ICLR 2026 Papers

Multimodal large language models (MLLMs) promise enhanced reasoning by integrating diverse inputs such as text, vision, and audio. Yet, despite their perceptual strengths, their reasoning ability across modalities remains underexplored, with conflicting reports on whether additional modalities help or harm performance. These inconsistencies stem from a lack of controlled evaluation frameworks and analysis of models' internals to isolate when and why modality interactions support or undermine reasoning. We address this gap through a logic-grounded evaluation framework that categorizes multimodal reasoning into six interaction patterns, varying how facts are distributed across modalities and logically combined. Empirically, additional modalities enhance reasoning only when they provide independent and sufficient reasoning paths, while redundant or chained entailment support often hurts performance. Besides, models recognize cross-modal facts reliably and always reason on text effectively. Moreover, reasoning degrades in three systematic ways: weaker modalities drag down overall performance, conflicts bias preference toward certain modalities, and joint signals from different modalities fail to be integrated effectively. Therefore, we identify two core failures: task-composition bottleneck, where recognition and reasoning cannot be jointly executed in one pass, and fusion bottleneck, where early integration introduces bias. For further investigation, we find that attention patterns fail to encode fact usefulness, but a simple two-step prompting (recognize then reason) restores performance, confirming the task-composition bottleneck. Moreover, modality identity remains recoverable in early layers, and softening attention in early fusion improves reasoning, highlighting biased fusion as another failure mode. Overall, our findings show that integration, not perception, is the main barrier to multimodal reasoning, suggesting composition-aware training and early fusion control as promising directions.

Reasoning over long contexts is essential for large language models. While reinforcement learning (RL) enhances short-context reasoning by inducing "Aha" moments in chain-of-thought, the advanced thinking patterns required for long-context reasoning remain largely unexplored, and high-difficulty RL data are scarce. In this paper, we introduce LoongRL, a data-driven RL method for advanced long-context reasoning. Central to LoongRL is KeyChain, a synthesis approach that transforms short multi-hop QA into high-difficulty long-context tasks by inserting UUID chains that hide the true question among large collections of distracting documents. Solving these tasks requires the model to trace the correct chain step-by-step, identify the true question, retrieve relevant facts and reason over them to answer correctly. RL training on KeyChain data induces an emergent plan–retrieve–reason–recheck reasoning pattern that generalizes far beyond training length. Models trained at 16K effectively solve 128K tasks without prohibitive full-length RL rollout costs. On Qwen2.5-7B and 14B, LoongRL substantially improves long-context multi-hop QA accuracy by +23.5% and +21.1% absolute gains. The resulting LoongRL-14B reaches a score of 74.2, rivaling much larger frontier models such as o3-mini (74.5) and DeepSeek-R1 (74.9). It also improves long-context retrieval, passes all 128K needle-in-a-haystack stress tests, and preserves short-context reasoning capabilities.

While Large Vision-Language Models (LVLMs) have achieved substantial progress in video understanding, their application to long video reasoning is hindered by uniform frame sampling and static textual reasoning, which are inefficient and struggle to handle visually intensive video tasks. To overcome these challenges, in this paper, we introduce the concept of thinking with long videos and propose a novel framework FrameThinker. Within this framework, LVLMs are able to iteratively interrogate video content. Developing such video reasoning capabilities in LVLMs presents notable challenges, particularly in adapting the model to new video actions (e.g. select frame), and designing reward functions to guide LVLMs to adopt the newly introduced action. To solve these challenges, we propose a two-phase training strategy, first employing Supervised Fine-Tuning (SFT) to instill fundamental action capabilities, followed by Reinforcement Learning (RL) to optimize a strategic decision-making policy. Notably, in this RL phase, we conduct an in-depth and comprehensive exploration of the reward design for each action and format reward. Extensive experiments on reasoning benchmarks like Video-Holmes, LongVideo-Reason, and long-video understanding benchmarks such as LongVideoBench, MLVU, VideoMME, and LVBench, demonstrate that FrameThinker gets a significant average improvement of +10.4\% over baselines while drastically reducing the number of processed frames. Most notably, our 7B model, FrameThinker establishes a new state-of-the-art on LongVideo-Reason, achieving 76.1\% accuracy using an average of only 20.6 frames. This not only outperforms the competitive LongVILA-R1 (72.0\%) but does so with over 20x fewer frames (vs. 512), demonstrating unparalleled efficiency and effectiveness. Our code is available at: \url{https://anonymous.4open.science/r/FrameThinker-B5FD}.

