NeurIPS 2025 Papers

Large Language Models (LLMs) have recently demonstrated strong capabilities in code-related tasks, but their robustness in code reasoning under perturbations remains underexplored. We introduce CodeCrash, a stress-testing framework with 1,279 questions from CRUXEVAL and LIVECODEBENCH, designed to evaluate reasoning reliability under structural perturbations and misleading natural language (NL) contexts. Through a systematic evaluation of 17 LLMs, we find that models often shortcut reasoning by over-relying on NL cues, leading to an average performance degradation of 23.2% in output prediction tasks. Even with Chain-of-Thought reasoning, models on average still have a 13.8% drop due to distractibility and rationalization, revealing a lack of critical reasoning capability to distinguish the actual code behaviors. While Large Reasoning Models with internal reasoning mechanisms improve robustness by fostering critical thinking, plausible yet incorrect hints can trigger pathological self-reflection, causing 2-3 times token consumption and even catastrophic cognitive dissonance in extreme cases for QwQ-32B. We refer to this phenomenon as Reasoning Collapse. CodeCrash provides a rigorous benchmark for evaluating robustness in code reasoning, guiding future research and development toward more reliable and resilient models.

Topology reasoning, which unifies perception and structured reasoning, plays a vital role in understanding intersections for autonomous driving. However, its performance heavily relies on the accuracy of lane detection, particularly at connected lane endpoints. Existing methods often suffer from lane endpoints deviation, leading to incorrect topology construction. To address this issue, we propose TopoPoint, a novel framework that explicitly detects lane endpoints and jointly reasons over endpoints and lanes for robust topology reasoning. During training, we independently initialize point and lane query, and proposed Point-Lane Merge Self-Attention to enhance global context sharing through incorporating geometric distances between points and lanes as an attention mask . We further design Point-Lane Graph Convolutional Network to enable mutual feature aggregation between point and lane query. During inference, we introduce Point-Lane Geometry Matching algorithm that computes distances between detected points and lanes to refine lane endpoints, effectively mitigating endpoint deviation. Extensive experiments on the OpenLane-V2 benchmark demonstrate that TopoPoint achieves state-of-the-art performance in topology reasoning (48.8 on OLS). Additionally, we propose DET$_p$ to evaluate endpoint detection, under which our method significantly outperforms existing approaches (52.6 v.s. 45.2 on DET$_p$). The code is released at https://github.com/Franpin/TopoPoint.

Imagination in world models is crucial for enabling agents to learn long-horizon policy in a sample-efficient manner. Existing recurrent state-space model (RSSM)-based world models depend on single-step statistical inference to capture the environment dynamics, and, hence, they are unable to perform long-term imagination tasks due to the accumulation of prediction errors. Inspired by the dual-process theory of human cognition, we propose a novel dual-mind world model (DMWM) framework that integrates logical reasoning to enable imagination with logical consistency. DMWM is composed of two components: an RSSM-based System 1 (RSSM-S1) component that handles state transitions in an intuitive manner and a logic-integrated neural network-based System 2 (LINN-S2) component that guides the imagination process through hierarchical deep logical reasoning. The inter-system feedback mechanism is designed to ensure that the imagination process follows the logical rules of the real environment. The proposed framework is evaluated on benchmark tasks that require long-term planning from the DMControl suite and robotic environment. Extensive experimental results demonstrate that the proposed framework yields significant improvements in terms of logical coherence, trial efficiency, data efficiency and long-term imagination over the state-of-the-art world models.

