ICLR 2026 论文列表

Large Language Models (LLMs) have shown strong performance in complex reasoning tasks, especially when guided by Chain-of-Thought (CoT) prompting. However, conventional CoT reasoning in the discrete token space suffers from high computational and memory costs due to verbose intermediate steps. Recent work has explored latent reasoning in the embedding space to improve efficiency, but often at the cost of clarity and performance. In this work, we propose $\underline{Hy}$brid $\underline{Rea}$soning ($\texttt{HyRea}$), a unified framework that enables LLMs to dynamically switch between explicit (token-based) and latent (embedding-based) reasoning during inference. To train the model to make these decisions effectively, we introduce a two-stage training pipeline: (1) a supervised cold-start phase that introduces latent reasoning by replacing low-entropy CoT steps with embeddings, and (2) a reinforcement learning phase using Group Relative Policy Optimization (GRPO) to fine-tune the model’s reasoning strategy based on task-specific rewards. Experiments on mathematical reasoning benchmarks show that \texttt{HyRea} achieves significant reductions in token usage while maintaining or improving accuracy, offering an effective and scalable solution for efficient multi-step reasoning in LLMs.

Large language models (LLMs) increasingly rely on Chain-of-Thought (CoT) prompting to improve problem-solving and provide seemingly transparent explanations. However, growing evidence shows that CoT often fail to faithfully represent the underlying reasoning process, raising concerns about their reliability in high-risk applications. Although prior studies have focused on mechanism-level analyses showing that CoTs can be unfaithful, they leave open the practical challenge of deciding whether a specific trajectory is faithful to the internal reasoning of the model. To address this gap, we introduce FaithCoT-Bench, a unified benchmark for instance-level CoT unfaithfulness detection. Our framework establishes a rigorous task formulation that formulates unfaithfulness detection as a discriminative decision problem, and provides FINE-CoT (Faithfulness instance evaluation for Chain-of-Thought), an expert-annotated collection of over 1,000 trajectories generated by four representative LLMs across four domains, including more than 300 unfaithful instances with fine-grained causes and step-level evidence. We further conduct a systematic evaluation of eleven representative detection methods spanning counterfactual, logit-based, and LLM-as-judge paradigms, deriving empirical insights that clarify the strengths and weaknesses of existing approaches and reveal the increased challenges of detection in knowledge-intensive domains and with more advanced models. To the best of our knowledge, FaithCoT-Bench establishes the first comprehensive benchmark for instance-level CoT faithfulness, setting a solid basis for future research toward more interpretable and trustworthy reasoning in LLMs.

Recent advances in large language models (LLMs) for mathematical reasoning have largely focused on tasks with easily verifiable final answers while generating and verifying natural language math proofs remains an open challenge. We identify the absence of a reliable, fine-grained evaluator for LLM-generated math proofs as a critical gap. To address this, we propose a systematic methodology for developing and validating evaluators that assign fine-grained scores on a 0-7 scale to model-generated math proofs. To enable this study, we introduce ProofBench, the first expert-annotated dataset of fine-grained proof ratings, spanning 145 problems from six major math competitions (USAMO, IMO, Putnam, etc) and 435 LLM-generated solutions from Gemini-2.5-Pro, o3, and DeepSeek-R1. Using ProofBench as a testbed, we systematically explore the evaluator design space across key axes: the backbone model, input context, instructions and evaluation workflow. Our analysis delivers ProofGrader, an evaluator that combines a strong reasoning backbone LM, rich context from reference solutions and marking schemes, and a simple ensembling method; it achieves a low Mean Absolute Error (MAE) of 0.926 against expert scores, significantly outperforming naive baselines. Finally, we demonstrate its practical utility in a best-of-$n$ selection task: at $n=16$, ProofGrader achieves an average score of 4.14/7, closing 78\% of the gap between a naive binary evaluator (2.48) and the human oracle (4.62), highlighting its potential to advance downstream proof generation.

