LLM Inference Acceleration

LLM: Large Language Model

Follow this topic Ask about this topic

This field studies how to make large language models answer faster and use less memory and computing power, especially during prompt processing and token-by-token generation. Common approaches include speculative decoding, which drafts and verifies tokens; sparse attention and KV-cache compression for long contexts; low-bit quantization; and adaptive computation that skips or reallocates work. Researchers test these methods on long-context and retrieval benchmarks such as LongBench and RULER, as well as reasoning and coding tasks, measuring both quality and serving metrics such as time to first token and throughput. A central challenge is turning theoretical reductions in computation or memory into real speedups on particular hardware without sacrificing accuracy or changing the model’s output distribution.

Latest papers

1,101 papers202 papers in the last four weeks, up 51% on the four weeks before. 1.2% of all new papers.

Mar 30, 2026

AdaptToken: Entropy-based Adaptive Token Selection for MLLM Long Video Understanding

Long video understanding remains challenging for Multi-modal Large Language Models (MLLMs) due to high memory costs and context-length limits. Prior approaches mitigate this by scoring and selecting frames/tokens within short clips, but they lack a principled mechanism to (i) compare relevance across distant video clips and (ii) stop processing once sufficient evidence has been gathered. We propose AdaptToken, a training-free framework that turns an MLLM's self-uncertainty into a global control signal for long-video token selection. AdaptToken splits a video into groups, extracts cross-modal attention to rank tokens within each group, and uses the model's response entropy to estimate each group's prompt relevance. This entropy signal enables a global token budget allocation across groups and further supports early stopping (AdaptToken-Lite), skipping the remaining groups when the model becomes sufficiently certain. Across four long-video benchmarks (VideoMME, LongVideoBench, LVBench, and MLVU) and multiple base MLLMs (7B-72B), AdaptToken consistently improves accuracy (e.g., +6.7 on average over Qwen2.5-VL 7B) and continues to benefit from extremely long inputs (up to 10K frames), while AdaptToken-Lite reduces inference time by about half with comparable performance. Project page: https://haozheqi.github.io/adapt-token

Haozhe Qi, Kevin Qu, Mahdi Rad +3

Microsoft Spatial AI Lab · EPFL · ETH Zurich

Long-Video Understanding Multimodal Large Language Models LLM Inference Acceleration

Mar 30, 2026

HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention

Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical key for each query through a lightweight indexer, then computing attention only on the selected subset. While the downstream sparse attention itself scales favorably, the indexer must still scan the entire prefix for every query, introducing an per-layer bottleneck that grows prohibitively with context length. We propose HISA (Hierarchical Indexed Sparse Attention), a plug-and-play replacement for the indexer that rewrites the search path from a flat token scan into a two-stage hierarchical procedure: (1) a block-level coarse filtering stage that scores pooled block representations to discard irrelevant regions, followed by (2) a token-level refinement stage that applies the original indexer exclusively within the retained candidate blocks. HISA preserves the identical token-level top-sparse pattern consumed by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves up to speedup at 64K context. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 and GLM-5 with our HISA indexer, without any finetuning. HISA closely matches the original DSA in quality, while substantially outperforming block-sparse baselines.

Yufei Xu, Fanxu Meng, Fan Jiang +11

Xiaohongshu

Self-Attention LLM Inference Acceleration Efficient Attention Sparse Attention Attention Mechanisms

Mar 30, 2026

QCFuse: Query-Centric Cache Fusion for Efficient RAG Inference

Cache fusion accelerates generation process of LLMs equipped with RAG through KV caching and selective token recomputation, thereby reducing computational costs and improving efficiency. However, existing methods primarily rely on local perspectives for token selection and lack global awareness from the user query. Utilizing this global awareness is challenging due to the high cost of obtaining context-aware query representations and the strict pipeline constraints required for efficient attention analysis. Thus, this demonstration introduces QCFuse, an innovative KV cache fusion system centered on the user query. QCFuse leverages semantic summary anchors to enhance query representations and selectively recomputes query-related tokens to improve accuracy, updating tokens based on the attention distribution of the most critical Transformer layer to preserve the high efficiency of the pipeline structure. Evaluations on real-world datasets demonstrate that QCFuse significantly improves the response efficiency of LLMs by 40% while maintaining equivalent accuracy compared to current methods. Additionally, in certain scenarios, QCFuse achieves an attention denoising effect that yields higher response accuracy, demonstrating substantial potential in the optimization of LLM inference.

Jianxin Yan, Zeheng Qian, Wangze Ni +6

Zhejiang University, Hangzhou, China · The University of Sydney, Sydney, Australia · Ant Group, Shanghai, Cina +4

LLM Inference Acceleration KV-Cache Management Retrieval-Augmented Generation

Mar 30, 2026

CSAttention: Centroid-Scoring Attention for Accelerating LLM Inference

Long-context LLMs increasingly rely on extended, reusable prefill prompts for agents and domain Q&A, pushing attention and KV-cache to become the dominant decode-time bottlenecks. While sparse attention reduces computation and transfer costs, it often struggles to maintain accuracy at high sparsity levels due to the inherent distribution shift between Queries and Keys. We propose Centroid-Scoring Attention (CSAttention), a training-free sparse attention method optimized for high-throughput serving of reusable contexts. CSAttention adopts a storage-for-computation strategy tailored to the offline-prefill/online-decode setting: it front-loads computation into a one-time offline prefill phase that can be amortized across multiple queries, while aggressively optimizing per-step decoding latency. Specifically, CSAttention constructs query-centric lookup tables during offline prefill, whose size remains fixed during decoding, and enables online decoding to replace full-context scans with efficient table lookups and GPU-friendly score accumulation. Extensive experiments demonstrate that CSAttention achieves near-identical accuracy to full attention. Under high sparsity (95%) and long-context settings (32K-128K), CSAttention consistently outperforms state-of-the-art sparse attention methods in both model accuracy and inference speed, achieving up to 4.6x inference speedup over the most accurate baseline at a context length of 128K.

