Daily Papers

Large reasoning models enhanced by reinforcement learning with verifiable rewards have achieved significant performance gains by extending their chain-of-thought. However, this paradigm incurs substantial deployment costs as models often exhibit excessive verbosity on simple queries. Existing efficient reasoning methods relying on explicit length penalties often introduce optimization conflicts and leave the generative mechanisms driving overthinking largely unexamined. In this paper, we identify a phenomenon termed length shift where models increasingly generate unnecessary reasoning on trivial inputs during training. To address this, we introduce Dynamic Outlier Truncation (DOT), a training-time intervention that selectively suppresses redundant tokens. This method targets only the extreme tail of response lengths within fully correct rollout groups while preserving long-horizon reasoning capabilities for complex problems. To complement this intervention and ensure stable convergence, we further incorporate auxiliary KL regularization and predictive dynamic sampling. Experimental results across multiple model scales demonstrate that our approach significantly pushes the efficiency-performance Pareto frontier outward. Notably, on the AIME-24, our method reduces inference token usage by 78% while simultaneously increasing accuracy compared to the initial policy and surpassing state-of-the-art efficient reasoning methods.

Large Language Models (LLMs) have shown remarkable performance on complex reasoning tasks, especially when equipped with long chain-of-thought (CoT) reasoning. However, eliciting long CoT typically requires large-scale reinforcement learning (RL) training, while often leading to overthinking with redundant intermediate steps. To improve learning and reasoning efficiency, while preserving or even enhancing performance, we propose TACLer, a model-tailored curriculum reinforcement learning framework that gradually increases the complexity of the data based on the model's proficiency in multi-stage RL training. TACLer features two core components: (i) tailored curriculum learning that determines what knowledge the model lacks and needs to learn in progressive stages; (ii) a hybrid Thinking/NoThinking reasoning paradigm that balances accuracy and efficiency by enabling or disabling the Thinking mode. Our experiments show that TACLer yields a twofold advantage in learning and reasoning: (i) it reduces computational cost, cutting training compute by over 50% compared to long thinking models and reducing inference token usage by over 42% relative to the base model; and (ii) it improves accuracy by over 9% on the base model, consistently outperforming state-of-the-art Nothinking and Thinking baselines across four math datasets with complex problems.

Multi-turn user-facing agents must infer user intent from incomplete requests, collect missing information through dialogue and tools, and execute valid actions. A training trajectory records this process as an interleaved sequence of user messages, agent responses, tool calls, etc. Synthesizing sufficiently complex trajectory has become a central route to train agents: existing pipelines often increase difficulty by composing multiple user requests into longer tasks, producing write-intensive trajectories that train sequential execution. We argue that a single write decision can itself be difficult when the agent must gather and compare substantial read-tool evidence before its arguments become identifiable, a challenge that write-intensive data alone cannot address. Guided by this insight, we propose WRIT (Write-Read Intensive Trajectory Synthesis), a pipeline for synthesizing multi-turn agent training trajectories along two complexity axes: the number of write decisions in a task and the evidence burden of each individual decision. WRIT first generates write-intensive and read-heavy tasks. It then diversifies user behavior instructions to reflect realistic conversational variation, and finally simulates agent-user interactions in an executable environment to produce complete training trajectories. The resulting data trains agents not only for longer task execution, but also for robust, evidence-grounded decision making under high information load. With only 2K synthesized trajectories, a 4B model trained on WRIT outperforms GPT-5.1 no-think on τ^2-bench and substantially reduces inference-time token usage, showing that compact SFT data can convert part of expensive test-time reasoning into efficient agent behavior.

Vision-Language Models (VLMs) excel across diverse tasks but suffer from high inference costs in time and memory. Token sparsity mitigates inefficiencies in token usage, while neuron sparsity reduces high-dimensional computations, both offering promising solutions to enhance efficiency. Recently, these two sparsity paradigms have evolved largely in parallel, fostering the prevailing assumption that they function independently. However, a fundamental yet underexplored question remains: Do they truly operate in isolation, or is there a deeper underlying interplay that has yet to be uncovered? In this paper, we conduct the first comprehensive investigation into this question. By introducing and analyzing the matching mechanism between Core Neurons and Core Tokens, we found that key neurons and tokens for inference mutually influence and reinforce each other. Building on this insight, we propose CoreMatching, a co-adaptive sparse inference framework, which leverages the synergy between token and neuron sparsity to enhance inference efficiency. Through theoretical analysis and efficiency evaluations, we demonstrate that the proposed method surpasses state-of-the-art baselines on ten image understanding tasks and three hardware devices. Notably, on the NVIDIA Titan Xp, it achieved 5x FLOPs reduction and a 10x overall speedup. Code is released at https://github.com/wangqinsi1/2025-ICML-CoreMatching/tree/main.

Large Language Models (LLMs) demonstrate substantial accuracy gains when augmented with reasoning modes such as chain-of-thought and inference-time scaling. However, reasoning also incurs significant costs in inference latency and token usage, with environmental and financial impacts, which are unnecessary for many simple prompts. We present a semantic router that classifies queries based on their reasoning requirements and selectively applies reasoning only when beneficial. Our approach achieves a 10.2 percentage point improvement in accuracy on the MMLU-Pro benchmark while reducing response latency by 47.1% and token consumption by 48.5% compared to direct inference with vLLM. These results demonstrate that semantic routing offers an effective mechanism for striking a balance between accuracy and efficiency in open-source LLM serving systems

Recent reasoning Large Language Models (LLMs) demonstrate remarkable problem-solving abilities but often generate long thinking traces whose utility is unclear. Our work aims to improve their efficiency, enabling them to reach high performance without overthinking. First, we analyze the entropy of token probabilities in reasoning traces. Across three models, we observe a consistent U-shaped entropy pattern: high entropy on easy problems despite high accuracy, low entropy on problems with medium difficulty, and high entropy on hard problems reflecting uncertainty. Specifically, we notice 22--25\% entropy reduction from easy to medium difficulty regions, suggesting an {overthinking} phenomenon on easy instances. Building on these insights, we introduce DiffAdapt, a lightweight framework that selects Easy/Normal/Hard inference strategies per question based on their difficulty and reasoning trace entropy. Each inference strategy consists of a fixed prompt, temperature and maximum token length. In contrast to existing efficiency optimization methods, our approach does not fine-tune base LLM but a small probe that classifies LLM's final hidden state, allowing inexpensive adaptation. We comprehensively evaluate our method on five models and eight benchmarks. Our method achieves comparable or improved accuracy while reducing token usage by up to 22.4\%, establishing a practical path toward compute-efficient reasoning.

Large Reasoning Models (LRMs) significantly improve the reasoning ability of Large Language Models (LLMs) by learning to reason, exhibiting promising performance in complex task-solving. However, their deliberative reasoning process leads to inefficiencies in token usage, memory consumption, and inference time. Thus, this survey provides a review of efficient inference methods designed specifically for LRMs, focusing on mitigating token inefficiency while preserving the reasoning quality. First, we introduce a taxonomy to group the recent methods into two main categories: (a) explicit compact Chain-of-Thought (CoT), which reduces tokens while keeping the explicit reasoning structure, and (b) implicit latent CoT, which encodes reasoning steps within hidden representations instead of explicit tokens. Meanwhile, we discuss their strengths and weaknesses. Then, we conduct empirical analyses on existing methods from performance and efficiency aspects. Besides, we present open challenges in this field, including human-centric controllable reasoning, trade-off between interpretability and efficiency of reasoning, ensuring safety of efficient reasoning, and broader applications of efficient reasoning. In addition, we highlight key insights for enhancing LRMs' inference efficiency via techniques such as model merging, new architectures, and agent routers. We hope this work serves as a valuable guide, helping researchers overcome challenges in this vibrant fieldhttps://github.com/yueliu1999/Awesome-Efficient-Inference-for-LRMs.

Large Language Models (LLMs) exhibit nonlinear relationships between performance, cost, and token usage. This paper presents a quantitative study on structured prompting using BRAID (Bounded Reasoning for Au tonomous Inference and Decisions) across multiple GPT model tiers, eval uated on the AdvancedIF, GSM-Hard, and the SCALE MultiChallenge benchmark datasets. BRAID introduces a bounded reasoning framework using Mermaid-based instruction graphs that enable models to reason struc turally rather than through unbounded natural-language token expansion. We show that structured machine-readable prompts substantially increase reasoning accuracy and cost efficiency for agents in production systems. The findings establish BRAID as an effective and scalable technique for optimizing inference efficiency in autonomous agent systems. All datasets and detailed result logs are available at https://benchmark.openserv.ai.

Chain-of-Thought (CoT) prompting improves reasoning but often produces long and redundant traces that substantially increase inference cost. We present SyncThink, a training-free and plug-and-play decoding method that reduces CoT overhead without modifying model weights. We find that answer tokens attend weakly to early reasoning and instead focus on the special token "/think", indicating an information bottleneck. Building on this observation, SyncThink monitors the model's own reasoning-transition signal and terminates reasoning. Experiments on GSM8K, MMLU, GPQA, and BBH across three DeepSeek-R1 distilled models show that SyncThink achieves 62.00 percent average Top-1 accuracy using 656 generated tokens and 28.68 s latency, compared to 61.22 percent, 2141 tokens, and 92.01 s for full CoT decoding. On long-horizon tasks such as GPQA, SyncThink can further yield up to +8.1 absolute accuracy by preventing over-thinking.

Large language models (LLMs) excel at complex reasoning, yet their efficiency is limited by the surging cognitive overhead of long thought traces. In this paper, we propose LightThinker, a method that enables LLMs to dynamically compress intermediate thoughts into compact semantic representations. However, static compression often struggles with complex reasoning where the irreversible loss of intermediate details can lead to logical bottlenecks. To address this, we evolve the framework into LightThinker++, introducing Explicit Adaptive Memory Management. This paradigm shifts to behavioral-level management by incorporating explicit memory primitives, supported by a specialized trajectory synthesis pipeline to train purposeful memory scheduling. Extensive experiments demonstrate the framework's versatility across three dimensions. (1) LightThinker reduces peak token usage by 70% and inference time by 26% with minimal accuracy loss. (2) In standard reasoning, LightThinker++ slashes peak token usage by 69.9% while yielding a +2.42% accuracy gain under the same context budget for maximum performance. (3) Most notably, in long-horizon agentic tasks, it maintains a stable footprint beyond 80 rounds (a 60%-70% reduction), achieving an average performance gain of 14.8% across different complex scenarios. Overall, our work provides a scalable direction for sustaining deep LLM reasoning over extended horizons with minimal overhead.

Inference-time scaling techniques have significantly bolstered the reasoning capabilities of large language models (LLMs) by harnessing additional computational effort at inference without retraining. Similarly, Chain-of-Thought (CoT) prompting and its extension, Long CoT, improve accuracy by generating rich intermediate reasoning trajectories, but these approaches incur substantial token costs that impede their deployment in latency-sensitive settings. In this work, we first show that truncated CoT, which stops reasoning before completion and directly generates the final answer, often matches full CoT sampling while using dramatically fewer tokens. Building on this insight, we introduce Fractured Sampling, a unified inference-time strategy that interpolates between full CoT and solution-only sampling along three orthogonal axes: (1) the number of reasoning trajectories, (2) the number of final solutions per trajectory, and (3) the depth at which reasoning traces are truncated. Through extensive experiments on five diverse reasoning benchmarks and several model scales, we demonstrate that Fractured Sampling consistently achieves superior accuracy-cost trade-offs, yielding steep log-linear scaling gains in Pass@k versus token budget. Our analysis reveals how to allocate computation across these dimensions to maximize performance, paving the way for more efficient and scalable LLM reasoning.

Reasoning models improve their problem-solving ability through inference-time scaling, allocating more compute via longer token budgets. Identifying which reasoning traces are likely to succeed remains a key opportunity: reliably predicting productive paths can substantially reduce wasted computation and improve overall efficiency. We introduce Latent-Trajectory signals that characterize the temporal evolution of a model's internal representations during the generation of intermediate reasoning tokens. By measuring the overall change in latent representations between the start and end of reasoning, the change accumulated across intermediate steps, and the extent to which these changes advance toward the final state, we show that these signals predict solution accuracy more reliably than both cross-layer metrics and output-based confidence measures. When used to guide answer selection across multiple sampled generations, Latent-Trajectory signals make test-time scaling more effective and efficient than majority voting, reducing token usage by up to 70% while preserving and even improving accuracy by 2.6% on average. Moreover, these predictive signals often emerge early in the reasoning trace, enabling early selection and allocation of compute to the most promising candidates. Our findings contribute not only practical strategies for inference-time efficiency, but also a deeper interpretability perspective on how reasoning processes are represented and differentiated in latent space.

Chain-of-Thought (CoT) prompting improves LLM accuracy on complex tasks but often increases token usage and inference cost. Existing "Budget Forcing" methods reducing cost via fine-tuning with heuristic length penalties, suppress both essential reasoning and redundant filler. We recast efficient reasoning as a lossy compression problem under the Information Bottleneck (IB) principle, and identify a key theoretical gap when applying naive IB to transformers: attention violates the Markov property between prompt, reasoning trace, and response. To resolve this issue, we model CoT generation under the Conditional Information Bottleneck (CIB) principle, where the reasoning trace Z acts as a computational bridge that contains only the information about the response Y that is not directly accessible from the prompt X. This yields a general Reinforcement Learning objective: maximize task reward while compressing completions under a prior over reasoning traces, subsuming common heuristics (e.g., length penalties) as special cases (e.g., uniform priors). In contrast to naive token-counting-based approaches, we introduce a semantic prior that measures token cost by surprisal under a language model prior. Empirically, our CIB objective prunes cognitive bloat while preserving fluency and logic, improving accuracy at moderate compression and enabling aggressive compression with minimal accuracy drop.

Skill usage has become a core component of modern agent systems and can substantially improve agents' ability to complete complex tasks. In real-world settings, where agents must monitor and interact with numerous personal applications, web browsers, and other environment interfaces, skill libraries can scale to thousands of reusable skills. Scaling to larger skill sets introduces two key challenges. First, loading the full skill set saturates the context window, driving up token costs, hallucination, and latency. In this paper, we present Graph of Skills (GoS), an inference-time structural retrieval layer for large skill libraries. GoS constructs an executable skill graph offline from skill packages, then at inference time retrieves a bounded, dependency-aware skill bundle through hybrid semantic-lexical seeding, reverse-weighted Personalized PageRank, and context-budgeted hydration. On SkillsBench and ALFWorld, GoS improves average reward by 43.6% over the vanilla full skill-loading baseline while reducing input tokens by 37.8%, and generalizes across three model families: Claude Sonnet, GPT-5.2 Codex, and MiniMax. Additional ablation studies across skill libraries ranging from 200 to 2,000 skills further demonstrate that GoS consistently outperforms both vanilla skills loading and simple vector retrieval in balancing reward, token efficiency, and runtime.