Reasoning is not just about solving problems---it is also about evaluating which problems are worth solving at all. Evaluations of artificial intelligence (AI) systems primarily focused on problem solving, historically by studying how models play games such as chess and Go. In this paper, we advocate for a new paradigm that assesses AI systems' evaluation of games. First, we introduce a formalism for evaluating such evaluations. We then leverage a large-scale dataset of over 100 novel board games and over 450 human judgments to compare evaluations produced by modern language and reasoning models against those of people and symbolic computational agents. We consider two kinds of evaluative queries: assessing the payoff (or fairness) and the funness of games. These queries span two dimensions relevant to the design of evaluations of AI evaluations: how complex a query is to compute and how difficult a query is to quantify. Our results show that reasoning models are generally more aligned to people in their evaluations of games than non-reasoning language models. However, we observe a non-monotonic relationship: as models get closer to game-theoretic optimal, their fit to human data weakens. We also observe more "jaggedness" across models for assessing funness, in line with the greater difficulty of quantifying this query. Across queries and games, reasoning models show highly variable and unpredictable resource usage when assessing queries, pointing to the importance of imbuing more resource-rational meta-reasoning in language and reasoning models.

Neural network models with latent recurrent processing, where identical layers are recursively applied to the latent state, have gained attention as promising models for performing reasoning tasks. A strength of such models is that they enable test-time scaling, where the models can enhance their performance in the test phase without additional training. Models such as the Hierarchical Reasoning Model (HRM) and Artificial Kuramoto Oscillatory Neurons (AKOrN) can facilitate deeper reasoning by increasing the number of recurrent steps, thereby enabling the completion of challenging tasks, including Sudoku, Maze solving, and AGI benchmarks. In this work, we introduce confidence-based voting (C-voting), a test-time scaling strategy designed for recurrent models with multiple latent candidate trajectories. Initializing the latent state with multiple candidates using random variables, C-voting selects the one maximizing the average of top-1 probabilities of the predictions, reflecting the model’s confidence. Additionally, it yields $4.9\\%$ higher accuracy on Sudoku-hard than the energy-based voting strategy, which is specific to models with explicit energy functions. An essential advantage of C‑voting is its applicability: it can be applied to recurrent models without requiring an explicit energy function. Finally, we introduce a simple attention-based recurrent model with randomized initial values named ItrSA++, and demonstrate that when combined with C-voting, it outperforms HRM on Sudoku-extreme ($95.2\\%$ vs. $55.0\\%$) and Maze ($78.6\\%$ vs. $74.5\\%$) tasks.

Reinforcement learning (RL) with large language models shows promise in complex reasoning. However, its progress is hindered by the lack of large-scale training data that is sufficiently challenging, contamination-free and verifiable. To this end, we introduce DeepMath-103K, a large-scale mathematical dataset designed with high difficulty (primarily levels 5-9), rigorous decontamination against numerous benchmarks, and verifiable answers for rule-based RL reward. It further includes three distinct R1 solutions adaptable for diverse training paradigms such as supervised fine-tuning (SFT). Spanning a wide range of mathematical topics, DeepMath-103K fosters the development of generalizable and advancing reasoning. Notably, models trained on DeepMath-103K achieve leading results on challenging mathematical benchmarks and demonstrate generalization beyond math such as biology, physics and chemistry, underscoring its broad efficacy.

