Large Language Model Reasoning

Greenland iceberg discharge exhibits complex nonlinear dynamics with limited observability, challenging traditional predictive models. We present a Hybrid NARX-LLM framework that combines a nonlinear autoregressive model with exogenous inputs (NARX) and a large language model (LLM) for residual correction. We further propose a Physics-Informed Prompt (PIP) method that transforms unstructured physical knowledge into structured prompts for zero-shot in-context reasoning. The primary objective is to explore the corrective potential of this framework for modeling Greenland iceberg discharge, rather than merely optimizing predictive accuracy. The NARX component captures intrinsic temporal dependencies, while the LLM, guided by PIP, encodes glacier dynamics and environmental drivers and perceives key trend patterns to correct systematic prediction errors. This integration allows the model to reason about unmodeled factors and produce interpretable residuals, enhancing overall predictive accuracy. Applied to Greenland iceberg discharge time series, our approach addresses extreme events that are difficult to predict due to rare variations and nonstationary trends, a limitation often overlooked by traditional methods. By fusing structured time-series modeling with knowledge-driven foundation AI, the framework offers a scalable and interpretable pathway to bridge data-limited climate forecasting with physics-informed LLM reasoning. The code is available.

Reasoning capabilities of multimodal large language models (MLLMs) have improved considerably in recent years. Existing approaches typically rely on explicit chain-of-thought or continuous latent-space trajectories to enhance multi-step reasoning. However, these methods generally assume that an input admits a single latent interpretation and unfold reasoning along a fixed path or under a uniform computation budget. In real-world multimodal settings, visual observations are often subject to occlusion, blur, viewpoint variation, or semantic ambiguity, giving rise to multiple plausible interpretations. A uniform reasoning strategy not only limits the model's ability to explore multiple hypotheses but also incurs high memory usage and rollout cost. We present DLWM (Diverse Latent World Models), a multimodal reasoning framework that combines latent-space reasoning with reinforcement learning. First, we construct a set of diverse latent world hypotheses in continuous latent space, each capturing a different plausible interpretation of the visual input, and unfold latent reasoning independently on each hypothesis. An orthogonality-based diversity regularizer explicitly prevents hypothesis collapse. Second, we formulate the latent reasoning process as a resource-constrained sequential decision problem and introduce a resource-aware reinforcement learning policy that adaptively allocates computation across hypotheses, dynamically deciding whether to expand, terminate, or merge reasoning paths, thereby substantially reducing memory footprint and improving rollout efficiency. Experiments on multiple multimodal reasoning benchmarks demonstrate that DLWM outperforms existing methods by 2-5 points in accuracy while reducing memory usage by 24%.

Reasoning models are increasingly used in settings where the final answer is not the only object of review: educational tools may show students intermediate steps, decision-support systems may require human oversight, and audit workflows may inspect traces for misleading or biased input. In such settings, two responses can receive the same final-answer score while differing in whether the trace explicitly flags injected biasing content. Accuracy-only evaluation collapses these cases. We study this gap as a measurement blind spot for responsible evaluation and introduce a minimal trace-level diagnostic with two axes: \emph{susceptibility} (whether the bias breaks a previously correct answer) and \emph{acknowledgment} (whether the trace contains a rubric-defined surface reference to the injected content). Across thousands of biased GSM8K trials, GPT-4o and Claude Sonnet~4 have similar susceptibility rates ($1.3\%$ vs. $1.2\%$) but substantially different acknowledgment rates ($13.0\%$ vs. $75.0\%$) under the same rubric.

While Large Reasoning Models (LRMs) show strong performance in English, they often fail to reason in the language of the query, a phenomenon known as language collapse. Existing RL-based fixes typically add a binary language fidelity reward to the accuracy objective, yet still incur trade-off in accuracy, mid-trace code-switching, and excessive token usage. In this work, we propose AdaMame, a two-stage training recipe for multilingual mathematical reasoning that addresses these limitations by adaptively aligning the reasoning language to the query language without compromising accuracy. The first SFT stage fine-tunes on non-MT reasoning traces across five languages to establish multilingual reasoning capability. In the subsequent RL stage, we introduce AdaMame-GRPO, an adaptation of Group Relative Policy Optimization (GRPO) in which a query-conditioned alignment factor grows progressively during training, guiding the model to first explore diverse reasoning languages before exploiting reasoning in the query language. Evaluated across two benchmarks, two LRMs, and 12 languages, AdaMame-GRPO achieves Pareto-optimal performance across reasoning accuracy, language fidelity, and token efficiency over all baselines, with the strongest gains on out-of-domain, lower-resource languages.

By incorporating test-time compute scaling, large reasoning models (LRMs) can solve complex problems through explicit chain-of-thought (CoT) reasoning processes. However, they often suffer from overthinking, resulting in redundant token outputs and degraded accuracy. Current methods to mitigate this issue remain limited: training-based approaches require substantial computational resources, while training-free methods rely on well-crafted prompts or unreliable confidence signals. In this work, we investigate early stopping from the perspective of attention distributions and propose a simple method, ASAG, which infers the model's reasoning state and adaptively adjusts the generation strategy. The proposed framework is training-free and plug-and-play, enabling seamless integration into existing LRMs. Extensive experiments on nine benchmarks demonstrate consistent improvements across mainstream LRMs with varying parameter scales, including the DeepSeek-R1-Distill and Qwen3 series. Specifically, ASAG improves average accuracy by 3.2% while reducing the number of generated tokens by nearly 40% across all reasoning tasks on Qwen3-8B.

Chain-of-thought (CoT) reasoning can improve LLM performance, but high answer confidence may be misleading when the accompanying CoT rationale is plausible yet incomplete or poorly supported. We study confidence--rationale alignment: whether a model's confidence in its committed answer is justified by its generated rationale. We introduce a GRPO-based reinforcement learning framework that jointly rewards answer correctness, committed-answer probability, and rubric-based rationale support, where the rubric assesses grounding, coherence, task match, and connection to the selected answer without revealing the gold answer to the judge. Across MedQA, MathQA, and OpenBookQA using three open-weight LLMs, our method reduces the confidence--rationale alignment error by up to 26.51% compared with untuned checkpoints, SFT, and correctness-only GRPO, while maintaining competitive accuracy and often improving calibration. These results show that reliable CoT reasoning requires not only confident answers, but rationales that substantively support them.

Frontier reasoning-tuned language models still fail on deductive tasks at depth, and the cost of improved performance through extended internal reasoning scales poorly. Symbolic delegation offers a complementary route: a language model translates the problem, while a solver performs the inference. However, current autoformalization pipelines for logic programming are typically bespoke integrations tied to particular tasks or agents. We introduce PrologMCP, a task-agnostic, open-source server that exposes Prolog as a stateful tool through the Model Context Protocol (MCP). Its compact tool interface, structured error reporting, and per-session isolation make the translate-run-inspect-repair loop a reusable primitive for MCP-capable agents. We evaluate a formalizer agent enhanced with PrologMCP against standard and reasoning LLMs (Claude Sonnet 4.6, GPT-4.1, and o4-mini) on two subsets of PARARULE-Plus: a general-purpose sample and a more challenging one targeting a specific failure mode of natural-language reasoning. On the general sample, the formalizer matches or exceeds reasoning LLMs (accuracy 1.00 vs.\ 1.00 / 0.998), with the largest gains over standard models (0.762 for GPT-4.1). On the challenging subset, the formalizer remains near-perfect (1.00 / 0.99) while reasoning LLMs drop to 0.95 / 0.94. These results suggest that delegating inference to Prolog via MCP is a robust and inspectable alternative to extended natural-language reasoning.

Large reasoning models typically follow a read-then-think paradigm: they observe the complete input, reason over a static context, and then produce the answer. Yet many real-world scenarios are inherently dynamic, such as audio and video stream, where information arrives as a continuous stream and models must reason, update, and respond under partial observations. Recent streaming reasoning methods allow models to think while reading, but they largely rely on supervised imitation of pre-constructed trajectories, which limits their flexibility. In this paper, we propose AdaSR, an adaptive streaming reasoning framework that enables models to reason during input streaming and perform final deliberation once the stream is complete, learning when to think, and how much computation to allocate across different stages. To optimize this hierarchical reasoning process, we introduce Hierarchical Relative Policy Optimization (HRPO), which decomposes policy optimization into streaming reasoning and deep reasoning phases, providing more fine-grained advantage assignment instead of uniformly distributing a single sequence-level advantage over all tokens. HRPO integrates format, accuracy, and adaptive thinking rewards to enforce valid reasoning protocols, preserve final task performance, and encourage latency-aware computation allocation. Experiments show that AdaSR achieves a better balance among reasoning accuracy, computational efficiency, and streaming latency compared with supervised fine-tuning baseline. We release our code at https://github.com/EIT-NLP/StreamingLLM/tree/main/AdaSR.