Human cognition typically involves thinking through abstract, fluid concepts rather than strictly using discrete linguistic tokens. Current Large Language Models (LLMs), however, are constrained to reasoning within the boundaries of human language, processing discrete token embeddings that represent fixed points in semantic space. This discrete constraint restricts the expressive power and upper potential of such reasoning models, often causing incomplete exploration of reasoning paths, as standard Chain-of-Thought (CoT) methods rely on sampling one token per step. In this work, we introduce Soft Thinking, a training-free method that emulates human-like ``soft'' reasoning by generating abstract concept tokens in a continuous concept space. These concept tokens are created by the probability-weighted mixture of token embeddings, which span the continuous concept space, enabling smooth transitions and richer representations that transcend traditional discrete boundaries. In essence, each generated concept token encapsulates multiple meanings from related discrete tokens, implicitly exploring various reasoning paths to converge effectively toward the correct answer. Empirical evaluations on diverse mathematical and coding benchmarks consistently demonstrate the effectiveness and efficiency of Soft Thinking, improving pass@1 accuracy by up to 2.48 points while simultaneously reducing token usage by up to 22.4\% compared to standard CoT. Qualitative analysis further reveals that Soft Thinking outputs remain highly interpretable and readable, highlighting the potential of Soft Thinking to break the inherent limits of discrete language-based reasoning.

Charts are high-density visualization carriers for complex data, serving as a crucial medium for information extraction and analysis. Automated chart understanding poses significant challenges to existing multimodal large language models (MLLMs) due to the need for precise and complex visual reasoning. Current step-by-step reasoning models primarily focus on text-based logical reasoning for chart understanding. However, they struggle to refine or correct their reasoning when errors stem from flawed visual understanding, as they lack the ability to leverage multimodal interaction for deeper comprehension. Inspired by human cognitive behavior, we propose ChartSketcher, a multimodal feedback-driven step-by-step reasoning method designed to address these limitations. ChartSketcher is a chart understanding model that employs Sketch-CoT, enabling MLLMs to annotate intermediate reasoning steps directly onto charts using a programmatic sketching library, iteratively feeding these visual annotations back into the reasoning process. This mechanism enables the model to visually ground its reasoning and refine its understanding over multiple steps. We employ a two-stage training strategy: a cold start phase to learn sketch-based reasoning patterns, followed by off-policy reinforcement learning to enhance reflection and generalization. Experiments demonstrate that ChartSketcher achieves promising performance on chart understanding benchmarks and general vision tasks, providing an interactive and interpretable approach to chart comprehension.

Test-time scaling (TTS) enhances the performance of large language models (LLMs) by allocating additional compute resources during inference. However, existing research primarily investigates TTS in single-stage tasks; while many real-world problems are multi-stage complex tasks, composed of a sequence of heterogeneous subtasks with each subtask requires LLM of specific capability. Therefore, we study a novel problem: the test-time compute-optimal scaling in multi-stage complex tasks, aiming to select suitable models and allocate budgets per subtask to maximize overall performance. TTS in multi-stage tasks introduces two fundamental challenges: (i) The combinatorial search space of model and budget allocations, combined with the high cost of inference, makes brute-force search impractical. (ii) The optimal model and budget allocations across subtasks are interdependent, increasing the complexity of the compute-optimal search. To address this gap, we conduct extensive pilot experiments on four tasks across six datasets, deriving three empirical insights characterizing the behavior of LLMs in multi-stage complex tasks. Informed by these insights, we propose AgentTTS, an LLM-agent-based framework that autonomously searches for compute-optimal allocations through iterative feedback-driven interactions with the execution environment. Experimental results demonstrate that AgentTTS significantly outperforms traditional and other LLM-based baselines in search efficiency, and shows improved robustness to varying training set sizes and enhanced interpretability.

