NeurIPS 2025 Papers

Large language models (LLMs) have shown impressive performance on reasoning benchmarks like math and logic. While many works have largely assumed well-defined tasks, real-world queries are often underspecified and only solvable by acquiring missing information. We formalize this information-gathering problem as a constraint satisfaction problem (CSP) with missing variable assignments. Using a special case where only one necessary variable assignment is missing, we can evaluate an LLM's ability to identify the minimal necessary question to ask. We present QuestBench, a set of underspecified reasoning tasks solvable by asking at most one question, which includes: (1) Logic-Q: logical reasoning tasks with one missing proposition, (2) Planning-Q: PDDL planning problems with partially-observed initial states, (3) GSM-Q: human-annotated grade school math problems with one unknown variable, and (4) GSME-Q: equation-based version of GSM-Q. The LLM must select the correct clarification question from multiple options. While current models excel at GSM-Q and GSME-Q, they achieve only 40-50% accuracy on Logic-Q and Planning-Q. Analysis shows that the ability to solve well-specified reasoning problems is not sufficient for success on our benchmark: models struggle to identify the right question even when they can solve the fully specified version. This highlights the need for specifically optimizing models' information acquisition capabilities.

We introduce *OpenVLThinker*, one of the first open-source large vision–language models (LVLMs) to exhibit sophisticated chain-of-thought reasoning, achieving notable performance gains on challenging visual reasoning tasks. While text-based reasoning models (e.g., Deepseek R1) show promising results in text-only tasks, distilling their reasoning into LVLMs via supervised fine-tuning (SFT) often results in performance degradation due to imprecise visual grounding. Conversely, purely reinforcement learning (RL)-based methods face a large search space, hindering the emergence of reflective behaviors in smaller models (e.g., 7B LVLMs). Surprisingly, alternating between SFT and RL ultimately results in significant performance improvements after a few iterations. Our analysis reveals that the base model rarely exhibits reasoning behaviors initially, but SFT effectively surfaces these latent actions and narrows the RL search space, accelerating the development of reasoning capabilities. Each subsequent RL stage further refines the model's reasoning skills, producing higher-quality SFT data for continued self-improvement. OpenVLThinker-7B consistently advances performance across six benchmarks demanding mathematical and general reasoning, notably improving MathVista by 3.2\%, EMMA by 1.4\%, and HallusionBench by 2.7\%. Beyond demonstrating the synergy between SFT and RL for complex reasoning tasks, our findings provide early evidence towards achieving R1-style reasoning in multimodal contexts.

The advancement of Large Language Models (LLMs) has spurred significant interest in Role-Playing Agents (RPAs) for applications such as emotional companionship and virtual interaction. However, recent RPAs are often built on explicit dialogue data, lacking deep, human-like internal thought processes, resulting in superficial knowledge and style expression. While Large Reasoning Models (LRMs) can be employed to simulate character thought, their direct application is hindered by attention diversion (i.e., RPAs forget their role) and style drift (i.e., overly formal and rigid reasoning rather than character-consistent reasoning). To address these challenges, this paper introduces a novel Role-Aware Reasoning (RAR) method, which consists of two important stages: Role Identity Activation (RIA) and Reasoning Style Optimization (RSO). RIA explicitly guides the model with character profiles during reasoning to counteract attention diversion, and then RSO aligns reasoning style with the character and scene via LRM distillation to mitigate style drift. Extensive experiments demonstrate that the proposed RAR significantly enhances the performance of RPAs by effectively addressing attention diversion and style drift.

