Daily Papers

Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, achieving impressive results across a wide range of tasks. Despite these advances, significant reasoning failures persist, occurring even in seemingly simple scenarios. To systematically understand and address these shortcomings, we present the first comprehensive survey dedicated to reasoning failures in LLMs. We introduce a novel categorization framework that distinguishes reasoning into embodied and non-embodied types, with the latter further subdivided into informal (intuitive) and formal (logical) reasoning. In parallel, we classify reasoning failures along a complementary axis into three types: fundamental failures intrinsic to LLM architectures that broadly affect downstream tasks; application-specific limitations that manifest in particular domains; and robustness issues characterized by inconsistent performance across minor variations. For each reasoning failure, we provide a clear definition, analyze existing studies, explore root causes, and present mitigation strategies. By unifying fragmented research efforts, our survey provides a structured perspective on systemic weaknesses in LLM reasoning, offering valuable insights and guiding future research towards building stronger, more reliable, and robust reasoning capabilities. We additionally release a comprehensive collection of research works on LLM reasoning failures, as a GitHub repository at https://github.com/Peiyang-Song/Awesome-LLM-Reasoning-Failures, to provide an easy entry point to this area.

Answering multi-hop reasoning questions requires retrieving and synthesizing information from diverse sources. Large Language Models (LLMs) struggle to perform such reasoning consistently. Here we propose an approach to pinpoint and rectify multi-hop reasoning failures through targeted memory injections on LLM attention heads. First, we analyze the per-layer activations of GPT-2 models in response to single and multi-hop prompts. We then propose a mechanism that allows users to inject pertinent prompt-specific information, which we refer to as "memories," at critical LLM locations during inference. By thus enabling the LLM to incorporate additional relevant information during inference, we enhance the quality of multi-hop prompt completions. We show empirically that a simple, efficient, and targeted memory injection into a key attention layer can often increase the probability of the desired next token in multi-hop tasks, by up to 424%.

This paper investigates the mathematical reasoning capabilities of large language models (LLMs) using 50 newly constructed high-school-level word problems. Unlike prior studies that focus solely on answer correctness, we rigorously analyze both final answers and solution steps to identify reasoning failures. Evaluating eight state-of-the-art models - including Mixtral, Llama, Gemini, GPT-4o, and OpenAI's o1 variants - we find that while newer models (e.g., o3-mini, deepseek-r1) achieve higher accuracy, all models exhibit errors in spatial reasoning, strategic planning, and arithmetic, sometimes producing correct answers through flawed logic. Common failure modes include unwarranted assumptions, over-reliance on numerical patterns, and difficulty translating physical intuition into mathematical steps. Manual analysis reveals that models struggle with problems requiring multi-step deduction or real-world knowledge, despite possessing broad mathematical knowledge. Our results underscore the importance of evaluating reasoning processes, not just answers, and caution against overestimating LLMs' problem-solving proficiency. The study highlights persistent gaps in LLMs' generalization abilities, emphasizing the need for targeted improvements in structured reasoning and constraint handling.

Recent advancements in large language models (LLMs) have propelled Artificial Intelligence (AI) to new heights, enabling breakthroughs in various tasks such as writing assistance, code generation, and machine translation. A significant distinction of advanced LLMs, such as ChatGPT, is their demonstrated ability to "reason." However, evaluating the reasoning ability of LLMs remains a challenge as most existing evaluations focus on their accuracy on the downstream tasks rather than directly assessing their reasoning processes. Efforts have been made to develop benchmarks and metrics to assess reasoning in LLMs, but they suffer from data leakage or limited scope. In this paper, we introduce LogicAsker, an automatic approach that comprehensively evaluates and improves the logical reasoning abilities of LLMs under a set of atomic reasoning skills based on propositional and predicate logic. The results provide insights into LLMs' reasoning abilities and reveal the logical rules the LLMs did not learn well. We evaluate LogicAsker on six widely deployed LLMs, including GPT-3, ChatGPT, GPT-4, Bard, Vicuna, and Guanaco. The results show that test cases from LogicAsker can find logical reasoning failures in different LLMs with a rate of 25\% - 94\%. In addition, the test cases of LogicAsker can be further used to design demonstration examples for in-context learning, which effectively improves the logical reasoning ability of LLMs, e.g., 10\% for GPT-4. As far as we know, our work is the first to create prompts based on testing results to improve LLMs' formal reasoning ability effectively. All the code, data, and results will be released for reproduction and future research.

Root cause analysis (RCA) is essential for diagnosing failures within complex software systems to ensure system reliability. The highly distributed and interdependent nature of modern cloud-based systems often complicates RCA efforts, particularly for multi-hop fault propagation, where symptoms appear far from their true causes. Recent advancements in Large Language Models (LLMs) present new opportunities to enhance automated RCA. However, their practical value for RCA depends on the fidelity of reasoning and decision-making. Existing work relies on historical incident corpora, operates directly on high-volume telemetry beyond current LLM capacity, or embeds reasoning inside complex multi-agent pipelines -- conditions that obscure whether failures arise from reasoning itself or from peripheral design choices. We present a focused empirical evaluation that isolates an LLM's reasoning behavior. We design a controlled experimental framework that foregrounds the LLM by using a simplified experimental setting. We evaluate six LLMs under two agentic workflows (ReAct and Plan-and-Execute) and a non-agentic baseline on two real-world case studies (GAIA and OpenRCA). In total, we executed 48,000 simulated failure scenarios, totaling 228 days of execution time. We measure both root-cause accuracy and the quality of intermediate reasoning traces. We produce a labeled taxonomy of 16 common RCA reasoning failures and use an LLM-as-a-Judge for annotation. Our results clarify where current open-source LLMs succeed and fail in multi-hop RCA, quantify sensitivity to input data modalities, and identify reasoning failures that predict final correctness. Together, these contributions provide transparent and reproducible empirical results and a failure taxonomy to guide future work on reasoning-driven system diagnosis.

Reasoning failures in large language models (LLMs) are typically measured only at the end of a generation, yet many failures manifest as a process-level breakdown: the model "loses the thread" mid-reasoning. We study whether such breakdowns are detectable from inference-time observables available in standard APIs (token log probabilities), without any training or fine-tuning. We define a simple instability signal that combines consecutive-step distributional shift (JSD) and uncertainty (entropy), summarize each trace by its peak instability strength, and show that this signal reliably predicts failure. Across GSM8K and HotpotQA, instability strength predicts wrong answers with above-chance AUC and yields monotonic bucket-level accuracy decline at scale across model sizes. Crucially, we show that instability is not uniformly harmful: early instability can reflect subsequent stabilization and a correct final answer (corrective instability), whereas late instability is more often followed by failure (destructive instability), even at comparable peak magnitudes, indicating that recoverability depends not only on how strongly the distribution changes but also on when such changes occur relative to the remaining decoding horizon. The method is model-agnostic, training-free, and reproducible, and is presented as a diagnostic lens rather than a corrective or control mechanism.

While Large Language Models (LLMs) achieve high performance on standard mathematical benchmarks, their underlying reasoning processes remain highly overfit to standard textual formatting. We propose a perturbation pipeline consisting of 14 techniques to evaluate robustness of LLM reasoning. We apply this pipeline to AIME 2024 dataset and evalute 8 state-of-the-art models on the resulting benchmark. While frontier models exhibit resilience, open weights reasoning models suffer catastrophic collapses (up to 55% average accuracy drops across perturbations and up to 100% on some), exposing structural fragility. To further disentangle mechanical parsing failures from downstream reasoning failures, we strictly isolate the models' working memory capacity by forcing models to solve multiple unperturbed mathematical problems sequentially within a single context window. Our results indicate that open weight models ranging from 7B to 120B parameters and Claude Opus 4.6 exhibit accuracy decay on subsequent problems. This degradation demonstrates that intermediate reasoning steps permanently pollute standard dense attention mechanisms. We argue that to achieve reliable reasoning, future reasoning architectures must integrate explicit contextual resets within a model's own Chain-of-Thought, leading to fundamental open questions regarding the optimal granularity of atomic reasoning tasks.

Understanding how Large Language Models (LLMs) perform complex reasoning and their failure mechanisms is a challenge in interpretability research. To provide a measurable geometric analysis perspective, we define the concept of the Reasoning Manifold, a latent low-dimensional geometric structure formed by the internal representations corresponding to all correctly reasoned generations. This structure can be conceptualized as the embodiment of the effective thinking paths that the model has learned to successfully solve a given task. Based on this concept, we build REMA, a framework that explains the origins of failures by quantitatively comparing the spatial relationships of internal model representations corresponding to both erroneous and correct reasoning samples. Specifically, REMA first quantifies the geometric deviation of each erroneous representation by calculating its k-nearest neighbors distance to the approximated manifold formed by correct representations, thereby providing a unified failure signal. It then localizes the divergence points where these deviations first become significant by tracking this deviation metric across the model's layers and comparing it against a baseline of internal fluctuations from correct representations, thus identifying where the reasoning chain begins to go off-track. Our extensive experiments on diverse language and multimodal models and tasks demonstrate the low-dimensional nature of the reasoning manifold and the high separability between erroneous and correct reasoning representations. The results also validate the effectiveness of the REMA framework in analyzing the origins of reasoning failures. This research connects abstract reasoning failures to measurable geometric deviations in representations, providing new avenues for in-depth understanding and diagnosis of the internal computational processes of black-box models.

The emergence of reasoning models and their integration into practical AI chat bots has led to breakthroughs in solving advanced math, deep search, and extractive question answering problems that requires a complex and multi-step thought process. Yet, a complete understanding of why these models hallucinate more than general purpose language models is missing. In this investigative study, we systematicallyexplore reasoning failures of contemporary language models on multi-hop question answering tasks. We introduce a novel, nuanced error categorization framework that examines failures across three critical dimensions: the diversity and uniqueness of source documents involved ("hops"), completeness in capturing relevant information ("coverage"), and cognitive inefficiency ("overthinking"). Through rigorous hu-man annotation, supported by complementary automated metrics, our exploration uncovers intricate error patterns often hidden by accuracy-centric evaluations. This investigative approach provides deeper insights into the cognitive limitations of current models and offers actionable guidance toward enhancing reasoning fidelity, transparency, and robustness in future language modeling efforts.

Despite the success of test-time scaling, Large Reasoning Models (LRMs) frequently encounter repetitive loops that lead to computational waste and inference failure. In this paper, we identify a distinct failure mode termed Circular Reasoning. Unlike traditional model degeneration, this phenomenon manifests as a self-reinforcing trap where generated content acts as a logical premise for its own recurrence, compelling the reiteration of preceding text. To systematically analyze this phenomenon, we introduce LoopBench, a dataset designed to capture two distinct loop typologies: numerical loops and statement loops. Mechanistically, we characterize circular reasoning as a state collapse exhibiting distinct boundaries, where semantic repetition precedes textual repetition. We reveal that reasoning impasses trigger the loop onset, which subsequently persists as an inescapable cycle driven by a self-reinforcing V-shaped attention mechanism. Guided by these findings, we employ the Cumulative Sum (CUSUM) algorithm to capture these precursors for early loop prediction. Experiments across diverse LRMs validate its accuracy and elucidate the stability of long-chain reasoning.

Current evaluation of mathematical reasoning in language models relies primarily on answer accuracy, potentially masking fundamental failures in logical computation. We introduce a diagnostic framework that distinguishes genuine mathematical reasoning from superficial pattern matching through four complementary axes: forward-backward consistency, transitivity coverage, counterfactual sensitivity, and perturbation robustness. Through a case study applying this framework to Qwen3-0.6B on the MenatQA dataset, we reveal a striking disconnect between surface performance and reasoning fidelity. While the model achieves reasonable answer accuracy (70%+), it demonstrates poor backward consistency (15%), limited transitivity coverage (32.2%), and brittle sensitivity to perturbations. Our diagnostics expose reasoning failures invisible to traditional accuracy metrics, suggesting that this small model relies heavily on pattern matching rather than genuine logical computation. While our empirical findings are based on a single 600M-parameter model, the diagnostic framework itself is model-agnostic and generalizable. We release our evaluation protocols to enable the research community to assess reasoning fidelity across different model scales and architectures, moving beyond surface-level accuracy toward verifiable mathematical reasoning.

Large language models (LLMs) have recently shown strong progress on scientific reasoning, yet two major bottlenecks remain. First, explicit retrieval fragments reasoning, imposing a hidden "tool tax" of extra tokens and steps. Second, multi-agent pipelines often dilute strong solutions by averaging across all candidates. We address these challenges with a unified framework that combines implicit retrieval and structured collaboration. At its foundation, a Monitor-based retrieval module operates at the token level, integrating external knowledge with minimal disruption to reasoning. On top of this substrate, Hierarchical Solution Refinement (HSR) iteratively designates each candidate as an anchor to be repaired by its peers, while Quality-Aware Iterative Reasoning (QAIR) adapts refinement to solution quality. On Humanity's Last Exam (HLE) Bio/Chem Gold, our framework achieves 48.3\% accuracy -- the highest reported to date, surpassing the strongest agent baseline by 13.4 points and leading frontier LLMs by up to 18.1 points, while simultaneously reducing token usage by 53.5\% and agent steps by 43.7\%. Results on SuperGPQA and TRQA confirm robustness across domains. Error analysis shows that reasoning failures and knowledge gaps co-occur in over 85\% of cases, while diversity analysis reveals a clear dichotomy: retrieval tasks benefit from solution variety, whereas reasoning tasks favor consensus. Together, these findings demonstrate how implicit augmentation and structured refinement overcome the inefficiencies of explicit tool use and uniform aggregation. Code is available at: https://github.com/tangxiangru/Eigen-1.

Large language models systematically fail when a salient surface cue conflicts with an unstated feasibility constraint. We study this through a diagnose-measure-bridge-treat framework. Causal-behavioral analysis of the ``car wash problem'' across six models reveals approximately context-independent sigmoid heuristics: the distance cue exerts 8.7 to 38 times more influence than the goal, and token-level attribution shows patterns more consistent with keyword associations than compositional inference. The Heuristic Override Benchmark (HOB) -- 500 instances spanning 4 heuristic by 5 constraint families with minimal pairs and explicitness gradients -- demonstrates generality across 14 models: under strict evaluation (10/10 correct), no model exceeds 75%, and presence constraints are hardest (44%). A minimal hint (e.g., emphasizing the key object) recovers +15 pp on average, suggesting the failure lies in constraint inference rather than missing knowledge; 12/14 models perform worse when the constraint is removed (up to -39 pp), revealing conservative bias. Parametric probes confirm that the sigmoid pattern generalizes to cost, efficiency, and semantic-similarity heuristics; goal-decomposition prompting recovers +6 to 9 pp by forcing models to enumerate preconditions before answering. Together, these results characterize heuristic override as a systematic reasoning vulnerability and provide a benchmark for measuring progress toward resolving it.

Multimodal Large Language Models (MLLMs) strive to achieve a profound, human-like understanding of and interaction with the physical world, but often exhibit a shallow and incoherent integration when acquiring information (Perception) and conducting reasoning (Cognition). This disconnect leads to a spectrum of reasoning failures, with hallucination being the most prominent. Collectively, these issues expose a fundamental challenge: the ability to process pixels does not yet confer the ability to construct a coherent, credible internal world model. To systematically dissect and address this challenge, this survey introduces a novel and unified analytical framework: ``From Perception to Cognition." We deconstruct the complex process of vision-language interactive understanding into two interdependent layers: Perception, the foundational ability to accurately extract visual information and achieve fine-grained alignment with textual instructions; and Cognition, the higher-order capability for proactive, multi-step, goal-oriented reasoning built upon this perceptual foundation, the core of which is the formation of a dynamic observe-think-verify reasoning loop. Guided by this framework, this paper systematically analyzes the key bottlenecks of current MLLMs at both layers. It surveys the landscape of cutting-edge methods designed to address these challenges, spanning from techniques that enhance low-level visual representations to those that improve high-level reasoning paradigms. Furthermore, we review critical benchmarks and delineate future research directions. This survey aims to provide the research community with a clear, structured perspective for understanding the intrinsic limitations of current MLLMs and to illuminate the path toward building next-generation models capable of deep reasoning and a genuine understanding of the world.

