Daily Papers

Video object removal must eliminate not only the target object but also its induced effects while maintaining high-fidelity and spatiotemporally coherent restoration. Existing methods mainly learn object-effect correspondences implicitly from predefined effect categories and fixed data distributions, limiting their generalization to complex real-world scenes involving compositional effects, spatially detached or weakly correlated effects, long-tail physical phenomena, and dynamically evolving interactions. We propose EffectLearner, a semantic-reasoning-enhanced framework that combines a VLM-based Object-Effect Reasoner with a DiT-based Video Eraser. Guided by a structured effect-analysis prompt, the Reasoner performs cross-modal reasoning over a target-highlighted video and extracts compact effect-aware context, which guides the Video Eraser toward comprehensive object-effect removal. Motion-aware mask guidance and motion-consistency supervision further improve removal coverage and spatiotemporal stability under object motion and evolving scene dynamics. To fully exploit the framework in challenging real-world scenarios, we further construct EffectWorld, a paired video dataset specifically designed for complex object-induced effects, and introduce a progressive training curriculum that combines common supervision with complex-effect data. On the standard ROSE-Bench, EffectLearner outperforms existing baselines on most metrics and achieves clear advantages on both EffectWorld-Eval and the challenging EffectWorld-Wild, demonstrating its ability to deliver high-quality video object removal in complex real-world scenes.

Cause-and-effect reasoning in video is a significant challenge for Vision-Language Models (VLMs), as it requires going beyond surface-level perception to a deeper understanding of causal mechanisms. However, existing benchmarks rarely provide the fine-grained, grounded evidence needed to rigorously evaluate this capability. To address this gap, we introduce CaST-Bench, a benchmark for Causal Chain-Grounded Spatio-Temporal Video Reasoning. CaST-Bench presents complex causal questions that require models to identify and localize a chain of multiple spatio-temporal evidences. Through a human-AI collaborative pipeline, we construct a high-quality dataset of 2,066 questions over 1,015 videos, with causal chains annotated by temporal segments and bounding-box tracks. Furthermore, we design a comprehensive evaluation suite with novel metrics that assess not only answer correctness but also the capability for visual evidence grounded reasoning. This grounding is crucial for improving accuracy by mitigating spurious correlations and for enhancing user trust by making models more transparent. Our experiments show that current VLMs struggle with causal questions, largely due to their limited ability to construct precise and grounded causal chains. This highlights an important direction for improving future VLMs.

Existing auto-research benchmarks often entangle multiple sources of improvement, including training frameworks, hyperparameters, compute budgets, and data, making it difficult to attribute why one frontier agent outperforms another to specific research capabilities. In this work, we isolate and systematically evaluate Data Intelligence: an agent's ability to understand, manipulate, and improve the data that shapes model capabilities. We introduce AutoDataBench, a controlled testbed built on a conceptual framework of data intelligence spanning data diagnosis, data organization, and data construction, instantiated through three highly curated optimization tasks while holding non-data factors fixed. Across tool use, retrieval, and knowledge injection, we evaluate frontier LLMs' ability to improve training data through iterative experimentation under task-specific resource budgets. Beyond optimization performance, we ask: do LLMs understand what their data interventions do? We compare predictions made before training with observed outcomes to seek evidence of data-effect reasoning beyond trial and error, and explore whether iterative feedback helps LLMs better understand how changes to training data affect model performance. Finally, we show that reusing AutoDataBench trajectories for mid-training improves downstream coding performance, highlighting its value in both evaluating data intelligence and generating high-quality training data. Code and resources are available at https://github.com/AutoDataBench/AutoDataBench.

Embodied cognition argues that intelligence arises from sensorimotor interaction rather than passive observation. It raises an intriguing question: do modern vision-language models (VLMs), trained largely in a disembodied manner, exhibit signs of embodied cognition? We introduce ENACT, a benchmark that casts evaluation of embodied cognition as world modeling from egocentric interaction in a visual question answering (VQA) format. Framed as a partially observable Markov decision process (POMDP) whose actions are scene graph changes, ENACT comprises two complementary sequence reordering tasks: forward world modeling (reorder shuffled observations given actions) and inverse world modeling (reorder shuffled actions given observations). While conceptually simple, solving these tasks implicitly demands capabilities central to embodied cognition-affordance recognition, action-effect reasoning, embodied awareness, and interactive, long-horizon memory from partially observable egocentric input, while avoiding low-level image synthesis that could confound the evaluation. We provide a scalable pipeline that synthesizes QA pairs from robotics simulation (BEHAVIOR) and evaluates models on 8,972 QA pairs spanning long-horizon home-scale activities. Experiments reveal a performance gap between frontier VLMs and humans that widens with interaction horizon. Models consistently perform better on the inverse task than the forward one and exhibit anthropocentric biases, including a preference for right-handed actions and degradation when camera intrinsics or viewpoints deviate from human vision. Website at https://enact-embodied-cognition.github.io/.

Affective computing has progressed from categorical emotion recognition to open-ended affective analysis with large multimodal models. Yet affective science describes emotion as an unfolding process shaped by appraisal, regulation, and social interpretation, which remains underexplored computationally. We propose TRACE, a cognition-oriented framework that formalizes an affective episode through three interrelated stages: Condition, Affect, and Effect, integrating observable cues with cognitive factors such as internal stance and regulation of emotional display. Based on this formulation, TRACE-Bench evaluates multimodal models in real-world social scenes through five tasks spanning grounded affect recognition, regulation decoding, cause reasoning, effect reasoning, and full-chain reconstruction, with 3,746 structured question-answer pairs over 646 videos. A matched human-model comparison reveals a substantial performance gap, while affect-specialized models also generally lag behind general-purpose MLLMs. Model outputs show recurring failures, including treating displayed behavior as genuine feeling and fabricating unsupported events during long-chain generation. We further propose TRACER, a cognition-grounded structured reasoning method that couples each inference with explicit premises from factual observations, cognitive appraisals, and established upstream conclusions, forming a traceable graph of intermediate and target conclusions. TRACER outperforms all evaluated model baselines on each of the five tasks. Project page: https://cogaffc.github.io/TRACE

With the widespread adoption of large language models (LLMs), hallucinations, which are non-factual fabrications in model outputs, have become serious concerns. Reasoning capabilities have received attention as a self-verification process to improve output reliability. However, the effect of reasoning within a closed system where LLMs cannot rely on external tools or knowledge has yet to be clarified. We therefore conduct experiments under strict constraints (recommending peer-reviewed journal articles in computer science) to examine the effect of reasoning across multiple models (GPT-5.2 and Gemini 3 Flash). Our results reveal a problematic trade-off between constraint compliance and factual accuracy. Non-reasoning models exhibit high constraint violation rates (66-75%) but maintain factual accuracy, while reasoning models reduce violations (13-26%) but systematically distort known facts to satisfy constraints and increase complete fabrication. This trade-off pattern is consistent across both models despite different architectures, indicating a fundamental limitation of reasoning. Furthermore, reasoning does not uniformly improve output authenticity: effects diverge by model, reflecting different allocations of the compliance-truthfulness trade-off. These findings challenge the assumption that reasoning universally improves reliability: reasoning models trade honest constraint violations for detection-resistant distortions.

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%.

Recent advances in Vision Language Models (VLMs) have driven significant progress in visual reasoning. However, open-source VLMs still lag behind proprietary systems, largely due to the lack of high-quality reasoning data. Existing datasets offer limited coverage of challenging domains such as STEM diagrams and visual puzzles, and lack consistent, long-form Chain-of-Thought (CoT) annotations essential for eliciting strong reasoning capabilities. To bridge this gap, we introduce MMFineReason, a large-scale multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring high-quality reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking. The dataset is established via a systematic three-stage pipeline: (1) large-scale data collection and standardization, (2) CoT rationale generation, and (3) comprehensive selection based on reasoning quality and difficulty awareness. The resulting dataset spans STEM problems, visual puzzles, games, and complex diagrams, with each sample annotated with visually grounded reasoning traces. We fine-tune Qwen3-VL-Instruct on MMFineReason to develop MMFineReason-2B/4B/8B versions. Our models establish new state-of-the-art results for their size class. Notably, MMFineReason-4B succesfully surpasses Qwen3-VL-8B-Thinking, and MMFineReason-8B even outperforms Qwen3-VL-30B-A3B-Thinking while approaching Qwen3-VL-32B-Thinking, demonstrating remarkable parameter efficiency. Crucially, we uncover a "less is more" phenomenon via our difficulty-aware filtering strategy: a subset of just 7\% (123K samples) achieves performance comparable to the full dataset. Notably, we reveal a synergistic effect where reasoning-oriented data composition simultaneously boosts general capabilities.

The prevailing paradigm for enhancing the reasoning abilities of LLMs revolves around post-training on high-quality, reasoning-intensive data. While emerging literature suggests that reasoning data is increasingly incorporated also during the mid-training stage-a practice that is relatively more proprietary and less openly characterized-the role of such data in pretraining remains unclear. In particular, due to the opaqueness of pretraining corpora in most frontier models, the effect of reasoning data introduced at different phases of pre- and/or post-training is relatively less reported in the scientific literature. This raises several important questions: Is adding reasoning data earlier during pretraining any better than introducing it during post-training? Could earlier inclusion risk overfitting and harm generalization, or instead establish durable foundations that later fine-tuning cannot recover? We conduct the first systematic study of how reasoning data-varying in scale, diversity, and quality-affects LLM performance when introduced at different stages of training. We find that front-loading reasoning data into pretraining is critical (19% avg gain), establishing foundational capabilities that cannot be fully replicated by later-stage SFT, even with more data. We uncover an asymmetric principle for optimal data allocation: pretraining benefits most from broad diversity in reasoning patterns (11% avg gain), while SFT is more sensitive to data quality (15% avg gain). We show that high-quality pretraining data has latent effects, activated only after SFT, and that naively scaling SFT data can be detrimental, washing away the benefits of early reasoning injection. Our results challenge the conventional separation of language modeling and reasoning, providing a principled guide for strategically allocating data across the entire training pipeline to build more capable models.

Reinforcement Learning from Verifiable Rewards (RLVR) on chain-of-thought reasoning has become a standard part of language model post-training recipes. A common assumption is that the reasoning chains trained through RLVR reliably represent how a model gets to its answer. In this paper, we develop two metrics for critically examining this assumption: Causal Importance of Reasoning (CIR), which measures the cumulative effect of reasoning tokens on the final answer, and Sufficiency of Reasoning (SR), which measures whether a verifier can arrive at an unambiguous answer based on the reasoning alone. Through experiments with the Qwen2.5 model series and ReasoningGym tasks, we find that: (1) while RLVR does improve task accuracy, it does not reliably improve CIR or SR, calling the role of reasoning in model performance into question; (2) a small amount of SFT before RLVR can be a remedy for low CIR and SR; and (3) CIR and SR can be improved even without SFT by applying auxiliary CIR/SR rewards on top of the outcome-based reward. This joint reward matches the accuracy of RLVR while also leading to causally important and sufficient reasoning. These results show that RLVR does not always lead models to rely on reasoning in the way that is commonly thought, but this issue can be remedied with simple modifications to the post-training procedure.

An enterprise agent that answers under a written policy must choose an action before it says anything: answer, ask for a missing fact, ask for confirmation, call one or several tools, escalate, or refuse. Holding one 27B Korean enterprise model fixed, we vary whether the decision is a generated JSON object or a single constrained token, and whether the model may reason first. On 2800 in-distribution items we did not detect an accuracy difference between the two deployed paths (93.46% vs 94.14%), while they differ about tenfold in restart stability (3.00% of actions flip vs 0.29%); they also differ in reasoning, so this is an operational comparison. On a sealed benchmark of 300 items whose labels come from a hidden executable policy, reasoning is worth +17.3 points [+12.3, +23.0] on exact action. With reasoning off, the format difference was 0.0 points [-4.3, +4.0], a failure to detect an advantage for either interface, not equivalence; with reasoning on it is unresolved. Against a token-matched random control, boundary-targeted supervision moves the two covered boundaries by +17.5 and +20.0 points and leaves a held-out boundary unchanged. In a 3-seed training x reasoning factorial, the sealed-rendering interaction is -4.9 [-8.1, -1.7], consistent with partial substitution. All supervision was Korean. On independently rendered English versions, the model is 0.82 points lower [-2.68, +0.93] with reasoning on, the same order as re-rendering in Korean; the reasoning-off gain from targeted supervision is not stable in either language, so we do not claim transfer. The original text stays private as the confirmatory set. An open-model re-rendering made after these results reproduced the reasoning effect and the English verdict; the interaction kept its sign but did not replicate under item-clustered intervals. We release the model and that re-rendering with gold labels.

Reasoning models are evaluated on single-turn benchmarks but deployed in multi-turn dialogue, where users push back on correct answers. Under sustained adversarial pressure we find a previously undocumented failure mode: the chain-of-thought stays factually correct from first turn to last while the emitted answer flips wrong. We call this unfaithful capitulation (UC) and isolate it with a 2times 2 latent-versus-behavioral framework that flip-rate metrics and single-turn faithfulness probes both miss. Across three datasets (MT-Consistency, MMLU-Pro, GSM8K), the latent-correct rate at the behavioral flip clusters near 50% in think mode and collapses to 11-15% under no_think -- paired, within-model causal evidence that reasoning creates the gap. Across models the effect tracks the reasoning channel (high in Qwen3-32B and GPT-OSS-20B, low in inline-CoT Gemma-4-31B-it). An independent GPT-4o judge corroborates 86% of UC labels; a token-level probe shows the answer-slot argmax is correct in 84% of UC cells; and a naive trace-anchored defense backfires. We release all trajectories, traces, and judge labels.

Large Language Models (LLMs) perform well on reasoning benchmarks, but it remains unclear whether this reflects genuine contextual reasoning or reliance on facts memorized in their parameters. We investigate this by distinguishing two possibilities: a broad memorization bias, where familiar content improves reasoning performance, and the Strong Parametric Shortcut Hypothesis, where models skip reasoning entirely and recall stored answers. To test these effects, we introduce MemoReason, a human-curated benchmark that pairs factual reasoning tasks with structurally identical versions where real entities like people, companies, or dates are systematically replaced by ones of the same type. This preserves task structure and specified reasoning operations while varying the familiarity of the context, allowing controlled measurement of how the parametric memory affects reasoning. Our evaluation of recent LLMs reveals consistent and statistically significant performance drops of up to 15.7\% in the fictitious setting, demonstrating a clear memorization bias. However, a targeted analysis of questions failed in the fictitious setting shows that models rarely respond with the corresponding factual answer, indicating that direct parametric shortcuts are not the dominant failure mode. These findings suggest that parametric memory influences reasoning through mechanisms more complex than simple factual recall. MemoReason provides a controlled framework for studying these mechanisms and for extending paired factual-fictitious{} evaluation to broader reasoning settings.

