for accurate and efficient LLM decoding. Specifically, at the algorithmic axis, SparseDecoding constructs calibration matrices from layer-wise activations collected during the dense-model autoregressive generation, excluding prefill, thereby aligning the pruning objective with the decoding activations. At the system axis, we develop an optimized N:M sparse matrix-vector kernel with bitmask indexing and fixed-step traversal. Substantial empirical results on representative LLMs (Llama-3.1-8B, Llama-3.3-70B, Qwen3-14B / 32B) demonstrate that our method consistently outperforms standard fixed-text calibration on the long-form generation benchmarks while achieving up to 1.48x end-to-end wall-clock decoding speedup on A100 GPUs.

Low-latency decoding for large language models (LLMs) is crucial for applications like chatbots and code assistants, yet generating long outputs remains slow in single-query settings. Prior work on speculative decoding (which combines a small draft model with a larger target model) and tensor parallelism has each accelerated decoding. However, conventional approaches fail to apply both simultaneously due to imbalanced compute requirements (between draft and target models), KV-cache inconsistencies, and communication overheads under small-batch tensor-parallelism. This paper introduces SwiftSpec, a system that targets ultra-low latency for LLM decoding. SwiftSpec redesigns the speculative decoding pipeline in an asynchronous and disaggregated manner, so that each component can be scaled flexibly and remove draft overhead from the critical path. To realize this design, SwiftSpec proposes parallel tree generation, tree-aware KV cache management, and fused, latency-optimized kernels to overcome the challenges listed above. Across 5 model families and 6 datasets, SwiftSpec achieves an average of 1.75x speedup over state-of-the-art speculative decoding systems and, as a highlight, serves Llama3-70B at 348 tokens/s on 8 Nvidia Hopper GPUs, making it the fastest known system for low-latency LLM serving at this scale.

Autoregressive decoding of large language models (LLMs) is memory bandwidth bounded, resulting in high latency and significant wastes of the parallel processing power of modern accelerators. Existing methods for accelerating LLM decoding often require a draft model (e.g., speculative decoding), which is nontrivial to obtain and unable to generalize. In this paper, we introduce Lookahead decoding, an exact, parallel decoding algorithm that accelerates LLM decoding without needing auxiliary models or data stores. It allows trading per-step log(FLOPs) to reduce the number of total decoding steps, is more parallelizable on single or multiple modern accelerators, and is compatible with concurrent memory-efficient attention (e.g., FlashAttention). Our implementation of Lookahead decoding can speed up autoregressive decoding by up to 1.8x on MT-bench and 4x with strong scaling on multiple GPUs in code completion tasks. Our code is avialable at https://github.com/hao-ai-lab/LookaheadDecoding

Large language models (LLMs) have driven significant advancements across diverse NLP tasks, with long-context models gaining prominence for handling extended inputs. However, the expanding key-value (KV) cache size required by Transformer architectures intensifies the memory constraints, particularly during the decoding phase, creating a significant bottleneck. Existing sparse attention mechanisms designed to address this bottleneck have two limitations: (1) they often fail to reliably identify the most relevant tokens for attention, and (2) they overlook the spatial coherence of token selection across consecutive Transformer layers, which can lead to performance degradation and substantial overhead in token selection. This paper introduces TidalDecode, a simple yet effective algorithm and system for fast and accurate LLM decoding through position persistent sparse attention. TidalDecode leverages the spatial coherence of tokens selected by existing sparse attention methods and introduces a few token selection layers that perform full attention to identify the tokens with the highest attention scores, while all other layers perform sparse attention with the pre-selected tokens. This design enables TidalDecode to substantially reduce the overhead of token selection for sparse attention without sacrificing the quality of the generated results. Evaluation on a diverse set of LLMs and tasks shows that TidalDecode closely matches the generative performance of full attention methods while reducing the LLM decoding latency by up to 2.1x.

The proliferation of long-context large language models (LLMs) exposes a key bottleneck: the rapidly expanding key-value cache during decoding, which imposes heavy memory and latency costs. While recent approaches attempt to alleviate this by sharing a single set of crucial tokens across layers, such coarse-grained sharing undermines model performance by neglecting the functional diversity of attention heads. To address this, we propose LycheeDecode, an efficient decoding method centered on a fine-grained hybrid-head attention mechanism that employs a hardware-efficient top-k selection strategy. Specifically, the novel HardKuma-based mechanism partitions attention heads into a small subset of retrieval heads that dynamically identify crucial tokens and a majority of sparse heads that reuse them for efficient computation. Through extensive experiments on leading models like Llama3 and Qwen3 across diverse benchmarks for long-context understanding (e.g., LongBench, RULER) and complex reasoning (e.g., AIME24, OlympiadBench), we demonstrate that LycheeDecode achieves generative quality comparable to, and at times surpassing even the full-attention baseline. Crucially, this is accomplished with up to a 2.7x speedup at a 128K context length. By preserving the functional diversity of attention heads, our fine-grained strategy overcomes the performance bottlenecks of existing methods, providing a powerful and validated pathway to both efficient and high-quality long-context LLM inference.

We introduce a principled probabilistic framework for reward-guided decoding in large language models, addressing the limitations of standard decoding methods that optimize token-level likelihood rather than sequence-level quality. Our method defines a reward-augmented target distribution over complete sequences by combining model transition probabilities with prefix-dependent reward potentials. Importantly, the approach is training-free: it leaves model weights unchanged and instead modifies the inference distribution via reward potentials, with all gains arising purely from inference-time sampling. To sample from this distribution, we develop Sequential Monte Carlo algorithms, including a computationally efficient prefix-only variant and a lookahead variant whose intermediate targets match the exact marginals of the full sequence distribution. The framework also integrates resample-move updates with Metropolis-Hastings rejuvenation and supports block-wise generation, subsuming common decoding strategies such as temperature sampling and power-tempered objectives. Empirical results across three 7B models show significant gains. On code generation (HumanEval), our method improves base performance by up to 54.9% and surpasses the strongest sampling baselines by 9.1%-15.3%. On mathematical reasoning (MATH500), it achieves gains of up to 8.8%. Notably, it reaches 87.8% on HumanEval and 78.4% on MATH500 with Qwen2.5-7B, consistently outperforming the reinforcement learning method GRPO.

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

At each step, decoding one sequence with a large language model rereads the projection weights, whose traffic is fixed, and the key-value (KV) cache, whose traffic grows with context. Activation sparsity trims the first term and KV-cache sparsity the second, yet their reported speedups are hard to compare because each depends on context length and on the dense attention kernel it is measured against. We derive a byte crossover, the context length at which the two savings are equal, together with ideal speedup bounds for each branch and for their composition, from model dimensions and keep ratios alone. We then time both branches and their composition from 2K to 128K tokens on two GPUs after a dense prefill of real text, with dense and sparse modes reading the cache through the same split-K attention kernel. The projection branch leads at short context and the KV branch at long context, with speedups that follow their byte bounds up to fixed kernel costs. Adding these costs, measured in separate sweeps, lets the byte account predict the measured crossings of three keep-ratio pairs, a second model, and a second GPU to within 4.1K tokens. Timing the dense baseline with masked instead of split-K attention inflates the apparent speedup of the same KV policy about fivefold. An attention-scored KV selection answers the same passkey and multi-key placements as dense decoding up to 127K tokens, whereas a KV window misses most of them. Under matched perplexity budgets, activation sparsity composed with this selection decodes 14 to 26% faster than the best single branch on both GPUs. Code is available at https://github.com/js-lee-AI/ByteCross.

As LLMs scale to multi-million-token KV histories, real-time autoregressive decoding under tight Token-to-Token Latency (TTL) constraints faces growing pressure. Two core bottlenecks dominate: accessing Feed-Forward Network (FFN) weights and reading long KV caches. While Tensor Parallelism (TP) helps mitigate the cost of FFN weight reads, it does not scale well for attention. When TP width exceeds the number of KV heads, it leads to inefficient KV duplication, limits parallelism, and constrains batch size. Simultaneously, DRAM reads for long KV histories scale linearly with batch size, further capping efficiency. We introduce Helix Parallelism, a hybrid execution strategy that applies KV parallelism during attention to shard KV caches across GPUs, then reuses the same GPUs for TP in dense LLMs or TPxExpert Parallel (EP) in MoEs during FFN computation. To preserve exact attention behavior, Helix includes a lightweight communication step. To minimize the exposed communication cost, we introduce Helix HOP-B. Helix HOP-B effectively minimizes communication overhead through batchwise overlap, preserving low TTL while improving GPU efficiency. Compared to conventional parallelism approaches, Helix reduces TTL by up to 1.5x at fixed batch sizes and supports up to 32x larger batches under the same latency budget for DeepSeek-R1, pushing forward the throughput-latency Pareto on Blackwell and making real-time inference with ultra-long-sequence practical.

Speculative decoding accelerates memory-bound LLM decoding with a lightweight drafter, and parallel drafters extend this speedup by drafting a whole block in one forward pass. However, it remains unclear whether their released systems (stock) attain the full speedup that parallel drafting permits. We identify three challenges. First, with one request, each step accepts few tokens but reads every target weight. Second, with batching, per-request verification work roughly halves the speedup over autoregressive (AR) decoding. Third, speculative decoding is deployment-specific, and no universal optimization fits every serving regime, drafter, target model and GPU. To address these challenges, we present SpecPilot, an autonomous agentic loop that discovers algorithms and systems optimizations for each deployment. In each round, a cost model locates the largest gaps between the current system and its ceiling. Parallel agents explore directions and integrate the successful ones. Token forcing checks every output token against stock to detect reward hacking. A structured memory carries validated mechanisms to new rounds and deployments. SpecPilot discovered a general lossless 12-bit weight format, tree verification for hybrid recurrent targets, two-level verification and three batching mechanisms. We evaluate it across four target models, three drafters, two GPU platforms and five benchmarks. Without training or quality loss, it raises the speedup over AR from stock's 4.46times to 10.87times with one request, and from 1.94times to 3.93times with batching.

Large language model (LLM) decoding involves generating a sequence of tokens based on a given context, where each token is predicted one at a time using the model's learned probabilities. The typical autoregressive decoding method requires a separate forward pass through the model for each token generated, which is computationally inefficient and poses challenges for deploying LLMs in latency-sensitive scenarios. The main limitations of current decoding methods stem from their inefficiencies and resource demands. Existing approaches either necessitate fine-tuning smaller models, which is resource-intensive, or rely on fixed retrieval schemes to construct drafts for the next tokens, which lack adaptability and fail to generalize across different models and contexts. To address these issues, we introduce a novel methodology called ADED, which accelerates LLM decoding without requiring fine-tuning. Our approach involves an adaptive draft-verification process that evolves over time to improve efficiency. We utilize a tri-gram matrix-based LLM representation to dynamically approximate the output distribution of the LLM, allowing the model to adjust to changing token probabilities during the decoding process. Additionally, we implement a draft construction mechanism that effectively balances exploration and exploitation, ensuring that the drafts generated are both diverse and close to the true output distribution of the LLM. The importance of this design lies in its ability to optimize the draft distribution adaptively, leading to faster and more accurate decoding. Through extensive experiments on various benchmark datasets and LLM architectures, we demonstrate that ADED significantly accelerates the decoding process while maintaining high accuracy, making it suitable for deployment in a wide range of practical applications.

Extensive LLM applications demand efficient structured generations, particularly for LR(1) grammars, to produce outputs in specified formats (e.g., JSON). Existing methods primarily parse LR(1) grammars into a pushdown automaton (PDA), leading to runtime execution overhead for context-dependent token processing, especially inefficient under large inference batches. To address these issues, we propose Pre^3 that exploits deterministic pushdown automata (DPDA) to optimize the constrained LLM decoding efficiency. First, by precomputing prefix-conditioned edges during the preprocessing, Pre^3 enables ahead-of-time edge analysis and thus makes parallel transition processing possible. Second, by leveraging the prefix-conditioned edges, Pre^3 introduces a novel approach that transforms LR(1) transition graphs into DPDA, eliminating the need for runtime path exploration and achieving edge transitions with minimal overhead. Pre^3 can be seamlessly integrated into standard LLM inference frameworks, reducing time per output token (TPOT) by up to 40% and increasing throughput by up to 36% in our experiments. Our code is available at https://github.com/ModelTC/lightllm.

Large language model (LLM) decoding is latency-sensitive and often bottlenecked by fragmented operator execution and repeated off-chip materialization of intermediate tensors. Prior work expands fusion scope by leveraging thread-block clusters and on-chip inter-block collectives to fuse attention-side operators such as QKV projection, attention, and output projection. We develop ClusterFusion++, a CUDA-level extension that broadens fusion to the full Transformer decoder block for GPT-NeoX/Pythia models: LayerNorm -> QKV -> RoPE -> decode attention -> output projection -> Post-LN -> MLP -> residual. We additionally engineer a CUDA-Graph-compatible execution mode with persistent Tensor Memory Accelerator (TMA) descriptors to reduce per-step overhead. On an NVIDIA RTX 5090-class GPU, ClusterFusion++ improves throughput by 1.34x for Pythia-2.8B and yields similar gains for Pythia-6.9B, while maintaining high output fidelity (near-token-identical generation, with minor non-determinism from FP16 atomics).

Deployment of autoregressive large language models (LLMs) is costly, and as these models increase in size, the associated costs will become even more considerable. Consequently, different methods have been proposed to accelerate the token generation process and reduce costs. Speculative decoding (SD) is among the most promising approaches to speed up the LLM decoding process by verifying multiple tokens in parallel and using an auxiliary smaller draft model to generate the possible tokens. In SD, usually, one draft model is used to serve a specific target model; however, in practice, LLMs are diverse, and we might need to deal with many target models or more than one target model simultaneously. In this scenario, it is not clear which draft model should be used for which target model, and searching among different draft models or training customized draft models can further increase deployment costs. In this paper, we first introduce a novel multi-target scenario for the deployment of draft models for faster inference. Then, we present a novel, more efficient sorted speculative decoding mechanism that outperforms regular baselines in multi-target settings. We evaluated our method on Spec-Bench in different settings, including base models such as Vicuna 7B, 13B, and LLama Chat 70B. Our results suggest that our draft models perform better than baselines for multiple target models at the same time.