Video foundation models generate visually realistic and temporally coherent content, but their reliability as world simulators depends on whether they capture physical, logical, and spatial constraints. Existing metrics such as Frechet Video Distance (FVD) emphasize perceptual quality and overlook reasoning failures, including violations of causality, physics, and global consistency. We introduce MMGR (Multi-Modal Generative Reasoning Evaluation and Benchmark), a principled evaluation framework based on five reasoning abilities: Physical, Logical, 3D Spatial, 2D Spatial, and Temporal. MMGR evaluates generative reasoning across three domains: Abstract Reasoning (ARC-AGI, Sudoku), Embodied Navigation (real-world 3D navigation and localization), and Physical Commonsense (sports and compositional interactions). MMGR applies fine-grained metrics that require holistic correctness across both video and image generation. We benchmark leading video models (Veo-3, Sora-2, Wan-2.2) and image models (Nano-banana, Nano-banana Pro, GPT-4o-image, Qwen-image), revealing strong performance gaps across domains. Models show moderate success on Physical Commonsense tasks but perform poorly on Abstract Reasoning (below 10 percent accuracy on ARC-AGI) and struggle with long-horizon spatial planning in embodied settings. Our analysis highlights key limitations in current models, including overreliance on perceptual data, weak global state consistency, and objectives that reward visual plausibility over causal correctness. MMGR offers a unified diagnostic benchmark and a path toward reasoning-aware generative world models.

Large language models (LLMs) excel at solving problems with clear and complete statements, but often struggle with nuanced environments or interactive tasks which are common in most real-world scenarios. This highlights the critical need for developing LLMs that can effectively engage in logically consistent multi-turn dialogue, seek information and reason with incomplete data. To this end, we introduce a novel benchmark comprising a suite of multi-turn tasks each designed to test specific reasoning, interactive dialogue, and information-seeking abilities. These tasks have deterministic scoring mechanisms, thus eliminating the need for human intervention. Evaluating frontier models on our benchmark reveals significant headroom. Our analysis shows that most errors emerge from poor instruction following, reasoning failures, and poor planning. This benchmark provides valuable insights into the strengths and weaknesses of current LLMs in handling complex, interactive scenarios and offers a robust platform for future research aimed at improving these critical capabilities.

Large language models (LLMs) reason over discrete token ID sequences, yet modern subword tokenizers routinely produce non-unique encodings: multiple token ID sequences can detokenize to identical surface strings. This representational mismatch creates an unmeasured fragility wherein reasoning processes can fail. LLMs may treat two internal representations as distinct "words" even when they are semantically identical at the text level. In this work, we show that tokenization can betray LLM reasoning through one-to-many token ID mappings. We introduce a tokenization-consistency probe that requires models to replace designated target words in context while leaving all other content unchanged. The task is intentionally simple at the surface level, enabling us to attribute failures to tokenizer-detokenizer artifacts rather than to knowledge gaps or parameter limitations. Through analysis of over 11000 replacement trials across state-of-the-art open-source LLMs, we find a non-trivial rate of outputs exhibit phantom edits: cases where models operate under the illusion of correct reasoning, a phenomenon arising from tokenizer-induced representational defects. We further analyze these cases and provide a taxonomy of eight systematic tokenizer artifacts, including whitespace-boundary shifts and intra-word resegmentation. These findings indicate that part of apparent reasoning deficiency originates in the tokenizer layer, motivating tokenizer-level remedies before incurring the cost of training ever-larger models on ever-larger corpora.

Reasoning-focused large language models (LLMs) have advanced in many NLP tasks, yet their evaluation remains challenging: final answers alone do not expose the intermediate reasoning steps, making it difficult to determine whether a model truly reasons correctly and where failures occur, while existing multi-hop QA benchmarks lack step-level annotations for diagnosing reasoning failures. To address this gap, we propose Omanic, an open-domain multi-hop QA resource that provides decomposed sub-questions and intermediate answers as structural annotations for analyzing reasoning processes. It contains 10,296 machine-generated training examples (OmanicSynth) and 967 expert-reviewed human-annotated evaluation examples (OmanicBench). Systematic evaluations show that state-of-the-art LLMs achieve only 73.11% multiple-choice accuracy on OmanicBench, confirming its high difficulty. Stepwise analysis reveals that CoT's performance hinges on factual completeness, with its gains diminishing under knowledge gaps and errors amplifying in later hops. Additionally, supervised fine-tuning on OmanicSynth brings substantial transfer gains (7.41 average points) across six reasoning and math benchmarks, validating the dataset's quality and further supporting the effectiveness of OmanicSynth as supervision for reasoning-capability transfer. We release the data at https://huggingface.co/datasets/li-lab/Omanic and the code at https://github.com/XiaojieGu/Omanic.

Large reasoning models (e.g., R1, o3) have demonstrated remarkable mathematical problem-solving abilities. However, the high reported accuracy of these advanced models on popular datasets, reliance on purely numerical evaluation and potential benchmark leakage, often masks their true reasoning shortcomings. To address this, we propose leveraging the inherent rigor and methodological complexity of mathematical proofs as a diagnostic tool to expose these hidden failures. Specifically, we introduce the RFMDataset (Reveal Failure Modes), a collection of 200 diverse mathematical proof problems, and thoroughly evaluate advanced models' performance on it. Our in-depth analysis of their failures uncovers 10 fine-grained error types, which shows fundamental limitations in current large reasoning models: 1) large reasoning models grapple profoundly with mathematical proofs, with some generating entirely correct proofs for less than 20% of problems and failing even on basic ones; 2) models exhibit a diverse spectrum of reasoning failures, prominently demonstrating the lack of guarantees for the correctness and rigor of single-step reasoning; and 3) models show hallucination and incompleteness during the reasoning process. Our findings reveal that models' self-reflection is insufficient to resolve the current logical dilemmas, necessitating formalized and fine-grained logical training.

Recent multimodal large language models have shown great promise in clinical image reasoning, but existing post-training pipelines remain predominantly outcome-centric, relying on final answer correctness or sequence-level preferences. This suffers from sparse credit assignment, making it difficult to optimize the reasoning process essential for clinical applications. Our analysis reveals that cascading errors from early-stage reasoning failures are a leading cause of incorrect predictions in medical visual question answering (VQA) benchmarks. Motivated by this, we propose Medical Reasoning-aware Policy Optimization (MRPO), an RL algorithm that incorporates step-wise process rewards. When the final answer is incorrect, MRPO assigns exponentially larger penalties to tokens in earlier invalid reasoning steps, breaking failure cascades without compromising successful paths. Across three multimodal LLM backbones, MRPO consistently outperforms standard GRPO and a recent RL baseline, and on Qwen3-VL-8B-Instruct even surpasses substantially larger medical MLLMs such as HuatuoGPT-Vision-34B by 2.79 points. Moreover, MRPO reduces early-stage reasoning failures from 64.0% to 13.0%, showing that targeted mitigation of cascading failures improves both reasoning quality and final answer accuracy. Our code is available at https://github.com/dmis-lab/MRPO

Building trustworthy medical multimodal large language models (MLLMs) is critical for reliable clinical decision support. Existing medical hallucination benchmarks mainly focus on data collection, but often ignore where hallucinations originate within the reasoning process. We find that hallucination sources vary across samples: errors may arise from visual misrecognition, incorrect medical knowledge recall, or flawed reasoning integration. To enable source-level hallucination diagnosis, we introduce ClinHallu, a benchmark for stage-wise hallucination diagnosis in medical MLLM reasoning. ClinHallu contains 7,031 validated instances, where each instance is augmented with a structured reasoning trace decomposed into Visual Recognition, Knowledge Recall, and Reasoning Integration. We also use stage-replacement interventions to measure how correcting specific stages affects the final answer. Beyond evaluation, we show that trace-supervised fine-tuning reduces stage-wise hallucinations. ClinHallu provides a fine-grained hallucination testbed for diagnosing and mitigating reasoning failures in medical MLLMs. The benchmark is publicly available at https://github.com/alibaba-damo-academy/ClinHallu.

On-Policy Distillation (OPD) improves large language model reasoning by training a student model on trajectories sampled from its own policy under teacher supervision. Although OPD operates on trajectories, its learning signal remains token-level: it identifies deviations through high-loss tokens and repairs them through local reverse-KL correction. We show that this "trajectory-sampled but token-learned" mechanism cannot reliably bridge student trajectories toward teacher trajectories. About 30% of high-loss tokens fall into the low-divergence regime, indicating that many are surface-form mismatches rather than real reasoning forks. Moreover, even truly divergent tokens are difficult to repair with isolated token-level supervision, since reasoning failures often unfold as short-horizon distributional drift. We propose Trajectory-aware OPD (TOPD), which uses near-future trajectory information to identify real divergent states and distribute guidance across multiple future tokens. Experiments show that suppressing non-divergent high-loss tokens improves standard OPD from 47.8% to 48.2% average accuracy, while TOPD further improves performance to 52.2%, with gains on AIME24 from 60.0% to 63.3% and AIME25 from 46.7% to 53.3%.

Multimodal hallucination in multimodal large language models (MLLMs) restricts the correctness of MLLMs. However, multimodal hallucinations are multi-sourced and arise from diverse causes. Existing benchmarks fail to adequately distinguish between perception-induced hallucinations and reasoning-induced hallucinations. This failure constitutes a significant issue and hinders the diagnosis of multimodal reasoning failures within MLLMs. To address this, we propose the {\dataset} benchmark, which isolates reasoning hallucinations by constructing questions where input images are correctly perceived by MLLMs yet reasoning errors persist. {\dataset} introduces multi-granular evaluation metrics: accuracy, factuality, and LLMs hallucination score for hallucination quantification. Our analysis reveals that (1) the model scale, data scale, and training stages significantly affect the degree of logical, fabrication, and factual hallucinations; (2) current MLLMs show no effective improvement on spatial hallucinations caused by misinterpreted spatial relationships, indicating their limited visual reasoning capabilities; and (3) question types correlate with distinct hallucination patterns, highlighting targeted challenges and potential mitigation strategies. To address these challenges, we propose {\method}, a method that combines curriculum reinforcement fine-tuning to encourage models to generate logic-consistent reasoning chains by stepwise reducing learning difficulty, and collaborative hint inference to reduce reasoning complexity. {\method} establishes a baseline on {\dataset}, and reduces the logical hallucinations in original base models.

This paper explores multi-turn visual reasoning and observes that MLLMs repeatedly fail to localize the target, leading to long, redundant trajectories. We attribute this failure to the entanglement of reasoning and perception within a single model, the MLLM reasons and localizes simultaneously, and inaccurate localization triggers additional reasoning turns that bloat the trajectory. To solve this problem, we propose PixelEyes, a multi-turn visual reasoning agent that explicitly decouples reasoning from perception, i.e., the reasoner decides what to look for, while a specialized perception tool answers where it is. Specifically, PixelEyes introduces 1) Mask-guided Visual Search. A referring segmentation model is invoked to provide mask-precise localization, freeing the reasoner from the need to compensate for imprecise grounding. 2) Semantic-region Breadth-first Search (BFS). To eliminate redundant loops caused by repeatedly cropping incorrect sub-regions, we organize exploration as a breadth-first search over semantic regions. To internalize these capabilities, we construct the PixelEyes-6K dataset by resynthesizing expert trajectories from existing data. This explicitly embeds our mask-guided search and BFS logic into the model. We further introduce Pinpoint-Bench, a zero-hint visual search benchmark, i.e., no location cues are provided in the question, with instance-level masks and bounding boxes that separate localization failures from reasoning failures, enabling fine-grained analysis of failure modes such as inattentional blindness. Recent state-of-the-art MLLMs and visual reasoning agents leave large headroom on Pinpoint-Bench, demonstrating its quality and difficulty. Code and models are open-sourced.

Large language models (LLMs) solve complex problems yet fail on simpler variants, suggesting they achieve correct outputs through mechanisms fundamentally different from human reasoning. To understand this gap, we synthesize cognitive science research into a taxonomy of 28 cognitive elements spanning reasoning invariants, meta-cognitive controls, representations for organizing reasoning & knowledge, and transformation operations. We introduce a fine-grained evaluation framework and conduct the first large-scale empirical analysis of 192K traces from 18 models across text, vision, and audio, complemented by 54 human think-aloud traces, which we make publicly available. We find that models under-utilize cognitive elements correlated with success, narrowing to rigid sequential processing on ill-structured problems where diverse representations and meta-cognitive monitoring are critical. Human traces show more abstraction and conceptual processing, while models default to surface-level enumeration. Meta-analysis of 1.6K LLM reasoning papers reveals the research community concentrates on easily quantifiable elements (sequential organization: 55%, decomposition: 60%) but neglecting meta-cognitive controls (self-awareness: 16%) that correlate with success. Models possess behavioral repertoires associated with success but fail to deploy them spontaneously. Leveraging these patterns, we develop test-time reasoning guidance that automatically scaffold successful structures, improving performance by up to 66.7% on complex problems. By establishing a shared vocabulary between cognitive science and LLM research, our framework enables systematic diagnosis of reasoning failures and principled development of models that reason through robust cognitive mechanisms rather than spurious shortcuts, while providing tools to test theories of human cognition at scale.

Tool-augmented Language Models (TaLMs) can invoke external tools to solve problems beyond their parametric capacity. However, it remains unclear whether these tool-enabled gains reflect trustworthy reasoning. Focusing on the Code Interpreter tool, we show that even when tools are selected and executed correctly, TaLMs treat tool outputs as substitutes for reasoning, producing solutions that appear correct but lack coherent justification. We term this failure mode Tool-Induced Myopia (TIM), and study it using PYMATH, a benchmark of 1,679 competition-level mathematical problems for which Python code is helpful but not sufficient. We further develop a multi-dimensional evaluation suite to quantify reasoning degradation in TaLMs relative to their non-tool counterparts. Our findings reveal that while TaLMs achieve up to a 19.3 percentage point gain in final-answer accuracy, their reasoning behavior consistently deteriorates (e.g., non-tool LLMs win up to 41.5% more often in pairwise comparisons of the reasoning process). This degradation intensifies with tool use; the more frequently a model invokes tools, the less coherent its reasoning becomes. Moreover, tool use shifts errors from arithmetic mistakes toward global reasoning failures (logic, assumption, creativity); with TIM present in ~55% of high-risk cases. Finally, we propose a preference-optimization-based framework that realigns TaLMs to use tools as assistive evidence, improving both final-answer accuracy and reasoning depth under tool use. Codes and data are available at: https://github.com/megagonlabs/TIM.

Recent advances in large language models (LLMs) have led to strong reasoning capabilities; however, evaluating such models in low-resource languages remains challenging due to the lack of standardized benchmarks. In particular, Urdu reasoning evaluation has been limited by the sensitivity of machine translation and an emphasis on general language tasks rather than reasoning benchmarks. In this paper, we propose a contextually ensembled translation framework with human-in-the-loop validation that leverages multiple translation systems to develop Urdu reasoning benchmarks while preserving contextual and structural integrity. Using this framework, we translate widely adopted reasoning and question-answering benchmarks, including MGSM, MATH-500, CommonSenseQA, and OpenBookQA, into Urdu, collectively referred to as UrduBench, and conduct a comprehensive evaluation of both reasoning-oriented and instruction-tuned LLMs across multiple prompting strategies. Our analysis reveals performance differences across (1) four datasets, (2) five task difficulty levels, (3) diverse model architectures, (4) multiple model scaling settings, and (5) language consistency tests. We find that multi-step and symbolic reasoning tasks pose significant challenges in Urdu, and that stable language alignment is a critical prerequisite for robust reasoning. Overall, our work establishes a scalable methodology for standardized reasoning evaluation in Urdu and provides empirical insights into multilingual reasoning failures. This experimental setup is also broadly applicable to other low-resource languages. The code and datasets will be publicly released.

Graph-RAG systems achieve strong multi-hop question answering by indexing documents into knowledge graphs, but strong retrieval does not guarantee strong answers. Evaluating KET-RAG, a leading Graph-RAG system, on three multi-hop QA benchmarks (HotpotQA, MuSiQue, 2WikiMultiHopQA), we find that 77% to 91% of questions have the gold answer in the retrieved context, yet accuracy is only 35% to 78%, and 73% to 84% of errors are reasoning failures. We propose two augmentations: (i) SPARQL chain-of-thought prompting, which decomposes questions into triple-pattern queries aligned with the entity-relationship context, and (ii) graph-walk compression, which compresses the context by ~60% via knowledge-graph traversal with no LLM calls. SPARQL CoT improves accuracy by +2 to +14 pp; graph-walk compression adds +6 pp on average when paired with structured prompting on smaller models. Surprisingly, we show that, with question-type routing, a fully augmented budget open-weight Llama-8B model matches or exceeds the unaugmented Llama-70B baseline on all three benchmarks at ~12x lower cost. A replication on LightRAG confirms that our augmentations transfer across Graph-RAG systems.