Visual effects (VFX) are essential for enhancing the expressiveness and creativity of video content, yet producing high-quality effects typically requires expert knowledge and costly production pipelines. Existing AIGC systems face significant challenges in VFX generation due to the scarcity of effect-specific data and the inherent difficulty of modeling supernatural or stylized effects. Moreover, these approaches often require per-effect fine-tuning, which severely limits their scalability and generalization to novel VFX. In this work, we present EffectMaker, a unified reasoning-generation framework that enables reference-based VFX customization. EffectMaker employs a multimodal large language model to interpret high-level effect semantics and reason about how they should adapt to a target subject, while a diffusion transformer leverages in-context learning to capture fine-grained visual cues from reference videos. These two components form a semantic-visual dual-path guidance mechanism that enables accurate, controllable, and effect-consistent synthesis without per-effect fine-tuning. Furthermore, we construct EffectData, the largest high-quality synthetic dataset containing 130k videos across 3k VFX categories, to improve generalization and scalability. Experiments show that EffectMaker achieves superior visual quality and effect consistency over state-of-the-art baselines, offering a scalable and flexible paradigm for customized VFX generation. Project page: https://effectmaker.github.io

Recent advances in large language models (LLMs) demonstrate that their capabilities are comparable, or even superior, to humans in many tasks in natural language processing. Despite this progress, LLMs are still inadequate at social-cognitive reasoning, which humans are naturally good at. Drawing inspiration from psychological research on the links between certain personality traits and Theory-of-Mind (ToM) reasoning, and from prompt engineering research on the hyper-sensitivity of prompts in affecting LLMs capabilities, this study investigates how inducing personalities in LLMs using prompts affects their ToM reasoning capabilities. Our findings show that certain induced personalities can significantly affect the LLMs' reasoning capabilities in three different ToM tasks. In particular, traits from the Dark Triad have a larger variable effect on LLMs like GPT-3.5, Llama 2, and Mistral across the different ToM tasks. We find that LLMs that exhibit a higher variance across personality prompts in ToM also tends to be more controllable in personality tests: personality traits in LLMs like GPT-3.5, Llama 2 and Mistral can be controllably adjusted through our personality prompts. In today's landscape where role-play is a common strategy when using LLMs, our research highlights the need for caution, as models that adopt specific personas with personalities potentially also alter their reasoning abilities in an unexpected manner.

The Identifiable Victim Effect (IVE) - the tendency to allocate greater resources to a specific, narratively described victim than to a statistically characterized group facing equivalent hardship - is one of the most robust findings in moral psychology and behavioural economics. As large language models (LLMs) assume consequential roles in humanitarian triage, automated grant evaluation, and content moderation, a critical question arises: do these systems inherit the affective irrationalities present in human moral reasoning? We present the first systematic, large-scale empirical investigation of the IVE in LLMs, comprising N=51,955 validated API trials across 16 frontier models spanning nine organizational lineages (Google, Anthropic, OpenAI, Meta, DeepSeek, xAI, Alibaba, IBM, and Moonshot). Using a suite of ten experiments - porting and extending canonical paradigms from Small et al. (2007) and Kogut and Ritov (2005) - we find that the IVE is prevalent but strongly modulated by alignment training. Instruction-tuned models exhibit extreme IVE (Cohen's d up to 1.56), while reasoning-specialized models invert the effect (down to d=-0.85). The pooled effect (d=0.223, p=2e-6) is approximately twice the single-victim human meta-analytic baseline (dapprox0.10) reported by Lee and Feeley (2016) - and likely exceeds the overall human pooled effect by a larger margin, given that the group-victim human effect is near zero. Standard Chain-of-Thought (CoT) prompting - contrary to its role as a deliberative corrective - nearly triples the IVE effect size (from d=0.15 to d=0.41), while only utilitarian CoT reliably eliminates it. We further document psychophysical numbing, perfect quantity neglect, and marginal in-group/out-group cultural bias, with implications for AI deployment in humanitarian and ethical decision-making contexts.

Longitudinal single cell atlases now capture matched pre treatment and post treatment states from responders and non responders, presenting an opportunity to mechanistically explain why two patients on the same drug diverge. We introduce scDEFT (single cell Drug EFfect Transducer), which treats a drug as a conditioning operator on cell representations, enabling prediction and explanation. In scDEFT, feature wise linear modulation produces drug conditioned cell latents, learned under abundant per cell supervision and then frozen. Two independent heads aggregate those latents over shared transcriptional neighborhoods to predict drug induced state change and responder status. A backward stage ranks the latent dimensions by how strongly they separate responders from non responders and maps them to genes under a cell composition control. On a harmonized inflammatory bowel disease atlas of 1.16 million cells, three cohorts and two drug classes, scDEFT predicts state change at 45% of the baseline to reproducibility ceiling headroom and stratifies responders before treatment at AUROC 0.70, where standard predictors remain at chance. These predictions and the drivers behind them support target and co target nomination, patient stratification, and counterfactual prediction of unseen drug cohort effects.

Large Language Models (LLMs) achieve strong performance on natural language tasks but remain unreliable in mathematical reasoning, frequently generating fluent yet logically inconsistent solutions. We present NeuroProlog, a neurosymbolic framework that ensures verifiable reasoning by compiling math word problems into executable Prolog programs with formal verification guarantees. We propose a multi-task Cocktail training strategy that jointly optimizes three synergistic objectives in a unified symbolic representation space: (i) mathematical formula-to-rule translation (KB), (ii) natural language-to-program synthesis (SOLVE), and (iii) program-answer alignment. This joint supervision enables positive transfer, where symbolic grounding in formula translation directly improves compositional reasoning capabilities. At inference, we introduce an execution-guided decoding pipeline with fine-grained error taxonomy that enables iterative program repair and quantifies model self-debugging capacity. Comprehensive evaluation on GSM8K across four model scales (3B--32B parameters) demonstrates consistent improvements: cocktail training achieves significant accuracy gains of +5.23\% (Qwen-32B, p < 0.01), +3.43\% (GPT-OSS-20B, p < 0.01), and +5.54\% (Llama-3B, p < 0.05) over single-task baselines. Systematic error analysis reveals scale-dependent learning dynamics: at 32B scale, cocktail training transforms unfixable type errors (12\% repair rate) into correctable domain errors (96\% repair rate), achieving 92.7\% overall correction; at 8B scale, the same training eliminates syntactic errors but introduces semantic failures, revealing a critical capacity threshold for type-safe symbolic reasoning.

Chain-of-thought (CoT) instructions can distort multiple-choice VLM evaluation when a scorer appends a reasoning cue but reads answer-label logits before the model generates any rationale. We call this CoT-prefix scoring. On ScienceQA, Qwen2.5-VL-7B drops from 80.76% to 45.48%, and across five option-content permutations 93.54% of CoT-prefix predictions select the first slot. Condition-matched linear probes recover 78.94% from the same hidden states, while free generation restores 75.24%, showing that the answer often survives the prefix and the immediate readout fails. Vocabulary and layer diagnostics explain the mismatch: probability mass moves toward continuation tokens, while answer information remains linearly accessible in late layers. The effect recurs with varying severity across datasets and models, though not universally. These results show that CoT-prefix scoring can confound model knowledge with an evaluation-interface mismatch and should be avoided unless the requested and scored output events are aligned.

Reasoning distillation has emerged as an efficient and powerful paradigm for enhancing the reasoning capabilities of large language models. However, reasoning distillation may inadvertently cause benchmark contamination, where evaluation data included in distillation datasets can inflate performance metrics of distilled models. In this work, we formally define the task of distillation data detection, which is uniquely challenging due to the partial availability of distillation data. Then, we propose a novel and effective method Token Probability Deviation (TBD), which leverages the probability patterns of the generated output tokens. Our method is motivated by the analysis that distilled models tend to generate near-deterministic tokens for seen questions, while producing more low-probability tokens for unseen questions. Our key idea behind TBD is to quantify how far the generated tokens' probabilities deviate from a high reference probability. In effect, our method achieves competitive detection performance by producing lower scores for seen questions than for unseen questions. Extensive experiments demonstrate the effectiveness of our method, achieving an AUC of 0.918 and a TPR@1% FPR of 0.470 on the S1 dataset.

Vision-language models (VLMs) are increasingly deployed in high-stakes settings where reliable uncertainty quantification (UQ) is as important as predictive accuracy. Extended reasoning via chain-of-thought (CoT) prompting or reasoning-trained models has become ubiquitous in modern VLM pipelines, yet its effect on UQ reliability remains poorly understood. We show that reasoning consistently degrades the quality of most uncertainty estimates, even when it improves task accuracy. We identify implicit answer conditioning as the primary mechanism: as reasoning traces converge on a conclusion before the final answer is generated, token probabilities increasingly reflect consistency with the model's own reasoning trace rather than uncertainty about correctness. In effect, the model becomes overconfident in its answer. In contrast, agreement-based consistency remains robust and often improves under reasoning, making it a practical choice for uncertainty estimation in reasoning-enabled VLMs.

This work presents code to procedurally generate examples for the ARC training tasks. For each of the 400 tasks, an example generator following the transformation logic of the original examples was created. In effect, the assumed underlying distribution of examples for any given task was reverse engineered by implementing a means to sample from it. An attempt was made to cover an as large as reasonable space of possible examples for each task. That is, whenever the original examples of a given task may be limited in their diversity e.g. by having the dimensions of the grids, the set of symbols or number of objects constant or within tight bounds, even though the transformation does not require it, such constraints were lifted. Having access to not just a few examples per task, as the case for ARC, but instead very many, should enable a wide range of experiments that may be important stepping stones towards making leaps on the benchmark.

Large language models (LLMs) have accomplished remarkable reasoning performance in various domains. However, in the domain of reasoning tasks, we discover a frailty: LLMs are surprisingly brittle to the ordering of the premises, despite the fact that such ordering does not alter the underlying task. In particular, we observe that LLMs achieve the best performance when the premise order aligns with the context required in intermediate reasoning steps. For example, in deductive reasoning tasks, presenting the premises in the same order as the ground truth proof in the prompt (as opposed to random ordering) drastically increases the model's accuracy. We first examine the effect of premise ordering on deductive reasoning on a variety of LLMs, and our evaluation shows that permuting the premise order can cause a performance drop of over 30%. In addition, we release the benchmark R-GSM, based on GSM8K, to examine the ordering effect for mathematical problem-solving, and we again observe a significant drop in accuracy, relative to the original GSM8K benchmark.

With the rise of Large Language Models(LLMs), it has become crucial to understand their capabilities and limitations in deciphering and explaining the complex web of causal relationships that language entails. Current methods use either explicit or implicit causal reasoning, yet there is a strong need for a unified approach combining both to tackle a wide array of causal relationships more effectively. This research proposes a novel architecture called Context Aware Reasoning Enhancement with Counterfactual Analysis(CARE CA) framework to enhance causal reasoning and explainability. The proposed framework incorporates an explicit causal detection module with ConceptNet and counterfactual statements, as well as implicit causal detection through LLMs. Our framework goes one step further with a layer of counterfactual explanations to accentuate LLMs understanding of causality. The knowledge from ConceptNet enhances the performance of multiple causal reasoning tasks such as causal discovery, causal identification and counterfactual reasoning. The counterfactual sentences add explicit knowledge of the not caused by scenarios. By combining these powerful modules, our model aims to provide a deeper understanding of causal relationships, enabling enhanced interpretability. Evaluation of benchmark datasets shows improved performance across all metrics, such as accuracy, precision, recall, and F1 scores. We also introduce CausalNet, a new dataset accompanied by our code, to facilitate further research in this domain.

We present LogiGAN, an unsupervised adversarial pre-training framework for improving logical reasoning abilities of language models. Upon automatic identifying logical reasoning phenomena in massive text corpus via detection heuristics, we train language models to predict the masked-out logical statements. Inspired by the facilitation effect of reflective thinking in human learning, we analogically simulate the learning-thinking process with an adversarial Generator-Verifier architecture to assist logic learning. LogiGAN implements a novel sequential GAN approach that (a) circumvents the non-differentiable challenge of the sequential GAN by leveraging the Generator as a sentence-level generative likelihood scorer with a learning objective of reaching scoring consensus with the Verifier; (b) is computationally feasible for large-scale pre-training with arbitrary target length. Both base and large size language models pre-trained with LogiGAN demonstrate obvious performance improvement on 12 datasets requiring general reasoning abilities, revealing the fundamental role of logic in broad reasoning, as well as the effectiveness of LogiGAN. Ablation studies on LogiGAN components reveal the relative orthogonality between linguistic and logic abilities and suggest that reflective thinking's facilitation effect might also generalize to machine learning.

Video Models have achieved remarkable success in high-fidelity video generation with coherent motion dynamics. Analogous to the development from text generation to text-based reasoning in language modeling, the development of video models motivates us to ask: Can video models reason via video generation? Compared with the discrete text corpus, video grounds reasoning in explicit spatial layouts and temporal continuity, which serves as an ideal substrate for spatial reasoning. In this work, we explore the reasoning via video paradigm and introduce VR-Bench -- a comprehensive benchmark designed to systematically evaluate video models' reasoning capabilities. Grounded in maze-solving tasks that inherently require spatial planning and multi-step reasoning, VR-Bench contains 7,920 procedurally generated videos across five maze types and diverse visual styles. Our empirical analysis demonstrates that SFT can efficiently elicit the reasoning ability of video model. Video models exhibit stronger spatial perception during reasoning, outperforming leading VLMs and generalizing well across diverse scenarios, tasks, and levels of complexity. We further discover a test-time scaling effect, where diverse sampling during inference improves reasoning reliability by 10--20%. These findings highlight the unique potential and scalability of reasoning via video for spatial reasoning tasks.

Scaling up multimodal models has enabled remarkable advances in visual understanding and reasoning, but practical demands call for smaller, efficient systems. In this work, we conduct a principled analysis of downscaling intelligence in multimodal models, examining how reduced large language model (LLM) capacity affects multimodal capabilities. Our initial findings reveal an interesting trend: LLM downscaling disproportionately affects visual capabilities, rather than abilities inherited from the LLM. We then examine whether this drop mainly reflects the expected decline in visual reasoning or a more fundamental loss of perceptual abilities. Isolating the effect of LLM downscaling on perception, we find performance still drops sharply, often matching or exceeding the impact on reasoning. To address this bottleneck, we introduce visual extraction tuning, which explicitly trains the model to extract instruction-relevant visual details consistently across tasks. With these extracted visual details, we then apply step-by-step reasoning to generate answers. Together, these components form our Extract+Think approach, setting a new standard for efficiency and performance in this space.