Scaling inference-time compute for Large Language Models (LLMs) has unlocked unprecedented reasoning capabilities. However, existing inference-time scaling methods typically rely on inefficient and suboptimal discrete search algorithms or trial-and-error prompting to improve the online policy. In this paper, we propose nabla-Reasoner, an iterative generation framework that integrates differentiable optimization over token logits into the decoding loop to refine the policy on the fly. Our core component, Differentiable Textual Optimization (DTO), leverages gradient signals from both the LLM's likelihood and a reward model to refine textual representations. nabla-Reasoner further incorporates rejection sampling and acceleration design to robustify and speed up decoding. Theoretically, we show that performing inference-time gradient descent in the sample space to maximize reward is dual to aligning an LLM policy via KL-regularized reinforcement learning. Empirically, nabla-Reasoner achieves over 20% accuracy improvement on a challenging mathematical reasoning benchmark, while reducing number of model calls by approximately 10-40% compared to strong baselines. Overall, our work introduces a paradigm shift from zeroth-order search to first-order optimization at test time, offering a cost-effective path to amplify LLM reasoning.

Top-k sparse attention makes long-context LLM decoding cheap to compute: each step reads only a few thousand selected KV entries rather than the full context. Serving systems, however, typically keep the entire KV cache in GPU HBM so that every position stays selectable, so a request's memory bill still grows with its full context length--decoding hits a capacity wall long before it runs out of compute, and a context whose KV cache exceeds HBM cannot be served at all. We present HiSparse, an exact, indexer-agnostic hierarchical KV cache for sparse-attention serving. HiSparse keeps each request's full KV history in host memory and bounds its decode footprint with a small, fixed-size GPU cache; a fused CUDA kernel resolves each layer's selections--hit detection, LRU replacement, and host-to-device fetches--inside the decode CUDA graph; and, for models that share selections across layers, exact layer-wise prefetching hides roughly half of the remaining miss overhead. Because only KV placement changes, model outputs are unchanged. HiSparse is merged into upstream SGLang and evaluated across three sparse-attention families (DSA, NSA, and Quest) on H200, B200, and GH200 platforms: it improves peak generation throughput by up to 4.7x on long-context workloads while preserving comparable per-token latency and reducing time-to-first-token at high load--and a no-IO oracle shows the resolution mechanism itself adds no measurable per-token cost, leaving host-device IO as the only price of bounded residency.

Test-time scaling (TTS) boosts LLM reasoning via long chain-of-thought (CoT), but the linear KV-cache growth amplifies the memory-bound bottleneck of LLM decoding. Query-aware page-level sparse decoding can achieve state-of-the-art performance under constrained FLOPs budgets, but is limited by both sequential-dependent page filtering and coarse-grained token selection, hampering serving efficiency and model performance on TTS tasks under high concurrency and long CoT scenarios (consuming even higher runtime than the forward pipeline itself). In this paper, we first find that the current-step query state can be accurately approximated in a unified manner from a short window of recent queries, enabling training-free query-aware sparsity without waiting in the decoding loop. We propose AsyncSpade, an asynchronous framework for efficient TTS built on two core components: (1) a novel light-weight temporal-regressive module that predicts the next-token query state; (2) an asynchronous and disaggregated framework that decouples the KV cache filtering from the auto-regressive decoding loop, overlapping the token-level KV selection with the forward inference computation through asynchronism. To our knowledge, AsyncSpade is the first to eliminate the sequential dependence without sacrificing model performance. We validate the effectiveness of AsyncSpade on common LLM serving setups with an A100 node, where AsyncSpade fully overlaps KV-cache operations with the inference pipeline, achieving theoretical optimal time-per-output-token (TPOT). Specifically, AsyncSpade delivers over 20% reduction on TPOT compared to SoTA baseline (i.e. Quest) and at least 50% TPOT reduction compared to full attention on Qwen3-8B and Qwen3-32B models, while matching or surpassing their accuracy on various TTS benchmarks (AIME-24/25, GPQA-Diamond, MATH-500).

Sparse attention reduces the KV-cache traffic that dominates long-context LLM decoding, but practical methods largely operate at coarse granularity. Token-level selection offers finer control, yet presents two challenges: low-dimensional scoring may lose query--key structure needed for accurate ranking, while gathering scattered selected tokens can offset the resulting latency benefit. We introduce Joint Subspace Sparse Attention (JSSA) to address both challenges. For selection quality, JSSA learns a shared low-dimensional query--key subspace by minimizing token-score preservation error. We upper-bound the resulting non-convex objective with a tractable marginal objective that admits a closed-form spectral solution; direct refinement of the exact objective provides no measurable improvement in the tested settings. For execution efficiency, we develop a gather-free token-level indexed-attention kernel that supports distinct per-KV-head selections under the standard grouped-query attention (GQA) KV-cache layout. Across standard GQA, hybrid sliding-window/global attention, and Multi-head Latent Attention (MLA) models, JSSA consistently remains close to full-KV accuracy while avoiding the substantial generation-length inflation and architecture-dependent failures observed with competing methods. On an NVIDIA B200 GPU, JSSA achieves up to 6.43times the decoding throughput of full-KV attention at 128K context. JSSA achieves reasoning accuracy comparable to model-native DSA without model training and reaches 96--99\% of its decoding throughput even at 700B scale large models.

Large language models (LLMs) have shown impressive capabilities in adapting to various tasks when provided with task-specific instructions. However, LLMs using standard decoding strategies often struggle with deviations from the inputs. Intuitively, compliant LLM outputs should reflect the information present in the input, which can be measured by point-wise mutual information (PMI) scores. Therefore, we propose Diver, a novel approach that enhances LLM Decoding through span-level PMI verification. During inference, Diver first identifies divergence steps that may lead to multiple candidate spans. Subsequently, it calculates the PMI scores by assessing the log-likelihood gains of the input if the candidate spans are generated. Finally, the optimal span is selected based on the PMI re-ranked output distributions. We evaluate our method across various downstream tasks, and empirical results demonstrate that Diver significantly outperforms existing decoding methods in both performance and versatility.

Speculative Decoding (SD) accelerates low-concurrency LLM inference by employing a draft-then-verify paradigm. However, mainstream methods typically rely on multi-token prediction, which introduces escalating prediction difficulty and serial drafting latency. To address these, we propose Speculative Pipeline Decoding (SPD), a groundbreaking framework that unlocks the true potential of pipeline parallelism. By partitioning the target LLM into n pipeline stages, SPD allows LLM to process n tokens in parallel to accelerate decoding. To continuous fill the pipeline in single sequence decoding, a speculation module aggregates intermediate features across different pipeline depths to predict the next token, executing strictly in parallel with the target model's pipeline step, to realize bounded difficulty, higher acceptance rates, and zero latency bubbles. Our experiments demonstrate that SPD achieves a significantly higher theoretical speedup compared to mainstream baselines, offering a highly scalable solution for LLM decoding acceleration. Our code is available at https://github.com/yuyijiong/speculative_pipeline_decoding

Large language models (LLMs) and large vision-language models (LVLMs) have demonstrated impressive generative capabilities, yet ensuring their outputs align with user intent is still challenging. While most existing approaches address this issue at the training stage, inference-time approaches like decoding methods offer a more efficient and scalable solution. Decoding methods control model generation by guiding token-level selection, performing sequence-level generation, or generating tokens in parallel to accelerate the process. In this survey, we identify three emerging paradigms from recent works on decoding methods for LLMs and LVLMs, provide a systematic review of these methods, highlight ongoing challenges, and discuss potential future research directions. Our goal is to underscore the efficiency and effectiveness of decoding methods and offer a practical view of their applications. Paper lists and more resources on decoding methods for LLMs and LVLMs can be found at https://github.com/wang2226/Awesome-LLM-Decoding.

The "end-to-end" label for LLMs is a misnomer. In practice, they depend on a non-differentiable decoding process that requires laborious, hand-tuning of hyperparameters like temperature and top-p. This paper introduces AutoDeco, a novel architecture that enables truly "end-to-end" generation by learning to control its own decoding strategy. We augment the standard transformer with lightweight heads that, at each step, dynamically predict context-specific temperature and top-p values alongside the next-token logits. This approach transforms decoding into a parametric, token-level process, allowing the model to self-regulate its sampling strategy within a single forward pass. Through extensive experiments on eight benchmarks, we demonstrate that AutoDeco not only significantly outperforms default decoding strategies but also achieves performance comparable to an oracle-tuned baseline derived from "hacking the test set"-a practical upper bound for any static method. Crucially, we uncover an emergent capability for instruction-based decoding control: the model learns to interpret natural language commands (e.g., "generate with low randomness") and adjusts its predicted temperature and top-p on a token-by-token basis, opening a new paradigm for steerable and interactive LLM decoding.

The complexity of large language model (LLM) serving workloads has substantially increased due to the integration with external tool invocations, such as ChatGPT plugins. In this paper, we identify a new opportunity for efficient LLM serving for requests that trigger tools: tool partial execution alongside LLM decoding. To this end, we design Conveyor, an efficient LLM serving system optimized for handling requests involving external tools. We introduce a novel interface for tool developers to expose partial execution opportunities to the LLM serving system and a request scheduler that facilitates partial tool execution. Our results demonstrate that tool partial execution can improve request completion latency by up to 38.8%.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

Optimizing recommender systems for objectives beyond accuracy, such as diversity, novelty, and personalization, is crucial for long-term user satisfaction. To this end, industrial practitioners have accumulated vast amounts of structured domain knowledge, which we term human priors (e.g., item taxonomies, temporal patterns). This knowledge is typically applied through post-hoc adjustments during ranking or post-ranking. However, this approach remains decoupled from the core model learning, which is particularly undesirable as the industry shifts to end-to-end generative recommendation foundation models. On the other hand, many methods targeting these beyond-accuracy objectives often require architecture-specific modifications and discard these valuable human priors by learning user intent in a fully unsupervised manner. Instead of discarding the human priors accumulated over years of practice, we introduce a backbone-agnostic framework that seamlessly integrates these human priors directly into the end-to-end training of generative recommenders. With lightweight, prior-conditioned adapter heads inspired by efficient LLM decoding strategies, our approach guides the model to disentangle user intent along human-understandable axes (e.g., interaction types, long- vs. short-term interests). We also introduce a hierarchical composition strategy for modeling complex interactions across different prior types. Extensive experiments on three large-scale datasets demonstrate that our method significantly enhances both accuracy and beyond-accuracy objectives. We also show that human priors allow the backbone model to more effectively leverage longer context lengths and larger model sizes.

Semi-structured 2:4 sparsity is widely supported by modern accelerators, providing up to a 2x theoretical speedup. However, its strict 50% sparsity constraint often causes non-negligible accuracy degradation under post-training pruning. Meanwhile, existing relaxed sparsity formats either require specialized compiler support or introduce runtime overheads that limit end-to-end speedup. We propose Spense, a practical hybrid sparse-dense format that splits each weight matrix into a 2:4 sparse region and a dense region. This design relaxes the effective sparsity constraint while remaining compatible with existing high-performance sparse and dense GEMM libraries, avoiding both custom compiler support and input activation expansion. Building on this format, we introduce SpenseGPT, a one-shot post-training pruning method that produces sparse and dense regions. Notably, we show that selecting the right dense regions is important, and we devise two different strategies to choose them. Experiments on Qwen3-32B and Seed-OSS-36B demonstrate that our method achieves up to 1.2x end-to-end decoding speedup on B200 GPUs with FP8 precision, while preserving accuracy. To the best of our knowledge, this is the first one-shot pruning demonstration of real-world end-to-end LLM decoding speedup from semi-structured sparse tensor cores on recent GPUs such as B200s, while maintaining model quality.

Recent work on activation and latent steering has demonstrated that modifying internal representations can effectively guide large language models (LLMs) toward improved reasoning and efficiency without updating model parameters. However, most existing approaches rely on fixed steering policies and static intervention strengths, which limit their robustness across problem instances and often result in over- or under-steering. We propose Adaptive Test-time Latent Steering (ATLAS), a lightweight framework that dynamically controls steering decisions at inference time using a trained, lightweight verifier over the latent states. Given intermediate hidden states, the verifier predicts the quality of ongoing reasoning and adaptively selects which steering action to apply, enabling per-example and per-step adjustment with minimal overhead. ATLAS provides a unified framework for combining learned latent verification with test-time activation steering, enabling adaptive reasoning control without additional LLM decoding or inference-time process reward model calls. Experiments on multiple mathematical and coding reasoning benchmarks show that ATLAS consistently outperforms both vanilla decoding and fixed steering baselines, achieving higher accuracy while substantially reducing test-time token usage. These results demonstrate that verifier-guided latent adaptation provides an effective and scalable mechanism for controlling reasoning efficiency without sacrificing solution quality. All source code will be publicly available.

Large language models (LLMs) are increasingly used for both open-ended and structured tasks, yet their inference-time behavior is still largely dictated by heuristic decoding strategies such as greedy search, sampling, or reranking. These methods provide limited control and do not explicitly optimize for task-specific objectives. We introduce MEMETRON, a task-agnostic framework that formulates LLM decoding as a discrete black-box optimization problem. MEMETRON leverages hybrid metaheuristic algorithms, GENETRON and ANNETRON, to search the response space, guided by reward models and contextual operations performed by the LLM itself. This approach enables efficient discovery of high-reward responses without requiring model retraining or gradient access. The framework is modular and generalizes across diverse tasks, requiring only a reward function and lightweight prompt templates. We evaluate our framework on the critical human preference alignment task and demonstrate that it significantly outperforms standard decoding and reranking methods, highlighting its potential to improve alignment without model retraining.

Large Language Models (LLMs), constrained by their auto-regressive nature, suffer from slow decoding. Speculative decoding methods have emerged as a promising solution to accelerate LLM decoding, attracting attention from both systems and AI research communities. Recently, the pursuit of better draft quality has driven a trend toward parametrically larger draft models, which inevitably introduces substantial computational overhead. While existing work attempts to balance the trade-off between prediction accuracy and compute latency, we address this fundamental dilemma through architectural innovation. We propose PRISM, which disaggregates the computation of each predictive step across different parameter sets, refactoring the computational pathways of draft models to successfully decouple model capacity from inference cost. Through extensive experiments, we demonstrate that PRISM outperforms all existing draft architectures, achieving exceptional acceptance lengths while maintaining minimal draft latency for superior end-to-end speedup. We also re-examine scaling laws with PRISM, revealing that PRISM scales more effectively with expanding data volumes than other draft architectures. Through rigorous and fair comparison, we show that PRISM boosts the decoding throughput of an already highly optimized inference engine by more than 2.6x.