Recent advances in reasoning techniques have substantially improved the performance of large language models (LLMs), raising expectations for their ability to provide accurate, truthful, and reliable information. However, emerging evidence suggests that iterative reasoning may foster belief entrenchment and confirmation bias, rather than enhancing truth-seeking behavior. In this study, we propose a systematic evaluation framework for *belief entrenchment* in LLM reasoning by leveraging the Martingale property from Bayesian statistics. This property implies that, under rational belief updating, the expected value of future beliefs should remain equal to the current belief, i.e., belief updates are unpredictable from the current belief. We propose the unsupervised, regression-based *Martingale Score* to measure violations of this property, which signal deviation from the Bayesian ability of updating on new evidence. In open-ended problem domains including event forecasting, value-laden questions, and academic paper review, we find such violations to be widespread across models and setups, where the current belief positively predicts future belief updates, a phenomenon which we term *belief entrenchment*. We identify the models, reasoning techniques, and domains more prone to belief entrenchment. Finally, we validate the Martingale Score by showing that it predicts ground-truth accuracy on problem domains where ground truth labels are available. This indicates that, while designed as an unsupervised metric that operates even in domains without access to ground truth, the Martingale Score is a useful proxy of the truth-seeking ability of a reasoning process.

Test-time scaling enhances large language model performance by allocating additional compute resources during decoding. Best-of-$N$ (BoN) sampling serves as a common sampling-based scaling technique, broadening the search space in parallel to find better solutions from the model distribution. However, its cost–performance trade-off is still underexplored. Two main challenges limit the efficiency of BoN sampling: (1) Generating $N$ full samples consumes substantial GPU memory, reducing inference capacity under limited resources. (2) Reward models add extra memory and latency overhead, and training strong reward models introduces potential training data costs. Although some studies have explored efficiency improvements, none have addressed both challenges at once. To address this gap, we propose **Self-Truncation Best-of-$N$ (ST-BoN)**, a decoding method that avoids fully generating all $N$ samples and eliminates the need for reward models. It leverages early sampling consistency in the model’s internal states to identify the most promising path and truncate suboptimal ones. In terms of cost, ST-BoN reduces dynamic GPU memory usage by over 80% and inference latency by 50%. In terms of cost–performance trade-off, ST-BoN achieves the same performance as Full-BoN while saving computational cost by 70%–80%, and under the same cost, it can improve accuracy by 3–4 points.

In this paper, we introduce **PolyMath**, a multilingual mathematical reasoning benchmark covering 18 languages and 4 easy-to-hard difficulty levels. Our benchmark ensures difficulty comprehensiveness, language diversity, and high-quality translation, making it a highly discriminative multilingual mathematical benchmark in the era of reasoning LLMs. We conduct a comprehensive evaluation for advanced LLMs and find that even Qwen-3-235B-A22B-Thinking and Gemini-2.5-pro, achieve only 54.6 and 52.2 benchmark scores, with about 40% accuracy under the highest level. From a language perspective, our benchmark reveals several key challenges of LLMs in multilingual reasoning: (1) Reasoning performance varies widely across languages for current LLMs; (2) Input-output language consistency is low in reasoning LLMs and may be correlated with performance; (3) The thinking length differs significantly by language for current LLMs. Additionally, we demonstrate that controlling the output language in the instructions has the potential to affect reasoning performance, especially for some low-resource languages, suggesting a promising direction for improving multilingual capabilities in LLMs.

In the field of AI-driven human-GUI interaction automation, while rapid advances in multimodal large language models and reinforcement fine-tuning techniques have yielded remarkable progress, a fundamental challenge persists: their interaction logic significantly deviates from natural human-GUI communication patterns. To address this gap, we propose Blink–Think–Link (BTL), a brain-inspired framework for human-GUI interaction that mimics the human cognitive process between users and graphical interfaces. The system decomposes interactions into three biologically plausible phases: (1) \textbf{Blink} - rapid detection and attention to relevant screen areas, analogous to saccadic eye movements; (2) \textbf{Think} - higher-level reasoning and decision-making, mirroring cognitive planning; and (3) \textbf{Link} - generation of executable commands for precise motor control, emulating human action selection mechanisms. Additionally, we introduce two key technical innovations for BTL framework: (1) Blink Data Generation - an automated annotation pipeline specifically optimized for blink data, and (2) {BTL Reward – the first rule-based reward mechanism that enables reinforcement learning driven by both process and outcome.} Building upon this framework, we develop a GUI agent model named BTL-UI, which demonstrates competitive performance across both static GUI understanding and dynamic interaction tasks in comprehensive benchmarks. These results provide conclusive empirical validation of the framework's efficacy in developing advanced GUI agents.