Multi-agent systems (MAS) built on large language models are typically organized around roles, pipelines, and turn schedules, while the content that agents pass to one another is often left as unconstrained natural language. However, this free-form communication can rapidly inflate token usage, consume the shared context window, and ultimately affect both system performance and inference cost. We analyze five common inter-agent communication strategies across two MAS topologies, finding that no fixed strategy is universally optimal. Instead, effective inter-agent messages consistently preserve action-centered information needed by downstream agents. Building on this, we propose the PACT (Protocolized Action-state Communication and Transmission), which treats inter-agent communication as a public state-update problem and projects each raw agent output into a compact action-state record before it enters shared history. Across different MAS topologies, PACT consistently improves the performance-cost trade-off, achieving comparable or stronger task performance with substantially fewer tokens. The gains extend to production coding harnesses: PACT lifts OpenHands' resolve rate at -10% tokens-per-resolved, and is resolve-neutral on SWE-agent while halving input tokens. Our code is publicly available at https://github.com/iNLP-Lab/PACT.

Reward models (RMs) play a critical role in aligning large language models (LLMs) with human preferences. Yet in the domain of tool learning, the lack of RMs specifically designed for function-calling tasks has limited progress toward more capable agentic AI. We introduce ToolRM, a family of lightweight generative RMs tailored for general tool-use scenarios. To build these models, we propose a novel pipeline that constructs pairwise preference data using rule-based scoring and multidimensional sampling. This yields ToolPref-Pairwise-30K, a diverse, balanced, and challenging dataset of critique tasks that supports reinforcement learning with verifiable feedback. To evaluate tool-use RMs, we also introduce TRBench_{BFCL}, a benchmark built on the agentic evaluation suite BFCL. Trained on our constructed data, models from the Qwen3-4B/8B series achieve up to 14.28% higher accuracy, substantially outperforming frontier models such as Claude 4 and OpenAI o3 in pairwise reward judgments. Beyond training objectives, ToolRM generalizes to broader critique tasks, including Best-of-N sampling and self-correction. Experiments on ACEBench highlight its effectiveness and efficiency, enabling inference-time scaling and reducing output token usage by over 66%. We release data and model checkpoints to facilitate future research.

Large language models (LLMs) have achieved impressive results in natural language understanding, yet their reasoning capabilities remain limited when operating as single agents. Multi-Agent Debate (MAD) has been proposed to address this limitation by enabling collaborative reasoning among multiple models in a round-table debate manner. While effective, MAD introduces substantial computational overhead due to the number of agents involved and the frequent communication required. In this paper, we propose MARS (Multi-Agent Review System), a role-based collaboration framework inspired by the review process. In MARS, an author agent generates an initial solution, reviewer agents provide decisions and comments independently, and a meta-reviewer integrates the feedback to make the final decision and guide further revision. This design enhances reasoning quality while avoiding costly reviewer-to-reviewer interactions, thereby controlling token consumption and inference time. We compared MARS with both MAD and other state-of-the-art reasoning strategies across multiple benchmarks. Extensive experiments with different LLMs show that MARS matches the accuracy of MAD while reducing both token usage and inference time by approximately 50\%. Code is available at https://github.com/xwang97/MARS.

Despite notable advancements in prompting methods for Large Language Models (LLMs), such as Chain-of-Thought (CoT), existing strategies still suffer from excessive token usage and limited generalisability across diverse reasoning tasks. To address these limitations, we propose an Adaptive Causal Prompting with Sketch-of-Thought (ACPS) framework, which leverages structural causal models to infer the causal effect of a query on its answer and adaptively select an appropriate intervention (i.e., standard front-door and conditional front-door adjustments). This design enables generalisable causal reasoning across heterogeneous tasks without task-specific retraining. By replacing verbose CoT with concise Sketch-of-Thought, ACPS enables efficient reasoning that significantly reduces token usage and inference cost. Extensive experiments on multiple reasoning benchmarks and LLMs demonstrate that ACPS consistently outperforms existing prompting baselines in terms of accuracy, robustness, and computational efficiency.

Recently, long-thought reasoning LLMs, such as OpenAI's O1, adopt extended reasoning processes similar to how humans ponder over complex problems. This reasoning paradigm significantly enhances the model's problem-solving abilities and has achieved promising results. However, long-thought reasoning process leads to a substantial increase in inference time. A pressing challenge is reducing the inference overhead of long-thought LLMs while ensuring accuracy. In this paper, we experimentally demonstrate that long-thought reasoning models struggle to effectively allocate token budgets based on problem difficulty and reasoning redundancies. To address this, we propose Length-Harmonizing Fine-Tuning (O1-Pruner), aiming at minimizing reasoning overhead while maintaining accuracy. This effective fine-tuning method first estimates the LLM's baseline performance through pre-sampling and then uses RL-style fine-tuning to encourage the model to generate shorter reasoning processes under accuracy constraints. This allows the model to achieve efficient reasoning with lower redundancy while maintaining accuracy. Experiments on various mathematical reasoning benchmarks show that O1-Pruner not only significantly reduces inference overhead but also achieves higher accuracy, providing a novel and promising solution to this challenge. Our code is coming soon at https://github.com/StarDewXXX/O1-Pruner

Recursive or looped language models have recently emerged as a new scaling axis by iteratively refining the same model computation over latent states to deepen reasoning. We extend such scaling principle from a single model to multi-agent systems, and ask: Can agent collaboration itself be scaled through recursion? To this end, we introduce RecursiveMAS, a recursive multi-agent framework that casts the entire system as a unified latent-space recursive computation. RecursiveMAS connects heterogeneous agents as a collaboration loop through the lightweight RecursiveLink module, enabling in-distribution latent thoughts generation and cross-agent latent state transfer. To optimize our framework, we develop an inner-outer loop learning algorithm for iterative whole-system co-optimization through shared gradient-based credit assignment across recursion rounds. Theoretical analyses of runtime complexity and learning dynamics establish that RecursiveMAS is more efficient than standard text-based MAS and maintains stable gradients during recursive training. Empirically, we instantiate RecursiveMAS under 4 representative agent collaboration patterns and evaluate across 9 benchmarks spanning mathematics, science, medicine, search, and code generation. In comparison with advanced single/multi-agent and recursive computation baselines, RecursiveMAS consistently delivers an average accuracy improvement of 8.3%, together with 1.2times-2.4times end-to-end inference speedup, and 34.6%-75.6% token usage reduction. Code and Data are provided in https://recursivemas.github.io.

While existing multi-agent systems (MAS) can handle complex problems by enabling collaboration among multiple agents, they are often highly task-specific, relying on manually crafted agent roles and interaction prompts, which leads to increased architectural complexity and limited reusability across tasks. Moreover, most MAS communicate primarily through natural language, making them vulnerable to error accumulation and instability in long-context, multi-stage interactions within internal agent histories. In this work, we propose Agent Primitives, a set of reusable latent building blocks for LLM-based MAS. Inspired by neural network design, where complex models are built from reusable components, we observe that many existing MAS architectures can be decomposed into a small number of recurring internal computation patterns. Based on this observation, we instantiate three primitives: Review, Voting and Selection, and Planning and Execution. All primitives communicate internally via key-value (KV) cache, which improves both robustness and efficiency by mitigating information degradation across multi-stage interactions. To enable automatic system construction, an Organizer agent selects and composes primitives for each query, guided by a lightweight knowledge pool of previously successful configurations, forming a primitive-based MAS. Experiments show that primitives-based MAS improve average accuracy by 12.0-16.5\% over single-agent baselines, reduce token usage and inference latency by approximately 3times-4times compared to text-based MAS, while incurring only 1.3times-1.6times overhead relative to single-agent inference and providing more stable performance across model backbones.

Large Language Models (LLMs) increasingly rely on chain-of-thought (CoT) reasoning to improve accuracy on complex tasks. However, always generating lengthy reasoning traces is inefficient, leading to excessive token usage and higher inference costs. This paper introduces the Hybrid Policy Optimization (i.e., HiPO), a framework for adaptive reasoning control that enables LLMs to selectively decide when to engage in detailed reasoning (Think-on) and when to respond directly (Think-off). Specifically, HiPO combines a hybrid data pipelineproviding paired Think-on and Think-off responseswith a hybrid reinforcement learning reward system that balances accuracy and efficiency while avoiding over-reliance on detailed reasoning. Experiments across mathematics and coding benchmarks demonstrate that HiPO can substantially reduce token length while maintaining or improving accuracy. Finally, we hope HiPO a can be a principled approach for efficient adaptive reasoning, advancing the deployment of reasoning-oriented LLMs in real-world, resource-sensitive settings.

Agentic retrieval improves multi-hop question answering by giving language models autonomy to iteratively gather evidence. Recent work augments these systems with knowledge graphs for structured traversal, but this combination introduces significant cost: expensive graph construction at index time and compounding token usage at inference time. We introduce Graph Agentic Search over Propositions (GRASP), an agentic system that simultaneously optimizes for high accuracy and minimal token usage in multi-hop question answering. Rather than executing a rigid, singular query, GRASP actively coordinates its retrieval strategy by decomposing multi-hop queries into dependency-aware plans. This enables GRASP to dynamically scale the number of sub-agents according to the complexity of the problem. Each sub-agent resolves its single-hop query by exploring a novel three-layer hierarchical graph of entities, propositions, and passages, using the entity layer for targeted traversal and the proposition layer for high-recall passage retrieval via reciprocal-rank voting. We evaluate GRASP on MuSiQue, 2WikiMultihopQA, and HotpotQA under two settings: open-corpus retrieval and extended context reasoning (LongBench). GRASP achieves the highest QA accuracy in the open retrieval setting on MuSiQue and 2Wiki while using 40-50 percent fewer tokens than IRCoT+HippoRAG2. Furthermore, GRASP leads on EM and F1 across all three datasets in the LongBench setting while using 30 percent fewer tokens than the next most accurate method. Finally, we introduce success economy - the amortized token cost per correct answer, weighted by difficulty - and advocate for efficiency-aware evaluation as a standard practice for agentic QA.

Test-time scaling (TTS), which involves dynamic allocation of compute during inference, offers a promising way to improve reasoning in large language models. While existing TTS methods work well, they often rely on long decoding paths or require a large number of samples to be generated, increasing the token usage and inference latency. We observe the surprising fact that for reasoning tasks, shorter traces are much more likely to be correct than longer ones. Motivated by this, we introduce First Finish Search (FFS), a training-free parallel decoding strategy that launches n independent samples and returns as soon as any one completes. We evaluate FFS alongside simple decoding, beam search, majority voting, and budget forcing on four reasoning models (DeepSeek-R1, R1-Distill-Qwen-32B, QwQ-32B and Phi-4-Reasoning-Plus) and across four datasets (AIME24, AIME25-I, AIME25-II and GPQA Diamond). With DeepSeek-R1, FFS achieves 82.23% accuracy on the AIME datasets, a 15% improvement over DeepSeek-R1's standalone accuracy, nearly matching OpenAI's o4-mini performance. Our theoretical analysis explains why stopping at the shortest trace is likely to yield a correct answer and identifies the conditions under which early stopping may be suboptimal. The elegance and simplicity of FFS demonstrate that straightforward TTS strategies can perform remarkably well, revealing the untapped potential of simple approaches at inference time.

Recent Large Reasoning Language Models (LRLMs) employ long chain-of-thought reasoning with complex reflection behaviors, typically signaled by specific trigger words (e.g., "Wait" and "Alternatively") to enhance performance. However, these reflection behaviors can lead to the overthinking problem where the generation of redundant reasoning steps that unnecessarily increase token usage, raise inference costs, and reduce practical utility. In this paper, we propose Certainty-Guided Reflection Suppression (CGRS), a novel method that mitigates overthinking in LRLMs while maintaining reasoning accuracy. CGRS operates by dynamically suppressing the model's generation of reflection triggers when it exhibits high confidence in its current response, thereby preventing redundant reflection cycles without compromising output quality. Our approach is model-agnostic, requires no retraining or architectural modifications, and can be integrated seamlessly with existing autoregressive generation pipelines. Extensive experiments across four reasoning benchmarks (i.e., AIME24, AMC23, MATH500, and GPQA-D) demonstrate CGRS's effectiveness: it reduces token usage by an average of 18.5% to 41.9% while preserving accuracy. It also achieves the optimal balance between length reduction and performance compared to state-of-the-art baselines. These results hold consistently across model architectures (e.g., DeepSeek-R1-Distill series, QwQ-32B, and Qwen3 family) and scales (4B to 32B parameters), highlighting CGRS's practical value for efficient reasoning.

Inference-time scaling has emerged as a powerful alternative to parameter scaling for improving language model performance on complex reasoning tasks. While existing methods have shown strong performance gains under fixed compute budgets, there has been little focus on optimally utilizing that budget during inference. In this work, we introduce A*-decoding, a search-based inference-time strategy that builds on the A* search algorithm to optimally utilize a fixed compute budget by prioritizing high-quality reasoning paths during generation. We frame language model decoding as a structured search in a state space of partial solutions, applying the A* transition model to identify promising continuations guided by an external process supervision signal. In our experiments, A*-decoding reaches the performance levels of strong inference scaling baselines like best-of-N and particle filtering while using up to 3x fewer tokens and 30% fewer PRM passes under equivalent compute budgets. On the MATH500 and AIME 2024 benchmarks, A*-decoding enables Llama-3.2-1B-Instruct to match the performance of the 70x larger Llama-3.1-70B-Instruct, and allows Qwen3-1.7B to reach o1-like reasoning accuracy. These results highlight the power of structured search in decoding, offering an alternative to brute-force sampling or scale-driven gains. Our work demonstrates how thoughtful inference-time strategies can enhance reasoning in SLMs, pointing toward future advances in more efficient and scalable language model deployment.

Long-context inference for Large Language Models (LLMs) is heavily limited by high computational demands. While several existing methods optimize attention computation, they still process the full set of hidden states at each layer, limiting overall efficiency. In this work, we propose SlimInfer, an innovative framework that aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass. Our key insight is an information diffusion phenomenon: As information from critical tokens propagates through layers, it becomes distributed across the entire sequence. This diffusion process suggests that LLMs can maintain their semantic integrity when excessive tokens, even including these critical ones, are pruned in hidden states. Motivated by this, SlimInfer introduces a dynamic fine-grained pruning mechanism that accurately removes redundant tokens of hidden state at intermediate layers. This layer-wise pruning naturally enables an asynchronous KV cache manager that prefetches required token blocks without complex predictors, reducing both memory usage and I/O costs. Extensive experiments show that SlimInfer can achieve up to 2.53times time-to-first-token (TTFT) speedup and 1.88times end-to-end latency reduction for LLaMA3.1-8B-Instruct on a single RTX 4090, without sacrificing performance on LongBench. Our code will be released upon acceptance.