Reinforcement learning with verifiable rewards (RLVR) has successfully elicited the reasoning capabilities of large language models, motivating its extension to multimodal scenarios. Existing methods primarily focus on improving the visual coverage of reasoning traces and mitigating visual hallucinations, but underestimate the semantic inconsistency between the reasoning process and the final answer. In this paper, we delve into thinking-answer inconsistency in RLVR for large vision-language models (LVLMs), showing thorough analyses of rollouts collected throughout Group Relative Policy Optimization (GRPO) training process and post-RLVR evaluation outputs that this issue persists during training and remains present during inference. Motivated by the analysis, we propose Consistency-Oriented Reasoning Alignment (CORA), which introduces thinking-answer semantic consistency into RLVR through a lightweight plug-and-play consistency reward model, and further incorporates Hybrid Reward Advantage Splitting (HRAS) to stably coordinate task and consistency optimization. Extensive experiments across representative multimodal reasoning benchmarks and mainstream LVLMs show that CORA improves task performance while effectively mitigating thinking-answer inconsistency, leading to more faithful reasoning traces.

How to define a good explanation is a long-standing philosophical debate which has found recent renewed interest in the context of AI outputs. Explainability is crucial for AI adoption in many contexts, but in order to produce good explanations of AI systems, we must first have an understanding of what good explanations are. In this paper we propose a definition inspired by the notion of counterfactual explanations, however we argue that one must also take into account the interlocutor's prior beliefs in each fact that could be offered in an explanation. We explore the ramifications of this definition for AI explainability and, in particular, why LLM outputs are difficult to produce good explanations for.

Large Audio-Language Models (LALMs) have shown strong performance on a wide range of audio understanding tasks, yet they still struggle with complex audio reasoning. A practical way to improve such capabilities is post-training, whose effectiveness critically depends on the quality and diversity of training data. However, existing audio-language datasets often contain substantial redundancy, where many samples are highly similar in acoustic content and thus provide overlapping supervisory signals. Such redundancy not only increases annotation cost, but also limits corpus diversity and reduces the effectiveness of post-training. To address this issue, we propose a redundancy-aware data construction pipeline for building reasoning-oriented supervision for LALMs. Specifically, we first perform acoustic similarity-based deduplication across raw audio datasets to improve corpus diversity. We then integrate existing audio captions and question-answer pairs into a unified multiple-choice format. Based on these unified annotations, we leverage Qwen3-30B to generate chain-of-thought (CoT) rationales for reasoning-oriented supervision. Based on this pipeline, we construct AudioDER, a reasoning-oriented post-training dataset containing approximately 191k samples spanning sound, speech, and music. Each sample consists of an audio clip, a multiple-choice question, four answer candidates, an audio caption, and a CoT rationale. Extensive experiments show that post-training on AudioDER consistently improves the performance of Qwen2-Audio-7B-Instruct on multiple audio reasoning benchmarks, including MMAU-mini, MMSU, and MMAR. We hope AudioDER can serve as a valuable resource for advancing audio reasoning research and the development of more capable LALMs.

Large Language Models (LLMs) are undergoing a fundamental transformation from conversational generators into integrated AI systems capable of reasoning, action, memory, and self-improvement. We conceptualize this transition as a shift from Chatbot to Digital Colleague: from conversational answers to persistent work. We organize this transition along two tightly coupled dimensions. First, at the cognitive core level, LLMs are advancing from Chatbot-era "fast thinking" systems driven by next-token prediction toward Thinking LLMs that leverage inference-time computation, Chain-of-Thought reasoning, reflection, process supervision, and reinforcement learning to support more deliberate and reliable cognition. Second, at the tool-augmented task execution level, LLMs are progressing from tool-calling Agents that invoke external resources in an ad hoc manner toward OpenClaw-style workstation systems (OpenClaw) equipped with persistent Workspaces, skills, verification loops, and governance. The "Workspace + Skill" paradigm makes episodic tool use colleague-like via state persistence, reusable procedures, task closure, and experience reuse. We examine data construction shifts from instruction-response pairs to State-Action-Observation trajectories and evaluation from static benchmarks to sandboxed, auditable, self-evolving AI ecosystems.

We present CacheRL, a system for training small agent foundation models that achieves 92 percent process accuracy on multi-step tool-calling tasks, approaching GPT-5's 94 percent while requiring 100 times less compute. Our approach addresses three challenges in practical agent training: transferring tool-calling knowledge from large models at scale, enabling reinforcement learning without costly live tool execution, and learning robustly from noisy cached environments. CacheRL introduces three key innovations. First, a hybrid thinking trajectory pipeline augments agent trajectories with LLM-generated reasoning traces, producing training examples that teach models not only what tools to call but also why. Second, the CacheAgentLoop eliminates live execution costs through a three-tier fuzzy cache while preserving trajectory fidelity using token-level masking. Third, a cache-tier-aware reward dynamically adjusts answer-quality weights to avoid penalizing models for cache-induced limitations. Through iterative supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO), CacheRL improves Qwen3-4B-Thinking's validation reward from 0.43 to 0.78. On public agentic tool-calling benchmarks, our model achieves competitive performance against frontier models such as GPT-5. Ablation studies show that removing knowledge transfer reduces performance by 41 percent, while cache-aware rewards contribute a 17 percent improvement. Interestingly, reinforcement learning improves training stability but yields limited gains beyond strong supervised fine-tuning, suggesting that data quality and reward design play a more important role than complex optimization methods in building practical small agent models.

Understanding student errors in the programming is a cornerstone of programming education, yet obtaining a representative set of student errors for any newly designed task remains slow and costly, since authentic submissions only accumulate after extensive classroom deployment. This paper explores whether large language models (LLMs) can serve as scalable proxies for students by simulating realistic logical errors in code submissions. Using the CodeWorkout dataset of 74,000+ unique student Java submissions across 37 problems, we evaluate five LLMs under three mainstream prompting strategies: Input-Output (IO), Chain-of-Thought (CoT), and iterative Self-Refine. We assess performance along two key dimensions: diversity (the range of distinct error patterns) and alignment (alignment with authentic student mistakes), and examine how these vary by struggling level of programming tasks. Our quantitative findings reveal that while all models generate diverse errors, their alignment to human submissions diverges: Claude Sonnet 4 achieves the most balanced performance. In addition, we conducted a blinded expert annotation study (N = 401) comparing synthetic and authentic errors. This qualitative analysis confirms that the generated errors are functionally indistinguishable from authentic student errors. Moreover, higher-struggling-level problems elicit more diverse but less student-like errors. These results highlight trade-offs in using LLMs to simulate human learners and suggest design considerations for integrating synthetic errors into teachable agents, intelligent tutoring systems, and large-scale learning analytics.

Sampling plays an important role in long-form language-model reasoning. Over thousands of decoding steps, small changes in the candidate token set can compound into different reasoning trajectories, stability profiles, and final answers. Existing truncation methods such as top-$p$, min-$p$, and fixed top-$nσ$ sampling improve over unrestricted sampling, but they rely on fixed thresholds that cannot adapt to changes in entropy, task difficulty, training stage, or generation budget. We introduce Adaptive Nucleus Truncation Sampling (ANTS), which extends top-$nσ$ sampling from a fixed decoding rule into an adaptive rollout-control mechanism for long-form generation. ANTS selects standardized neighborhoods around the maximum logit before temperature scaling, adapts the truncation width using an entropy-conditioned controller, and retains a no-truncation fallback arm to stabilize training when truncation becomes unsafe. On a 33B-total / 4B-active sparse Mixture-of-Experts reasoning model, ANTS improves average performance over percentage-based benchmarks by +1.9, +3.8, and +5.2 points at 8K, 16K, and 32K generation budgets, respectively. The strongest gains appear on instruction following and mathematical reasoning, with IFBench improving by more than 10 points at 32K and AIME 2025 improving by 7 points. Code generation reveals an important budget interaction. On Codeforces, ANTS trails the baseline at 8K, but reverses this gap and substantially improves ELO at 16K and 32K. These results suggest that sampler design should be treated not just as a decoding hyperparameter, but as part of how we stabilize and scale long-budget reasoning.