When people reason about cause and effect, they often consider many competing "what if" scenarios before deciding which explanation fits best. Analogously, advanced language models capable of causal inference can consider multiple interventions and counterfactuals to judge the validity of causal claims. Crucially, this type of reasoning is less like a single calculation and more like an internal dialogue between alternative hypotheses. In this paper, we make this dialogue explicit through a dual-agent debate framework where one model provides a structured causal inference, and the other critically examines this reasoning for logical flaws. When disagreements arise, agents attempt to persuade each other, challenging each other's logic and revising their conclusions until they converge on a mutually agreed answer. To take advantage of this deliberative process, we specifically use reasoning language models, whose strengths in both causal inference and adversarial debate remain under-explored relative to standard large language models. We evaluate our approach on the CLadder dataset, a benchmark linking natural language questions to formally defined causal graphs across all three rungs of Pearl's ladder of causation. With Qwen3 and DeepSeek-R1 as debater agents, we demonstrate that multi-agent debate improves DeepSeek-R1's overall accuracy in causal inference from 78.03% to 87.45%, with the counterfactual category specifically improving from 67.94% to 80.04% accuracy. Similarly, Qwen3's overall accuracy improves from 84.16% to 89.41%, and counterfactual questions from 71.53% to 80.35%, showing that strong models can still benefit greatly from debate with weaker agents. Our results highlight the potential of reasoning models as building blocks for multi-agent systems in causal inference, and demonstrate the importance of diverse perspectives in causal problem-solving.

Supervised fine-tuning (SFT) on chain-of-thought (CoT) trajectories demonstrations is a common approach for enabling reasoning in large language models. Standard practices typically only retain trajectories with correct final answers (positives) while ignoring the rest (negatives). We argue that this paradigm discards substantial supervision and exacerbates overfitting, limiting out-of-domain (OOD) generalization. Specifically, we surprisingly find that incorporating negative trajectories into SFT yields substantial OOD generalization gains over positive-only training, as these trajectories often retain valid intermediate reasoning despite incorrect final answers. To understand this effect in depth, we systematically analyze data, training dynamics, and inference behavior, identifying 22 recurring patterns in negative chains that serve a dual role: they moderate loss descent to mitigate overfitting during training and boost policy entropy by 35.67% during inference to facilitate exploration. Motivated by these observations, we further propose Gain-based LOss Weighting (GLOW), an adaptive, sample-aware scheme that exploits such distinctive training dynamics by rescaling per-sample loss based on inter-epoch progress. Empirically, GLOW efficiently leverages unfiltered trajectories, yielding a 5.51% OOD gain over positive-only SFT on Qwen2.5-7B and boosting MMLU from 72.82% to 76.47% as an RL initialization.

Causal reasoning, the ability to identify cause-and-effect relationship, is crucial in human thinking. Although large language models (LLMs) succeed in many NLP tasks, it is still challenging for them to conduct complex causal reasoning like abductive reasoning and counterfactual reasoning. Given the fact that programming code may express causal relations more often and explicitly with conditional statements like ``if``, we want to explore whether Code-LLMs acquire better causal reasoning abilities. Our experiments show that compared to text-only LLMs, Code-LLMs with code prompts are significantly better in causal reasoning. We further intervene on the prompts from different aspects, and discover that the programming structure is crucial in code prompt design, while Code-LLMs are robust towards format perturbations.

Recent advancements in Chain of Thought (COT) generation have significantly improved the reasoning capabilities of Large Language Models (LLMs), with reinforcement learning (RL) emerging as an effective post-training approach. Multimodal Large Language Models (MLLMs) inherit this reasoning potential but remain underexplored in tasks requiring both perception and logical reasoning. To address this, we introduce SEED-Bench-R1, a benchmark designed to systematically evaluate post-training methods for MLLMs in video understanding. It includes intricate real-world videos and complex everyday planning tasks in the format of multiple-choice questions, requiring sophisticated perception and reasoning. SEED-Bench-R1 assesses generalization through a three-level hierarchy: in-distribution, cross-environment, and cross-environment-task scenarios, equipped with a large-scale training dataset with easily verifiable ground-truth answers. Using Qwen2-VL-Instruct-7B as a base model, we compare RL with supervised fine-tuning (SFT), demonstrating RL's data efficiency and superior performance on both in-distribution and out-of-distribution tasks, even outperforming SFT on general video understanding benchmarks like LongVideoBench. Our detailed analysis reveals that RL enhances visual perception but often produces less logically coherent reasoning chains. We identify key limitations such as inconsistent reasoning and overlooked visual cues, and suggest future improvements in base model reasoning, reward modeling, and RL robustness against noisy signals.

Cognitive diversity, reflected in variations of language, perspective, and reasoning, is essential to creativity and collective intelligence. This diversity is rich and grounded in culture, history, and individual experience. Yet as large language models (LLMs) become deeply embedded in people's lives, they risk standardizing language and reasoning. This Review synthesizes evidence across linguistics, cognitive, and computer science to show how LLMs reflect and reinforce dominant styles while marginalizing alternative voices and reasoning strategies. We examine how their design and widespread use contribute to this effect by mirroring patterns in their training data and amplifying convergence as all people increasingly rely on the same models across contexts. Unchecked, this homogenization risks flattening the cognitive landscapes that drive collective intelligence and adaptability.

Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in complex problem-solving through Chain-of-Thought (CoT) reasoning. However, the multi-step nature of CoT introduces new safety challenges that extend beyond conventional language model alignment. We identify a failure mode in current safety CoT tuning methods: the snowball effect, where minor reasoning deviations progressively amplify throughout the thought process, leading to either harmful compliance or excessive refusal. This effect stems from models being trained to imitate perfect reasoning scripts without learning to self-correct. To address this limitation, we propose AdvChain, an alignment paradigm that teaches models dynamic self-correction through adversarial CoT tuning. Our method involves constructing a dataset containing Temptation-Correction and Hesitation-Correction samples, where models learn to recover from harmful reasoning drifts and unnecessary cautions. Extensive experiments show that AdvChain significantly enhances robustness against jailbreak attacks and CoT hijacking while substantially reducing over-refusal on benign prompts, achieving a superior safety-utility balance without compromising reasoning capabilities. Our work establishes a new direction for building more robust and reliable reasoning 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

Do large language models reason morally, or do they merely sound like they do? We investigate whether LLM responses to moral dilemmas exhibit genuine developmental progression through Kohlberg's stages of moral development, or whether alignment training instead produces reasoning-like outputs that superficially resemble mature moral judgment without the underlying developmental trajectory. Using an LLM-as-judge scoring pipeline validated across three judge models, we classify more than 600 responses from 13 LLMs spanning a range of architectures, parameter scales, and training regimes across six classical moral dilemmas, and conduct ten complementary analyses to characterize the nature and internal coherence of the resulting patterns. Our results reveal a striking inversion: responses overwhelmingly correspond to post-conventional reasoning (Stages 5-6) regardless of model size, architecture, or prompting strategy, the effective inverse of human developmental norms, where Stage 4 dominates. Most strikingly, a subset of models exhibit moral decoupling: systematic inconsistency between stated moral justification and action choice, a form of logical incoherence that persists across scale and prompting strategy and represents a direct reasoning consistency failure independent of rhetorical sophistication. Model scale carries a statistically significant but practically small effect; training type has no significant independent main effect; and models exhibit near-robotic cross-dilemma consistency producing logically indistinguishable responses across semantically distinct moral problems. We posit that these patterns constitute evidence for moral ventriloquism: the acquisition, through alignment training, of the rhetorical conventions of mature moral reasoning without the underlying developmental trajectory those conventions are meant to represent.

The fundamental goal of agentic visual reasoning is to improve the success rate of multimodal large language models (MLLMs) on complex tasks, rather than merely equipping them with a sophisticated yet inefficient reasoning paradigm. In this work, we rethink agentic visual reasoning through two key dimensions of tool use: Mode Adaptiveness (MA) and Tool Effect (TE). Mode Adaptiveness characterizes whether an MLLM can recognize when tools are truly necessary and invoke them accordingly, thereby avoiding unnecessary computational overhead while improving performance on challenging problems that require tool assistance. Tool Effect characterizes the actual impact of tool use: tools should extend the model's capabilities on problems unsolvable through text-only reasoning, while avoiding additional errors on problems that the model can already solve without tools. We conduct a comprehensive analysis to quantify these two properties and empirically reveal that existing agentic visual reasoning models exhibit limited Mode Adaptiveness, while the gains produced by tool use on hard examples are largely offset by the harm introduced on easy examples that the models can already solve. Motivated by these observations, we propose Beacon, a novel agentic visual reasoning model that achieves stronger overall performance, improved Mode Adaptiveness, and genuine tool-induced performance gains. At the core of Beacon are the Necessity-Aware Adaptive Reward and the Hint-Guided Capability Expansion mechanism in the reinforcement learning stage, which respectively encourage adaptive tool invocation based on task necessity and strengthen the model's tool-use capability on the most challenging problems. Extensive experiments across diverse benchmarks demonstrate the strong overall performance of Beacon and its substantial improvements in both Mode Adaptiveness and Tool Effect.

ETAS is a programming language for agent systems that treats model-backed agents, tool calls, prompts, typed memory, human approvals, policies, and execution traces as semantic program elements rather than library conventions. It separates deterministic computation from agentic nondeterminism and externally visible actions while preserving a direct programming style. We present the core design of ETAS. Its static semantics assigns ordinary types through spec conformance and tracks each computation with two behavioral indices: an escaping effect row and a persistent abstraction of the typed action trace it may request. Specs form a terminating compile-time constraint calculus: type specs provide evidence for polymorphism and resource facts, callable specs constrain function and stage shapes, and trace specs express allow, deny, and temporal constraints. Typing checks requested traces against compiled monitors and emits residual obligations when dynamic resources preclude a complete static proof. The dynamic semantics distinguish requested, handled, denied, and committed events; handlers interpret typed actions without making their requests invisible to authorization or audit. We formalize a core calculus and state preservation, progress, type/effect soundness, handler trace-transparency, and policy safety. We also implement ETAS in Rust with a command-line interface, typed HIR checks, effect and policy diagnostics, handler checks, and trace-aware execution hooks. ETAS provides a programming-language foundation for reasoning about authorization, nondeterminism, recovery, and audit evidence before and during agent execution.

As large language models (LLMs) grow larger and more sophisticated, assessing their "reasoning" capabilities in natural language grows more challenging. Recent question answering (QA) benchmarks that attempt to assess reasoning are often limited by a narrow scope of covered situations and subject matters. We introduce WikiWhy, a QA dataset built around a novel auxiliary task: explaining why an answer is true in natural language. WikiWhy contains over 9,000 "why" question-answer-rationale triples, grounded on Wikipedia facts across a diverse set of topics. Each rationale is a set of supporting statements connecting the question to the answer. WikiWhy serves as a benchmark for the reasoning capabilities of LLMs because it demands rigorous explicit rationales for each answer to demonstrate the acquisition of implicit commonsense knowledge, which is unlikely to be easily memorized. GPT-3 baselines achieve only 38.7% human-evaluated correctness in the end-to-end answer & explain condition, leaving significant room for future improvements.

Large language models have emerged as powerful tools for general intelligence, showcasing advanced natural language processing capabilities that find applications across diverse domains. Despite their impressive performance, recent studies have highlighted the potential for significant enhancements in LLMs' task-specific performance through fine-tuning strategies like Reinforcement Learning with Human Feedback (RLHF), supervised fine-tuning (SFT), and Quantized Low-Rank Adapters (Q-LoRA) method. However, previous works have shown that while fine-tuning offers significant performance gains, it also leads to challenges such as catastrophic forgetting and privacy and safety risks. To this end, there has been little to no work in understanding the impact of fine-tuning on the reasoning capabilities of LLMs. Our research investigates the effect of fine-tuning on the reasoning abilities of LLMs, addressing critical questions regarding the impact of task-specific fine-tuning on overall reasoning capabilities, the influence of fine-tuning on Chain-of-Thought (CoT) reasoning performance, and the implications for the faithfulness of CoT reasonings. By exploring these dimensions, our study shows the impact of fine-tuning on LLM reasoning capabilities, where the faithfulness of CoT reasoning, on average across four datasets, decreases, highlighting potential shifts in internal mechanisms of the LLMs resulting from fine-tuning processes.

Mathematical reasoning has been challenging for large language models (LLMs). However, the introduction of step-by-step Chain-of-Thought (CoT) inference has significantly advanced the mathematical capabilities of LLMs. Despite this progress, current approaches either necessitate extensive inference datasets for training or depend on few-shot methods that frequently compromise computational accuracy. To address these bottlenecks in mathematical reasoning, we propose a novel method called Step Guidied Reasoning, which is more stable and generalizable than few-shot methods and does not involve further fine-tuning of the model. In this approach, LLMs reflect on small reasoning steps, similar to how humans deliberate and focus attention on what to do next. By incorporating this reflective process into the inference stage, LLMs can effectively guide their reasoning from one step to the next. Through extensive experiments, we demonstrate the significant effect of Step Guidied Reasoning in augmenting mathematical performance in state-of-the-art language models. Qwen2-72B-Instruct outperforms its math-specific counterpart, Qwen2.5-72B-Math-Instruct, on MMLU- STEM with a score of 90.9%, compared to 87.3%. The average scores of Qwen2-7B-Instruct and Qwen2-72B-Instruct increase from 27.1% to 36.3% and from 36.5% to 47.4% on the mathematics domain, respectively.

Causal reasoning and compositional reasoning are two core aspirations in generative AI. Measuring the extent of these behaviors requires principled evaluation methods. We explore a unified perspective that considers both behaviors simultaneously, termed compositional causal reasoning (CCR): the ability to infer how causal measures compose and, equivalently, how causal quantities propagate through graphs. We instantiate a framework for the systematic evaluation of CCR for the average treatment effect and the probability of necessity and sufficiency. As proof of concept, we demonstrate the design of CCR tasks for language models in the LLama, Phi, and GPT families. On a math word problem, our framework revealed a range of taxonomically distinct error patterns. Additionally, CCR errors increased with the complexity of causal paths for all models except o1.

Data-centric distillation, including data augmentation, selection, and mixing, offers a promising path to creating smaller, more efficient student Large Language Models (LLMs) that retain strong reasoning abilities. However, there still lacks a comprehensive benchmark to systematically assess the effect of each distillation approach. This paper introduces DC-CoT, the first data-centric benchmark that investigates data manipulation in chain-of-thought (CoT) distillation from method, model and data perspectives. Utilizing various teacher models (e.g., o4-mini, Gemini-Pro, Claude-3.5) and student architectures (e.g., 3B, 7B parameters), we rigorously evaluate the impact of these data manipulations on student model performance across multiple reasoning datasets, with a focus on in-distribution (IID) and out-of-distribution (OOD) generalization, and cross-domain transfer. Our findings aim to provide actionable insights and establish best practices for optimizing CoT distillation through data-centric techniques, ultimately facilitating the development of more accessible and capable reasoning models. The dataset can be found at https://huggingface.co/datasets/rana-shahroz/DC-COT, while our code is shared in https://anonymous.4open.science/r/DC-COT-FF4C/.