Long-context LLM decoding reads the key-value (KV) cache at every step. Loading it takes longer than computing attention over it, so throughput is bandwidth-bound. Hence, reducing the cache size can raise both decoding speed and serving capacity. The challenge is to reduce cache size while preserving the attention products, keeping reconstruction cheap, and using a fixed per-token bit count. At two bits per element, the most competitive methods rely on orthogonal transforms. However, existing techniques are either data-oblivious or use the query statistics without deriving the transform from a distortion criterion. Moreover, they rely on transforms built on top of random or Hadamard rotations, which equalize variances across entries rather than compacting energy, and fixed-width scalar quantizers, which are suboptimal at low rates. In this paper, we formulate KV cache quantization as a transform coding problem in which distortion is the error in the attention products. We derive closed-form optimal transforms for keys and values from calibration statistics, under a high-resolution model. We show that the optimal key transform is not orthogonal and satisfies a generalized Parseval relation: the attention-aware distortion becomes mean-squared error (MSE) in the transform domain. Thus, we can use MSE-optimal vector quantizers applied directly to the transformed key coefficients. To meet the fixed-width layout requirement, we show that grouping coefficients into equal-volume partitions makes equal-size codebooks attain the variable-rate optimum under the same high-resolution model. At two bits per element, our method, termed NOVA-KV, recovers most of the long-context retrieval accuracy lost by scalar quantization methods at comparable throughput.

Large language models (LLMs) generally rely on sampling temperature to regulate the amount of randomness in their outputs. Mainstream methods for selecting temperature typically involve fixed, heuristic global settings and context or token-level calibration, rather than a geometrically interpretable representation of the communicative scenario. In this work, we introduce a simple and interpretable temperature selection framework, the Scenario-Actor, which adapts decoding to the inferred situational characteristics. Scenario-Actor extracts scenario-relevant structures from Transformer hidden states and determines the temperature from scenario embedding, a structured geometric representation of the communicative setting. Empirical results show that on Llama-3.1-8B-Instruct, Scenario-Actor achieves strong goal-completion performance in interactive social scenarios, significant gains on open-ended instruction following, and competitive performance on multi-turn conversation compared to multiple baselines.

Parallel decoding methods such as Jacobi decoding show promise for more efficient LLM inference as it breaks the sequential nature of the LLM decoding process and transforms it into parallelizable computation. However, in practice, it achieves little speedup compared to traditional autoregressive (AR) decoding, primarily because Jacobi decoding seldom accurately predicts more than one token in a single fixed-point iteration step. To address this, we develop a new approach aimed at realizing fast convergence from any state to the fixed point on a Jacobi trajectory. This is accomplished by refining the target LLM to consistently predict the fixed point given any state as input. Extensive experiments demonstrate the effectiveness of our method, showing 2.4times to 3.4times improvements in generation speed while preserving generation quality across both domain-specific and open-domain benchmarks.

Decoding methods for large language models (LLMs) usually struggle with the tradeoff between ensuring factuality and maintaining diversity. For example, a higher p threshold in the nucleus (top-p) sampling increases the diversity but decreases the factuality, and vice versa. In this paper, we propose REAL (Residual Entropy from Asymptotic Line) sampling, a decoding method that achieves improved factuality and diversity over nucleus sampling by predicting an adaptive threshold of p. Specifically, REAL sampling predicts the step-wise likelihood of an LLM to hallucinate, and lowers the p threshold when an LLM is likely to hallucinate. Otherwise, REAL sampling increases the p threshold to boost the diversity. To predict the step-wise hallucination likelihood without supervision, we construct a Token-level Hallucination Forecasting (THF) model to predict the asymptotic entropy (i.e., inherent uncertainty) of the next token by extrapolating the next-token entropies from a series of LLMs with different sizes. If a LLM's entropy is higher than the asymptotic entropy (i.e., the LLM is more uncertain than it should be), the THF model predicts a high hallucination hazard, which leads to a lower p threshold in REAL sampling. In the FactualityPrompts benchmark, we demonstrate that REAL sampling based on a 70M THF model can substantially improve the factuality and diversity of 7B LLMs simultaneously, judged by both retrieval-based metrics and human evaluation. After combined with contrastive decoding, REAL sampling outperforms 9 sampling methods, and generates texts that are more factual than the greedy sampling and more diverse than the nucleus sampling with p=0.5. Furthermore, the predicted asymptotic entropy is also a useful unsupervised signal for hallucination detection tasks.

Chain-of-Thought (CoT) has been proven effective in enhancing the reasoning capabilities of large language models (LLMs). Recent advancements, such as OpenAI's o1 and DeepSeek-R1, suggest that scaling up the length of CoT sequences during inference could further boost LLM reasoning performance. However, due to the autoregressive nature of LLM decoding, longer CoT outputs lead to a linear increase in inference latency, adversely affecting user experience, particularly when the CoT exceeds 10,000 tokens. To address this limitation, we analyze the semantic importance of tokens within CoT outputs and reveal that their contributions to reasoning vary. Building on this insight, we propose TokenSkip, a simple yet effective approach that enables LLMs to selectively skip less important tokens, allowing for controllable CoT compression. Extensive experiments across various models and tasks demonstrate the effectiveness of TokenSkip in reducing CoT token usage while preserving strong reasoning performance. Notably, when applied to Qwen2.5-14B-Instruct, TokenSkip reduces reasoning tokens by 40% (from 313 to 181) on GSM8K, with less than a 0.4% performance drop.

We introduce FUSION, a family of multimodal large language models (MLLMs) with a fully vision-language alignment and integration paradigm. Unlike existing methods that primarily rely on late-stage modality interaction during LLM decoding, our approach achieves deep, dynamic integration throughout the entire processing pipeline. To this end, we propose Text-Guided Unified Vision Encoding, incorporating textual information in vision encoding to achieve pixel-level integration. We further design Context-Aware Recursive Alignment Decoding that recursively aggregates visual features conditioned on textual context during decoding, enabling fine-grained, question-level semantic integration. To guide feature mapping and mitigate modality discrepancies, we develop Dual-Supervised Semantic Mapping Loss. Additionally, we construct a Synthesized Language-Driven Question-Answer (QA) dataset through a new data synthesis method, prioritizing high-quality QA pairs to optimize text-guided feature integration. Building on these foundations, we train FUSION at two scales-3B, 8B-and demonstrate that our full-modality integration approach significantly outperforms existing methods with only 630 vision tokens. Notably, FUSION 3B surpasses Cambrian-1 8B and Florence-VL 8B on most benchmarks. FUSION 3B continues to outperform Cambrian-1 8B even when limited to 300 vision tokens. Our ablation studies show that FUSION outperforms LLaVA-NeXT on over half of the benchmarks under same configuration without dynamic resolution, highlighting the effectiveness of our approach. We release our code, model weights, and dataset. https://github.com/starriver030515/FUSION

We present Statistical Calibrated Activation Pruning (SCAP), a post-training activation pruning framework that (1) generalizes sparsification by input activations of Fully-Connected layers for generic and flexible application across Transformers, and (2) features a simple Mode-Centering technique to pre-calibrate activation distributions for maximizing post-training sparsity. Our results demonstrate robust Pareto efficiency compared to prior methods, translating to a 1.5x additional LLM decoding speedup against CATS at iso model quality. SCAP effectiveness is empirically verified across a wide range of models, including recent Transformer Decoders, MoE, Mamba2, Encoding Transformer, and pre-quantized models, highlighting its practicality and scalability. The code is available at: https://github.com/IntelLabs/SCAP.

Large language models (LLMs) have demonstrated impressive reasoning abilities, but they still struggle with faithful reasoning due to knowledge gaps and hallucinations. To address these issues, knowledge graphs (KGs) have been utilized to enhance LLM reasoning through their structured knowledge. However, existing KG-enhanced methods, either retrieval-based or agent-based, encounter difficulties in accurately retrieving knowledge and efficiently traversing KGs at scale. In this work, we introduce graph-constrained reasoning (GCR), a novel framework that bridges structured knowledge in KGs with unstructured reasoning in LLMs. To eliminate hallucinations, GCR ensures faithful KG-grounded reasoning by integrating KG structure into the LLM decoding process through KG-Trie, a trie-based index that encodes KG reasoning paths. KG-Trie constrains the decoding process, allowing LLMs to directly reason on graphs and generate faithful reasoning paths grounded in KGs. Additionally, GCR leverages a lightweight KG-specialized LLM for graph-constrained reasoning alongside a powerful general LLM for inductive reasoning over multiple reasoning paths, resulting in accurate reasoning with zero reasoning hallucination. Extensive experiments on several KGQA benchmarks demonstrate that GCR achieves state-of-the-art performance and exhibits strong zero-shot generalizability to unseen KGs without additional training.

Large Language Models (LLMs) have achieved strong performance across natural language and multimodal tasks, yet their practical deployment remains constrained by inference latency and kernel launch overhead, particularly in interactive, short-sequence settings. This paper presents a hybrid runtime framework that combines Just-In-Time (JIT) compilation with CUDA Graph execution to reduce launch overhead while preserving runtime flexibility during autoregressive decoding. The framework partitions transformer inference into static components executed via CUDA Graph replay and dynamic components handled through JIT-compiled kernels, enabling asynchronous graph capture and reuse across decoding steps. We evaluate the proposed approach on LLaMA-2 7B using single-GPU, batch-size-one inference across prompt lengths from 10 to 500 tokens. Experimental results show that the hybrid runtime reduces Time-to-First-Token (TTFT) by up to 66.0% and achieves lower P99 latency compared with TensorRT-LLM in this regime. These results indicate that hybrid JIT-CUDA Graph execution can effectively reduce inference latency and variance for short-sequence LLM workloads, making it a practical optimization strategy for latency-sensitive AI applications.

Next token prediction paradigm has been prevailing for autoregressive models in the era of LLMs. The current default sampling choice for popular LLMs is temperature scaling together with nucleus sampling to balance diversity and coherence. Nevertheless, such approach leads to inferior performance in various NLP tasks when the model is not certain about testing questions. To this end, we propose a brand new training-free decoding strategy, dubbed as Cautious Next Token Prediction (CNTP). In the decoding process, if the model has comparatively high prediction entropy at a certain step, we sample multiple trials starting from the step independently and stop when encountering any punctuation. Then we select the trial with the lowest perplexity score viewed as the most probable and reliable trial path given the model's capacity. The trial number is negatively correlated with the prediction confidence, i.e., the less confident the model is, the more trials it should sample. This is consistent with human beings' behaviour: when feeling uncertain or unconfident, one tends to think more creatively, exploring multiple thinking paths, to cautiously select the path one feels most confident about. Extensive experiments on both LLMs and MLLMs show that our proposed CNTP approach outperforms existing standard decoding strategies consistently by a clear margin. Moreover, the integration of CNTP with self consistency can further improve over vanilla self consistency. We believe our proposed CNTP has the potential to become one of the default choices for LLM decoding. Code is available at https://github.com/wyzjack/CNTP.

Large language model (LLM) inference serving is increasingly constrained by memory rather than compute. As long-context and long-form reasoning workloads become more prevalent, the key-value (KV) cache dominates both memory footprint and memory traffic during LLM token generation, i.e., decode. In particular, HBM capacity has become a scarce and costly resource that heavily limits inference batch size and system throughput. This paper presents OasisKV, a memory-centric LLM inference system design that alleviates HBM capacity pressure by decoupling full KV-cache storage from HBM during LLM decoding. Because decode-time attention is naturally sparse, OasisKV keeps only the KV entries of the most relevant tokens in HBMs for attention computation. We observe that future important tokens can be predicted accurately in advance using lookahead tokens drafted by speculative decoding (SD). OasisKV employs an efficient attention background pipeline to identify important KV blocks. They are then prefetched from higher-capacity memory tiers (e.g., host or remote memory) and staged in HBMs before being used in the next decode step. We implement OasisKV based on vLLM. The lookahead prediction is accurate enough to keep accuracy within 0.7 points of full attention under a 2,048-token KV budget. This lets OasisKV turn sparsity into throughput gain: 1.69times over dense vLLM on the reasoning workload at 0.1 points of accuracy loss, and up to 2.1times on multi-GPU long-context serving. Under prefill--decode disaggregation, OasisKV reaches about 2times dense throughput while admitting each request with 6.5--9.7times less KV and holding 2.2-2.6 less decode-node host memory than full KV transfer.

Large Language Models (LLMs) rely on Human Preference Alignment (HPA) to ensure the generation of safe content. Due to the heavy cost associated with fine-tuning, fine-tuning-free methods have emerged, typically modifying LLM decoding with external auxiliary methods. However, these methods do not essentially enhance the LLM itself. In this paper, we rethink the derivation procedures of DPO, based on which we conversely build an instant scorer using the states of the LLM before and after In-context Learning (ICL). Accordingly, we propose a novel approach called In-Context Direct Preference Optimization (ICDPO). It enables LLMs to borrow the HPA capabilities from superior LLMs with ICL, generating well-aligned responses as estimated by the aforementioned instant scorer, thereby enhancing the final performance. ICDPO can be further enhanced with a two-stage retriever and an upgraded scorer, both offering benefits. Extensive experiments show its effectiveness, particularly in outperforming two fine-tuning-free baselines, and it exhibits competitiveness with SFT + LoRA. We also conduct detailed analyses to offer comprehensive insights into ICDPO.

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 15.4times acceleration in self-attention operations and 3.9times acceleration in end-to-end per token latency in long context LLM decoding.