Long Chain-of-Thought (CoT) reasoning improves LLM problem-solving but is computationally expensive due to sequential token generation. While recent works explore reasoning in continuous latent spaces to bypass discrete token generation, they often struggle with training stability and fail to scale to complex, long-horizon tasks due to lack of supervision signal. We propose SuperThoughts, which compresses pairs of consecutive CoT tokens into single latent representations and decodes two tokens per step via a lightweight Multi-Token Prediction (MTP) module. This preserves discrete token supervision at training time while doubling throughput at inference time. We finetune Qwen2.5-Math-1.5B-Instruct, Qwen2.5-Math-7B-Instruct, Qwen2.5-Math-14B-Instruct, and evaluate on MATH500, AMC, OlympiadBench, and GPQA-Diamond. With a confidence-based adaptive mechanism that falls back to standard decoding when uncertain, SuperThoughts achieves $\sim$20--30% CoT length reduction while maintaining accuracy with minimal degradation (1-2 points accuracy drop on most tasks).

While test-time scaling has revolutionized reasoning in large language models, generative video reasoning remains bottlenecked by a single-shot paradigm. We demonstrate that searching over denoising steps cannot rescue logically flawed rollouts because spatial trajectories commit early in the diffusion process. Root-level Best-of-N (BoN) sampling is similarly inefficient: reasoning errors cluster early in the temporal axis, and resampling blindly discards verified upstream progress. To unlock effective test-time scaling for video models, we introduce Temporal Backtracking Search (TBS), which shifts the search space to the temporal axis. TBS transforms video generation into an iterative generate-verify-restart loop via three core mechanisms: (1) variable-K conditioning to resume generation from arbitrary clean prefixes; (2) temporal process verification to localize failures and extract valid restart anchors; and (3) prefix-based search to reallocate compute toward extending correct trajectories rather than root resampling. Across algorithmic, navigation, and robotics domains, TBS Pareto-dominates matched-budget BoN. In a strict out-of-distribution setting where one-shot generation collapses (0.7% for BoN), TBS achieves 22.7%, with every solved episode stemming from a restarted branch. Ultimately, TBS reveals that the local reasoning competence of video models far exceeds what single-shot rollouts indicate, providing a scalable test-time framework to unlock it.

Large Language Models (LLMs) have made notable progress in automated theorem proving, yet existing formal benchmarks remain limited in both mathematical coverage and difficulty. Most are concentrated in areas that are easier to formalize, such as algebra and elementary number theory, and provide limited coverage of subfields that require deeper reasoning, including mathematical analysis. To address this gap, we introduce MA-ProofBench, to the best of our knowledge, the first formal theorem-proving benchmark dedicated to Mathematical Analysis. The benchmark contains 200 formalized theorems covering 6 core topics and 27 subcategories, including measure and integration theory, complex analysis, and functional analysis. The problems are divided into two difficulty levels, an undergraduate level (Level I, 100 problems) and a Ph.D. qualifying level (Level II, 100 problems), to evaluate how well LLMs perform formal reasoning at different mathematical depths. Each problem is constructed through a human-led, LLM-assisted formalization pipeline followed by independent expert review, ensuring that the formal statements remain faithful to the original mathematics. We evaluate a range of recent general-purpose reasoning models and formal theorem provers on MA-ProofBench. However, most models perform poorly: even the best-performing model, GPT-5.5, achieves only 16% Pass@8 on Level I and 5% on Level II, while most models stay close to 0% on Level II. Further analysis identifies Mathlib hallucinations and incomplete proofs as the two dominant failure modes, while an evaluation on the natural-language version of the benchmark exposes a clear gap between informal and formal reasoning. MA-ProofBench is intended to serve as a reliable reference for tracking progress in formal mathematical reasoning in advanced domains.

Detecting LLM reasoning failures at inference time without ground-truth labels has motivated a wide range of confidence baselines, including self-consistency, semantic entropy, and P(True), built on within-question sampling and self-evaluation. Operad theory, the formalism for systems built by iterated substitution, suggests a complementary diagnostic: a model's direct answer to a compositional query should agree with the answer it produces by composing a stated decomposition of the same query. We instantiate this idea as operadic consistency (OC), a per-question signal. Across twelve instruction-tuned LLMs (4B to 671B parameters, open-weights and closed-source) on four multi-hop QA datasets, OC is strongly correlated with accuracy on every dataset (Pearson $r \in [0.86, 0.94]$, all $p \leq 0.0004$), and is the only signal we evaluate with $r \geq 0.85$ uniformly across all four datasets. Chain-of-thought self-consistency (CoT-SC; Wang et al., 2023) matches OC on HotpotQA and DROP ($r = 0.93, 0.87$) but drops to $r \approx 0.45$ on MuSiQue and StrategyQA. At the per-question level, OC contributes information beyond CoT-SC and semantic entropy on every dataset (cluster-robust $p \leq 10^{-16}$ for the OC coefficient), and the conclusion is robust to additionally controlling for constructed decomposition-aware baselines ($p \leq 10^{-13}$). The same signal yields selective-prediction improvements (accuracy at fixed coverage) over a tuned CoT-SC baseline at the equal-cost $K = 3$ budget (AUARC lifts of +0.086 to +0.096 and AUROC lifts of +0.092 to +0.164; 95% CIs exclude zero on every cell). On five frontier thinking models, where the decomposition is extracted from the model's own chain of thought, the same equal-cost comparison gives positive selective-prediction point-estimate lift on all 16 (dataset, budget, metric) cells tested, with 95% CIs excluding zero on 12 of the 16.

Question decomposition, i.e. breaking a complex query into simpler sub-queries whose answers are composed to produce a final answer, is a widely used strategy for improving LLM reasoning, yet it currently lacks a rigorous mathematical foundation. In this paper, we propose operads, mathematical structures that model many-in, one-out operations and compositions thereof, as a natural framework for describing question decomposition. We define the questions operad $Q$, in which operations correspond to question templates and composition corresponds to substitution of sub-answers, and show how QA models can be interpreted as algebras over $Q$. Beyond reframing existing practice, this operadic perspective points toward new methods, in particular a notion of operadic consistency, which measures whether a QA model's answers agree across the partial collapses of a question decomposition tree. Empirical evaluation of operadic consistency is reported in our companion paper (Bottman, Liu, and Richardson, 2026), which finds it strongly correlated with accuracy across twelve LLMs and four multi-hop QA datasets and outperforming standard temperature-based self-consistency baselines. We argue that operads are the natural mathematical home for question decomposition, and that invariants such as operadic consistency open new directions for analyzing and improving the reliability of multi-step reasoning.

When large language models (LLMs) fail to generalize or make haphazard errors in reasoning, it is often taken as evidence that LLMs are not truly reasoning, but rather performing a kind of pattern matching. The implication is that people's behavior does not exhibit the same types of failures because human reasoning uses principled and abstract world models. We evaluate human participants and 25 LLMs on their ability to engage in common-sense reasoning about a variety of everyday situations and observe similar patterns of errors in both people and models. We then identify the set of attention heads driving LLM responses and find that these heads implement a form of pattern-matching. These attention heads allow us to predict seemingly inexplicable reasoning errors in people caused by ostensibly irrelevant prompt details. Taken together, our results suggest that everyday causal reasoning in people and LLMs is more consistent with a form of pattern-matching than with abstract world models.

Reinforcement Learning with Verifiable Rewards (RLVR) is a central technique for improving long-horizon reasoning in Large Language Models (LLMs). However, existing RLVR methods often encourage unnecessarily long reasoning rollouts, which can degrade reasoning coherence and exhaust the available context budget. Existing approaches to long-context organization often depend on external mechanisms to organize rollouts, rather than enabling the model to manage its own reasoning trajectory. To address this limitation, we propose ReSum, a novel RLVR framework that enables LLMs to compress and organize their reasoning trajectories through self-summarization. Our pilot studies show that self-summarization stabilizes generation by lowering token-level entropy, and that introducing a ``summarization'' phrase can substantially mitigate errors propagated from an incorrect rollout prefix. Motivated by these findings, ReSum adopts a summarization-aware adaptive rollout mechanism that contrastively evaluates whether self-summarization benefits the ongoing reasoning process. Specifically, when the model spontaneously triggers self-summarization, ReSum masks the summarization phrase to create a contrastive branch; for non-summarization positions, it instead randomly injects the phrase to create a matched branch. We further design a summarization-aware advantage to enable finer-grained comparison between contrastive rollout trajectories. Extensive experiments show that ReSum improves performance at an average of 4% while reducing rollout length by 18.6%.