The recent paradigm shift towards training large language models (LLMs) using DeepSeek-R1-Zero-style reinforcement learning (RL) on verifiable rewards has led to impressive advancements in code and mathematical reasoning. However, this methodology is limited to tasks where rule-based answer verification is possible and does not naturally extend to real-world domains such as chemistry, healthcare, engineering, law, biology, business, and economics. Current practical workarounds use an additional LLM as a model-based verifier; however, this introduces issues such as reliance on a strong verifier LLM, susceptibility to reward hacking, and the practical burden of maintaining the verifier model in memory during training. To address this and extend DeepSeek-R1-Zero-style training to general reasoning domains, we propose a verifier-free method (**VeriFree**) that bypasses answer verification and instead directly maximizes the probability of generating the reference answer, derived in a principled way from the RL objective. We compare VeriFree with verifier-based methods and demonstrate that, in addition to its significant practical benefits and reduced compute requirements, VeriFree matches and even surpasses verifier-based methods on extensive evaluations across MMLU-Pro, GPQA, SuperGPQA, and math-related benchmarks.

Chain-of-thought explanations are widely used to inspect the decision process of large language models (LLMs) and to evaluate the trustworthiness of model outputs, making them important for effective collaboration between LLMs and humans. We demonstrate that preference optimization -- a key step in the alignment phase -- can inadvertently reduce the faithfulness of these explanations. This occurs because the reward model (RM), which guides alignment, is tasked with optimizing both the expected quality of the response and the appropriateness of the explanations (e.g., minimizing bias or adhering to safety standards), creating potential conflicts. The RM lacks a mechanism to assess the consistency between the model’s internal decision process and the generated explanation. Consequently, the LLM may engage in ``reward hacking'' by producing a final response that scores highly while giving an explanation tailored to maximize reward rather than accurately reflecting its reasoning. To address this issue, we propose enriching the RM’s input with a causal attribution of the prediction, allowing the RM to detect discrepancies between the generated self-explanation and the model's decision process. In controlled settings, we show that this approach reduces the tendency of the LLM to generate misleading explanations.

Reinforcement Learning with Verifiable Rewards (RLVR) offers a robust mechanism for enhancing mathematical reasoning in large models. However, we identify a systematic lack of emphasis on more challenging questions in existing methods from both algorithmic and data perspectives, despite their importance for refining underdeveloped capabilities. Algorithmically, widely used Group Relative Policy Optimization (GRPO) suffers from an implicit imbalance where the magnitude of policy updates is lower for harder questions. Data-wise, augmentation approaches primarily rephrase questions to enhance diversity without systematically increasing intrinsic difficulty. To address these issues, we propose a two-dual MathForge framework to improve mathematical reasoning by targeting harder questions from both perspectives, which comprises a Difficulty-Aware Group Policy Optimization (DGPO) algorithm and a Multi-Aspect Question Reformulation (MQR) strategy. Specifically, DGPO first rectifies the implicit imbalance in GRPO via difficulty-balanced group advantage estimation, and further prioritizes harder questions by difficulty-aware question-level weighting. Meanwhile, MQR reformulates questions across multiple aspects to increase difficulty while maintaining the original gold answer. Overall, MathForge forms a synergistic loop: MQR expands the data frontier, and DGPO effectively learns from the augmented data. Extensive experiments show that MathForge significantly outperforms existing methods on various mathematical reasoning tasks. The code and augmented data are all available at https://github.com/AMAP-ML/MathForge.

Value trade-offs are an integral part of human decision-making and language use, however, current tools for interpreting such dynamic and multi-faceted notions of values in LLMs are limited. In cognitive science, so-called “cognitive models” provide formal accounts of such trade-offs in humans, by modeling the weighting of a speaker’s competing utility functions in choosing an action or utterance. Here we use a leading cognitive model of polite speech to systematically evaluate value trade-offs in two encompassing model settings: degrees of reasoning “effort” in frontier black-box models, and RL post-training dynamics of open-source models. Our results highlight patterns of higher informational utility than social utility in reasoning models’ default behavior, and demonstrate that these patterns shift in predictable ways when models are prompted to prioritize certain goals over others. Our findings from LLMs’ training dynamics suggest large shifts in utility values early on in training with persistent effects of the choice of base model and pretraining data, compared to feedback dataset or alignment method. We show that our method is responsive to diverse aspects of the rapidly evolving LLM landscape, with insights for forming hypotheses about other social behaviors such as sycophancy, and shaping training regimes that better control trade-offs between values during model development