Understanding and reasoning about the physical world is the foundation of intelligent behavior, yet state-of-the-art vision-language models (VLMs) still fail at causal physical reasoning, often producing plausible but incorrect answers. To address this gap, we introduce CausalPhys, a benchmark of over 3,000 carefully curated video- and image-based questions spanning four domains: Perception, Anticipation, Intervention, and Goal Orientation. Each question is paired with an expert-annotated causal graph capturing object-attribute-event dependencies, enabling interpretable and fine-grained evaluation of causal understanding. Building on this, we formulate a causal-graph-grounded metric that quantitatively measures how well a model's chain-of-thought reasoning aligns with the correct causal relations, moving beyond answer-only accuracy and enabling systematic diagnosis of VLMs' causal reasoning failures. Using this metric, we conduct a comprehensive analysis of leading VLMs, revealing systematic gaps in capturing causal dependencies and underscoring the need for causality-aware learning. To address these limitations, we further propose Causal Rationale-informed Fine-Tuning (CRFT), which explicitly aligns VLM reasoning with causal structures. Extensive experiments demonstrate that CRFT substantially enhances both reasoning accuracy and interpretability across multiple model backbones. By unifying dataset curation, causal evaluation, and causality-informed learning, CausalPhys establishes a strong foundation for advancing modern VLMs toward causally grounded physical reasoning.

Analogical reasoning is at the core of human cognition, serving as an important foundation for a variety of intellectual activities. While prior work has shown that LLMs can represent task patterns and surface-level concepts, it remains unclear whether these models can encode high-level relational concepts and apply them to novel situations through structured comparisons. In this work, we explore this fundamental aspect using proportional and story analogies, and identify three key findings. First, LLMs effectively encode the underlying relationships between analogous entities; both attributive and relational information propagate through mid-upper layers in correct cases, whereas reasoning failures reflect missing relational information within these layers. Second, unlike humans, LLMs often struggle not only when relational information is missing, but also when attempting to apply it to new entities. In such cases, strategically patching hidden representations at critical token positions can facilitate information transfer to a certain extent. Lastly, successful analogical reasoning in LLMs is marked by strong structural alignment between analogous situations, whereas failures often reflect degraded or misplaced alignment. Overall, our findings reveal that LLMs exhibit emerging but limited capabilities in encoding and applying high-level relational concepts, highlighting both parallels and gaps with human cognition.

Recent Vision-Language-Action (VLA) models report impressive success rates on standard robotic benchmarks, fueling optimism about general-purpose physical intelligence. However, recent evidence suggests a systematic misalignment between standard benchmark success and true embodied reasoning, raising the question of whether these high scores reflect genuine cognitive capability. To address this gap, we introduce BeTTER, a diagnostic Benchmark for Testing True Embodied Reasoning in robotic policies. BeTTER applies targeted causal interventions (e.g., spatial layout shifts, temporal extrapolation) while enforcing kinematic isolation to explicitly decouple high-level reasoning failures from low-level execution limits. Through systematic evaluation, we reveal that state-of-the-art VLAs catastrophically fail in dynamic scenarios, exhibiting severe lexical-kinematic shortcuts, behavioral inertia, and semantic feature collapse. Crucially, our mechanistic analysis traces these symptoms to fundamental architectural bottlenecks - such as capacity compression and myopic downsampling - which systematically degrade the model's foundational semantic representation. We demonstrate that highly static evaluation protocols effectively mask this degradation by allowing optimization to overfit to sensorimotor priors. Supported by real-world robotic validation, our findings confirm that this representational breakdown is not a simulation artifact, highlighting the critical need for future VLA paradigms to resolve the structural tension between high-frequency control and high-level reasoning.

Spatial reasoning, the ability to ground language in 3D understanding, remains a persistent challenge for Vision-Language Models (VLMs). We identify two fundamental bottlenecks: inadequate 3D understanding capabilities stemming from 2D-centric pre-training, and reasoning failures induced by redundant 3D information. To address these, we first construct a Minimal Sufficient Set (MSS) of information before answering a given question: a compact selection of 3D perception results from expert models. We introduce MSSR (Minimal Sufficient Spatial Reasoner), a dual-agent framework that implements this principle. A Perception Agent programmatically queries 3D scenes using a versatile perception toolbox to extract sufficient information, including a novel SOG (Situated Orientation Grounding) module that robustly extracts language-grounded directions. A Reasoning Agent then iteratively refines this information to pursue minimality, pruning redundant details and requesting missing ones in a closed loop until the MSS is curated. Extensive experiments demonstrate that our method, by explicitly pursuing both sufficiency and minimality, significantly improves accuracy and achieves state-of-the-art performance across two challenging benchmarks. Furthermore, our framework produces interpretable reasoning paths, offering a promising source of high-quality training data for future models. Source code is available at https://github.com/gyj155/mssr.

Shojaee et al. (2025) report that Large Reasoning Models (LRMs) exhibit "accuracy collapse" on planning puzzles beyond certain complexity thresholds. We demonstrate that their findings primarily reflect experimental design limitations rather than fundamental reasoning failures. Our analysis reveals three critical issues: (1) Tower of Hanoi experiments systematically exceed model output token limits at reported failure points, with models explicitly acknowledging these constraints in their outputs; (2) The authors' automated evaluation framework fails to distinguish between reasoning failures and practical constraints, leading to misclassification of model capabilities; (3) Most concerningly, their River Crossing benchmarks include mathematically impossible instances for N > 5 due to insufficient boat capacity, yet models are scored as failures for not solving these unsolvable problems. When we control for these experimental artifacts, by requesting generating functions instead of exhaustive move lists, preliminary experiments across multiple models indicate high accuracy on Tower of Hanoi instances previously reported as complete failures. These findings highlight the importance of careful experimental design when evaluating AI reasoning capabilities.

Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 16% of the KV cache. This KV-cache reduction also leads to a 90% memory saving and a 6.6X throughput over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.

Large language models (LLMs) have made impressive strides in mathematical reasoning, often fine-tuned using rejection sampling that retains only correct reasoning trajectories. While effective, this paradigm treats supervision as a binary filter that systematically excludes teacher-generated errors, leaving a gap in how reasoning failures are modeled during training. In this paper, we propose TrajFusion, a fine-tuning strategy that reframes rejection sampling as a structured supervision construction process. Specifically, TrajFusion forms fused trajectories that explicitly model trial-and-error reasoning by interleaving selected incorrect trajectories with reflection prompts and correct trajectories. The length of each fused sample is adaptively controlled based on the frequency and diversity of teacher errors, providing richer supervision for challenging problems while safely reducing to vanilla rejection sampling fine-tuning (RFT) when error signals are uninformative. TrajFusion requires no changes to the architecture or training objective. Extensive experiments across multiple math benchmarks demonstrate that TrajFusion consistently outperforms RFT, particularly on challenging and long-form reasoning problems.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a leading paradigm for enhancing visual reasoning in Multimodal Large Language Models (MLLMs). However, existing RLVR methods optimize primarily for the reasoning outcome, fundamentally overlooking the fine-grained cross-modal coordination required during the generation process. Through token-level analyses and controlled interventions, we reveal that during Chain-of-Thought (CoT) reasoning, MLLMs frequently fail to dynamically alternate between extracting visual evidence and synthesizing textual context-a coordination breakdown that is causally linked to reasoning failures. Motivated by these findings, we propose DyCo-RL, which integrates dynamic cross-modal coordination into RLVR optimization. Specifically, DyCo-RL uses the Fisher-Rao geodesic distance to measure within-modality attention shifts, assigning tokens to either visually-oriented or text-oriented functional roles. It then evaluates the alignment between a token's actual attention allocation and its assigned role, leveraging this score for alignment-guided advantage reweighting during policy optimization. Extensive experiments demonstrate that the algorithm-agnostic DyCo-RL, when applied to Qwen2.5-VL-3B/7B, consistently improves four representative RLVR algorithms across seven benchmarks spanning visual-centric and mathematical reasoning.

LLMs for clinical decision support often fail under small but clinically meaningful input shifts such as masking a symptom or negating a finding, despite high performance on static benchmarks. These reasoning failures frequently go undetected by standard NLP metrics, which are insensitive to latent representation shifts that drive diagnosis instability. We propose a geometry-aware evaluation framework, LAPD (Latent Agentic Perturbation Diagnostics), which systematically probes the latent robustness of clinical LLMs under structured adversarial edits. Within this framework, we introduce Latent Diagnosis Flip Rate (LDFR), a model-agnostic diagnostic signal that captures representational instability when embeddings cross decision boundaries in PCA-reduced latent space. Clinical notes are generated using a structured prompting pipeline grounded in diagnostic reasoning, then perturbed along four axes: masking, negation, synonym replacement, and numeric variation to simulate common ambiguities and omissions. We compute LDFR across both foundation and clinical LLMs, finding that latent fragility emerges even under minimal surface-level changes. Finally, we validate our findings on 90 real clinical notes from the DiReCT benchmark (MIMIC-IV), confirming the generalizability of LDFR beyond synthetic settings. Our results reveal a persistent gap between surface robustness and semantic stability, underscoring the importance of geometry-aware auditing in safety-critical clinical AI.

Large Language Models (LLMs) are increasingly deployed in real-world fact-checking systems, yet existing evaluations focus predominantly on claim verification and overlook the broader fact-checking workflow, including claim extraction and evidence retrieval. This narrow focus prevents current benchmarks from revealing systematic reasoning failures, factual blind spots, and robustness limitations of modern LLMs. To bridge this gap, we present FactArena, a fully automated arena-style evaluation framework that conducts comprehensive, stage-wise benchmarking of LLMs across the complete fact-checking pipeline. FactArena integrates three key components: (i) an LLM-driven fact-checking process that standardizes claim decomposition, evidence retrieval via tool-augmented interactions, and justification-based verdict prediction; (ii) an arena-styled judgment mechanism guided by consolidated reference guidelines to ensure unbiased and consistent pairwise comparisons across heterogeneous judge agents; and (iii) an arena-driven claim-evolution module that adaptively generates more challenging and semantically controlled claims to probe LLMs' factual robustness beyond fixed seed data. Across 16 state-of-the-art LLMs spanning seven model families, FactArena produces stable and interpretable rankings. Our analyses further reveal significant discrepancies between static claim-verification accuracy and end-to-end fact-checking competence, highlighting the necessity of holistic evaluation. The proposed framework offers a scalable and trustworthy paradigm for diagnosing LLMs' factual reasoning, guiding future model development, and advancing the reliable deployment of LLMs in safety-critical fact-checking applications.

Large Audio Language Models (LALMs) achieve strong performance on a variety of audio understanding tasks but continue to struggle with temporal reasoning, a fundamental capability central to human auditory perception. Understanding the causes of these failures remains challenging as existing benchmarks report performance gaps without probing underlying mechanisms. To address this, we introduce a benchmark with 1,657 questions across three foundational tasks designed specifically for mechanistic analysis. Examining model outputs across varying input settings (behavioral analysis) reveals that models often under-utilize audio when textual cues are available. We also provide the first causal mechanistic analysis of temporal reasoning failures in LALMs. Comparing attention upweighting against scaling, we find that redistributing attention across audio tokens is more effective than increasing audio attention. Targeting task-relevant tokens yields further gains. These findings suggest that modality imbalance alone cannot explain failures. Attention scaling at bottleneck layers improves accuracy from 55.9% to 59.1% without fine-tuning, demonstrating a promising direction for future work.

The standard post-training recipe for large multimodal models (LMMs) applies supervised fine-tuning (SFT) on curated demonstrations followed by reinforcement learning with verifiable rewards (RLVR). However, SFT introduces distributional drift that neither preserves the model's original capabilities nor faithfully matches the supervision distribution. This problem is further amplified in multimodal reasoning, where perception errors and reasoning failures follow distinct drift patterns that compound during subsequent RL. We introduce PRISM, a three-stage pipeline that mitigates this drift by inserting an explicit distribution-alignment stage between SFT and RLVR. Building on the principle of on-policy distillation (OPD), PRISM casts alignment as a black-box, response-level adversarial game between the policy and a Mixture-of-Experts (MoE) discriminator with dedicated perception and reasoning experts, providing disentangled corrective signals that steer the policy toward the supervision distribution without requiring access to teacher logits. While 1.26M public demonstrations suffice for broad SFT initialization, distribution alignment demands higher-fidelity supervision; we therefore curate 113K additional demonstrations from Gemini 3 Flash, featuring dense visual grounding and step-by-step reasoning on the hardest unsolved problems. Experiments on Qwen3-VL show that PRISM consistently improves downstream RLVR performance across multiple RL algorithms (GRPO, DAPO, GSPO) and diverse multimodal benchmarks, improving average accuracy by +4.4 and +6.0 points over the SFT-to-RLVR baseline on 4B and 8B, respectively. Our code, data, and model checkpoints are publicly available at https://github.com/XIAO4579/PRISM.

Conditioning a language model on additional context, such as feedback on a previous attempt, typically improves its response. Self-distillation trains the model to retain this improvement when the context is not present. The method works by matching the model's output distribution under two settings: a student that sees only the question, and a self-teacher that also sees the context. What the model learns therefore depends on what context the self-teacher receives, yet the design of this context remains largely unexplored. We study context design for self-distillation by training a solver on feedback from a frozen critic. We compare three conditions: (i) a binary reward (GRPO), (ii) the reference solution, and (iii) a step-by-step critique aligned to the solver's reasoning trace. Step-aligned critique yields the largest gains, outperforming GRPO by 16.11 points and reference-solution-conditioned self-distillation by 5.27 points (Avg@12). Per-token advantage analysis reveals why: step-aligned feedback targets only the tokens where reasoning fails, leaving correct behavior intact. Conditioning on the reference solution, by contrast, pressures the model to change its behavior at every token (even correct steps) because an alternative derivation inevitably differs in phrasing and approach. This suggests that structural alignment between feedback and the solver's reasoning is a key driver of self-distillation effectiveness.

Medical vision-language models (VLMs) and AI agents have made significant progress in learning to analyze and reason about clinical images. However, existing medical visual question answering (VQA) benchmarks collapse model capabilities into a single accuracy score, obscuring where and why models fail. We propose DeepTumorVQA, a hierarchical benchmark that follows the multi-stage evidence chain in tumor diagnosis and decomposes 3D CT reasoning into four stages: recognition, measurement, visual reasoning, and medical reasoning. Higher-level questions remain independently scorable, while their ground-truth evidence chains are defined over lower-level primitives. The benchmark contains 476K questions across 42 clinical subtypes on 9,262 3D CT volumes. In addition to a direct reasoning mode for VLMs, DeepTumorVQA provides tool-interaction environments for agent evaluation, where a model can call external tools, including segmentation models, measurement programs, and medical knowledge modules, before answering the question. Evaluating over 30 model configurations, we find that reliable quantitative measurement is the primary bottleneck, making later-stage visual and medical reasoning harder for VLMs, while tool augmentation substantially mitigates this issue. When tools are available, leveraging medical knowledge and tools to reason about medical images becomes a new challenge. We further show that ground-truth step-by-step tool-use traces from DeepTumorVQA can supervise agents and reduce tool-use and reasoning failures. This stage-wise progression from recognition to measurement to visual and medical reasoning provides a concrete roadmap for future medical VLM and AI agent studies. All data and code are released at https://github.com/Schuture/DeepTumorVQA.

Large language models (LLMs) often produce errors, including factual inaccuracies, biases, and reasoning failures, collectively referred to as "hallucinations". Recent studies have demonstrated that LLMs' internal states encode information regarding the truthfulness of their outputs, and that this information can be utilized to detect errors. In this work, we show that the internal representations of LLMs encode much more information about truthfulness than previously recognized. We first discover that the truthfulness information is concentrated in specific tokens, and leveraging this property significantly enhances error detection performance. Yet, we show that such error detectors fail to generalize across datasets, implying that -- contrary to prior claims -- truthfulness encoding is not universal but rather multifaceted. Next, we show that internal representations can also be used for predicting the types of errors the model is likely to make, facilitating the development of tailored mitigation strategies. Lastly, we reveal a discrepancy between LLMs' internal encoding and external behavior: they may encode the correct answer, yet consistently generate an incorrect one. Taken together, these insights deepen our understanding of LLM errors from the model's internal perspective, which can guide future research on enhancing error analysis and mitigation.

Autonomous data analysis agents are increasingly expected to conduct exploratory analysis over underexplored data environments. This burden is especially salient in complex financial analytics, where relevant evidence is rarely pre-specified. However, existing benchmarks typically evaluate such agents in prior-guided settings, providing selected data sources, explicit data schemas, or cleaned data, thereby understating the exploratory burden. We introduce DataClawBench, a benchmark for exploratory real-world financial data analysis under limited prior guidance. DataClawBench contains approximately 2.06 million real-world records across enterprise, industry, and policy domains, with native data noise preserved. It further includes 492 cross-domain tasks derived from think-tank consulting scenarios, each annotated with intermediate milestones that diagnose exploration and reasoning failures beyond outcome accuracy. A systematic evaluation of eight advanced LLMs under the OpenClaw agent reveals that exploratory data analysis breaks agent reliability: more exploration does not reliably translate into task-relevant progress or correct final answers.