Reasoning about failures is crucial for building reliable and trustworthy robotic systems. Prior approaches either treat failure reasoning as a closed-set classification problem or assume access to ample human annotations. Failures in the real world are typically subtle, combinatorial, and difficult to enumerate, whereas rich reasoning labels are expensive to acquire. We address this problem by introducing ARMOR: Adaptive Round-based Multi-task mOdel for Robotic failure detection and reasoning. We formulate detection and reasoning as a multi-task self-refinement process, where the model iteratively predicts detection outcomes and natural language reasoning conditioned on past outputs. During training, ARMOR learns from heterogeneous supervision - large-scale sparse binary labels and small-scale rich reasoning annotations - optimized via a combination of offline and online imitation learning. At inference time, ARMOR generates multiple refinement trajectories and selects the most confident prediction via a self-certainty metric. Experiments across diverse environments show that ARMOR achieves state-of-the-art performance by improving over the previous approaches by up to 30\% on failure detection rate and up to 100\% in reasoning measured through LLM fuzzy match score, demonstrating robustness to heterogeneous supervision and open-ended reasoning beyond predefined failure modes. We provide dditional visualizations on our website: https://sites.google.com/utexas.edu/armor.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, but their high computational costs limit their widespread practical application. We argue that this inefficiency arises from the tight coupling of high-level cognitive planning (devising the solution strategy) and low-level linguistic realization (generating step-by-step text). To address this challenge, we propose a novel collaborative framework that decouples these two processes through Latent Guidance. Our approach implements a division of labor: a large model acts as an Implicit Thinker, performing high-level cognitive planning and compressing its solution strategy into a set of compact latent guidance vectors. A small, efficient model then serves as an Explicit Executor, which receives this latent guidance to generate a concise and effective reasoning chain. This process is enabled by a dual-loss training objective, grounded in information-theoretic principles, where a reconstruction loss explicitly compels the latent guidance to become a high-fidelity representation of the full reasoning chain. Extensive experiments on 8 diverse reasoning benchmarks demonstrate that our method substantially enhances the reasoning capabilities of small models across various scales (from 0.5B to 8B), allowing them to outperform strong baselines and exhibit superior generalization. Notably, our framework boosts small model accuracy by up to 13.9% and its speed by 2x over its standalone baseline, while being up to 4x faster than the large model. Our work introduces a new, theoretically-grounded paradigm for empowering small models with large-model thinking, substantially improving the performance-cost trade-off for complex reasoning.

Multimodal Large Language Models (MLLMs) have recently emerged as general architectures capable of reasoning over diverse modalities. Benchmarks for MLLMs should measure their ability for cross‑modal integration. However, current benchmarks are filled with shortcut questions, which can be solved using only single modality, and thereby yielding unreliable rankings. For example, in vision-language cases, we can find the correct answer without either the image or the text. These low-quality questions unnecessarily increase the size and computational requirements of benchmarks. We introduce a multi-modal and multidimensional item response theory framework (M$^3$-IRT) that extends classical IRT by decomposing both model ability and item difficulty into image‑only, text‑only, and cross‑modal components. M$^3$-IRT estimates cross‑modal ability of MLLMs and each question’s cross‑modal difficulty, enabling compact, high‑quality subsets that better reflect multimodal reasoning. Across 24 VLMs on three benchmarks, M$^3$-IRT prioritizes genuinely cross‑modal questions over shortcuts and preserves ranking fidelity even when 50\% of items are artificially generated low‑quality questions, thereby reducing evaluation cost while improving reliability. M$^3$-IRT thus offers a practical tool for assessing cross‑modal reasoning and refining multimodal benchmarks.

LLMs demonstrate performance comparable to human abilities in complex tasks such as mathematical reasoning, but their robustness in mathematical reasoning under minor input perturbations still lacks systematic investigation. Existing methods generally suffer from limited scalability, weak semantic preservation, and high costs. Therefore, we propose MSCR, an automated adversarial attack method based on multi-source candidate replacement. By combining three information sources including cosine similarity in the embedding space of LLMs, the WordNet dictionary, and contextual predictions from a masked language model, we generate for each word in the input question a set of semantically similar candidates, which are then filtered and substituted one by one to carry out the attack. We conduct large-scale experiments on LLMs using the GSM8K and MATH500 benchmarks. The results show that even a slight perturbation involving only a single word can significantly reduce the accuracy of all models, with the maximum drop reaching 49.89\% on GSM8K and 35.40\% on MATH500, _while preserving the high semantic consistency of the perturbed questions._ Further analysis reveals that perturbations not only lead to incorrect outputs but also substantially increase the average response length, which results in more redundant reasoning paths and higher computational resource consumption. These findings highlight the robustness deficiencies and efficiency bottlenecks of current LLMs in mathematical reasoning tasks.