Chuxu Song, Zhencan Peng, Jiuqi Wei +1

Department of Computer Science, Rutgers University · Oceanbase, Ant Group.

LLM Inference Acceleration Long-Context Language Model Inference Attention Mechanisms Sparse Attention

Mar 30, 2026

ITQ3_S: High-Fidelity 3-bit LLM Inference via Interleaved Ternary Quantization with Rotation-Domain Smoothing

We present ITQ3_S (Interleaved Ternary Quantization -- Specialized), a novel 3-bit weight quantization format for LLMs integrating TurboQuant (TQ), a rotation-domain strategy based on the Fast Walsh-Hadamard Transform (FWHT). Conventional 3-bit methods suffer precision loss from heavy-tailed weight distributions and inter-channel outliers. ITQ3_S pre-rotates the weight space via FWHT before quantization, spreading outlier energy across the vector and inducing a near-Gaussian distribution amenable to uniform ternary coding. We derive a rigorous dequantization procedure fusing a 256-point Inverse FWHT into the CUDA shared-memory loading stage, ensuring reconstruction error is bounded exclusively by the ternary quantization grid with no additional error from the transform inversion. For any weight vector

\mathbf{w} \in \mathbb{R}^{256}

, the reconstruction satisfies

\|\hat{\mathbf{w}} - \mathbf{w}\|_2 \leq ε_q

, strictly smaller than uniform 3-bit baselines that do not exploit rotation-induced distribution normalization. TurboQuant lacks a native CUDA kernel, precluding direct deployment; naively composing TQ with existing weight quantizers introduces domain mismatch errors that accumulate across layers, degrading quality below standard 3-bit baselines. ITQ3_S resolves this by co-designing the FWHT rotation and quantization kernel as a unified pipeline grounded in the IQ3_S weight format, with the inverse transform fused into the CUDA MMQ kernel. Empirically, on the NVIDIA RTX 5090 (Blackwell), ITQ3_S achieves perplexity competitive with FP16 while delivering throughput exceeding 1.5x that of 4-bit alternatives via optimized DP4A and Tensor Core scheduling. Our results establish ITQ3_S as a practical, mathematically grounded solution for high-fidelity LLM deployment on consumer hardware.

Edward J. Yoon

LLM Quantization Ternary Quantization LLM Inference Acceleration Weight-Only Quantization Rotation-Based Quantization

Mar 29, 2026

Probabilistic Language Tries: A Unified Framework for Compression, Decision Policies, and Execution Reuse

We introduce probabilistic language tries (PLTs), a unified representation that makes explicit the prefix structure implicitly defined by any generative model over sequences. By assigning to each outgoing edge the conditional probability of the corresponding token or action, a PLT simultaneously serves as: (i) an optimal lossless compressor via frequency-weighted interval encoding, generalizing arithmetic coding to model-conditioned distributions; (ii) a policy representation for sequential decision problems including games, search, and robotic control; and (iii) a memoization index that lets repeated inference queries be answered by structured retrieval rather than full model execution. The central technical result is a prior-guided caching theorem: under a stationary generative distribution, a PLT-guided cache achieves strictly lower expected inference cost than any empirical-frequency cache for all query counts below a threshold that grows with the concentration of the prior. This converts O(n^2) transformer attention cost into an expected cost of p_r * O(log N) + (1 - p_r) * O(n^2), where p_r is the prior-estimated reuse probability and N is the artifact store size. We further introduce a hybrid compression architecture decomposing any dataset into a PLT-covered majority and a sparse residual store, connecting arithmetic coding with Kolmogorov-style program representations and rate-distortion theory. We instantiate the framework across chess, web search, robotics, organizational workflows, and LLM inference, demonstrating that compression, decision making, and computational reuse are all derived from a single probability measure on sequence space.

Gregory Magarshak

Lossless Data Compression LLM Inference Acceleration LLM Compression

Mar 29, 2026

KVSculpt: KV Cache Compression as Distillation

KV cache compression is critical for efficient long-context LLM inference. Approaches that reduce the per-pair footprint -- quantization and low-rank decomposition -- are orthogonal to those that reduce the sequence length of the cache. Along the sequence-length dimension, existing methods range from pure eviction -- selecting which KV pairs to keep -- to merging, which combines similar pairs into fewer ones. Both remain anchored to the original cache entries. We propose KVSculpt, which moves to the other end of this spectrum: instead of selecting or combining original pairs, we optimize a smaller set of unconstrained KV pairs in continuous embedding space to preserve each layer's attention behavior. Keys are optimized via L-BFGS and values are solved in closed form via least squares, alternating every few steps. On top of this, we introduce adaptive budget allocation, which uses a cheap pilot compression run to redistribute the compression budget across layers and KV heads based on per-component difficulty. On Qwen2.5-1.5B-Instruct with 2048-token contexts, KVSculpt reduces KL divergence by 3.5-4.1x compared to Select+Fit -- attention-score eviction with least-squares value fitting -- across compression ratios r in {0.3, 0.5, 0.7}. Adaptive allocation provides an additional 1.3x KL reduction at no extra inference cost. Analysis reveals that compression difficulty is highly non-uniform: per-layer pilot MSE varies by up to 100x across layers, and the two KV heads within a single layer can differ by up to 467x -- demonstrating that fine-grained budget allocation is essential.

