ICML 2026 论文列表

While Large Language Models (LLMs) excel in certain reasoning tasks, they struggle in multi-agent games where the final outcome depends on the joint strategies of all agents. In multi-agent games, the non-stationarity of other agents brings significant challenges on the evaluation of the reasoning process and the credit assignment over multiple reasoning steps. Existing single-agent reinforcement learning (RL) approaches and their multi-agent extensions fail to address these challenges as they do not incorporate other agents in the reasoning process. In this work, we propose Strat-Reasoner, a novel RL-based framework that improves LLMs' strategic reasoning ability in multi-agent games. We introduce a novel recursive reasoning paradigm where an agent's reasoning also integrates other agents' reasoning processes. To provide effective reward signals for the intermediate reasoning sequences, we employ a centralized Chain-of-Thought (CoT) comparison module to evaluate the reasoning quality. Finally, we compute an accurate hybrid advantage and develop a group-relative RL approach to optimize the LLM policy. Experimental results show that Strat-Reasoner substantially improves strategic abilities of underlying LLMs, achieving 22.1\% average performance improvements across various multi-agent games. Code is publicly available at https://github.com/ydhe1012/Strat-Reasoner.

While visually grounded Chain-of-Thought (CoT) has emerged as a promising paradigm to enhance fine-grained perception in multimodal large language models (MLLMs), its efficacy during the inference phase remains underexplored. In this work, we empirically find that mandating explicit object boxes in visually grounded CoT during inference often degrades performance compared to standard textual CoT, which reasons without explicit visual grounding. We hypothesize that the visual localization capability can be internalized into the textual CoT and that the mandatory explicit grounding introduces unnecessary interference with the model's primary objective of answer prediction. To address this problem, we propose Internalizing Visually Grounded Reasoning (**iVGR**), a novel reinforcement learning framework that transfers localization capabilities into the textual reasoning process. We employ a dual-stream training strategy, where a textual stream is aligned with a high-quality visually grounded stream via a proposed consistency reward, enabling the model to localize accurately without explicit grounding during inference. Extensive experiments demonstrate that our method significantly outperforms existing baselines on fine-grained benchmarks, while maintaining the flexibility to support tool-assisted inference workflows. Project page: https://visual-ai.github.io/ivgr/

Transformers can acquire Chain-of-Thought (CoT) capabilities to solve reasoning tasks via fine-tuning. Reinforcement learning (RL) and supervised fine-tuning (SFT) are two primary approaches to this end. In this work, we examine RL with verifiable process rewards and SFT for learning $k$-sparse Boolean functions with a one-layer transformer through intermediate reasoning steps akin to CoT. In particular, we consider Boolean functions that can be recursively decomposed into fixed 2-sparse Boolean functions. We first analyze the learning dynamics of RL fine-tuning with verifiable process rewards and SFT in a unified way, allowing us to identify sufficient conditions under which the transformer provably learns these functions. We then verify that these conditions hold for three basic examples, including $k$-PARITY, $k$-AND, and $k$-OR, thus demonstrating their learnability via both RL and SFT. Notably, we reveal that RL and SFT exhibit distinct learning behaviors: RL learns the whole CoT chain simultaneously, whereas SFT learns the CoT step-by-step. Overall, our findings provide insights on the mechanisms underlying RL and SFT and how they differ in triggering the CoT capabilities of transformers, and suggest that the comparison between RL and SFT may need to consider the reward design and the use of teacher forcing.

Current multimodal benchmarks often conflate reasoning with domain knowledge, making it difficult to isolate and evaluate reasoning abilities in non-expert settings. To address this, we introduce VisualPuzzles, a benchmark that targets visual reasoning while deliberately minimizing reliance on specialized knowledge. VisualPuzzles consists of diverse puzzles spanning 5 categories: algorithmic, analogical, deductive, inductive, and spatial reasoning. One major source of our questions is manually translated logical reasoning questions from the Chinese Civil Service Examination. Experiments show that VisualPuzzles requires significantly less intensive domain-specific knowledge and more complex reasoning compared to benchmarks like MMMU, enabling us to better evaluate genuine multimodal reasoning. Evaluations show that state-of-the-art multimodal large language models consistently lag behind human performance on VisualPuzzles, and that strong performance on knowledge-intensive benchmarks does not necessarily translate to success on reasoning-focused, knowledge-light tasks. Additionally, reasoning enhancements such as scaling up inference compute (with "thinking" modes) yield inconsistent gains across models and task types, and we observe no clear correlation between model size and performance. We also found that models exhibit different reasoning and answering patterns on VisualPuzzles compared to benchmarks with heavier emphasis on knowledge. VisualPuzzles offers a clearer lens through which to evaluate reasoning capabilities beyond factual recall and domain knowledge.