Tables condense key transactional and administrative information into compact layouts, but practical extraction requires more than text recognition: systems must also recover structure (rows, columns, merged cells, headers) and interpret roles such as line items, subtotals, and totals under common capture artifacts. Many existing resources for table structure recognition and TableVQA are built from clean digital-born sources or rendered tables, and therefore only partially reflect noisy administrative conditions. We introduce DenTab, a dataset of 2{,}000 cropped table images from dental estimates with high-quality HTML annotations, enabling evaluation of table recognition (TR) and table visual question answering (TableVQA) on the same inputs. DenTab includes 2{,}208 questions across eleven categories spanning retrieval, aggregation, and logic/consistency checks. We benchmark 16 systems, including 14 vision--language models (VLMs) and two OCR baselines. Across models, strong structure recovery does not consistently translate into reliable performance on multi-step arithmetic and consistency questions, and these reasoning failures persist even when using ground-truth HTML table inputs. To improve arithmetic reliability without training, we propose the Table Router Pipeline, which routes arithmetic questions to deterministic execution. The pipeline combines (i) a VLM that produces a baseline answer, a structured table representation, and a constrained table program with (ii) a rule-based executor that performs exact computation over the parsed table. The source code and dataset will be made publicly available at https://github.com/hamdilaziz/DenTab.

Modern information retrieval (IR) is no longer consumed primarily by humans but increasingly by large language models (LLMs) via retrieval-augmented generation (RAG) and agentic search. Unlike human users, LLMs are constrained by limited attention budgets and are uniquely vulnerable to noise; misleading or irrelevant information is no longer just a nuisance, but a direct cause of hallucinations and reasoning failures. In this perspective paper, we argue that denoising-maximizing usable evidence density and verifiability within a context window-is becoming the primary bottleneck across the full information access pipeline. We conceptualize this paradigm shift through a four-stage framework of IR challenges: from inaccessible to undiscoverable, to misaligned, and finally to unverifiable. Furthermore, we provide a pipeline-organized taxonomy of signal-to-noise optimization techniques, spanning indexing, retrieval, context engineering, verification, and agentic workflow. We also present research works on information denoising in domains that rely heavily on retrieval such as lifelong assistant, coding agent, deep research, and multimodal understanding.

Large Language Model (LLM) agents trained with reinforcement learning (RL) show great promise for solving complex, multi-step tasks. However, their performance is often crippled by "Context Explosion", where the accumulation of long text outputs overwhelms the model's context window and leads to reasoning failures. To address this, we introduce CoDA, a Context-Decoupled hierarchical Agent, a simple but effective reinforcement learning framework that decouples high-level planning from low-level execution. It employs a single, shared LLM backbone that learns to operate in two distinct, contextually isolated roles: a high-level Planner that decomposes tasks within a concise strategic context, and a low-level Executor that handles tool interactions in an ephemeral, isolated workspace. We train this unified agent end-to-end using PECO (Planner-Executor Co-Optimization), a reinforcement learning methodology that applies a trajectory-level reward to jointly optimize both roles, fostering seamless collaboration through context-dependent policy updates. Extensive experiments demonstrate that CoDA achieves significant performance improvements over state-of-the-art baselines on complex multi-hop question-answering benchmarks, and it exhibits strong robustness in long-context scenarios, maintaining stable performance while all other baselines suffer severe degradation, thus further validating the effectiveness of our hierarchical design in mitigating context overload.

Robotic manipulation in open-world settings requires not only task execution but also the ability to detect and learn from failures. While recent advances in vision-language models (VLMs) and large language models (LLMs) have improved robots' spatial reasoning and problem-solving abilities, they still struggle with failure recognition, limiting their real-world applicability. We introduce AHA, an open-source VLM designed to detect and reason about failures in robotic manipulation using natural language. By framing failure detection as a free-form reasoning task, AHA identifies failures and provides detailed, adaptable explanations across different robots, tasks, and environments. We fine-tuned AHA using FailGen, a scalable framework that generates the first large-scale dataset of robotic failure trajectories, the AHA dataset. FailGen achieves this by procedurally perturbing successful demonstrations from simulation. Despite being trained solely on the AHA dataset, AHA generalizes effectively to real-world failure datasets, robotic systems, and unseen tasks. It surpasses the second-best model (GPT-4o in-context learning) by 10.3% and exceeds the average performance of six compared models including five state-of-the-art VLMs by 35.3% across multiple metrics and datasets. We integrate AHA into three manipulation frameworks that utilize LLMs/VLMs for reinforcement learning, task and motion planning, and zero-shot trajectory generation. AHA's failure feedback enhances these policies' performances by refining dense reward functions, optimizing task planning, and improving sub-task verification, boosting task success rates by an average of 21.4% across all three tasks compared to GPT-4 models.

Reasoning-enhanced large language models (RLLMs), whether explicitly trained for reasoning or prompted via chain-of-thought (CoT), have achieved state-of-the-art performance on many complex reasoning tasks. However, we uncover a surprising and previously overlooked phenomenon: explicit CoT reasoning can significantly degrade instruction-following accuracy. Evaluating 15 models on two benchmarks: IFEval (with simple, rule-verifiable constraints) and ComplexBench (with complex, compositional constraints), we consistently observe performance drops when CoT prompting is applied. Through large-scale case studies and an attention-based analysis, we identify common patterns where reasoning either helps (e.g., with formatting or lexical precision) or hurts (e.g., by neglecting simple constraints or introducing unnecessary content). We propose a metric, constraint attention, to quantify model focus during generation and show that CoT reasoning often diverts attention away from instruction-relevant tokens. To mitigate these effects, we introduce and evaluate four strategies: in-context learning, self-reflection, self-selective reasoning, and classifier-selective reasoning. Our results demonstrate that selective reasoning strategies, particularly classifier-selective reasoning, can substantially recover lost performance. To our knowledge, this is the first work to systematically expose reasoning-induced failures in instruction-following and offer practical mitigation strategies.

Robust robotic manipulation requires reliable failure detection and recovery. Although recent Vision-Language Models (VLMs) show promise in robot failure detection, their generalization is severely limited by the scarcity and narrow coverage of failure data. To address this bottleneck, we propose an automatic framework for generating diverse robotic planning and execution failures across both simulated and real-world environments. Our approach perturbs successful manipulation trajectories to synthesize failures that reflect realistic failure distributions, and leverages VLMs to produce structured step-by-step reasoning traces. This yields FailCoT, a large-scale failure reasoning dataset built upon the RLBench simulator and the BridgeDataV2 real-robot dataset. Using FailCoT, we train Guardian, a multi-view reasoning VLM for unified planning and execution verification. Guardian achieves state-of-the-art performance on three unseen real-world benchmarks: RoboFail, RoboVQA, and our newly introduced UR5-Fail. When integrated with a state-of-the-art LLM-based manipulation policy, it consistently boosts task success rates in both simulation and real-world deployment. These results demonstrate that scaling high-quality failure reasoning data is critical for improving generalization in robotic failure detection. Code, Data, and Models available at https://www.di.ens.fr/willow/research/guardian/.

The ability of large language models (LLMs) to follow user instructions is central to their reliability, safety, and usefulness. While prior studies assess instruction adherence in the model's main responses, we argue that it is also critical for large reasoning models (LRMs) to follow user instructions throughout their reasoning process. Reasoning instruction following makes LRMs more controllable and transparent, while reducing risks of undesirable shortcuts, hallucinations, or reward hacking within reasoning traces. To evaluate this dimension, we introduce ReasonIF, a systematic benchmark for assessing reasoning instruction following. ReasonIF includes six categories of instruction prompts, spanning multilingual reasoning, formatting and length control. Across many open-source LRMs including GPT-OSS, Qwen3, and DeepSeek-R1, we find substantial failures in reasoning instruction adherence: the highest instruction following score (IFS) remains below 0.25, meaning that fewer than 25% of reasoning traces comply with the given instructions. Notably, as task difficulty increases, reasoning instruction following degrades further. We also explore two strategies to enhance reasoning instruction fidelity. (1) multi-turn reasoning and (2) Reasoning Instruction Finetuning (RIF) using synthetic data. RIF improves the IFS of GPT-OSS-20B from 0.11 to 0.27, indicating measurable progress but leaving ample room for improvement.

When faced with complex spatial problems, humans naturally sketch layouts to organize their thinking, and the act of drawing further sharpens their understanding. In this work, we ask whether a similar principle holds for Large Language Models (LLMs): can learning to construct explicit visual layouts from spatial descriptions instill genuine spatial understanding? We introduce Text2Space, a dataset that pairs natural language descriptions with ground-truth ASCII grid layouts and spatial QA pairs, enabling us to separate failures in constructing spatial representations from failures in reasoning over them. We adopt ASCII because it is human-readable, operates entirely within the token space of language models, and encodes spatial relations in a structurally verifiable form. Our evaluation reveals a pronounced "Read-Write Asymmetry": LLMs interpret ASCII representations effectively but struggle to produce them from text, and these construction errors propagate to incorrect answers downstream. To address this limitation, we train models on layout construction (TextrightarrowASCII) and find that it significantly improves spatial reasoning from text alone, even without producing any ASCII at inference time. Combining construction with comprehension training further amplifies these gains. Crucially, these improvements transfer to three external spatial reasoning benchmarks, demonstrating that, much as sketching sharpens human spatial thinking, learning to construct explicit layouts instills spatial understanding that generalizes beyond the training format.

Frontier AI systems are increasingly capable and deployed in high-stakes multi-agent environments. However, existing AI safety benchmarks largely evaluate single agents, leaving multi-agent risks such as coordination failure and conflict poorly understood. We introduce GT-HarmBench, a benchmark of 1,535 high-stakes scenarios spanning game-theoretic structures such as the Prisoner's Dilemma, Stag Hunt and Chicken. Scenarios are drawn from realistic AI risk contexts in the MIT AI Risk Repository. Across 15 frontier models, agents fail to choose socially beneficial actions in 38% of high-stakes cases, such as military escalation, election manipulation, and medical malpractice. We measure sensitivity to game-theoretic prompt framing and ordering, and analyze reasoning patterns driving failures. We further show that game-theoretic interventions improve socially beneficial outcomes by up to 18%. Our results highlight substantial reliability gaps and provide a broad standardized testbed for studying alignment in multi-agent environments. The benchmark and code are available at https://github.com/causalNLP/gt-harmbench.

This paper presents ConvBench, a novel multi-turn conversation evaluation benchmark tailored for Large Vision-Language Models (LVLMs). Unlike existing benchmarks that assess individual capabilities in single-turn dialogues, ConvBench adopts a three-level multimodal capability hierarchy, mimicking human cognitive processes by stacking up perception, reasoning, and creativity. Each level focuses on a distinct capability, mirroring the cognitive progression from basic perception to logical reasoning and ultimately to advanced creativity. ConvBench comprises 577 meticulously curated multi-turn conversations encompassing 215 tasks reflective of real-world demands. Automatic evaluations quantify response performance at each turn and overall conversation level. Leveraging the capability hierarchy, ConvBench enables precise attribution of conversation mistakes to specific levels. Experimental results reveal a performance gap between multi-modal models, including GPT4-V, and human performance in multi-turn conversations. Additionally, weak fine-grained perception in multi-modal models contributes to reasoning and creation failures. ConvBench serves as a catalyst for further research aimed at enhancing visual dialogues.

Current paradigms for code verification rely heavily on external mechanisms-such as execution-based unit tests or auxiliary LLM judges-which are often labor-intensive or limited by the judging model's own capabilities. This raises a fundamental, yet unexplored question: Can an LLM's functional correctness be assessed purely from its internal computational structure? Our primary objective is to investigate whether the model's neural dynamics encode internally decodable signals that are predictive of logical validity during code generation. Inspired by mechanistic interpretability, we propose to treat code verification as a mechanistic diagnostic task, mapping the model's explicit algorithmic trajectory into line-level attribution graphs. By decomposing complex residual flows, we aim to identify the structural signatures that distinguish sound reasoning from logical failure within the model's internal circuits. Analysis across Python, C++, and Java confirms that intrinsic correctness signals are robust across diverse syntaxes. Topological features from these internal graphs predict correctness more reliably than surface heuristics and enable targeted causal interventions to fix erroneous logic. These findings establish internal introspection as a decodable property for verifying generated code. Our code is at https:// github.com/bruno686/CodeCircuit.

Modern AI systems struggle most in environments where reliability is critical - scenes with smoke, poor visibility, and structural deformation. Each year, tens of thousands of firefighters are injured on duty, often due to breakdowns in situational perception. We introduce Fire360, a benchmark for evaluating perception and reasoning in safety-critical firefighting scenarios. The dataset includes 228 360-degree videos from professional training sessions under diverse conditions (e.g., low light, thermal distortion), annotated with action segments, object locations, and degradation metadata. Fire360 supports five tasks: Visual Question Answering, Temporal Action Captioning, Object Localization, Safety-Critical Reasoning, and Transformed Object Retrieval (TOR). TOR tests whether models can match pristine exemplars to fire-damaged counterparts in unpaired scenes, evaluating transformation-invariant recognition. While human experts achieve 83.5% on TOR, models like GPT-4o lag significantly, exposing failures in reasoning under degradation. By releasing Fire360 and its evaluation suite, we aim to advance models that not only see, but also remember, reason, and act under uncertainty. The dataset is available at: https://uofi.box.com/v/fire360dataset.

We present Omni-I2C, a comprehensive benchmark designed to evaluate the capability of Large Multimodal Models (LMMs) in converting complex, structured digital graphics into executable code. We argue that this task represents a non-trivial challenge for the current generation of LMMs: it demands an unprecedented synergy between high-fidelity visual perception -- to parse intricate spatial hierarchies and symbolic details -- and precise generative expression -- to synthesize syntactically sound and logically consistent code. Unlike traditional descriptive tasks, Omni-I2C requires a holistic understanding where any minor perceptual hallucination or coding error leads to a complete failure in visual reconstruction. Omni-I2C features 1080 meticulously curated samples, defined by its breadth across subjects, image modalities, and programming languages. By incorporating authentic user-sourced cases, the benchmark spans a vast spectrum of digital content -- from scientific visualizations to complex symbolic notations -- each paired with executable reference code. To complement this diversity, our evaluation framework provides necessary depth; by decoupling performance into perceptual fidelity and symbolic precision, it transcends surface-level accuracy to expose the granular structural failures and reasoning bottlenecks of current LMMs. Our evaluation reveals a substantial performance gap among leading LMMs; even state-of-the-art models struggle to preserve structural integrity in complex scenarios, underscoring that multimodal code generation remains a formidable challenge. Data and code are available at https://github.com/MiliLab/Omni-I2C.

Large reasoning models (LRMs) increasingly expose chain-of-thought-like reasoning for transparency, verification, and deliberate problem solving. This creates a safety blind spot: harmful or policy-violating content may appear in reasoning traces even when final answers appear safe. We test whether final-answer safety is a sufficient proxy for the full reasoning-answer trajectory by scoring both stages under a unified twenty-principle safety rubric. Using prompts from seven public harmfulness and jailbreak sources, plus four out-of-distribution (OOD) sources, we evaluate 15 open-weight and API-based LRMs across 41K prompts per model. Reasoning traces consistently reveal additional safety risks beyond final answers, especially in high-severity stage-wise failures: leak cases, where unsafe reasoning precedes a safe-looking answer, and escape cases, where benign-looking reasoning precedes an unsafe final response. Principle-level analysis shows that risk concentrates in misinformation, legal compliance, discrimination, physical harm, and psychological harm. We further propose adaptive multi-principle steering, a white-box test-time mitigation that learns one unsafe-to-safe activation direction per safety principle and activates only directions whose current hidden state is closer to the unsafe than safe centroid. On three steerable open reasoning models, adaptive steering reduces unsafe counts in both reasoning traces and final answers on held-out and OOD benchmarks. DeepSeek-R1-Qwen-7B achieves a 40.8% average unsafe-count reduction while retaining 97.7% macro-averaged accuracy on BBH, GSM8K, and MMLU. These results suggest that LRM safety should be evaluated and mitigated over the full exposed reasoning-answer trajectory, not only at the final-answer stage.