Despite their impressive capabilities, LLMs exhibit a basic generalization failure known as the *Reversal Curse*, where they struggle to learn reversible factual associations. Understanding why this occurs could help identify weaknesses in current models and advance their generalization and robustness. In this paper, we conjecture that the Reversal Curse in LLMs is a manifestation of the long-standing *binding problem* in cognitive science, neuroscience and AI. Specifically, we hypothesize two primary causes of the Reversal Curse stemming from transformers' limitations in conceptual binding: the *inconsistency* and *entanglements* of concept representations. We perform a series of experiments that support these conjectures. Our exploration leads to a model design based on JEPA (Joint-Embedding Predictive Architecture) that for the first time breaks the Reversal Curse without side-stepping it with specialized data augmentation or non-causal masking, and moreover, generalization could be further improved by incorporating special memory layers that support disentangled concept representations. Our research opens up the broader fundamental challenge of designing models capable of learning systematic conceptual binding with less human scaffolding.

Large language models (LLMs) with explicit reasoning capabilities excel at mathematical reasoning yet still commit process errors, such as incorrect calculations, brittle logic, and superficially plausible but invalid steps. In this paper, we introduce Generative Adversarial Reasoner, an on-policy joint training framework designed to enhance reasoning by co-evolving an LLM reasoner and an LLM-based discriminator through adversarial reinforcement learning. A compute-efficient review schedule partitions each reasoning chain into logically complete slices of comparable length, and the discriminator evaluates each slice’s soundness with concise, structured justifications. Learning couples complementary signals: the LLM reasoner is rewarded for logically consistent steps that yield correct answers, while the discriminator earns rewards for correctly detecting errors or distinguishing traces in the reasoning process. This produces dense, well-calibrated, on-policy step-level rewards that supplement sparse exact-match signals, improving credit assignment, increasing sample efficiency, and enhancing overall reasoning quality of LLMs. Across various mathematical benchmarks, the method delivers consistent gains over strong baselines with standard RL post-training. Specifically, on AIME24, we improve DeepSeek-R1-Distill-Qwen-7B from 54.0 to 61.3 (+7.3) and DeepSeek-R1-Distill-Llama-8B from 43.7 to 53.7 (+10.0). The modular discriminator also enables flexible reward shaping for objectives such as teacher distillation, preference alignment, and mathematical proof-based reasoning.

Reinforcement Learning, particularly through policy gradient methods, has played a central role in enabling reasoning capabilities of Large Language Models. However, the optimization stability of policy gradients in this setting remains understudied. As a result, existing implementations often resort to conservative hyperparameter choices to ensure stability, which requires more training samples and increases computational costs. Hence, developing models for reliably tracking the underlying optimization dynamics and leveraging them into training enables more sample-efficient regimes and further unleashes scalable post-training. We address this gap by formalizing the stochastic optimization problem of policy gradients with explicit consideration of second-order geometry. We propose a tractable computational framework that tracks and leverages curvature information during policy updates. We further employ this framework to design interventions in the optimization process through data selection. The resultant algorithm, Curvature-Aware Policy Optimization (CAPO), identifies samples that contribute to unstable updates and masks them out. Theoretically, we establish monotonic improvement guarantees under realistic assumptions. On standard math reasoning benchmarks, we empirically show that CAPO ensures stable updates under aggressive learning regimes where baselines catastrophically fail. With minimal intervention (rejecting fewer than 8% of tokens), CAPO achieves up to 30$\times$ improvement in sample efficiency over standard GRPO for LLM reasoning.