Current chemical large language models (LLMs) predominantly rely on explicit Chain-of-Thought (CoT) to solve complex reasoning problems. However, forcing nonverbal tacit chemical logic into discrete natural language imposes a fundamental ``modality mismatch,'' creating an artificial bottleneck for reasoning. We introduce LatentChem, a reasoning interface that decouples chemical logic from linguistic generation, enabling the model to process information via continuous thought vectors and dynamic perception. Our investigation reveals a pivotal emergent behavior: spontaneous internalization, defined here as self-selected under outcome-only optimization. When optimized for task success, the model abandons verbose textual derivations in favor of implicit latent computation, suggesting that it identifies the continuous manifold as a more native substrate for chemical logic. This paradigm shift also proves to be a superior computational strategy: LatentChem achieves a 59.88\% non-tie win rate against the strong CoT baseline on the rigorous ChemCoTBench, while delivering a broad 10.84$\times$ average reduction in reasoning step overhead (5.96$\times$ wall-clock speedup) across all evaluated benchmarks. Our results provide empirical evidence that chemical reasoning is more naturally and effectively realized as continuous latent dynamics rather than discretized linguistic trajectories.

Long-form video understanding remains a formidable challenge due to the complexity of modeling long-range temporal dependencies and multi-event narratives. Existing methods often rely on static reasoning or external Visual-Language Models (VLMs), resulting in high computational complexity and sub-optimal performance. In this paper, we propose Video-MTR, a reinforced multi-turn reasoning framework that operates solely through data-efficient, pure RL post-training. Video-MTR reformulates video understanding as a dynamic decision-making process, where the agent iteratively selects key segments conditioned on the evolving context of previously processed frames and the query. To ensure effective intermediate reasoning and training stability, we introduce a novel gated bi-level reward system, which synergizes trajectory-level rewards (answer correctness) with turn-level rewards (frame-query relevance). This mechanism eliminates the need for data-intensive supervised fine-tuning, thereby substantially reducing reliance on large-scale datasets. Remarkably, Video-MTR achieves competitive or superior performance using only $\sim$8K training samples, compared to existing approaches that require 257K to 4.4M examples. Extensive experiments on benchmarks including VideoMME, MLVU, LongVideoBench, LVBench, and EgoSchema demonstrate that Video-MTR surpasses state-of-the-art methods in both accuracy and efficiency. Code is available at https://github.com/Xyuan13/Video-MTR.

Large Language Models (LLMs) are pretrained on massive datasets and later instruction-tuned via supervised fine-tuning (SFT) or reinforcement learning (RL). Best practices emphasize large, diverse pretraining data, whereas post-training operates differently: SFT relies on smaller, high-quality datasets, while RL benefits more from scale, with larger amounts of feedback often outweighing label quality. Yet it remains unclear why pretraining and RL require large datasets, why SFT excels on smaller ones, and what defines high-quality SFT data. In this work, we theoretically analyze transformers trained on an in-context weight prediction task for linear regression. Our analysis reveals several key findings: $(i)$ balanced pretraining data can induce latent capabilities later activated during post-training, and $(ii)$ SFT learns best from a small set of examples challenging for the pretrained model, while excessively large SFT datasets may dilute informative pretraining signals. In contrast, RL is most effective on large-scale data that is not overly difficult for the pretrained model. We validate these theoretical insights with experiments on large nonlinear transformer architectures.