Multimodal Misinformation Detection (MMD) in realistic, mixed-sourced scenarios must incorporate robust reasoning capabilities to handle the social complexity and diverse types of forgeries. While MLLM-based agents are increasingly used for MMD task due to their powerful reasoning abilities, they suffer from a critical trade-off: on one hand, single-agent methods provide only the limited, single-view analysis; on the other hand, multi-agent methods introduce high computational costs and significant optimization difficulties. To address this gap, we propose a novel Multi-Social-Agent Self-Distillation framework that internalizes collective social reasoning capabilities into a unified model. Our framework consists of two core stages: (1) we simulate multi-perspective judgments from a diverse society of MLLM agents and synthesize their collective feedback into high-quality Social Chain-of-Thought (SCoT) data; (2) Building on this, we propose the Social Correction Value-Driven Preference Optimization (SCPO), a new alignment algorithm that leverages the degree of social misjudgment as a verifiable signal to dynamically focus training on the most challenging samples. Extensive experiments on the challenging MFC-Bench and MMFakeBench benchmarks demonstrate the effectiveness of our framework. Our 7B Qwen2-VL-based model significantly outperforms various MLLM baselines, multi-agent methods, and even competes or surpasses proprietary models like GPT-4o and Claude, facilitating advanced multimodal misinformation reasoning and detection via thinking as society.

The performance of fine-tuned language models is heavily influenced by the quality and quantity of their fine-tuning data. While scaling laws suggest that larger models benefit from more data during pretraining, the Less-is-More hypothesis highlights that downstream fine-tuning often requires only a small but high-quality dataset to effectively elicit a model’s pretrained knowledge. However, identifying such premium data, particularly in terms of difficulty and diversity, typically relies on human expertise, and existing methods offer limited guidance for automatic selection from large unannotated corpora. This work presents a novel quantitative framework that formalizes the interplay between question difficulty and diversity, and introduces *Difficulty–Diversity Collaborative Filtering* (DDCF): an automated approach that tailors data selection to the unique characteristics of each language model via collaborative filtering. By leveraging a small seed dataset to predict correctness across a large unannotated corpus, our method reduces the annotation cost by $100-200\times$, while maintaining downstream performance comparable to full-corpus fine-tuning.

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet optimally integrating Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through a comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from an entropy-based perspective, we reveal key differences between these paradigms: SFT induces coarse-grained, global shifts to policy distributions, while RL performs fine-grained, selective optimizations. Our analysis further establishes entropy as a critical indicator of training efficacy. Building on these observations, we introduce **S**upervised **R**einforcement **F**ine-**T**uning (**SRFT**), a single-stage framework that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. SRFT simultaneously applies SFT and RL to directly optimize LLMs using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT outperforms zero-RL baselines by **9.0%** on five mathematical reasoning benchmarks and by **10.9%** on three out-of-distribution benchmarks. Moreover, by leveraging demonstration data, SRFT maintains a more stable policy entropy, facilitating sustained policy improvement.

Graph-based Retrieval-Augmented Generation (RAG) systems leverage interconnected knowledge structures to capture complex relationships that flat retrieval struggles with, enabling multi-hop reasoning. Yet most existing graph-based methods suffer from (i) heuristic designs lacking theoretical guarantees for subgraph quality or relevance and/or (ii) the use of static exploration strategies that ignore the query's holistic meaning, retrieving neighborhoods or communities regardless of intent. We propose *Query-Aware Flow Diffusion RAG* (QAFD-RAG), a training-free framework that dynamically adapts graph traversal to each query's holistic semantics. The central innovation is *query-aware traversal*: during graph exploration, edges are dynamically weighted by how well their endpoints align with the query's embedding, guiding flow along semantically relevant paths while avoiding structurally connected but irrelevant regions. These query-specific reasoning subgraphs enable the first statistical guarantees for query-aware graph retrieval, showing that QAFD-RAG recovers relevant subgraphs with high probability under mild signal-to-noise conditions. The algorithm converges exponentially fast, with complexity scaling with the retrieved subgraph size rather than the full graph. Experiments on question answering and text-to-SQL tasks demonstrate consistent improvements over state-of-the-art graph-based RAG methods.