Chain-of-thought reasoning has demonstrated significant improvements in large language models and vision language models, yet its potential for audio language models remains largely unexplored. In this technical report, we take a preliminary step towards closing this gap. For better assessment of sound reasoning, we propose AF-Reasoning-Eval, a benchmark targeting common-sense reasoning and the ability to discriminate among closely related choices. To prepare training corpus for sound reasoning abilities, we propose automatic pipelines that transform existing audio question answering and classification data into explicit reasoning chains, yielding AF-CoT-Train with 1.24M samples. We study the effect of finetuning Audio Flamingo series on AF-CoT-Train and observe considerable improvements on several reasoning benchmarks, validating the effectiveness of chain-of-thought finetuning on advanced sound understanding.

Self-improvement has emerged as a mainstream paradigm for advancing the reasoning capabilities of large vision-language models (LVLMs), where models explore and learn from successful trajectories iteratively. However, we identify a critical issue during this process: the model excels at generating high-quality trajectories for simple queries (i.e., head data) but struggles with more complex ones (i.e., tail data). This leads to an imbalanced optimization that drives the model to prioritize simple reasoning skills, while hindering its ability to tackle more complex reasoning tasks. Over iterations, this imbalance becomes increasingly pronounced--a dynamic we term the "Matthew effect"--which ultimately hinders further model improvement and leads to performance bottlenecks. To counteract this challenge, we introduce four efficient strategies from two perspectives: distribution-reshaping and trajectory-resampling, to achieve head-tail re-balancing during the exploration-and-learning self-improvement process. Extensive experiments on Qwen2-VL-7B-Instruct and InternVL2.5-4B models across visual reasoning tasks demonstrate that our methods consistently improve visual reasoning capabilities, outperforming vanilla self-improvement by 3.86 points on average.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key method for improving Large Language Models' reasoning capabilities, yet recent evidence suggests it may paradoxically shrink the reasoning boundary rather than expand it. This paper investigates the shrinkage issue of RLVR by analyzing its learning dynamics and reveals two critical phenomena that explain this failure. First, we expose negative interference in RLVR, where learning to solve certain training problems actively reduces the likelihood of correct solutions for others, leading to the decline of Pass@k performance, or the probability of generating a correct solution within k attempts. Second, we uncover the winner-take-all phenomenon: RLVR disproportionately reinforces problems with high likelihood, correct solutions, under the base model, while suppressing other initially low-likelihood ones. Through extensive theoretical and empirical analysis on multiple mathematical reasoning benchmarks, we show that this effect arises from the inherent on-policy sampling in standard RL objectives, causing the model to converge toward narrow solution strategies. Based on these insights, we propose a simple yet effective data curation algorithm that focuses RLVR learning on low-likelihood problems, achieving notable improvement in Pass@k performance. Our code is available at https://github.com/mail-research/SELF-llm-interference.

Do reasoning models have "Aha!" moments? Prior work suggests that models like DeepSeek-R1-Zero undergo sudden mid-trace realizations that lead to accurate outputs, implying an intrinsic capacity for self-correction. Yet, it remains unclear whether such intrinsic shifts in reasoning strategy actually improve performance. Here, we study mid-reasoning shifts and instrument training runs to detect them. Our analysis spans 1M+ reasoning traces, hundreds of training checkpoints, three reasoning domains, and multiple decoding temperatures and model architectures. We find that reasoning shifts are rare, do not become more frequent with training, and seldom improve accuracy, indicating that they do not correspond to prior perceptions of model insight. However, their effect varies with model uncertainty. Building on this finding, we show that artificially triggering extrinsic shifts under high entropy reliably improves accuracy. Our results show that mid-reasoning shifts are symptoms of unstable inference behavior rather than an intrinsic mechanism for self-correction.

State-of-the-art embedding models are increasingly derived from decoder-only Large Language Model (LLM) backbones adapted via contrastive learning. Given the emergence of reasoning models trained via Reinforcement Learning with Verifiable Rewards (RLVR), a natural question arises: do enhanced reasoning translate to superior semantic representations when these models serve as embedding initializations? Contrary to expectation, our evaluation on MTEB and BRIGHT reveals a **null effect**: embedding models initialized from RLVR-tuned backbones yield no consistent performance advantage over their base counterparts when subjected to identical training recipes. To unpack this paradox, we introduce **H**ierarchical **R**epresentation **S**imilarity **A**nalysis (HRSA), a framework that decomposes similarity across representation, geometry, and function levels. HRSA reveals that while RLVR induces irreversible latent manifold's local geometry reorganization and reversible coordinate basis drift, it preserves the global manifold geometry and linear readout. Consequently, subsequent contrastive learning drives strong alignment between base- and reasoning-initialized models, a phenomenon we term **Manifold Realignment**. Empirically, our findings suggest that unlike Supervised Fine-Tuning (SFT), RLVR optimizes trajectories within an existing semantic landscape rather than fundamentally restructuring the landscape itself.

Reasoning prefixes shape the future trajectory of LLM problem solving, yet existing process reward models usually evaluate them through local step correctness. We argue that correctness is a useful but indirect proxy for the effect we ultimately care about: whether a prefix increases the probability of successful completion. We define this effect as prefix gain, the solve-rate improvement induced by conditioning lightweight student model group on a prefix, and use it to train a Prefix Utility Model (PUM) with a simple pairwise ranking objective. PUM learns outcome-grounded prefix utility and can score both complete trajectories and partial reasoning prefixes. Across Best-of-N selection, beam search, and reinforcement learning on mathematical reasoning, PUM provides a strong prefix-level supervision signal, especially when candidate pools are large, search budgets increase, or rule-based rewards are sparse. We release all data, models, and code at https://zhiqix.github.io/pum-project-page.

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.

Detecting AI-generated images is only half the task: a deployed detector must also justify its verdict, yet existing detectors inherit three failure modes from their training data: real and fake images collected from different sources invite provenance shortcuts, supervised explanation corpora teach templated rationales, and a static forgery corpus leaves the decision boundary standing still while generators keep moving. We introduce , an adversarial reinforcement learning framework that pits two heterogeneous models against each other. A diffusion image editor learns to edit real photographs into fake counterparts of those same photographs that fool the current detector, while a reasoning MLLM learns to expose them with a verdict grounded in free-form reasoning. Both rewards are shortcut-proof by design: the attacker is credited only when its edit is faithfully executed, and the defender only when its verdict is correct. As the two models alternate, each round's attacker regenerates a harder training pool aimed at the current detector's blind spots, so the detector must generalize rather than memorize any fixed artifact distribution. Although the explanation is never rewarded, its quality rises round over round as a side effect of accuracy-only training. A detector trained within this loop improves monotonically across rounds on each of three external benchmarks.

Reactive programming frameworks such as React allow developers to build interactive applications by declaratively specifying how outputs depend on changing inputs. Although this model makes it easy to reason about what an application computes, the temporal behavior of reactive programs remains difficult to understand and verify. Applications implicitly rely on timing assumptions buried in framework runtimes, leading to subtle bugs such as stale reads, transient inconsistencies, order-dependent behavior, and unintended feedback cycles. To address these challenges, this paper presents Willow, a core calculus for reactive programming inspired by React. Willow gives a time-aware operational semantics that models computation in terms of renders, the fundamental evaluation step in which components produce user interface descriptions, and pairs it with a novel type-and-effect system that statically tracks timing behavior as effects. A "next" modality expresses delays measured not only in renders but in any unit the host environment exposes--renders, network requests, or milliseconds. A family of modalities tracks the lifecycle of event handlers: when they are registered, when they fire, when pending events are canceled and when handlers are removed. A key insight is that the resulting effects form a temporal dependency graph, letting standard graph algorithms statically detect render cascades and inter-render loops that cause non-termination or performance degradation. We formalize Willow and prove preservation of the effect system with respect to the time-aware semantics. We also implement a prototype checker with automatic effect inference and evaluate it on representative reactive patterns such as debouncing, form inputs, and API-driven updates. Our results demonstrate that time-aware typing provides a practical foundation for reasoning about the temporal correctness of reactive programs.

A key component of successfully reading a passage of text is the ability to apply knowledge gained from the passage to a new situation. In order to facilitate progress on this kind of reading, we present ROPES, a challenging benchmark for reading comprehension targeting Reasoning Over Paragraph Effects in Situations. We target expository language describing causes and effects (e.g., "animal pollinators increase efficiency of fertilization in flowers"), as they have clear implications for new situations. A system is presented a background passage containing at least one of these relations, a novel situation that uses this background, and questions that require reasoning about effects of the relationships in the background passage in the context of the situation. We collect background passages from science textbooks and Wikipedia that contain such phenomena, and ask crowd workers to author situations, questions, and answers, resulting in a 14,322 question dataset. We analyze the challenges of this task and evaluate the performance of state-of-the-art reading comprehension models. The best model performs only slightly better than randomly guessing an answer of the correct type, at 61.6% F1, well below the human performance of 89.0%.

Fuzzy reasoning is vital due to the frequent use of imprecise information in daily contexts. However, the ability of current large language models (LLMs) to handle such reasoning remains largely uncharted. In this paper, we introduce a new benchmark, FRoG, for fuzzy reasoning, featuring real-world mathematical word problems that incorporate generalized quantifiers. Our experimental findings reveal that fuzzy reasoning continues to pose significant challenges for LLMs. Moreover, we find that existing methods designed to enhance reasoning do not consistently improve performance in tasks involving fuzzy logic. Additionally, our results show an inverse scaling effect in the performance of LLMs on FRoG. Interestingly, we also demonstrate that strong mathematical reasoning skills are not necessarily indicative of success on our benchmark.

The task of video-based commonsense captioning aims to generate event-wise captions and meanwhile provide multiple commonsense descriptions (e.g., attribute, effect and intention) about the underlying event in the video. Prior works explore the commonsense captions by using separate networks for different commonsense types, which is time-consuming and lacks mining the interaction of different commonsense. In this paper, we propose a Hybrid Reasoning Network (HybridNet) to endow the neural networks with the capability of semantic-level reasoning and word-level reasoning. Firstly, we develop multi-commonsense learning for semantic-level reasoning by jointly training different commonsense types in a unified network, which encourages the interaction between the clues of multiple commonsense descriptions, event-wise captions and videos. Then, there are two steps to achieve the word-level reasoning: (1) a memory module records the history predicted sequence from the previous generation processes; (2) a memory-routed multi-head attention (MMHA) module updates the word-level attention maps by incorporating the history information from the memory module into the transformer decoder for word-level reasoning. Moreover, the multimodal features are used to make full use of diverse knowledge for commonsense reasoning. Experiments and abundant analysis on the large-scale Video-to-Commonsense benchmark show that our HybridNet achieves state-of-the-art performance compared with other methods.

Large Language Models (LLMs) have attained human-level accuracy on medical question-answer (QA) benchmarks. However, their limitations in navigating open-ended clinical scenarios have recently been shown, raising concerns about the robustness and generalizability of LLM reasoning across diverse, real-world medical tasks. To probe potential LLM failure modes in clinical problem-solving, we present the medical abstraction and reasoning corpus (M-ARC). M-ARC assesses clinical reasoning through scenarios designed to exploit the Einstellung effect -- the fixation of thought arising from prior experience, targeting LLM inductive biases toward inflexible pattern matching from their training data rather than engaging in flexible reasoning. We find that LLMs, including current state-of-the-art o1 and Gemini models, perform poorly compared to physicians on M-ARC, often demonstrating lack of commonsense medical reasoning and a propensity to hallucinate. In addition, uncertainty estimation analyses indicate that LLMs exhibit overconfidence in their answers, despite their limited accuracy. The failure modes revealed by M-ARC in LLM medical reasoning underscore the need to exercise caution when deploying these models in clinical settings.

Effective evaluation of the reasoning capabilities of large language models (LLMs) are susceptible to overestimation due to data exposure of evaluation benchmarks. We introduce a framework for producing linguistic reasoning problems that reduces the effect of memorisation in model performance estimates and apply this framework to develop LINGOLY-TOO, a challenging evaluation benchmark for linguistic reasoning. By developing orthographic templates, we dynamically obfuscate the writing systems of real languages to generate numerous question variations. These variations preserve the reasoning steps required for each solution while reducing the likelihood of specific problem instances appearing in model training data. Our experiments demonstrate that frontier models, including OpenAI o1-preview and DeepSeem R1, struggle with advanced reasoning. Our analysis also shows that LLMs exhibit noticeable variance in accuracy across permutations of the same problem, and on average perform better on questions appearing in their original orthography. Our findings highlight the opaque nature of response generation in LLMs and provide evidence that prior data exposure contributes to overestimating the reasoning capabilities of frontier models.

Unlocking deep, interpretable biological reasoning from complex genomic data is a major AI challenge hindering scientific discovery. Current DNA foundation models, despite strong sequence representation, struggle with multi-step reasoning and lack inherent transparent, biologically intuitive explanations. We introduce BioReason, a pioneering architecture that, for the first time, deeply integrates a DNA foundation model with a Large Language Model (LLM). This novel connection enables the LLM to directly process and reason with genomic information as a fundamental input, fostering a new form of multimodal biological understanding. BioReason's sophisticated multi-step reasoning is developed through supervised fine-tuning and targeted reinforcement learning, guiding the system to generate logical, biologically coherent deductions. On biological reasoning benchmarks including KEGG-based disease pathway prediction - where accuracy improves from 88% to 97% - and variant effect prediction, BioReason demonstrates an average 15% performance gain over strong single-modality baselines. BioReason reasons over unseen biological entities and articulates decision-making through interpretable, step-by-step biological traces, offering a transformative approach for AI in biology that enables deeper mechanistic insights and accelerates testable hypothesis generation from genomic data. Data, code, and checkpoints are publicly available at https://github.com/bowang-lab/BioReason

Empirical scaling laws have driven the evolution of large language models (LLMs), yet their coefficients shift whenever the model architecture or data pipeline changes. Mixture-of-Experts (MoE) models, now standard in state-of-the-art systems, introduce a new sparsity dimension that current dense-model frontiers overlook. We investigate how MoE sparsity influences two distinct capability regimes: memorization and reasoning. We train families of MoE Transformers that systematically vary total parameters, active parameters, and top-k routing while holding the compute budget fixed. For every model we record pre-training loss, downstream task loss, and task accuracy, allowing us to separate the train-test generalization gap from the loss-accuracy gap. Memorization benchmarks improve monotonically with total parameters, mirroring training loss. By contrast, reasoning performance saturates and can even regress despite continued gains in both total parameters and training loss. Altering top-k alone has little effect when active parameters are constant, and classic hyperparameters such as learning rate and initialization modulate the generalization gap in the same direction as sparsity. Neither post-training reinforcement learning (GRPO) nor extra test-time compute rescues the reasoning deficit of overly sparse models. Our model checkpoints, code and logs are open-source at https://github.com/rioyokotalab/optimal-sparsity.