3D Visual Grounding (3DVG) faces persistent challenges due to coarse scene-level observations and logically inconsistent annotations, which introduce ambiguities that compromise data quality and hinder effective model supervision. To address these challenges, we introduce Refer-Judge, a novel framework that harnesses the reasoning capabilities of Multimodal Large Language Models (MLLMs) to identify and mitigate toxic data. At the core of Refer-Judge is a Jury-and-Judge Chain-of-Thought paradigm, inspired by the deliberative process of the judicial system. This framework targets the root causes of annotation noise: jurors collaboratively assess 3DVG samples from diverse perspectives, providing structured, multi-faceted evaluations. Judges then consolidate these insights using a Corroborative Refinement strategy, which adaptively reorganizes information to correct ambiguities arising from biased or incomplete observations. Through this two-stage deliberation, Refer-Judge significantly enhances the reliability of data judgments. Extensive experiments demonstrate that our framework not only achieves human-level discrimination at the scene level but also improves the performance of baseline algorithms via data purification. Code is available at https://github.com/Hermione-HKX/Refer_Judge.

Test-Time Discovery (TTD) addresses the critical challenge of identifying and adapting to novel classes during inference while maintaining performance on known classes, which is a capability essential for dynamic real-world environments such as healthcare and autonomous driving. Recent TTD methods adopt training-free, memory-based strategies but rely on frozen models and static representations, resulting in poor generalization. In this paper, we propose a Discrepancy-Amplifying Adapter (DAA), a trainable module that enables real-time adaptation by amplifying feature-level discrepancies between known and unknown classes. During training, DAA is optimized using simulated unknowns and a novel warm-up strategy to enhance its discriminative capacity. To ensure continual adaptation at test time, we introduce a Short-Term Memory Renewal (STMR) mechanism, which maintains a queue-based memory for unknown classes and selectively refreshes prototypes using recent, reliable samples. DAA is further updated through self-supervised learning, promoting knowledge retention for known classes while improving discrimination of emerging categories. Extensive experiments show that our method maintains high adaptability and stability, and significantly improves novel class discovery performance. Our code will be available.

Multimodal large language models (MLLMs) have begun to demonstrate robust reasoning capabilities on general tasks, yet their application in the medical domain remains in its early stages. Constructing chain-of-thought (CoT) training data is essential for bolstering the reasoning abilities of medical MLLMs. However, existing approaches exhibit a deficiency in offering a comprehensive framework for searching and evaluating effective reasoning paths towards critical diagnosis. To address this challenge, we propose Mentor-Intern Collaborative Search (MICS), a novel reasoning-path searching scheme to generate rigorous and effective medical CoT data. MICS first leverages mentor models to initialize the reasoning, one step at a time, then prompts each intern model to continue the thinking along those initiated paths, and finally selects the optimal reasoning path according to the overall reasoning performance of multiple intern models. The reasoning performance is determined by an MICS-Score, which assesses the quality of generated reasoning paths. Eventually, we construct MMRP, a multi-task medical reasoning dataset with ranked difficulty, and Chiron-o1, a new medical MLLM devised via a curriculum learning strategy, with robust visual question-answering and generalizable reasoning capabilities. Extensive experiments demonstrate that Chiron-o1, trained on our CoT dataset constructed using MICS, achieves state-of-the-art performance across a list of medical visual question answering and reasoning benchmarks. Codes are available at https://github.com/Yankai96/Chiron-o1

This paper introduces an approach for training o1-like RAG models that retrieve and reason over relevant information step by step before generating the final answer. Conventional RAG methods usually perform a single retrieval step before the generation process, which limits their effectiveness in addressing complex queries due to imperfect retrieval results. In contrast, our proposed method, CoRAG (Chain-of-Retrieval Augmented Generation), allows the model to dynamically reformulate the query based on the evolving state. To train CoRAG effectively, we utilize rejection sampling to automatically generate intermediate retrieval chains, thereby augmenting existing RAG datasets that only provide the correct final answer. At test time, we propose various decoding strategies to scale the model's test-time compute by controlling the length and number of sampled retrieval chains. Experimental results across multiple benchmarks validate the efficacy of CoRAG, particularly in multi-hop question answering tasks, where we observe more than $10$ points improvement in EM score compared to strong baselines. On the KILT benchmark, CoRAG establishes a new state-of-the-art performance across a diverse range of knowledge-intensive tasks. Furthermore, we offer comprehensive analyses to understand the scaling behavior of CoRAG, laying the groundwork for future research aimed at developing factual and grounded foundation models.