Recent advances in robotic manipulation have integrated low-level robotic control into Vision-Language Models (VLMs), extending them into Vision-Language-Action (VLA) models. Although state-of-the-art VLAs achieve strong performance in downstream robotic applications, supported by large-scale crowd-sourced robot training data, they still inevitably encounter failures during execution. Enabling robots to reason and recover from unpredictable and abrupt failures remains a critical challenge. Existing robotic manipulation datasets, collected in either simulation or the real world, primarily provide only ground-truth trajectories, leaving robots unable to recover once failures occur. Moreover, the few datasets that address failure detection typically offer only textual explanations, which are difficult to utilize directly in VLA models. To address this gap, we introduce FailSafe, a novel failure generation and recovery system that automatically produces diverse failure cases paired with executable recovery actions. FailSafe can be seamlessly applied to any manipulation task in any simulator, enabling scalable creation of failure action data. To demonstrate its effectiveness, we fine-tune LLaVa-OneVision-7B (LLaVa-OV-7B) to build FailSafe-VLM. Experimental results show that FailSafe-VLM successfully helps robotic arms detect and recover from potential failures, improving the performance of three state-of-the-art VLA models (pi0-FAST, OpenVLA, OpenVLA-OFT) by up to 22.6% on average across several tasks in Maniskill. Furthermore, FailSafe-VLM could generalize across different spatial configurations, camera viewpoints, object and robotic embodiments. We plan to release the FailSafe code to the community.

Mathematical reasoning models are widely deployed in education, automated tutoring, and decision support systems despite exhibiting fundamental computational instabilities. We demonstrate that state-of-the-art models (Qwen2.5-Math-7B) achieve 61% accuracy through a mixture of reliable and unreliable reasoning pathways: 18.4% of correct predictions employ stable, faithful reasoning while 81.6% emerge through computationally inconsistent pathways. Additionally, 8.8% of all predictions are silent failures -- confident yet incorrect outputs. Through comprehensive analysis using novel faithfulness metrics, we reveal: (1) reasoning quality shows weak negative correlation with correctness (r=-0.21, p=0.002), reflecting a binary classification threshold artifact rather than a monotonic inverse relationship; (2) scaling from 1.5B to 7B parameters (4.7x increase) provides zero accuracy benefit on our evaluated subset (6% of GSM8K), requiring validation on the complete benchmark; and (3) latent reasoning employs diverse computational strategies, with ~20% sharing CoT-like patterns. These findings highlight that benchmark accuracy can mask computational unreliability, demanding evaluation reforms measuring stability beyond single-sample metrics.

Chain-of-thought (CoT) monitoring is a promising tool for detecting misbehaviors and understanding the motivations of modern reasoning models. However, if models can control what they verbalize in their CoT, it could undermine CoT monitorability. To measure this undesirable capability -- CoT controllability -- we introduce the CoT-Control evaluation suite, which includes tasks that require models to solve problems while adhering to CoT instructions, e.g., reasoning about a genetics question without using the word 'chromosome'. We show that reasoning models possess significantly lower CoT controllability than output controllability; for instance, Claude Sonnet 4.5 can control its CoT only 2.7% of the time but 61.9% when controlling its final output. We also find that CoT controllability is higher for larger models and decreases with more RL training, test-time compute, and increased problem difficulty. CoT controllability failures extend even to situations in which models are given incentives (as opposed to direct requests) to evade CoT monitors, although models exhibit slightly higher controllability when they are told they are being monitored. Similarly, eliciting controllability by adversarially optimizing prompts does not meaningfully increase controllability. Our results leave us cautiously optimistic that CoT controllability is currently unlikely to be a failure mode of CoT monitorability. However, the mechanism behind low controllability is not well understood. Given its importance for maintaining CoT monitorability, we recommend that frontier labs track CoT controllability in future models.

LLM failures in causal reasoning, including sycophancy, rung collapse, and miscalibrated refusal, are well-documented, yet progress on remediation is slow because no benchmark enables systematic diagnosis. We introduce CausalT5K, a diagnostic benchmark of over 5,000 cases across 10 domains that tests three critical capabilities: (1) detecting rung collapse, where models answer interventional queries with associational evidence; (2) resisting sycophantic drift under adversarial pressure; and (3) generating Wise Refusals that specify missing information when evidence is underdetermined. Unlike synthetic benchmarks, CausalT5K embeds causal traps in realistic narratives and decomposes performance into Utility (sensitivity) and Safety (specificity), revealing failure modes invisible to aggregate accuracy. Developed through a rigorous human-machine collaborative pipeline involving 40 domain experts, iterative cross-validation cycles, and composite verification via rule-based, LLM, and human scoring, CausalT5K implements Pearl's Ladder of Causation as research infrastructure. Preliminary experiments reveal a Four-Quadrant Control Landscape where static audit policies universally fail, a finding that demonstrates CausalT5K's value for advancing trustworthy reasoning systems. Repository: https://github.com/genglongling/CausalT5kBench

A prevailing narrative in LLM post-training holds that supervised finetuning (SFT) memorizes while reinforcement learning (RL) generalizes. We revisit this claim for reasoning SFT with long chain-of-thought (CoT) supervision and find that cross-domain generalization is not absent but conditional, jointly shaped by optimization dynamics, training data, and base-model capability. Some reported failures are under-optimization artifacts: cross-domain performance first degrades before recovering and improving with extended training (a dip-and-recovery pattern), so shorttraining checkpoints can underestimate generalization. Data quality and structure both matter: low-quality solutions broadly hurt generalization,while verified long-CoT traces yield consistent cross-domain gains. Model capability is essential: stronger models internalize transferable procedural patterns (e.g., backtracking) even from a toy arithmetic game, while weaker ones imitate surface verbosity. This generalization is asymmetric, however: reasoning improves while safety degrades, reframing the question from whether reasoning SFT generalizes to under what conditions and at what cost.

We find that the response length of reasoning LLMs, whether trained by reinforcement learning or supervised learning, drastically increases for ill-posed questions with missing premises (MiP), ending up with redundant and ineffective thinking. This newly introduced scenario exacerbates the general overthinking issue to a large extent, which we name as the MiP-Overthinking. Such failures are against the ``test-time scaling law'' but have been widely observed on multiple datasets we curated with MiP, indicating the harm of cheap overthinking and a lack of critical thinking. Surprisingly, LLMs not specifically trained for reasoning exhibit much better performance on the MiP scenario, producing much shorter responses that quickly identify ill-posed queries. This implies a critical flaw of the current training recipe for reasoning LLMs, which does not encourage efficient thinking adequately, leading to the abuse of thinking patterns. To further investigate the reasons behind such failures, we conduct fine-grained analyses of the reasoning length, overthinking patterns, and location of critical thinking on different types of LLMs. Moreover, our extended ablation study reveals that the overthinking is contagious through the distillation of reasoning models' responses. These results improve the understanding of overthinking and shed novel insights into mitigating the problem.

Failures in multi-turn reasoning models are largely invisible to terminal-score evaluation. A model can lock onto an unsafe stance early in a long dialogue, yet its final-turn refusal rate may appear indistinguishable from a robustly aligned baseline. To expose these hidden temporal dynamics, we propose a trace-level diagnostic - the CoT-Output 2x2 safety matrix. This framework labels every turn along two independent axes (internal reasoning and visible output), yielding four operationally defined failure cells: robust alignment, alignment faking, overt jailbreak, and a distinct failure mode we term context-injection failure (where the CoT maintains safe reasoning, but the visible output produces harm, highlighting a multi-turn manifestation of reasoning unfaithfulness). We evaluate three distilled reasoning targets against a fixed attacker across five oversight conditions, collecting 6750 turn-level observations on the Information-Hazard scenario. Our analysis reveals two reproducible vulnerabilities: an oversight paradox where explicit monitoring cues paradoxically increase alignment-faking rates rather than suppress them, and a context-injection failure where models lock onto unsafe external outputs despite safe internal states. We release the full dataset of multi-turn dialogues and CoT traces to support follow-up trace-diagnostic research.

Large Reasoning Models (LRMs) exhibit strong performance, yet often produce rationales that sound plausible but fail to reflect their true decision process, undermining reliability and trust. We introduce a formal framework for reasoning faithfulness, defined by two testable conditions: stance consistency (a coherent stance linking reasoning to answer) and causal influence (the stated reasoning causally drives the answer under output-level interventions), explicitly decoupled from accuracy. To operationalize this, we present RFEval, a benchmark of 7,186 instances across seven tasks that probes faithfulness via controlled, output-level counterfactual interventions. Evaluating twelve open-source LRMs, we find unfaithfulness in 49.7% of outputs, predominantly from stance inconsistency. Failures are concentrated in brittle, convergent domains such as math and code, and correlate more with post-training regimes than with scale: within-family ablations indicate that adding current RL-style objectives on top of supervised fine-tuning can reduce reasoning faithfulness, even when accuracy is maintained. Crucially, accuracy is neither a sufficient nor a reliable proxy for faithfulness: once controlling for model and task, the accuracy-faithfulness link is weak and statistically insignificant. Our work establishes a rigorous methodology for auditing LRM reliability and shows that trustworthy AI requires optimizing not only for correct outcomes but also for the structural integrity of the reasoning process. Our code and dataset can be found at project page: https://aidaslab.github.io/RFEval/}{https://aidaslab.github.io/RFEval/

The reliability of Large Language Models (LLMs) in high-stakes domains such as healthcare, law, and scientific discovery is often compromised by hallucinations. These failures typically stem from two sources: data-driven hallucinations and reasoning-driven hallucinations. However, existing detection methods usually address only one source and rely on task-specific heuristics, limiting their generalization to complex scenarios. To overcome these limitations, we introduce the Hallucination Risk Bound, a unified theoretical framework that formally decomposes hallucination risk into data-driven and reasoning-driven components, linked respectively to training-time mismatches and inference-time instabilities. This provides a principled foundation for analyzing how hallucinations emerge and evolve. Building on this foundation, we introduce HalluGuard, an NTK-based score that leverages the induced geometry and captured representations of the NTK to jointly identify data-driven and reasoning-driven hallucinations. We evaluate HalluGuard on 10 diverse benchmarks, 11 competitive baselines, and 9 popular LLM backbones, consistently achieving state-of-the-art performance in detecting diverse forms of LLM hallucinations.

Spatial reasoning is a core aspect of human intelligence that allows perception, inference and planning in 3D environments. However, current vision-language models (VLMs) struggle to maintain geometric coherence and cross-view consistency for spatial reasoning in multi-view settings. We attribute this gap to the lack of fine-grained benchmarks that isolate multi-view reasoning from single-view perception and temporal factors. To address this, we present ReMindView-Bench, a cognitively grounded benchmark for evaluating how VLMs construct, align and maintain spatial mental models across complementary viewpoints. ReMindView-Bench systematically varies viewpoint spatial pattern and query type to probe key factors of spatial cognition. Evaluations of 15 current VLMs reveals consistent failures in cross-view alignment and perspective-taking in multi-view spatial reasoning, motivating deeper analysis on the reasoning process. Explicit phase-wise analysis using LLM-as-a-judge and self-consistency prompting shows that VLMs perform well on in-frame perception but degrade sharply when integrating information across views. Implicit analysis, including linear probing and entropy dynamics, further show progressive loss of task-relevant information and uncertainty separation between correct and incorrect trajectories. These results provide a cognitively grounded diagnosis of VLM spatial reasoning and reveal how multi-view spatial mental models are formed, degraded and destabilized across reasoning phases. The ReMindView-Bench benchmark is available at https://huggingface.co/datasets/Xue0823/ReMindView-Bench, and the source codes of benchmark construction and VLM reasoning analysis are available at https://github.com/pittisl/ReMindView-Bench.

Large language models (LLMs) generate fluent and complex outputs but often fail to recognize their own mistakes and hallucinations. Existing approaches typically rely on external judges, multi-sample consistency, or text-based self-critique, which incur additional compute or correlate weakly with true correctness. We ask: can LLMs predict their own failures by inspecting internal states during inference? We introduce Gnosis, a lightweight self-awareness mechanism that enables frozen LLMs to perform intrinsic self-verification by decoding signals from hidden states and attention patterns. Gnosis passively observes internal traces, compresses them into fixed-budget descriptors, and predicts correctness with negligible inference cost, adding only ~5M parameters and operating independently of sequence length. Across math reasoning, open-domain question answering, and academic knowledge benchmarks, and over frozen backbones ranging from 1.7B to 20B parameters, Gnosis consistently outperforms strong internal baselines and large external judges in both accuracy and calibration. Moreover, it generalizes zero-shot to partial generations, enabling early detection of failing trajectories and compute-aware control. These results show that reliable correctness cues are intrinsic to generation process and can be extracted efficiently without external supervision.

Large Language Models have demonstrated strong performance on many established reasoning benchmarks. However, these benchmarks primarily evaluate structured skills like quantitative problem-solving, leaving a gap in assessing flexible, multifaceted reasoning abilities that are central to human intelligence. These abilities require integrating logical deduction with spatial awareness and constraint satisfaction, which current evaluations do not measure well. To address this, we introduce RiddleBench, a benchmark of 1,737 challenging puzzles in English designed to probe these core reasoning capabilities. Evaluation of state-of-the-art models on RiddleBench shows fundamental weaknesses. Even top proprietary models like Gemini 2.5 Pro, o3, and Claude 4 Sonnet achieve accuracy just above 60% (60.30%, 63.37%, and 63.16%). Analysis further reveals deep failures, including hallucination cascades (accepting flawed reasoning from other models) and poor self-correction due to a strong self-confirmation bias. Their reasoning is also fragile, with performance degrading significantly when constraints are reordered or irrelevant information is introduced. RiddleBench functions as a diagnostic tool for these issues and as a resource for guiding the development of more robust and reliable language models.

Failure attribution in LLM multi-agent systems-identifying the agent and step responsible for task failures-provides crucial clues for systems debugging but remains underexplored and labor-intensive. In this paper, we propose and formulate a new research area: automated failure attribution for LLM multi-agent systems. To support this initiative, we introduce the Who&When dataset, comprising extensive failure logs from 127 LLM multi-agent systems with fine-grained annotations linking failures to specific agents and decisive error steps. Using the Who&When, we develop and evaluate three automated failure attribution methods, summarizing their corresponding pros and cons. The best method achieves 53.5% accuracy in identifying failure-responsible agents but only 14.2% in pinpointing failure steps, with some methods performing below random. Even SOTA reasoning models, such as OpenAI o1 and DeepSeek R1, fail to achieve practical usability. These results highlight the task's complexity and the need for further research in this area. Code and dataset are available at https://github.com/mingyin1/Agents_Failure_Attribution

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an effective paradigm for improving the reasoning capabilities of large language models. However, RLVR training is often hindered by sparse binary rewards and weak credit assignment, resulting in ambiguous optimization signals and underutilization of the useful information embedded in failed trajectories. To address this challenge, we propose Correction-Oriented Policy Optimization (CIPO), a simple and effective extension to RLVR that converts on-policy failed trajectories into correction-oriented supervision, without relying on any external signals. By jointly optimizing correction samples derived from the model's own failed attempts together with the standard RLVR objective, CIPO improves learning effectiveness while explicitly enhancing the model's ability to correct its own errors. Extensive experiments across 11 benchmarks spanning mathematical reasoning and code generation demonstrate that CIPO consistently and significantly outperforms strong baselines in both reasoning and correction performance. Moreover, CIPO yields stronger pass@K gains, indicating that it improves the model's intrinsic reasoning capacity rather than merely redistributing probability mass over existing correct answers.

Running LLMs with extended reasoning on every problem is expensive, but determining which inputs actually require additional compute remains challenging. We investigate whether their own likelihood of success is recoverable from their internal representations before generation, and if this signal can guide more efficient inference. We train linear probes on pre-generation activations to predict policy-specific success on math and coding tasks, substantially outperforming surface features such as question length and TF-IDF. Using E2H-AMC, which provides both human and model performance on identical problems, we show that models encode a model-specific notion of difficulty that is distinct from human difficulty, and that this distinction increases with extended reasoning. Leveraging these probes, we demonstrate that routing queries across a pool of models can exceed the best-performing model whilst reducing inference cost by up to 70\% on MATH, showing that internal representations enable practical efficiency gains even when they diverge from human intuitions about difficulty. Our code is available at: https://github.com/KabakaWilliam/llms_know_difficulty

With the growing adoption of Large Language Model (LLM) agents in persistent, real-world roles, they naturally encounter continuous streams of tasks and inevitable failures. A key limitation, however, is their inability to systematically learn from these mistakes, forcing them to repeat identical errors in similar contexts. Unlike prior training-free methods that primarily store raw instance-level experience or focus on retrieving successful trajectories, we propose Mistake Notebook Learning (MNL), a novel memory framework that enables agents to self-curate generalizable guidance from batch-clustered failures. This mechanism allows agents to distill shared error patterns into structured "mistake notes," updating an external memory only when batch performance improves to ensure stability. To further amplify adaptability, we integrate MNL with test-time scaling, leveraging aggregated failure patterns to actively steer the search process away from known pitfalls. Experiments on mathematical reasoning, Text-to-SQL, and interactive agent benchmarks show that MNL achieves competitive performance compared to existing memory mechanisms and in-context methods in both effectiveness and efficiency. These findings position structured mistake abstraction as a critical lever for robust agent evolution, enabling continuous improvement without the cost of parameter updates. The code is available at https://github.com/Bairong-Xdynamics/MistakeNotebookLearning/tree/main.