Recent diffusion editors perform diverse instruction-based edits while conditioning on the source image at every denoising step. Yet persistent source-image conditioning can limit how fully an edit is executed and how natural the result appears, especially when the target scene diverges substantially from the input. We introduce DuET (Dual Expert Trajectories), a training-free inference method that temporarily relaxes source-image conditioning by transitioning through a text-to-image phase before returning to edit mode, allowing the denoising trajectory to move toward the target distribution while retaining the structural benefits of image-conditioned editing. Without modifying model weights or increasing sampling cost, DuET consistently improves instruction relevance, semantic fidelity, and perceptual quality across diverse models and benchmarks. In some cases, these gains come with a modest reduction in source-image preservation, revealing a predictable trade-off between source preservation and edit fidelity.

Multi-agent debate (MAD) can improve large language model reasoning, but fixed debate pipelines often waste computation and can amplify correlated errors among similar agents. We propose ARMOR-MAD, a training-free heterogeneous MAD framework that treats debate as conditional computation. ARMOR-MAD combines three components: Pre-debate Agreement Routing (PAR) decides whether independently generated Round-0 answers require debate; Early Agreement Stopping Evaluator (EASE) stops debate after convergence; and Semantic Outlier Detection (SOD) down-weights abnormal final answers during aggregation. Across MATH Level 5, GSM8K, MMLU, and MMLU-Pro, ARMOR-MAD consistently improves over fixed-round heterogeneous debate with the same model pool, reaching 65.5%, 96.5%, 90.0%, and 81.5% accuracy, respectively. The results suggest that genuine model heterogeneity and agreement-based control are both important for making MAD more accurate and efficient.

Mental health problems such as anxiety, depression, and suicide remain urgent global challenges, where timely and accurate assessment is critical for effective intervention. Recently, large language models have been explored for mental health assessment. However, existing general-purpose post-training methods do not align with the cognitive processes of human assessment, which may lead to unreliable reasoning outcomes. To bridge this gap, we propose Cognitive Relative Policy Optimization (CRPO), a reinforcement learning framework tailored for the mental health domain. CRPO extends group relative policy optimization by integrating stage-dependent uncertainty modeling into the policy optimization process. Specifically, we introduce a stage-wise entropy regularization mechanism that encourages broad exploration in early reasoning phases and progressively enforces confident decision-making in later stages, mimicking the human cognitive shift from uncertainty to certainty. In addition, inspired by cognitive appraisal theory, we formalize cognitive reasoning stages, thereby guiding theory-grounded interpretable inference. Experiments on 8 mental health datasets show that CRPO achieves an average improvement of 10.4 percentage points in weighted F1-score over the best reinforcement learning baseline. Furthermore, the CRPO-trained model Mental-R1 demonstrates clear advantages compared with existing large language models on reasoning-intensive cases, suggesting that CRPO enhances reasoning capabilities for mental health assessment.

The rapid updates of online news make tracking event developments challenging, highlighting the need for timeline summarization (TLS). Hallucinations, where LLM-generated content deviates from source news, still remain a critical issue in LLM-based TLS and are not well studied in existing works. To bridge this gap, we identify two primary types of hallucinations: unfaithful content during news summarization and information omission in date-event summarization. Then, we propose NTS-CoT, a novel framework that leverages Chain-of-Thought (CoT) reasoning to mitigate hallucinations in TLS. The framework consists of three key modules: i) Element-CoT to capture essential news elements for faithful summarization, ii) Date Selection to combine temporal saliency and event prominence for timestamp selection, and iii) Causal-CoT to infer causal relationships and reduce omissions in date-event summarization. Extensive experiments, including quantitative analysis on three TLS benchmarks and human evaluation, demonstrate that NTS-CoT outperforms state-of-the-art baselines, effectively mitigating hallucinations and improving LLM-based TLS performance. Our source code is available at https://anonymous.4open.science/r/NTS-CoT .

Reinforcement learning has rapidly emerged as a key component in the training of reasoning and coding models, yet it remains poorly understood from a mechanistic perspective. We study how and through what underlying processes capabilities are acquired or enhanced via reinforcement learning post-training. Our analysis, based on controlled math reasoning experiments with Qwen-2.5-1.5B, reveals two core mechanisms: strategy selection and strategy improvement. Our results highlight the role of SFT data and reinforcement learning data in activating these mechanisms, in particular showing how supervising the model on diverse reasoning strategies can enable strategy selection and how increasing difficulty in reinforcement learning data can enable strategy improvement. Taken together, our results provide mechanistic insight into RL training and suggest practical interventions to continue scaling reasoning capabilities.

Latent chain-of-thought compresses reasoning by replacing visible reasoning traces with continuous hidden-state recurrence, but existing formulations are difficult to optimize with standard on-policy reinforcement learning (RL) and hard to interpret causally. Our key insight is that a single pair of explicit boundary tokens can address both issues at once: discrete entry and exit anchors make the latent block compatible with standard on-policy RL, and the same anchors offer a natural foothold for mechanistic analysis. Motivated by this, we propose SWITCH, a switchable latent reasoning framework. The model emits <swi> to enter latent mode and </swi> to exit. Because the boundaries are ordinary discrete tokens, the GRPO policy ratio is well-defined at every decision point. The same anchors also expose the latent steps to direct probing and causal intervention. We train the model with a visible-to-latent curriculum and a Switch-GRPO objective that propagates gradients through recurrent latent computation. SWITCH consistently outperforms prior hidden-state-recurrence latent reasoning approaches at similar scale. Mechanistic analysis through the boundary tokens further reveals three findings: (i) <swi> is a sharply localised, learned switching policy rather than a stylistic artefact; (ii) the latent step it opens performs problem-specific, causally important computation rather than acting as an inert placeholder; and (iii) that computation is concentrated at a single hidden-state transition on entry. Together, these results show that hidden-state-recurrence latent reasoning is both RL-trainable and open to direct mechanistic analysis, including of how on-policy RL itself improves the model from the inside.

Reasoning-driven universal multimodal embedding has advanced rapidly by introducing Chain-of-Thought (CoT) reasoning into the embedding pipeline. Despite the strong performance across both general and complex tasks, this paradigm suffers from two core limitations: (i) autoregressive CoT reasoning incurs high computational cost, making it impractical for low-latency retrieval; and (ii) embedding performance is heavily coupled with CoT annotation quality, making large-scale training unreliable. These raise fundamental questions: Is textual CoT the optimal form of reasoning for embedding, and can effective embedding reasoning be accomplished in latent space? To this end, we propose LaME (Latent Reasoning Multimodal Embedding), which formulates embedding-oriented latent reasoning as a weakly supervised information bottleneck. LaME employs K learnable reason tokens as a fixed-capacity bottleneck, completing all reasoning within a single forward pass. The two weak supervision signals structurally decouple contrastive from autoregressive objectives and eliminate dependence on CoT annotations, while a two-stage training pipeline ensures stable convergence. Experiments on MMEB-v2 and MRMR show that LaME achieves competitive performance, surpassing some explicit CoT-based models, while delivering 60x faster inference than explicit CoT methods and 2x faster than latent baselines with throughput comparable to discriminative embedding models. Code is available at https://github.com/PeppaWu/LaME.

Three paradigmatic forms of inference recur across scientific reasoning: deduction, induction, and causal abduction. Reliably evaluating LLMs on these in scientific settings is currently out of reach: scientific benchmarks built on human annotations are costly and lack mechanistic ground truth, while synthetic logical-reasoning benchmarks do not resemble real scientific documents. We introduce SciR, a benchmark that combines multi-paradigm reasoning with controllable scientific rendering, anchored on three paradigmatic scientific problems. Tasks are generated from formal objects (deduction tree, inductive rule hypothesis, causal graph) to guarantee verifiable answers, then rendered into multi-document scientific discourse via per-track domain-tuned genres. The construction lets us independently vary two difficulty axes: how hard it is to extract the key information needed for inference, and how hard the principled inference itself is. We test six models. Both axes hurt every model, and their effects compound. The rendering even hurts neurosymbolic pipelines, which hand inference to a verified solver. The two axes yield a per-model extraction-vs-inference profile: for instance, reasoning models like deepseek-r1 mostly surpass non-reasoning instruct models on the inference axis. To our knowledge, SciR is the first multi-paradigm scientific-reasoning benchmark with parametric control on both extraction and inference difficulty.