Agents in the real world must make not only logical but also *timely* judgments. This requires continuous awareness of the dynamic environment: hazards emerge, opportunities arise, and other agents act, while the agent's reasoning is still unfolding. Despite advances in language model reasoning, existing approaches fail to account for this dynamic nature. We introduce *real-time reasoning* as a new problem formulation for agents in evolving environments and build **Real-Time Reasoning Gym** to demonstrate it. We study two paradigms for deploying language models in agents: (1) **Reactive agents**, which employ language models with *bounded reasoning computation for rapid responses*. (2) **Planning agents**, which allow *extended reasoning computation for complex problems*. Our experiments show that even state-of-the-art models struggle with making logical and timely judgments in either paradigm. To address this limitation, we propose **AgileThinker**, which simultaneously engages *both reasoning paradigms*. **AgileThinker** consistently outperforms agents engaging only one reasoning paradigm as task difficulty and time pressure rise, effectively balancing reasoning depth and response latency. Our work establishes real-time reasoning as a critical testbed for developing practical agents and provides a foundation for research in temporally constrained AI systems, highlighting a path toward real-time capable agents.

We propose Guided Speculative Inference (GSI), a novel algorithm for efficient reward-guided decoding in large language models. GSI combines soft best-of-$n$ test-time scaling with a reward model $r(x,y)$ and speculative samples from a small auxiliary model $\pi_S(y\mid x)$. We provably approximate both the optimal tilted policy $\pi_{\beta,B}(y\mid x) \propto \pi_B(y\mid x)\exp(\beta\,r(x,y))$ of soft best-of-$n$ under the base model $\pi_B$, as well as the expected reward under the optimal policy. In experiments on reasoning benchmarks (MATH500, OlympiadBench, Minerva Math, MMLU-STEM, GSM8K) and across different model families, our method achieves higher accuracy than standard soft best-of-$n$ with $\pi_S$ and reward-guided speculative decoding (Liao et al., 2025), and in certain settings even outperforms soft best-of-$n$ with $\pi_B$, while reducing end-to-end latency by up to 28%.

Large Language Models (LLMs) can enhance their reasoning capabilities by using external tools. However, many tasks lack predefined tools. Prior works have explored instructing LLMs to generate tools on their own, but such approaches depend heavily on internal knowledge and struggle when tasks fall outside the model’s knowledge scope. To address this limitation, we propose RefTool, a reference-guided framework for automatic tool creation that leverages external materials, such as textbooks and knowledge snippets. RefTool consists of two modules: (1) tool creation, where LLMs generate executable tools from reference content, validate them using illustrative examples, and organize them hierarchically into a toolbox; and (2) tool utilization, where LLMs navigate the toolbox structure to select and apply the appropriate tools to solve problems. Experiments on causality, physics, and chemistry benchmarks demonstrate that RefTool outperforms existing tool-creation and domain-specific reasoning methods by 12.3% on average accuracy, while being cost-efficient and broadly generalizable to non-scientific tasks, e.g., extremely low-resource language translation. Analyses reveal that grounding tool creation in references produces accurate and faithful tools, and that the hierarchical structure facilitates effective tool selection. RefTool enables LLMs to overcome internal knowledge limitations, advancing generalizable reasoning in knowledge-intensive domains.

Visual reasoning is a core component of human intelligence and a critical capability for advanced multimodal models. Yet current reasoning evaluations of multimodal large language models (MLLMs) often rely on text descriptions and allow language-based reasoning shortcuts, failing to measure genuine vision-centric reasoning. To address this, we introduce VisuLogic: a benchmark of 1,000 human-verified problems across six categories (e.g., quantitative shifts, spatial relations, attribute comparisons). These various types of questions can be evaluated to assess the visual reasoning capabilities of MLLMs from multiple perspectives. We evaluate leading MLLMs on this benchmark and analyze their results to identify common failure modes. Most models score below 30\% accuracy—only slightly above the 25\% random baseline and far below the 51.4\% achieved by humans—revealing significant gaps in visual reasoning.

Test-Time Adaptation (TTA) methods are often computationally expensive, require a large amount of data for effective adaptation, or are brittle to hyperparameters. Based on a theoretical foundation of the geometry of the latent space, we are able to significantly improve the alignment between source and distribution-shifted samples by re-centering target data embeddings at the origin. This insight motivates NEO – a hyperparameter-free fully TTA method, that adds no significant compute compared to vanilla inference. NEO is able to improve the classification accuracy of ViT-Base on ImageNet-C from 55.6\% to 59.2\% after adapting on just one batch of 64 samples. When adapting on 512 samples NEO beats all 7 TTA methods we compare against on ImageNet-C, ImageNet-R and ImageNet-S and beats 6/7 on CIFAR-10-C, while using the least amount of compute. NEO performs well on model calibration metrics and additionally is able to adapt from 1 class to improve accuracy on 999 other classes in ImageNet-C. On Raspberry Pi and Jetson Orin Nano devices, NEO reduces inference time by 63\% and memory usage by 9\% compared to baselines. Our results based on 3 ViT architectures and 4 datasets show that NEO can be used efficiently and effectively for TTA.