We study how different Chain-of-Thought (CoT) designs affect the acquisition of the generalizable visual reasoning ability in vision-language models (VLMs). While CoT data, especially long or visual CoT such as "think with image", has been widely used to supervise intermediate reasoning, it remains unclear why specific CoT designs help and which ones truly support generalizable reasoning. To systematically evaluate this, we focus on a controlled maze-solving benchmark where reasoning rules are fully visual, difficulty can be tuned by grid size, and all the intermediate steps can be automatically generated. Using Qwen2.5-VL-7B under a standard SFT-then-RL pipeline, we compare three representative CoT formats: Language CoT, Grounding CoT (with spatial coordinate trajectories), and Visual CoT (with image manipulations). Our experiments reveal that visual and longer CoT mainly accelerate convergence but do not lift the final performance ceiling; concise CoT containing only essential grounding steps outperforms longer traces; and, strikingly, CoT retaining only the minimal grounding results generalizes best across different maze sizes. We further validate these insights on other vision-centric tasks. These findings highlight a "short is long" effect and provide practical guidance for constructing more generalizable SFT datasets for visual reasoning.

Reinforcement Learning (RL) has enabled Large Language Models (LLMs) to achieve remarkable reasoning in domains like mathematics and coding, where verifiable rewards provide clear signals. However, extending this paradigm to financial decision is challenged by the market's stochastic nature: rewards are verifiable but inherently noisy, causing standard RL to degenerate into reward hacking. To address this, we propose Trade-R1, a model training framework that bridges verifiable rewards to stochastic environments via process-level reasoning verification. Our key innovation is a verification method that transforms the problem of evaluating reasoning over lengthy financial documents into a structured Retrieval-Augmented Generation (RAG) task. We construct a triangular consistency metric, assessing pairwise alignment between retrieved evidence, reasoning chains, and decisions to serve as a validity filter for noisy market returns. We explore two reward integration strategies: Fixed-effect Semantic Reward (FSR) for stable alignment signals, and Dynamic-effect Semantic Reward (DSR) for coupled magnitude optimization. Experiments on different country asset selection demonstrate that our paradigm reduces reward hacking, with DSR achieving superior cross-market generalization while maintaining the highest reasoning consistency.

Most work interpreting reasoning models studies only a single chain-of-thought (CoT), yet these models define distributions over many possible CoTs. We argue that studying a single sample is inadequate for understanding causal influence and the underlying computation. Though fully specifying this distribution is intractable, it can be understood by sampling. We present case studies using resampling to investigate model decisions. First, when a model states a reason for its action, does that reason actually cause the action? In "agentic misalignment" scenarios, we resample specific sentences to measure their downstream effects. Self-preservation sentences have small causal impact, suggesting they do not meaningfully drive blackmail. Second, are artificial edits to CoT sufficient for steering reasoning? These are common in literature, yet take the model off-policy. Resampling and selecting a completion with the desired property is a principled on-policy alternative. We find off-policy interventions yield small and unstable effects compared to resampling in decision-making tasks. Third, how do we understand the effect of removing a reasoning step when the model may repeat it post-edit? We introduce a resilience metric that repeatedly resamples to prevent similar content from reappearing downstream. Critical planning statements resist removal but have large effects when eliminated. Fourth, since CoT is sometimes "unfaithful", can our methods teach us anything in these settings? Adapting causal mediation analysis, we find that hints that have a causal effect on the output without being explicitly mentioned exert a subtle and cumulative influence on the CoT that persists even if the hint is removed. Overall, studying distributions via resampling enables reliable causal analysis, clearer narratives of model reasoning, and principled CoT interventions.

Medical image segmentation plays a vital role in clinical decision-making, enabling precise localization of lesions and guiding interventions. Despite significant advances in segmentation accuracy, the black-box nature of most deep models has raised growing concerns about their trustworthiness in high-stakes medical scenarios. Current explanation techniques have primarily focused on classification tasks, leaving the segmentation domain relatively underexplored. We introduced an explanation model for segmentation task which employs the causal inference framework and backpropagates the average treatment effect (ATE) into a quantification metric to determine the influence of input regions, as well as network components, on target segmentation areas. Through comparison with recent segmentation explainability techniques on two representative medical imaging datasets, we demonstrated that our approach provides more faithful explanations than existing approaches. Furthermore, we carried out a systematic causal analysis of multiple foundational segmentation models using our method, which reveals significant heterogeneity in perceptual strategies across different models, and even between different inputs for the same model. Suggesting the potential of our method to provide notable insights for optimizing segmentation models. Our code can be found at https://github.com/lcmmai/PdCR.

While reasoning in LLMs plays a natural role in math, code generation, and multi-hop factual questions, its effect on simple, single-hop factual questions remains unclear. Such questions do not require step-by-step logical decomposition, making the utility of reasoning highly counterintuitive. Nevertheless, we find that enabling reasoning substantially expands the capability boundary of the model's parametric knowledge recall, unlocking correct answers that are otherwise effectively unreachable. Why does reasoning aid parametric knowledge recall when there are no complex reasoning steps to be done? To answer this, we design a series of hypothesis-driven controlled experiments, and identify two key driving mechanisms: (1) a computational buffer effect, where the model uses the generated reasoning tokens to perform latent computation independent of their semantic content; and (2) factual priming, where generating topically related facts acts as a semantic bridge that facilitates correct answer retrieval. Importantly, this latter generative self-retrieval mechanism carries inherent risks: we demonstrate that hallucinating intermediate facts during reasoning increases the likelihood of hallucinations in the final answer. Finally, we show that our insights can be harnessed to directly improve model accuracy by prioritizing reasoning trajectories that contain hallucination-free factual statements.

Agentic Reinforcement Learning (ARL) focuses on training large language models (LLMs) to interleave reasoning with external tool execution to solve complex tasks. Most existing ARL methods train a single shared model parameters to support both reasoning and tool use behaviors, implicitly assuming that joint training leads to improved overall agent performance. Despite its widespread adoption, this assumption has rarely been examined empirically. In this paper, we systematically investigate this assumption by introducing a Linear Effect Attribution System(LEAS), which provides quantitative evidence of interference between reasoning and tool-use behaviors. Through an in-depth analysis, we show that these two capabilities often induce misaligned gradient directions, leading to training interference that undermines the effectiveness of joint optimization and challenges the prevailing ARL paradigm. To address this issue, we propose Disentangled Action Reasoning Tuning(DART), a simple and efficient framework that explicitly decouples parameter updates for reasoning and tool-use via separate low-rank adaptation modules. Experimental results show that DART consistently outperforms baseline methods with averaged 6.35 percent improvements and achieves performance comparable to multi-agent systems that explicitly separate tool-use and reasoning using a single model.

This paper assesses the ability of large language models (LLMs) to translate texts that include inter-sentential dependencies. We use the English-French DiscEvalMT benchmark (Bawden et al., 2018) with pairs of sentences containing translation challenges for pronominal anaphora and lexical cohesion. We evaluate 12 LLMs from the DeepSeek-R1, GPT, Llama, Mistral and Phi families on two tasks: (1) distinguish a correct translation from a wrong but plausible one; and (2) generate a correct translation. We compare prompts that encourage chain-of-thought reasoning with those that do not. The best models take advantage of reasoning and reach about 90% accuracy on the first task and COMET scores of about 92% on the second task, with GPT-4, GPT-4o and Phi standing out. Moreover, we observe a "wise get wiser" effect: the improvements through reasoning are larger for models that already perform well without reasoning.

As Large Reasoning Models (LRMs) are increasingly deployed, auditing their chain-of-thought (CoT) traces for safety becomes critical. Recent work has reported that monitorability--the degree to which CoT faithfully and informatively reflects internal computation--can appear as a "free gift" during the early stages of Reinforcement Learning with Verifiable Rewards (RLVR). We make this observation concrete through a systematic evaluation across model families and training domains. Our results show that this effect is not universal: monitorability improvements are strongly data-dependent. In particular, we demonstrate the critical role of data diversity and instruction-following data during RLVR training. We further show that monitorability is orthogonal to capability--improvements in reasoning performance do not imply increased transparency. Through mechanistic analysis, we attribute monitorability gains primarily to response distribution sharpening (entropy reduction) and increased attention to the prompt, rather than stronger causal reliance on reasoning traces. We also reveal how monitorability dynamics vary with controlled training and evaluation difficulty. Together, these findings provide a holistic view of how monitorability emerges under RLVR, clarifying when gains are likely to occur and when they are not.

We introduce SeePhys Pro, a fine-grained modality transfer benchmark that studies whether models preserve the same reasoning capability when critical information is progressively transferred from text to image. Unlike standard vision-essential benchmarks that evaluate a single input form, SeePhys Pro features four semantically aligned variants for each problem with progressively increasing visual elements. Our evaluation shows that current frontier models are far from representation-invariant reasoners: performance degrades on average as information moves from language to diagrams, with visual variable grounding as the most critical bottleneck. Motivated by this inference-time fragility, we further develop large training corpora for multimodal RLVR and use blind training as a diagnostic control, finding that RL with all training images masked can still improve performance on unmasked validation sets. To analyze this effect, text-deletion, image-mask-rate, and format-saturation controls suggest that such gains can arise from residual textual and distributional cues rather than valid visual evidence. Our results highlight the need to evaluate multimodal reasoning not only by final-answer accuracy, but also by robustness under modality transfer and by diagnostics that test whether improvements rely on task-critical visual evidence.

Large reasoning models (LRMs) improve complex problem-solving by generating long intermediate reasoning traces, but this substantially increases inference costs. NVFP4 inference offers a promising approach to reduce both computational and memory costs through hardware-supported low-precision execution. However, directly applying NVFP4 to LRMs introduces two practical limitations: reasoning accuracy degrades under quantization, and existing NVFP4 kernels do not fully realize latency benefits in small-batch autoregressive decoding. In this work, we analyze the effect of NVFP4 quantization on token-level uncertainty during reasoning. We show that quantization increases incorrect sampling at low-entropy symbolic tokens, while causing over-concentration on a small set of tokens in high-uncertainty reasoning steps. Based on this observation, we propose ReSET, a reasoning-step entropy-based temperature-scaling method that estimates step-level uncertainty online and adapts the decoding temperature using both token-level and step-level entropy signals. To address the latency gap, we further design a CUDA-core small-M NVFP4 kernel for latency-critical autoregressive decoding. Across reasoning benchmarks and model scales, ReSET improves NVFP4 reasoning accuracy by up to sim!2 points over the NVFP4 baseline. Our CUDA-core small-M kernel further improves latency-critical decoding, delivering up to 2.5!times kernel-level speedup over NVFP4 vLLM and approximately 2!times end-to-end decoding speedup over BF16. Code is available at https://github.com/aiha-lab/ReSET.

Large language models (LLMs) can now solve complex problems through long chain-of-thought (CoT) reasoning, but the trade-off between performance and token cost remains a central challenge. To address this issue, supervised fine-tuning (SFT) often uses compressed reasoning data, where CoT traces are shortened into compact forms. However, the effect of such compressed reasoning data on post-training remains poorly understood. In this paper, we propose a taxonomy of CoT consisting of Explicit CoT, which outputs all operations without aggregation, Composed CoT, which combines multiple operations into a single step, and Implicit CoT, which omits intermediate operations. We construct a synthetic compositional reasoning task that allows controlled variation of difficulty, compression granularity, and data size, and conducted a comprehensive set of experiments across different model families and sizes. Notably, we find that (i) coarser CoT requires more SFT data, (ii) compared with Explicit CoT, Composed CoT and Implicit CoT benefit more from data scaling, while Composed CoT benefits from data repetition and Implicit CoT tends to lead to memorization, (iii) unlike SFT, subsequent reinforcement learning (RL) with verifiable rewards (RLVR) decomposes compressed steps learned during SFT, and (iv) unidirectional CoT ordering shows stronger generalization on longer sequential tasks. Our findings provide implications for CoT design under data resource constraints and offer important insights into the mechanisms of SFT and RL in LLM post-training.

Agentic AI development today runs on token maxing: buying capability with tokens -- longer reasoning traces, more turns, wider tool payloads, bigger replayed contexts -- so tokens per task grow faster than task value. Falling per-token prices mask the pattern; total spend rises anyway. We argue the decisive lever against token maxing is the harness: the orchestration layer that assembles context, exposes tools, sequences turns, delegates work, and carries enterprise observability and governance. We isolate it with a controlled swap: 22 locked evaluation tasks, six foundation models (Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, Palmyra X6), changing only the orchestration layer -- a frozen conventional production loop versus the Writer Agent Harness. Holding models constant, the harness cuts blended cost per task 41% (0.21->0.12), median wall-clock 44% (48s->27s), and tokens per task 38% (14.2k->8.8k), with task-completion quality at parity (0.78->0.81, directional at this sample size). Efficiency is model-invariant -- every model gets cheaper (33-61%) -- while quality gains are capability-dependent: a model's gain correlates almost perfectly with its baseline strength (r=0.99, n=6), a phenomenon we term harness leverage. Quality per dollar rises 82%; task-completions per million tokens rise from 54.9 to 92.0. On this workload the orchestration layer moved cost per task more than the full spread of the model menu did. We formalize token economics at the orchestration layer (including effective input price under prompt caching), detail the six mechanism families behind the effect -- cache-shape discipline to failure-spend governance -- compare six widely used agent systems on the same axes, and argue the harness is the one component whose efficiency multiplies across every model an organization runs -- present and future.

Supervised fine-tuning (SFT) is widely used to adapt large language models to downstream tasks, but its effect on behavioral diversity in sequential decision-making remains under-explored. We study this question in a controlled suite of deterministic board games based on tic-tac-toe variants, where optimal actions are exactly computable and diversity can be measured directly. Across state-level evaluation, arena gameplay, and training trajectories, we find that reasoning-mode generation frequently suppresses action diversity without uniformly improving action accuracy. Furthermore, standard SFT improves accuracy but often induces premature diversity collapse, which exceeds what is minimally required by the accuracy-diversity tradeoff. We then show that action augmentation, which trains on all optimal actions per state rather than a single demonstrated action, would partially mitigates this effect. Our results identify narrow-support imitation as a source of policy collapse in LLM decision-making and suggest that preserving action support during SFT is important for maintaining exploratory behavior.

Large language models (LLMs) can solve every intermediate step of a multi-step math problem on its own and still fail the full problem, even when given a roadmap of the steps and all of their answers. We introduce OracleLadder, a diagnostic evaluation that locates where LLM math reasoning fails by giving the model increasing levels of oracle help. For each problem, a teacher model writes a fixed roadmap of intermediate sub-goals (milestones), and a deterministic symbolic verifier grades every answer. Testing the model with no help, with the roadmap, with the roadmap plus the milestone answers, and on each milestone alone sorts each failure into one of five reasoning gaps. On 354 NuminaMath problems and six models from 8B to 671B parameters (Qwen3, gpt-oss, Llama 3.3, DeepSeek-V3.1), the largest gap for every model is the composition gap, a stricter form of the compositionality gap. It covers 33-48% of problems, and 24-37% after removing problems that an LLM review flags as grading errors. Accuracy and milestone-help recovery rank the two strongest models differently, and two RLVR runs with similar accuracy gains move problems differently. The roadmap effect replicates on MATH500 and AIME 2024/25, per-problem recovery agrees for 83-87% of problems under an independent second teacher, and the help ladder carries over to code generation. We release the data, roadmaps, prompts, and code at https://github.com/slark-prime/OracleLadder.