Multimodal large language models have experienced rapid growth, and numerous different models have emerged. The interpretability of LVLMs remains an under-explored area. Especially when faced with more complex tasks such as chain-of-thought reasoning, its internal mechanisms still resemble a black box that is difficult to decipher. By studying the interaction and information flow between images and text, we noticed that in models such as LLaVA1.5, image tokens that are semantically related to text are more likely to have information flow convergence in the LLM decoding layer, and these image tokens receive higher attention scores. However, those image tokens that are less relevant to the text do not have information flow convergence, and they only get very small attention scores. To efficiently utilize the image information, we propose a new image token reduction method, Simignore, which aims to improve the complex reasoning ability of LVLMs by computing the similarity between image and text embeddings and ignoring image tokens that are irrelevant and unimportant to the text. Through extensive experiments, we demonstrate the effectiveness of our method for complex reasoning tasks. The paper's source code can be accessed from https://github.com/FanshuoZeng/Simignore.

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.

Can a large language model (LLM) improve at code generation using only its own raw outputs, without a verifier, a teacher model, or reinforcement learning? We answer in the affirmative with simple self-distillation (SSD): sample solutions from the model with certain temperature and truncation configurations, then fine-tune on those samples with standard supervised fine-tuning. SSD improves Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench v6, with gains concentrating on harder problems, and it generalizes across Qwen and Llama models at 4B, 8B, and 30B scale, including both instruct and thinking variants. To understand why such a simple method can work, we trace these gains to a precision-exploration conflict in LLM decoding and show that SSD reshapes token distributions in a context-dependent way, suppressing distractor tails where precision matters while preserving useful diversity where exploration matters. Taken together, SSD offers a complementary post-training direction for improving LLM code generation.

As Large Language Models (LLMs) scale, weight-only quantization (W4A16: 4-bit weights, 16-bit activations) becomes critical for reducing memory footprint with minimal accuracy loss. However, its efficient deployment on Huawei's Ascend 910 Neural Processing Unit (NPU) is challenging due to limited native mixed-precision support and the accelerator's decoupled compute architecture. To enable quantization on such architecture, we present the first practical W4A16 matrix multiplication kernel tailored for the Ascend 910 NPU. Our design leverages vector cores for on-the-fly INT4-to-FP16 dequantization, cube cores for high-throughput GEMM, and Split-K parallelization to mitigate memory latency. Performance evaluations across diverse matrix shapes and batch sizes show our method outperforms data-parallel approaches when K >> N, a typical scenario in LLM decoding. Specially, our method can achieve a speedup ranging from 1.01x to 1.74x. In addition, our profile reveals the primary bottleneck is not dequantization compution itself, but extra global memory transfer for the weight, making W4A16 only reaching a maximum speedup of 1.48x over native FP16xFP16 matrix multiplication in PyTorch. In the long run, our method lays a solid foundation and provides insightful views for the efficient deployment of quantized large language models on various domain-specific accelerators.

Languages with rich static semantics, such as Rust, provide stronger guarantees for AI-generated code, but their strictness makes generation more difficult. Off-the-shelf compilers can provide useful feedback post-generation, but does not guide intermediate generation steps, such as those during autoregressive LLM decoding. Constrained decoding intervenes earlier by rejecting invalid tokens during sampling, but requires white-box model access and costly reimplementation for semantic constraints.We introduce generative compilation, the first approach to obtaining compiler feedback on partial programs during generation. The core technical device is a sealor: a lightweight, mostly syntax-guided transformation that converts partial programs into complete ones that standard compilers can diagnose. It is designed such that possible-to-complete partial programs are never rejected, while preserving enough code context to catch genuine dead ends early. We construct such a sealor on a core Rust-like calculus and prove that it satisfies these properties, all mechanized in Lean. We extend it to the first partial-program checker for real Rust. We evaluate our method on challenging repository-level Rust coding tasks, across both frontier black-box and open-weight models. We show that generative compilation reduces non-compiling outputs and improves functional correctness, relative to standard post-generation feedback. It does so by detecting a broad range of errors close to their source and early during generation, thereby reducing errors cascades and enabling focused diagnostics. More broadly, generative compilation is a step toward making compilers a first-class citizen of AI-assisted programming active during generation, rather than a separate post-generation check.

Despite their capabilities, Large Language Models (LLMs) remain opaque with limited understanding of their internal representations. Current interpretability methods, such as direct logit attribution (DLA) and sparse autoencoders (SAEs), provide restricted insight due to limitations such as the model's output vocabulary or unclear feature names. This work introduces Hyperdimensional Probe, a novel paradigm for decoding information from the LLM vector space. It combines ideas from symbolic representations and neural probing to project the model's residual stream into interpretable concepts via Vector Symbolic Architectures (VSAs). This probe combines the strengths of SAEs and conventional probes while overcoming their key limitations. We validate our decoding paradigm with controlled input-completion tasks, probing the model's final state before next-token prediction on inputs spanning syntactic pattern recognition, key-value associations, and abstract inference. We further assess it in a question-answering setting, examining the state of the model both before and after text generation. Our experiments show that our probe reliably extracts meaningful concepts across varied LLMs, embedding sizes, and input domains, also helping identify LLM failures. Our work advances information decoding in LLM vector space, enabling extracting more informative, interpretable, and structured features from neural representations.

Speculative decoding, which combines a draft model with a target model, has emerged as an effective approach to accelerate large language model (LLM) inference. However, existing methods often face a trade-off between the acceptance rate and decoding speed in downstream tasks due to the limited capacity of the draft model, making it difficult to ensure efficiency across diverse tasks. To address this problem, we propose a speculative decoding algorithm tailored for downstream task optimization. It includes an automatic task partitioning and assigning method, which automatically categorizes downstream tasks into different sub-tasks and assigns them to a set of heterogeneous draft models. Each draft model is aligned with the target model using task-specific data, thereby enhancing the consistency of inference results. In addition, our proposed method incorporates an online lightweight prompt classifier to dynamically route prompts to the appropriate draft model. Experimental results demonstrate that the proposed method improves draft accuracy by 6% to 50% over vanilla speculative decoding, while achieving a speedup of 1.10x to 2.64x in LLM inference.

As Large Language Models (LLMs) become increasingly integrated into our daily lives, the potential harms from deceptive behavior underlie the need for faithfully interpreting their decision-making. While traditional probing methods have shown some effectiveness, they remain best for narrowly scoped tasks while more comprehensive explanations are still necessary. To this end, we investigate meta-models-an architecture using a "meta-model" that takes activations from an "input-model" and answers natural language questions about the input-model's behaviors. We evaluate the meta-model's ability to generalize by training them on selected task types and assessing their out-of-distribution performance in deceptive scenarios. Our findings show that meta-models generalize well to out-of-distribution tasks and point towards opportunities for future research in this area. Our code is available at https://github.com/acostarelli/meta-models-public .

Recent advancements in generative large language models (LLMs) have significantly boosted the performance in natural language processing tasks. However, their efficiency is hampered by the inherent limitations in autoregressive token generation. While parallel decoding with token tree verification, e.g., Medusa, has been proposed to improve decoding parallelism and efficiency, it often struggles with maintaining contextual relationships due to its independent token prediction approach and incurs significant verification overhead, especially with large tree sizes and batch processing. In this paper, we propose ProPD, an efficient LLM parallel decoding framework based on dynamic token tree pruning and generation. ProPD features an advanced early pruning mechanism to efficiently eliminate unpromising token sequences to improve verification efficiency. Additionally, it introduces a dynamic token tree generation algorithm to balance the computation and parallelism of the verification phase in real-time and maximize the overall efficiency across different batch sizes, sequence lengths, and tasks, etc. We verify ProPD across a diverse set of datasets, LLMs, and batch sizes and demonstrate ProPD consistently outperforms existing decoding algorithms by 1.1-3.2x.

Deploying large language models (LLMs) in mobile and edge computing environments is constrained by limited on-device resources, scarce wireless bandwidth, and frequent model evolution. Although edge-cloud collaborative inference with speculative decoding (SD) can reduce end-to-end latency by executing a lightweight draft model at the edge and verifying it with a cloud-side target model, existing frameworks fundamentally rely on tight coupling between the two models. Consequently, repeated model synchronization introduces excessive communication overhead, increasing end-to-end latency, and ultimately limiting the scalability of SD in edge environments. To address these limitations, we propose FlexSpec, a communication-efficient collaborative inference framework tailored for evolving edge-cloud systems. The core design of FlexSpec is a shared-backbone architecture that allows a single and static edge-side draft model to remain compatible with a large family of evolving cloud-side target models. By decoupling edge deployment from cloud-side model updates, FlexSpec eliminates the need for edge-side retraining or repeated model downloads, substantially reducing communication and maintenance costs. Furthermore, to accommodate time-varying wireless conditions and heterogeneous device constraints, we develop a channel-aware adaptive speculation mechanism that dynamically adjusts the speculative draft length based on real-time channel state information and device energy budgets. Extensive experiments demonstrate that FlexSpec achieves superior performance compared to conventional SD approaches in terms of inference efficiency.

Large language models (LLMs) typically employ sampling or beam search, accompanied by prompts such as Chain-of-Thought (CoT), to boost reasoning and decoding ability. Recent work like Tree-of-Thought (ToT) and Reasoning via Planning (RAP) aim to augment the reasoning capabilities of LLMs by utilizing tree-search algorithms to guide multi-step reasoning. These methods mainly focus on LLMs' reasoning ability during inference and heavily rely on human-designed prompts to activate LLM as a value function, which lacks general applicability and scalability. To address these limitations, we present an AlphaZero-like tree-search framework for LLMs (termed TS-LLM), systematically illustrating how tree-search with a learned value function can guide LLMs' decoding ability. TS-LLM distinguishes itself in two key ways: (1) Leveraging a learned value function, our approach can be generally applied to different tasks beyond reasoning (such as RLHF alignment), and LLMs of any size, without prompting advanced, large-scale models. (2) It can guide LLM's decoding during both inference and training. Empirical evaluations across reasoning, planning, and RLHF alignment tasks validate the effectiveness of TS-LLM, even on trees with a depth of 64.

Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks--referred to as Streaming Heads--do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU. Code is provided in https://github.com/mit-han-lab/duo-attention.

Automated transcription of parliamentary proceedings faces significant hurdles due to demographic bias, dialectal variation, and technical artifacts such as utterance truncation during segmentation. This paper introduces the ROManian PARliamentary Speech Corpus (ROMPAR) dataset, a 17.80-hour corpus of Romanian and Moldavian parliamentary speech, featuring double-annotated ground truth and explicit labels for reconstructed word fragments. To build a robust ASR system, we propose a multi-task adversarial training framework that enforces demographic invariance across age, gender, and dialect. We address the inherent instability of adversarial objectives in generative architectures by introducing an exponential decay mechanism for the adversarial coefficients. Furthermore, we implement an LLM-guided decoding strategy with position-dependent weighting to facilitate morphological completion of truncated terminal words. Our results demonstrate that the proposed framework significantly reduces WER and achieves an F1-score of 96.6% in morphological reconstruction.

Large language models (LLMs) achieve strong performance on mathematical reasoning benchmarks, yet the mathematically meaningful skills underlying their reasoning remain underexplored. We introduce SHAPE, a framework that analyzes Chain-of-Thought (CoT) trajectories through two lenses developed in mathematics education: (1) semantic spaces: the model's evolving mathematical interpretations of a problem (e.g., algebraic, geometric), and (2) heuristics: the specific mathematical actions taken within those spaces (e.g., simplifying the problem, working backward). We first use SHAPE to analyze the reasoning patterns of various models. Our findings reveal that the mathematical heuristics employed by a model better explain final answer correctness than traditional CoT features. Furthermore, models are likely to reach correct solutions by concentrating their reasoning effort within a few semantic spaces rather than exploring many disparate ones -- a pattern consistent with human behavior. Next, we utilize the SHAPE lens to evaluate whether post-training truly enhances mathematical proficiency. We find that reinforcement learning induces mode-seeking in heuristic usage. Lastly, we post-train LLMs by promoting diverse heuristics and demonstrate its effectiveness in improving accuracy. Overall, SHAPE provides a theoretically-grounded diagnostic framework for decoding LLM reasoning and offers a new path toward post-training LLMs for math reasoning. The code for our model is available at https://github.com/holi-lab/SHAPE-of-CoT

Transformer-based large language models (LLMs) exhibit limitations such as generating unsafe responses, unreliable reasoning, etc. Existing inference intervention approaches attempt to mitigate these issues by finetuning additional models to produce calibration signals (such as rewards) that guide the LLM's decoding process. However, this solution introduces substantial time and space overhead due to the separate models required. This work proposes Non-disruptive parameters insertion (Otter), inserting extra parameters into the transformer architecture to predict calibration signals along with the original LLM output. Otter offers state-of-the-art performance on multiple demanding tasks while saving up to 86.5\% extra space and 98.5\% extra time. Furthermore, Otter seamlessly integrates with existing inference engines, requiring only a one-line code change, and the original model response remains accessible after the parameter insertion. Our code is publicly available at https://github.com/chenhan97/Otter

Speculative decoding is an inference-acceleration method for large language models (LLMs) where a small language model generates a draft-token sequence which is further verified by the target LLM in parallel. Recent works have advanced this method by establishing a draft-token tree, achieving superior performance over a single-sequence speculative decoding. However, those works independently generate tokens at each level of the tree, not leveraging the tree's entire diversifiability. Besides, their empirical superiority has been shown for fixed length of sequences, implicitly granting more computational resource to LLM for the tree-based methods. None of the existing works has conducted empirical studies with fixed target computational budgets despite its importance to resource-bounded devices. We present Recursive Speculative Decoding (RSD), a novel tree-based method that samples draft tokens without replacement and maximizes the diversity of the tree. During RSD's drafting, the tree is built by either Gumbel-Top-k trick that draws tokens without replacement in parallel or Stochastic Beam Search that samples sequences without replacement while early-truncating unlikely draft sequences and reducing the computational cost of LLM. We empirically evaluate RSD with Llama 2 and OPT models, showing that RSD outperforms the baseline methods, consistently for fixed draft sequence length and in most cases for fixed computational budgets at LLM.