Long-context LLM inference is bottlenecked by the quadratic attention complexity and linear KV cache growth. Prior approaches mitigate this via post-hoc selection or eviction but overlook the root inefficiency: indiscriminate writing to memory. In this paper, we formalize KV cache management as a causal system of three primitives: KV Admission, Selection, and Eviction. We instantiate KV Admission via Write-Gated KV (WG-KV), a lightweight mechanism that learns to predict token utility before cache entry. By filtering out low-utility states early to maintain a compact global cache alongside a sliding local cache, WG-KV reduces memory usage by 46-68% and delivers 3.03-3.70x prefill and 1.85-2.56x decode speedups on Llama and Qwen models, while maintaining compatibility with FlashAttention and Paged-KV systems. These results demonstrate that learning what to write is a principled and practical recipe for efficient long-context inference. Code is available at https://github.com/EMCLab-Sinica/WG-KV.

Since ChatGPT released its API for public use, the number of applications built on top of commercial large language models (LLMs) increase exponentially. One popular usage of such models is leveraging its in-context learning ability and generating responses given user queries leveraging knowledge obtained by retrieval augmentation. One problem of deploying commercial retrieval-augmented LLMs is the cost due to the additionally retrieved context that largely increases the input token size of the LLMs. To mitigate this, we propose a token compression scheme that includes two methods: summarization compression and semantic compression. The first method applies a T5-based model that is fine-tuned by datasets generated using self-instruct containing samples with varying lengths and reduce token size by doing summarization. The second method further compresses the token size by removing words with lower impact on the semantic. In order to adequately evaluate the effectiveness of the proposed methods, we propose and utilize a dataset called Food-Recommendation DB (FRDB) focusing on food recommendation for women around pregnancy period or infants. Our summarization compression can reduce 65% of the retrieval token size with further 0.3% improvement on the accuracy; semantic compression provides a more flexible way to trade-off the token size with performance, for which we can reduce the token size by 20% with only 1.6% of accuracy drop.

High-resolution Vision-Language Models (VLMs) have been widely used in multimodal tasks to enhance accuracy by preserving detailed image information. However, these models often generate excessive visual tokens due to encoding multiple partitions of the input image. Processing these excessive visual tokens is computationally challenging, especially in resource-constrained environments with commodity GPUs. To support high-resolution images while meeting resource constraints, we propose High-Resolution Early Dropping (HiRED), a token-dropping scheme that operates within a fixed token budget before the Large Language Model (LLM) stage. HiRED can be integrated with existing high-resolution VLMs in a plug-and-play manner, as it requires no additional training while still maintaining superior accuracy. We strategically use the vision encoder's attention in the initial layers to assess the visual content of each image partition and allocate the token budget accordingly. Then, using the attention in the final layer, we select the most important visual tokens from each partition within the allocated budget, dropping the rest. Empirically, when applied to LLaVA-Next-7B on NVIDIA TESLA P40 GPU, HiRED with a 20% token budget increases token generation throughput by 4.7, reduces first-token generation latency by 15 seconds, and saves 2.3 GB of GPU memory for a single inference.

Large language models (LLMs) are known for their exceptional performance across a range of natural language processing tasks, but their deployment comes at a high computational and financial cost. On the other hand, smaller language models (SLMs), which can be deployed on lower-cost edge devices, struggle to match the performance of their larger counterparts. This paper presents a novel hybrid inference approach that leverages the strengths of both model types while minimizing reliance on costly cloud-based LLMs. Unlike existing methods that route entire queries to either an SLM or a cloud LLM, our approach introduces a reward-based mechanism to dynamically determine the involvement of the cloud LLM during token generation. Specifically, each token predicted by the SLM is evaluated against a reward score, and only when this score falls below a certain threshold is the cloud LLM consulted for assistance in the next token prediction. This method not only reduces the traffic to the cloud LLM, thereby lowering costs, but also allows for flexible control over response quality depending on the reward score threshold. Experimental results demonstrate that our approach significantly reduces cloud LLM usage with minimal impact on overall response quality, offering a cost-effective solution for deploying high-performance language models

Many LLM tasks are performed in large batches or even offline, and the performance indictor for which is throughput. These tasks usually show the characteristic of prefix sharing, where different prompt input can partially show the common prefix. However, the existing LLM inference engines tend to optimize the streaming requests and show limitations of supporting the large batched tasks with the prefix sharing characteristic. The existing solutions use the LRU-based cache to reuse the KV context of common prefix. The KV context that is about to be reused may prematurely be evicted with the implicit cache management. Even if not evicted, the lifetime of the shared KV context is extended since requests sharing the same context are not scheduled together, resulting in larger memory usage. These streaming oriented systems schedule the requests in the first-come-first-serve or similar order. As a result, the requests with larger ratio of decoding steps may be scheduled too late to be able to mix with the prefill chunks to increase the hardware utilization. Besides, the token and request number based batching can limit the size of token-batch, which keeps the GPU from saturating for the iterations dominated by decoding tokens. We propose BatchLLM to address the above problems. BatchLLM explicitly identifies the common prefixes globally. The requests sharing the same prefix will be scheduled together to reuse the KV context the best, which also shrinks the lifetime of common KV memory. BatchLLM reorders the requests and schedules the requests with larger ratio of decoding first to better mix the decoding tokens with the latter prefill chunks and applies memory-centric token batching to enlarge the token-batch sizes, which helps to increase the GPU utilization. Extensive evaluation shows that BatchLLM outperforms vLLM by 1.1x to 2x on a set of microbenchmarks and two typical industry workloads.

Multimodal large language models (MLLMs) have recently demonstrated strong capabilities in understanding and generating responses from diverse visual inputs, including high-resolution images and long video sequences. As these models scale to richer visual representations, inference increasingly relies on storing large numbers of vision tokens in the key-value (KV) cache, making memory consumption a central bottleneck. Existing methods address this issue by identifying redundancy in vision tokens and compressing the cache, but such compression is typically applied only after all inputs are processed, resulting in high peak memory usage during the prefill stage. In this work, we show that MLLMs exhibit inherent structural regularities and representational redundancy that can be exploited to control memory growth throughout inference. Based on this insight, we propose a sequential input-compression mechanism that enforces a fixed memory budget by performing structure-aware key-value cache compression during the prefill process. This approach substantially reduces peak memory usage while maintaining generative performance with only minimal degradation, enabling more practical and memory-efficient multimodal inference.

Optimizing the Key-Value (KV) cache of the Large Language Model (LLM) has been considered critical to saving the cost of inference. Most of the existing KV-cache compression algorithms attempted to sparsify the sequence of tokens by taking advantage of the different importance of tokens. In this work, we found that by identifying the importance of attention layers, we could optimize the KV-cache jointly from two dimensions. Based on our observations regarding layer-wise importance in inference, we propose SqueezeAttention to precisely optimize the allocation of KV-cache budget among layers on-the-fly and then incorporate three representative token sparsification algorithms to compress the KV-cache for each layer with its very own budget. By optimizing the KV-cache from both sequence's and layer's dimensions, SqueezeAttention achieves around 30% to 70% of the memory reductions and up to 2.2 times of throughput improvements in a wide range of LLMs and benchmarks. The code is available at https://github.com/hetailang/SqueezeAttention.

Large Vision-Language Models (LVLMs) have recently demonstrated strong multimodal understanding, yet their fine-grained visual perception is often constrained by low input resolutions. A common remedy is to partition high-resolution images into multiple sub-images for separate encoding, but this approach drastically inflates the number of visual tokens and introduces prohibitive inference overhead. To overcome this challenge, we propose Pyramid Token Pruning (PTP), a training-free strategy that hierarchically integrates bottom-up visual saliency at both region and token levels with top-down instruction-guided relevance. Inspired by human visual cognition, PTP selectively preserves more tokens from salient regions while further emphasizing those most relevant to task instructions. Extensive experiments on 13 diverse benchmarks show that PTP substantially reduces computational cost, memory usage, and inference latency, with negligible performance degradation.

Large Language Models (LLMs) have demonstrated remarkable capabilities in handling long context inputs, but this comes at the cost of increased computational resources and latency. Our research introduces a novel approach for the long context bottleneck to accelerate LLM inference and reduce GPU memory consumption. Our research demonstrates that LLMs can identify relevant tokens in the early layers before generating answers to a query. Leveraging this insight, we propose an algorithm that uses early layers of an LLM as filters to select and compress input tokens, significantly reducing the context length for subsequent processing. Our method, GemFilter, demonstrates substantial improvements in both speed and memory efficiency compared to existing techniques, such as standard attention and SnapKV/H2O. Notably, it achieves a 2.4times speedup and 30\% reduction in GPU memory usage compared to SOTA methods. Evaluation on the Needle in a Haystack task shows that GemFilter significantly outperforms standard attention, SnapKV and demonstrates comparable performance on the LongBench challenge. GemFilter is simple, training-free, and broadly applicable across different LLMs. Crucially, it provides interpretability by allowing humans to inspect the selected input sequence. These findings not only offer practical benefits for LLM deployment, but also enhance our understanding of LLM internal mechanisms, paving the way for further optimizations in LLM design and inference. Our code is available at https://github.com/SalesforceAIResearch/GemFilter.

A method is presented for accelerating inference in transformer language models by exploiting the low effective rank of the token activation manifold at each layer. The method decomposes each activation vector into a subspace component and a residual, computes the linear-layer output on the subspace component via a cached low-rank weight image at reduced memory bandwidth, and applies a per-token gate that determines whether the residual correction is computed or skipped. The gate ensures that the output distribution is preserved to within a controllable tolerance. Validation on three model families (GPT-2 124M, GPT-J 6B, OPT 6.7B) on AMD MI300X demonstrates effective speedups of 3.0x to 10.5x on linear-layer weight reads with perplexity ratios below 1.00 and top-1 token agreement above 98%. The method requires no retraining, no architectural modification, and no approximation of the attention mechanism. At the operating point (k = 256, ε = 0.05) on GPT-J 14 6B, the accelerated model produces character-for-character identical output to the baseline.

Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.

In Transformer architectures, tokens\textemdash discrete units derived from raw data\textemdash are formed by segmenting inputs into fixed-length chunks. Each token is then mapped to an embedding, enabling parallel attention computations while preserving the input's essential information. Due to the quadratic computational complexity of transformer self-attention mechanisms, token reduction has primarily been used as an efficiency strategy. This is especially true in single vision and language domains, where it helps balance computational costs, memory usage, and inference latency. Despite these advances, this paper argues that token reduction should transcend its traditional efficiency-oriented role in the era of large generative models. Instead, we position it as a fundamental principle in generative modeling, critically influencing both model architecture and broader applications. Specifically, we contend that across vision, language, and multimodal systems, token reduction can: (i) facilitate deeper multimodal integration and alignment, (ii) mitigate "overthinking" and hallucinations, (iii) maintain coherence over long inputs, and (iv) enhance training stability, etc. We reframe token reduction as more than an efficiency measure. By doing so, we outline promising future directions, including algorithm design, reinforcement learning-guided token reduction, token optimization for in-context learning, and broader ML and scientific domains. We highlight its potential to drive new model architectures and learning strategies that improve robustness, increase interpretability, and better align with the objectives of generative modeling.

Optimizing inference for long-context Large Language Models (LLMs) is increasingly important due to the quadratic compute and linear memory complexity of Transformers. Existing approximation methods, such as key-value (KV) cache dropping, sparse attention, and prompt compression, typically rely on rough predictions of token or KV pair importance. We propose a novel framework for approximate LLM inference that leverages small draft models to more accurately predict the importance of tokens and KV pairs. Specifically, we introduce two instantiations of our proposed framework: (i) SpecKV, which leverages a draft output to accurately assess the importance of each KV pair for more effective KV cache dropping, and (ii) SpecPC, which uses the draft model's attention activations to identify and discard unimportant prompt tokens. To the best of our knowledge, this is the first work to use draft models for approximate LLM inference acceleration, extending their utility beyond traditional lossless speculative decoding. We motivate our methods with theoretical and empirical analyses, and show a strong correlation between the attention patterns of draft and target models. Extensive experiments on long-context benchmarks show that our methods consistently achieve higher accuracy than existing baselines, while preserving the same improvements in memory usage, latency, and throughput. Our code is available at https://github.com/furiosa-ai/draft-based-approx-llm.

Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.

Visual Language Models require substantial computational resources for inference due to the additional input tokens needed to represent visual information. However, these visual tokens often contain redundant and unimportant information, resulting in an unnecessarily high number of tokens. To address this, we introduce PACT, a method that reduces inference time and memory usage by pruning irrelevant tokens and merging visually redundant ones at an early layer of the language model. Our approach uses a novel importance metric to identify unimportant tokens without relying on attention scores, making it compatible with FlashAttention. We also propose a novel clustering algorithm, called Distance Bounded Density Peak Clustering, which efficiently clusters visual tokens while constraining the distances between elements within a cluster by a predefined threshold. We demonstrate the effectiveness of PACT through extensive experiments.

While Large Language Models (LLMs) have achieved remarkable success in various fields, the efficiency of training and inference remains a major challenge. To address this issue, we propose SUBLLM, short for Subsampling-Upsampling-Bypass Large Language Model, an innovative architecture that extends the core decoder-only framework by incorporating subsampling, upsampling, and bypass modules. The subsampling modules are responsible for shortening the sequence, while the upsampling modules restore the sequence length, and the bypass modules enhance convergence. In comparison to LLaMA, the proposed SUBLLM exhibits significant enhancements in both training and inference speeds as well as memory usage, while maintaining competitive few-shot performance. During training, SUBLLM increases speeds by 26% and cuts memory by 10GB per GPU. In inference, it boosts speeds by up to 37% and reduces memory by 1GB per GPU. The training and inference speeds can be enhanced by 34% and 52% respectively when the context window is expanded to 8192. Our code is available at https://github.com/XiaoMi/subllm.

This paper introduces VideoScan, an efficient vision-language model (VLM) inference framework designed for real-time video interaction that effectively comprehends and retains streamed video inputs while delivering rapid and accurate responses. A longstanding challenge in video understanding--particularly for long-term or real-time applications--stems from the substantial computational overhead caused by the extensive length of visual tokens. To address this, VideoScan employs a single semantic carrier token to represent each frame, progressively reducing computational and memory overhead during its two-phase inference process: prefilling and decoding. The embedding of the semantic carrier token is derived from an optimized aggregation of frame-level visual features, ensuring compact yet semantically rich representations. Critically, the corresponding key-value pairs are trained to retain contextual semantics from prior frames, enabling efficient memory management without sacrificing temporal coherence. During inference, the visual tokens of each frame are processed only once during the prefilling phase and subsequently discarded in the decoding stage, eliminating redundant computations. This design ensures efficient VLM inference even under stringent real-time constraints. Comprehensive experiments on diverse offline and online benchmarks demonstrate that LLaVA-Video, supported by our method, achieves up to sim 5times and 1.29times speedups compared to its original version and previous efficient streaming video understanding approaches, respectively. Crucially, these improvements are attained while maintaining competitive performance and ensuring stable GPU memory consumption (consistently sim 18GB, independent of video duration).