Language is not only used to transmit neutral information; we often seek to persuade by arguing in favor of a particular view. Persuasion raises a number of challenges for classical accounts of belief updating, as information cannot be taken at face value. How should listeners account for a speaker's "hidden agenda" when incorporating new information? Here, we extend recent probabilistic models of recursive social reasoning to allow for persuasive goals and show that our model provides a pragmatic account for why weakly favorable arguments may backfire, a phenomenon known as the weak evidence effect. Critically, this model predicts a systematic relationship between belief updates and expectations about the information source: weak evidence should only backfire when speakers are expected to act under persuasive goals and prefer the strongest evidence. We introduce a simple experimental paradigm called the Stick Contest to measure the extent to which the weak evidence effect depends on speaker expectations, and show that a pragmatic listener model accounts for the empirical data better than alternative models. Our findings suggest further avenues for rational models of social reasoning to illuminate classical decision-making phenomena.

We present a conditional text generation framework that posits sentential expressions of possible causes and effects. This framework depends on two novel resources we develop in the course of this work: a very large-scale collection of English sentences expressing causal patterns CausalBank; and a refinement over previous work on constructing large lexical causal knowledge graphs Cause Effect Graph. Further, we extend prior work in lexically-constrained decoding to support disjunctive positive constraints. Human assessment confirms that our approach gives high-quality and diverse outputs. Finally, we use CausalBank to perform continued training of an encoder supporting a recent state-of-the-art model for causal reasoning, leading to a 3-point improvement on the COPA challenge set, with no change in model architecture.

This paper explores the impact of extending input lengths on the capabilities of Large Language Models (LLMs). Despite LLMs advancements in recent times, their performance consistency across different input lengths is not well understood. We investigate this aspect by introducing a novel QA reasoning framework, specifically designed to assess the impact of input length. We isolate the effect of input length using multiple versions of the same sample, each being extended with padding of different lengths, types and locations. Our findings show a notable degradation in LLMs' reasoning performance at much shorter input lengths than their technical maximum. We show that the degradation trend appears in every version of our dataset, although at different intensities. Additionally, our study reveals that traditional perplexity metrics do not correlate with performance of LLMs' in long input reasoning tasks. We analyse our results and identify failure modes that can serve as useful guides for future research, potentially informing strategies to address the limitations observed in LLMs.

Recent advances in visual reasoning (VR), particularly with the aid of Large Vision-Language Models (VLMs), show promise but require access to large-scale datasets and face challenges such as high computational costs and limited generalization capabilities. Compositional visual reasoning approaches have emerged as effective strategies; however, they heavily rely on the commonsense knowledge encoded in Large Language Models (LLMs) to perform planning, reasoning, or both, without considering the effect of their decisions on the visual reasoning process, which can lead to errors or failed procedures. To address these challenges, we introduce HYDRA, a multi-stage dynamic compositional visual reasoning framework designed for reliable and incrementally progressive general reasoning. HYDRA integrates three essential modules: a planner, a Reinforcement Learning (RL) agent serving as a cognitive controller, and a reasoner. The planner and reasoner modules utilize an LLM to generate instruction samples and executable code from the selected instruction, respectively, while the RL agent dynamically interacts with these modules, making high-level decisions on selection of the best instruction sample given information from the historical state stored through a feedback loop. This adaptable design enables HYDRA to adjust its actions based on previous feedback received during the reasoning process, leading to more reliable reasoning outputs and ultimately enhancing its overall effectiveness. Our framework demonstrates state-of-the-art performance in various VR tasks on four different widely-used datasets.

We introduce ReXTime, a benchmark designed to rigorously test AI models' ability to perform temporal reasoning within video events. Specifically, ReXTime focuses on reasoning across time, i.e. human-like understanding when the question and its corresponding answer occur in different video segments. This form of reasoning, requiring advanced understanding of cause-and-effect relationships across video segments, poses significant challenges to even the frontier multimodal large language models. To facilitate this evaluation, we develop an automated pipeline for generating temporal reasoning question-answer pairs, significantly reducing the need for labor-intensive manual annotations. Our benchmark includes 921 carefully vetted validation samples and 2,143 test samples, each manually curated for accuracy and relevance. Evaluation results show that while frontier large language models outperform academic models, they still lag behind human performance by a significant 14.3% accuracy gap. Additionally, our pipeline creates a training dataset of 9,695 machine generated samples without manual effort, which empirical studies suggest can enhance the across-time reasoning via fine-tuning.

Large Language Models (LLMs) have showcased exceptional ability in causal reasoning from textual information. However, will these causalities remain straightforward for Vision Large Language Models (VLLMs) when only visual hints are provided? Motivated by this, we propose a novel Multimodal Causal Reasoning benchmark, namely MuCR, to challenge VLLMs to infer semantic cause-and-effect relationship when solely relying on visual cues such as action, appearance, clothing, and environment. Specifically, we introduce a prompt-driven image synthesis approach to create siamese images with embedded semantic causality and visual cues, which can effectively evaluate VLLMs' causal reasoning capabilities. Additionally, we develop tailored metrics from multiple perspectives, including image-level match, phrase-level understanding, and sentence-level explanation, to comprehensively assess VLLMs' comprehension abilities. Our extensive experiments reveal that the current state-of-the-art VLLMs are not as skilled at multimodal causal reasoning as we might have hoped. Furthermore, we perform a comprehensive analysis to understand these models' shortcomings from different views and suggest directions for future research. We hope MuCR can serve as a valuable resource and foundational benchmark in multimodal causal reasoning research. The project is available at: https://github.com/Zhiyuan-Li-John/MuCR

This work revisits the dominant supervised fine-tuning (SFT) then reinforcement learning (RL) paradigm for training Large Vision-Language Models (LVLMs), and reveals a key finding: SFT can significantly undermine subsequent RL by inducing ``pseudo reasoning paths'' imitated from expert models. While these paths may resemble the native reasoning paths of RL models, they often involve prolonged, hesitant, less informative steps, and incorrect reasoning. To systematically study this effect, we introduce VLAA-Thinking, a new multimodal dataset designed to support reasoning in LVLMs. Constructed via a six-step pipeline involving captioning, reasoning distillation, answer rewrite and verification, VLAA-Thinking comprises high-quality, step-by-step visual reasoning traces for SFT, along with a more challenging RL split from the same data source. Using this dataset, we conduct extensive experiments comparing SFT, RL and their combinations. Results show that while SFT helps models learn reasoning formats, it often locks aligned models into imitative, rigid reasoning modes that impede further learning. In contrast, building on the Group Relative Policy Optimization (GRPO) with a novel mixed reward module integrating both perception and cognition signals, our RL approach fosters more genuine, adaptive reasoning behavior. Notably, our model VLAA-Thinker, based on Qwen2.5VL 3B, achieves top-1 performance on Open LMM Reasoning Leaderboard (https://huggingface.co/spaces/opencompass/Open_LMM_Reasoning_Leaderboard) among 4B scale LVLMs, surpassing the previous state-of-the-art by 1.8%. We hope our findings provide valuable insights in developing reasoning-capable LVLMs and can inform future research in this area.

Large language models (LLMs) face significant challenges in specialized biomedical tasks due to the inherent complexity of medical reasoning and the sensitive nature of clinical data. Existing LLMs often struggle with intricate medical terminology and the need for accurate clinical insights, leading to performance reduction when quantized for resource-constrained deployment. To address these issues, we propose Quantized Medical Tree of Thought (QM-ToT), a path-based reasoning framework. QM-ToT leverages a Tree of Thought (ToT) reasoning approach to decompose complex medical problems into manageable subtasks, coupled with evaluator assessment layers. This framework facilitates substantial performance improvements in INT4-quantized models on the challenging MedQAUSMLE dataset. Specifically, we demonstrate a remarkable accuracy increase from 34% to 50% for the LLaMA2-70b model and from 58.77% to 69.49% for LLaMA-3.1-8b. Besides, we also proposed an effect data distillation method based on ToT. Compared to the traditional distillation method, we achieved an improvement of 86. 27% while using only 3.9% of the data.This work, for the first time, showcases the potential of ToT to significantly enhance performance on complex biomedical tasks, establishing a crucial foundation for future advances in deploying high-performing quantized LLM in resource-limited medical settings.

Recent advances in Large Language Models (LLMs) have underscored the potential of Reinforcement Learning (RL) to facilitate the emergence of reasoning capabilities. Despite the encouraging results, a fundamental dilemma persists as RL improvement relies on learning from high-quality samples, yet the exploration for such samples remains bounded by the inherent limitations of LLMs. This, in effect, creates an undesirable cycle in which what cannot be explored cannot be learned. In this work, we propose Rubric-Scaffolded Reinforcement Learning (RuscaRL), a novel instructional scaffolding framework designed to break the exploration bottleneck for general LLM reasoning. Specifically, RuscaRL introduces checklist-style rubrics as (1) explicit scaffolding for exploration during rollout generation, where different rubrics are provided as external guidance within task instructions to steer diverse high-quality responses. This guidance is gradually decayed over time, encouraging the model to internalize the underlying reasoning patterns; (2) verifiable rewards for exploitation during model training, where we can obtain robust LLM-as-a-Judge scores using rubrics as references, enabling effective RL on general reasoning tasks. Extensive experiments demonstrate the superiority of the proposed RuscaRL across various benchmarks, effectively expanding reasoning boundaries under the best-of-N evaluation. Notably, RuscaRL significantly boosts Qwen-2.5-7B-Instruct from 23.6 to 50.3 on HealthBench-500, surpassing GPT-4.1. Furthermore, our fine-tuned variant on Qwen3-30B-A3B-Instruct achieves 61.1 on HealthBench-500, outperforming leading LLMs including OpenAI-o3.

The ripple effect poses a significant challenge in knowledge editing for large language models. Namely, when a single fact is edited, the model struggles to accurately update the related facts in a sequence, which is evaluated by multi-hop questions linked to a chain of related facts. Recent strategies have moved away from traditional parameter updates to more flexible, less computation-intensive methods, proven to be more effective in addressing the ripple effect. In-context learning (ICL) editing uses a simple demonstration `Imagine that + new fact` to guide LLMs, but struggles with complex multi-hop questions as the new fact alone fails to specify the chain of facts involved in such scenarios. Besides, memory-based editing maintains additional storage for all edits and related facts, requiring continuous updates to stay effective. As a result of these design limitations, the challenge remains, with the highest accuracy being only 33.8% on the MQuAKE-cf benchmarks for Vicuna-7B. To address this, we propose RippleCOT, a novel ICL editing approach integrating Chain-of-Thought (COT) reasoning. RippleCOT structures demonstrations as `newfact, question, thought, answer`, incorporating a thought component to identify and decompose the multi-hop logic within questions. This approach effectively guides the model through complex multi-hop questions with chains of related facts. Comprehensive experiments demonstrate that RippleCOT significantly outperforms the state-of-the-art on the ripple effect, achieving accuracy gains ranging from 7.8% to 87.1%.

Fine-tuning large language models (LLMs) on chain-of-thought (CoT) data shows that a small amount of high-quality data can outperform massive datasets. Yet, what constitutes "quality" remains ill-defined. Existing reasoning methods rely on indirect heuristics such as problem difficulty or trace length, while instruction-tuning has explored a broader range of automated selection strategies, but rarely in the context of reasoning. We propose to define reasoning data quality using influence functions, which measure the causal effect of individual CoT examples on downstream accuracy, and introduce influence-based pruning, which consistently outperforms perplexity and embedding-based baselines on math reasoning within a model family.

When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. Despite much work on CoT prompting, it is unclear if reasoning verbalized in a CoT is faithful to the models' parametric beliefs. We introduce a framework for measuring parametric faithfulness of generated reasoning, and propose Faithfulness by Unlearning Reasoning steps (FUR), an instance of this framework. FUR erases information contained in reasoning steps from model parameters, and measures faithfulness as the resulting effect on the model's prediction. Our experiments with four LMs and five multi-hop multi-choice question answering (MCQA) datasets show that FUR is frequently able to precisely change the underlying models' prediction for a given instance by unlearning key steps, indicating when a CoT is parametrically faithful. Further analysis shows that CoTs generated by models post-unlearning support different answers, hinting at a deeper effect of unlearning.

Despite achieving high accuracy on medical benchmarks, LLMs exhibit the Einstellung Effect in clinical diagnosis--relying on statistical shortcuts rather than patient-specific evidence, causing misdiagnosis in atypical cases. Existing benchmarks fail to detect this critical failure mode. We introduce MedEinst, a counterfactual benchmark with 5,383 paired clinical cases across 49 diseases. Each pair contains a control case and a "trap" case with altered discriminative evidence that flips the diagnosis. We measure susceptibility via Bias Trap Rate--probability of misdiagnosing traps despite correctly diagnosing controls. Extensive Evaluation of 17 LLMs shows frontier models achieve high baseline accuracy but severe bias trap rates. Thus, we propose ECR-Agent, aligning LLM reasoning with Evidence-Based Medicine standard via two components: (1) Dynamic Causal Inference (DCI) performs structured reasoning through dual-pathway perception, dynamic causal graph reasoning across three levels (association, intervention, counterfactual), and evidence audit for final diagnosis; (2) Critic-Driven Graph and Memory Evolution (CGME) iteratively refines the system by storing validated reasoning paths in an exemplar base and consolidating disease-specific knowledge into evolving illness graphs. Source code is to be released.

Large language models (LLMs) achieve strong reasoning performance by allocating substantial computation at inference time, often generating long and verbose reasoning traces. While recent work on efficient reasoning reduces this overhead through length-based rewards or pruning, many approaches are post-trained under a much shorter context window than base-model training, a factor whose effect has not been systematically isolated. We first show that short-context post-training alone, using standard GRPO without any length-aware objective, already induces substantial reasoning compression-but at the cost of increasingly unstable training dynamics and accuracy degradation. To address this, we propose Step-level Advantage Selection (SAS), which operates at the reasoning-step level and assigns a zero advantage to low-confidence steps in correct rollouts and to high-confidence steps in verifier-failed rollouts, where failures often arise from truncation or verifier issues rather than incorrect reasoning. Across diverse mathematical and general reasoning benchmarks, SAS improves average Pass@1 accuracy by 0.86 points over the strongest length-aware baseline while reducing average reasoning length by 16.3%, yielding a better accuracy-efficiency trade-off.