Speculative decoding accelerates LLM inference by using a draft model to look ahead, but gains are capped by the cost of autoregressive draft generation: increasing draft size elevates acceptance rates but introduces additional latency overhead exacerbating the speed-accuracy tradeoff. Prior methods (Medusa, Hydra, EAGLE) partially reduce draft cost but either degrade acceptance or introduce overheads that limit scaling. We present Mirror Speculative Decoding (Mirror-SD), an inference algorithm that breaks the latency-acceptance tradeoff. Mirror-SD launches branch-complete rollouts from early-exit signals in parallel with the target model's suffix and explicitly maps computation across heterogeneous accelerators (GPU and NPU) to exploit cross-device parallelism. The draft speculates forward continuations for the target to verify, while the target simultaneously speculates correction paths for the draft, converting speculation into two complementary execution pipelines. To further cut draft latency without weakening acceptance semantics, we add speculative streaming so the draft emits multiple tokens per step. This dual strategy of parallel heterogeneous execution plus multi-token speculative streaming pushes speculative decoding toward its ideal regime of high acceptance with low overhead. On SpecBench with server-scale models from 14B to 66B parameters, Mirror-SD delivers consistent end-to-end gains, achieving 2.8x-5.8x wall-time speedups across diverse tasks and a 30% average relative improvement over the strongest baseline, EAGLE3.

Large language models (LLMs) have revolutionized natural language processing and broadened their applicability across diverse commercial applications. However, the deployment of these models is constrained by high inference time in multilingual settings. To mitigate this challenge, this paper explores a training recipe of an assistant model in speculative decoding, which are leveraged to draft and-then its future tokens are verified by the target LLM. We show that language-specific draft models, optimized through a targeted pretrain-and-finetune strategy, substantially brings a speedup of inference time compared to the previous methods. We validate these models across various languages in inference time, out-of-domain speedup, and GPT-4o evaluation.

We propose self-speculative decoding for speech-aware LLMs by using the CTC encoder as a draft model to accelerate auto-regressive (AR) inference and improve ASR accuracy. Our three-step procedure works as follows: (1) if the frame entropies of the CTC output distributions are below a threshold, the greedy CTC hypothesis is accepted as final; (2) otherwise, the CTC hypothesis is verified in a single LLM forward pass using a relaxed acceptance criterion based on token likelihoods; (3) if verification fails, AR decoding resumes from the accepted CTC prefix. Experiments on nine corpora and five languages show that this approach can simultaneously accelerate decoding and reduce WER. On the HuggingFace Open ASR benchmark with a 1B parameter LLM and 440M parameter CTC encoder, we achieve a record 5.58% WER and improve the inverse real time factor by a factor of 4.4 with only a 12% relative WER increase over AR search. Code and model weights are publicly available under a permissive license.

Large language models (LLMs) have demonstrated immense utility across various industries. However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious instruction prompts. While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful responses from the prefill-level lacks utilization of the model's decoding outputs, leading to relatively lower effectiveness and robustness. 2) Rejecting potentially harmful responses based on a single evaluation can significantly impair the model's helpfulness.This paper examines the LLMs' capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens. Motivated by pilot experiment results, we design a robust defense mechanism at the decoding level. Our novel decoder-oriented, step-by-step defense architecture corrects harmful queries directly rather than rejecting them outright. We introduce speculative decoding to enhance usability and facilitate deployment to boost secure decoding speed. Extensive experiments demonstrate that our approach improves model security without compromising reasoning speed. Notably, our method leverages the model's ability to discern hazardous information, maintaining its helpfulness compared to existing methods.

Speculative decoding (SD) has emerged as a widely used paradigm to accelerate the inference of large language models (LLMs) without compromising generation quality. It works by first employing a compact model to draft multiple tokens efficiently and then using the target LLM to verify them in parallel. While this technique has achieved notable speedups, most existing approaches necessitate either additional parameters or extensive training to construct effective draft models, thereby restricting their applicability across different LLMs and tasks. To address this limitation, we explore a novel plug-and-play SD solution with layer-skipping, which skips intermediate layers of the target LLM as the compact draft model. Our analysis reveals that LLMs exhibit great potential for self-acceleration through layer sparsity and the task-specific nature of this sparsity. Building on these insights, we introduce SWIFT, an on-the-fly self-speculative decoding algorithm that adaptively selects intermediate layers of LLMs to skip during inference. SWIFT does not require auxiliary models or additional training, making it a plug-and-play solution for accelerating LLM inference across diverse input data streams. Our extensive experiments across a wide range of models and downstream tasks demonstrate that SWIFT can achieve over a 1.3x-1.6x speedup while preserving the original distribution of the generated text.

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 4.7times compared to the standard pipeline, all with negligible accuracy degradation.

Large language model (LLM)-based agent applications often incur high response time. Speculative decoding is a promising solution to improve the inference efficiency of LLM agents without impacting generation quality. However, state-of-the-art speculative decoding algorithms exhibit substantial speed degradation under large batch sizes, limiting their effectiveness to deploy in real-world agent applications. In this work, we first present a systematic analysis of speculative decoding for LLM agents and identify two dominant factors of speedup degradation: high rejection rate of speculative tokens, and under-utilization of dynamic token budgets.B ased on these observations, we propose AgentSpec, a speculative decoding algorithm that addresses the limitations of existing methods for LLM agents. AgentSpec incorporates structure-isolated drafting that constrains speculation to semantically coherent segments of the agent workflow, reducing the drafts of irrelevant semantic paths and achieving an extremely low rejection rate. Moreover, AgentSpec adopts redundancy-aware budget allocation that exploits agent-level information to better utilize the dynamically-free token budget during the agent inference. We implement and evaluate AgentSpec on five different workloads and four different models from four different LLM families in vLLM. Our results demonstrate the superiority of AgentSpec over state-of-the-arts.

Large language models (LLMs) often hallucinate and lack the ability to provide attribution for their generations. Semi-parametric LMs, such as kNN-LM, approach these limitations by refining the output of an LM for a given prompt using its nearest neighbor matches in a non-parametric data store. However, these models often exhibit slow inference speeds and produce non-fluent texts. In this paper, we introduce Nearest Neighbor Speculative Decoding (NEST), a novel semi-parametric language modeling approach that is capable of incorporating real-world text spans of arbitrary length into the LM generations and providing attribution to their sources. NEST performs token-level retrieval at each inference step to compute a semi-parametric mixture distribution and identify promising span continuations in a corpus. It then uses an approximate speculative decoding procedure that accepts a prefix of the retrieved span or generates a new token. NEST significantly enhances the generation quality and attribution rate of the base LM across a variety of knowledge-intensive tasks, surpassing the conventional kNN-LM method and performing competitively with in-context retrieval augmentation. In addition, NEST substantially improves the generation speed, achieving a 1.8x speedup in inference time when applied to Llama-2-Chat 70B.

Large Language Models (LLMs) exhibit high inference latency due to their autoregressive decoding nature. While the draft head in speculative decoding mitigates this issue, its full potential remains unexplored. In this paper, we introduce KOALA (K-layer Optimized Adversarial Learning Architecture), an orthogonal approach to the draft head. By transforming the conventional single-layer draft head into a multi-layer architecture and incorporating adversarial learning into the traditional supervised training, KOALA significantly improves the accuracy of the draft head in predicting subsequent tokens, thus more closely mirroring the functionality of LLMs. Although this improvement comes at the cost of slightly increased drafting overhead, KOALA substantially unlocks the draft head's potential, greatly enhancing speculative decoding. We conducted comprehensive evaluations of KOALA, including both autoregressive and non-autoregressive draft heads across various tasks, demonstrating a latency speedup ratio improvement of 0.24x-0.41x, which is 10.57%-14.09% faster than the original draft heads.

Aligning foundation models is essential for their safe and trustworthy deployment. However, traditional fine-tuning methods are computationally intensive and require updating billions of model parameters. A promising alternative, alignment via decoding, adjusts the response distribution directly without model updates to maximize a target reward r, thus providing a lightweight and adaptable framework for alignment. However, principled decoding methods rely on oracle access to an optimal Q-function (Q^*), which is often unavailable in practice. Hence, prior SoTA methods either approximate this Q^* using Q^{pi_{sft}} (derived from the reference SFT model) or rely on short-term rewards, resulting in sub-optimal decoding performance. In this work, we propose Transfer Q^*, which implicitly estimates the optimal value function for a target reward r through a baseline model rho_{BL} aligned with a baseline reward rho_{BL} (which can be different from the target reward r). Theoretical analyses of Transfer Q^* provide a rigorous characterization of its optimality, deriving an upper bound on the sub-optimality gap and identifying a hyperparameter to control the deviation from the pre-trained reference SFT model based on user needs. Our approach significantly reduces the sub-optimality gap observed in prior SoTA methods and demonstrates superior empirical performance across key metrics such as coherence, diversity, and quality in extensive tests on several synthetic and real datasets.

Generative retrieval has emerged as a powerful paradigm for LLM-based recommendation. However, industrial recommender systems often benefit from restricting the output space to a constrained subset of items based on business logic (e.g. enforcing content freshness or product category), which standard autoregressive decoding cannot natively support. Moreover, existing constrained decoding methods that make use of prefix trees (Tries) incur severe latency penalties on hardware accelerators (TPUs/GPUs). In this work, we introduce STATIC (Sparse Transition Matrix-Accelerated Trie Index for Constrained Decoding), an efficient and scalable constrained decoding technique designed specifically for high-throughput LLM-based generative retrieval on TPUs/GPUs. By flattening the prefix tree into a static Compressed Sparse Row (CSR) matrix, we transform irregular tree traversals into fully vectorized sparse matrix operations, unlocking massive efficiency gains on hardware accelerators. We deploy STATIC on a large-scale industrial video recommendation platform serving billions of users. STATIC produces significant product metric impact with minimal latency overhead (0.033 ms per step and 0.25% of inference time), achieving a 948x speedup over a CPU trie implementation and a 47-1033x speedup over a hardware-accelerated binary-search baseline. Furthermore, the runtime overhead of STATIC remains extremely low across a wide range of practical configurations. To the best of our knowledge, STATIC enables the first production-scale deployment of strictly constrained generative retrieval. In addition, evaluation on academic benchmarks demonstrates that STATIC can considerably improve cold-start performance for generative retrieval. Our code is available at https://github.com/youtube/static-constraint-decoding.

Although Large Language Models (LLMs) excel in reasoning and generation for language tasks, they are not specifically designed for multimodal challenges. Training Multimodal Large Language Models (MLLMs), however, is resource-intensive and constrained by various training limitations. In this paper, we propose the Modular-based Visual Contrastive Decoding (MVCD) framework to move this obstacle. Our framework leverages LLMs' In-Context Learning (ICL) capability and the proposed visual contrastive-example decoding (CED), specifically tailored for this framework, without requiring any additional training. By converting visual signals into text and focusing on contrastive output distributions during decoding, we can highlight the new information introduced by contextual examples, explore their connections, and avoid over-reliance on prior encoded knowledge. MVCD enhances LLMs' visual perception to make it see and reason over the input visuals. To demonstrate MVCD's effectiveness, we conduct experiments with four LLMs across five question answering datasets. Our results not only show consistent improvement in model accuracy but well explain the effective components inside our decoding strategy. Our code will be available at https://github.com/Pbhgit/MVCD.

Due to the high resource demands of Large Language Models (LLMs), achieving widespread deployment on consumer-grade devices presents significant challenges. Typically, personal or consumer-grade devices, including servers configured prior to the era of large-scale models, generally have relatively weak GPUs and relatively strong CPUs. However, most current methods primarily depend on GPUs for computation. Therefore, we propose Dovetail, an approach that deploys the draft model on the GPU to generate draft tokens while allowing the target model to perform parallel verification on the CPU, thereby improving the utilization of all available hardware resources and occupying less inter-device communication bandwidth. Accordingly, we have redesigned the draft model to better align with heterogeneous hardware characteristics. To this end, we implemented several optimizations: reducing the number of draft tokens to mitigate latency in parallel verification, increasing the depth of the draft model to enhance its predictive capacity, and introducing DGF (Dynamic Gating Fusion) to improve the integration of features and token embeddings. In the HumanEval benchmark, Dovetail achieved an inference speed of 5.86 tokens per second for LLaMA2-Chat-7B using 3GB of VRAM, representing an approximately 2.77x improvement over CPU-only inference. Furthermore, the inference speed was increased to 8 tokens per second when utilizing 7GB of VRAM.

We propose LLMA, an LLM accelerator to losslessly speed up Large Language Model (LLM) inference with references. LLMA is motivated by the observation that there are abundant identical text spans between the decoding result by an LLM and the reference that is available in many real world scenarios (e.g., retrieved documents). LLMA first selects a text span from the reference and copies its tokens to the decoder and then efficiently checks the tokens' appropriateness as the decoding result in parallel within one decoding step. The improved computational parallelism allows LLMA to achieve over 2x speed-up for LLMs with identical generation results as greedy decoding in many practical generation scenarios where significant overlap between in-context reference and outputs exists (e.g., search engines and multi-turn conversations).

Speculative decoding accelerates LLM inference by drafting a tree of candidate continuations and verifying it in one target forward. Existing drafters fall into two camps with opposite weaknesses. Autoregressive drafters such as EAGLE-3 preserve dependence along each draft path but call the drafter once per tree depth, making drafting a non-trivial share of per-iteration latency. Parallel drafters cut drafter calls by predicting multiple future positions in one forward, but each position is predicted without seeing the others, producing paths the verifier rejects. In this paper, we propose SpecBlock, a block-iterative drafter that combines path dependence with cheap drafting. Each drafter forward produces K dependent positions and we call this a block. The draft tree grows through repeated block expansions. Two mechanisms explicitly carry path dependence to keep later draft positions accurate. Within each block, a layer-wise shift carries the previous position's hidden state into every decoder layer. Across blocks, each new block can start from any position of the previous block, inheriting its hidden state to extend the path. To spend verifier budget where acceptance is likely, a co-trained rank head replaces the fixed top-k tree by allocating per-position branching during drafting. To avoid training the drafter on prefixes it never produces at inference, a valid-prefix mask drops the loss at later positions once an earlier one is wrong. Beyond static drafting, a cost-aware bandit at deployment uses free verifier feedback to update the drafter selectively, only when the expected throughput gain exceeds the update cost. Experiments show that SpecBlock improves mean speedup by 8-13% over EAGLE-3 at 44-52% of its drafting cost, and cost-aware adaptation extends this lead to 11-19%.

Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs. We identify a previously-unreported phenomenon we call attention drift: as the drafter generates successive tokens within a speculation chain, attention progressively moves from the prompt onto its own recently-generated tokens. We observe this across both EAGLE3 drafters and MTP heads, suggesting drift is a property of drafter designs. We trace this to the un-normalized residual path between chain steps: the drafter's hidden state magnitude grows monotonically with chain depth, which exhibits dynamics consistent with additional pre-norm transformer layers stacked on the target rather than as a standalone autoregressive predictor. In order to limit the growth, we propose two architectural changes: Post-norm on the drafter hidden states and per-hidden-state RMSNorm after capturing target hidden states. Our interventions improve acceptance length over the current leading model, pre-norm EAGLE3, by up to 2times under template perturbation, 1.18times on long-context tasks, and 1.10times on seven standard benchmarks spanning multi-turn chat, math, and coding. Our changes also allow shorter train-time-test depths to generalize over longer drafting sequences.

Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel with the target model. However, its practical speedup is constrained by the trade-off between draft quality and drafting cost: autoregressive drafters model causal dependencies among draft tokens but incur sequential overhead, while parallel drafters reduce drafting cost but weaken intra-block dependency modeling. In this paper, we propose Domino, a speculative decoding framework that decouples causal dependency modeling from expensive autoregressive draft execution. Domino first uses a parallel draft backbone to produce preliminary draft distributions for the entire block, and then applies a lightweight Domino head to refine them with prefix-dependent causal information. To stabilize teacher-forced causal encoding, we further introduce a base-anchored training curriculum that first strengthens the parallel backbone and then gradually shifts optimization toward the causally corrected final distribution. Experiments on Qwen3 models show that Domino achieves up to \(5.49\times\) end-to-end speedup under the Transformers backend and up to \(5.8\times\) throughput speedup under SGLang serving.

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.

Speculative decoding accelerates LLM inference by verifying multiple drafted tokens in parallel, allowing a single target forward pass to accept several tokens. By construction, verification computes representations for both accepted and rejected tokens. Yet, conventional drafters retain only the representations of accepted tokens, leaving the substantial verifier computation spent on rejected tokens effectively wasted. We find that these discarded hidden states generated during target forward retain useful information about future tokens that can improve subsequent drafts. However, realizing this opportunity poses two distinct challenges. At inference, recycling overhead can increase drafting latency, diminishing the speedup gained from increased acceptance length. During training, standard parallel drafter training does not produce inference-aligned rejected states, while obtaining them through sequential rollouts would sacrifice parallelism across training positions. We introduce Carryover Drafting, which addresses both challenges. Carryover recycles rejected target hidden states as temporary KV context, allowing the drafter to selectively attend to them. It reuses the drafter's existing interface and adds only a single learned embedding to distinguish rejected states from committed context. The additional KV context is replaced each drafting round, keeping its length bounded by one proposal block. We introduce parallel draft--verify--draft training that exposes the drafter to inference-aligned rejected states while preserving parallelism across training positions. Experiments with DFlash and a DSpark-derived semi-autoregressive drafter across two target models show that this simple Carryover mechanism improves average acceptance length by 6.5--14.7% and end-to-end vLLM speedup by 7.9--14.4% over the corresponding baselines, with speedup gains reaching 28.8% on translation.

Speculative decoding speeds up LLM inference by using a small draft model to propose multiple tokens that a target model verifies in parallel. Extending this idea to batches is essential for production serving, but it introduces the ragged tensor problem: sequences in the same batch accept different numbers of draft tokens, breaking right-alignment and corrupting position IDs, attention masks, and KV-cache state. We show that several existing batch implementations violate output equivalence-the fundamental requirement that speculative decoding must produce identical token sequences to standard autoregressive generation. These violations occur precisely due to improper handling of the ragged tensor problem. In response, we (1) characterize the synchronization requirements that guarantee correctness, (2) present a correctness-first batch speculative decoding EQSPEC that exposes realignment as consuming 40% of overhead, and (3) introduce EXSPEC, which maintains a sliding pool of sequences and dynamically forms same-length groups, to reduce the realignment overhead while preserving per-sequence speculative speedups. On the SpecBench dataset, across Vicuna-7B/68M, Qwen3-8B/0.6B, and GLM-4-9B/0.6B target/draft pairs, our approach achieves up to 3times throughput improvement at batch size 8 compared to batch size 1, with efficient scaling through batch size 8, while maintaining 95% output equivalence. Our method requires no custom kernels and integrates cleanly with existing inference stacks. Our code is available at https://github.com/eBay/spec_dec.

Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model's internal knowledge. However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. In this paper, we present \modelname, which flares out the narrow conditioning bottleneck of DFlash through a lightweight layer-wise fusion mechanism: each draft layer attends to its own learnable combination of a broad set of target layers at negligible overhead, simultaneously injecting richer target knowledge and providing every draft layer with a distinct input. This enhanced per-layer expressiveness enables scaling the draft model to deeper architectures with consistent gains. We further scale training data from 800K to 2.4M samples to fully exploit the enlarged capacity. On six benchmarks spanning mathematical reasoning, code generation, and conversation, \modelname attains average wall-clock speedups of 5.52x on Qwen3-4B, 5.46x on Qwen3-8B, and 3.91x on GPT-OSS-20B, improving over DFlash by roughly 11\%, 8\%, and 5\% respectively. Our code is available at https://github.com/Tencent/AngelSlim.

In this work, we propose FastCoT, a model-agnostic framework based on parallel decoding without any further training of an auxiliary model or modification to the LLM itself. FastCoT uses a size-varying context window whose size changes with position to conduct parallel decoding and auto-regressive decoding simultaneously, thus fully utilizing GPU computation resources. In FastCoT, the parallel decoding part provides the LLM with a quick glance of the future composed of approximate tokens, which could lead to faster answers compared to regular autoregressive decoding used by causal transformers. We also provide an implementation of parallel decoding within LLM, which supports KV-cache generation and batch processing. Through extensive experiments, we demonstrate that FastCoT saves inference time by nearly 20% with only a negligible performance drop compared to the regular approach. Additionally, we show that the context window size exhibits considerable robustness for different tasks.

Decoder-only Large Language Models (LLMs) are currently the model of choice for many Natural Language Processing (NLP) applications. Through instruction fine-tuning and prompting approaches, such LLMs have been efficiently used to solve both general and domain-specific tasks. However, they are costly to train and, to a certain extent, costly to use as well, and one can wonder whether LLMs can be replaced by domain-specific Tiny Language Models (TLMs), which typically contain less than 100M parameters. We address this question in this study by comparing the performance of an 88M TLM pretrained from scratch for 185 A100 hours on a specific domain with a domain-specific tokenizer (here, the Indian legal domain) with LLMs of various sizes between 1B and 8B for solving domain-specific tasks. We show in particular that our legal TLM, Ayn, can indeed outperform LLMs up to 80 times larger on the legal case judgment prediction task, rival LLMs up to 30 times larger on the summarization task, and still be competitive with these larger LLMs on general tasks.

To mitigate the high inference latency stemming from autoregressive decoding in Large Language Models (LLMs), Speculative Decoding has emerged as a novel decoding paradigm for LLM inference. In each decoding step, this method first efficiently drafts several future tokens and then verifies them in parallel. Unlike autoregressive decoding, Speculative Decoding facilitates the simultaneous decoding of multiple tokens per step, thereby accelerating inference. This paper presents a comprehensive overview and analysis of this promising decoding paradigm. We begin by providing a formal definition and formulation of Speculative Decoding. Then, we organize in-depth discussions on its key facets, including current leading techniques, the challenges faced, and potential future directions in this field. We aim for this work to serve as a catalyst for further research on Speculative Decoding, ultimately contributing to more efficient LLM inference.

General-purpose LLM judges capable of human-level evaluation provide not only a scalable and accurate way of evaluating instruction-following LLMs but also new avenues for supervising and improving their performance. One promising way of leveraging LLM judges for supervision is through Minimum Bayes Risk (MBR) decoding, which uses a reference-based evaluator to select a high-quality output from amongst a set of candidate outputs. In the first part of this work, we explore using MBR decoding as a method for improving the test-time performance of instruction-following LLMs. We find that MBR decoding with reference-based LLM judges substantially improves over greedy decoding, best-of-N decoding with reference-free judges and MBR decoding with lexical and embedding-based metrics on AlpacaEval and MT-Bench. These gains are consistent across LLMs with up to 70B parameters, demonstrating that smaller LLM judges can be used to supervise much larger LLMs. Then, seeking to retain the improvements from MBR decoding while mitigating additional test-time costs, we explore iterative self-training on MBR-decoded outputs. We find that self-training using Direct Preference Optimisation leads to significant performance gains, such that the self-trained models with greedy decoding generally match and sometimes exceed the performance of their base models with MBR decoding.

Decoding methods play an indispensable role in converting language models from next-token predictors into practical task solvers. Prior research on decoding methods, primarily focusing on task-specific models, may not extend to the current era of general-purpose large language models (LLMs). Moreover, the recent influx of decoding strategies has further complicated this landscape. This paper provides a comprehensive and multifaceted analysis of various decoding methods within the context of LLMs, evaluating their performance, robustness to hyperparameter changes, and decoding speeds across a wide range of tasks, models, and deployment environments. Our findings reveal that decoding method performance is notably task-dependent and influenced by factors such as alignment, model size, and quantization. Intriguingly, sensitivity analysis exposes that certain methods achieve superior performance at the cost of extensive hyperparameter tuning, highlighting the trade-off between attaining optimal results and the practicality of implementation in varying contexts.

Large language models (LLMs) have become ubiquitous in practice and are widely used for generation tasks such as translation, summarization and instruction following. However, their enormous size and reliance on autoregressive decoding increase deployment costs and complicate their use in latency-critical applications. In this work, we propose a hybrid approach that combines language models of different sizes to increase the efficiency of autoregressive decoding while maintaining high performance. Our method utilizes a pretrained frozen LLM that encodes all prompt tokens once in parallel, and uses the resulting representations to condition and guide a small language model (SLM), which then generates the response more efficiently. We investigate the combination of encoder-decoder LLMs with both encoder-decoder and decoder-only SLMs from different model families and only require fine-tuning of the SLM. Experiments with various benchmarks show substantial speedups of up to 4times, with minor performance penalties of 1-2% for translation and summarization tasks compared to the LLM.

The field of neural machine translation (NMT) has changed with the advent of large language models (LLMs). Much of the recent emphasis in natural language processing (NLP) has been on modeling machine translation and many other problems using a single pre-trained Transformer decoder, while encoder-decoder architectures, which were the standard in earlier NMT models, have received relatively less attention. In this paper, we explore translation models that are universal, efficient, and easy to optimize, by marrying the world of LLMs with the world of NMT. We apply LLMs to NMT encoding and leave the NMT decoder unchanged. We also develop methods for adapting LLMs to work better with the NMT decoder. Furthermore, we construct a new dataset involving multiple tasks to assess how well the machine translation system generalizes across various tasks. Evaluations on the WMT and our datasets show that results using our method match or surpass a range of baselines in terms of translation quality, but achieve 2.4 sim 6.5 times inference speedups and a 75% reduction in the memory footprint of the KV cache. It also demonstrates strong generalization across a variety of translation-related tasks.

Instruction tuning has emerged as the key in aligning large language models (LLMs) with specific task instructions, thereby mitigating the discrepancy between the next-token prediction objective and users' actual goals. To reduce the labor and time cost to collect or annotate data by humans, researchers start to explore the use of LLMs to generate instruction-aligned synthetic data. Recent works focus on generating diverse instructions and applying LLM to increase instruction complexity, often neglecting downstream use cases. It remains unclear how to tailor high-quality data to elicit better instruction-following abilities in different target instruction distributions and LLMs. To this end, we introduce CodecLM, a general framework for adaptively generating high-quality synthetic data for LLM alignment with different downstream instruction distributions and LLMs. Drawing on the Encode-Decode principles, we use LLMs as codecs to guide the data generation process. We first encode seed instructions into metadata, which are concise keywords generated on-the-fly to capture the target instruction distribution, and then decode metadata to create tailored instructions. We also introduce Self-Rubrics and Contrastive Filtering during decoding to tailor data-efficient samples. Extensive experiments on four open-domain instruction following benchmarks validate the effectiveness of CodecLM over the current state-of-the-arts.

Efficient inference in large language models (LLMs) has become a critical focus as their scale and complexity grow. Traditional autoregressive decoding, while effective, suffers from computational inefficiencies due to its sequential token generation process. Speculative decoding addresses this bottleneck by introducing a two-stage framework: drafting and verification. A smaller, efficient model generates a preliminary draft, which is then refined by a larger, more sophisticated model. This paper provides a comprehensive survey of speculative decoding methods, categorizing them into draft-centric and model-centric approaches. We discuss key ideas associated with each method, highlighting their potential for scaling LLM inference. This survey aims to guide future research in optimizing speculative decoding and its integration into real-world LLM applications.

Large Language Models (LLMs) have attracted extensive attention due to their remarkable performance across various tasks. However, the substantial computational and memory requirements of LLM inference pose challenges for deployment in resource-constrained scenarios. Efforts within the field have been directed towards developing techniques aimed at enhancing the efficiency of LLM inference. This paper presents a comprehensive survey of the existing literature on efficient LLM inference. We start by analyzing the primary causes of the inefficient LLM inference, i.e., the large model size, the quadratic-complexity attention operation, and the auto-regressive decoding approach. Then, we introduce a comprehensive taxonomy that organizes the current literature into data-level, model-level, and system-level optimization. Moreover, the paper includes comparative experiments on representative methods within critical sub-fields to provide quantitative insights. Last but not least, we provide some knowledge summary and discuss future research directions.

