ICML 2026 论文列表

LLMs achieve remarkable multi-step reasoning capabilities, yet effectively transferring these skills via post-training distillation remains challenging. Existing data selection methods, ranging from manual curation to heuristics based on length, entropy, or overall loss, fail to capture the causal importance of individual reasoning steps, limiting distillation efficiency. To address this, we propose Attention Influence for Reasoning (AIR), a principled, unsupervised and training-free framework that leverages mechanistic insights of the retrieval head to select high-value post-training data. AIR first identifies reasoning-critical attention heads of an off-the-shelf model, then constructs a weakened reference model with disabled head influence, and finally quantifies the resulting loss divergence as the Attention Influence Score. This score enables fine-grained assessment at both the step and sample levels, supporting step-level weighted fine-tuning and global sample selection. Experiments across multiple reasoning benchmarks show that AIR consistently improves reasoning accuracy, surpassing heuristic baselines and effectively isolating the most critical steps and samples. Our work establishes a mechanism-driven, data-efficient approach for reasoning distillation in LLMs.

This paper identifies a critical yet underexplored challenge in reasoning alignment from multiple multi-modal large language models (MLLMs): In non-stationary environments, the diverse reasoning distributions of source models often evolve unpredictably, transmitting systematic biases and drift to the target model. To address this, we formulate multi-source reasoning alignment as a constraint satisfaction problem under concept drift theory. We propose Autonomous Preference Optimization (APO), a novel framework that treats inter-model divergences not as noise, but as dynamic negative constraints. APO operates via a two-stage protocol: first, supervised bootstrapping projects the target model into the capability union of source models; second, constraint-aware optimization synthesizes a consistent consensus manifold by explicitly suppressing drifting trajectories via a multi-negative Plackett-Luce objective. Extensive experiments on chest X-ray interpretation demonstrate that our 7B model achieves superior robustness, outperforming even proprietary source models in average accuracy. Furthermore, we release CXR-MAX, a large-scale benchmark comprising 170,982 reasoning trajectories from seven large-scale MLLMs to facilitate research on reasoning alignment under drift. Code and data are available at: https://github.com/XiaoyuYoung/APO.

In single-stream autoregressive interfaces, the same tokens both update the model state and constitute an irreversible public commitment. This coupling creates a \emph{silence tax}: additional deliberation postpones the first \emph{task-relevant} content, while naive early streaming risks premature commitments that bias subsequent generations. We introduce \textbf{\emph{Side-by-Side (SxS)}} Interleaved Reasoning, which makes \emph{disclosure timing} a controllable decision within standard autoregressive generation. SxS interleaves partial disclosures with continued private reasoning in the same context, but releases content only when it is \emph{supported} by the reasoning so far. To learn such pacing without incentivizing filler, we construct entailment-aligned interleaved trajectories by matching answer prefixes to supporting reasoning prefixes, then train with SFT to acquire the dual-action semantics and RL to recover reasoning performance under the new format. Across two Qwen3 architectures/scales (MoE \textbf{Qwen3-30B-A3B}, dense \textbf{Qwen3-4B}) and both in-domain (AIME25) and out-of-domain (GPQA-Diamond) benchmarks, SxS improves accuracy--\emph{content-latency} Pareto trade-offs under token-level proxies (e.g., inter-update waiting).

Despite the advancing reasoning capabilities of large language models (LLMs), many reasoning evaluations are increasingly compromised by data contamination, which induces unreliable **contaminated reasoning** on leaked inputs. While this phenomenon is widely observed, its underlying mechanism remains poorly understood, hindering the ability to distinguish generalization from memorization and to develop effective solutions. In this work, we first identify a distinctive signal of contaminated reasoning, namely the mutual information decay between representations and gradients. Our mechanistic analysis reveals that contaminated models exhibit pronounced eigenspectrum concentration in their representations, leading to a low-dimensional computation regime. Under leaked inputs, this mechanism weakens the linear coupling between representations and gradients, manifested as a structural decay of the singular values in the whitened space. We show that this narrowing geometry mathematically implies a reduction in mutual information, and further demonstrate the practical utility of our analysis by successfully restoring the reasoning behavior of contaminated models, achieving up to an 11.03% absolute improvement in average consistency with the base model over the strongest baseline. Our code is available at https://github.com/jiakuan929/ContamReasoning.