Video large language models (VLLMs) have significantly advanced recently in processing complex video content, yet their inference efficiency remains constrained because of the high computational cost stemming from the thousands of visual tokens generated from the video inputs. We empirically observe that, unlike single image inputs, VLLMs typically attend visual tokens from different frames at different decoding iterations, making a one-shot pruning strategy prone to removing important tokens by mistake. Motivated by this, we present DyCoke, a training-free token compression method to optimize token representation and accelerate VLLMs. DyCoke incorporates a plug-and-play temporal compression module to minimize temporal redundancy by merging redundant tokens across frames, and applies dynamic KV cache reduction to prune spatially redundant tokens selectively. It ensures high-quality inference by dynamically retaining the critical tokens at each decoding step. Extensive experimental results demonstrate that DyCoke can outperform the prior SoTA counterparts, achieving 1.5X inference speedup, 1.4X memory reduction against the baseline VLLM, while still improving the performance, with no training.

Large Language Models (LLMs) have achieved remarkable progress across reasoning, generation, and decision-making tasks, yet deploying them on mobile, embedded, and edge devices remains particularly challenging. On-device LLM inference is heavily constrained by the memory and bandwidth overhead of the key-value (KV) cache, which grows linearly with context length and often dominates decoding cost. Existing KV-cache quantization schemes typically rely on fixed precision or hand-crafted heuristics, thereby wasting bits on low-impact tokens while over-compressing informative ones, leading to avoidable accuracy degradation. Inspired by Huffman coding's principle of variable-length allocation, we propose adaptive KV-cache quantization, a learned policy that assigns bit-width proportional to token importance, minimizing expected memory and latency without sacrificing competitive accuracy. Our framework extracts lightweight token-level features, including token frequency, quality score, attention variance, and entropy-based uncertainty, and feeds them into a compact data-driven controller that dynamically selects KV precision from {2-bit, 4-bit, 8-bit, FP16} during decoding. This adaptive precision policy reduces KV memory footprint and latency while improving accuracy compared to static KV quantization and rule-based baselines, and maintaining competitive accuracy close to FP16 inference across standard LLM benchmarks. Extensive experiments across multiple commonsense reasoning benchmarks using SmolLM-135M, SmolLM-360M, and SmolLM-1.7B demonstrate that our controller consistently improves the accuracy-latency trade-off. For instance, with SmolLM-360M on HellaSwag, our method reduces decoding latency (ms/token) by 17.75% relative to static KV quantization, improves accuracy by 7.60 points, and remains within only 0.30 points of FP16 inference.

As large language models increasingly gain popularity in real-world applications, processing extremely long contexts, often exceeding the model's pre-trained context limits, has emerged as a critical challenge. While existing approaches to efficient long-context processing show promise, recurrent compression-based methods struggle with information preservation, whereas random access approaches require substantial memory resources. We introduce REFORM, a novel inference framework that efficiently handles long contexts through a two-phase approach. First, it incrementally processes input chunks while maintaining a compressed KV cache, constructs cross-layer context embeddings, and utilizes early exit strategy for improved efficiency. Second, it identifies and gathers essential tokens via similarity matching and selectively recomputes the KV cache. Compared to baselines, REFORM achieves over 50% and 27% performance gains on RULER and BABILong respectively at 1M context length. It also outperforms baselines on Infinite-Bench and MM-NIAH, demonstrating flexibility across diverse tasks and domains. Additionally, REFORM reduces inference time by 30% and peak memory usage by 5%, achieving both efficiency and superior performance.

Generative recommendation has recently emerged as a transformative paradigm that directly generates target items, surpassing traditional cascaded approaches. It typically involves two components: a tokenizer that learns item identifiers and a recommender trained on them. Existing methods often decouple tokenization from recommendation or rely on asynchronous alternating optimization, limiting full end-to-end alignment. To address this, we unify the tokenizer and recommender under the ultimate recommendation objective via differentiable soft item identifiers, enabling joint end-to-end training. However, this introduces three challenges: training-inference discrepancy due to soft-to-hard mismatch, item identifier collapse from codeword usage imbalance, and collaborative signal deficiency due to an overemphasis on fine-grained token-level semantics. To tackle these challenges, we propose UniGRec, a unified generative recommendation framework that addresses them from three perspectives. UniGRec employs Annealed Inference Alignment during tokenization to smoothly bridge soft training and hard inference, a Codeword Uniformity Regularization to prevent identifier collapse and encourage codebook diversity, and a Dual Collaborative Distillation mechanism that distills collaborative priors from a lightweight teacher model to jointly guide both the tokenizer and the recommender. Extensive experiments on real-world datasets demonstrate that UniGRec consistently outperforms state-of-the-art baseline methods. Our codes are available at https://github.com/Jialei-03/UniGRec.

Current Vision-Language-Action (VLA) models rely on fixed computational depth, expending the same amount of compute on simple adjustments and complex multi-step manipulation. While Chain-of-Thought (CoT) prompting enables variable computation, it scales memory linearly and is ill-suited for continuous action spaces. We introduce Recurrent-Depth VLA (RD-VLA), an architecture that achieves computational adaptivity via latent iterative refinement rather than explicit token generation. RD-VLA employs a recurrent, weight-tied action head that supports arbitrary inference depth with a constant memory footprint. The model is trained using truncated backpropagation through time (TBPTT) to efficiently supervise the refinement process. At inference, RD-VLA dynamically allocates compute using an adaptive stopping criterion based on latent convergence. Experiments on challenging manipulation tasks show that recurrent depth is critical: tasks that fail entirely (0 percent success) with single-iteration inference exceed 90 percent success with four iterations, while simpler tasks saturate rapidly. RD-VLA provides a scalable path to test-time compute in robotics, replacing token-based reasoning with latent reasoning to achieve constant memory usage and up to 80x inference speedup over prior reasoning-based VLA models. Project page: https://rd-vla.github.io/

Contrastive Language-Image Pretraining (CLIP) excels at learning generalizable image representations but often falls short in zero-shot inference on certain downstream datasets. Test-time adaptation (TTA) mitigates this issue by adjusting components like normalization layers or context prompts, yet it typically requires large batch sizes and extensive augmentations, leading to high computational costs. This raises a key question: Can VLMs' performance drop in specific test cases be mitigated through efficient, training-free approaches? To explore the solution, we investigate token condensation (TC) techniques, originally designed to enhance vision transformer efficiency by refining token usage during inference. We observe that informative tokens improve visual-text alignment in VLMs like CLIP on unseen datasets. However, existing TC methods often fail to maintain in-distribution performance when reducing tokens, prompting us to ask: How can we transform TC into an effective ``free-lunch'' adaptation strategy for VLMs? To address this, we propose Token Condensation as Adaptation (TCA), a training-free adaptation method that takes a step beyond standard TC. Rather than passively discarding tokens, TCA condenses token representation by introducing reservoir-based domain anchor tokens for information-preserving token reduction and logits correction. TCA achieves up to a 21.4% performance improvement over the strongest baseline on cross-dataset benchmark and the CIFAR-100-Corrupted dataset while reducing GFLOPs by 12.2% to 48.9%, with minimal hyperparameter dependency on both CLIP and SigLIP series.

Large Language Models (LLMs) have shown strong capabilities in document re-ranking, a key component in modern Information Retrieval (IR) systems. However, existing LLM-based approaches face notable limitations, including ranking uncertainty, unstable top-k recovery, and high token cost due to token-intensive prompting. To effectively address these limitations, we propose REALM, an uncertainty-aware re-ranking framework that models LLM-derived relevance as Gaussian distributions and refines them through recursive Bayesian updates. By explicitly capturing uncertainty and minimizing redundant queries, REALM achieves better rankings more efficiently. Experimental results demonstrate that our REALM surpasses state-of-the-art re-rankers while significantly reducing token usage and latency, improving NDCG@10 by 0.7-11.9 and simultaneously reducing the number of LLM inferences by 23.4-84.4%, promoting it as the next-generation re-ranker for modern IR systems.

Reinforcement Learning with Verifiable Reward (RLVR) effectively solves complex tasks but demands extremely long context lengths during training, leading to substantial computational costs. While multi-stage training can partially mitigate this, starting with overly short contexts often causes irreversible performance degradation, ultimately failing to reduce overall training compute significantly. In this paper, we introduce **T**hinking-**F**ree **P**olicy **I**nitialization (**TFPI**), a simple yet effective adaptation to RLVR that bridges long Chain-of-Thought (CoT) distillation and standard RLVR. TFPI employs a simple *ThinkFree* operation, explicitly discarding the thinking content via a direct *</think>* append, to reduce token usage during inference. Training with *ThinkFree*-adapted inputs improves performance and lowers token consumption, even in the original slow-thinking mode. Extensive experiments across various benchmarks have shown that TFPI accelerates RL convergence, achieves a higher performance ceiling, and yields more token-efficient reasoning models without specialized rewards or complex training designs. With TFPI only, we train a 4B model to reach 89.0% accuracy on AIME24 and 65.5% on LiveCodeBench using less than 4K H20 hours.

Multi-agent systems (MAS) extend large language models (LLMs) from independent single-model reasoning to coordinative system-level intelligence. While existing LLM agents depend on text-based mediation for reasoning and communication, we take a step forward by enabling models to collaborate directly within the continuous latent space. We introduce LatentMAS, an end-to-end training-free framework that enables pure latent collaboration among LLM agents. In LatentMAS, each agent first performs auto-regressive latent thoughts generation through last-layer hidden embeddings. A shared latent working memory then preserves and transfers each agent's internal representations, ensuring lossless information exchange. We provide theoretical analyses establishing that LatentMAS attains higher expressiveness and lossless information preservation with substantially lower complexity than vanilla text-based MAS. In addition, empirical evaluations across 9 comprehensive benchmarks spanning math and science reasoning, commonsense understanding, and code generation show that LatentMAS consistently outperforms strong single-model and text-based MAS baselines, achieving up to 14.6% higher accuracy, reducing output token usage by 70.8%-83.7%, and providing 4x-4.3x faster end-to-end inference. These results demonstrate that our new latent collaboration framework enhances system-level reasoning quality while offering substantial efficiency gains without any additional training. Code and data are fully open-sourced at https://github.com/Gen-Verse/LatentMAS.

LLM inference is still evaluated mainly as a model or software problem: accuracy, latency, throughput, and hardware utilization. This is incomplete. At deployment scale, the relevant output is a quality-conditioned token produced under joint constraints from effective compute, delivered data-center power, cooling capacity, PUE, and utilization. We argue that the ML community should treat inference as energy-to-token production. We formalize this view with a dimensionally consistent Token Production Function in which token rate is bounded by both compute-per-token and energy-per-token ceilings. Listed API prices vary by over an order of magnitude across providers, but we use price dispersion only as directional motivation, not as causal evidence of marginal cost. The core physical question is instead: under fixed quality and service targets, when does the binding constraint move from theoretical peak compute toward delivered power, cooling, and operational efficiency? Under this framing, system optimizations -- latent KV-cache compression, sparse or heavily compressed attention, quantization, routing, and difficulty-adaptive reasoning -- are not merely local engineering tricks. They are energy-to-token levers because they reduce FLOPs/token, joules/token, memory traffic, or utilization losses under fixed (q^{*},s^{*}). We therefore call for inference papers and benchmarks to report Joules/token, active binding constraint, PUE-adjusted delivered power, and utilization-adjusted token output alongside accuracy and latency.

In recent years, large language models have demonstrated remarkable performance across various natural language processing (NLP) tasks. However, deploying these models for real-world applications often requires efficient inference solutions to handle the computational demands. In this paper, we explore the utilization of CPUs for accelerating the inference of large language models. Specifically, we introduce a parallelized approach to enhance throughput by 1) Exploiting the parallel processing capabilities of modern CPU architectures, 2) Batching the inference request. Our evaluation shows the accelerated inference engine gives an 18-22x improvement in the generated token per sec. The improvement is more with longer sequence and larger models. In addition to this, we can also run multiple workers in the same machine with NUMA node isolation to further improvement in tokens/s. Table 2, we have received 4x additional improvement with 4 workers. This would also make Gen-AI based products and companies environment friendly, our estimates shows that CPU usage for Inference could reduce the power consumption of LLMs by 48.9% while providing production ready throughput and latency.

Large Language Models (LLMs) have demonstrated the ability to tackle increasingly complex tasks through advanced reasoning, long-form content generation, and tool use. Solving these tasks often involves long inference-time computations. In human problem solving, a common strategy to expedite work is collaboration: by dividing the problem into sub-tasks, exploring different strategies concurrently, etc. Recent research has shown that LLMs can also operate in parallel by implementing explicit cooperation frameworks, such as voting mechanisms or the explicit creation of independent sub-tasks that can be executed in parallel. However, each of these frameworks may not be suitable for all types of tasks, which can hinder their applicability. In this work, we propose a different design approach: we run LLM "workers" in parallel , allowing them to synchronize via a concurrently-updated attention cache and prompt these workers to decide how best to collaborate. Our approach allows the instances to come up with their own collaboration strategy for the problem at hand, all the while "seeing" each other's partial progress in the concurrent cache. We implement this approach via Hogwild! Inference: a parallel LLM inference engine where multiple instances of the same LLM run in parallel with the same attention cache, with "instant" access to each other's generated tokens. Hogwild! inference takes advantage of Rotary Position Embeddings (RoPE) to avoid recomputation while improving parallel hardware utilization. We find that modern reasoning-capable LLMs can perform inference with shared Key-Value cache out of the box, without additional fine-tuning.

We present RWKV-7 "Goose", a new sequence modeling architecture, along with pre-trained language models that establish a new state-of-the-art in downstream performance at the 3 billion parameter scale on multilingual tasks, and match current SoTA English language performance despite being trained on dramatically fewer tokens than other top 3B models. Nevertheless, RWKV-7 models require only constant memory usage and constant inference time per token. RWKV-7 introduces a newly generalized formulation of the delta rule with vector-valued gating and in-context learning rates, as well as a relaxed value replacement rule. We show that RWKV-7 can perform state tracking and recognize all regular languages, while retaining parallelizability of training. This exceeds the capabilities of Transformers under standard complexity conjectures, which are limited to TC^0. To demonstrate RWKV-7's language modeling capability, we also present an extended open source 3.1 trillion token multilingual corpus, and train four RWKV-7 models ranging from 0.19 billion to 2.9 billion parameters on this dataset. To foster openness, reproduction, and adoption, we release our models and dataset component listing at https://huggingface.co/RWKV, and our training and inference code at https://github.com/RWKV/RWKV-LM all under the Apache 2.0 License.