Tool-Integrated Reasoning (TIR) enables large language models (LLMs) to enhance their internal reasoning ability by integrating external tools. However, models with TIR often exhibit suboptimal behaviors, including insufficient tool calls, excessive tool calls, and overthinking after receiving tool call results. How to empower LLMs to perform TIR efficiently and accurately, while stabilizing the reasoning process, remains an open challenge. In this paper, we first analyze the impact of tool calls on model reasoning from the perspective of information entropy. We find that when tool call results are provided, the information entropy of subsequent reasoning content will show a clear trend of change, and the overall information entropy of the reasoning chain will vary depending on the number of tool calls. Based on these observations, we propose Tool-Light, a framework designed to encourage LLMs to perform TIR efficiently and accurately. Our framework consists of dataset construction and multi-stage fine-tuning. For dataset construction, we use the trained model for continuous self-evolved sampling, integrating two methods: vanilla sampling and entropy-guided sampling. At the same time, during the sampling process, we design strict criteria for selecting positive-negative pairs. For the training process, we introduce a two-stage method, which includes a Supervised Fine-Tuning (SFT), and Self-Evolved Direct Preference Optimization (DPO). Test results on 10 datasets reveal the effectiveness of Tool-Light, significantly improving the efficiency and accuracy of the model in completing TIR tasks.

Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in complex problem-solving through Chain-of-Thought (CoT) reasoning. However, the multi-step nature of CoT introduces new safety challenges that extend beyond conventional language model alignment. We identify a failure mode in current safety CoT tuning methods: the snowball effect, where minor reasoning deviations progressively amplify throughout the thought process, leading to either harmful compliance or excessive refusal. This effect stems from models being trained to imitate perfect reasoning scripts without learning to self-correct. To address this limitation, we propose AdvChain, an alignment paradigm that teaches models dynamic self-correction through adversarial CoT tuning. Our method involves constructing a dataset containing Temptation-Correction and Hesitation-Correction samples, where models learn to recover from harmful reasoning drifts and unnecessary cautions. Extensive experiments show that AdvChain significantly enhances robustness against jailbreak attacks and CoT hijacking while substantially reducing over-refusal on benign prompts, achieving a superior safety-utility balance without compromising reasoning capabilities. Our work establishes a new direction for building more robust and reliable reasoning models.

Evaluating language models fairly is becoming harder as static benchmarks available on the internet risk contamination by training data. This makes it unclear whether models are truly reasoning or just recalling answers. In this paper, we introduce $\textbf{BeyondBench}$, an evaluation framework that avoids this problem by using $\textbf{algorithmic problem generation}$. Unlike traditional benchmarks that risk contamination from internet-scale training data, $\textbf{BeyondBench}$ creates mathematically grounded problems on the fly, ensuring each test remains fresh and uncontaminated. Our framework covers $\textbf{44 algorithmic tasks}$ with a total of $\textbf{117 variations}$, grouped into three difficulty levels: the $\textit{Easy Suite}$ (29 tasks) for basic arithmetic and statistics, the $\textit{Medium Suite}$ (5 tasks, 49 variations) for sequence patterns and reasoning, and the $\textit{Hard Suite}$ (10 tasks, 68 variations) tackling NP-complete and constraint satisfaction problems. Each task generates problems from a combinatorial space larger than $10^{15}$ unique instances, with solutions verified deterministically by mathematical proofs. We evaluated $\textbf{101 language models}$, including 85 open-source and 16 closed-source models, spanning sizes from 0.5B to 141B parameters and multiple quantization schemes. All evaluations use three-fold cross-validation to ensure statistical robustness. Our results show consistent reasoning deficiencies across model families, with performance degrading sharply as problem complexity increases from polynomial to exponential. In our Hard Suite evaluations, models such as Gemini-2.5-pro, Llama-3.3-70B, and Qwen2.5-72B achieved average accuracies of 56.38%, 26.91%, and 33.60%, respectively. Moreover, we observe that performance drops drastically without tool usage, with GPT-5, GPT-5-mini, and GPT-5-nano showing a $\textbf{decline}$ of 16.81% 28.05%, and 47.59% accuracy on the hard suite. The contamination resistance of $\textbf{BeyondBench}$ rests on three guarantees: (i) the problem space is vastly larger than any static dataset, (ii) every instance has a unique, verifiable solution, and (iii) isomorphic transformations generate semantically equivalent but syntactically new problems. $\textbf{BeyondBench}$ redefines reasoning evaluation through genuine algorithmic problem-solving capability, ensuring a fair and meaningful evaluation.