The escalating demand for efficient decoding in large language models (LLMs) is particularly critical for reasoning-intensive architectures like OpenAI-o3 and DeepSeek-R1, which depend on extended chain-of-thought reasoning. This study investigates speculative decoding techniques through dense LLM architectures to establish foundational insights for accelerating reasoning tasks. While speculative decoding methods leveraging parallel draft-verification cycles have emerged as promising acceleration techniques, the scaling laws governing decoding efficiency remain under-explored compared to conventional backbone LLMs developed through Pretraining->SFT->RLHF training paradigms. In this work, we discover Log-linear Scaling Laws (Theorem 1.1, 1.2 and 1.3) governing draft model acceptance rate (or decoding speed) across three dimensions: pretraining token volume, draft model capacity, and decoding batch size. Building on these laws, we achieve Scylla, which coordinates multi-dimensional scaling for popular LLMs (Llama2/3, Qwen2.5). Empirical validation shows Scylla achieves 1.5-2.2 higher acceptance rate than EAGLE2 and 0.3 higher than EAGLE3 at temperature T = 0, with peak performance gains on summarization and QA tasks (Figure 2). Industrial inference engine deployments demonstrate 2X decoding throughput improvements over EAGLE2 (Table 5), validating the transformative potential of systematic scaling for efficient LLM inference. Code will be released later.

Despite their remarkable achievements, modern Large Language Models (LLMs) encounter exorbitant computational and memory footprints. Recently, several works have shown significant success in training-free and data-free compression (pruning and quantization) of LLMs achieving 50-60% sparsity and reducing the bit-width down to 3 or 4 bits per weight, with negligible perplexity degradation over the uncompressed baseline. As recent research efforts are focused on developing increasingly sophisticated compression methods, our work takes a step back, and re-evaluates the effectiveness of existing SoTA compression methods, which rely on a fairly simple and widely questioned metric, perplexity (even for dense LLMs). We introduce Knowledge-Intensive Compressed LLM BenchmarK (LLM-KICK), a collection of carefully-curated tasks to re-define the evaluation protocol for compressed LLMs, which have significant alignment with their dense counterparts, and perplexity fail to capture subtle change in their true capabilities. LLM-KICK unveils many favorable merits and unfortunate plights of current SoTA compression methods: all pruning methods suffer significant performance degradation, sometimes at trivial sparsity ratios (e.g., 25-30%), and fail for N:M sparsity on knowledge-intensive tasks; current quantization methods are more successful than pruning; yet, pruned LLMs even at geq 50% sparsity are robust in-context retrieval and summarization systems; among others. LLM-KICK is designed to holistically access compressed LLMs' ability for language understanding, reasoning, generation, in-context retrieval, in-context summarization, etc. We hope our study can foster the development of better LLM compression methods. All our related codes are planed to be open-sourced.

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose ParallelBench, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We release our benchmark to help accelerate the development of truly efficient dLLMs.

Recent advances in interpretability suggest that large language models (LLMs) implicitly encode signals in their generated text that enable self-recognition of their outputs. We demonstrate that this capability is reliable, even in low-entropy scenarios, and that it can be amplified through targeted intervention. By steering the internal residual stream during generation with a random sparse vector, we create a detectable fingerprint that enables attribution of a given text to a specific LLM. This signal is recoverable from the activations of an LLM used as a detector, achieving over 98% accuracy across multiple detection settings while preserving the quality of generated text. As AI-generated content proliferates, this approach offers a practical alternative to traditional detectors by leveraging the model's natural representation structure for attribution rather than embedding a signal externally. Our contributions include: (i) establishing reliable self-recognition capabilities in LLMs, (ii) a simple steering mechanism enabling multi-LLM identification with no quality degradation, (iii) demonstrating that activation spaces contain exploitable structure for encoding signals without semantic interference.

Large Language Models (LLMs) have revolutionized natural language processing by understanding and generating human-like text. However, the increasing demand for more sophisticated LLMs presents significant computational challenges due to their scale and complexity. This paper introduces Hardware Accelerated Decoding (HADES), a novel approach to enhance the performance and energy efficiency of LLMs. We address the design of an LLM accelerator with hardware-level speculative decoding support, a concept not previously explored in existing literature. Our work demonstrates how speculative decoding can significantly improve the efficiency of LLM operations, paving the way for more advanced and practical applications of these models.

Text generation with Large Language Models (LLMs) is known to be memory bound due to the combination of their auto-regressive nature, huge parameter counts, and limited memory bandwidths, often resulting in low token rates. Speculative decoding has been proposed as a solution for LLM inference acceleration. However, since draft models are often unavailable in the modern open-source LLM families, e.g., for Llama 2 7B, training a high-quality draft model is required to enable inference acceleration via speculative decoding. In this paper, we propose a simple draft model training framework for direct alignment to chat-capable target models. With the proposed framework, we train Llama 2 Chat Drafter 115M, a draft model for Llama 2 Chat 7B or larger, with only 1.64\% of the original size. Our training framework only consists of pretraining, distillation dataset generation, and finetuning with knowledge distillation, with no additional alignment procedure. For the finetuning step, we use instruction-response pairs generated by target model for distillation in plausible data distribution, and propose a new Total Variation Distance++ (TVD++) loss that incorporates variance reduction techniques inspired from the policy gradient method in reinforcement learning. Our empirical results show that Llama 2 Chat Drafter 115M with speculative decoding achieves up to 2.3 block efficiency and 2.4times speed-up relative to autoregressive decoding on various tasks with no further task-specific fine-tuning.

Low-resource languages such as Sinhala are often overlooked by open-source Large Language Models (LLMs). In this research, we extend an existing multilingual LLM (Llama-3-8B) to better serve Sinhala. We enhance the LLM tokenizer with Sinhala specific vocabulary and perform continual pre-training on a cleaned 10 million Sinhala corpus, resulting in the SinLlama model. This is the very first decoder-based open-source LLM with explicit Sinhala support. When SinLlama was instruction fine-tuned for three text classification tasks, it outperformed base and instruct variants of Llama-3-8B by a significant margin.

Large Language Model (LLM) inference is hard. The autoregressive Decode phase of the underlying Transformer model makes LLM inference fundamentally different from training. Exacerbated by recent AI trends, the primary challenges are memory and interconnect rather than compute. To address these challenges, we highlight four architecture research opportunities: High Bandwidth Flash for 10X memory capacity with HBM-like bandwidth; Processing-Near-Memory and 3D memory-logic stacking for high memory bandwidth; and low-latency interconnect to speedup communication. While our focus is datacenter AI, we also review their applicability for mobile devices.

Speculative decoding is a relatively new decoding framework that leverages small and efficient draft models to reduce the latency of LLMs. In this study, we introduce GliDe and CaPE, two low-hassle modifications to vanilla speculative decoding to further improve the decoding speed of a frozen LLM. Specifically, GliDe is a modified draft model architecture that reuses the cached keys and values from the target LLM, while CaPE is a proposal expansion method that uses the draft model's confidence scores to help select additional candidate tokens for verification. Extensive experiments on different benchmarks demonstrate that our proposed GliDe draft model significantly reduces the expected decoding latency. Additional evaluation using walltime reveals that GliDe can accelerate Vicuna models up to 2.17x and further extend the improvement to 2.61x with CaPE. We will release our code, data, and the trained draft models.

Pre-trained language models based on masked language modeling (MLM) objective excel in natural language understanding (NLU) tasks. While fine-tuned MLM-based encoders consistently outperform causal language modeling decoders of comparable size, a recent trend of scaling decoder models to multiple billion parameters resulted in large language models (LLMs), making them competitive with MLM-based encoders. Although scale amplifies their prowess in NLU tasks, LLMs fall short of SOTA results in information extraction (IE) tasks, many framed as sequence labeling (SL). However, whether this is an intrinsic limitation of LLMs or whether their SL performance can be improved remains unclear. To address this, we explore strategies to enhance the SL performance of "open" LLMs (Llama2 and Mistral) on IE tasks. We investigate bidirectional information flow within groups of decoder blocks, applying layer-wise removal or enforcement of the causal mask (CM) during LLM fine-tuning. This approach yields performance gains competitive with SOTA SL models, matching or outperforming the results of CM removal from all blocks. Our findings hold for diverse SL tasks, proving that "open" LLMs with layer-dependent CM removal outperform strong MLM-based encoders and instruction-tuned LLMs. However, we observe no effect from CM removal on a small scale when maintaining an equivalent model size, pre-training steps, and pre-training and fine-tuning data.

Large decoder-only language models (LLMs) are the state-of-the-art models on most of today's NLP tasks and benchmarks. Yet, the community is only slowly adopting these models for text embedding tasks, which require rich contextualized representations. In this work, we introduce LLM2Vec, a simple unsupervised approach that can transform any decoder-only LLM into a strong text encoder. LLM2Vec consists of three simple steps: 1) enabling bidirectional attention, 2) masked next token prediction, and 3) unsupervised contrastive learning. We demonstrate the effectiveness of LLM2Vec by applying it to 3 popular LLMs ranging from 1.3B to 7B parameters and evaluate the transformed models on English word- and sequence-level tasks. We outperform encoder-only models by a large margin on word-level tasks and reach a new unsupervised state-of-the-art performance on the Massive Text Embeddings Benchmark (MTEB). Moreover, when combining LLM2Vec with supervised contrastive learning, we achieve state-of-the-art performance on MTEB among models that train only on publicly available data. Our strong empirical results and extensive analysis demonstrate that LLMs can be effectively transformed into universal text encoders in a parameter-efficient manner without the need for expensive adaptation or synthetic GPT-4 generated data.

Large Language Models have demonstrated remarkable abilities in reasoning and planning by breaking down complex problems into sequential steps. Despite their success in various domains like mathematical problem-solving and coding, LLMs face challenges in ensuring reliable and optimal planning due to their inherent myopic nature of autoregressive decoding. This paper revisits LLM reasoning from an optimal-control perspective, proposing a novel method, Predictive-Decoding, that leverages Model Predictive Control to enhance planning accuracy. By re-weighting LLM distributions based on foresight trajectories, Predictive-Decoding aims to mitigate early errors and promote non-myopic planning. Our experiments show significant improvements in a wide range of tasks for math, coding, and agents. Furthermore, Predictive-Decoding demonstrates computational efficiency, outperforming search baselines with reduced computational resources. This study provides insights into optimizing LLM planning capabilities.

Speculative decoding accelerates LLM inference by having a small drafter propose tokens that a larger target model verifies in parallel. Recent diffusion-based parallel drafters such as DFlash predict the full B-token block in one forward pass, enabling deeper drafters and longer accepted blocks. However, existing multi-token drafter objectives often use fixed position-dependent weighting schedules, such as head-dependent weights or block-position decays, which do not adapt as the positions limiting acceptance change during training. To address this, we derive per-position training weights from a differentiable surrogate of expected accepted draft length, matching the weight of each position to its log-probability gradient contribution. The resulting loss, D-PACE (Dynamic Position-Aware Cross-Entropy), shifts training signal toward positions that currently limit acceptance as the drafter improves. Across six benchmarks, two Qwen3-4B draft depths, two decoding temperatures, and two additional target models, D-PACE consistently improves both wall-clock speedup and average emitted length, with 2.3\% measured training-time overhead and no changes to the drafter architecture or inference procedure.

Transformer based Large Language Models (LLMs) have been widely used in many fields, and the efficiency of LLM inference becomes hot topic in real applications. However, LLMs are usually complicatedly designed in model structure with massive operations and perform inference in the auto-regressive mode, making it a challenging task to design a system with high efficiency. In this paper, we propose an efficient LLM inference solution with low latency and high throughput. Firstly, we simplify the LLM decoder layer by fusing data movement and element-wise operations to reduce the memory access frequency and lower system latency. We also propose a segment KV cache policy to keep key/value of the request and response tokens in separate physical memory for effective device memory management, helping enlarge the runtime batch size and improve system throughput. A customized Scaled-Dot-Product-Attention kernel is designed to match our fusion policy based on the segment KV cache solution. We implement our LLM inference solution on Intel GPU and publish it publicly. Compared with the standard HuggingFace implementation, the proposed solution achieves up to 7x lower token latency and 27x higher throughput for some popular LLMs on Intel GPU.

Speculative Decoding is a widely used technique to speed up inference for Large Language Models (LLMs) without sacrificing quality. When performing inference, speculative decoding uses a smaller draft model to generate speculative tokens and then uses the target LLM to verify those draft tokens. The speedup provided by speculative decoding heavily depends on the choice of the draft model. In this work, we perform a detailed study comprising over 350 experiments with LLaMA-65B and OPT-66B using speculative decoding and delineate the factors that affect the performance gain provided by speculative decoding. Our experiments indicate that the performance of speculative decoding depends heavily on the latency of the draft model, and the draft model's capability in language modeling does not correlate strongly with its performance in speculative decoding. Based on these insights we explore a new design space for draft models and design hardware-efficient draft models for speculative decoding. Our newly designed draft model for LLaMA-65B can provide 60% higher throughput than existing draft models and can generalize further to the LLaMA-2 model family and supervised fine-tuned models.

Modular vision-language models (Vision-LLMs) align pretrained image encoders with (pretrained) large language models (LLMs), representing a computationally much more efficient alternative to end-to-end training of large vision-language models from scratch, which is prohibitively expensive for most. Vision-LLMs instead post-hoc condition LLMs to `understand' the output of an image encoder. With the abundance of readily available high-quality English image-text data as well as monolingual English LLMs, the research focus has been on English-only Vision-LLMs. Multilingual vision-language models are still predominantly obtained via expensive end-to-end pretraining, resulting in comparatively smaller models, trained on limited multilingual image data supplemented with text-only multilingual corpora. In this work, we present mBLIP, the first multilingual Vision-LLM, which we obtain in a computationally efficient manner -- on consumer hardware using only a few million training examples -- by leveraging a pretrained multilingual LLM. To this end, we re-align an image encoder previously tuned to an English LLM to a new, multilingual LLM -- for this, we leverage multilingual data from a mix of vision-and-language tasks, which we obtain by machine-translating high-quality English data to 95 languages. On the IGLUE benchmark, mBLIP yields results competitive with state-of-the-art models. Moreover, in image captioning on XM3600, mBLIP (zero-shot) even outperforms PaLI-X (a model with 55B parameters). Compared to these very large multilingual vision-language models trained from scratch, we obtain mBLIP by training orders of magnitude fewer parameters on magnitudes less data. We release our model and code at https://github.com/gregor-ge/mBLIP.