LLM-based agents increasingly solve complex tasks through long trajectories involving reasoning steps, tool calls, and inter-agent communication. However, when these agents fail, it is often unclear which agent caused the failure and which step introduced the decisive error. This attribution problem is challenging because mistakes can propagate across the trajectory: later actions may appear incorrect, but only because they depend on an earlier corrupted state. Therefore, failure attribution cannot be treated as independent step-level classification. We propose FALAT, a diagnostic framework for failure attribution in LLM agent trajectories. FALAT frames attribution as a dependency-guided search problem. It first constructs an expectation of how the task should be solved and uses this expectation to identify suspicious regions in the trajectory. It then traces dependencies among decisions, tool outputs, and agent messages to distinguish error-introducing steps from steps that merely inherit or propagate prior mistakes. Finally, FALAT evaluates whether correcting a candidate step would be sufficient to recover the expected outcome, allowing it to identify both the responsible agent and the decisive failure step. We evaluate FALAT on the Who&When benchmark, which includes both algorithm-generated and hand-crafted multi-agent failure trajectories. The results show that FALAT consistently improves responsible-agent and decisive-step attribution. Its best configurations achieve 46.0% step-level accuracy on algorithm-generated trajectories and 29.1% on the more challenging hand-crafted trajectories, outperforming specialized attribution baselines and direct prompting with standalone LLMs. These findings suggest that dependency-aware reasoning is essential for reliable failure diagnosis in LLM agent systems.

We introduce FailureSensorIQ, a novel Multi-Choice Question-Answering (MCQA) benchmarking system designed to assess the ability of Large Language Models (LLMs) to reason and understand complex, domain-specific scenarios in Industry 4.0. Unlike traditional QA benchmarks, our system focuses on multiple aspects of reasoning through failure modes, sensor data, and the relationships between them across various industrial assets. Through this work, we envision a paradigm shift where modeling decisions are not only data-driven using statistical tools like correlation analysis and significance tests, but also domain-driven by specialized LLMs which can reason about the key contributors and useful patterns that can be captured with feature engineering. We evaluate the Industrial knowledge of over a dozen LLMs-including GPT-4, Llama, and Mistral-on FailureSensorIQ from different lens using Perturbation-Uncertainty-Complexity analysis, Expert Evaluation study, Asset-Specific Knowledge Gap analysis, ReAct agent using external knowledge-bases. Even though closed-source models with strong reasoning capabilities approach expert-level performance, the comprehensive benchmark reveals a significant drop in performance that is fragile to perturbations, distractions, and inherent knowledge gaps in the models. We also provide a real-world case study of how LLMs can drive the modeling decisions on 3 different failure prediction datasets related to various assets. We release: (a) expert-curated MCQA for various industrial assets, (b) FailureSensorIQ benchmark and Hugging Face leaderboard based on MCQA built from non-textual data found in ISO documents, and (c) LLMFeatureSelector, an LLM-based feature selection scikit-learn pipeline. The software is available at https://github.com/IBM/FailureSensorIQ.

Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved the reasoning abilities of large language models (LLMs), yet training often stalls as problems become saturated. We identify the core challenge as the poor accessibility of informative failures: learning signals exist but are rarely encountered during standard rollouts. To address this, we propose failure-prefix conditioning, a simple and effective method for learning from saturated problems. Rather than starting from the original question, our approach reallocates exploration by conditioning training on prefixes derived from rare incorrect reasoning trajectories, thereby exposing the model to failure-prone states. We observe that failure-prefix conditioning yields performance gains matching those of training on medium-difficulty problems, while preserving token efficiency. Furthermore, we analyze the model's robustness, finding that our method reduces performance degradation under misleading failure prefixes, albeit with a mild trade-off in adherence to correct early reasoning. Finally, we demonstrate that an iterative approach, which refreshes failure prefixes during training, unlocks additional gains after performance plateaus. Overall, our results suggest that failure-prefix conditioning offers an effective pathway to extend RLVR training on saturated problems.

Instruction-driven image editing with unified multimodal generative models has advanced rapidly, yet their underlying visual reasoning remains limited, leading to suboptimal performance on reasoning-centric edits. Reinforcement learning (RL) has been investigated for improving the quality of image editing, but it faces three key challenges: (1) limited reasoning exploration confined to denoising stochasticity, (2) biased reward fusion, and (3) unstable VLM-based instruction rewards. In this work, we propose ThinkRL-Edit, a reasoning-centric RL framework that decouples visual reasoning from image synthesis and expands reasoning exploration beyond denoising. To the end, we introduce Chain-of-Thought (CoT)-based reasoning sampling with planning and reflection stages prior to generation in online sampling, compelling the model to explore multiple semantic hypotheses and validate their plausibility before committing to a visual outcome. To avoid the failures of weighted aggregation, we propose an unbiased chain preference grouping strategy across multiple reward dimensions. Moreover, we replace interval-based VLM scores with a binary checklist, yielding more precise, lower-variance, and interpretable rewards for complex reasoning. Experiments show our method significantly outperforms prior work on reasoning-centric image editing, producing instruction-faithful, visually coherent, and semantically grounded edits.

Reasoning language models (RLMs) achieve strong performance on complex reasoning tasks, yet they still suffer from a multilingual reasoning gap, performing better in high-resource languages than in low-resource ones. While recent efforts have reduced this gap, its underlying causes remain largely unexplored. In this paper, we address this by showing that the multilingual reasoning gap largely stems from failures in language understanding-the model's inability to represent the multilingual input meaning into the dominant language (i.e., English) within its reasoning trace. This motivates us to examine whether understanding failures can be detected, as this ability could help mitigate the multilingual reasoning gap. To this end, we evaluate a range of detection methods and find that understanding failures can indeed be identified, with supervised approaches performing best. Building on this, we propose Selective Translation, a simple yet effective strategy that translates the multilingual input into English only when an understanding failure is detected. Experimental results show that Selective Translation bridges the multilingual reasoning gap, achieving near full-translation performance while using translation for only about 20% of inputs. Together, our work demonstrates that understanding failures are the primary cause of the multilingual reasoning gap and can be detected and selectively mitigated, providing key insight into its origin and a promising path toward more equitable multilingual reasoning. Our code and data are publicly available at https://github.com/deokhk/RLM_analysis.

What is reasoning? This question has driven centuries of philosophical inquiry, from Aristotle's syllogisms to modern computational complexity theory. In the age of large language models achieving superhuman performance on benchmarks like GSM8K (95\% accuracy) and HumanEval (90\% pass@1), we must ask: have these systems learned to reason, or have they learned to pattern-match over reasoning traces? This paper argues for a specific answer: reasoning is iterative operator application in state spaces, converging to fixed points. This definition is not merely philosophical -- it has concrete architectural implications that explain both the failures of current systems and the path to genuine reasoning capabilities. Our investigation begins with a puzzle (OpenXOR), progresses through theory (OpenOperator), and culminates in a working solution (OpenLM) that achieves 76\% accuracy where state-of-the-art LLMs achieve 0\%. This is not about criticizing existing systems, but about understanding what reasoning requires and building architectures that provide it.

Agentic coding assistants are increasingly given extra capabilities, such as browser based testing tools and design oriented system prompts, on the assumption that more capability yields better software. This study tested that assumption directly. Ninety independent agent runs built the same application, a real time retrospective board, from one detailed specification, each scored on a fixed 14 criterion functional rubric (42 point maximum) and a visual quality review. The runs spanned several model generations, two agent harnesses, two reasoning effort levels, a testing tool, and two design oriented prompts. Capability tier dominated: frontier models clustered near the ceiling while a low cost local model fell to 24 to 37 points. A criterion level analysis revealed what run totals conceal. Container deployment was the dominant defect, failing first try in 44 percent of runs, with its failure rate shifting sharply across model generations while mean totals moved less than a point. The testing tool raised cost by 42 to 68 percent without improving functional score or reliability, even on interface visible criteria. Raising reasoning effort from High to xHigh lifted first try perfect runs from 28 percent to 89 percent and cut corrective prompts about five fold, for 9 to 29 percent more cost. A design oriented prompt raised visual quality, 4.5 versus 3.0 on a 5 point scale, without lifting function, and a one paragraph paraphrase of its directive reproduced the entire lift. The practical lesson is to match the fix to the failure: most first run failures came from weak reasoning, which a stronger model or more effort prevents, not from visible flaws a checking tool would catch.

Automatic evaluation of long-form literary writing remains challenging, as generic LLM-as-Judge approaches may not fully capture creativity-related dimensions such as originality and flexibility. Although the Torrance Test of Creative Writing (TTCW) provides a structured creativity framework, and prior work has demonstrated reference-based TTCW evaluation at the pairwise level, no large-scale dataset exists for long-form TTCW-based literary review generation. We address this gap by constructing a dataset of 263,911 long-form stories, each annotated with scalar scores and meta-synthesised review comments across 14 TTCW-based dimensions. Using this dataset, we fine-tune Qwen3 models at two scales, 4B and 8B, under two conditions: with and without reasoning content. Results show that non-reasoning fine-tuning achieves stronger and more stable performance, with the best setting reaching an evaluation score of 0.6820. Further analysis shows that reasoning-supervised models are more prone to parse failures, often continuing with irrelevant or repetitive reasoning-style text rather than completing the required 14-metric review report. These results suggest that, for fixed-format rubric-based review generation, reasoning supervision is not straightforwardly beneficial, and precise metric-aligned scoring remains challenging even after task-specific fine-tuning.

Solving topological grid puzzles requires reasoning over global spatial invariants such as connectivity, loop closure, and region symmetry and remains challenging for even the most powerful large language models (LLMs). To study these abilities under controlled settings, we introduce TopoBench, a benchmark of six puzzle families across three difficulty levels. We evaluate strong reasoning LLMs on TopoBench and find that even frontier models solve fewer than one quarter of hard instances, with two families nearly unsolved. To investigate whether these failures stem from reasoning limitations or from difficulty extracting and maintaining spatial constraints, we annotate 750 chain of thought traces with an error taxonomy that surfaces four candidate causal failure modes, then test them with targeted interventions simulating each error type. These interventions show that certain error patterns like premature commitment and constraint forgetting have a direct impact on the ability to solve the puzzle, while repeated reasoning is a benign effect of search. Finally we study mitigation strategies including prompt guidance, cell-aligned grid representations and tool-based constraint checking, finding that the bottleneck lies in extracting constraints from spatial representations and not in reasoning over them. Code and data are available at github.com/mayug/topobench-benchmark.

Although Large Audio-Language Models (LALMs) deliver state-of-the-art (SOTA) performance, they frequently suffer from hallucinations, e.g. generating text not grounded in the audio input. We analyze these grounding failures and identify a distinct taxonomy: Event Omission, False Event Identity, Temporal Relation Error, and Quantitative Temporal Error. To address this, we introduce the AHA (Audio Hallucination Alignment) framework. By leveraging counterfactual hard negative mining, our pipeline constructs a high-quality preference dataset that forces models to distinguish strict acoustic evidence from linguistically plausible fabrications. Additionally, we establish AHA-Eval, a diagnostic benchmark designed to rigorously test these fine-grained temporal reasoning capabilities. We apply this data to align Qwen2.5-Omni. The resulting model, Qwen-Audio-AHA, achieves a 13.7% improvement on AHA-Eval. Crucially, this benefit generalizes beyond our diagnostic set. Our model shows substantial gains on public benchmarks, including 1.3% on MMAU-Test and 1.6% on MMAR, outperforming latest SOTA methods.

Tool-integrated reasoning (TIR) has become a key approach for improving large reasoning models (LRMs) on complex problems. Prior work has mainly studied when to invoke tools, while overlooking how tools are applied. We identify two common patterns: a calculator pattern that uses code for direct computation, and an algorithmic pattern that encodes problems as programs. Misaligned choices often cause failures even when reasoning is sound. We propose a two-stage framework that first builds code competence from both patterns and then aligns pattern selection with teacher preferences. Across challenging math datasets, our pattern-aware method substantially improves both code usage and accuracy, for instance raising Code@1 on MATH500 from 64.0% to 70.5% and on AIME24 from 26.7% to 50.0%. These gains highlight the effectiveness of a pattern-aware approach for tool-integrated reasoning.

Large Language Models (LLMs) exhibit positional bias, struggling to utilize information from the middle or end of long contexts. Our study explores LLMs' long-context reasoning by probing their hidden representations. We find that while LLMs encode the position of target information, they often fail to leverage this in generating accurate responses. This reveals a disconnect between information retrieval and utilization, a "know but don't tell" phenomenon. We further analyze the relationship between extraction time and final accuracy, offering insights into the underlying mechanics of transformer models.

Mathematical reasoning is a primary indicator of large language models (LLMs) intelligence. However, existing LLMs exhibit failures of robustness and generalization. This paper attributes these deficiencies to spurious reasoning, i.e., producing answers from superficial features. To address this challenge, we propose the AdaR framework to enable adaptive reasoning, wherein models rely on problem-solving logic to produce answers. AdaR synthesizes logically equivalent queries by varying variable values, and trains models with RLVR on these data to penalize spurious logic while encouraging adaptive logic. To improve data quality, we extract the problem-solving logic from the original query and generate the corresponding answer by code execution, then apply a sanity check. Experimental results demonstrate that AdaR improves robustness and generalization, achieving substantial improvement in mathematical reasoning while maintaining high data efficiency. Analysis indicates that data synthesis and RLVR function in a coordinated manner to enable adaptive reasoning in LLMs. Subsequent analyses derive key design insights into the effect of critical factors and the applicability to instruct LLMs. Our project is available at https://github.com/LaiZhejian/AdaR

Large language model (LLM)-based systems are increasingly deployed to conduct scientific research autonomously, yet whether their reasoning adheres to the epistemic norms that make scientific inquiry self-correcting is poorly understood. Here, we evaluate LLM-based scientific agents across eight domains, spanning workflow execution to hypothesis-driven inquiry, through more than 25,000 agent runs and two complementary lenses: (i) a systematic performance analysis that decomposes the contributions of the base model and the agent scaffold, and (ii) a behavioral analysis of the epistemological structure of agent reasoning. We observe that the base model is the primary determinant of both performance and behavior, accounting for 41.4% of explained variance versus 1.5% for the scaffold. Across all configurations, evidence is ignored in 68% of traces, refutation-driven belief revision occurs in 26%, and convergent multi-test evidence is rare. The same reasoning pattern appears whether the agent executes a computational workflow or conducts hypothesis-driven inquiry. They persist even when agents receive near-complete successful reasoning trajectories as context, and the resulting unreliability compounds across repeated trials in epistemically demanding domains. Thus, current LLM-based agents execute scientific workflows but do not exhibit the epistemic patterns that characterize scientific reasoning. Outcome-based evaluation cannot detect these failures, and scaffold engineering alone cannot repair them. Until reasoning itself becomes a training target, the scientific knowledge produced by such agents cannot be justified by the process that generated it.

With the rapid advancement of agent-based methods in recent years, Agentic RAG has undoubtedly become an important research direction. Multi-hop reasoning, which requires models to engage in deliberate thinking and multi-step interaction, serves as a critical testbed for assessing such capabilities. However, existing benchmarks typically provide only final questions and answers, while lacking the intermediate hop-level questions that gradually connect atomic questions to the final multi-hop query. This limitation prevents researchers from analyzing at which step an agent fails and restricts more fine-grained evaluation of model capabilities. Moreover, most current benchmarks are manually constructed, which is both time-consuming and labor-intensive, while also limiting scalability and generalization. To address these challenges, we introduce AgenticRAGTracer, the first Agentic RAG benchmark that is primarily constructed automatically by large language models and designed to support step-by-step validation. Our benchmark spans multiple domains, contains 1,305 data points, and has no overlap with existing mainstream benchmarks. Extensive experiments demonstrate that even the best large language models perform poorly on our dataset. For instance, GPT-5 attains merely 22.6\% EM accuracy on the hardest portion of our dataset. Hop-aware diagnosis reveals that failures are primarily driven by distorted reasoning chains -- either collapsing prematurely or wandering into over-extension. This highlights a critical inability to allocate steps consistent with the task's logical structure, providing a diagnostic dimension missing in traditional evaluations. We believe our work will facilitate research in Agentic RAG and inspire further meaningful progress in this area. Our code and data are available at https://github.com/YqjMartin/AgenticRAGTracer.