Majority voting has proven effective for close-ended question answering by aggregating parallel reasoning traces. However, it is not directly applicable to open-ended reasoning, where "majority" is undefined. We introduce THINKMERGE, a training-free, plug-and-play decoding strategy that runs K parallel reasoning traces and averages their next-token logits at synchronization points to produce a single coherent output. THINKMERGE integrates seamlessly with vLLM/SGLang and remains compatible with standard decoding techniques such as Top-p/Top-k. Empirically, it matches or surpasses majority voting on AIME and GPQA, while delivering consistent gains on open-ended coding tasks: on LiveCodeBench (hard), pass@1 improves by +8.28% for DeepCoder-14B-Preview and +7.58% for Qwen3-8B. These results demonstrate that parallel test-time scaling can benefit open-ended reasoning without relying on voting over complete outputs.

Neural theorem proving has advanced rapidly in the past year, reaching IMO gold-medalist capabilities and producing formal proofs that span thousands of lines. Although such proofs are mechanically verified by formal systems like Lean, their excessive length renders them difficult for humans to comprehend and limits their usefulness for mathematical insight. Proof simplification is therefore a critical bottleneck. Yet, training data for this task is scarce, and existing methods—mainly agentic scaffolding with off-the-shelf LLMs—struggle with the extremely long proofs generated by RL-trained provers. We introduce ProofOptimizer, the first language model trained to simplify Lean proofs without requiring additional human supervision. ProofOptimizer is trained via expert iteration and reinforcement learning, using Lean to verify simplifications and provide training signal. At inference time, it operates within an iterative proof-shortening workflow, progressively reducing proof length. Experiments show that ProofOptimizer substantially compresses proofs generated by state-of-the-art RL-trained provers on standard benchmarks, reducing proof length by 87% on miniF2F, 57% on PutnamBench, and 50% on Seed-Prover's IMO 2025 proofs. Beyond conciseness, the simplified proofs check faster in Lean and further improve downstream prover performance when reused as training data for supervised finetuning.

Reinforcement learning with verifiable rewards has significantly advanced the reasoning capabilities of large language models, yet how to explicitly steer training toward exploration or exploitation remains an open problem. We introduce Token Hidden Reward (THR), a token-level metric that quantifies each token’s influence on the likelihood of correct responses under Group Relative Policy Optimization (GRPO). We find that training dynamics are dominated by a small subset of tokens with high absolute THR values. Most interestingly, tokens with positive THR strengthen confidence in correct outputs, thus favoring exploitation, while tokens with negative THR preserve probability mass for alternative outputs, enabling exploration. This insight suggests a natural intervention: a THR-guided reweighting algorithm that modulates GRPO’s learning signals to explicitly bias training toward exploitation or exploration. We validate the efficacy of this algorithm on diverse math reasoning benchmarks. By amplifying tokens with positive THR value and weakening negative ones, our algorithm improves greedy-decoding accuracy, favoring exploitation. The reverse strategy yields consistent gains in Pass@K accuracy, favoring exploration. We further demonstrate that our algorithm integrates seamlessly with other RL objectives such as GSPO and generalizes across architectures including Llama. These findings establish THR as a principled and fine-grained mechanism for dynamically controlling exploration and exploitation in RL-tuned LLMs, providing new tools for targeted fine-tuning in reasoning-intensive applications.

Large language models achieve strong reasoning performance, but inference strategies such as Self-Consistency (SC) are computationally expensive, as they fully expand all reasoning traces. We introduce PoLR (Path of Least Resistance), the first inference-time method to leverage prefix self-consistency for compute-efficient reasoning. PoLR clusters short prefixes of reasoning traces, identifies the dominant cluster, and expands only a subset of promising paths, preserving the accuracy benefits of SC while substantially reducing token usage and latency. Our theoretical analysis, framed via mutual information and entropy, explains why early reasoning steps encode strong signals predictive of final correctness. Empirically, PoLR consistently matches or exceeds SC across GSM8K, Math500, AIME 2024/2025, and GPQA-Diamond, reducing token usage by up to 60% and wall-clock latency by up to 50%. Moreover, PoLR is fully complementary to adaptive inference methods (e.g., Adaptive Consistency, Early-Stopping SC) and can serve as a drop-in pre-filter, making SC substantially more efficient and scalable without requiring model fine-tuning.