Bo Jiang, Sian Jin

Temple University

KV-Cache Management KV-Cache Compression LLM Inference Acceleration

Mar 29, 2026

RSR-core: A High-Performance Engine for Low-Bit Matrix-Vector Multiplication

Matrix-vector multiplication is a fundamental building block in neural networks, vector databases, and large language models, particularly during inference. As a result, efficient matrix-vector multiplication engines directly translate into more efficient inference. Recent work has explored low-bit quantization of model weights, where matrices are represented using binary (1-bit) or ternary (1.58-bit) values while activation is kept in higher precision. These representations enable efficient hardware-level computation. In parallel, algorithms such as Redundant Segment Reduction (RSR) provide theoretical guarantees for accelerating low-bit matrix-vector multiplication. However, existing implementations operate at the application level and cannot be efficiently integrated into hardware kernels, limiting practical performance. To bridge this gap, we present RSR-core, a high-performance engine that implements the RSR algorithm as optimized low-level kernels for both CPU and CUDA environments. RSR-core supports efficient matrix-vector multiplication for binary and ternary weight matrices and general vectors while enabling practical deployment of RSR algorithm in real inference pipelines. RSR-core is provided as a production-ready engine with HuggingFace integration for preprocessing low-bit models and running accelerated inference. Experimental results demonstrate significant performance improvements over baseline HuggingFace PyTorch multiplication, achieving up to 62x speedup on CPU and up to 1.9x speedup for token generation on CUDA for popular ternary LLMs. The source code is publicly available at https://github.com/UIC-InDeXLab/RSR-core.

Mohsen Dehghankar, Abolfazl Asudeh

University of Illinois Chicago

LLM Inference Acceleration

Mar 28, 2026

ScoutAttention: Efficient KV Cache Offloading via Layer-Ahead CPU Pre-computation for LLM Inference

Large language models encounter critical GPU memory capacity constraints during long-context inference, where KV cache memory consumption severely limits decode batch sizes. While existing research has explored offloading KV cache to DRAM, these approaches either demand frequent GPU-CPU data transfers or impose extensive CPU computation requirements, resulting in poor GPU utilization as the system waits for I/O operations or CPU processing to complete. We propose ScoutAttention, a novel KV cache offloading framework that accelerates LLM inference through collaborative GPU-CPU attention computation. To prevent CPU computation from bottlenecking the system, ScoutAttention introduces GPU-CPU collaborative block-wise sparse attention that significantly reduces CPU load. Unlike conventional parallel computing approaches, our framework features a novel layer-ahead CPU pre-computation algorithm, enabling the CPU to initiate attention computation one layer in advance, complemented by asynchronous periodic recall mechanisms to maintain minimal CPU compute load. Experimental results demonstrate that ScoutAttention maintains accuracy within 2.4% of baseline while achieving 2.1x speedup compared to existing offloading methods.

Qiuyang Zhang, Kai Zhou, Ding Tang +5

Huazhong University of Science and Technology, Wuhan, China · Huawei Technologies, Hefei, China · Research Center for High Efficiency Computing Infrastructure, Zhejiang Lab, Hangzhou, China

KV-Cache Offloading LLM Inference Acceleration Long-Context Language Model Inference Heterogeneous Computing

Mar 27, 2026

TAPS: Task Aware Proposal Distributions for Speculative Sampling

Speculative decoding accelerates autoregressive generation by letting a lightweight draft model propose future tokens that a larger target model then verifies in parallel. In practice, however, draft models are usually trained on broad generic corpora, which leaves it unclear how much speculative decoding quality depends on the draft training distribution. We study this question with lightweight HASS and EAGLE-2 drafters trained on MathInstruct, ShareGPT, and mixed-data variants, evaluated on MT-Bench, GSM8K, MATH-500, and SVAMP. Measured by acceptance length, task-specific training yields clear specialization: MathInstruct-trained drafts are strongest on reasoning benchmarks, while ShareGPT-trained drafts are strongest on MT-Bench. Mixed-data training improves robustness, but larger mixtures do not dominate across decoding temperatures. We also study how to combine specialized drafters at inference time. Naive checkpoint averaging performs poorly, whereas confidence-based routing improves over single-domain drafts and merged-tree verification yields the highest acceptance length overall for both backbones. Finally, confidence is a more useful routing signal than entropy: rejected tokens tend to have higher entropy, but confidence produces much clearer benchmark-level routing decisions. These results show that speculative decoding quality depends not only on draft architecture, but also on the match between draft training data and downstream workload, and that specialized drafters are better combined at inference time than in weight space.

Mohamad Zbib, Mohamad Bazzi, Ammar Mohanna +2

King Abdullah University of Science and Technology (KAUST) · American University of Beirut (AUB)

Speculative Decoding LLM Inference Acceleration

Mar 27, 2026

Two-dimensional early exit optimisation of LLM inference

We introduce a two-dimensional (2D) early exit strategy that coordinates layer-wise and sentence-wise exiting for classification tasks in large language models. By processing input incrementally sentence-by-sentence while progressively activating deeper layers, our method achieves multiplicative computational savings that exceed those from optimizing either dimension independently. Experimental evaluation across four state-of-the-art LLMs (Llama 3.1, Llama 3.2, Gemma, Qwen; 3B-8B parameters) on three sentiment classification datasets demonstrates additional speed-ups of 1.4--2.3

\times

over optimal layer-wise early exit for simpler tasks with vanilla models, with graceful degradation on complex multi-class problems. Fine-tuning reduces but does not eliminate this advantage. The approach is model-agnostic, requires only lightweight classification adapters, and is orthogonal to complementary efficiency methods such as quantization and pruning. Our findings indicate that 2D early exit strategies excel when semantic information accumulates predictably across input structure, suggesting possible applicability to sequence-processing tasks beyond sentiment classification.