Compressing long chains of thought (CoT) into compact latent tokens is crucial for efficient reasoning with large language models (LLMs). Recent studies employ autoencoders to achieve this by reconstructing textual CoT from latent tokens, thus encoding CoT semantics. However, treating textual CoT as the reconstruction target forces latent tokens to preserve surface-level linguistic features (e.g., word choice and syntax), introducing a strong linguistic inductive bias that prioritizes linguistic form over reasoning structure and limits logical abstraction. Thus, we propose ImgCoT that replaces the reconstruction target from textual CoT to the visual CoT obtained by rendering CoT into images. This substitutes linguistic bias with spatial inductive bias, i.e., a tendency to model spatial layouts of the reasoning steps in visual CoT, enabling latent tokens to better capture global reasoning structure. Moreover, although visual latent tokens encode abstract reasoning structure, they may blur reasoning details. We thus propose a loose ImgCoT, a hybrid reasoning that augments visual latent tokens with a few key textual reasoning steps, selected based on low token log-likelihood. This design allows LLMs to retain both global reasoning structure and fine-grained reasoning details with fewer tokens than the complete CoT. Extensive experiments across multiple datasets and LLMs demonstrate the effectiveness of the two versions of ImgCoT.

We propose DynVLA, a driving VLA model that introduces a new CoT paradigm termed Dynamics CoT. DynVLA forecasts compact world dynamics before action generation, enabling more informed and physically grounded decision-making. To obtain compact dynamics representations, DynVLA introduces a Dynamics Tokenizer that compresses future evolution into a small set of dynamics tokens. Considering the rich environment dynamics in interaction-intensive driving scenarios, DynVLA decouples ego-centric and environment-centric dynamics, yielding more accurate world dynamics modeling. We then train DynVLA to generate dynamics tokens before actions through SFT and RFT, improving decision quality while maintaining latency-efficient inference. Compared to Textual CoT, which lacks fine-grained spatiotemporal understanding, and Visual CoT, which introduces substantial redundancy due to dense image prediction, Dynamics CoT captures the evolution of the world in a compact, interpretable, and efficient form. Extensive experiments on NAVSIM, Bench2Drive, and a large-scale in-house dataset demonstrate that DynVLA consistently outperforms Textual CoT and Visual CoT methods, validating the effectiveness and practical value of Dynamics CoT.

Reward-maximizing RL methods have shown to be capable of enhancing the reasoning performance of LLMs, but often lead to reduced generation diversity. Recent works address this issue by adopting GFlowNets, training LLMs to match a target distribution while jointly learning its partition function. In contrast to prior works that treat this partition function solely as a normalizer, we reinterpret it as a per-prompt expected-reward (i.e., online accuracy) signal, leveraging this unused information to improve sample efficiency. Specifically, we first establish a theoretical relationship between the partition function and per-prompt accuracy estimates. Building on this key insight, we propose \textbf{Pa}rtition Fun\textbf{c}tion-Guid\textbf{ed} \textbf{RL} (PACED-RL), a post-training framework that leverages accuracy estimates to prioritize informative question prompts during training, and further improves sample efficiency through an accuracy estimate error–prioritized replay. Crucially, both components reuse information already produced during GFlowNet training, effectively amortizing the compute overhead into the existing optimization process. Extensive experiments across diverse benchmarks demonstrate strong performance improvements over GRPO and prior GFlowNet approaches, highlighting PACED-RL as a promising direction for a more sample efficient distribution-matching training for LLMs.

Large Language Models (LLMs) are increasingly deployed for knowledge synthesis, yet their capacity for compositional generalization in scientific knowledge remains under-characterized. Existing benchmarks primarily focus on single-turn restricted scenarios, failing to capture the capability boundaries exposed by real-world interactive scientific workflows. To address this, we introduce XDomainBench, a diagnostic benchmark for interactive interdisciplinary scientific reasoning. We formalize the composition order and mixture structure to enable systematic stress-testing from single-discipline to inter-disciplinary, comprising 8,598 interactive sessions across 20 domains and 4 task categories, with 8 realistic trajectory patterns covering difficulty and domain-mixture dynamics, simulating real AI4S scenarios. Large-scale evaluation of LLMs reveals a systematic reasoning collapse as composition order increases, stemming from two root causes: (i) direct difficulty increases induced by domain composition, and (ii) indirect interaction-amplified failures where trajectory patterns trigger error accumulation, reasoning breaks, and domain confusion, ultimately leading to session collapse.