As Large Language Models (LLMs) gain agentic abilities, they will have to navigate complex multi-agent scenarios, interacting with human users and other agents in cooperative and competitive settings. This will require new reasoning skills, a crucial one being theory of mind (ToM), or the ability to reason about the ``mental'' states of other agents. However, ToM and other multi-agent abilities in LLMs are poorly understood, since existing benchmarks suffer from narrow scope, data leakage, saturation, and lack of interactivity. We thus propose Decrypto, a game-based benchmark for multi-agent reasoning and ToM drawing inspiration from cognitive science, computational pragmatics and multi-agent reinforcement learning. It is designed to be as easy as possible in all other dimensions, eliminating confounding factors common in other benchmarks. To our knowledge, it is also the first platform that isolates ToM evaluation in an interactive setting. We validate the benchmark design through comprehensive empirical evaluations of frontier LLMs, robustness studies, and human-AI cross-play experiments. We find that LLMs lag behind humans and simple word-embedding baselines on key game metrics. We then create variants of two classic cognitive science experiments within Decrypto to evaluate three distinct ToM abilities. Surprisingly, our results show that state-of-the-art reasoning models are significantly worse at those tasks than their older counterparts. This demonstrates that Decrypto addresses a crucial gap in current reasoning and ToM evaluations, and paves the path towards better artificial agents. Code at https://github.com/facebookresearch/decrypto.

Most video reasoning models only generate textual reasoning traces without indicating when and where key evidence appears. Recent models such as OpenAI-o3 have sparked wide interest in evidence-centered reasoning for images, yet extending this ability to videos is more challenging due to the need for joint temporal tracking and spatial localization across dynamic scenes. We introduce Open-o3-Video, a non-agent framework that integrates explicit spatio-temporal evidence into video reasoning by highlighting key timestamps, objects, and bounding boxes, making the reasoning process traceable and verifiable. To enable this capability, we first construct high-quality datasets STGR that provide unified spatio-temporal supervision, which is absent in existing resources. We further adopt a cold-start reinforcement learning strategy with specially designed rewards that jointly encourage answer accuracy, temporal alignment, and spatial precision. On the V-STAR benchmark, Open-o3-Video achieves state-of-the-art performance, improving mAM by 14.4% and mLGM by 24.2% over the Qwen2.5-VL baseline, and shows consistent gains across a range of video understanding benchmarks. Beyond accuracy, the grounded reasoning traces produced by Open-o3-Video support confidence-aware test-time scaling, improving answer reliability. The code, model and datasets are publicly available at https://marinero4972.github.io/projects/Open-o3-Video/.

Existing multimodal reasoning approaches predominantly follow two paradigms: converting visual inputs into text prior to reasoning, or performing end-to-end reasoning within a unified vision–language representation space. Despite their empirical progress, both paradigms suffer from fundamental structural limitations. The former relies on static visual-to-text conversion, which tends to compress and lose fine-grained visual details. The latter is prone to linguistic dominance induced by joint optimization and attention mechanisms, leading to systematically weakened faithfulness to visual evidence during reasoning. In this work, we argue that a central challenge is how and when visual evidence is introduced into the reasoning process. Motivated by this insight, we propose CSMR, a multimodal reasoning framework in which a language model controls the reasoning process by deciding when to invoke an independent visual perception module to acquire task-relevant visual evidence. Experiments across multiple multimodal reasoning benchmarks show that CSMR consistently outperforms representative baseline methods in accuracy under a zero-shot setting. Further experimental analysis confirms that these advantages primarily arise from the proposed cognitive scheduling mechanism.