Jan Hůla, David Adamczyk, Tomáš Filip +2

Institute for Research and Applications of Fuzzy Modelling, University of Ostrava, 70200 Ostrava, Czech Republic · Institute of Computer Science, Faculty of Philosophy and Science, Silesian University in Opava, 74601 Opava, Czech Republic

LLM Inference Acceleration Early-Exit Neural Networks

Mar 27, 2026

LLaDA-TTS: Unifying Speech Synthesis and Zero-Shot Editing via Masked Diffusion Modeling

Large language model (LLM)-based text-to-speech (TTS) systems achieve remarkable naturalness via autoregressive (AR) decoding, but require N sequential steps to generate N speech tokens. We present LLaDA-TTS, which replaces the AR LLM with a masked diffusion model that completes generation in a fixed number of parallel steps, decoupling inference latency from sequence length. Remarkably, using only 50 hours of fine-tuning data, we successfully transfer a pretrained AR checkpoint to the masked diffusion paradigm via bidirectional attention. At 64 steps, LLaDA-TTS achieves 0.98% CER (zh) and 1.96% WER (en) on Seed-TTS-Eval, matching the original CosyVoice 3 baseline performance while delivering a 2x LLM-stage speedup--a notable acceleration achieved despite the absence of KV cache, an optimization the AR baseline heavily relies on. Beyond acceleration, the bidirectional architecture naturally enables zero-shot speech editing--including word-level insertion, deletion, and substitution--without any additional training. Theoretically, we prove that AR-pretrained weights are near-optimal for bidirectional masked prediction under the locality property of acoustic tokens, explaining this rapid convergence. This general method modifies only the attention mask and objective, applying seamlessly to any LLM-based AR TTS system. Code and audio samples will be available at https://deft-piroshki-b652b5.netlify.app/.

Xiaoyu Fan, Huizhi Xie, Wei Zou +1

BRVoice Team, Bairong, Inc., China

TTS Synthesis Zero-Shot TTS LLM Inference Acceleration Diffusion Language Models Speech Editing

Mar 27, 2026

TTKV: Temporal-Tiered KV Cache for Long-Context LLM Inference

Key-value (KV) caching is critical for efficient inference in large language models (LLMs), yet its memory footprint scales linearly with context length, resulting in a severe scalability bottleneck. Existing approaches largely treat KV states as equally important across time, implicitly assuming uniform precision and accessibility. However, this assumption contrasts with human memory systems, where memories vary in clarity, recall frequency, and relevance with temporal proximity.Motivated by this insight, we propose TTKV, a KV cache management framework that maps the human memory system onto the KV cache. TTKV partitions the KV cache into temporal tiers with heterogeneous capacity and precision. The design addresses three aspects: (1) Tier Layout, decoupling fast and slow memory using HBM and DRAM; (2) Tier Content, assigning more recent KV states to faster, higher-precision tiers based on temporal proximity; and (3) Tier Interaction, employing block-wise streaming attention to overlap communication and computation when accessing slow tiers. Experiments show that TTKV reduces cross-tier traffic by 5.94x on 128K-context tasks, achieving up to 76% latency reduction and 2x throughput improvement over strong baselines.

Gradwell Dzikanyanga, Weihao Yang, Hao Huang +4

Harbin Institute of Technology, Shenzhen · Guangzhou University

KV-Cache Management Long-Context Language Model Inference LLM Inference Acceleration KV-Cache Offloading

Mar 27, 2026

Accelerating PayPal's Commerce Agent with Speculative Decoding: An Empirical Study on EAGLE3 with Fine-Tuned Nemotron Models

We evaluate speculative decoding with EAGLE3 as an inference-time optimization for PayPal's Commerce Agent, powered by a fine-tuned llama3.1-nemotron-nano-8B-v1 model. Building on prior work (NEMO-4-PAYPAL) that reduced latency and cost through domain-specific fine-tuning, we benchmark EAGLE3 via vLLM against NVIDIA NIM on identical 2xH100 hardware across 40 configurations spanning speculative token counts (gamma=3, gamma=5), concurrency levels (1-32), and sampling temperatures (0, 0.5). Key findings: (1) gamma=3 achieves 22-49% throughput improvement and 18-33% latency reduction at zero additional hardware cost; (2) acceptance rates remain stable at approximately 35.5% for gamma=3 across all conditions; (3) gamma=5 yields diminishing returns (approximately 25% acceptance rate); (4) LLM-as-Judge evaluation confirms fully preserved output quality; and (5) speculative decoding on a single H100 matches or exceeds NIM on two H100s, enabling 50% GPU cost reduction.