Large Vision-Language Models (LVLMs) exhibit remarkable general capabilities but struggle significantly with spatial reasoning tasks. In this paper, we uncover a critical representation-output misalignment via linear probing: LVLMs correctly encode spatial features internally, but generate incorrect results in the final text. To address this, we pioneer the Inference-time Geometric Manifold Adaptation paradigm and propose **GRASP** (**G**eometric **R**ectification for **A**ctive **S**patial **P**erception), a training-free framework to awaken these latent capabilities. GRASP employs Manifold Differential Search to identify optimal geometric counterfactuals, which then drive a dual-level rectification mechanism: Implicit Trajectory Correction to rectify attenuated intrinsic geometric features in intermediate decoder layers, and Explicit Distribution Alignment to break the dominance of language priors at the output layer. Extensive experiments spanning diverse architectures (LLaVA, Qwen 2.5/3-VL) and positional encoding paradigms (1D APE, 2D/3D RoPE) across image and video benchmarks (WhatsUp, VSR, VSI-Bench) demonstrate that GRASP significantly mitigates spatial hallucinations without parameter updates, achieving accuracy gains of up to 26.1% on image benchmarks and 9.7% on video reasoning tasks, consistently outperforming baseline methods.

Self-interpretation methods prompt language models to describe their own internal states, but remain unreliable due to hyperparameter sensitivity. We show that training lightweight adapters on interpretability artifacts, while keeping the LM entirely frozen, yields reliable self-interpretation across tasks and model families. A scalar affine adapter with just $d_\text{model}+1$ parameters suffices: trained adapters generate sparse autoencoder feature labels that outperform the training labels themselves (70% vs 50% generation scoring at 70B scale), identify topics with 94% recall@1 versus 1% for untrained baselines, and decode bridge entities in multi-hop reasoning that appear in neither prompt nor response, surfacing implicit reasoning without chain-of-thought. The learned bias vector alone accounts for 85% of improvement, and simpler adapters generalize better than more expressive alternatives. Controlling for model knowledge via prompted descriptions, we find self-interpretation gains outpace capability gains from 7B to 72B parameters. Our results demonstrate that self-interpretation improves with scale, without modifying the model being interpreted.

Large Language Models (LLMs) show remarkable semantic understanding but often struggle with structural understanding when processing graph topologies in a serialized format. Existing solutions rely on training external graph-based adapters or fine-tuning, which incur high costs and lost generalizability. In this work, we investigate the internal mechanisms of LLMs and present a critical finding: *LLMs spontaneously reconstruct the graph's topology internally*, evidenced by a distinct "sawtooth" pattern in their attention maps that structurally aligns with the "token-level adjacency matrix". However, this intrinsic structural understanding is diluted by the attention sink. We theoretically formalize this dilution as a representation bottleneck, stemming from a fundamental conflict: the model's anisotropic bias, essential for language tasks, suppresses the topology-aware local aggregation required for graph reasoning. To address this, we propose a training-free solution, named **S**tructura**L** **A**ttention **SH**arpening (SLASH), which amplifies this internal structural understanding via a plug-and-play attention redistribution. Experiments on pure graph tasks and molecular prediction validate that SLASH delivers significant and consistent performance gains across diverse LLMs.

Test-time scaling methods such as majority vote aggregation and iterative refinement (e.g., self-reflection or multi-agent inference) improve reasoning performance by leveraging multiple solution samples. However, their efficacy depends not only on raw performance, but critically on the distribution of errors across samples. When errors concentrate, (a) aggregation accuracy degrades, as the majority vote may select a shared mistake, and (b) confidence in common mistakes may suppress exploration in iterative refinement. We argue that improving correctness alone is not sufficient to mitigate these issues; to this end, we propose to explicitly shape error distributions to improve aggregation. First, we introduce a theoretically grounded \textbf{diverse failure reward} that incentivizes calibrated disagreement within model errors. We prove that this reward directly optimizes majority-vote accuracy: policies achieving higher reward attain higher expected majority-vote performance, and vice versa. We further show that this theoretical property generalizes to iterative refinement. Second, we introduce \textbf{anti-votes}, in which the model predicts the most common mistake alongside its solution, allowing probability mass on dominant errors to be explicitly reweighted. We identify conditions under which anti-votes are guaranteed to improve majority-vote accuracy. Empirically, across three model families of varying sizes and four benchmarks, we show that both approaches substantially improve majority vote and iterative refinement performance without degrading single-sample accuracy.