Current online reinforcement learning (RL) algorithms like GRPO share a key limitation in LLM reasoning: they cannot learn from problems that are "unsolvable" to the model. In other words, they can only improve performance on problems where the model is capable of exploring the correct answer. If a problem is too difficult -- such that even hundreds of attempts never produce a correct solution -- the model cannot learn from it. Consequently, the model's "upper limit" remains unchanged after RL training, even though the likelihood of solving easier, solvable problems may increase. These hard, unsolvable samples -- though potentially rich in learning signal -- cannot contribute to training, as no rollouts yield rewards and thus no gradients are produced. To unlock learning from these hard samples, we propose NuRL, a "nudging" method that aims to push the upper bound of LLM reasoning using self-generated hints, i.e., abstract cues that help reduce the problem difficulty for the model. Given a question and its gold answer, the model generates a Chain-of-Thought (CoT) and then produces a hint containing the core knowledge needed to solve the problem. During online RL training, we generate G rollouts from the base policy and use the pass rate to decide whether the hint should be injected. For hard samples with a 0% pass rate, we inject the offline-generated hint and regenerate a new batch of trajectories. This yields two benefits: (1) the hint boosts pass rates (from 0% to non-zero), thereby introducing training signals for previously unsolvable samples, and (2) the hints are self-generated (conditioned on the gold answer), avoiding distributional shift and do not rely on external models. Compared to standard GRPO, NuRL achieves consistent improvements across six diverse benchmarks and three models, while remaining complementary to test-time scaling. Notably, NuRL can raise the model's upper limit, whereas GRPO leaves pass@1024 unchanged from the base model. Furthermore, we present a systematic study of what makes an effective hint and when hints are most useful. Interestingly, the best hints are abstract and high-level -- as revealing gold answers actually hurt performance -- and are most beneficial when applied necessarily and after GRPO has converged.

Reasoning LLMs are trained to verbalize their thinking process, yielding strong gains on complex tasks. This transparency also opens a promising direction: multiple reasoners can directly collaborate on each other's thinking on a shared trajectory, yielding better inference efficiency and exploration. A key prerequisite, however, is the ability to assess usefulness and build on another model's partial thinking —we call this *off-trajectory reasoning*. Our paper investigates a critical question: can standard *solo-reasoning* training pipelines yield desired *off-trajectory* behaviors? We propose twin tests that capture the two extremes of the off-trajectory spectrum, namely **Recoverability**, which tests whether LLMs can backtrack from ``distractions'' induced by misleading reasoning traces, and **Guidability**, which tests their ability to build upon correct reasoning from stronger collaborators. Our study evaluates 15 open-weight LLMs (1.5B—32B) and reveals a counterintuitive finding — "stronger" LLMs on benchmarks are often more fragile under distraction. Moreover, all models tested fail to effectively leverage guiding steps from collaborators on problems beyond their inherent capabilities, with solve rates remaining under 9.2%. Finally, we conduct control studies to isolate the effects of three factors in post-training on these behaviors: the choice of distillation teacher, the use of RL, and data selection strategy. Our results provide actionable insights for training natively strong reasoning collaborators; e.g., we find that sub-optimal recoverability behaviors of teacher models are transferred to distilled students even if the distilled data trajectories are correct. Taken together, this work lays the groundwork for evaluating multi-model collaborations under shared reasoning, while revealing limitations of off-the-shelf reasoning LLMs.