Ally Qin, Jian Wan, Sarat Mudunuri +1

LLM Inference LLM Inference Acceleration Speculative Decoding

Mar 26, 2026

GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs

Quantization techniques such as BitsAndBytes, AWQ, and GPTQ are widely used as a standard method in deploying large language models but often degrades accuracy when using low-bit representations, e.g., 4 bits. Low-rank correction methods (e.g., LQER, QERA, ASER) has been proposed to mitigate this issue, however, they restore all layers and insert error-correction modules into every decoder block, which increases latency and memory overhead. To address this limitation, we propose GlowQ, a group-shared low-rank approximation for quantized LLMs that caches a single shared right factor per input-sharing group and restores only the groups or layers that yield the highest accuracy benefit. GlowQ computes the high-precision projection once per input-sharing group and reuses it across its modules, reducing parameter and memory overhead, and retaining the expressivity of layer-specific corrections. We also propose a selective variant, GlowQ-S, that applies the cached shared module only where it provides the largest benefit. Compared with strong baselines, our approach reduces TTFB by (5.6%) and increases throughput by (9.6%) on average, while reducing perplexity on WikiText-2 by (0.17%) and increasing downstream accuracy by 0.42 percentage points. The selective model GlowQ-S further reduces latency, cutting TTFB by (23.4%) and increasing throughput by (37.4%), while maintaining accuracy within 0.2 percentage points on average. Code is available at https://github.com/ahnselim/GlowQ.

Selim An, Il hong Suh, Yeseong Kim

Department of Artificial Intelligence, DGIST, Korea · COGA robotics, Korea · Department of Electrical Engineering, POSTECH, Korea

LLM Quantization LLM Inference Acceleration Low-Bit Quantization

Mar 26, 2026

Large Language Model as Token Compressor and Decompressor

In this paper, we study whether an off-the-shelf LLM can be adapted into a discrete, variable-length token compressor and decompressor for long-context processing. To this end, we design a self-expressive autoencoding framework that fine-tunes a pretrained LLM with lightweight LoRA adapters to map long texts into compact sequences of learned latent codes, termed Z-tokens, and to decode them back into natural language or task outputs. The resulting representation is content-adaptive: less predictable or information-dense segments can receive more Z-tokens, while redundant regions can be represented more compactly through a budget-aware length regularizer. Our method is evaluated on long-context datasets such as Wikipedia, CNN/DailyMail, HotpotQA, and QuALITY, showing that it preserves reconstruction quality and downstream performance while reducing effective context length, generation-stage memory usage, and end-to-end latency. This simple design supports both direct decoding from compressed contexts and autoregressive generation in the Z-token space, providing a practical interface for efficient long-context inference.

Wenbing Li, Yiran Wang, Zikai Song +4

LLM Context Compression LLM Compression LLM Inference Acceleration Long-Context Language Modeling LLM Fine-Tuning

Mar 25, 2026

Prune as You Generate: Online Rollout Pruning for Faster and Better RLVR

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, methods such as GRPO and DAPO suffer from substantial computational cost, since they rely on sampling many rollouts for each prompt. Moreover, in RLVR the relative advantage is often sparse: many samples become nearly all-correct or all-incorrect, yielding low within-group reward variance and thus weak learning signals. In this paper, we introduce arrol (Accelerating RLVR via online Rollout Pruning), an online rollout pruning method that prunes rollouts during generation while explicitly steering the surviving ones more correctness-balanced to enhance learning signals. Specifically, arrol trains a lightweight quality head on-the-fly to predict the success probability of partial rollouts and uses it to make early pruning decisions. The learned quality head can further weigh candidates to improve inference accuracy during test-time scaling. To improve efficiency, we present a system design that prunes rollouts inside the inference engine and re-batches the remaining ones for log-probability computation and policy updates. Across GRPO and DAPO on Qwen-3 and LLaMA-3.2 models (1B-8B), arrol improves average accuracy by +2.30 to +2.99 while achieving up to 1.7x training speedup, and yielding up to +8.33 additional gains in average accuracy in test-time scaling. The code is available at https://github.com/Hsu1023/ARRoL.

Haobo Xu, Sirui Chen, Ruizhong Qiu +5

University of Illinois at Urbana-Champaign · Amazon

RL for Language Model Reasoning Reinforcement Learning with Verifiable Rewards LLM Inference Acceleration

Mar 25, 2026

Self-Distillation for Multi-Token Prediction

As Large Language Models (LLMs) scale up, inference efficiency becomes a critical bottleneck. Multi-Token Prediction (MTP) could accelerate LLM inference by predicting multiple future tokens in parallel. However, existing MTP approaches still face two challenges: limited acceptance rates of MTP heads, and difficulties in jointly training multiple MTP heads. Therefore, we propose MTP-D, a simple yet effective self-distillation method with minimal additional training cost, which boosts MTP head acceptance rates (+7.5%) while maximumly preserving main-head performance. We also introduce a looped extension strategy for MTP-D, enabling effective and economical MTP head extension and further significant inference speedup to 1-head MTP (+220.4%). Moreover, we systematically explore and validate key insights on the distillation strategies and the potential scalability of MTP through extensive experiments on seven benchmarks. These results demonstrate that our MTP-D and looped extension strategy effectively enhance MTP-head performance and inference efficiency, facilitating the practical usage of MTP in LLMs.

Guoliang Zhao, Ruobing Xie, An Wang +3

Large Language Model Department, Tencent

Multi-Token Prediction LLM Inference Acceleration Self-Distillation Language Model Distillation

Mar 24, 2026

The Diminishing Returns of Early-Exit Decoding in Modern LLMs

In Large Language Model (LLM) inference, early-exit refers to stopping computation at an intermediate layer once the prediction is sufficiently confident, thereby reducing latency and cost. However, recent LLMs adopt improved pretraining recipes and architectures that reduce layer redundancy, potentially limiting early-exit opportunities. We re-evaluate layer-wise early-exit in modern LLMs and analyze how intermediate representations evolve during training. We introduce a metric to quantify a model's intrinsic suitability for early-exit and propose a benchmark for researchers to explore the potential early-exit benefits on different models and workloads. Our results show a diminishing trend in early-exit effectiveness across newer model generations. We further find that dense transformers generally offer greater early-exit potential than Mixture-of-Experts and State Space Models. In addition, larger models, particularly those with more than 20 billion parameters, and base pretrained models without specialized tuning tend to exhibit higher early-exit potential.