Prevailing wisdom posits that Multi-Agent Systems (MAS) are superior to Single-Agent Systems (SAS), citing advantages like context protection, parallel processing and distributed decision-making. However, empirical support for this claim relies primarily on comparisons with SAS baselines using benchmarks that prioritize isolated reasoning tasks, which do not adequately assess these advantages. Focusing on automatically generated MAS that are designed for enhanced generalizability over manually-designed counterparts, we perform a rigorous, systematic evaluation against SAS, specifically Chain-of-Thought with Self-Consistency (CoT-SC). Across traditional reasoning datasets and tasks with interactive multi-step workflows (e.g., BrowseComp-Plus), we demonstrate that automatic MAS consistently underperform CoT-SC despite being up to 10x more expensive. To isolate these failures from limitations inherent to task structure, we introduce a diagnostic synthetic dataset tailored for MAS featuring explicit task decomposition, context separation and parallelization potential. We show that expert-architected MAS consistently outperforms automatically generated architectures in both raw performance and cost-efficiency on this dataset, demonstrating that existing evaluation frameworks mask critical architectural gaps and inefficiencies of complex MAS by failing to account for the marginal utility of increased computational cost. Critically, systematic deconstruction of the generated MAS architectures reveals that current automated design paradigms produce architectural bloat that prioritizes superficial complexity which does not translate into functional utility, exposing a fundamental misalignment with multi-agent principles.

Parallel test-time scaling samples many reasoning traces and majority-votes their answers, improving LLM accuracy but requiring traces to run to completion, incurring substantial computational overhead. We observe that probing partial traces at intermediate checkpoints can extract current answers without disrupting generation, revealing an evolving aggregate vote. Based on this observation, we introduce MARS, a margin-adversarial stopping rule that estimates which active traces are likely to change their answers and stops once the leader remains safe under a conservative bound on future vote movement. The rule separates two sources of uncertainty. It learns the trace-level switch probabilities that determine how much of the current margin is likely to be retained, while handling the harder question of where switching traces land through an adversarial bound calibrated from warmup traces. With true switch probabilities, MARS guarantees with high probability that the early-stopped answer matches the full-budget vote. In practice, a five-feature logistic model closely matches oracle switching behavior. Across three reasoning models and three competition-math benchmarks, MARS saves 25-47% of self-consistency tokens and 14-29% on top of DeepConf Online, a strong confidence-weighted baseline that already filters and truncates weak traces, while matching the accuracy of the corresponding full-budget baselines.

Evaluating procedural reasoning in AI-supported learning systems requires question-answer datasets that are both learner-like and grounded in the instructional knowledge the system is expected to use. We study how TMK-based question generation strategies affect dataset quality for procedural and multi-hop reasoning. We compare three strategies: strict generation from Task-Method-Knowledge (TMK) models, transcript-first generation with post-hoc TMK filtering, and TMK-aware generation that combines transcripts with structured guidance. To evaluate generated items, we introduce a grounding validation framework based on closed-set evidence units extracted from TMK models. The framework measures whether answers are supported by the underlying representation, whether questions are self-contained, and whether they target multi-hop procedural reasoning. Across 23 instructional topics and 690 generated question-answer pairs, strict TMK generation achieves the strongest overall quality, with 96.5% grounded questions and 92.6% usable questions. Transcript-first generation produces more learner-like questions but more context-dependent or weakly grounded items, while TMK-aware generation yields high raw multi-hop coverage but lower grounding. These results show that procedural richness and natural phrasing do not guarantee representational grounding, motivating explicit representation-aware validation for evaluation datasets in AI-supported learning.

Vision-language-action (VLA) models generate chain-of-thought (CoT) reasoning alongside driving trajectories, but existing benchmarks evaluate only trajectory quality and do not assess whether the CoT is relevant, consistent, or causally connected to the driving action. We introduce VLADriveBench, a framework that combines observational metrics (mentioning, hallucination, contradiction, action alignment) with a CoT intervention protocol to provide complementary views of the CoT-action relationship. Applying VLADriveBench to three models across two architectures, we find that the two analyses can diverge sharply: ORION scores highest on observational alignment yet its CoT is epiphenomenal, while Alpamayo v1.5 scores lower yet its CoT is strongly causal, with visual salience gating the extent of CoT influence.

Latent reasoning models (LRMs) replace explicit chain-of-thought with continuous thoughts. Recent work treats observable latent-state patterns, such as BFS-like frontiers and decodable arithmetic computation, as evidence for internal reasoning mechanisms. Evaluating two LRMs (Coconut and CODI) against controls lacking the proposed recurrence or curriculum, we find these patterns also appear in the controls and do not always causally affect behavior. Causal interventions reveal that latent-thought utilization is not binary but graded, scaling with a thought's causal effect on model behavior. Geometric analyses reveal this effect concentrates in low-rank directions whose step-to-step geometry grows more structured as their behavioral influence increases. Latent thoughts should therefore be treated as hidden computation, not hidden explanation: decodability, attention, or static structure alone cannot establish mechanism. LRM interpretability thus requires matched controls and causal tests.

Robust lie detectors for language models could enable powerful techniques for auditing, monitoring, and post-hoc investigation of model behaviour, but evaluating them requires testbeds where models verifiably believe the opposite of what they say. We show that existing trained model organisms often fail this requirement, leaving prior positive and negative detection results difficult to interpret. We address this with 13 reasoning model organisms whose hidden beliefs are verified in chain-of-thought and shown to generalise to held-out tasks, alongside Varied Deception, a prompted-lying testbed covering a broad range of lie-inducing motivations. On these testbeds we evaluate four detectors: a chain-of-thought judge, a logprob classifier, and two activation probes, including Did-You-Lie (DYL), a new method for training follow-up probes. On prompted lying, across 31 open-weight models spanning 2B to 1T parameters, all four detectors show positive scaling with model capability. However, every activation- and logprob-based detector drops sharply on our trained model organisms, with DYL retaining the most signal; only the chain-of-thought judge remains strong, with 0.82 balanced accuracy, partly as an artefact of our verification process favouring CoT-readable beliefs. Current lie detectors therefore cannot support high-confidence claims about model beliefs, and we suggest research directions that may address some of their current limitations. We release our datasets, model organisms, and trained detectors.

Conversational shopping assistants now serve hundreds of millions of customers, yet no existing benchmark jointly evaluates the open-ended multi-turn reasoning, domain expertise, and criterion-level quality that real shopping conversations demand. Shopping reasoning is unique among language model applications. Unlike factual question answering or verifiable code generation, it requires balancing subjective preferences, budget constraints, and cross-product trade-offs across multi-turn dialogue, capabilities absent from previous e-commerce and general-purpose benchmarks. We introduce the Shopping Reasoning Bench, an expert-authored benchmark of 525 missions (232 single-turn, 293 multi-turn) with 10863 importance-weighted binary rubrics authored by retail domain experts. These criteria are organized under a taxonomy of five reasoning categories and fifteen subcategories covering diverse demands such as preference refinement, trade-off analysis, and compatibility assessment. An evaluation of nine models across three families (GPT, Claude, Gemini) shows that pass rates reach only 57--77% overall. On multi-turn missions, all models score 13--29 points lower on optional above-and-beyond criteria than on required ones, and performance degrades 4--18 points as conversations progress. These gaps show that current models handle basic shopping assistance but fall short of expert-level advice, making Shopping Reasoning Bench a challenging testbed for future shopping assistant development.