Prompt optimization methods either analyze individual failures in isolation or compare prompt variants across examples, operating on single execution traces with no access to the reasoning process distinguishing success from failure on the same input. We introduce ContraPrompt, built on the observation that when a model fails but succeeds on a retry with feedback, the difference between its two chain-of-thought traces constitutes an optimization signal not captured by prior methods. Unlike prior contrastive methods, we compare complete intermediate reasoning processes: the two traces share model, input, and base prompt, so remaining differences reflect reasoning strategy and appended error feedback -- we call this dyadic reasoning trace analysis. The multi-attempt solving phase is an instrumented agentic retry loop that generates contrastive data automatically without human annotation. Extracted rules are organized into an input-aware decision tree routing instructions by observable input characteristics. On four reasoning and compliance benchmarks, ContraPrompt outperforms GEPA (Agrawal et al., 2026) on all four, with absolute gains of +8.29 pp on HotPotQA (+20.8% rel.), +2.21 pp on GDPR-Bench (+18.2% rel.), +7.14 pp on GPQA Diamond (+10.6% rel.), and +0.74 pp on BBH (+0.85% rel.). Ablations confirm dyadic trace contrastivity is the critical component, with a -16% relative average drop upon its removal. On 53 EvalSet black-box optimization problems, ContraPrompt beats GEPA on 11, ties on 41, and loses on 1 at equal budget. On FiNER-139 financial named entity recognition (Loukas et al., 2022), ContraPrompt achieves +7.77 pp over the unoptimized baseline (+11.6% rel.) and +1.94 pp over GEPA (+2.66% rel.), with branch conditions aligning with standard US GAAP financial-instrument categories.

Large language models (LLMs) struggle with multi-step reasoning, where inference-time scaling has emerged as a promising strategy for performance improvement. Verifier-guided search outperforms repeated sampling when sample size is limited by selecting and prioritizing valid reasoning paths. However, we identify a critical limitation: scaling flaws, prevalent across different models (Mistral 7B and DeepSeekMath 7B), benchmarks (GSM8K and MATH), and verifiers (outcome value models and process reward models). As sample size increases, verifier-guided search exhibits diminishing advantages and eventually underperforms repeated sampling. Our analysis attributes this to verifier failures, where imperfect verifiers misrank candidates and erroneously prune all valid paths. These issues are further exacerbated in challenging and out-of-distribution problems, restricting search effectiveness. To mitigate verifier failures, we explore reducing reliance on verifiers and conduct preliminary investigations using two simple methods. Our findings reveal fundamental limitations in verifier-guided search and suggest future directions.

We endow Large Language Models (LLMs) with fine-grained self-evaluation to refine multi-step reasoning inference. We propose an effective prompting approach that integrates self-evaluation guidance through stochastic beam search. Our approach explores the reasoning search space using a well-calibrated automatic criterion. This enables an efficient search to produce higher-quality final predictions. With the self-evaluation guided stochastic beam search, we also balance the quality-diversity trade-off in the generation of reasoning chains. This allows our approach to adapt well with majority voting and surpass the corresponding Codex-backboned baselines by 6.34%, 9.56%, and 5.46% on the GSM8K, AQuA, and StrategyQA benchmarks, respectively, in few-shot accuracy. Analysis of our decompositional reasoning finds it pinpoints logic failures and leads to higher consistency and robustness. Our code is publicly available at https://github.com/YuxiXie/SelfEval-Guided-Decoding.

Large language models often produce hallucinated answers that violate prompt-level constraints. A key diagnostic question is whether these failures reflect missing knowledge, or whether the model has the relevant information but follows the wrong inference path. We study this phenomenon as inference misalignment: a mismatch between the answer supported by the prompt and the answer favored by statistically salient latent associations. We formalize this view with a latent key-task model, in which pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. We introduce TrapQA, a controlled diagnostic testbed with two components. ScientistQA tests disambiguation among similar scientists with supplementary factual probes, while Real-Life Constrained QA tests everyday constraint following under salient shortcuts. Our results show that hallucination can arise from biased latent inference rather than absent knowledge alone.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal reasoning. However, prevailing reinforcement learning (RL) paradigms lack explicit counterfactual enhancement and causal learning mechanisms. This fundamental deficiency results in severe grounding failures, manifesting as a tendency to ignore visual evidence in favor of language priors or exhibiting hallucination drift during long chain-of-thought reasoning. To address this root cause, we propose CounterFactual Policy Optimization (CFPO), a novel framework that enforces causal consistency between visual perception and textual reasoning. CFPO introduces a cross-modal counterfactual enhancement mechanism, which regularizes the policy by maximizing the discrepancy between the model's predictions and those from a counterfactual state where critical visual cues are suppressed. This approach seamlessly integrates with standard algorithms like GRPO and DAPO without requiring external reward models or additional supervision. Extensive experiments demonstrate that CFPO significantly improves reasoning fidelity, achieving consistent gains of 3.17%-6.25% over standard RL baselines and 1.32%-2.13% over the state-of-the-art perception-aware method (PAPO). Code is available at https://github.com/Raven-July/CFPO.

Disasters are inevitable and increasingly costly, and effective response depends on querying structured tabular data: precise, information-dense records of hazard, exposure, vulnerability, and lifeline infrastructure that underpin disaster management. Current text-to-SQL methods enable natural-language access to such tables but transfer poorly to the disaster domain, where queries span heterogeneous geospatial schemas and require reasoning over causal relations. We introduce DisasterLex, a knowledge-graph-mediated framework that inserts an Expert Knowledge Graph (EKG) of curated concepts and typed causal edges between the user query and the database, bridged to schema by concept-to-table links. The orchestration runs four stages (identifying query entities, routing to the operational domain, planning over causal edges, and grounding the SQL), restricting the schema passed to the model at each step. We instantiate it on a disaster-analytics database (36 geospatial tables, 150 columns) with an EKG of 107 concepts, 117 causal edges, and 52 concept-to-schema links, evaluated on a 75-query test set. On all seven base models spanning proprietary and open-weight families, DisasterLex beats four state-of-the-art baselines (LightRAG, HippoRAG 2, ReFoRCE, CHESS) by 1.4x to 2.75x, with absolute scores of 1.65 to 3.56 (of 5.0). Error analysis shows baseline failures cluster in routing and multi-table SQL composition, the operations our orchestration explicitly addresses. Code, data, and the EKG artifact are available at https://github.com/YimingXiao98/DisasterLex and on Zenodo at https://doi.org/10.5281/zenodo.20388029.

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a powerful paradigm for enhancing the complex reasoning capabilities of Large Reasoning Models. However, standard outcome-based supervision suffers from a critical limitation that penalizes trajectories that are largely correct but fail due to several missteps as heavily as completely erroneous ones. This coarse feedback signal causes the model to discard valuable largely correct rollouts, leading to a degradation in rollout diversity that prematurely narrows the exploration space. Process Reward Models have demonstrated efficacy in providing reliable step-wise verification for test-time scaling, naively integrating these signals into RLVR as dense rewards proves ineffective.Prior methods attempt to introduce off-policy guided whole-trajectory replacement that often outside the policy model's distribution, but still fail to utilize the largely correct rollouts generated by the model itself and thus do not effectively mitigate the narrowing of the exploration space. To address these issues, we propose SCOPE (Step-wise Correction for On-Policy Exploration), a novel framework that utilizes Process Reward Models to pinpoint the first erroneous step in suboptimal rollouts and applies fine-grained, step-wise off-policy rectification. By applying precise refinement on partially correct rollout, our method effectively salvages partially correct trajectories and increases diversity score by 13.5%, thereby sustaining a broad exploration space. Extensive experiments demonstrate that our approach establishes new state-of-the-art results, achieving an average accuracy of 46.6% on math reasoning and exhibiting robust generalization with 53.4% accuracy on out-of-distribution reasoning tasks.

Modern analyst agents must reason over complex, high token inputs, including dozens of retrieved documents, tool outputs, and time sensitive data. While prior work has produced tool calling benchmarks and examined factuality in knowledge augmented systems, relatively little work studies their intersection: settings where LLMs must integrate large volumes of dynamic, structured and unstructured multi tool outputs. We investigate LLM failure modes in this regime using crypto as a representative high data density domain. We introduce (1) CryptoAnalystBench, an analyst aligned benchmark of 198 production crypto and DeFi queries spanning 11 categories; (2) an agentic harness equipped with relevant crypto and DeFi tools to generate responses across multiple frontier LLMs; and (3) an evaluation pipeline with citation verification and an LLM as a judge rubric spanning four user defined success dimensions: relevance, temporal relevance, depth, and data consistency. Using human annotation, we develop a taxonomy of seven higher order error types that are not reliably captured by factuality checks or LLM based quality scoring. We find that these failures persist even in state of the art systems and can compromise high stakes decisions. Based on this taxonomy, we refine the judge rubric to better capture these errors. While the judge does not align with human annotators on precise scoring across rubric iterations, it reliably identifies critical failure modes, enabling scalable feedback for developers and researchers studying analyst style agents. We release CryptoAnalystBench with annotated queries, the evaluation pipeline, judge rubrics, and the error taxonomy, and outline mitigation strategies and open challenges in evaluating long form, multi tool augmented systems.

Lane segment topology reasoning provides comprehensive bird's-eye view (BEV) road scene understanding, which can serve as a key perception module in planning-oriented end-to-end autonomous driving systems. Existing lane topology reasoning methods often fall short in effectively leveraging temporal information to enhance detection and reasoning performance. Recently, stream-based temporal propagation method has demonstrated promising results by incorporating temporal cues at both the query and BEV levels. However, it remains limited by over-reliance on historical queries, vulnerability to pose estimation failures, and insufficient temporal propagation. To overcome these limitations, we propose FASTopoWM, a novel fast-slow lane segment topology reasoning framework augmented with latent world models. To reduce the impact of pose estimation failures, this unified framework enables parallel supervision of both historical and newly initialized queries, facilitating mutual reinforcement between the fast and slow systems. Furthermore, we introduce latent query and BEV world models conditioned on the action latent to propagate the state representations from past observations to the current timestep. This design substantially improves the performance of temporal perception within the slow pipeline. Extensive experiments on the OpenLane-V2 benchmark demonstrate that FASTopoWM outperforms state-of-the-art methods in both lane segment detection (37.4% v.s. 33.6% on mAP) and centerline perception (46.3% v.s. 41.5% on OLS).

Reinforcement Learning with Verifiable Rewards (RLVR) has advanced LLM reasoning, but remains constrained by inefficient exploration under limited rollout budgets, leading to low sampling success and unstable training in complex tasks. We find that many exploration failures arise not from problem difficulty, but from a small number of prompt tokens that introduce interference. Building on this insight, we propose the Less Noise Sampling Framework (LENS), which first prompts by identifying and removing interference tokens. then transfers successful rollouts from the purification process to supervise policy optimization on the original noisy prompts, enabling the model to learn to ignore interference in the real-world, noisy prompting settings. Experimental results show that LENS significantly outperforms GRPO, delivering higher performance and faster convergence, with a 3.88% average gain and over 1.6times speedup. Our work highlights the critical role of pruning interference tokens in improving rollout efficiency, offering a new perspective for RLVR research.

Large Language Models (LLMs) have shown tremendous potential as agents, excelling at tasks that require multiple rounds of reasoning and interactions. Rejection Sampling Fine-Tuning (RFT) has emerged as an effective method for finetuning LLMs as agents: it first imitates expert-generated successful trajectories and further improves agentic skills through iterative fine-tuning on successful, self-generated trajectories. However, since the expert (e.g., GPT-4) succeeds primarily on simpler subtasks and RFT inherently favors simpler scenarios, many complex subtasks remain unsolved and persistently out-of-distribution (OOD). Upon investigating these challenging subtasks, we discovered that previously failed expert trajectories can often provide valuable guidance, e.g., plans and key actions, that can significantly improve agent exploration efficiency and acquisition of critical skills. Motivated by these observations, we propose Exploring Expert Failures (EEF), which identifies beneficial actions from failed expert trajectories and integrates them into the training dataset. Potentially harmful actions are meticulously excluded to prevent contamination of the model learning process. By leveraging the beneficial actions in expert failures, EEF successfully solves some previously unsolvable subtasks and improves agent tuning performance. Remarkably, our approach achieved a 62\% win rate in WebShop, outperforming RFT (53. 6\%) and GPT-4 (35. 6\%), and to the best of our knowledge, setting a new state-of-the-art as the first method to surpass a score of 0.81 in WebShop and exceed 81 in SciWorld.

Video generation models are increasingly used as world simulators for storytelling, simulation, and embodied AI. As these models advance, a key question arises: do generated videos obey the physical laws of the real world? Existing evaluations largely rely on automated metrics or coarse human judgments such as preferences or rubric-based checks. While useful for assessing perceptual quality, these methods provide limited insight into when and why generated dynamics violate real-world physical constraints. We introduce Physion-Eval, a large-scale benchmark of expert human reasoning for diagnosing physical realism failures in videos generated by five state-of-the-art models across egocentric and exocentric views, containing 10,990 expert reasoning traces spanning 22 fine-grained physical categories. Each generated video is derived from a corresponding real-world reference video depicting a clear physical process, and annotated with temporally localized glitches, structured failure categories, and natural-language explanations of the violated physical behavior. Using this dataset, we reveal a striking limitation of current video generation models: in physics-critical scenarios, 83.3% of exocentric and 93.5% of egocentric generated videos exhibit at least one human-identifiable physical glitch. We hope Physion-Eval will set a new standard for physical realism evaluation and guide the development of physics-grounded video generation. The benchmark is publicly available at https://huggingface.co/datasets/PhysionLabs/Physion-Eval.

Large Language Models (LLMs) display striking surface fluency yet systematically fail at tasks requiring symbolic reasoning, arithmetic accuracy, and logical consistency. This paper offers a structural diagnosis of such failures, revealing a persistent gap between comprehension and competence. Through controlled experiments and architectural analysis, we demonstrate that LLMs often articulate correct principles without reliably applying them--a failure rooted not in knowledge access, but in computational execution. We term this phenomenon the computational split-brain syndrome, where instruction and action pathways are geometrically and functionally dissociated. This core limitation recurs across domains, from mathematical operations to relational inferences, and explains why model behavior remains brittle even under idealized prompting. We argue that LLMs function as powerful pattern completion engines, but lack the architectural scaffolding for principled, compositional reasoning. Our findings delineate the boundary of current LLM capabilities and motivate future models with metacognitive control, principle lifting, and structurally grounded execution. This diagnosis also clarifies why mechanistic interpretability findings may reflect training-specific pattern coordination rather than universal computational principles, and why the geometric separation between instruction and execution pathways suggests limitations in neural introspection and mechanistic analysis.

Auto-formalization (AF) translates natural-language reasoning problems into solver-executable programs, enabling symbolic solvers to perform sound logical deduction. In practice, however, AF pipelines are currently brittle: programs may fail to execute, or execute but encode incorrect semantics. While prior work largely mitigates syntactic failures via repairs based on solver feedback, reducing semantics failures remains a major bottleneck. We propose Draft-and-Prune (D&P), an inference-time framework that improves AF-based logical reasoning via diversity and verification. D&P first drafts multiple natural-language plans and conditions program generation on them. It further prunes executable but contradictory or ambiguous formalizations, and aggregates predictions from surviving paths via majority voting. Across four representative benchmarks (AR-LSAT, ProofWriter, PrOntoQA, LogicalDeduction), D&P substantially strengthens AF-based reasoning without extra supervision. On AR-LSAT, in the AF-only setting, D&P achieves 78.43% accuracy with GPT-4 and 78.00% accuracy with GPT-4o, significantly outperforming the strongest AF baselines MAD-LOGIC and CLOVER. D&P then attains near-ceiling performance on the other benchmarks, including 100% on PrOntoQA and LogicalDeduction.

Post-training quantization (PTQ) is widely used to deploy large language models efficiently, but its effect on reasoning models is not well understood. Across math, coding, and science QA, we find that aggressive PTQ reduces accuracy while increasing chain-of-thought (CoT) length. Surprisingly, we show that in up to 52% of the quantized models' failures, models reach the right answer in intermediate reasoning steps but do not output it as a final answer. To understand why quantization leads to this increase in overthinking errors, we measure the token-level KL divergence between quantized and full-precision output distributions. Positions with high KL divergence correlate strongly with high next-token entropy, and at these positions quantized models disproportionately sample overthinking markers such as "wait", "but", and "alternatively". We show that simply introducing a training-free logit penalty on a curated set of overthinking markers can reduce CoT length by 12--23% while preserving or improving accuracy across 5 models (1.5B-32B parameters), 3 quantization methods, and 5 benchmarks, yielding a favorable Pareto frontier of accuracy against reasoning cost compared to penalizing other token sets. Overthinking errors produced by quantized models are particularly reduced by up to 58%.