Large language models (LLMs) are widely applied in chatbots, code generators, and search engines. Workloads such as chain-of-thought, complex reasoning, and agent services significantly increase the inference cost by invoking the model repeatedly. Optimization methods such as parallelism, compression, and caching have been adopted to reduce costs, but the diverse service requirements make it hard to select the right method. Recently, specialized LLM inference engines have emerged as a key component for integrating the optimization methods into service-oriented infrastructures. However, a systematic study on inference engines is still lacking. This paper provides a comprehensive evaluation of 25 open-source and commercial inference engines. We examine each inference engine in terms of ease-of-use, ease-of-deployment, general-purpose support, scalability, and suitability for throughput- and latency-aware computation. Furthermore, we explore the design goals of each inference engine by investigating the optimization techniques it supports. In addition, we assess the ecosystem maturity of open source inference engines and handle the performance and cost policy of commercial solutions. We outline future research directions that include support for complex LLM-based services, support of various hardware, and enhanced security, offering practical guidance to researchers and developers in selecting and designing optimized LLM inference engines. We also provide a public repository to continually track developments in this fast-evolving field: https://github.com/sihyeong/Awesome-LLM-Inference-Engine

Cross Attention is a popular method for retrieving information from a set of context tokens for making predictions. At inference time, for each prediction, Cross Attention scans the full set of O(N) tokens. In practice, however, often only a small subset of tokens are required for good performance. Methods such as Perceiver IO are cheap at inference as they distill the information to a smaller-sized set of latent tokens L < N on which cross attention is then applied, resulting in only O(L) complexity. However, in practice, as the number of input tokens and the amount of information to distill increases, the number of latent tokens needed also increases significantly. In this work, we propose Tree Cross Attention (TCA) - a module based on Cross Attention that only retrieves information from a logarithmic O(log(N)) number of tokens for performing inference. TCA organizes the data in a tree structure and performs a tree search at inference time to retrieve the relevant tokens for prediction. Leveraging TCA, we introduce ReTreever, a flexible architecture for token-efficient inference. We show empirically that Tree Cross Attention (TCA) performs comparable to Cross Attention across various classification and uncertainty regression tasks while being significantly more token-efficient. Furthermore, we compare ReTreever against Perceiver IO, showing significant gains while using the same number of tokens for inference.

One of the most striking findings in modern research on large language models (LLMs) is that scaling up compute during training leads to better results. However, less attention has been given to the benefits of scaling compute during inference. This survey focuses on these inference-time approaches. We explore three areas under a unified mathematical formalism: token-level generation algorithms, meta-generation algorithms, and efficient generation. Token-level generation algorithms, often called decoding algorithms, operate by sampling a single token at a time or constructing a token-level search space and then selecting an output. These methods typically assume access to a language model's logits, next-token distributions, or probability scores. Meta-generation algorithms work on partial or full sequences, incorporating domain knowledge, enabling backtracking, and integrating external information. Efficient generation methods aim to reduce token costs and improve the speed of generation. Our survey unifies perspectives from three research communities: traditional natural language processing, modern LLMs, and machine learning systems.

Transformers have emerged as the underpinning architecture for Large Language Models (LLMs). In generative language models, the inference process involves two primary phases: prompt processing and token generation. Token generation, which constitutes the majority of the computational workload, primarily entails vector-matrix multiplications and interactions with the Key-Value (KV) Cache. This phase is constrained by memory bandwidth due to the overhead of transferring weights and KV cache values from the memory system to the computing units. This memory bottleneck becomes particularly pronounced in applications that require long-context and extensive text generation, both of which are increasingly crucial for LLMs. This paper introduces "Keyformer", an innovative inference-time approach, to mitigate the challenges associated with KV cache size and memory bandwidth utilization. Keyformer leverages the observation that approximately 90% of the attention weight in generative inference focuses on a specific subset of tokens, referred to as "key" tokens. Keyformer retains only the key tokens in the KV cache by identifying these crucial tokens using a novel score function. This approach effectively reduces both the KV cache size and memory bandwidth usage without compromising model accuracy. We evaluate Keyformer's performance across three foundational models: GPT-J, Cerebras-GPT, and MPT, which employ various positional embedding algorithms. Our assessment encompasses a variety of tasks, with a particular emphasis on summarization and conversation tasks involving extended contexts. Keyformer's reduction of KV cache reduces inference latency by 2.1x and improves token generation throughput by 2.4x, while preserving the model's accuracy.

Language models are often trained to maximize the likelihood of the next token given past tokens in the training dataset. However, during inference time, they are utilized differently, generating text sequentially and auto-regressively by using previously generated tokens as input to predict the next one. Marginal differences in predictions at each step can cascade over successive steps, resulting in different distributions from what the models were trained for and potentially leading to unpredictable behavior. This paper proposes two simple approaches based on model own generation to address this discrepancy between the training and inference time. Our first approach is Batch-Scheduled Sampling, where, during training, we stochastically choose between the ground-truth token from the dataset and the model's own generated token as input to predict the next token. This is done in an offline manner, modifying the context window by interleaving ground-truth tokens with those generated by the model. Our second approach is Reference-Answer-based Correction, where we explicitly incorporate a self-correction capability into the model during training. This enables the model to effectively self-correct the gaps between the generated sequences and the ground truth data without relying on an external oracle model. By incorporating our proposed strategies during training, we have observed an overall improvement in performance compared to baseline methods, as demonstrated by our extensive experiments using summarization, general question-answering, and math question-answering tasks.

Frontier LLMs are increasingly utilised across academia, society and industry. A commonly used unit for comparing models, their inputs and outputs, and estimating inference pricing is the token. In general, tokens are used as a stable currency, assumed to be broadly consistent across tokenizers and contexts, enabling direct comparisons. However, tokenization varies significantly across models and domains of text, making naive interpretation of token counts problematic. We quantify this variation by providing a comprehensive empirical analysis of tokenization, exploring the compression of sequences to tokens across different distributions of textual data. Our analysis challenges commonly held heuristics about token lengths, finding them to be overly simplistic. We hope the insights of our study add clarity and intuition toward tokenization in contemporary LLMs.

Chain-of-thought (CoT) prompting enhances reasoning in large language models (LLMs) but often leads to verbose and redundant outputs, thus increasing inference cost. We hypothesize that many reasoning steps are unnecessary for producing correct answers. To investigate this, we start with a systematic study to examine what is the minimum reasoning required for a model to reach a stable decision. We find that on math reasoning tasks like math, models typically converge to their final answers after 60\% of the reasoning steps, suggesting substantial redundancy in the remaining content. Based on these insights, we propose three inference-time strategies to improve efficiency: (1) early stopping via answer consistency, (2) boosting the probability of generating end-of-reasoning signals, and (3) a supervised method that learns when to stop based on internal activations. Experiments across five benchmarks and five open-weights LLMs show that our methods significantly reduce token usage with little or no accuracy drop. In particular, on NaturalQuestions, Answer Consistency reduces tokens by over 40\% while further improving accuracy. Our work underscores the importance of cost-effective reasoning methods that operate at inference time, offering practical benefits for real-world applications.

Vision Language Models (VLMs) have demonstrated strong capabilities across various visual understanding and reasoning tasks. However, their real-world deployment is often constrained by high latency during inference due to substantial compute required to process the large number of input tokens (predominantly from the image) by the LLM. To reduce inference costs, one can either downsize the LLM or reduce the number of input image-tokens, the latter of which has been the focus of many recent works around token compression. However, it is unclear what the optimal trade-off is, as both the factors directly affect the VLM performance. We first characterize this optimal trade-off between the number of visual tokens and LLM parameters by establishing scaling laws that capture variations in performance with these two factors. Our results reveal a surprising trend: for visual reasoning tasks, the inference-optimal behavior in VLMs, i.e., minimum downstream error at any given fixed inference compute, is achieved when using the largest LLM that fits within the inference budget while minimizing visual token count - often to a single token. While the token reduction literature has mainly focused on maintaining base model performance by modestly reducing the token count (e.g., 5-10times), our results indicate that the compute-optimal inference regime requires operating under even higher token compression ratios. Based on these insights, we take some initial steps towards building approaches tailored for high token compression settings. Code is available at https://github.com/locuslab/llava-token-compression.

The recent surge of open-source large language models (LLMs) enables developers to create AI-based solutions while maintaining control over aspects such as privacy and compliance, thereby providing governance and ownership of the model deployment process. To utilize these LLMs, inference engines are needed. These engines load the model's weights onto available resources, such as GPUs, and process queries to generate responses. The speed of inference, or performance, of the LLM, is critical for real-time applications, as it computes millions or billions of floating point operations per inference. Recently, advanced inference engines such as vLLM have emerged, incorporating novel mechanisms such as efficient memory management to achieve state-of-the-art performance. In this paper, we analyze the performance, particularly the throughput (tokens generated per unit of time), of 20 LLMs using two inference libraries: vLLM and HuggingFace's pipelines. We investigate how various hyperparameters, which developers must configure, influence inference performance. Our results reveal that throughput landscapes are irregular, with distinct peaks, highlighting the importance of hyperparameter optimization to achieve maximum performance. We also show that applying hyperparameter optimization when upgrading or downgrading the GPU model used for inference can improve throughput from HuggingFace pipelines by an average of 9.16% and 13.7%, respectively.

Streaming visual transformers like StreamVGGT achieve strong 3D perception but suffer from unbounded growth of key value (KV) memory, which limits scalability. We propose a training-free, inference-time token eviction policy that bounds memory by discarding redundant tokens while keeping the most informative ones. Our method uses significantly less memory with little to no drop in accuracy: on 7-Scenes with long sequences it reduces peak memory from 18.63 GB to 9.39 GB while accuracy and completeness drop by only 0.003. Under strict memory budgets, eviction enables denser frame sampling, which improves reconstruction accuracy compared to the baseline. Experiments across video depth estimation (Sintel, KITTI), 3D reconstruction (7-Scenes, NRGBD), and camera pose estimation (Sintel, TUM-dynamics) show that our approach closely matches StreamVGGT at a fraction of the memory and makes long-horizon streaming inference more practical.

Transformer models have revolutionized natural language processing, achieving state-of-the-art performance and demonstrating remarkable scalability. However, their memory demands, particularly due to maintaining full context in memory, pose significant challenges for inference. In this paper, we present FlashEVA, an efficient implementation of EVA (Efficient Attention via Control Variates), and demonstrate how to finetune transformers to adapt to FlashEVA attention. Our method enables fine-tuning of Transformer models with as few as 1.5B tokens while preserving effectiveness across various downstream tasks. Notably, FlashEVA achieves up to 6.7x higher throughput and 5x lower peak GPU memory usage during inference compared to standard Transformer implementations. Despite these improvements, we observe limitations in retrieval-focused tasks. Our implementation offers control over the trade-off between throughput and accuracy through adjustable hyperparameters, providing flexibility for diverse use cases. This work represents a significant step towards more efficient and adaptable Transformer-based models for inference.

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.

Large language models face significant computational and memory challenges when processing long contexts. During inference, efficient management of the key-value (KV) cache, which stores intermediate activations for autoregressive generation, is critical to reducing memory overhead and improving computational efficiency. Traditional token-level efficient KV caching methods overlook semantic information, treating tokens independently without considering their semantic relationships. Meanwhile, existing semantic-preserving KV cache management approaches often suffer from substantial memory usage and high time-to-first-token. To address these limitations, we propose SentenceKV, a novel sentence-level semantic KV caching approach designed to enhance inference efficiency while preserving semantic coherence. During prefilling, SentenceKV groups tokens based on sentence-level semantic similarity, compressing sentence representations into concise semantic vectors stored directly on the GPU, while individual KV pairs are offloaded to CPU. During decoding, SentenceKV generates tokens by selectively retrieving semantically relevant sentence-level KV entries, leveraging the semantic similarity between the prefilling-stage semantic vectors and decoding-stage queries. This ensures efficient and contextually accurate predictions, minimizing the loading of redundant or irrelevant data into GPU memory and significantly reducing memory overhead while maintaining stable inference latency, even for extremely long contexts. Extensive evaluations on benchmarks including PG-19, LongBench, and Needle-In-A-Haystack demonstrate that SentenceKV significantly outperforms state-of-the-art methods in both efficiency and memory usage, without compromising model accuracy.

The increasing context window size in Large Language Models (LLMs), such as the GPT and LLaMA series, has improved their ability to tackle complex, long-text tasks, but at the cost of inference efficiency, particularly regarding memory and computational complexity. Existing methods, including selective token retention and window-based attention, improve efficiency but risk discarding important tokens needed for future text generation. In this paper, we propose an approach that enhances LLM efficiency without token loss by reducing the memory and computational load of less important tokens, rather than discarding them.We address two challenges: 1) investigating the distribution of important tokens in the context, discovering recent tokens are more important than distant tokens in context, and 2) optimizing resources for distant tokens by sharing attention scores across layers. The experiments show that our method saves 35% KV cache without compromising the performance.

While Chain-of-Thought (CoT) prompting improves reasoning in large language models (LLMs), the excessive length of reasoning tokens increases latency and KV cache memory usage, and may even truncate final answers under context limits. We propose ThinkLess, an inference-efficient framework that terminates reasoning generation early and maintains output quality without modifying the model. Atttention analysis reveals that answer tokens focus minimally on earlier reasoning steps and primarily attend to the reasoning terminator token, due to information migration under causal masking. Building on this insight, ThinkLess inserts the terminator token at earlier positions to skip redundant reasoning while preserving the underlying knowledge transfer. To prevent format discruption casued by early termination, ThinkLess employs a lightweight post-regulation mechanism, relying on the model's natural instruction-following ability to produce well-structured answers. Without fine-tuning or auxiliary data, ThinkLess achieves comparable accuracy to full-length CoT decoding while greatly reducing decoding time and memory consumption.

Large language models (LLMs) rely on key-value (KV) caches for efficient autoregressive decoding; however, cache size grows linearly with context length and model depth, becoming a major bottleneck in long-context inference. Prior KV cache compression methods either enforce rigid heuristics, disrupt tensor layouts with per-attention-head variability, or require specialized compute kernels. We propose a simple, yet effective, KV cache compression framework based on attention-guided, layer-adaptive composite tokens. Our method aggregates attention scores to estimate token importance, selects head-specific tokens independently, and aligns them into composite tokens that respect the uniform cache structure required by existing inference engines. A global allocation mechanism further adapts retention budgets across layers, assigning more capacity to layers with informative tokens. This approach achieves significant memory reduction while preserving accuracy, consistently outperforming prior structured and semi-structured methods. Crucially, our approach remains fully compatible with standard inference pipelines, offering a practical and scalable solution for efficient long-context LLM deployment.