Reinforcement Learning (RL) has become a cornerstone for enhancing the reasoning capabilities of Large Language Models (LLMs). However, standard actor-critic methods, such as PPO, rely on scalar value functions that estimate only the expectation of cumulative returns. This reduction inherently discards higher-order statistical information (e.g., variance and multimodality), leading to inaccurate value estimation and suboptimal credit assignment in complex tasks. While Distributional RL offers a solution by modeling the full return distribution, its application to LLMs remains challenging due to the computational intractability of value-based operations over large vocabularies and the instability and memory burden of off-policy replay mechanisms. In this paper, we propose DisPPO, a novel on-policy framework that seamlessly integrates non-parametric quantile regression into PPO. Theoretically, we prove that our distributional update operator---composed of the $\lambda$-return Bellman operator and quantile projection---is a contraction mapping in the Wasserstein metric, guaranteeing convergence to a unique fixed point. Empirically, we evaluate DisPPO using Llama and Qwen models across diverse benchmarks, including mathematical reasoning and Text-to-SQL generation. DisPPO consistently outperforms standard PPO and recent group-based baselines in both Pass@1 and Pass@$k$ metrics, demonstrating that distributional critics provide a richer, more robust learning signal for large-scale reasoning models.

Tree-search decoding is an effective form of test-time scaling for large language models (LLMs), but real-world deployment often imposes a fixed per-query token budget that varies across settings. Existing tree-search policies are largely budget-agnostic, treating the budget merely as a termination condition, thereby risking late-stage over-branching or premature termination. We propose Budget-Guided MCTS (BG-MCTS), a tree-search decoding algorithm that aligns its search policy with the remaining token budget: it starts with broad exploration, then prioritizes refinement and answer completion as the remaining budget decreases while reducing late-stage branching from shallow nodes. BG-MCTS consistently outperforms budget-agnostic tree-search baselines across inference budgets on mathematical reasoning benchmarks and an additional physics reasoning benchmark with open-weight LLMs.

Large language models (LLMs) demonstrate strong reasoning abilities via Chain-of-Thought (CoT), but their token-level generation encourages local decisions and lacks global planning, often leading to redundant or inaccurate reasoning. Existing methods, such as tree-based search and reinforcement learning (RL), attempt to address this issue but incur high computational costs and still struggle to produce reliable reasoning trajectories. To address these challenges, we propose Plan-Then-Action Enhanced Reasoning with Group Relative Policy Optimization (PTA-GRPO), a two-stage framework designed to jointly improve high-level planning and fine-grained CoT reasoning. Specifically, in the first stage, a given LLM is responsible for summarizing CoT reasoning into compact high-level guidance, which is then leveraged for supervised fine-tuning. Then, we introduce a guidance-aware reinforcement learning method that jointly optimizes the final output and the quality of guidance, enhancing reasoning effectiveness. We evaluate PTA-GRPO on ten reasoning benchmarks across mathematics and natural sciences, using five diverse base models spanning multiple data modalities. The results show that PTA-GRPO consistently delivers stable and significant improvements across models and tasks, demonstrating strong effectiveness and generalization.

While LLMs have seen substantial improvement in reasoning capabilities, they also sometimes overthink, generating unnecessary reasoning steps, particularly under uncertainty, given ill-posed or ambiguous queries. We introduce statistically principled early stopping methods that monitor uncertainty signals during generation to mitigate this issue. Our first approach is nonparametric and provides finite-sample guarantees on the probability of halting too early on well-posed queries. Our second approach is parametric: it models inter-arrival times of uncertainty keywords as a renewal process and applies sequential testing for stopping. We conduct empirical evaluations on reasoning tasks across several domains and models. Our results indicate that uncertainty-aware early stopping can improve both efficiency and reliability in LLM reasoning. The performance varies across domains, and we observe especially significant gains for math reasoning.