We present Prototypical Pair Network (ProtoPairNet), a novel interpretable architecture that combines deep learning with case-based reasoning to predict continuous targets. While prototype-based models have primarily addressed image classification with discrete outputs, extending these methods to continuous targets, such as regression, poses significant challenges. Existing architectures which rely heavily on one-to-one comparison with prototypes lack the directional information necessary for continuous predictions. Our method redefines the role of prototypes in such tasks by incorporating prototypical pairs into the reasoning process. Predictions are derived based on the input's relative dissimilarities to these pairs, leveraging an intuitive geometric interpretation. Our method further reduces the complexity of the reasoning process by relying on the single most relevant pair of prototypes, rather than all prototypes in the model as was done in prior works. Our model is versatile enough to be used in both vision-based regression and continuous control in reinforcement learning. Our experiments demonstrate that ProtoPairNet achieves performance on par with its black-box counterparts across these tasks. Comprehensive analyses confirm the meaningfulness of prototypical pairs and the faithfulness of our model’s interpretations, and extensive user studies highlight our model's improved interpretability over existing methods.

Recent large language models (LLMs) with long-chain-of-thought reasoning—such as DeepSeek-R1—have achieved impressive results on Olympiad-level mathematics benchmarks. However, they often rely on a narrow set of strategies and struggle with problems that require a novel way of thinking. To systematically investigate these limitations, we introduce OMEGA—Out-of-distribution Math Problems Evaluation with 3 Generalization Axes—a controlled yet diverse bench- mark designed to evaluate three axes of out-of-distribution generalization, inspired by Boden’s typology of creativity: (1) Exploratory—applying known problem- solving skills to more complex instances within the same problem domain; (2) Com- positional—combining distinct reasoning skills, previously learned in isolation, to solve novel problems that require integrating these skills in new and coherent ways; and (3) Transformative—adopting novel, often unconventional strategies by moving beyond familiar approaches to solve problems more effectively. OMEGA consists of programmatically generated training–test pairs derived from templated problem generators across geometry, number theory, algebra, combinatorics, logic, and puzzles, with solutions verified using symbolic, numerical, or graphical methods. We evaluate frontier (or top-tier) LLMs and observe sharp performance degradation as problem complexity increases. Moreover, we fine-tune the Qwen-series models across all generalization settings and observe notable improvements in exploratory generalization, while compositional generalization remains limited, and transformative reasoning shows little to no improvement. By isolating and quantifying these fine-grained failures, OMEGA lays the groundwork for advancing LLMs toward genuine mathematical creativity beyond mechanical proficiency. Our code and dataset are available at https://github.com/sunblaze-ucb/omega.

Modern large reasoning models (LRMs) exhibit impressive multi-step problem-solving via chain-of-thought (CoT) reasoning. However, this iterative thinking mechanism introduces a new vulnerability surface. We present the Deadlock Attack, a resource exhaustion method that hijacks an LRM's generative control flow by training a malicious adversarial embedding to induce perpetual reasoning loops. Specifically, the optimized embedding encourages transitional tokens (e.g., “Wait”, “But”) after reasoning steps, preventing the model from concluding its answer. A key challenge we identify is the continuous-to-discrete projection gap: naïve projections of adversarial embeddings to token sequences nullify the attack. To overcome this, we introduce a backdoor implantation strategy, enabling reliable activation through specific trigger tokens. Our method achieves a 100\% attack success rate across four advanced LRMs (Phi-RM, Nemotron-Nano, R1-Qwen, R1-Llama) and three math reasoning benchmarks, forcing models to generate up to their maximum token limits. The attack is also stealthy (in terms of causing negligible utility loss on benign user inputs) and remains robust against existing strategies trying to mitigate the overthinking issue. Our findings expose a critical and underexplored security vulnerability in LRMs from the perspective of reasoning (in)efficiency.