Large Language Models (LLMs) are increasingly deployed in scenarios demanding ultra-long context reasoning, such as agentic workflows and deep research understanding. However, long-context inference is constrained by the KV cache, a transient memory structure that grows linearly with sequence length and batch size, quickly dominating GPU memory usage. Existing memory reduction techniques, including eviction and quantization, often rely on static heuristics and suffer from degraded quality under tight budgets. In this paper, we propose ARKV, a lightweight and adaptive framework that dynamically allocates precision levels to cached tokens based on per-layer attention dynamics and token-level importance. During a short prefill phase, ARKV estimates the original quantization (OQ) ratio of each layer by computing statistical scores such as attention entropy, variance and kurtosis. During decoding, tokens are assigned to one of three states, Original (full precision), Quantization (low precision), or Eviction, according to a fast heavy-hitter scoring strategy. Our experiments on LLaMA3 and Qwen3 models across diverse long- and short-context tasks demonstrate that ARKV preserves ~97% of baseline accuracy on long-context benchmarks while reducing KV memory usage by 4x, with minimal throughput loss. On short-context tasks, ARKV matches full-precision baselines; on GSM8K math reasoning, it significantly outperforms uniform quantization. These results highlight the practical viability of ARKV for scalable LLM deployment, offering fine-grained, data-driven memory control without retraining or architectural modifications. The source code and artifacts can be found in: https://github.com/Large-scale-Sustainable-Computing-LSC/ARKV

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H_2). Through a comprehensive investigation, we find that (i) the emergence of H_2 is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H_2O), a KV cache eviction policy that dynamically retains a balance of recent and H_2 tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H_2O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29times, 29times, and 3times on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9times. The code is available at https://github.com/FMInference/H2O.

As large language models (LLMs) process increasing context windows, the memory usage of KV cache has become a critical bottleneck during inference. The mainstream KV compression methods, including KV pruning and KV quantization, primarily focus on either token or precision dimension and seldom explore the efficiency of their combination. In this paper, we comprehensively investigate the token-precision trade-off in KV cache compression. Experiments demonstrate that storing more tokens in the KV cache with lower precision, i.e., quantized pruning, can significantly enhance the long-context performance of LLMs. Furthermore, in-depth analysis regarding token-precision trade-off from a series of key aspects exhibit that, quantized pruning achieves substantial improvements in retrieval-related tasks and consistently performs well across varying input lengths. Moreover, quantized pruning demonstrates notable stability across different KV pruning methods, quantization strategies, and model scales. These findings provide valuable insights into the token-precision trade-off in KV cache compression. We plan to release our code in the near future.

As demand for LLM inference grows, it is becoming increasingly important that providers and their customers can verify that inference processes are performed correctly, without errors or tampering. However, re-running the same inference process twice often leads to different results due to benign numerical noise, making it difficult to distinguish legitimate variation from actual problems. To address this problem, we introduce Token-DiFR (Token-Divergence-From-Reference), a method for verifying inference outputs by comparing generated tokens against predictions made by a trusted reference implementation conditioned on the same random seed. Sampling seed synchronization tightly constrains valid outputs, leaving providers minimal room to deviate from correct inference, which allows output tokens themselves to serve as auditable evidence of correctness at zero additional cost to the provider. Token-DiFR reliably identifies sampling errors, simulated bugs, and model quantization, detecting 4-bit quantization with AUC > 0.999 within 300 output tokens. For applications requiring sample-efficient forward-pass verification, we additionally introduce Activation-DiFR, a scheme that uses random orthogonal projections to compress activations into compact fingerprints for subsequent verification. Activation-DiFR detects 4-bit quantization with AUC > 0.999 using just 2 output tokens, while reducing communication overhead by 25-75% relative to existing methods. We release an open-source integration with vLLM to accelerate practical deployment of verifiable inference.

Video large language models (VideoLLMs) have demonstrated the capability to process longer video inputs and enable complex reasoning and analysis. However, due to the thousands of visual tokens from the video frames, key-value (KV) cache can significantly increase memory requirements, becoming a bottleneck for inference speed and memory usage. KV cache quantization is a widely used approach to address this problem. In this paper, we find that 2-bit KV quantization of VideoLLMs can hardly hurt the model performance, while the limit of KV cache quantization in even lower bits has not been investigated. To bridge this gap, we introduce VidKV, a plug-and-play KV cache quantization method to compress the KV cache to lower than 2 bits. Specifically, (1) for key, we propose a mixed-precision quantization strategy in the channel dimension, where we perform 2-bit quantization for anomalous channels and 1-bit quantization combined with FFT for normal channels; (2) for value, we implement 1.58-bit quantization while selectively filtering semantically salient visual tokens for targeted preservation, for a better trade-off between precision and model performance. Importantly, our findings suggest that the value cache of VideoLLMs should be quantized in a per-channel fashion instead of the per-token fashion proposed by prior KV cache quantization works for LLMs. Empirically, extensive results with LLaVA-OV-7B and Qwen2.5-VL-7B on six benchmarks show that VidKV effectively compresses the KV cache to 1.5-bit and 1.58-bit precision with almost no performance drop compared to the FP16 counterparts.

Prompt tuning (PT), where a small amount of trainable soft (continuous) prompt vectors is affixed to the input of language models (LM), has shown promising results across various tasks and models for parameter-efficient fine-tuning (PEFT). PT stands out from other PEFT approaches because it maintains competitive performance with fewer trainable parameters and does not drastically scale up its parameters as the model size expands. However, PT introduces additional soft prompt tokens, leading to longer input sequences, which significantly impacts training and inference time and memory usage due to the Transformer's quadratic complexity. Particularly concerning for Large Language Models (LLMs) that face heavy daily querying. To address this issue, we propose Decomposed Prompt Tuning (DePT), which decomposes the soft prompt into a shorter soft prompt and a pair of low-rank matrices that are then optimised with two different learning rates. This allows DePT to achieve better performance while saving over 20% memory and time costs compared to vanilla PT and its variants, without changing trainable parameter sizes. Through extensive experiments on 23 natural language processing (NLP) and vision-language (VL) tasks, we demonstrate that DePT outperforms state-of-the-art PEFT approaches, including the full fine-tuning baseline in some scenarios. Additionally, we empirically show that DEPT grows more efficient as the model size increases. Our further study reveals that DePT integrates seamlessly with parameter-efficient transfer learning in the few-shot learning setting and highlights its adaptability to various model architectures and sizes.

State Space Models (SSMs) have emerged as a promising alternative to Transformers for long-context sequence modeling, offering linear O(N) computational complexity compared to the Transformer's quadratic O(N^2) scaling. This paper presents a comprehensive benchmarking study comparing the Mamba SSM against the LLaMA Transformer on long-context sequences, using dyadic therapy sessions as a representative test case. We evaluate both architectures across two dimensions: (1) computational efficiency, where we measure memory usage and inference speed from 512 to 8,192 tokens, and (2) representational efficiency, where we analyze hidden state dynamics and attention patterns. Our findings provide actionable insights for practitioners working with long-context applications, establishing precise conditions under which SSMs offer advantages over Transformers.

Large language models (LLMs) are widely used to tackle complex tasks with autonomous workflows. Recently, reusable natural language skills have emerged as a popular paradigm to inject procedural knowledge into LLM applications. Since popular skills are often invoked repeatedly, placing their full text in every context significantly increases prefill cost and latency. While text compression techniques have the potential to solve this problem, most existing methods are designed to compress factual knowledge in documents instead of procedural knowledge, making them insufficient for skill compression. In this paper, we argue that an effective skill compression method should: 1) preserve logical dependencies among workflows and tool protocols, 2) enable lightweight, offline compression for frequently updated community skills, and 3) be adaptable to varying complexities across skills. To address this, we present SKIM (SKIll coMpression), an adaptive multi-resolution soft token compression framework for procedural skills. Depending on the complexity of each skill, SKIM creates different numbers of soft tokens that not only improve the efficiency of LLM inference, but also preserve the effectiveness of skill usage. Experiments indicate that SKIM compresses skills to 30 to 60 percent of their original token length while preserving task performance better than existing compression methods.We have released our code at https://github.com/bebr2/SKIM .

As Large Language Models (LLMs) have made significant advancements across various tasks, such as question answering, translation, text summarization, and dialogue systems, the need for accuracy in information becomes crucial, especially for serious financial products serving billions of users like Alipay. To address this, Alipay has developed a Retrieval-Augmented Generation (RAG) system that grounds LLMs on the most accurate and up-to-date information. However, for a real-world product serving millions of users, the inference speed of LLMs becomes a critical factor compared to a mere experimental model. Hence, this paper presents a generic framework for accelerating the inference process, resulting in a substantial increase in speed and cost reduction for our RAG system, with lossless generation accuracy. In the traditional inference process, each token is generated sequentially by the LLM, leading to a time consumption proportional to the number of generated tokens. To enhance this process, our framework, named lookahead, introduces a multi-branch strategy. Instead of generating a single token at a time, we propose a Trie-based Retrieval (TR) process that enables the generation of multiple branches simultaneously, each of which is a sequence of tokens. Subsequently, for each branch, a Verification and Accept (VA) process is performed to identify the longest correct sub-sequence as the final output. Our strategy offers two distinct advantages: (1) it guarantees absolute correctness of the output, avoiding any approximation algorithms, and (2) the worst-case performance of our approach is equivalent to the conventional process. We conduct extensive experiments to demonstrate the significant improvements achieved by applying our inference acceleration framework. Code is avaliable: https://github.com/alipay/PainlessInferenceAcceleration.

In real world, large language models (LLMs) can serve as the assistant to help users accomplish their jobs, and also support the development of advanced applications. For the wide application of LLMs, the inference efficiency is an essential concern, which has been widely studied in existing work, and numerous optimization algorithms and code libraries have been proposed to improve it. Nonetheless, users still find it challenging to compare the effectiveness of all the above methods and understand the underlying mechanisms. In this work, we perform a detailed coarse-to-fine analysis of the inference performance of various code libraries. To evaluate the overall effectiveness, we examine four usage scenarios within two practical applications. We further provide both theoretical and empirical fine-grained analyses of each module in the Transformer architecture. Our experiments yield comprehensive results that are invaluable for researchers to evaluate code libraries and improve inference strategies.

Although Large Vision-Language Models (LVLMs) have achieved impressive results, their high computational cost poses a significant barrier to wider application. To enhance inference efficiency, most existing approaches depend on parameter-dependent or token-dependent strategies to reduce computational demands. However, these methods typically require complex training processes and struggle to consistently select the most relevant tokens. In this paper, we systematically analyze the above challenges and provide a series of valuable insights for inference acceleration. Based on these findings, we propose a novel framework, the Pruning All-Rounder (PAR). Different from previous works, PAR develops a meta-router to adaptively organize pruning flows across both tokens and layers. With a self-supervised learning manner, our method achieves a superior balance between performance and efficiency. Notably, PAR is highly flexible, offering multiple pruning versions to address a range of pruning scenarios. The code for this work will be made publicly available.

Performing inference on hundreds of thousands of samples with large language models (LLMs) can be computationally and financially costly. We propose batch prompting, a simple alternative prompting approach that enables the LLM to run inference in batches, instead of one sample at a time. Our method reduces both token and time costs while retaining downstream performance. We theoretically demonstrate that under a few-shot in-context learning setting, the inference costs decrease almost inverse linearly with the number of samples in each batch. We extensively validate the effectiveness of batch prompting on ten datasets across commonsense QA, arithmetic reasoning, and NLI/NLU: batch prompting significantly~(up to 5times with six samples in batch) reduces the LLM (Codex) inference token and time costs while achieving better or comparable performance. Our analysis shows that the number of samples in each batch and the complexity of tasks affect its performance. Further, batch prompting can be applied across different LLMs and reasoning methods.

Reasoning and inference are central to human and artificial intelligence. Modeling inference in human language is very challenging. With the availability of large annotated data (Bowman et al., 2015), it has recently become feasible to train neural network based inference models, which have shown to be very effective. In this paper, we present a new state-of-the-art result, achieving the accuracy of 88.6% on the Stanford Natural Language Inference Dataset. Unlike the previous top models that use very complicated network architectures, we first demonstrate that carefully designing sequential inference models based on chain LSTMs can outperform all previous models. Based on this, we further show that by explicitly considering recursive architectures in both local inference modeling and inference composition, we achieve additional improvement. Particularly, incorporating syntactic parsing information contributes to our best result---it further improves the performance even when added to the already very strong model.

Progress on training and architecture strategies has enabled LLMs with millions of tokens in context length. However, empirical evidence suggests that such long-context LLMs can consume far more text than they can reliably use. On the other hand, it has been shown that inference-time compute can be used to scale performance of LLMs, often by generating thinking tokens, on challenging tasks involving multi-step reasoning. Through controlled experiments on sandbox long-context tasks, we find that such inference-time strategies show rapidly diminishing returns and fail at long context. We attribute these failures to score dilution, a phenomenon inherent to static self-attention. Further, we show that current inference-time strategies cannot retrieve relevant long-context signals under certain conditions. We propose a simple method that, through targeted gradient updates on the given context, provably overcomes limitations of static self-attention. We find that this shift in how inference-time compute is spent leads to consistently large performance improvements across models and long-context benchmarks. Our method leads to large 12.6 and 14.1 percentage point improvements for Qwen3-4B on average across subsets of LongBench-v2 and ZeroScrolls benchmarks. The takeaway is practical: for long context, a small amount of context-specific training is a better use of inference compute than current inference-time scaling strategies like producing more thinking tokens.

Membership inference attacks (MIA) attempt to verify the membership of a given data sample in the training set for a model. MIA has become relevant in recent years, following the rapid development of large language models (LLM). Many are concerned about the usage of copyrighted materials for training them and call for methods for detecting such usage. However, recent research has largely concluded that current MIA methods do not work on LLMs. Even when they seem to work, it is usually because of the ill-designed experimental setup where other shortcut features enable "cheating." In this work, we argue that MIA still works on LLMs, but only when multiple documents are presented for testing. We construct new benchmarks that measure the MIA performances at a continuous scale of data samples, from sentences (n-grams) to a collection of documents (multiple chunks of tokens). To validate the efficacy of current MIA approaches at greater scales, we adapt a recent work on Dataset Inference (DI) for the task of binary membership detection that aggregates paragraph-level MIA features to enable MIA at document and collection of documents level. This baseline achieves the first successful MIA on pre-trained and fine-tuned LLMs.