While language models have shown remarkable performance across diverse tasks, they still encounter challenges in complex reasoning scenarios. Recent research suggests that language models trained on linearized search traces toward solutions, rather than solely on the final solutions, exhibit improved generalization, despite the search traces being potentially noisy or suboptimal. However, relying on such imperfect traces can result in inefficient use of test-time compute. To address this, we propose guided reinforced self-training (Guided-ReST), a fine-tuning algorithm designed to improve the model’s capability for effective search during inference. The key insight behind Guided-ReST is that optimal solutions can serve as valuable step-by-step landmarks to guide the model’s search process. Based on this insight, we introduce a novel data generation method that seamlessly incorporates optimal solutions into the model’s search procedure, enabling the generation of high-quality search traces. By fine-tuning the model on these search traces, we effectively distill improved search strategies into the model. Our method significantly enhances the search capabilities of language models on arithmetic reasoning and code self-repair tasks, including Countdown, CodeContests, and CodeForces. We release the source code at https://github.com/snu-mllab/guided-rest.

Tabular reasoning requires complex, multi-step information extraction and logical inference, such as aggregation, comparison, or calculation over tabular data. While recent advances have leveraged large language models (LLMs) for reasoning over structured text tables, such high-quality textual representations are often unavailable in real-world settings, where tables typically appear as images. In this paper, we tackle the task of tabular reasoning directly from table images. Our core strategy is to leverage privileged structured information---specifically, the ground-truth structured table data available during training but inaccessible at test time---to enhance multimodal large language models (MLLMs). The key challenges lie in: accurately aligning visual representations with the structured information, particularly mapping the visual evidence to logical steps; and effectively transferring the reasoning skills learned during training to the MLLM for visual inference. To address these, we introduce {\sc Turbo} (TabUlar Reasoning with Bridged infOrmation), a new framework for multimodal tabular reasoning using privileged information. {\sc Turbo} benefits from a structure-aware reasoning trace generator based on DeepSeek-R1, which contributes to high-quality modality-bridged information. On this basis, {\sc Turbo} repeatedly generates and selects advantageous reasoning traces, further enhancing the model's tabular reasoning ability. Experimental results demonstrate that, with limited (9k) data, {\sc Turbo} achieves state-of-the-art performance ($+7.2\%$ vs. previous SOTA) across multiple datasets.

Scaling model size and training data has led to great advances in the performance of Large Language Models (LLMs). However, the diminishing returns of this approach necessitate alternative methods to improve model capabilities, particularly in tasks requiring advanced reasoning. Large reasoning models, which leverage long chain-of-thoughts, bring unprecedented breakthroughs in problem-solving capabilities but at a substantial deployment cost associated to longer generations. Reducing inference costs is crucial for the economic feasibility, user experience, and environmental sustainability of these models. In this work, we propose to train large reasoning models to reason efficiently. Our method incentivizes models to minimize unnecessary computational overhead while largely maintaining accuracy, thereby achieving substantial deployment efficiency gains. It enables the derivation of a family of reasoning models with varying efficiency levels, controlled via a single hyperparameter. Experiments on two open-weight large reasoning models demonstrate significant reductions in inference cost while preserving most of the accuracy.