Scientific problem solving poses unique challenges for LLMs, requiring both deep domain knowledge and the ability to apply such knowledge through complex reasoning. While automated scientific reasoners hold great promise for assisting human scientists, there is currently no widely adopted holistic benchmark for evaluating scientific reasoning, and few approaches systematically disentangle the distinct roles of knowledge and reasoning in these tasks. To address these gaps, we introduce **SciReas**, a diverse suite of existing benchmarks for scientific reasoning tasks, and **SciReas-Pro**, a selective subset that requires more complex reasoning. Our holistic evaluation surfaces insights about scientific reasoning performance that remain hidden when relying on individual benchmarks alone. We then propose **KRUX**, a probing framework for studying the distinct roles of reasoning and knowledge in scientific tasks. Combining the two, we conduct an in-depth analysis that yields several key findings: (1) Retrieving task-relevant knowledge from model parameters is a critical bottleneck for LLMs in scientific reasoning; (2) Reasoning models consistently benefit from external knowledge added in-context on top of the reasoning enhancement; (3) Enhancing verbalized reasoning improves LLMs' ability to surface task-relevant knowledge.

Masked diffusion language models decode by iteratively unmasking tokens, where the unmasking order defines an ``order of thought'' that strongly influences generation quality yet is typically chosen heuristically. We derive a tractable upper bound on the sequential decoding mismatch, measured by the Kullback–Leibler divergence and expressed in terms of the model’s pathwise log-likelihood, with tightness under sufficient model expressivity. This bound induces a dense self-aware reward for a target sequence $x$ and unmasking order $\sigma$, over ordered paths, casting order selection as a principled policy optimization problem with a frozen denoiser. We instantiate this idea as **Self-Aware Scheduling (SAS)**, which learns a lightweight order policy using Group Relative Policy Optimization and applies seamlessly to both sequential and semi-autoregressive decoding. On Sudoku with 1B MDM, SAS improves puzzle accuracy from $82.0\%$ (best heuristic schedule) to $91.8\%$, and reaches $97.9\%$ with second-stage fine-tuning along learned trajectories. On LLaDA-8B, SAS improves pass@1 on GSM8K from $64\%$ to $76\%$ (full diffusion) and on MBPP from $39.5\%$ to $41\%$, while consistently matching or exceeding heuristic schedules across generation lengths and block sizes.

Large language models (LLMs) demonstrate strong capabilities in high-level semantic reasoning and strategic planning, making them appealing for complex decision-making tasks; however, their quantitative reasoning remains unreliable despite recent progress in tool-augmented and structured inference. To address this limitation, we decouple reasoning from computation by assigning LLMs to planning, analysis, and result interpretation, while delegating numerical computation and statistical inference to specialized external tools. Rather than being hard-coded, these tools are constructed in a constrained and structured manner during planning as explicit intermediate reasoning artifacts, enabling adaptive and scenario-dependent quantitative reasoning. LLMs iteratively analyze tool outputs under diverse market conditions and leverage performance-based feedback to refine subsequent tool selection and construction, thereby forming a bounded self-evolving loop. We instantiate this process through self-play in a controllable digital twin market, DecoupledMarket, where LLM agents continuously test, compare, and adapt their strategies. By coupling high-level planning with robust quantitative execution, the proposed framework enhances the quantitative reliability of LLM-driven decision-making. All code and data are available at https://github.com/MTMQuantAI/Agent-Trading-Arena.git.

Reasoning is a core capability of language models (LMs), yet it remains unclear how much model capacity is necessary to support reasoning during pretraining. In this work, we study the minimal parameter budget required for implicit reasoning, defined as the ability to infer new facts from learned knowledge without explicit chain-of-thought supervision. To isolate this phenomenon, we pretrain LMs from scratch in a controlled synthetic environment that mimics the structure and distribution of real-world knowledge graphs, and evaluate their ability to complete missing edges via multi-hop inference. From both a theoretical and an empirical perspective, we identify a scaling law linking this optimal parameter budget to a graph search entropy measure. Across a wide range of model sizes, training steps, and graph complexities, we show that an optimally sized language model can reliably reason over approximately 0.008 bits of information per parameter at most. Our results characterize the minimal sufficient capacity for implicit reasoning during pretraining. Our findings provide principled guidance for matching model size to data complexity and offer new insights into the scaling behavior of reasoning in large language models.