Large Language Models (LLMs) show great promise in complex reasoning, with Reinforcement Learning with Verifiable Rewards (RLVR) being a key enhancement strategy. However, a prevalent issue is ``superficial self-reflection'', where models fail to robustly verify their own outputs. We introduce RISE (Reinforcing Reasoning with Self-Verification), a novel online RL framework designed to tackle this. RISE explicitly and simultaneously trains an LLM to improve both its problem-solving and self-verification abilities within a single, integrated RL process. The core mechanism involves leveraging verifiable rewards from an outcome verifier to provide on-the-fly feedback for both solution generation and self-verification tasks. In each iteration, the model generates solutions, then critiques its own on-policy generated solutions, with both trajectories contributing to the policy update. Extensive experiments on diverse mathematical reasoning benchmarks show that RISE consistently improves model's problem-solving accuracy while concurrently fostering strong self-verification skills. Our analyses highlight the advantages of online verification and the benefits of increased verification compute. Additionally, RISE models exhibit more frequent and accurate self-verification behaviors during reasoning. These advantages reinforce RISE as a flexible and effective path towards developing more robust and self-aware reasoners.

Recently, Multimodal Large Language Models (MLLMs) have made rapid progress, particularly in enhancing their reasoning capabilities. However, existing reasoning benchmarks still primarily assess language-based reasoning, often treating visual input as replaceable context. To address this gap, we introduce BLINK-Twice, a vision-centric reasoning benchmark grounded in challenging perceptual tasks. Instead of relying on external knowledge, our tasks require models to reason from visual content alone, shifting the focus from language-based to image-grounded reasoning. Compared to prior perception benchmarks, it moves beyond shallow perception ("see") and requires fine-grained observation and analytical reasoning ("observe"). BLINK-Twice integrates three core components: seven types of visual challenges for testing visual reasoning, natural adversarial image pairs that enforce reliance on visual content, and annotated reasoning chains for fine-grained evaluation of the reasoning process rather than final answers alone. We evaluate 20 leading MLLMs, including 12 foundation models and 8 reasoning-enhanced models. BLINK-Twice poses a significant challenge to current models. While existing reasoning strategies in the language space—such as chain-of-thought or self-criticism can improve performance, they often result in unstable and redundant reasoning. We observe that repeated image observation improves performance across models, and active visual interaction, as demonstrated by models like o3, highlights the need for a new paradigm for vision reasoning. The dataset is publicly available at https://github.com/PicoTrex/BLINK-Twice.

Real-world objects frequently undergo state transformations. From an apple being cut into pieces to a butterfly emerging from its cocoon, tracking through these changes is important for understanding real-world objects and dynamics. However, existing methods often lose track of the target object after transformation, due to significant changes in object appearance. To address this limitation, we introduce the task of Track Any State: tracking objects through transformations while detecting and describing state changes, accompanied by a new benchmark dataset, VOST-TAS. To tackle this problem, we present TubeletGraph, a zero-shot system that recovers missing objects after transformation and maps out how object states are evolving over time. TubeletGraph first identifies potentially overlooked tracks, and determines whether they should be integrated based on semantic and proximity priors. Then, it reasons about the added tracks and generates a state graph describing each observed transformation. TubeletGraph achieves state-of-the-art tracking performance under transformations, while demonstrating deeper understanding of object transformations and promising capabilities in temporal grounding and semantic reasoning for complex object transformations. Code, additional results, and the benchmark dataset are available at https://tubelet-graph.github.io.