Visual reasoning in multimodal large language models (MLLMs) has primarily been studied in passive, static settings, limiting their effectiveness in real-world physical environments where an embodied agent must contend with incomplete information due to occlusion or a limited field of view. Humans, in contrast, leverage their embodiment to actively explore and interact with their environment—moving, examining, and manipulating objects—to gather information through a closed-loop process integrating perception, reasoning, and action. Inspired by this capability, we introduce the Active Visual Reasoning (AVR) task, extending visual reasoning to a paradigm of embodied interaction in partially observable environments. AVR necessitates embodied agents to: (1) actively acquire information via sequential physical actions, (2) integrate observations across multiple steps for coherent reasoning, and (3) dynamically adjust decisions based on evolving visual feedback. To rigorously evaluate AVR, we introduce CLEVR-AVR, a simulation benchmark featuring multi-round interactive environments designed to assess both reasoning correctness and information-gathering efficiency. We present AVR-152k, a large-scale dataset that offers rich Chain-of-Thought (CoT) annotations detailing iterative reasoning for uncertainty identification, action-conditioned information gain prediction, and information-maximizing action selection, crucial for training agents in a higher-order Markov Decision Process. Building on this, we develop PhysVLM-AVR, an embodied MLLM achieving state-of-the-art performance on CLEVR-AVR, embodied reasoning (OpenEQA, RoboVQA), and passive visual reasoning (GeoMath, Geometry30K). Our analysis also reveals that current embodied MLLMs, despite detecting information incompleteness, struggle to actively acquire and integrate new information through interaction, highlighting a fundamental gap in active reasoning capabilities.

Large Language Models (LLMs) achieve superior performance through Chain-of-Thought (CoT) reasoning, but these token-level reasoning chains are computationally expensive and inefficient. In this paper, we introduce Compressed Latent Reasoning (CoLaR), a novel framework that dynamically compresses reasoning processes in latent space through a two-stage training approach. First, during supervised fine-tuning, CoLaR extends beyond next-token prediction by incorporating an auxiliary next compressed embedding prediction objective. This process merges embeddings of consecutive tokens using a compression factor $c$ randomly sampled from a predefined range, and trains a specialized latent head to predict distributions of subsequent compressed embeddings. Second, we enhance CoLaR through reinforcement learning (RL) that leverages the latent head's non-deterministic nature to explore diverse reasoning paths and exploit more compact ones. This approach enables CoLaR to: i) **perform reasoning at a dense latent level** (i.e., silently), substantially reducing reasoning chain length, and ii) **dynamically adjust reasoning speed** at inference time by simply prompting the desired compression factor. Extensive experiments across four mathematical reasoning datasets demonstrate that CoLaR achieves 14.1% higher accuracy than latent-based baseline methods at comparable compression ratios, and reduces reasoning chain length by 53.3% with only 4.8% performance degradation compared to explicit CoT method. Moreover, when applied to more challenging mathematical reasoning tasks, our RL-enhanced CoLaR demonstrates performance gains of up to 5.4% while dramatically reducing latent reasoning chain length by 82.8%. The code and models will be released upon acceptance.

Despite significant advances in Vision Language Models (VLMs), they remain constrained by the complexity and redundancy of visual input. When images contain large amounts of irrelevant information, VLMs are susceptible to interference, thus generating excessive task-irrelevant reasoning processes or even hallucinations. This limitation stems from their inability to discover and process the required regions during reasoning precisely. To address this limitation, we present the Chain of Foresight-Focus Thought (CoFFT), a novel training-free approach that enhances VLMs' visual reasoning by emulating human visual cognition. Each Foresight-Focus Thought consists of three stages: (1) Diverse Sample Generation: generates diverse reasoning samples to explore potential reasoning paths, where each sample contains several reasoning steps; (2) Dual Foresight Decoding: rigorously evaluates these samples based on both visual focus and reasoning progression, adding the first step of optimal sample to the reasoning process; (3) Visual Focus Adjustment: precisely adjust visual focus toward regions most beneficial for future reasoning, before returning to stage (1) to generate subsequent reasoning samples until reaching the final answer. These stages function iteratively, creating an interdependent cycle where reasoning guides visual focus and visual focus informs subsequent reasoning. Empirical results across multiple benchmarks using Qwen2.5-VL, InternVL-2.5, and Llava-Next demonstrate consistent performance improvements of 3.1-5.8\% with controllable increasing computational overhead.