Self-consistency improves LLM reasoning by sampling multiple outputs and selecting the most consistent answer, but existing formulations largely rely on exact matching and therefore remain limited to tasks with categorical outputs. In this work, we study self-consistency in open-ended generation tasks such as code synthesis and text summarization. We hypothesize that consistency can be understood as a geometric property of the generation space, where semantically compatible generations concentrate in similar regions of representation space. To study this hypothesis, we introduce Embedding-Based Agreement (EBA), a simple training-free operationalization that estimates agreement by clustering sampled generations in embedding space. Through experiments on mathematical reasoning, code generation, and summarization, we show that agreement in representation space provides a robust and scalable signal of self-consistency for open-ended tasks. In particular, EBA consistently outperforms random selection and exhibits more stable scaling behavior than recent selection approaches based on LLM evaluation or uncertainty estimation. We further show that these agreement signals remain stable across model families and embedding spaces, even with native hidden representations. Finally, our analysis shows that the geometric location occupied by sampled generations is strongly correlated with generation quality: generations concentrated near central regions of representation space tend to correspond to more reliable outputs, whereas peripheral generations are substantially less accurate. Overall, our findings support viewing self-consistency as a property of the geometric organization of sampled generations rather than exact symbolic overlap.

Trusted monitoring is a cornerstone of AI control. However, as frontier models grow more capable, the increasing capabilities gap between trusted and untrusted models may render trusted models unreliable monitors. We introduce \emph{bootstrapped monitoring}, a protocol that addresses this by inserting a stronger, intermediate untrusted model with transparent chain-of-thought reasoning into the oversight chain. The untrusted monitor ($U_m$) evaluates the agent's actions, while a weaker trusted model ($T$) oversees $U_m$'s reasoning to detect collusion. We evaluate bootstrapped monitoring on multi-turn software engineering tasks (BashArena) across multiple agents and monitors. Bootstrapped monitoring substantially improves catch rates over trusted-only monitoring, even when the untrusted monitor actively colludes with the agent, provided we have access to its raw chain-of-thought. Our results suggest that bootstrapped monitoring can extend the useful lifetime of trusted models in control as AI capabilities advance.

Hateful videos have become prevalent on online platforms, highlighting an urgent need for effective detection. However, existing studies primarily focus on binary classification and fail to provide contextual rationales that reveal the implicit meanings behind these judgments, significantly undermining model explainability. To fill this gap, we aim to achieve explainable hateful video detection, enabling models to provide contextual rationales that integrate relevant evidence and logical reasoning alongside decisions. This approach can comprehensively enhance the understanding of video content and the explainability of the decision-making process. We first introduce two datasets, Ex-HateMM and Ex-ImpliHateVid, for explainable hateful video detection. Each dataset provides fine-grained annotations of multimodal harmful elements, along with contextual rationales. We then propose an Information Augmentation and Reasoning Enhancement (IARE) framework designed for explainable detection. The framework employs an information augmentation phase that leverages the multimodal chain-of-thought to integrate harmful elements, thereby enriching rationale evidence. Additionally, IARE incorporates a reasoning enhancement phase, in which Direct Preference Optimization guides the model toward correct reasoning paths and away from incorrect ones, thereby improving the logical coherence of its justifications. We conduct extensive experiments on the two datasets, comparing multiple baselines with our proposed IARE framework. The results demonstrate that IARE achieves state-of-the-art performance while also generating accurate rationales.

Current Large Reasoning Models (LRMs) exhibit remarkable general capabilities but significantly underperform in spatial reasoning tasks. Existing approaches treat this gap as a knowledge deficit, relying on supervised fine-tuning (SFT) to ingest labeled spatial data from external vision sources or synthetic engines. In contrast, we argue that for many tasks, spatial reasoning capabilities are already present in pre-trained LRMs but require alignment through logical coherence under geometric 2D and 3D constraints. In this work, we propose a self-supervised reinforcement learning (RL) framework that targets the internal reasoning process without requiring ground-truth annotations. By formalizing the notion of consistency verifiers -- reward functions that check for geometric and semantic consistency under transformations -- we demonstrate that models can improve their spatial reasoning abilities. We use both image transformations, like flipping, and textual transformations, like swapping the order of objects in the question, and propose a new optimal transport-based RL strategy, OT-GRPO, which is a minimal-matching variant of group relative policy optimization tailored to pairwise verifiers. We show that this label-free consistency training approaches the accuracy of models trained with ground-truth supervision and achieves similar generalization across diverse tasks and data domains.

The correspondence between large language models (LLMs) and the neural mechanisms underlying human higher-order cognition remains insufficiently characterized. Given that language and reasoning in the human brain appear dissociable, an open question is whether LLMs align with neural signals from reasoning-related regions and whether such signals can improve them. Here, focusing on deductive reasoning, we show that LLM internal representations are not only partially aligned with task-fMRI activity but can also be directly enhanced by these signals. Using a neural-predictivity metric, we find that LLMs explain a substantial fraction of the explainable variance in reasoning-related regions at the aggregate level, whereas predictivity within specific reasoning types is lower, indicating both alignment and divergence. Building on this, we propose a brain-guided framework: we steer model representations along directions induced by the joint structure of model and brain representations, applying intervention at inference and fine-tuning during training. We demonstrate that task-evoked brain signals can directly enhance LLM reasoning, yielding gains orthogonal to language-only supervision across 10 LLMs (1.5B-72B), with transfer across reasoning types and up to 13% absolute accuracy gain. Our results advance LLM-brain correspondences from correlation to guidance, establishing a brain-signal-driven pathway toward more robust and cognitively aligned AI.

Large language models (LLMs) have demonstrated strong reasoning and instruction-following capabilities, making them potentially powerful tools for time-series analysis. However, time series lie outside their native textual modality, raising a fundamental question: how should time series be represented so that LLMs can reason about them effectively? Existing work typically serializes raw numerical sequences or fine-tunes pre-trained LLMs on time-series data. These approaches place the burden of extracting temporal structure directly on the LLM, creating a modality mismatch that often degrades performance on long sequences and introduces substantial computational overhead. In this work, we introduce Time-Series-to-Structured-Program representation (T2SP), a deterministic, training-free method that represents a time series as a structured symbolic program. T2SP decomposes time series into trends, periods, and salient events, expressing them in a program-friendly format aligned with the textual and code-like modalities on which LLMs are natively trained. By shifting temporal-structure extraction from the model to the representation itself, T2SP enables off-the-shelf LLMs to leverage their existing reasoning capabilities for time-series understanding. We evaluate T2SP on three reasoning tasks -- editing, captioning, and question answering -- where it consistently improves performance, reduces reasoning time, and lowers failure rates compared with raw-string representations. Our results demonstrate that T2SP provides an effective interface between time series and LLMs.

Touch is a key modality for embodied agents to understand the physical world. Although recent work has incorporated tactile signals into language systems for tactile commonsense reasoning, scaling such systems to realistic open-world settings remains challenging due to two key bottlenecks: (1) current tactile reasoning datasets remain limited in format and scale, providing insufficient supervision for reasoning from tactile observations to physical commonsense and hindering the learning of transferable tactile commonsense; (2) Tactile signals are inherently redundant and action-specific, yet existing methods often overlook these properties, resulting in inefficient representations with limited semantic expressiveness. To address these limitations, we propose TouchThinker, a tactile-language framework that scales tactile commonsense reasoning to the open world from both data and representation perspectives. First, we construct TouchThinker-1M, a million-scale, multi-source tactile reasoning dataset covering \textbf{415} objects, \textbf{8} scenarios, and \textbf{7} sensor types, providing a solid data foundation for open-world generalization. We further introduce TouchThinker-Bench, an open-world benchmark with more realistic and diverse tasks. Then, we propose action-aware modeling mechanism to improve tactile representation efficiency and enable efficient reasoning. Experimental results demonstrate that TouchThinker achieves competitive performance against state-of-the-art models across multiple datasets. Our code and dataset will be made available at: https://github.com/lvkailin0118/TouchThinker.