Rui Wei, Rui Du, Hanfei Yu +4

Stevens Institute of Technology · Northeastern University · Stony Brook University

Early-Exit Neural Networks LLM Inference Acceleration LLM Evaluation

Mar 24, 2026

SpecEyes: Accelerating Agentic Multimodal LLMs via Speculative Perception and Planning

Agentic multimodal large language models (MLLMs) (e.g., OpenAI o3 and Gemini Agentic Vision) achieve remarkable reasoning capabilities through iterative visual tool invocation. However, the cascaded perception, reasoning, and tool-calling loops introduce significant sequential overhead. This overhead, termed agentic depth, incurs prohibitive latency and seriously limits system-level concurrency. To this end, we propose SpecEyes, an agentic-level speculative acceleration framework that breaks this sequential bottleneck. Our key insight is that a lightweight, tool-free MLLM can serve as a speculative planner to predict the execution trajectory, enabling early termination of expensive tool chains without sacrificing accuracy. To regulate this speculative planning, we introduce a cognitive gating mechanism based on answer separability, which quantifies the model's confidence for self-verification without requiring oracle labels. Furthermore, we design a heterogeneous parallel funnel that exploits the stateless concurrency of the small model to mask the stateful serial execution of the large model, maximizing system throughput. Extensive experiments on V* Bench, HR-Bench, and POPE demonstrate that SpecEyes achieves 1.1-3.35x speedup over the agentic baseline while preserving or even improving accuracy (up to +6.7%), thereby boosting serving throughput under concurrent workloads.

Haoyu Huang, Jinfa Huang, Zhongwei Wan +3

Xiamen University, Xiamen, Fujian, China · University of Rochester, Rochester, NY, USA · The Ohio State University, Columbus, OH, USA

Efficient Multimodal Inference LLM Inference Acceleration Multimodal Agents

Mar 24, 2026

Sparser, Faster, Lighter Transformer Language Models

Scaling autoregressive large language models (LLMs) has driven unprecedented progress but comes with vast computational costs. In this work, we tackle these costs by leveraging unstructured sparsity within an LLM's feedforward layers, the components accounting for most of the model parameters and execution FLOPs. To achieve this, we introduce a new sparse packing format and a set of CUDA kernels designed to seamlessly integrate with the optimized execution pipelines of modern GPUs, enabling efficient sparse computation during LLM inference and training. To substantiate our gains, we provide a quantitative study of LLM sparsity, demonstrating that simple L1 regularization can induce over 99% sparsity with negligible impact on downstream performance. When paired with our kernels, we show that these sparsity levels translate into substantial throughput, energy efficiency, and memory usage benefits that increase with model scale. We will release all code and kernels under an open-source license to promote adoption and accelerate research toward establishing sparsity as a practical axis for improving the efficiency and scalability of modern foundation models.

Edoardo Cetin, Stefano Peluchetti, Emilio Castillo +3

Sakana AI · NVIDIA

LLM Inference Acceleration Sparse Neural Networks LLM Compression

Mar 24, 2026

APreQEL: Adaptive Mixed Precision Quantization For Edge LLMs

Today, large language models have demonstrated their strengths in various tasks ranging from reasoning, code generation, and complex problem solving. However, this advancement comes with a high computational cost and memory requirements, making it challenging to deploy these models on edge devices to ensure real-time responses and data privacy. Quantization is one common approach to reducing memory use, but most methods apply it uniformly across all layers. This does not account for the fact that different layers may respond differently to reduced precision. Importantly, memory consumption and computational throughput are not necessarily aligned, further complicating deployment decisions. This paper proposes an adaptive mixed precision quantization mechanism that balances memory, latency, and accuracy in edge deployment under user-defined priorities. This is achieved by analyzing the layer-wise contribution and by inferring how different quantization types behave across the target hardware platform in order to assign the most suitable quantization type to each layer. This integration ensures that layer importance and the overall performance trade-offs are jointly respected in this design. Our work unlocks new configuration designs that uniform quantization cannot achieve, expanding the solution space to efficiently deploy the LLMs on resource-constrained devices.

Meriem Bouzouad, Yuan-Hao Chang, Jalil Boukhobza

Lab-STICC, CNRS UMR 6285 , ENSTA, Institut Polytechnique de Paris, Brest, France · National Taiwan University, Taipei, Taiwan

Hardware-Aware Quantization LLM Quantization Neural Network Quantization Mixed-Precision Quantization On-Device Language Model Inference LLM Inference Acceleration

Mar 24, 2026

EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction

The increasing memory demand of the Key-Value (KV) cache poses a significant bottleneck for Large Language Models (LLMs) in long-context applications. Existing low-rank KV compression methods reduce this footprint by modifying model projections, limiting the flexibility to switch back to standard full-cache inference when sufficient memory is available. In this paper, we propose EchoKV, a flexible KV cache compression framework that supports on-demand transitions from full KV caching to compressed caching. Unlike traditional compression-decompression paradigms, EchoKV utilizes a lightweight network to reconstruct the discarded KV components from a partial subset, exploiting intrinsic inter-layer and intra-layer similarities among attention heads. We further introduce a lightweight two-stage fine-tuning strategy, requiring only a few minutes on a single A100 GPU for a 7B model. Experimental results on LongBench and RULER demonstrate that EchoKV consistently outperforms existing methods across multiple compression ratios and backbone models while preserving the throughput of full-cache inference in short-context scenarios.