Large language models (LLMs) possess strong semantic understanding, driving significant progress in data mining applications. This is further enhanced by large reasoning models (LRMs), which provide explicit multi-step reasoning traces. On the other hand, the growing need for the right to be forgotten has driven the development of machine unlearning techniques, which aim to eliminate the influence of specific data from trained models without full retraining. However, unlearning may also introduce new security vulnerabilities by exposing additional interaction surfaces. Although many studies have investigated unlearning attacks, there is no prior work on LRMs. To bridge the gap, we first in this paper propose LRM unlearning attack that forces incorrect final answers while generating convincing but misleading reasoning traces. This objective is challenging due to non-differentiable logical constraints, weak optimization effect over long rationales, and discrete forget set selection. To overcome these challenges, we introduce a bi-level exact unlearning attack that incorporates a differentiable objective function, influential token alignment, and a relaxed indicator strategy. To demonstrate the effectiveness and generalizability of our attack, we also design novel optimization frameworks and conduct comprehensive experiments in both white-box and black-box settings, aiming to raise awareness of the emerging threats to LRM unlearning pipelines.

We present a machine-checked formalization of structurally governed AI workflow architectures and prove that effect-level governance can be imposed without reducing internal computational expressivity. Using Interaction Trees in Rocq 8.19, we define a governance operator G that mediates all effectful directives, including memory access, external calls, and oracle (LLM) queries. Our development compiles with 0 admitted lemmas and consists of 36 modules, ~12,000 lines of Rocq, and 454 theorems. We establishseven properties: (P1) governed Turing completeness, (P2) governed oracle expressivity, (P3) a decidability boundary in which governance predicates are total and closed under Boolean composition while semantic program properties remain non-trivial and undecidable by governance, (P4) goal preservation for permitted executions, (P5) expressive minimality of primitive capabilities (compute, memory, reasoning, external call, observability), (P6) subsumption asymmetry showing structural governance strictly subsumes content-level filtering, and (P7) semantic transparency: on all executions where governance permits, the governed interpretation is observationally equivalent (modulo governance-only events) to the ungoverned interpretation. Together, these results show that governance and computational expressivity are orthogonal dimensions: governance constrains the effect boundary of programs while remaining semantically transparent to internal computation.

Reasoning-trained language models often spend more tokens on harder problems, but longer chains of thought do not show whether a model is merely computing for more steps or following a different internal trajectory. We study this distinction through hidden-state trajectories during chain-of-thought generation across competitive programming, mathematics, and Boolean satisfiability. Raw trajectory geometry is strongly shaped by generation length: longer generations mechanically alter path statistics, so difficulty-dependent comparisons are misleading without adjustment. After residualizing trajectory statistics on length, difficulty remains systematically coupled to corrected trajectory geometry across all domains studied. The clearest reasoning-specific separation appears in the code domain, where harder problems show more direct corrected trajectories and less heterogeneous local curvature in reasoning-trained models than in matched instruction-tuned baselines. Corrected difficulty-geometry coupling is weaker, but still present, in mathematics and Boolean satisfiability. Prompt-stage linear probes do not mirror the code-domain separation, and behavioral annotations show that stronger corrected coupling co-occurs with strategy shifts and uncertainty monitoring. Together, these findings establish length correction as a prerequisite for generation-time trajectory analysis and show that reasoning training can be associated with distinct corrected trajectory geometry, with the strength of the effect depending on the domain.

Agent harnesses have become the operational infrastructure of modern large language model agents, coordinating context, tools, verification, and execution control to translate latent model capability into reliable long-horizon behavior. However, reliable long-horizon behavior requires harness control to adapt to task demands, execution environments, and evolving execution states, whereas current harnesses predominantly rely on hand-crafted or globally fixed policies; this mismatch manifests as unnecessary computational overhead and, in adverse cases, reduced task success. To address this limitation, we formulate the task of enabling adaptive orchestration in harness systems as a causal learning problem and propose Counterfactual Harness Intervention Learning for Long-Horizon Agents (CHILL-Harness). CHILL-Harness intervenes at the orchestration layer to enable advantage-guided workflow adaptation, thereby improving reasoning and execution efficiency while preserving task performance. Specifically, we develop causal intervention effect learning as the effect-estimation component of CHILL-Harness to estimate intervention-relative workflow advantage from confidence-weighted execution evidence and identify advantageous workflow adaptations. We further introduce advantage-realizing causal orchestration as its realization component to adaptively allocate counterfactual reasoning and realize only workflow adjustments supported by sufficient expected advantage. Finally, we incorporate a success-preserving objective and advantage-margin authorization constraints into CHILL-Harness to promote reliable adaptation. Extensive experiments on heterogeneous long-horizon tasks spanning information seeking, software engineering, and terminal interaction show that CHILL-Harness consistently preserves or improves task success while substantially reducing token consumption and execution time.

The prevailing approach to improving robot foundation models (RFMs) relies on larger models, more robot demonstrations, and costly training at scale. We show that there exists an effective and efficient complementary approach: the right reasoning recipe can substantially improve the zero-shot task performance of existing state-of-the-art RFMs. We refer to this recipe as ARC. It consists of three key ingredients: a reasoning trace, a scalable automatic labeling pipeline, and a strategy for adapting pretrained RFMs to use these traces for control. First, we find that effective reasoning traces should be grounded in the robot's next action and explain its causal structure: why the action is appropriate and what effect it should produce. Second, we show that these traces can be generated automatically from existing demonstrations, enabling us to construct ARC-Trace-DROID from DROID without collecting new robot data. Third, we show how state-of-the-art VLAs such as π_{0.5} and WAMs such as Cosmos3-Nano-Policy can learn to use these traces for control, with fine-tuning and inference tailored to each model's architecture and capabilities. Using ARC, we obtain gains in zero-shot RFM performance that, to our knowledge, are unprecedented without additional robot demonstrations or foundation-scale training. The adapted models establish a new state of the art on RoboLab-120 and MolmoSpaces, with gains of up to 50 percentage points on RoboLab-Reasoning-50. On real robots, ARC improves π_{0.5}'s task success by 82.2 percentage points. Project website: https://arc-robot-reasoning.github.io/

The success of chain-of-thought (CoT) in complex reasoning can be attributed to a textual scaffolding effect: explicit intermediate results can be examined, reused, and composed to support subsequent reasoning, much as scaffolding supports a building under construction. However, mainstream multimodal LLMs lack an equivalent scaffold for visual distinctions that are difficult to express in concise verbal descriptions. We therefore ask: Can native visual representations become first-class citizens alongside language, providing a visual scaffold for reasoning? We define native multimodal reasoning as reasoning in which native visual representations and language jointly carry reusable intermediate states. As a concrete implementation, the visual scaffolding mechanism we proposed lets the MLLM autonomously initiate a retrieval action at any reasoning step. A lightweight retriever selects native visual features, which are then incorporated directly into the evolving CoT and remain available to subsequent steps. To evaluate this mechanism under controlled visual demands, we introduce NMR-Bench, comprising Burden Counting and Path Tracing. With 88M added parameters (approximately 2%) on Qwen3-VL-4B-Thinking, our model achieves 66.43% and 72.50% accuracy, respectively, surpassing the evaluated GPT-5.5 and Gemini 3.1 Pro baselines. Controlled interventions show that changing selected visual content affects task accuracy and can redirect subsequent visual selection even when intervening text is held fixed. These findings show that native visual states can function as intermediate reasoning material and point toward reusable scaffolds for additional modalities.

Videos often capture objects, their visible properties, their motion, and the interactions between different objects. Objects also have physical properties such as mass, which the imaging pipeline is unable to directly capture. However, these properties can be estimated by utilizing cues from relative object motion and the dynamics introduced by collisions. In this paper, we introduce CRIPP-VQA, a new video question answering dataset for reasoning about the implicit physical properties of objects in a scene. CRIPP-VQA contains videos of objects in motion, annotated with questions that involve counterfactual reasoning about the effect of actions, questions about planning in order to reach a goal, and descriptive questions about visible properties of objects. The CRIPP-VQA test set enables evaluation under several out-of-distribution settings -- videos with objects with masses, coefficients of friction, and initial velocities that are not observed in the training distribution. Our experiments reveal a surprising and significant performance gap in terms of answering questions about implicit properties (the focus of this paper) and explicit properties of objects (the focus of prior work).

Generating a Chain of Thought (CoT) has been shown to consistently improve large language model (LLM) performance on a wide range of NLP tasks. However, prior work has mainly focused on logical reasoning tasks (e.g. arithmetic, commonsense QA); it remains unclear whether improvements hold for more diverse types of reasoning, especially in socially situated contexts. Concretely, we perform a controlled evaluation of zero-shot CoT across two socially sensitive domains: harmful questions and stereotype benchmarks. We find that zero-shot CoT reasoning in sensitive domains significantly increases a model's likelihood to produce harmful or undesirable output, with trends holding across different prompt formats and model variants. Furthermore, we show that harmful CoTs increase with model size, but decrease with improved instruction following. Our work suggests that zero-shot CoT should be used with caution on socially important tasks, especially when marginalized groups or sensitive topics are involved.

Large language models (LLMs) are capable of answering knowledge-intensive complex questions with chain-of-thought (CoT) reasoning. However, they tend to generate factually incorrect reasoning steps when the required knowledge is not available or up-to-date in models' parameters. Recent works turn to retrieving external knowledge to augment CoT reasoning. Despite being promising, these chain-based methods suffer from: 1) Negative retrieval. Unnecessary or incorrect retrieval may mislead the reasoning; 2) Limited sight. Lacking the ability to look backward or forward, a local error in one step will propagate along the chain. In this paper, we propose a novel approach: Probabilistic Tree-of-thought Reasoning (ProbTree). First, LLMs translate a complex question into a query tree, in which each non-root node denotes a sub-question of its parent node. Then, probabilistic reasoning is conducted over the tree, by solving questions from leaf to root considering the confidence of both question decomposing and answering. During reasoning, for leaf nodes, LLMs choose a more confident answer from Closed-book QA that employs parametric knowledge and Open-book QA that employs retrieved external knowledge, thus eliminating the negative retrieval problem. For non-leaf nodes, with the hierarchical structure, LLMs have broader sights and are able to globally reason with the information from child nodes, thus recovering from local errors. The experiments on three Complex QA datasets under the open-domain setting show that our approach outperforms SOTA methods significantly, demonstrating the effect of probabilistic tree-of-thought reasoning.

Recent works have showcased the ability of LLMs to embody diverse personas in their responses, exemplified by prompts like 'You are Yoda. Explain the Theory of Relativity.' While this ability allows personalization of LLMs and enables human behavior simulation, its effect on LLMs' capabilities remains unclear. To fill this gap, we present the first extensive study of the unintended side-effects of persona assignment on the ability of LLMs to perform basic reasoning tasks. Our study covers 24 reasoning datasets, 4 LLMs, and 19 diverse personas (e.g. an Asian person) spanning 5 socio-demographic groups. Our experiments unveil that LLMs harbor deep rooted bias against various socio-demographics underneath a veneer of fairness. While they overtly reject stereotypes when explicitly asked ('Are Black people less skilled at mathematics?'), they manifest stereotypical and erroneous presumptions when asked to answer questions while adopting a persona. These can be observed as abstentions in responses, e.g., 'As a Black person, I can't answer this question as it requires math knowledge', and generally result in a substantial performance drop. Our experiments with ChatGPT-3.5 show that this bias is ubiquitous - 80% of our personas demonstrate bias; it is significant - some datasets show performance drops of 70%+; and can be especially harmful for certain groups - some personas suffer statistically significant drops on 80%+ of the datasets. Overall, all 4 LLMs exhibit this bias to varying extents, with GPT-4-Turbo showing the least but still a problematic amount of bias (evident in 42% of the personas). Further analysis shows that these persona-induced errors can be hard-to-discern and hard-to-avoid. Our findings serve as a cautionary tale that the practice of assigning personas to LLMs - a trend on the rise - can surface their deep-rooted biases and have unforeseeable and detrimental side-effects.

Large language models (LLMs) have revolutionized many areas (e.g. natural language processing, software engineering, etc.) by achieving state-of-the-art performance on extensive downstream tasks. Aiming to achieve robust and general artificial intelligence, there has been a surge of interest in investigating the reasoning ability of the LLMs. Whereas the textual and numerical reasoning benchmarks adopted by previous works are rather shallow and simple, it is hard to conclude that the LLMs possess strong reasoning ability by merely achieving positive results on these benchmarks. Recent efforts have demonstrated that the LLMs are poor at solving sequential decision-making problems that require common-sense planning by evaluating their performance on the reinforcement learning benchmarks. In this work, we conduct an in-depth assessment of several state-of-the-art LLMs' reasoning ability based on the inductive logic programming (ILP) benchmark, which is broadly recognized as a representative and challenging measurement for evaluating logic program induction/synthesis systems as it requires inducing strict cause-effect logic to achieve robust deduction on independent and identically distributed (IID) and out-of-distribution (OOD) test samples. Our evaluations illustrate that compared with the neural program induction systems which are much smaller in model size, the state-of-the-art LLMs are much poorer in terms of reasoning ability by achieving much lower performance and generalization using either natural language prompting or truth-value matrix prompting.

This is the first work to look at the application of large language models (LLMs) for the purpose of model space edits in automated planning tasks. To set the stage for this union, we explore two different flavors of model space problems that have been studied in the AI planning literature and explore the effect of an LLM on those tasks. We empirically demonstrate how the performance of an LLM contrasts with combinatorial search (CS) -- an approach that has been traditionally used to solve model space tasks in planning, both with the LLM in the role of a standalone model space reasoner as well as in the role of a statistical signal in concert with the CS approach as part of a two-stage process. Our experiments show promising results suggesting further forays of LLMs into the exciting world of model space reasoning for planning tasks in the future.

Instruction tuning is an important step in making language models useful for direct user interaction. However, many legal tasks remain out of reach for most open LLMs and there do not yet exist any large scale instruction datasets for the domain. This critically limits research in this application area. In this work, we curate LawInstruct, a large legal instruction dataset, covering 17 jurisdictions, 24 languages and a total of 12M examples. We present evidence that domain-specific pretraining and instruction tuning improve performance on LegalBench, including improving Flan-T5 XL by 8 points or 16\% over the baseline. However, the effect does not generalize across all tasks, training regimes, model sizes, and other factors. LawInstruct is a resource for accelerating the development of models with stronger information processing and decision making capabilities in the legal domain.