Inference accounts for the majority of latency and energy consumption in large language model (LLM) deployments, often exceeding 90% of total cost. While training-time efficiency has seen extensive progress, runtime optimization remains a key bottleneck, particularly under autoregressive decoding. Existing approaches -- such as pruning, quantization, early exits, and speculative decoding -- often require retraining, architectural changes, or disrupt decoding compatibility. We introduce QuickSilver, a modular, token-level framework that enables semantic adaptivity at inference time without altering model weights or structure. QuickSilver integrates four synergistic mechanisms: (i) Dynamic Token Halting, which halts computation for tokens with converged representations; (ii) KV Cache Skipping, which selectively suppresses memory writes to reduce attention overhead; and (iii) Contextual Token Fusion, which collapses redundant tokens into shared paths to shrink sequence length. Unlike speculative decoding or MoE routing, QuickSilver operates entirely on frozen, dense models and requires no auxiliary networks. Applied to GPT-2 and Llama-2 across WikiText-103 and C4, QuickSilver achieves up to 39.6% FLOP reduction with negligible perplexity degradation (<=0.2).

Large Language Models (LLMs) have demonstrated strong performance in handling complex tasks requiring both extensive knowledge and reasoning abilities. However, the existing LLM inference pipeline operates as an opaque process without explicit separation between knowledge retrieval and reasoning steps, making the model's decision-making process unclear and disorganized. This ambiguity can lead to issues such as hallucinations and knowledge forgetting, which significantly impact the reliability of LLMs in high-stakes domains. In this paper, we propose a new inference paradigm that decomposes the complex inference process into two distinct and clear actions: (1) memory recall: which retrieves relevant knowledge, and (2) reasoning: which performs logical steps based on the recalled knowledge. To facilitate this decomposition, we introduce two special tokens memory and reason, guiding the model to distinguish between steps that require knowledge retrieval and those that involve reasoning. Our experiment results show that this decomposition not only improves model performance but also enhances the interpretability of the inference process, enabling users to identify sources of error and refine model responses effectively. The code is available at https://github.com/MingyuJ666/Disentangling-Memory-and-Reasoning.

Test-time scaling has emerged as an effective approach for improving language model performance by utilizing additional compute at inference time. Recent studies have shown that overriding end-of-thinking tokens (e.g., replacing "</think>" with "Wait") can extend reasoning steps and improve accuracy. In this work, we explore whether a dedicated continue-thinking token can be learned to trigger extended reasoning. We augment a distilled version of DeepSeek-R1 with a single learned "<|continue-thinking|>" token, training only its embedding via reinforcement learning while keeping the model weights frozen. Our experiments show that this learned token achieves improved accuracy on standard math benchmarks compared to both the baseline model and a test-time scaling approach that uses a fixed token (e.g., "Wait") for budget forcing. In particular, we observe that in cases where the fixed-token approach enhances the base model's accuracy, our method achieves a markedly greater improvement. For example, on the GSM8K benchmark, the fixed-token approach yields a 1.3% absolute improvement in accuracy, whereas our learned-token method achieves a 4.2% improvement over the base model that does not use budget forcing.

Large Language Models (LLMs) have made significant strides in problem-solving by incorporating reasoning processes. However, this enhanced reasoning capability results in an increased number of output tokens during inference, leading to higher computational costs. To address this challenge, we propose TwT (Thinking without Tokens), a method that reduces inference-time costs through habitual reasoning distillation with multi-teachers' guidance, while maintaining high performance. Our approach introduces a Habitual Reasoning Distillation method, which internalizes explicit reasoning into the model's habitual behavior through a Teacher-Guided compression strategy inspired by human cognition. Additionally, we propose Dual-Criteria Rejection Sampling (DCRS), a technique that generates a high-quality and diverse distillation dataset using multiple teacher models, making our method suitable for unsupervised scenarios. Experimental results demonstrate that TwT effectively reduces inference costs while preserving superior performance, achieving up to a 13.6% improvement in accuracy with fewer output tokens compared to other distillation methods, offering a highly practical solution for efficient LLM deployment.

A membership inference attack allows an adversary to query a trained machine learning model to predict whether or not a particular example was contained in the model's training dataset. These attacks are currently evaluated using average-case "accuracy" metrics that fail to characterize whether the attack can confidently identify any members of the training set. We argue that attacks should instead be evaluated by computing their true-positive rate at low (e.g., <0.1%) false-positive rates, and find most prior attacks perform poorly when evaluated in this way. To address this we develop a Likelihood Ratio Attack (LiRA) that carefully combines multiple ideas from the literature. Our attack is 10x more powerful at low false-positive rates, and also strictly dominates prior attacks on existing metrics.

While modern Transformer-based language models (LMs) have achieved major success in multi-task generalization, they often struggle to capture long-range dependencies within their context window. This work introduces a novel approach using meta-tokens, special tokens injected during pre-training, along with a dedicated meta-attention mechanism to guide LMs to use these tokens. We pre-train a language model with a modified GPT-2 architecture equipped with meta-attention in addition to causal multi-head attention, and study the impact of these tokens on a suite of synthetic tasks. We find that data-efficient language model pre-training on fewer than 100B tokens utilizing meta-tokens and our meta-attention mechanism achieves strong performance on these tasks after fine-tuning. We suggest that these gains arise due to the meta-tokens sharpening the positional encoding. This enables them to operate as trainable, content-based landmarks, implicitly compressing preceding context and "caching" it in the meta-token. At inference-time, the meta-token points to relevant context, facilitating length generalization up to 2times its context window, even after extension with YaRN. We provide further evidence of these behaviors by visualizing model internals to study the residual stream, and assessing the compression quality by information-theoretic analysis on the rate-distortion tradeoff. Our findings suggest that pre-training LMs with meta-tokens offers a simple, data-efficient method to enhance long-context language modeling performance, while introducing new insights into the nature of their behavior towards length generalization.

Latent tokens are gaining attention for enhancing reasoning in large language models (LLMs), yet their internal mechanisms remain unclear. This paper examines the problem from a reliability perspective, uncovering fundamental weaknesses: latent tokens function as uninterpretable placeholders rather than encoding faithful reasoning. While resistant to perturbation, they promote shortcut usage over genuine reasoning. We focus on Chain-of-Continuous-Thought (COCONUT), which claims better efficiency and stability than explicit Chain-of-Thought (CoT) while maintaining performance. We investigate this through two complementary approaches. First, steering experiments perturb specific token subsets, namely COCONUT and explicit CoT. Unlike CoT tokens, COCONUT tokens show minimal sensitivity to steering and lack reasoning-critical information. Second, shortcut experiments evaluate models under biased and out-of-distribution settings. Results on MMLU and HotpotQA demonstrate that COCONUT consistently exploits dataset artifacts, inflating benchmark performance without true reasoning. These findings reposition COCONUT as a pseudo-reasoning mechanism: it generates plausible traces that conceal shortcut dependence rather than faithfully representing reasoning processes.

In domains with high stakes such as law, recruitment, and healthcare, learning models frequently rely on sensitive user data for inference, necessitating the complete set of features. This not only poses significant privacy risks for individuals but also demands substantial human effort from organizations to verify information accuracy. This paper asks whether it is necessary to use all input features for accurate predictions at inference time. The paper demonstrates that, in a personalized setting, individuals may only need to disclose a small subset of their features without compromising decision-making accuracy. The paper also provides an efficient sequential algorithm to determine the appropriate attributes for each individual to provide. Evaluations across various learning tasks show that individuals can potentially report as little as 10\% of their information while maintaining the same accuracy level as a model that employs the full set of user information.

Emerging Large Language Model (LLM) applications require long input prompts to perform complex downstream tasks like document analysis and code generation. For these long context length applications, the length of the input prompt poses a significant challenge in terms of inference efficiency since the inference costs increase linearly with sequence length. However, for many of these applications, much of the context in the prompt is fixed across different user inputs, thereby providing the opportunity to perform offline optimizations to process user inputs quickly, as they are received. In this work, we propose Squeezed Attention as a mechanism to accelerate LLM applications where a large portion of the input prompt is fixed. We first leverage K-means clustering offline to group the keys for the fixed context based on semantic similarity and represent each cluster with a single centroid value. During inference, we compare query tokens from the user input with the centroids to predict which of the keys from the fixed context are semantically relevant and need to be loaded during inference. We then compute exact attention using only these important keys from the fixed context, thereby reducing bandwidth and computational costs. We also extend our method to use a hierarchical centroid lookup to identify important keys, which can reduce the complexity of attention from linear to logarithmic with respect to the context length. We implement optimized Triton kernels for centroid comparison and sparse FlashAttention with important keys, achieving more than 4x speedups during both the prefill and generation phases for long-context inference. Furthermore, we have extensively evaluated our method on various long-context benchmarks including LongBench, where it achieves a 3x reduction in KV cache budget without accuracy loss and up to an 8x reduction with <0.5 point accuracy gap for various models.

Latent reasoning models (LRMs) have attracted significant research interest due to their low inference cost (relative to explicit reasoning models) and theoretical ability to explore multiple reasoning paths in parallel. However, these benefits come at the cost of reduced interpretability: LRMs are difficult to monitor because they do not reason in natural language. This paper presents an investigation into LRM interpretability by examining two state-of-the-art LRMs. First, we find that latent reasoning tokens are often unnecessary for LRMs' predictions; on logical reasoning datasets, LRMs can almost always produce the same final answers without using latent reasoning at all. This underutilization of reasoning tokens may partially explain why LRMs do not consistently outperform explicit reasoning methods and raises doubts about the stated role of these tokens in prior work. Second, we demonstrate that when latent reasoning tokens are necessary for performance, we can decode gold reasoning traces up to 65-93% of the time for correctly predicted instances. This suggests LRMs often implement the expected solution rather than an uninterpretable reasoning process. Finally, we present a method to decode a verified natural language reasoning trace from latent tokens without knowing a gold reasoning trace a priori, demonstrating that it is possible to find a verified trace for a majority of correct predictions but only a minority of incorrect predictions. Our findings highlight that current LRMs largely encode interpretable processes, and interpretability itself can be a signal of prediction correctness.

Large Reasoning Models (LRMs) achieve remarkable performance by explicitly generating multi-step chains of thought, but this capability incurs substantial inference latency and computational cost. Collaborative inference offers a promising solution by selectively allocating work between lightweight and large models, yet a fundamental challenge remains: determining when a reasoning step requires the capacity of a large model or the efficiency of a small model. Existing routing strategies either rely on local token probabilities or post-hoc verification, introducing significant inference overhead. In this work, we propose a novel perspective on step-wise collaboration: the difficulty of a reasoning step can be inferred from its very first token. Inspired by the "Aha Moment" phenomenon in LRMs, we show that the entropy of the initial token serves as a strong predictor of step difficulty. Building on this insight, we introduce GlimpRouter, a training-free step-wise collaboration framework. GlimpRouter employs a lightweight model to generate only the first token of each reasoning step and routes the step to a larger model only when the initial token entropy exceeds a threshold. Experiments on multiple benchmarks demonstrate that our approach significantly reduces inference latency while preserving accuracy. For instance, GlimpRouter attains a substantial 10.7% improvement in accuracy while reducing inference latency by 25.9% compared to a standalone large model on AIME25. These results suggest a simple yet effective mechanism for reasoning: allocating computation based on a glimpse of thought rather than full-step evaluation.

When using agent-task datasets to enhance agent capabilities for Large Language Models (LLMs), current methodologies often treat all tokens within a sample equally. However, we argue that tokens serving different roles - specifically, reasoning tokens versus boilerplate tokens (e.g., those governing output format) - differ significantly in importance and learning complexity, necessitating their disentanglement and distinct treatment. To address this, we propose a novel Shuffle-Aware Discriminator (SHAD) for adaptive token discrimination. SHAD classifies tokens by exploiting predictability differences observed after shuffling input-output combinations across samples: boilerplate tokens, due to their repetitive nature among samples, maintain predictability, whereas reasoning tokens do not. Using SHAD, we propose the Reasoning-highlighted Fine-Tuning (RFT) method, which adaptively emphasizes reasoning tokens during fine-tuning, yielding notable performance gains over common Supervised Fine-Tuning (SFT).

Large language models excel at reasoning but lack key aspects of introspection, including anticipating their own success and the computation required to achieve it. Humans use real-time introspection to decide how much effort to invest, when to make multiple attempts, when to stop, and when to signal success or failure. Without this, LLMs struggle to make intelligent meta-cognition decisions. Test-time scaling methods like Best-of-N drive up cost and latency by using a fixed budget of samples regardless of the marginal benefit of each one at any point in generation, and the absence of confidence signals can mislead people, prevent appropriate escalation to better tools, and undermine trustworthiness. Learned verifiers or reward models can provide confidence estimates, but do not enable adaptive inference and add substantial cost by requiring extra models or forward passes. We present ZIP-RC, which equips models with zero-overhead introspective predictions of reward and cost. At every token, ZIP-RC reuses reserved or unused logits in the same forward pass as next-token prediction to output a joint distribution over final reward and remaining length -- no extra models, architecture change, or inference overhead. This full joint distribution is used to compute a sampling utility which is the linear combination of the expected maximum reward, total compute, and latency of set of samples if generated to completion. During inference, we maximize this utility with meta-actions that determine which prefix of tokens to continue or initiate sampling from. On mixed-difficulty mathematical benchmarks, ZIP-RC improves accuracy by up to 12% over majority voting at equal or lower average cost, and traces smooth Pareto frontiers between quality, compute, and latency. By providing real-time reward-cost introspection, ZIP-RC enables adaptive, efficient reasoning.

We present Prompt Cache, an approach for accelerating inference for large language models (LLM) by reusing attention states across different LLM prompts. Many input prompts have overlapping text segments, such as system messages, prompt templates, and documents provided for context. Our key insight is that by precomputing and storing the attention states of these frequently occurring text segments on the inference server, we can efficiently reuse them when these segments appear in user prompts. Prompt Cache employs a schema to explicitly define such reusable text segments, called prompt modules. The schema ensures positional accuracy during attention state reuse and provides users with an interface to access cached states in their prompt. Using a prototype implementation, we evaluate Prompt Cache across several LLMs. We show that Prompt Cache significantly reduce latency in time-to-first-token, especially for longer prompts such as document-based question answering and recommendations. The improvements range from 8x for GPU-based inference to 60x for CPU-based inference, all while maintaining output accuracy and without the need for model parameter modifications.

An increasing number of companies have begun providing services that leverage cloud-based large language models (LLMs), such as ChatGPT. However, this development raises substantial privacy concerns, as users' prompts are transmitted to and processed by the model providers. Among the various privacy protection methods for LLMs, those implemented during the pre-training and fine-tuning phrases fail to mitigate the privacy risks associated with the remote use of cloud-based LLMs by users. On the other hand, methods applied during the inference phrase are primarily effective in scenarios where the LLM's inference does not rely on privacy-sensitive information. In this paper, we outline the process of remote user interaction with LLMs and, for the first time, propose a detailed definition of a general pseudonymization framework applicable to cloud-based LLMs. The experimental results demonstrate that the proposed framework strikes an optimal balance between privacy protection and utility. The code for our method is available to the public at https://github.com/Mebymeby/Pseudonymization-Framework.