We present SeePhys, a large-scale multimodal benchmark for LLM reasoning grounded in physics questions ranging from middle school to PhD qualifying exams. The benchmark covers 7 fundamental domains spanning the physics discipline, incorporating 21 categories of highly heterogeneous diagrams. In contrast to prior works where visual elements mainly serve auxiliary purposes, our benchmark features a substantial proportion of vision-essential problems (75%) that mandate visual information extraction for correct solutions. Through extensive evaluation, we observe that even the most advanced visual reasoning models (e.g., Gemini-2.5-pro and o4-mini) achieve sub-60% accuracy on our benchmark. These results reveal fundamental challenges in current large language models' visual understanding capabilities, particularly in: (i) establishing rigorous coupling between diagram interpretation and physics reasoning, and (ii) overcoming their persistent reliance on textual cues as cognitive shortcuts. Project Page: github.com/SeePhys/seephys-project Hugging Face: huggingface.co/datasets/SeePhys/SeePhys

Recent advances in reinforcement learning (RL) have strengthened the reasoning capabilities of vision-language models (VLMs). However, enhancing policy exploration to better scale test-time compute remains largely underexplored. In addition, VLMs continue to struggle with imperfect visual perception, which in turn affects the subsequent reasoning process. To this end, we propose **NoisyRollout**, a simple yet effective data augmentation method that mixes trajectories from both clean and moderately distorted images during RL training. By injecting targeted diversity in visual perception and the resulting reasoning patterns, NoisyRollout promotes better policy exploration through vision-oriented inductive biases, ultimately leading to more robust reasoning behaviors. We further adopt a noise annealing schedule that gradually reduces distortion strength over training, leveraging noisy signals early on while ensuring training stability in later stages. Crucially, our method is easy-to-adopt—**requiring no additional training cost and no modifications to the RL objective**. Extensive experiments on $2$ distinct training datasets demonstrate that NoisyRollout achieves state-of-the-art performance among open-source RL-tuned models across $5$ out-of-domain reasoning and perception benchmarks. Furthermore, we validate the effectiveness of NoisyRollout across model sizes ($7$B and $32$B) and data scales (from $1$K to $6$K), highlighting its generalizability and scalability.

Despite promising performance on open-source large vision-language models (LVLMs), transfer-based targeted attacks often fail against closed-source commercial LVLMs. Analyzing failed adversarial perturbations reveals that the learned perturbations typically originate from a uniform distribution and lack clear semantic details, resulting in unintended responses. This critical absence of semantic information leads commercial black-box LVLMs to either ignore the perturbation entirely or misinterpret its embedded semantics, thereby causing the attack to fail. To overcome these issues, we propose to refine semantic clarity by encoding explicit semantic details within local regions, thus ensuring the capture of finer-grained features and inter-model transferability, and by concentrating modifications on semantically rich areas rather than applying them uniformly. To achieve this, we propose *a simple yet highly effective baseline*: at each optimization step, the adversarial image is cropped randomly by a controlled aspect ratio and scale, resized, and then aligned with the target image in the embedding space. While the naive source-target matching method has been utilized before in the literature, we are the first to provide a tight analysis, which establishes a close connection between perturbation optimization and semantics. Experimental results confirm our hypothesis. Our adversarial examples crafted with local-aggregated perturbations focused on crucial regions exhibit surprisingly good transferability to commercial LVLMs, including GPT-4.5, GPT-4o, Gemini-2.0-flash, Claude-3.5/3.7-sonnet, and even reasoning models like o1, Claude-3.7-thinking and Gemini-2.0-flash-thinking. Our approach achieves success rates exceeding 90\% on GPT-4.5, 4o, and o1, significantly outperforming all prior state-of-the-art attack methods with lower $\ell_1/\ell_2$ perturbations. Our optimized adversarial examples under different configurations are available at https://huggingface.co/datasets/MBZUAI-LLM/M-Attack_AdvSamples and our training code at https://github.com/VILA-Lab/M-Attack.