Formal reasoning and automated theorem proving constitute a challenging subfield of machine learning, in which machines are tasked with proving mathematical theorems using formal languages like Lean. A formal verification system can check whether a formal proof is correct or not almost instantaneously, but generating a completely correct formal proof with large language models (LLMs) remains a formidable task. The usual approach in the literature is to prompt the LLM many times (up to several thousands) until one of the generated proofs passes the verification system. In this work, we present APOLLO (**A**utomated **P**r**O**of repair via**L**LM and **L**ean c**O**llaboration), a modular, model‑agnostic agentic framework that combines the strengths of the Lean compiler with an LLM’s reasoning abilities to achieve better proof‐generation results at a low token and sampling budgets. _Apollo_ directs a fully automated process in which the LLM generates proofs for theorems, a set of agents analyze the proofs, fix the syntax errors, identify the mistakes in the proofs using Lean, isolate failing sub‑lemmas, utilize automated solvers, and invoke an LLM on each remaining goal with a low top‑K budget. The repaired sub‑proofs are recombined and reverified, iterating up to a user‑controlled maximum number of attempts. On the miniF2F benchmark, we establish a new state‑of‑the‑art accuracy of 84.9% among sub 8B‑parameter models (as of August 2025) while keeping the sampling budget below one hundred. Moreover, _Apollo_ raises the state‑of‑the‑art accuracy for Goedel‑Prover‑SFT to 65.6% while cutting sample complexity from 25,600 to a few hundred. General‑purpose models (o3‑mini, o4‑mini) jump from 3–7% to over 40% accuracy. Our results demonstrate that targeted, compiler‑guided repair of LLM outputs yields dramatic gains in both efficiency and correctness, suggesting a general paradigm for scalable automated theorem proving. The codebase is available at https://github.com/aziksh-ospanov/APOLLO

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for training language models (LMs) on reasoning tasks that elicit emergent long chains of thought (CoTs). Unlike supervised learning, it updates the model using both correct and incorrect samples via policy gradients. To better understand its mechanism, we decompose the learning signal into reinforcing correct responses and penalizing incorrect ones, referred to as **P**ositive and **N**egative **S**ample **R**einforcement (**PSR** and **NSR**), respectively. We train `Qwen2.5-Math-7B`, `Qwen3-4B` and `Llama-3.1-8B-Instruct` on a mathematical reasoning dataset and uncover a surprising result: training with only negative samples — without reinforcing correct responses — can be highly effective: it consistently improves performance over the base model across the entire Pass@$k$ spectrum $k$ up to 256), often matching or surpassing PPO and GRPO. In contrast, reinforcing only correct responses improves Pass@1 but degrades performance at higher $k$, due to reduced diversity. These inference-scaling trends highlight that solely penalizing incorrect responses may contribute more to performance than previously recognized. Through gradient analysis, we show that NSR works by suppressing incorrect generations and redistributing probability mass toward other plausible candidates, guided by the model's prior beliefs. It refines the model's existing knowledge rather than introducing entirely new behaviors. Building on this insight, we propose a simple variant of the RL objective that upweights NSR, and show that it consistently improves overall Pass@$k$ performance on MATH, AIME 2025, and AMC23. Our code is available at [`https://github.com/TianHongZXY/RLVR-Decomposed`](https://github.com/TianHongZXY/RLVR-Decomposed).

Vision–language models (VLMs) exhibit strong zero-shot generalization on natural images and show early promise in interpretable medical image analysis. However, existing benchmarks do not systematically evaluate whether these models truly reason like human clinicians or merely imitate superficial patterns. To address this gap, we propose DrVD-Bench, the first multimodal benchmark for clinical visual reasoning. DrVD-Bench consists of three modules: Visual Evidence Comprehension, Reasoning Trajectory Assessment, and Report Generation Evaluation, comprising a total of 7,789 image–question pairs. Our benchmark covers 20 task types, 17 diagnostic categories, and five imaging modalities—CT, MRI, ultrasound, radiography, and pathology. DrVD-Bench is explicitly structured to reflect the clinical reasoning workflow from modality recognition to lesion identification and diagnosis. We benchmark 19 VLMs, including general-purpose and medical-specific, open-source and proprietary models, and observe that performance drops sharply as reasoning complexity increases. While some models begin to exhibit traces of human-like reasoning, they often still rely on shortcut correlations rather than grounded visual understanding. DrVD-Bench offers a rigorous and structured evaluation framework to guide the development of clinically trustworthy VLMs.