Vision-language models (VLMs) now support both direct Instruct and explicit-reasoning Thinking modes, yet practitioners still lack principled ways to decide when reasoning actually improves performance, or how much computation to spend at test time, so we investigate whether VLMs encode meta-cognitive signals for adaptive inference. We derive oracle labels for two properties: (1) reasoning helpfulness, namely whether explicit reasoning improves accuracy, and (2) desired generation length, the minimal token budget needed for a correct answer. Probing final-layer representations in InternVL and Qwen3-VL models, we find Thinking models encode these signals more linearly than Instruct models, suggesting that reasoning-oriented training enhances meta-cognitive structure. Head-wise attribution reveals two circuits: length-control heads in lower layers and reasoning/difficulty heads in higher layers. Causal interventions confirm these roles across benchmarks: scaling length heads controls output length with minimal accuracy loss, while scaling reasoning heads enables a perception-reasoning trade-off, improving accuracy by up to 5.3\%. Our results demonstrate that reasoning-tuned VLMs develop localized, manipulable circuits for meta-cognitive control, enabling test-time steering of computation and reasoning without retraining.

Large Reasoning Models (LRMs) are increasingly integrated into systems requiring reliable multi-step inference, yet this growing dependence exposes new vulnerabilities related to computational availability. In particular, LRMs exhibit a tendency to “overthink’’—producing excessively long and redundant reasoning traces—when confronted with incomplete or logically inconsistent inputs. This behavior significantly increases inference latency and energy consumption, forming a potential vector for denial-of-service (DoS) style resource exhaustion. In this work, we investigate this attack surface and propose an automated black-box framework that induces overthinking in LRMs by systematically perturbing the logical structure of input problems. Our method employs a hierarchical genetic algorithm (HGA) operating on structured problem decompositions, and optimizes a composite fitness function designed to maximize both response length and reflective overthinking markers. Across four state-of-the-art reasoning models, the proposed method substantially amplifies output length, achieving up to a 26.1× increase on the MATH benchmark and consistently outperforming benign and manually crafted missing-premise baselines. We further demonstrate strong transferability, showing that adversarial inputs evolved using a small proxy model retain high effectiveness against large commercial LRMs. These findings highlight overthinking as a shared and exploitable vulnerability in modern reasoning systems, underscoring the need for more robust defenses.

Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex tasks but suffer from high computational costs and latency. While selective thinking strategies improve efficiency by routing easy queries to non-thinking models, existing approaches often incur uncontrollable errors, especially in online settings where the performance loss of a non-thinking model is only partially observed and data are non-stationary. To address this, we propose *Betting Probably Approximately Correct* (B-PAC) *reasoning*, a principled method that enables anytime safe and efficient online reasoning under partial feedback. Specifically, we utilize inverse propensity scoring estimators to construct test supermartingales for candidate thresholds, and then dynamically adjust the routing threshold based on the accumulated statistical evidence of safety. Theoretically, we establish the anytime-valid performance loss control and the efficiency of B-PAC reasoning. Extensive experiments demonstrate that B-PAC reasoning significantly reduces computational overhead, decreasing thinking model usage by up to 81.01\%, while controlling the performance loss below the user-specified level.

Reinforcement learning (RL) has demonstrated potential for enhancing reasoning in large language models (LLMs). However, effective RL training, which requires medium-difficulty training samples, faces two fundamental challenges: Effective Data Scarcity and Dynamic Difficulty Shifts, where medium-difficulty samples are scarce and become trivial as models improve. Existing methods mitigate this scarcity to some extent by generating training samples. However, these approaches suffer from anchor-free generation, ignoring co-evolution, and difficulty mismatch. To address these issues, we propose D²Evo, a Dual Difficulty-aware self-Evolution RL framework. In each iteration, our method mines medium-difficulty anchors based on the current Solver's capability, trains the Questioner to generate diverse questions at appropriate difficulty levels, and jointly optimizes both components to enable progressive reasoning gains. Extensive experiments demonstrate that D²Evo outperforms existing methods on mathematical reasoning benchmarks with fewer than 2K real mathematical samples, and exhibits strong generalization on general reasoning benchmarks.