Shiyu Ji, Yixuan Wang, Yijun Liu +2

Research Center for Social Computing and Interactive Robotics, Harbin Institute of Technology, China

KV-Cache Compression LLM Inference Acceleration

Mar 24, 2026

Caterpillar of Thoughts: The Optimal Test-Time Algorithm for Large Language Models

Large language models (LLMs) can often produce substantially better outputs when allowed to use additional test-time computation, such as sampling, chain of thought, backtracking, or revising partial solutions. Despite the growing empirical success of such techniques, there is limited theoretical understanding of how inference time computation should be structured, or what constitutes an optimal use of a fixed computation budget. We model test-time computation as an algorithm interacting with a Markov chain: at any point, the algorithm may resume generation from any previously observed state. That is, unlike standard Markov chains where the states are drawn passively, we allow the algorithm to backtrack to any previously observed state of the Markov chain at any time. Many of the existing test-time algorithms, such as Chain-of-Thought (CoT) (Wei et al., 2023), Tree-of-Thoughts (ToT) (Yao et al., 2023), or Best-of-

k

(Brown et al., 2024) could be seen as specific algorithms in this model. We prove that while backtracking can reduce the number of generations exponentially, a very limited form of backtracking is theoretically sufficient. Namely, we show that the optimal algorithm always generates a caterpillar tree. That is, if we remove the leaves of the state tree generated by the optimal algorithm, we obtain a path. Motivated by our characterization of the optimal algorithm, we present Caterpillar of Thoughts (CaT), a new test-time computation algorithm, reducing the number of token/state generations. Our empirical evaluation shows that CaT, compared to ToT, achieves a better success rate while also reducing the number of token generations.

Amir Azarmehr, Soheil Behnezhad, Alma Ghafari

Northeastern University

Inference-Time Optimization Efficient Language Model Reasoning Test-Time Scaling LLM Inference Acceleration

Mar 23, 2026

ROM: Real-time Overthinking Mitigation via Streaming Detection and Intervention

Large reasoning models (LRMs) often suffer from overthinking, continuing with redundant verification and repeated attempts even after reaching a correct solution. This wastes computation and can overturn an initially correct answer. We show that the transition from productive to redundant reasoning is decodable from late-layer hidden states. These representations separate the two phases around first-correct-solution boundaries beyond what position and lexical controls explain, and distinguish correct from incorrect solution endings. Building on this finding, we propose ROM (Real-time Overthinking Mitigation), which uses a lightweight streaming detector (0.07--0.25% of backbone parameters) to monitor a frozen LRM and intervene at well-formed reasoning boundaries without answer extraction or probe decoding. Counterfactual Self-Correction (CSC) augments detector training with wrong

\rightarrow

correct trajectories to preserve useful self-correction. Heads are trained separately for each backbone using shared MATH500 supervision and transfer zero-shot to four unseen benchmarks. Evaluated against ten baselines across five backbones from three families and five benchmarks, ROM

_{\text{CSC}}

reduces response length by 28--77% (45% on average), with significant reductions in all 25 settings and accuracy matching or exceeding vanilla decoding in 22. Measurements on two backbones further show 20--70% lower end-to-end latency with less than 0.5% detector overhead per decoding step. Code is available at https://github.com/SaFo-Lab/ROM.

Xinyan Wang, Xiaogeng Liu, Ming Pei +1

University of Wisconsin–Madison · Johns Hopkins University

Early Stopping for Language Model Reasoning Overthinking in Language Models Efficient Language Model Reasoning Large Reasoning Models LLM Inference Acceleration

Mar 23, 2026

PRISM: Breaking the O(n) Memory Wall in Long-Context LLM Inference via O(1) Photonic Block Selection

Long-context LLM inference is bottlenecked not by compute but by the O(n) memory bandwidth cost of scanning the KV cache at every decode step -- a wall that no amount of arithmetic scaling can break. Recent photonic accelerators have demonstrated impressive throughput for dense attention computation; however, these approaches inherit the same O(n) memory scaling as electronic attention when applied to long contexts. We observe that the real leverage point is the coarse block-selection step: a memory-bound similarity search that determines which KV blocks to fetch. We identify, for the first time, that this task is structurally matched to the photonic broadcast-and-weight paradigm -- the query fans out to all candidates via passive splitting, signatures are quasi-static (matching electro-optic MRR programming), and only rank order matters (relaxing precision to 4-6 bits). Crucially, the photonic advantage grows with context length: as N increases, the electronic scan cost rises linearly while the photonic evaluation remains O(1). We instantiate this insight in PRISM (Photonic Ranking via Inner-product Similarity with Microring weights), a thin-film lithium niobate (TFLN) similarity engine. Hardware-impaired needle-in-a-haystack evaluation on Qwen2.5-7B confirms 100% accuracy from 4K through 64K tokens at k=32, with 16x traffic reduction at 64K context. PRISM achieves a four-order-of-magnitude energy advantage over GPU baselines at practical context lengths (n >= 4K).