Large Language Models (LLMs) have achieved strong performance across natural language processing tasks, yet reliable reasoning remains an open challenge. Although modern LLMs show progress in structured inference, multi-step problem solving, and contextual understanding, their reasoning behavior is often inconsistent and sensitive to prompting strategies, task design, and model scale. This survey provides a systematic analysis of more than 300 recent papers from arXiv, Semantic Scholar, Google Scholar, Papers with Code, and the ACL Anthology to examine how reasoning capabilities emerge in LLMs and where they fail. We make three main contributions. First, we introduce a structured taxonomy of LLM reasoning research, covering Chain-of-Thought reasoning, multi-hop reasoning, mathematical reasoning, common sense reasoning, visual and temporal reasoning, code and algorithmic reasoning, retrieval-augmented reasoning, tool-augmented and agentic reasoning, and reinforcement learning-based reasoning. Second, we analyze methodological trends across these paradigms, including prompting methods, model architectures, training objectives, reward modeling, and evaluation benchmarks. Third, we synthesize recurring limitations and failure modes, such as reasoning hallucinations, brittle multi-step inference, weak causal abstraction, and poor cross-domain generalization. By organizing a rapidly expanding literature, this survey offers a unified view of the current capabilities and limitations of reasoning in LLMs. We also identify emerging research directions, including meta-reasoning, self-evolving reasoning frameworks, multimodal reasoning, and socially grounded reasoning. Overall, this work aims to serve as a reference for developing more robust, interpretable, and generalizable reasoning systems in future language models.

Trust in an AI system is often anchored by explanations of how it works, which one then uses to forecast its behavior on new inputs. For large reasoning models (LRMs), this conventional route is particularly difficult to follow: explanation methods for single token generations do not naturally generalize to long trajectories, and the trajectories themselves are often not faithful when read as natural language. We propose an alternative that bypasses the explanation step: treat behavior forecasting as a learnable task and train Behavior Forecasters that operates on a single reasoning trajectory to make the same forecasts one would typically seek from an explanation. The forecaster's training data is obtained by querying the LRM with no human annotation, and its inference is done in a single forward pass. We instantiate this approach on two tasks: how likely the LRM is to repeat its answer on re-runs, and how removing parts of the input changes its answer. We evaluate this approach on both tasks across three diverse reasoning datasets and find that trained Behavior Forecasters are more accurate than GPT-5.4 and Claude Opus-4.6 reading the same trajectories as naive readers, at a small fraction of their inference cost. We find that fine-tuning the backbone end-to-end and initializing it from the target LRM are each necessary for strong performance. These results show that the reasoning trajectory carries information about the LRM's future behavior that goes beyond what naive reading conveys.

Deployed large reasoning models (LRMs) often behave unexpectedly. Test-time steering controls LRM outputs by intervening on their hidden representations, but it can degrade output quality. We argue that prior steering work implicitly relies on internal features that detect behavior in already generated text. We show that these detection features are poor predictors of future behavioral outcomes, and thus not the natural intervention target. Instead, we train activation probes to predict future behavior likelihoods from intermediate reasoning steps. These probes predict the most likely behavior with 64%-91% accuracy, revealing a separate type of internal prediction features. Building on these prediction features, we introduce a text-level steering method, Future Probe Controlled Generation. FPCG samples multiple candidate sentences and chooses the best one according to a probe predicting the future behavior likelihood. This enables steering with almost no output quality degradation. FPCG also enables steering in several evaluations where activation steering fails. These results show that distinguishing detection and prediction features enables a more nuanced approach to controlling LRM behaviors.

Long chain-of-thought (CoT) trajectories in large language model (LLM) reasoning cause severe inference bottlenecks due to rapid key-value (KV) cache growth. Current decoding-time compression methods mitigate this issue via token eviction, but typically assume a uniform budget distribution across all layers and heads. In contrast, existing non-uniform budget allocation methods are predominantly designed for the static prompt prefill phase, and they do not capture the stepwise context demands of autoregressive reasoning. To bridge this gap, we propose ReasonAlloc, a training-free framework that recasts decoding-time KV compression as a hierarchical budget allocation problem. ReasonAlloc operates at two complementary levels: an offline layer-wise preallocation strategy captures an architecture-driven demand pattern which we call ``\textit{Reasoning Wave}'', while an online head-wise strategy reallocates resources during decoding to information-rich heads based on real-time utility. Evaluations on mathematical reasoning benchmarks (MATH-500, AIME~2024) using DeepSeek-R1-Distill-Llama-8B, DeepSeek-R1-Distill-Qwen-14B, and AceReason-14B show that ReasonAlloc outperforms uniform-budget R-KV, SnapKV, and Pyramid-RKV (a baseline enforcing a static, monotonically decreasing layer budget), with the largest gains at small budgets (128-512 tokens). ReasonAlloc is plug-and-play with existing token-eviction policies and introduces negligible inference-time overhead.

Chain-of-thought (CoT) supervised fine-tuning (SFT) is widely adopted to improve reasoning ability, yet we find that it systematically degrades long-context recall in hybrid linear-attention models. Across architectures including HypeNet and Jet-Nemotron, retrieval performance on Needle-In-A-Haystack (NIAH) deteriorates substantially after CoT-SFT, and the degradation becomes more severe under harder retrieval settings and longer context windows. For example, HypeNet-9B on NIAH-S2@256K decreases from $67.2\%$ to $9.4\%$. We attribute this to CoT-SFT biasing attention gradients toward short-range patterns, disrupting query-key projections ($W_Q, W_K$) that are responsible for long-range routing. Motivated by this observation, we propose QK-Restore, a training-free method that restores only $W_Q$ and $W_K$ from the pre-SFT checkpoint while preserving all other post-SFT parameters. We further introduce a Procrustes variant to balance routing preservation and reasoning adaptation. Across architectures, QK-Restore consistently restores long-context capability at zero training cost while preserving reasoning performance; for instance, on HypeNet-5B it improves S3@256K from $65.4\%$ to $76.4\%$ while maintaining strong reasoning performance.

Instruction-tuned LLMs are increasingly converted into reasoning models through post-training to improve multi-step task performance. This conversion is usually optimized for reasoning accuracy, without explicitly preserving the alignment behavior of the instruction-tuned model, such as safe refusal, bias avoidance, and privacy protection. We ask: does this conversion preserve alignment? We study this question through a trustworthiness audit and find that it is not behavior-preserving by default. For a systematic analysis, we compare reasoning models produced via supervised fine-tuning, RL-based post-training, and distillation against matched instruction-tuned baselines across six trustworthiness dimensions: safety, toxicity, stereotyping and bias, machine ethics, privacy, and out-of-distribution robustness. We observe that reasoning models often improve on reasoning benchmarks but exhibit alignment regressions, including increased toxicity, amplified stereotyping, miscalibrated refusal, and contextual privacy leakage. These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence. Overall, our results point to the broader conclusion that trustworthiness metrics are essential for evaluating reasoning models and should be reported alongside gains in reasoning capability.

AI compilers aggressively rewrite computation graphs through normalization, lowering, and optimization, making it difficult to track the provenance of tensors and operators across compilation. Reliable provenance is essential for attaching platform-specific postprocessing, debugging compiler behavior, and validating transformations, yet existing solutions are either invasive or ad hoc under non-injective graph rewrites. We present a lightweight, generative approach to provenance tracking based on observational semantics. Instead of propagating identifiers through compiler passes, we observe graph transformations and reason about provenance in terms of observable computational actions. We formalize this approach using a coalgebraic model and bisimulation, which preserves provenance even when intermediate nodes are eliminated. Furthermore, we implement this approach in a prototype AI compiler COVAN, demonstrating stable provenance across compilation pipelines with minimal engineering overhead.

Large Language Models (LLMs) struggle to rigorously verify complex mathematical proofs. Standard global evaluation approaches suffer from "context poisoning," in which superficially plausible statements mask subtle logical flaws, leading to hallucination or over-skepticism. To address this, we shift from global evaluation to strict step-level verification: our framework maintains detailed context for each deduction step and strictly constrains the sources of applied theorems. We evaluate on a carefully curated adversarial diagnostic suite of research-level proofs drawn from the FirstProof challenge. A systematic ablation study demonstrates that these deductive constraints are indispensable, as unconstrained global prompting consistently fails to localize subtle logical errors. Beyond outperforming global evaluation, our approach fundamentally alters the failure taxonomy. Error analysis reveals that, rather than exhibiting severe logical hallucinations, remaining rejections are primarily instances of "pedantic hyper-rigor" stemming from unstated domain conventions, effectively exposing implicit ambiguities within the expert benchmark itself. Our findings suggest that prompting agents to organize their verification notes in a cautious, human-mathematician-like manner can substantially improve their ability to distinguish rigorous proofs from flawed ones, with the potential to strengthen agentic reasoning on frontier mathematical concepts that the base model does not already know well, and to lay a theoretical foundation for future automated proof-review systems. Code and prompts are available at GitHub.