Extended chain-of-thought reasoning can degrade performance on deterministic state-tracking tasks, not due to preference biases, but limits rooted in the information-theoretic capacity of decoder-only attention. We establish: (1) an Attention Bottleneck Theorem with a complementary achievability construction, bounding state-tracking capacity as $O(H \cdot \log(L/H) \cdot \sqrt{d_h})$; (2) a context-dependent error model yielding super-exponential accuracy decay; (3) the State-Space Jaccard metric distinguishing capability from preference failures; (4) a Deterministic Horizon $d^* \in [19, 31]$ beyond which tool delegation becomes necessary. Across 12 models and 8 task domains (including SWE-Bench, WebArena, and SQL-Multi), tool-integrated reasoning consistently outperforms neural chain-of-thought; on the primary model suite it reaches 86–94% accuracy versus 24–42% for neural chain-of-thought. Fine-tuning on optimal-length traces yields $<$5% improvement, confirming an architectural ceiling, and high cross-model correlation ($r = 0.81$–$0.91$) indicates these failures are architectural rather than training-specific. Our results provide principled guidance for when pure neural reasoning should yield to hybrid approaches in agentic systems.

Large Language Models (LLMs) have achieved strong complex reasoning capabilities through Chain-of-Thought (CoT) reasoning. However, their reasoning patterns remain too complicated to analyze. While Sparse Autoencoders (SAEs) have emerged as a powerful tool for interpretability, existing approaches predominantly operate at the token level, creating a granularity mismatch when capturing more critical step-level information, such as reasoning direction and semantic transitions. In this work, we propose step-level sparse autoencoder (SSAE), which serves as an analytical tool to disentangle different aspects of LLMs' reasoning steps into sparse features. Specifically, by precisely controlling the sparsity of a step feature conditioned on its context, we form an information bottleneck in step reconstruction, which splits incremental information from background information and disentangles it into several sparsely activated dimensions. Experiments on multiple base models and reasoning tasks show the effectiveness of the extracted features. By linear probing, we can easily predict surface-level information, such as generation length and first token distribution, as well as more complicated properties, such as the correctness and logicality of the step. These observations indicate that LLMs should already at least partly know about these properties during generation, which provides the foundation for the self-verification ability of LLMs. Our code is available at https://github.com/Miaow-Lab/SSAE.

Chain-of-Thought (CoT) reasoning successfully enhances the reasoning capabilities of Large Language Models (LLMs), yet it incurs substantial computational overhead for inference. Existing CoT compression methods often suffer from a critical loss of logical fidelity at high compression ratios, resulting in significant performance degradation. To achieve high-fidelity, fast reasoning, we propose a novel EXTreme-RAtio Chain-of-Thought Compression framework, termed Extra-CoT, which aggressively reduces the token budget while preserving answer accuracy. To generate reliable, high-fidelity supervision, we first train a dedicated semantically-preserved compressor on mathematical CoT data with fine-grained annotations. An LLM is then fine-tuned on these compressed pairs via a mixed-ratio supervised fine-tuning (SFT), teaching it to follow a spectrum of compression budgets and providing a stable initialization for reinforcement learning (RL). We further propose Constrained and Hierarchical Ratio Policy Optimization (CHRPO) to explicitly incentivize question-solving ability under lower budgets by a hierarchical reward. Experiments on three mathematical reasoning benchmarks show the superiority of Extra-CoT. For example, on MATH-500 using Qwen3-1.7B, Extra-CoT achieves over 73\% token reduction with an accuracy improvement of 0.6\%, significantly outperforming state-of-the-art (SOTA) methods. Our source codes are released in the Supplementaries.

Self-correction in language models remains elusive. In this work, we explore whether language models can explicitly localize errors in incorrect reasoning, as a path toward building AI systems that can effectively correct themselves. We introduce a prompting method that structures reasoning as discrete, semantically coherent thought steps, and show that models can localize errors more reliably within this structure than in conventional, unstructured chain-of-thought reasoning. Motivated by how the human brain monitors errors at discrete decision points and resamples alternatives, we introduce Iterative Correction Sampling of Thoughts (Thought-ICS), a self-correction framework. Thought-ICS iteratively prompts the model to generate reasoning one discrete and complete thought at a time—where each thought represents a deliberate decision by the model—creating natural boundaries for precise error localization. Upon verification, the model localizes the first erroneous step, and the system backtracks to generate alternative reasoning from the last correct point. When asked to correct reasoning verified as incorrect by an oracle, Thought-ICS achieves 20-40% self-correction lift. In a completely autonomous setting without external verification, it outperforms contemporary self-correction baselines.