Assembly code analysis and comprehension play critical roles in applications like reverse engineering, yet they face substantial challenges due to low information density and a lack of explicit syntactic structures. While traditional masked language modeling (MLM) approaches do not explicitly focus on natural language interaction, emerging decoder-focused large language models (LLMs) demonstrate partial success in binary analysis yet remain underexplored for holistic comprehension. We present Assembly Augmented Tuning, an end-to-end structural-semantic instruction tuning framework that synergizes encoder architecture with decoder-based LLMs through a projector module, where the assembly encoder extracts hardware-level structural features, the projector bridges representations with the semantic space, and the instruction-tuned LLM preserves natural language capabilities. Experimental results demonstrate three key advantages: (1) State-of-the-art performance in assembly comprehension with +39.7% Recall@1 and +17.8% MRR improvements over GPT-4-Turbo, (2) Consistent enhancements across base models (24.6-107.4% Recall@1 and 15.2-106.3% MRR on Qwen2.5-Coder, Deepseek-Coder and CodeLlama variants), and (3) Superior instruction-following capabilities (41.5%-118% improvements) with controlled code generation degradation (-8.9% to -35% across architectures).

Interpretable machine learning has exploded as an area of interest over the last decade, sparked by the rise of increasingly large datasets and deep neural networks. Simultaneously, large language models (LLMs) have demonstrated remarkable capabilities across a wide array of tasks, offering a chance to rethink opportunities in interpretable machine learning. Notably, the capability to explain in natural language allows LLMs to expand the scale and complexity of patterns that can be given to a human. However, these new capabilities raise new challenges, such as hallucinated explanations and immense computational costs. In this position paper, we start by reviewing existing methods to evaluate the emerging field of LLM interpretation (both interpreting LLMs and using LLMs for explanation). We contend that, despite their limitations, LLMs hold the opportunity to redefine interpretability with a more ambitious scope across many applications, including in auditing LLMs themselves. We highlight two emerging research priorities for LLM interpretation: using LLMs to directly analyze new datasets and to generate interactive explanations.

Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM's autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. We first construct a structured latent reasoning space using a Variational Autoencoder (VAE) that encodes text reasoning steps into blocks of thought tokens, preserving semantic information and interpretability while offering compact but expressive representations. Subsequently, we utilize a latent diffusion model that learns to denoise a block of latent thought tokens with a blockwise bidirectional attention mask, enabling longer horizon and iterative refinement with adaptive test-time compute. This design allows efficient parallel generation of diverse reasoning trajectories, allowing the model to plan and revise the reasoning process holistically. We conduct evaluations on a suite of mathematical reasoning and planning benchmarks. Empirical results show that LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods, revealing a new paradigm for text reasoning with latent diffusion.

With the help of simple fine-tuning, one can artificially embed hidden text into large language models (LLMs). This text is revealed only when triggered by a specific query to the LLM. Two primary applications are LLM fingerprinting and steganography. In the context of LLM fingerprinting, a unique text identifier (fingerprint) is embedded within the model to verify licensing compliance. In the context of steganography, the LLM serves as a carrier for hidden messages that can be disclosed through a designated trigger. Our work demonstrates that embedding hidden text in the LLM via fine-tuning, though seemingly secure due to the vast number of potential triggers (any sequence of characters or tokens could serve as a trigger), is susceptible to extraction through analysis of the LLM's output decoding process. We propose a novel approach to extraction called Unconditional Token Forcing. It is premised on the hypothesis that iteratively feeding each token from the LLM's vocabulary into the model should reveal sequences with abnormally high token probabilities, indicating potential embedded text candidates. Additionally, our experiments show that when the first token of a hidden fingerprint is used as an input, the LLM not only produces an output sequence with high token probabilities, but also repetitively generates the fingerprint itself. We also present a method to hide text in such a way that it is resistant to Unconditional Token Forcing, which we named Unconditional Token Forcing Confusion.

Large language models (LLMs) have shown promising performance across diverse domains. Many practical applications of LLMs, such as code completion and structured data extraction, require adherence to syntactic constraints specified by a formal language. Yet, due to their probabilistic nature, LLM output is not guaranteed to adhere to such formal languages. Prior work has proposed constrained decoding as a means to restrict LLM generation to particular formal languages. However, existing works are not applicable to the emerging paradigm of diffusion LLMs, when used in practical scenarios such as the generation of formally correct C++ or JSON output. In this paper we address this challenge and present the first constrained decoding method for diffusion models, one that can handle formal languages captured by context-free grammars. We begin by reducing constrained decoding to the more general additive infilling problem, which asks whether a partial output can be completed to a valid word in the target language. This problem also naturally subsumes the previously unaddressed multi-region infilling constrained decoding. We then reduce this problem to the task of deciding whether the intersection of the target language and a regular language is empty and present an efficient algorithm to solve it for context-free languages. Empirical results on various applications, such as C++ code infilling and structured data extraction in JSON, demonstrate that our method achieves near-perfect syntactic correctness while consistently preserving or improving functional correctness. Importantly, our efficiency optimizations ensure that the computational overhead remains practical.

Although large language models (LLMs) have demonstrated their effectiveness in a wide range of applications, they have also been observed to perpetuate unwanted biases present in the training data, potentially leading to harm for marginalized communities. In this paper, we mitigate bias by leveraging small biased and anti-biased expert models to obtain a debiasing signal that will be added to the LLM output at decoding-time. This approach combines resource efficiency with interpretability and can be optimized for mitigating specific types of bias, depending on the target use case. Experiments on mitigating gender, race, and religion biases show a reduction in bias on several local and global bias metrics while preserving language model performance.

Empowering LLMs with the ability to utilize useful information from a long context is crucial for many downstream applications. However, achieving long context lengths with the conventional transformer architecture requires substantial training and inference resources. In this paper, we present FocusLLM, a framework designed to extend the context length of any decoder-only LLM, enabling the model to focus on relevant information from very long sequences. FocusLLM processes long text inputs by dividing them into chunks based on the model's original context length to alleviate the issue of attention distraction. Then, it appends the local context to each chunk as a prompt to extract essential information from each chunk based on a novel parallel decoding mechanism, and ultimately integrates the extracted information into the local context. FocusLLM stands out for great training efficiency and versatility: trained with an 8K input length with much less training cost than previous methods, FocusLLM exhibits superior performance across downstream long-context tasks and maintains strong language modeling ability when handling extensive long texts, even up to 400K tokens. Our code is available at https://github.com/leezythu/FocusLLM.

Multimodal large language models (MLLMs) extend LLMs with visual understanding through a three-stage pipeline: multimodal preprocessing, vision encoding, and LLM inference. While these stages enhance capability, they introduce significant system bottlenecks. First, multimodal preprocessing-especially video decoding-often dominates Time-to-First-Token (TTFT). Most systems rely on CPU-based decoding, which severely limits throughput, while existing GPU-based approaches prioritize throughput-oriented parallelism and fail to meet the latency-sensitive requirements of MLLM inference. Second, the vision encoder is a standalone, compute-intensive stage that produces visual embeddings and cannot be co-batched with LLM prefill or decoding. This heterogeneity forces inter-stage blocking and increases token-generation latency. Even when deployed on separate GPUs, these stages underutilize available compute and memory resources, reducing overall utilization and constraining system throughput. To address these challenges, we present FlashCodec and UnifiedServe, two complementary designs that jointly optimize the end-to-end MLLM pipeline. FlashCodec accelerates the multimodal preprocessing stage through collaborative multi-GPU video decoding, reducing decoding latency while preserving high throughput. UnifiedServe optimizes the vision-to-text and inference stages using a logically decoupled their execution to eliminate inter-stage blocking, yet physically sharing GPU resources to maximize GPU system utilization. By carefully orchestrating execution across stages and minimizing interference, UnifiedServe Together, our proposed framework forms an end-to-end optimized stack that can serve up to 3.0times more requests or enforce 1.5times tighter SLOs, while achieving up to 4.4times higher throughput compared to state-of-the-art systems.

Diffusion Language Models (DLMs) have recently demonstrated remarkable capabilities in natural language processing tasks. However, the potential of Retrieval-Augmented Generation (RAG), which shows great successes for enhancing large language models (LLMs), has not been well explored, due to the fundamental difference between LLM and DLM decoding. To fill this critical gap, we systematically test the performance of DLMs within the RAG framework. Our findings reveal that DLMs coupled with RAG show promising potentials with stronger dependency on contextual information, but suffer from limited generation precision. We identify a key underlying issue: Response Semantic Drift (RSD), where the generated answer progressively deviates from the query's original semantics, leading to low precision content. We trace this problem to the denoising strategies in DLMs, which fail to maintain semantic alignment with the query throughout the iterative denoising process. To address this, we propose Semantic-Preserving REtrieval-Augmented Diffusion (SPREAD), a novel framework that introduces a query-relevance-guided denoising strategy. By actively guiding the denoising trajectory, SPREAD ensures the generation remains anchored to the query's semantics and effectively suppresses drift. Experimental results demonstrate that SPREAD significantly enhances the precision and effectively mitigates RSD of generated answers within the RAG framework.

Video understanding has rapidly evolved toward video large language models (VideoLLMs): systems that couple video representations with pretrained large language models and condition generation on a textual prompt. Their strong performance on captioning, question answering, retrieval and temporal grounding comes at a computation and memory cost that grows with frame count and context length, limiting deployment in real-time, mobile and resource-constrained settings. This survey covers inference-efficiency mechanisms for visual and audiovisual VideoLLMs that report concrete reductions in parameter count, FLOPs per input, latency, memory, or visual and audio token count. We analyze bottlenecks across frame sampling, modality encoding, connector-level token reduction, and LLM prefilling and decoding. We organize methods by the pipeline stage at which they act, covering VideoLLMs developed since late 2022 together with earlier frame-sampling and vision-encoder mechanisms that remain components of current pipelines. We assemble literature-reported accuracy--cost comparisons under shared host models and input protocols wherever available, distinguish them from heterogeneous cross-paper evidence, and identify gaps in audiovisual efficiency and standardized evaluation. We maintain a repository at https://github.com/momentslab/awesome-efficient-videollm.

Large language models (LLMs) have shown extraordinary performance in various language tasks, but high computational requirements hinder their widespread deployment. Speculative decoding, which uses amateur models to predict the generation of expert models, has been proposed as a way to accelerate LLM inference. However, speculative decoding focuses on acceleration instead of making the best use of the token distribution from amateur models. We proposed Speculative Contrastive Decoding (SCD), an accelerated decoding method leveraging the natural contrast between expert and amateur models in speculative decoding. Comprehensive evaluations on four benchmarks show that SCD can achieve similar acceleration factors as speculative decoding while further improving the generation quality as the contrastive decoding. The analysis of token probabilities further demonstrates the compatibility between speculative and contrastive decoding. Overall, SCD provides an effective approach to enhance the decoding quality of LLMs while saving computational resources.

Recent advances with large language models (LLM) illustrate their diverse capabilities. We propose a novel algorithm, staged speculative decoding, to accelerate LLM inference in small-batch, on-device scenarios. We address the low arithmetic intensity of small-batch inference by improving upon previous work in speculative decoding. First, we restructure the speculative batch as a tree, which reduces generation costs and increases the expected tokens per batch. Second, we add a second stage of speculative decoding. Taken together, we reduce single-batch decoding latency by 3.16x with a 762M parameter GPT-2-L model while perfectly preserving output quality.

Existing Multimodal Large Language Models (MLLMs) process a large number of visual tokens, leading to significant computational costs and inefficiency. Previous works generally assume that all visual tokens are necessary in the shallow layers of LLMs, and therefore token compression typically occurs in intermediate layers. In contrast, our study reveals an interesting insight: with proper selection, token compression is feasible at the input stage of LLM with negligible performance loss. Specifically, we reveal that explainability methods can effectively evaluate the importance of each visual token with respect to the given instruction, which can well guide the token compression. Furthermore, we propose to learn a mapping from the attention map of the first LLM layer to the explanation results, thereby avoiding the need for a full inference pass and facilitating practical deployment. Interestingly, this mapping can be learned using a simple and lightweight convolutional network, whose training is efficient and independent of MLLMs. Extensive experiments on 10 image and video benchmarks across three leading MLLMs (Qwen2-VL, LLaVA-OneVision, and VILA1.5) demonstrate the effectiveness of our approach, e.g., pruning 50% visual tokens while retaining more than 96% of the original performance across all benchmarks for all these three MLLMs. It also exhibits strong generalization, even when the number of tokens in inference far exceeds that used in training.

Large language models (LLMs) have demonstrated impressive translation capabilities even without being explicitly trained on parallel data. This remarkable property has led some to believe that parallel data is no longer necessary for building multilingual language models. While some attribute this to the emergent abilities of LLMs due to scale, recent work suggests that it is actually caused by incidental bilingual signals present in the training data. Various methods have been proposed to maximize the utility of parallel data to enhance the multilingual capabilities of multilingual encoder-based and encoder-decoder language models. However, some decoder-based LLMs opt to ignore parallel data instead. In this work, we conduct a systematic study on the impact of adding parallel data on LLMs' multilingual capabilities, focusing specifically on translation and multilingual common-sense reasoning. Through controlled experiments, we demonstrate that parallel data can significantly improve LLMs' multilingual capabilities.

The emergence of accurate open large language models (LLMs) has led to a race towards quantization techniques for such models enabling execution on end-user devices. In this paper, we revisit the problem of "extreme" LLM compression--defined as targeting extremely low bit counts, such as 2 to 3 bits per parameter, from the point of view of classic methods in Multi-Codebook Quantization (MCQ). Our work builds on top of Additive Quantization, a classic algorithm from the MCQ family, and adapts it to the quantization of language models. The resulting algorithm advances the state-of-the-art in LLM compression, outperforming all recently-proposed techniques in terms of accuracy at a given compression budget. For instance, when compressing Llama 2 models to 2 bits per parameter, our algorithm quantizes the 7B model to 6.93 perplexity (a 1.29 improvement relative to the best prior work, and 1.81 points from FP16), the 13B model to 5.70 perplexity (a .36 improvement) and the 70B model to 3.94 perplexity (a .22 improvement) on WikiText2. We release our implementation of Additive Quantization for Language Models AQLM as a baseline to facilitate future research in LLM quantization.