Hyoseok Park, Yeonsang Park

Department of Physics, Chungnam National University, Daejeon 34134, Republic of Korea

LLM Inference Acceleration Long-Context LLM Inference Memory-Efficient Inference Photonic Computing KV-Cache Management

Mar 22, 2026

Cross-Family Speculative Decoding for Polish Language Models on AppleSilicon: An Empirical Evaluation of Bielik11B with UAG-Extended MLX-LM

Speculative decoding accelerates LLM inference by using a small draft model to propose k candidate tokens for a target model to verify. While effective for same-tokenizer pairs on high-bandwidth GPUs, its applicability to cross-family pairs with mismatched tokenizers and consumer-grade unified memory remains underexplored. We extend the MLX-LM framework with Universal Assisted Generation (UAG) to enable cross-tokenizer speculative decoding on Apple Silicon. We evaluate Bielik 11B-Instruct (Mistral-based) as the target model, paired with three draft models: Bielik 1.5B (Qwen-based with custom tokenizer), Qwen2.5-1.5B, and Llama 3.2-1B. Experiments on three Polish-language datasets (Wikipedia, pl_alpaca, synthetic) use draft lengths k in {2, 4, 6} to compare naive and context-aware token translation. Results show: (1) context-aware translation consistently improves acceptance rates across all configurations; (2) the Polish-specialized Bielik 1.5B achieves lower acceptance than general-purpose Qwen2.5 and Llama 3.2 drafters; (3) throughput on Apple Silicon is content-dependent, reaching 1.7x speedup for structured text but failing for varied instructions; and (4) verification cost on unified memory does not amortize as theory predicts because both models are memory-bandwidth bound, making sequential drafting expensive relative to batched verification. We propose a hardware-aware speedup formula and characterize conditions for cross-family speculative decoding on Apple Silicon. This is the first systematic evaluation of cross-family speculative decoding for Polish LLMs and the first empirical study of UAG-based decoding on unified memory architectures.

Krzysztof Fonal

Wrocław University of Science and Technology, Poland

Speculative Decoding LLM Inference Acceleration Text Generation Language Modeling

Mar 22, 2026

TIDE: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference

Large language models run every token through every layer, regardless of difficulty. We present TIDE, a post-training system that attaches tiny learned routers at periodic checkpoint layers and, at inference time, selects the earliest layer whose hidden state has converged for each token. TIDE requires no model retraining, works with any HuggingFace causal LM, auto-detects GPU architecture, and supports float32, float16, and bfloat16 through fused CUDA kernels. On an NVIDIA A100 with DeepSeek R1 Distill 8B, TIDE achieves 100% prefill exit rate (5% of tokens exit at layer 11, the remaining at layer 31), reduces prefill latency by 7.2%, and increases single-batch throughput by 6.6%. During autoregressive decoding, 98-99% of tokens exit early while the model correctly solves a multi-step math problem with 95 unique output tokens. On Qwen3 8B (36 layers), throughput improves by 8.1% at batch size 8. Calibration on 2,000 WikiText samples takes under 3 minutes and produces a ~4 MB router checkpoint. The system comprises 1,308 lines of Python and 1,081 lines of CUDA/C++ with 74 passing tests. Code: https://github.com/RightNow-AI/TIDE

Jaber Jaber, Osama Jaber

RightNow AI

LLM Inference LLM Inference Acceleration Layer Skipping in Language Model Decoding

Mar 22, 2026

Compressed-Sensing-Guided, Inference-Aware Structured Reduction for Large Language Models

Large language models deliver strong generative performance but at the cost of massive parameter counts, memory use, and decoding latency. Prior work has shown that pruning and structured sparsity can preserve accuracy under substantial compression, while prompt-compression methods reduce latency by removing redundant input tokens. However, these two directions remain largely separate. Most model-compression methods are static and optimized offline, and they do not exploit the fact that different prompts and decoding steps activate different latent computational pathways. Prompt-compression methods reduce sequence length, but they do not adapt the executed model subnetwork. We propose a unified compressed-sensing-guided framework for dynamic LLM execution. Random measurement operators probe latent model usage, sparse recovery estimates task-conditioned and token-adaptive support sets, and the recovered supports are compiled into hardware-efficient sparse execution paths over blocks, attention heads, channels, and feed-forward substructures. The framework introduces five key contributions: task-conditioned measurements, so different prompts induce different sparse supports; token-adaptive recovery, so active substructures are re-estimated during decoding; formal sample-complexity bounds under restricted isometry or mutual incoherence assumptions; compile-to-hardware constraints that restrict recovery to GPU-efficient structures; and a joint objective that unifies prompt compression with model reduction. Together, these components recast LLM inference as a measurement-and-recovery problem with explicit approximation guarantees and deployment-oriented speedup constraints.

Andrew Kiruluta

UC Berkeley School of Information

LLM Inference Acceleration LLM Pruning Neural Network Pruning Prompt Compression

Mar 21, 2026

Beyond Token Eviction: Mixed-Dimension Budget Allocation for Efficient KV Cache Compression

Key-value (KV) caching is widely used to accelerate transformer inference, but its memory cost grows linearly with input length, limiting long-context deployment. Existing token eviction methods reduce memory by discarding less important tokens, which can be viewed as a coarse form of dimensionality reduction that assigns each token either zero or full dimension. We propose MixedDimKV, a mixed-dimension KV cache compression method that allocates dimensions to tokens at a more granular level, and MixedDimKV-H, which further integrates head-level importance information. Experiments on long-context benchmarks show that MixedDimKV outperforms prior KV cache compression methods that do not rely on head-level importance profiling. When equipped with the same head-level importance information, MixedDimKV-H consistently outperforms HeadKV. Notably, our approach achieves comparable performance to full attention on LongBench with only 6.25% of the KV cache. Furthermore, in the Needle-in-a-Haystack test, our solution maintains 100% accuracy at a 50K context length while using as little as 0.26% of the cache.

Ruijie Miao, Zhiming Wang, Wang Li +4

Peking University · ByteDance Seed

KV-Cache Management KV-Cache Compression LLM Inference Acceleration Long-Context Language Modeling