As large language models gain widespread adoption, running them efficiently becomes crucial. Recent works on LLM inference use speculative decoding to achieve extreme speedups. However, most of these works implicitly design their algorithms for high-end datacenter hardware. In this work, we ask the opposite question: how fast can we run LLMs on consumer machines? Consumer GPUs can no longer fit the largest available models (50B+ parameters) and must offload them to RAM or SSD. When running with offloaded parameters, the inference engine can process batches of hundreds or thousands of tokens at the same time as just one token, making it a natural fit for speculative decoding. We propose SpecExec (Speculative Execution), a simple parallel decoding method that can generate up to 20 tokens per target model iteration for popular LLM families. It utilizes the high spikiness of the token probabilities distribution in modern LLMs and a high degree of alignment between model output probabilities. SpecExec takes the most probable tokens continuation from the draft model to build a "cache" tree for the target model, which then gets validated in a single pass. Using SpecExec, we demonstrate inference of 50B+ parameter LLMs on consumer GPUs with RAM offloading at 4-6 tokens per second with 4-bit quantization or 2-3 tokens per second with 16-bit weights.

We explore the Iterative Inference Hypothesis (IIH) within the context of transformer-based language models, aiming to understand how a model's latent representations are progressively refined and whether observable differences are present between correct and incorrect generations. Our findings provide empirical support for the IIH, showing that the nth token embedding in the residual stream follows a trajectory of decreasing loss. Additionally, we observe that the rate at which residual embeddings converge to a stable output representation reflects uncertainty in the token generation process. Finally, we introduce a method utilizing cross-entropy to detect this uncertainty and demonstrate its potential to distinguish between correct and incorrect token generations on a dataset of idioms.

Reasoning has become a central paradigm for large language models (LLMs), consistently boosting accuracy across diverse benchmarks. Yet its suitability for precision-sensitive tasks remains unclear. We present the first systematic study of reasoning for classification tasks under strict low false positive rate (FPR) regimes. Our analysis covers two tasks--safety detection and hallucination detection--evaluated in both fine-tuned and zero-shot settings, using standard LLMs and Large Reasoning Models (LRMs). Our results reveal a clear trade-off: Think On (reasoning-augmented) generation improves overall accuracy, but underperforms at the low-FPR thresholds essential for practical use. In contrast, Think Off (no reasoning during inference) dominates in these precision-sensitive regimes, with Think On surpassing only when higher FPRs are acceptable. In addition, we find token-based scoring substantially outperforms self-verbalized confidence for precision-sensitive deployments. Finally, a simple ensemble of the two modes recovers the strengths of each. Taken together, our findings position reasoning as a double-edged tool: beneficial for average accuracy, but often ill-suited for applications requiring strict precision.

While subword tokenizers such as BPE and WordPiece are typically used to build vocabularies for NLP models, the method of decoding text into a sequence of tokens from these vocabularies is often left unspecified, or ill-suited to the method in which they were constructed. We provide a controlled analysis of seven tokenizer inference methods across four different algorithms and three vocabulary sizes, performed on a novel intrinsic evaluation suite we curated for English, combining measures rooted in morphology, cognition, and information theory. We show that for the most commonly used tokenizers, greedy inference performs surprisingly well; and that SaGe, a recently-introduced contextually-informed tokenizer, outperforms all others on morphological alignment.

Large language models (LLMs) have played a fundamental role in various natural language processing tasks with powerful prompt techniques. However, in real-world applications, there are often similar prompt components for repeated queries, which causes significant computational burdens during inference. Existing prompt compression and direct fine-tuning methods aim to tackle these challenges, yet they frequently struggle to strike an optimal balance between cost-efficiency and performance effectiveness, especially in complex tasks such as NL2Code. In this paper, we propose a novel method namely PromptIntern to internalize the prompt knowledge into model parameters via progressive fine-tuning. Our method enables LLMs to emulate the human learning process for a new task, where detailed templates and examples in a prompt are gradually internalized and phased out progressively as the model grows accustomed to the task. Extensive experiments demonstrate that our method reduces inference tokens over 90%, speedups inference by 4.2 times, and saves 88.3% monetary cost.

The proliferation of large language models (LLMs) in the real world has come with a rise in copyright cases against companies for training their models on unlicensed data from the internet. Recent works have presented methods to identify if individual text sequences were members of the model's training data, known as membership inference attacks (MIAs). We demonstrate that the apparent success of these MIAs is confounded by selecting non-members (text sequences not used for training) belonging to a different distribution from the members (e.g., temporally shifted recent Wikipedia articles compared with ones used to train the model). This distribution shift makes membership inference appear successful. However, most MIA methods perform no better than random guessing when discriminating between members and non-members from the same distribution (e.g., in this case, the same period of time). Even when MIAs work, we find that different MIAs succeed at inferring membership of samples from different distributions. Instead, we propose a new dataset inference method to accurately identify the datasets used to train large language models. This paradigm sits realistically in the modern-day copyright landscape, where authors claim that an LLM is trained over multiple documents (such as a book) written by them, rather than one particular paragraph. While dataset inference shares many of the challenges of membership inference, we solve it by selectively combining the MIAs that provide positive signal for a given distribution, and aggregating them to perform a statistical test on a given dataset. Our approach successfully distinguishes the train and test sets of different subsets of the Pile with statistically significant p-values < 0.1, without any false positives.

The inference phase of Large Language Models (LLMs) is very expensive. An ideal inference stage of LLMs could utilize fewer computational resources while still maintaining its capabilities (e.g., generalization and in-context learning ability). In this paper, we try to answer the question, "During LLM inference, can we use shallow layers for easy instances; and deep layers for hard ones?" To answer this question, we first indicate that Not all Layers are Necessary during Inference by statistically analyzing the activated layers across tasks. Then, we propose a simple algorithm named AdaInfer to determine the inference termination moment based on the input instance adaptively. More importantly, AdaInfer does not alter LLM parameters and maintains generalizability across tasks. Experiments on well-known LLMs (i.e., Llama2 series and OPT) show that AdaInfer saves an average of 14.8% of computational resources, even up to 50% on sentiment tasks, while maintaining comparable performance. Additionally, this method is orthogonal to other model acceleration techniques, potentially boosting inference efficiency further.

There is growing demand for performing inference with hundreds of thousands of input tokens on trained transformer models. Inference at this extreme scale demands significant computational resources, hindering the application of transformers at long contexts on commodity (i.e not data center scale) hardware. To address the inference time costs associated with running self-attention based transformer language models on long contexts and enable their adoption on widely available hardware, we propose a tunable mechanism that reduces the cost of the forward pass by attending to only the most relevant tokens at every generation step using a top-k selection mechanism. We showcase the efficiency gains afforded by our method by performing inference on context windows up to 1M tokens using approximately 16GB of GPU RAM. Our experiments reveal that models are capable of handling the sparsity induced by the reduced number of keys and values. By attending to less than 2% of input tokens, we achieve over 95% of model performance on common benchmarks (RULER, AlpacaEval, and Open LLM Leaderboard).

Inference of Large Language Models (LLMs) across computer clusters has become a focal point of research in recent times, with many acceleration techniques taking inspiration from CPU speculative execution. These techniques reduce bottlenecks associated with memory bandwidth, but also increase end-to-end latency per inference run, requiring high speculation acceptance rates to improve performance. Combined with a variable rate of acceptance across tasks, speculative inference techniques can result in reduced performance. Additionally, pipeline-parallel designs require many user requests to maintain maximum utilization. As a remedy, we propose PipeInfer, a pipelined speculative acceleration technique to reduce inter-token latency and improve system utilization for single-request scenarios while also improving tolerance to low speculation acceptance rates and low-bandwidth interconnects. PipeInfer exhibits up to a 2.15times improvement in generation speed over standard speculative inference. PipeInfer achieves its improvement through Continuous Asynchronous Speculation and Early Inference Cancellation, the former improving latency and generation speed by running single-token inference simultaneously with several speculative runs, while the latter improves speed and latency by skipping the computation of invalidated runs, even in the middle of inference.

Inference-time techniques are emerging as highly effective tools to enhance large language model (LLM) capabilities. However, best practices for developing systems that combine these techniques remain underdeveloped due to our limited understanding of the utility of individual inference-time techniques and the interactions between them. Additionally, efficiently and automatically searching the space of model choices, inference-time techniques, and their compositions is challenging due to the large design space. To address these challenges, we introduce Archon, a modular framework for selecting, combining, and stacking layers of inference-time techniques to construct optimized LLM systems for target benchmarks. Rather than relying on a single LLM called once, we leverage a diverse set of LLMs and inference-time techniques, creating LLM systems greater than the sum of their parts. Archon defines an extensible design space, encompassing techniques such as generation ensembling, repeated sampling, ranking, fusion, critiquing, verification, and unit testing. It transforms the problem of building LLM systems into a hyperparameter optimization objective. Given the available LLMs, inference-time techniques, and compute budget, Archon utilizes hyperparameter search techniques to discover optimized architectures for target benchmark(s). We evaluate Archon architectures across a range of instruction-following, reasoning, and coding benchmarks, including MT-Bench, Arena-Hard-Auto, AlpacaEval 2.0, MixEval, MixEval Hard, MATH, and CodeContests. Archon architectures outperform frontier models, such as GPT-4o and Claude 3.5 Sonnet, on these benchmarks, achieving an average accuracy increase of 15.1 percentage points by using all available LLMs. We make our code and datasets available publicly on Github: https://github.com/ScalingIntelligence/Archon.

Language models generate responses by producing a series of tokens in immediate succession: the (K+1)^{th} token is an outcome of manipulating K hidden vectors per layer, one vector per preceding token. What if instead we were to let the model manipulate say, K+10 hidden vectors, before it outputs the (K+1)^{th} token? We operationalize this idea by performing training and inference on language models with a (learnable) pause token, a sequence of which is appended to the input prefix. We then delay extracting the model's outputs until the last pause token is seen, thereby allowing the model to process extra computation before committing to an answer. We empirically evaluate pause-training on decoder-only models of 1B and 130M parameters with causal pretraining on C4, and on downstream tasks covering reasoning, question-answering, general understanding and fact recall. Our main finding is that inference-time delays show gains when the model is both pre-trained and finetuned with delays. For the 1B model, we witness gains on 8 of 9 tasks, most prominently, a gain of 18% EM score on the QA task of SQuAD, 8% on CommonSenseQA and 1% accuracy on the reasoning task of GSM8k. Our work raises a range of conceptual and practical future research questions on making delayed next-token prediction a widely applicable new paradigm.

Large Language Models (LLMs) have exhibited remarkable performance on reasoning tasks. They utilize autoregressive token generation to construct reasoning trajectories, enabling the development of a coherent chain of thought. In this work, we explore the impact of individual tokens on the final outcomes of reasoning tasks. We identify the existence of ``critical tokens'' that lead to incorrect reasoning trajectories in LLMs. Specifically, we find that LLMs tend to produce positive outcomes when forced to decode other tokens instead of critical tokens. Motivated by this observation, we propose a novel approach - cDPO - designed to automatically recognize and conduct token-level rewards for the critical tokens during the alignment process. Specifically, we develop a contrastive estimation approach to automatically identify critical tokens. It is achieved by comparing the generation likelihood of positive and negative models. To achieve this, we separately fine-tune the positive and negative models on various reasoning trajectories, consequently, they are capable of identifying identify critical tokens within incorrect trajectories that contribute to erroneous outcomes. Moreover, to further align the model with the critical token information during the alignment process, we extend the conventional DPO algorithms to token-level DPO and utilize the differential likelihood from the aforementioned positive and negative model as important weight for token-level DPO learning.Experimental results on GSM8K and MATH500 benchmarks with two-widely used models Llama-3 (8B and 70B) and deepseek-math (7B) demonstrate the effectiveness of the propsoed approach cDPO.

Recent Large Reasoning Models (LRMs), such as DeepSeek-R1 and OpenAI o1, have demonstrated strong performance gains by scaling up the length of Chain-of-Thought (CoT) reasoning during inference. However, a growing concern lies in their tendency to produce excessively long reasoning traces, which are often filled with redundant content (e.g., repeated definitions), over-analysis of simple problems, and superficial exploration of multiple reasoning paths for harder tasks. This inefficiency introduces significant challenges for training, inference, and real-world deployment (e.g., in agent-based systems), where token economy is critical. In this survey, we provide a comprehensive overview of recent efforts aimed at improving reasoning efficiency in LRMs, with a particular focus on the unique challenges that arise in this new paradigm. We identify common patterns of inefficiency, examine methods proposed across the LRM lifecycle, i.e., from pretraining to inference, and discuss promising future directions for research. To support ongoing development, we also maintain a real-time GitHub repository tracking recent progress in the field. We hope this survey serves as a foundation for further exploration and inspires innovation in this rapidly evolving area.

Reasoning methods such as chain-of-thought prompting and self-consistency have shown immense potential to improve the accuracy of large language models across various reasoning tasks. However such methods involve generation of lengthy reasoning chains, which substantially increases token consumption, computational cost, and latency. To address this inefficiency, we propose ProofSketch, a verification-guided reasoning framework that integrates symbolic closure computation, lexicographic verification and adaptive sketch generation. Our experiments show that ProofSketch consistently reduces token usage while improving accuracy, demonstrating that this approach offers a promising path for efficient and trustworthy reasoning.

Carefully engineered system prompts play a critical role in guiding the behavior of LLM agents, but their considerable length introduces significant drawbacks, including increased inference latency, higher computational cost, and reduced effective context length. This raises the question of whether such lengthy prompts can be replaced by a drastically reduced number of tokens while preserving their behavioral effect on downstream tasks. To enable this, we propose a lightweight three-stage training framework that learns a single prompt-specific Behavior-Equivalent token ([BE]). The framework first trains [BE] to encode the natural-language content of the original system prompt via reconstruction, and then distills the prompt 's downstream behavior into this single token. Importantly, our method requires no access to model internals, no auxiliary compression models, and no labeled responses. Empirical evaluations on three datasets show that a single [BE] token achieves up to a 3000x reduction in prompt length, while retaining about 98% of the downstream performance of the original system prompts. This substantially reduces inference cost and leaves almost the entire context window available for user inputs.

Being prompted to engage in reasoning has emerged as a core technique for using large language models (LLMs), deploying additional inference-time compute to improve task performance. However, as LLMs increase in both size and adoption, inference costs are correspondingly becoming increasingly burdensome. How, then, might we optimize reasoning's cost-performance tradeoff? This work introduces a novel approach based on computational models of metareasoning used in cognitive science, training LLMs to selectively use intermediate reasoning steps only when necessary. We first develop a reward function that incorporates the Value of Computation by penalizing unnecessary reasoning, then use this reward function with Expert Iteration to train the LLM. Compared to few-shot chain-of-thought prompting and STaR, our method significantly reduces inference costs (20-37\% fewer tokens generated across three models) while maintaining task performance across diverse datasets.

Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by "some" as entailments. For some presupposition triggers like "only", BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.