Large Language Models (LLMs), such as OpenAI’s o1 and DeepSeek’s R1, excel at advanced reasoning tasks like math and coding via Reinforcement Learning with Verifiable Rewards (RLVR), but still struggle with puzzles solvable by humans without domain knowledge. We introduce ENIGMATA, the first comprehensive suite tailored for improving LLMs with puzzle reasoning skills. It includes 36 tasks across 7 categories, each with: 1) a generator that produces unlimited examples with controllable difficulty, and 2) a rule-based verifier for automatic evaluation. This generator-verifier design supports scalable, multi-task RL training, fine-grained analysis, and seamless RLVR integration. We further propose ENIGMATA-Eval, a rigorous benchmark, and develop optimized multi-task RLVR strategies. Our trained model, Qwen2.5-32B-ENIGMATA, consistently surpasses o3-mini-high and o1 on the puzzle reasoning benchmarks like ENIGMATA-Eval, ARC-AGI (32.8%), and ARC-AGI 2 (0.6%). It also generalizes well to out-of-domain puzzle benchmarks and mathematical reasoning, with little multi-tasking trade-off. When trained on larger models like Seed1.5-Thinking (20B activated parameters and 200B total parameters), puzzle data from ENIGMATA further boosts SoTA performance on advanced math and STEM reasoning tasks such as AIME (2024-2025), BeyondAIME and GPQA (Diamond), showing nice generalization benefits of ENIGMATA. This work offers a unified, controllable framework for advancing logical reasoning in LLMs. Project page: https://seed-enigmata.github.io.

Multilingual reasoning remains a significant challenge for large language models (LLMs), with performance disproportionately favoring high-resource languages. Drawing inspiration from cognitive neuroscience, which suggests that human reasoning functions largely independently of language processing, we hypothesize that LLMs similarly encode reasoning and language as separable components that can be disentangled to enhance multilingual reasoning. To evaluate this, we perform a causal intervention by ablating language-specific representations at inference time. Experiments on 10 open-weight LLMs spanning 11 typologically diverse languages show that this language-specific ablation consistently boosts multilingual reasoning performance. Layer-wise analyses further confirm that language and reasoning representations can be effectively disentangled throughout the model, yielding improved multilingual reasoning capabilities, while preserving top-layer language features remains essential for maintaining linguistic fidelity. Compared to post-training methods such as supervised fine-tuning or reinforcement learning, our training-free language-reasoning disentanglement achieves comparable or superior results with minimal computational overhead. These findings shed light on the internal mechanisms underlying multilingual reasoning in LLMs and suggest a lightweight and interpretable strategy for improving cross-lingual generalization.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks. MMAR comprises 1,000 meticulously curated audio-question-answer triplets, collected from real-world internet videos and refined through iterative error corrections and quality checks to ensure high quality. Unlike existing benchmarks that are limited to specific domains of sound, music, or speech, MMAR extends them to a broad spectrum of real-world audio scenarios, including mixed-modality combinations of sound, music, and speech. Each question in MMAR is hierarchically categorized across four reasoning layers: Signal, Perception, Semantic, and Cultural, with additional sub-categories within each layer to reflect task diversity and complexity. To further foster research in this area, we annotate every question with a Chain-of-Thought (CoT) rationale to promote future advancements in audio reasoning. Each item in the benchmark demands multi-step deep reasoning beyond surface-level understanding. Moreover, a part of the questions requires graduate-level perceptual and domain-specific knowledge, elevating the benchmark's difficulty and depth. We evaluate MMAR using a broad set of models, including Large Audio-Language Models (LALMs), Large Audio Reasoning Models (LARMs), Omni Language Models (OLMs), Large Language Models (LLMs), and Large Reasoning Models (LRMs), with audio caption inputs. The performance of these models on MMAR highlights the benchmark's challenging nature, and our analysis further reveals critical limitations of understanding and reasoning capabilities among current models. These findings underscore the urgent need for greater research attention in audio-language reasoning, including both data and algorithm innovation. We hope MMAR will serve as a catalyst for future advances in this important but little-explored area.