Automatic Depression Detection (ADD) from clinical interviews is a pivotal task in computational mental health, yet it remains challenging due to two critical obstacles: 1) difficulty in modeling complex but sparsely distributed depression clues within lengthy, multi-topic clinical interviews, leading to superficial and unreliable reasoning; 2) scarcity of labeled data due to clinical privacy, together with high cost of training and fine-tuning, limiting the deployment of supervised ADD systems. To jointly address these challenges, we propose Dep-LLM, a training-free framework that mirrors the step-by-step reasoning of clinical psychiatrists and operates entirely on frozen off-the-shelf foundation LLMs. Dep-LLM comprises three stages. First, a Chain-of-Thought (CoT) Depression Multi-factor Analysis module structurally decomposes the long dialogue into five clinically aligned themes and produces evidence-grounded rationales, effectively handling long-context dependencies. Second, we introduce Confidence Analysis and Modulation module that quantifies the epistemic reliability from token-level entropy of each rationale and applies an intra-label and inter-theme modulation that amplifies trustworthy signals while suppressing uncertain ones without extra training. Third, a Collaborative Multi-factor Prediction module dynamically integrates multi-factor signals weighted by confidence into the final diagnosis. Extensive experiments on the DAIC-WOZ and E-DAIC datasets demonstrate the effectiveness and generalizability of Dep-LLM: it surpasses zero-shot baseline on nearly all 21 foundation LLMs across 9 metrics such as accuracy, macro F1 and weighted-average F1, and further outperforms state-of-the-art supervised domain-specific LLMs as well as the latest closed-source commercial LLMs, while requiring no extra training.

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.

Token-level credit assignment remains a key obstacle for reinforcement learning (RL) in large language models (LLMs), where RL recipes typically treat all tokens equally, failing to distinguish decisive reasoning steps from routine formatting or fluent filler. Recent attempts leverage model-internal signals to assign finer-grained credit, but these are often point-wise heuristics that ignore the global structure of information propagation. We propose FlowTracer, an RL framework that traces answer-targeted reasoning flow on an attention-induced directed acyclic graph in which nodes correspond to tokens and edge capacities come from aggregated attention weights and derives token credit from this global structure. The edge capacities are reweighted to retain only the influence that can reach the answer region, while enforcing local flow conservation so intermediate tokens neither lose nor gain effective mass due to path length or irrelevant branches. On this graph, FlowTracer extracts an information-flow backbone connecting the question to the answer and scores tokens by flow throughput, revealing high-impact hubs and aggregation checkpoints that mediate long-range dependencies. These derived importances are used to shape token-level rewards, enabling learning signals to focus precisely on the tokens that route information toward (or away from) correct answers and delivering consistent performance gains across a range of reasoning tasks.

3D Gaussian Splatting (3DGS) is a powerful technique for creating high-fidelity 3D assets. However, the widespread sharing and iterative modification of 3DGS models across digital platforms create pressing challenges for intellectual property protection and forensic traceability. To address this, we propose GaussTrace, a novel framework for constructing directed provenance graphs for 3DGS models. GaussTrace formulates provenance analysis as an evidence-based reasoning problem. It builds upon attribute-wise statistical profiling of 3DGS parameters to capture intrinsic properties. Moreover, we introduce hypothesis-driven editing simulations of common operations to provide auxiliary evidence for plausible transformation pathways. These statistical and simulated cues jointly enable a Large Language Model (LLM) to perform structured Chain-of-Thought (CoT) reasoning, yielding directional provenance inferences and explainable edge reasons. Experimental results demonstrate that GaussTrace effectively constructs evolutionary relationships among diverse 3DGS models, delivering accurate, interpretable, and robust provenance graphs without requiring model training or access to editing histories. Project page: https://haolianghan.github.io/GaussTrace.

Combinatorics is central to Olympiad-level mathematical problem solving, requiring deep discrete reasoning, creative constructions, and rigorous structural insight. Recent evidence suggests that even today's strongest frontier models remain uneven on Olympiad combinatorics, revealing a gap in creative mathematical reasoning. We introduce ComBench, an Olympiad-level combinatorics benchmark for evaluating and diagnosing the combinatorial reasoning capabilities of large language models. ComBench contains 100 human-annotated competition-level problems organized around two complementary settings: analysis-centric problems, which primarily require rigorous mathematical arguments, and construction-centric problems, which require explicit constructions in addition to correctness justifications. The evaluation protocol combines rubric-guided proof grading with deterministic construction verification, exposing cases where proof quality and construction validity diverge. Experiments on frontier open- and closed-source models show that ComBench is far from saturated: the strongest model reaches 65.4% overall Avg. and 75.3% overall Best@4. We further find that Rigorous Proof Reasoning and Constructive Realization are distinct capabilities: Kimi-K2.6 trails GPT-5.5 on analysis-centric proof grading but surpasses it on construction-centric Best@4, while Existence and Construction problems remain consistently hardest across representative frontier models.

Autonomous web navigation remains challenging for LLM agents, and the strongest generalist systems rely on proprietary reasoning models whose inference cost is prohibitive for the repetitive tasks where such agents would be most useful. We argue this gap stems not from insufficient model capability but from agent architectures that fail to replicate three human cognitive advantages: selective attention to relevant page regions, persistent memory of website structure, and procedural fluency with common interaction patterns. We introduce WebChallenger, a web agent framework that addresses each gap through architecture design rather than model scale, built around PageMem: a structured page representation deterministically constructed from the DOM that exposes each page as a hierarchy of semantic sections with short summaries. On this shared substrate we build three mechanisms that mirror the three cognitive advantages: a divide-and-conquer observation pipeline that lets the agent skim section summaries and extract details only from task-relevant regions; a lightweight exploration and memory system that traverses each website once to build a reusable map of pages and element behaviors; and compound action workflows that collapse common multi-step interactions into single agent actions, handling partial state changes automatically. Because all three operate over PageMem, the framework generalizes across websites without site-specific adapters. Using off-the-shelf open-weight models without fine-tuning, our system achieves 56.3% on WebArena, 48.7% on VisualWebArena, 51.0% on Online-Mind2Web, and 70.9% on WorkArena, approaching frontier proprietary systems at a fraction of the cost. Our code is released at https://github.com/jayoohwang1/webchallenger

Math reasoning benchmarks have proliferated, yet most lack a per-item difficulty signal grounded in actual human performance. We introduce KCSAT-ML, a decade (2014-2025) of Korean College Scholastic Ability Test (KCSAT; Suneung) mathematics: 664 problems with a 339-item core set carrying official per-item error rates from nationwide cohorts of hundreds of thousands of examinees. We pair the benchmark with Difficulty-aligned Reasoning Gain (DRG): a score-orthogonal metric that asks whether a model's mistakes concentrate on the items humans found hard, or on items humans found easy. Together they expose, across a wide range of VLMs (and LLMs via OCR), three patterns: (i) low-budget accuracy collapses on the high-human-error tail at every model size; (ii) test-time scaling (TTS) raises token use roughly linearly with cohort error rate, while accuracy gains follow a non-monotonic curve; (iii) within a single family, TTS flips between anti-scaling on the hardest items and overthinking on easier ones -- two faces of the same alignment failure. On DRG, models with near-identical accuracy can sit at near-opposite values: one model gets wrong what humans also find hard, while another solves the hardest items yet fails on items humans find easy -- a contrast that aggregate accuracy hides. Our code and dataset builder will be open-sourced at https://github.com/naver-ai/KCSAT-ML.

Reinforcement learning has become a key paradigm for eliciting reasoning abilities in large language models, where exploration is crucial for discovering effective solution trajectories. Existing exploration methods typically encourage diversity in semantic or gradient spaces, without distinguishing what drives this diversity. A trajectory may appear novel because it follows a new reasoning process, or because it varies memorized patterns and shortcuts. Rewarding both cases equally may steer exploration toward memorization rather than genuine reasoning improvement. In this paper, we propose DiRL, a Direction-Aware Reinforcement Learning framework that anchors exploration to an internal reasoning-memorization direction of the policy. Specifically, DiRL extracts this direction from model representations, constructs direction-weighted gradient features to characterize rollout updates, and shapes rewards to amplify reasoning-aligned exploration while suppressing memorization-aligned variations. DiRL integrates seamlessly into standard Group Relative Policy Optimization (GRPO). Extensive experiments on mathematical and general reasoning benchmarks demonstrate the effectiveness of DiRL, showing significant improvements over various existing exploration methods.