Large language model (LLM) agents frequently fail on multi-step tasks involving reasoning, tool use, and environment interaction. While such failures are typically logged or retried heuristically, they contain structured signals about where execution broke down. We introduce CausalFlow, an interventional framework that converts failed agent traces into minimal counterfactual repairs and reusable supervision. CausalFlow models execution traces as sequential chains of dependent steps and computes Causal Responsibility Scores(CRS) via step-level counterfactual intervention to identify failure-inducing steps. For these steps, we generate minimally edited repairs that flip the final outcome to success, producing validated contrastive pairs of the form (wrong step, corrected step). CausalFlow supports two complementary uses: targeted test-time repair that recovers from failures with minimal behavioral drift, and training-time supervision suitable for offline preference optimization or reward modeling. Across four benchmarks spanning mathematical reasoning, code generation, question answering, and medical browsing, CausalFlow converts failed executions into validated minimal repairs with high minimality and causal-consensus scores, and demonstrates that causal attribution is necessary for reliable improvement across diverse agent tasks, outperforming heuristic refinement in complex retrieval settings while producing more localized repairs throughout. These results demonstrate that interventional analysis over structured execution traces provides a principled and scalable mechanism for transforming agent failures into reliability gains and learning-ready supervision.

Recent advances in Reinforcement Learning (RL) have underscored its potential for incentivizing reasoning capabilities of Large Language Models (LLMs). However, existing step-level efforts suffer from costly annotations that limit domain coverage, while scalar scores further impose an information bottleneck, offering insufficient semantic bandwidth to improve intermediate decisions. Alternative language-critique approaches, which rely on frozen or external critics, provide richer textual feedback but lack the scalability needed for sustained policy improvement. In this work, we propose language-driven stepwise trajectory redirection, termed as STRIDE, a novel training framework that shifts process supervision from scalar rewards to learnable stepwise language feedback. Specifically, we co-train a generator and a generative verifier using only outcome-based rewards, eliminating external annotations, while delivering sustained policy improvement through jointly aligned verifier training. The verifier's stepwise language critiques explicitly localize and explain failures, enabling the generator to redirect reasoning trajectories at intermediate steps toward alternative decisions. The trajectory redirection design guarantees harmless policy improvement, even under noisy or suboptimal verifier feedback. Experiments on diverse reasoning benchmarks show that STRIDE significantly outperforms state-of-the-art baselines, as well as achieving breakthroughs on zero-pass-rate problems where scalar methods yield no learning signal in our ablation studies, demonstrating the effectiveness of learnable stepwise language feedback for enhancing LLM reasoning.

Modeling plausible student misconceptions is critical for AI in education. In this work, we examine how large language models (LLMs) reason about misconceptions when generating multiple-choice distractors, a task that requires modeling incorrect yet plausible answers by coordinating solution knowledge, simulating student misconceptions, and evaluating plausibility. We introduce a taxonomy for analyzing the strategies used by state-of-the-art LLMs, examining their reasoning procedures and comparing them to established best practices in the learning sciences. Our structured analysis reveals a surprising alignment between their processes and best practices: the models typically solve the problem correctly first, then articulate and simulate multiple potential misconceptions, and finally select a set of distractors. An analysis of failure modes reveals that errors arise primarily from failures in recovering the correct solution and selecting among response candidates, rather than simulating errors or structuring the process. Consistent with these results, we find that providing the correct solution in the prompt improves alignment with human-authored distractors by 8%, highlighting the critical role of anchoring to the correct solution when generating plausible incorrect student reasoning. Overall, our analysis offers a structured and interpretable lens into LLMs' ability to model incorrect student reasoning and produce high-quality distractors.

Recent advances in reasoning-centric models promise improved robustness through mechanisms such as chain-of-thought prompting and test-time scaling. However, their ability to withstand misleading user input remains underexplored. In this paper, we conduct a systematic evaluation of three state-of-the-art reasoning models, i.e., OpenAI's o4-mini, Claude-3.7-Sonnet and Gemini-2.5-Flash, across three multimodal benchmarks: MMMU, MathVista, and CharXiv. Our evaluation reveals significant accuracy drops (25-29% on average) following gaslighting negation prompts, indicating that even top-tier reasoning models struggle to preserve correct answers under manipulative user feedback. Built upon the insights of the evaluation and to further probe this vulnerability, we introduce GaslightingBench-R, a new diagnostic benchmark specifically designed to evaluate reasoning models' susceptibility to defend their belief under gaslighting negation prompt. Constructed by filtering and curating 1,025 challenging samples from the existing benchmarks, GaslightingBench-R induces even more dramatic failures, with accuracy drops exceeding 53% on average. Our findings reveal fundamental limitations in the robustness of reasoning models, highlighting the gap between step-by-step reasoning and belief persistence.

While large language models (LLMs) have shown strong performance in math and logic reasoning, their ability to handle combinatorial optimization (CO) -- searching high-dimensional solution spaces under hard constraints -- remains underexplored. To bridge the gap, we introduce NLCO, a Natural Language Combinatorial Optimization benchmark that evaluates LLMs on end-to-end CO reasoning: given a language-described decision-making scenario, the model must output a discrete solution without writing code or calling external solvers. NLCO covers 43 CO problems and is organized using a four-layer taxonomy of variable types, constraint families, global patterns, and objective classes, enabling fine-grained evaluation. We provide solver-annotated solutions and comprehensively evaluate LLMs by feasibility, solution optimality, and reasoning efficiency. Experiments across a wide range of modern LLMs show that high-performing models achieve strong feasibility and solution quality on small instances, but both degrade as instance size grows, even if more tokens are used for reasoning. We also observe systematic effects across the taxonomy: set-based tasks are relatively easy, whereas graph-structured problems and bottleneck objectives lead to more frequent failures.

Recent advancements in reinforcement learning (RL) for large language models (LLMs), exemplified by DeepSeek R1, have shown that even a simple question-answering task can substantially improve an LLM's reasoning capabilities. In this work, we extend this approach by modifying the task into a multi-attempt setting. Instead of generating a single response per question, the model is given multiple attempts, with feedback provided after incorrect responses. The multi-attempt task encourages the model to refine its previous attempts and improve search efficiency. Experimental results show that even a small LLM trained on a multi-attempt task achieves significantly higher accuracy when evaluated with more attempts, improving from 45.6% with 1 attempt to 52.5% with 2 attempts on the math benchmark. In contrast, the same LLM trained on a standard single-turn task exhibits only a marginal improvement, increasing from 42.3% to 43.2% when given more attempts during evaluation. The results indicate that, compared to the standard single-turn task, an LLM trained on a multi-attempt task achieves slightly better performance on math benchmarks while also learning to refine its responses more effectively based on user feedback. Full code is available at https://github.com/DualityRL/multi-attempt

Recent generations of language models have introduced Large Reasoning Models (LRMs) that generate detailed thinking processes before providing answers. While these models demonstrate improved performance on reasoning benchmarks, their fundamental capabilities, scaling properties, and limitations remain insufficiently understood. Current evaluations primarily focus on established math and coding benchmarks, emphasizing final answer accuracy. However, this evaluation paradigm often suffers from contamination and does not provide insights into the reasoning traces. In this work, we systematically investigate these gaps with the help of controllable puzzle environments that allow precise manipulation of complexity while maintaining consistent logical structures. This setup enables the analysis of not only final answers but also the internal reasoning traces, offering insights into how LRMs think. Through extensive experiments, we show that LRMs face a complete accuracy collapse beyond certain complexities. Moreover, they exhibit a counterintuitive scaling limit: their reasoning effort increases with problem complexity up to a point, then declines despite having remaining token budget. By comparing LRMs with their standard LLM counterparts under same inference compute, we identify three performance regimes: (1) low-complexity tasks where standard models outperform LRMs, (2) medium-complexity tasks where LRMs demonstrates advantage, and (3) high-complexity tasks where both models face complete collapse. We found that LRMs have limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across scales. We also investigate the reasoning traces in more depth, studying the patterns of explored solutions and analyzing the models' computational behavior, shedding light on their strengths, limitations, and raising questions about their reasoning capabilities.

Multimodal symbolic logical reasoning, which aims to deduce new facts from multimodal input via formal logic, is critical in high-stakes applications such as autonomous driving and medical diagnosis, as its rigorous, deterministic reasoning helps prevent serious consequences. To evaluate such capabilities of current state-of-the-art vision language models (VLMs), we introduce the first benchmark MuSLR for multimodal symbolic logical reasoning grounded in formal logical rules. MuSLR comprises 1,093 instances across 7 domains, including 35 atomic symbolic logic and 976 logical combinations, with reasoning depths ranging from 2 to 9. We evaluate 7 state-of-the-art VLMs on MuSLR and find that they all struggle with multimodal symbolic reasoning, with the best model, GPT-4.1, achieving only 46.8%. Thus, we propose LogiCAM, a modular framework that applies formal logical rules to multimodal inputs, boosting GPT-4.1's Chain-of-Thought performance by 14.13%, and delivering even larger gains on complex logics such as first-order logic. We also conduct a comprehensive error analysis, showing that around 70% of failures stem from logical misalignment between modalities, offering key insights to guide future improvements. All data and code are publicly available at https://llm-symbol.github.io/MuSLR.

Text-to-image (T2I) generation models have achieved remarkable progress in producing visually realistic images from natural language prompts. Yet it remains unclear whether their success reflects genuine causal understanding or sophisticated pattern matching over visual-textual correlations. Inspired by Russell's inductivist turkey, we introduce Counterfactual-World (CF-World), a counterfactual benchmark designed to investigate whether text-to-image models can generate images under rules that systematically contradict real-world priors. CF-World organizes each scenario into three progressive levels: factual generation under ordinary world knowledge, explicit counterfactual generation with direct visual instructions, and implicit counterfactual generation requiring causal deduction from altered rules. We evaluate both open-source and closed-source T2I models using a Vision Language Model (VLM)-based evaluator (CF-Eval). Furthermore, we introduce two metrics: Prior Resistance Rate (PRR), which measures a model's ability to overcome entrenched real-world priors, and Reasoning Retention Rate (RRR), which assesses whether models can maintain reasoning-dependent counterfactual generation without explicit visual cues. Experiments show that all models exhibit sharp degradation from factual to counterfactual settings. Further analyses suggest that these failures arise because current T2I models encode world knowledge and visual appearances as tightly coupled patterns. Consequently, their heavy reliance on frequent visual co-occurrences within the training data forces them to default to familiar commonsense priors when tasked with rendering counterfactual worlds.

Natural language understanding often depends on meanings that are implied rather than explicitly stated, requiring pragmatic reasoning. Despite strong performance on math and logical reasoning, large language models (LLMs) still struggle with making pragmatic inferences, often choosing literal interpretations. To improve LLM pragmatic reasoning, we introduce PragReST, a self-supervised framework that constructs pragmatic QA data, generates counterfactual reasoning traces, and trains models to internalize them through supervised fine-tuning and reinforcement learning, without human-labeled training data or distillation from a stronger teacher. Across four pragmatic benchmarks (PragMega, Ludwig, MetoQA, and AltPrag), PragReST improves over backbone models, task-specific pragmatic tuning baselines, and non-counterfactual variants of the same pipeline. On accuracy-based benchmarks, PragReST improves over the instruct backbone by 5.37 and 5.50% (absolute) for Qwen3-8B and Qwen3-14B, respectively. Our error analysis and ablations underscore the importance of counterfactual reasoning: PragReST primarily reduces errors caused by failures to contrast observed utterances with plausible alternatives, and removing counterfactual reasoning substantially reduces performance. Moreover, our training preserves out-of-domain performance on general-knowledge and mathematical reasoning benchmarks.

Current Chain-of-Thought (CoT) verification methods predict reasoning correctness based on outputs (black-box) or activations (gray-box), but offer limited insight into why a computation fails. We introduce a white-box method: Circuit-based Reasoning Verification (CRV). We hypothesize that attribution graphs of correct CoT steps, viewed as execution traces of the model's latent reasoning circuits, possess distinct structural fingerprints from those of incorrect steps. By training a classifier on structural features of these graphs, we show that these traces contain a powerful signal of reasoning errors. Our white-box approach yields novel scientific insights unattainable by other methods. (1) We demonstrate that structural signatures of error are highly predictive, establishing the viability of verifying reasoning directly via its computational graph. (2) We find these signatures to be highly domain-specific, revealing that failures in different reasoning tasks manifest as distinct computational patterns. (3) We provide evidence that these signatures are not merely correlational; by using our analysis to guide targeted interventions on individual transcoder features, we successfully correct the model's faulty reasoning. Our work shows that, by scrutinizing a model's computational process, we can move from simple error detection to a deeper, causal understanding of LLM reasoning.

Hybrid reasoning systems that combine learned components with model-based inference are increasingly deployed in tool-augmented decision loops, yet their runtime behavior under partial observability and sustained evidence mismatch remains poorly understood. In practice, failures often arise as gradual divergence of internal reasoning dynamics rather than as isolated prediction errors. This work studies runtime stability in hybrid reasoning systems from a Kalman-inspired perspective. We model reasoning as a stochastic inference process driven by an internal innovation signal and introduce cognitive drift as a measurable runtime phenomenon. Stability is defined in terms of detectability, bounded divergence, and recoverability rather than task-level correctness. We propose a runtime stability framework that monitors innovation statistics, detects emerging instability, and triggers recovery-aware control mechanisms. Experiments on multi-step, tool-augmented reasoning tasks demonstrate reliable instability detection prior to task failure and show that recovery, when feasible, re-establishes bounded internal behavior within finite time. These results emphasize runtime stability as a system-level requirement for reliable reasoning under uncertainty.

Hierarchical reasoning model (HRM) achieves extraordinary performance on various reasoning tasks, significantly outperforming large language model-based reasoners. To understand the strengths and potential failure modes of HRM, we conduct a mechanistic study on its reasoning patterns and find three surprising facts: (a) Failure of extremely simple puzzles, e.g., HRM can fail on a puzzle with only one unknown cell. We attribute this failure to the violation of the fixed point property, a fundamental assumption of HRM. (b) "Grokking" dynamics in reasoning steps, i.e., the answer is not improved uniformly, but instead there is a critical reasoning step that suddenly makes the answer correct; (c) Existence of multiple fixed points. HRM "guesses" the first fixed point, which could be incorrect, and gets trapped there for a while or forever. All facts imply that HRM appears to be "guessing" instead of "reasoning". Leveraging this "guessing" picture, we propose three strategies to scale HRM's guesses: data augmentation (scaling the quality of guesses), input perturbation (scaling the number of guesses by leveraging inference randomness), and model bootstrapping (scaling the number of guesses by leveraging training randomness). On the practical side, by combining all methods, we develop Augmented HRM, boosting accuracy on Sudoku-Extreme from 54.5% to 96.9%. On the scientific side, our analysis provides new insights into how reasoning models "reason".

Reliable clinical decision support requires medical AI agents capable of safe, multi-step reasoning over structured electronic health records (EHRs). While large language models (LLMs) show promise in healthcare, existing benchmarks inadequately assess performance on action-based tasks involving threshold evaluation, temporal aggregation, and conditional logic. We introduce ART, an Action-based Reasoning clinical Task benchmark for medical AI agents, which mines real-world EHR data to create challenging tasks targeting known reasoning weaknesses. Through analysis of existing benchmarks, we identify three dominant error categories: retrieval failures, aggregation errors, and conditional logic misjudgments. Our four-stage pipeline -- scenario identification, task generation, quality audit, and evaluation -- produces diverse, clinically validated tasks grounded in real patient data. Evaluating GPT-4o-mini and Claude 3.5 Sonnet on 600 tasks shows near-perfect retrieval after prompt refinement, but substantial gaps in aggregation (28--64%) and threshold reasoning (32--38%). By exposing failure modes in action-oriented EHR reasoning, ART advances toward more reliable clinical agents, an essential step for AI systems that reduce cognitive load and administrative burden, supporting workforce capacity in high-demand care settings

Multimodal large language models (MLLMs) promise enhanced reasoning by integrating diverse inputs such as text, vision, and audio. Yet cross-modal reasoning remains underexplored, with conflicting reports on whether added 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. 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.

The ability of large vision-language models (LVLMs) to critique and correct their reasoning is an essential building block towards their self-improvement. However, a systematic analysis of such capabilities in LVLMs is still lacking. We propose VISCO, the first benchmark to extensively analyze the fine-grained critique and correction capabilities of LVLMs. Compared to existing work that uses a single scalar value to critique the entire reasoning [4], VISCO features dense and fine-grained critique, requiring LVLMs to evaluate the correctness of each step in the chain-of-thought and provide natural language explanations to support their judgments. Extensive evaluation of 24 LVLMs demonstrates that human-written critiques significantly enhance the performance after correction, showcasing the potential of the self-improvement strategy. However, the model-generated critiques are less helpful and sometimes detrimental to the performance, suggesting that critique is the crucial bottleneck. We identified three common patterns in critique failures: failure to critique visual perception, reluctance to "say no", and exaggerated assumption of error propagation. To address these issues, we propose an effective LookBack strategy that revisits the image to verify each piece of information in the initial reasoning. LookBack significantly improves critique and correction performance by up to 13.5%.