Fully comprehending scientific papers by machines reflects a high level of Artificial General Intelligence, requiring the ability to reason across fragmented and heterogeneous sources of information, presenting a complex and practically significant challenge. While Vision-Language Models (VLMs) have made remarkable strides in various tasks, particularly those involving reasoning with evidence source from single image or text page, their ability to use cross-source information for reasoning remains an open problem. This work presents MMCR, a high-difficulty benchmark designed to evaluate VLMs' capacity for reasoning with cross-source information from scientific papers. The benchmark comprises 276 high-quality questions, meticulously annotated by humans across 7 subjects and 10 task types. Experiments with 18 VLMs demonstrate that cross-source reasoning presents a substantial challenge for existing models. Notably, even the top-performing model, GPT-4o, achieved only 48.55% overall accuracy, with only 20% accuracy in multi-table comprehension tasks, while the second-best model, Qwen2.5-VL-72B, reached 39.86% overall accuracy. Furthermore, we investigated the impact of the Chain-of-Thought (CoT) technique on cross-source reasoning and observed a detrimental effect on small models, whereas larger models demonstrated substantially enhanced performance. These results highlight the pressing need to develop VLMs capable of effectively utilizing cross-source information for reasoning.

The application of large language models (LLMs) in domain-specific contexts, including finance, has expanded rapidly. Domain-specific LLMs are typically evaluated based on their performance in various downstream tasks relevant to the domain. In this work, we present a detailed analysis of fine-tuning LLMs for such tasks. Somewhat counterintuitively, we find that in domain-specific cases, fine-tuning exclusively on the target task is not always the most effective strategy. Instead, multi-task finetuning - where models are trained on a cocktail of related tasks - can significantly enhance performance. We demonstrate how this approach enables a small model, such as Phi-3-Mini, to achieve state-of-the-art results, even surpassing the much larger GPT-4-o model on financial benchmarks. Our study involves a large-scale experiment, conducting over 200 training experiments using several widely adopted LLMs as baselines, and empirically confirms the benefits of multi-task fine-tuning. Additionally, we explore the use of general instruction data as a form of regularization, suggesting that it helps minimize performance degradation. We also investigate the inclusion of mathematical data, finding improvements in numerical reasoning that transfer effectively to financial tasks. Finally, we note that while fine-tuning for downstream tasks leads to targeted improvements in task performance, it does not necessarily result in broader gains in domain knowledge or complex domain reasoning abilities.

Large reasoning models (LRMs) have demonstrated impressive performance across a range of reasoning tasks, yet little is known about their internal reasoning processes in multilingual settings. We begin with a critical question: {\it In which language do these models reason when solving problems presented in different languages?} Our findings reveal that, despite multilingual training, LRMs tend to default to reasoning in high-resource languages (e.g., English) at test time, regardless of the input language. When constrained to reason in the same language as the input, model performance declines, especially for low-resource languages. In contrast, reasoning in high-resource languages generally preserves performance. We conduct extensive evaluations across reasoning-intensive tasks (MMMLU, MATH-500) and non-reasoning benchmarks (CulturalBench, LMSYS-toxic), showing that the effect of language choice varies by task type: input-language reasoning degrades performance on reasoning tasks but benefits cultural tasks, while safety evaluations exhibit language-specific behavior. By exposing these linguistic biases in LRMs, our work highlights a critical step toward developing more equitable models that serve users across diverse linguistic backgrounds.

Recent advances in post-training techniques have endowed Large Language Models (LLMs) with enhanced capabilities for tackling complex, logic-intensive tasks through the generation of supplementary planning tokens. This development raises a fundamental question: Are these models aware of what they "learn" and "think"? To address this, we define three core competencies: (1) awareness of learned latent policies, (2) generalization of these policies across domains, and (3) alignment between internal reasoning traces and final outputs. We empirically evaluate these abilities on several tasks, each designed to require learning a distinct policy. Furthermore, we contrast the profiles of models post-trained via Supervised Fine-Tuning (SFT), Direct Policy Optimization (DPO), and Group Relative Policy Optimization (GRPO). Our findings indicate that RL-trained models not only demonstrate greater awareness of their learned behaviors and stronger generalizability to novel, structurally similar tasks than SFT models but also often exhibit weak alignment between their reasoning traces and final outputs, an effect most pronounced in GRPO-trained models.

We characterize how memorization is represented in transformer models and show that it can be disentangled in the weights of both language models (LMs) and vision transformers (ViTs) using a decomposition based on the loss landscape curvature. This insight is based on prior theoretical and empirical work showing that the curvature for memorized training points is much sharper than non memorized, meaning ordering weight components from high to low curvature can reveal a distinction without explicit labels. This motivates a weight editing procedure that suppresses far more recitation of untargeted memorized data more effectively than a recent unlearning method (BalancedSubnet), while maintaining lower perplexity. Since the basis of curvature has a natural interpretation for shared structure in model weights, we analyze the editing procedure extensively on its effect on downstream tasks in LMs, and find that fact retrieval and arithmetic are specifically and consistently negatively affected, even though open book fact retrieval and general logical reasoning is conserved. We posit these tasks rely heavily on specialized directions in weight space rather than general purpose mechanisms, regardless of whether those individual datapoints are memorized. We support this by showing a correspondence between task data's activation strength with low curvature components that we edit out, and the drop in task performance after the edit. Our work enhances the understanding of memorization in neural networks with practical applications towards removing it, and provides evidence for idiosyncratic, narrowly-used structures involved in solving tasks like math and fact retrieval.

Self-consistency has emerged as a popular technique for improving large language model accuracy on reasoning tasks. The approach is straightforward: generate multiple reasoning paths and select the most common answer through majority voting. While this reliably boosts accuracy, it remains unclear whether these gains reflect genuine improvements in reasoning quality. We investigate a fundamental question that has not been studied before: does inference scaling improve reasoning faithfulness? We conduct a comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, and DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems. Our analysis employs bootstrap confidence intervals, McNemar's tests for paired comparisons, and Cohen's d effect sizes to quantify the effects rigorously. The results reveal striking differences across models that challenge common assumptions about self-consistency. GPT-5.2 shows the expected pattern: accuracy improves from 78% to 90% at N=5, with faithfulness remaining relatively stable (0.540 to 0.510). Claude Opus 4.5 tells a completely different story. Its accuracy actually drops from 78% to 74.3% while faithfulness jumps dramatically from 0.270 to 0.891 at N=5. DeepSeek-v3.2, already at 98% accuracy, shows ceiling effects with modest faithfulness gains (0.440 to 0.541). Gemini-3-flash improves from 81% to 86% accuracy with a slight faithfulness decrease (0.260 to 0.212). Problem difficulty analysis reveals that GPT-5.2 solves 82% of hard problems while breaking only 13% of easy ones. Claude, in contrast, breaks 23% of easy problems, explaining its accuracy decrease. These findings matter for practitioners: self-consistency is not universally beneficial, and teams should test their specific models before deployment. We release our code and provide practical recommendations for navigating these tradeoffs.

Quantum Zeno Effect (QZE) has been one of the most interesting phenomena in quantum mechanics ever since its discovery in 1977 by Misra and Sudarshan [J. Math. Phys. 18, 756 (1977)]. There have been many attempts for experimental realization of the same. Here, we present the first ever simulation of QZE on IBM quantum experience platform. We simulate a two-level system for Rabi-driven oscillation and then disturb the time evolution by intermediate repetitive measurements using quantum gates to increase the survival probability of the qubit in the initial state. The circuits are designed along with the added intermediate measurements and executed on IBM quantum simulator, and the outcomes are shown to be consistent with the predictions. The increasing survival probability with the number of intermediate measurements demonstrates QZE. Furthermore, some alternative explanations for the obtained results are provided which leads to some ambiguity in giving the exact reasoning for the observed outcomes.

Most humans use visual imagination to understand and reason about language, but models such as BERT reason about language using knowledge acquired during text-only pretraining. In this work, we investigate whether vision-and-language pretraining can improve performance on text-only tasks that involve implicit visual reasoning, focusing primarily on zero-shot probing methods. We propose a suite of visual language understanding (VLU) tasks for probing the visual reasoning abilities of text encoder models, as well as various non-visual natural language understanding (NLU) tasks for comparison. We also contribute a novel zero-shot knowledge probing method, Stroop probing, for applying models such as CLIP to text-only tasks without needing a prediction head such as the masked language modelling head of models like BERT. We show that SOTA multimodally trained text encoders outperform unimodally trained text encoders on the VLU tasks while being underperformed by them on the NLU tasks, lending new context to previously mixed results regarding the NLU capabilities of multimodal models. We conclude that exposure to images during pretraining affords inherent visual reasoning knowledge that is reflected in language-only tasks that require implicit visual reasoning. Our findings bear importance in the broader context of multimodal learning, providing principled guidelines for the choice of text encoders used in such contexts.

Test-Time Scaling (TTS) has significantly enhanced the capabilities of Large Reasoning Models (LRMs) but introduces a critical side-effect known as Overthinking. We conduct a preliminary study to rethink this phenomenon from a fine-grained perspective. We observe that LRMs frequently conduct repetitive self-verification without revision even after obtaining the final answer during the reasoning process. We formally define this specific position where the answer first stabilizes as the Reasoning Anchor. By analyzing pre- and post-anchor reasoning behaviors, we uncover the structural redundancy fixed in LRMs: the meaningless repetitive verification after deriving the first complete answer, which we term the Answer-Stable Tail (AST). Motivated by this observation, we propose Anchor-based Process Reward (APR), a structure-aware reward shaping method that localizes the reasoning anchor and penalizes exclusively the post-anchor AST. Leveraging the policy optimization algorithm suitable for length penalties, our APR models achieved the performance-efficiency Pareto frontier at 1.5B and 7B scales averaged across five mathematical reasoning datasets while requiring substantially fewer computational resources for RL training.

Latent visual reasoning aims to mimic human's imagination process by meditating through hidden states of Multimodal Large Language Models. While recognized as a promising paradigm for visual reasoning, the underlying mechanisms driving its effectiveness remain unclear. Motivated to demystify the true source of its efficacy, we investigate the validity of latent reasoning using Causal Mediation Analysis. We model the process as a causal chain: the input as the treatment, the latent tokens as the mediator, and the final answer as the outcome. Our findings uncover two critical disconnections: (a) Input-Latent Disconnect: dramatic perturbations on the input result in negligible changes to the latent tokens, suggesting that latent tokens do not effectively attend to the input sequence. (b) Latent-Answer Disconnect: perturbations on the latent tokens yield minimal impact on the final answer, indicating the limited causal effect latent tokens imposing on the outcome. Furthermore, extensive probing analysis reveals that latent tokens encode limited visual information and exhibit high similarity. Consequently, we challenge the necessity of latent reasoning and propose a straightforward alternative named CapImagine, which teaches the model to explicitly imagine using text. Experiments on vision-centric benchmarks show that CapImagine significantly outperforms complex latent-space baselines, highlighting the superior potential of visual reasoning through explicit imagination.

Recent soft prompt research has tried to improve reasoning by inserting trained vectors into LLM inputs, yet whether the gain comes from the learned content or from the act of injection itself has not been carefully separated. We study Random Soft Prompts (RSPs), which drop the training step entirely and append a freshly drawn sequence of random embedding vectors to the input. Each RSP vector is sampled from an isotropic Gaussian fitted to the entrywise mean and variance of the pretrained embedding table; the sequence carries no learned content, and yet reaches accuracy comparable to optimized soft prompts on math reasoning benchmarks in several settings. The mechanism unfolds in two stages: because attention has to absorb a never-seen-before random position, the distribution over the first few generated tokens flattens and reasoning trajectories branch, and as generation continues this influence dilutes naturally so the response commits to a single completion. We show that during inference RSPs lift early-stage token diversity and, combined with temperature sampling, widen Pass@N, the probability that at least one out of N attempts is correct. Beyond inference, we carry the same effect into DAPO training and demonstrate practical gains. Our contributions are: (i) RSP isolates the simplest form of soft prompt -- training-free, freshly resampled -- providing a unified lens for the structural effect of injection that variants otherwise differing in training and form all share; (ii) a theoretical and empirical validation of the underlying mechanism; and (iii) an extension from inference to training.

Self-consistency assumes the most frequent answer among sampled reasoning traces is the most reliable, but this can fail in causal reasoning: samples often repeat the same confounding error, and votes fragment across multiple valid answers, letting an invalid answer win despite a valid minority trace. We introduce CALVER (Causal Axiom-Level VERification), a training-free symbolic verifier that scores structured traces against Pearl's causal criteria, including -separation, backdoor adjustment, and intervention, and selects the highest-scoring candidate without consulting a reference answer. On CLEAR find-one-valid queries that admit multiple graph-valid answers, CALVER reaches 42.1% where plurality, a reward model, an LLM judge, and model confidence remain near 30% on identical frozen pools. Scaling the judge to 72B does not close the gap. In an audited clean-core subset, 11 of 21 graph-valid CALVER selections differ from the benchmark's listed answer while still satisfying the requested predicate. The advantage widens with the sampling budget and reproduces across ten published Bayesian networks, a second model family, and settings where the model must build the graph from text. CALVER also improves thresholded average-treatment-effect decisions against exact ground truth, generalizes to logic under a truth-table checker, and scores each candidate in milliseconds on CPU. CALVER needs only a causal structure, supplied outright or built from the text; wherever that holds, selection can aggregate via causal validity.

Activation steering has become a widely used approach for controlling language models during explicit chain-of-thought (CoT) reasoning, motivating its extension to latent CoT. However, we find that steering continuous thoughts produces substantially weaker effects on subsequent language generation than steering explicit CoT, even when the hidden representations are moved by comparable amounts. We first show that task information remains identifiable in continuous thoughts. Hence, we hypothesize a latent-to-language transition gap, in which an intervention effect in latent space fails to transfer to language generation. Two further results support this hypothesis: the output distribution changes abruptly at the transition boundary, and task-related directions exert much weaker bidirectional control in latent CoT than in explicit CoT. These findings identify the transition interface as a central target for evaluating and designing future latent-steering methods.

Multi-path reasoning methods such as self-consistency (SC) sample K reasoning paths and choose the most frequent answer. However, their gains quickly plateau as K increases, and existing methods do not predict when this saturation will occur. We formalize multi-path LLM reasoning as a diversity combining problem from wireless communications: each path is a noisy channel observation, and the pairwise correlation of path correctness caps the design-effect effective sample size of the vote at a finite ceiling. Generalized least squares (GLS) analysis shows that, under exchangeability, the optimal symmetric linear combiner of latent embeddings is uniform, supporting majority vote as the natural default in standard SC while leaving room for weighting or pruning under heterogeneous prompt-template branches. Across 5 models and 12 benchmarks, prompt-template diversity reduces path correlation in 55 of 57 valid cells, with the strongest effect on open-ended QA. We derive an Adaptive-K rule that uses a four-path pilot to select K^*, retaining 96--103% of MV@K{=}32 accuracy across Math, QA, and NLU.

This paper introduces a novel type theory and logic for probabilistic reasoning. Its logic is quantitative, with fuzzy predicates. It includes normalisation and conditioning of states. This conditioning uses a key aspect that distinguishes our probabilistic type theory from quantum type theory, namely the bijective correspondence between predicates and side-effect free actions (called instrument, or assert, maps). The paper shows how suitable computation rules can be derived from this predicate-action correspondence, and uses these rules for calculating conditional probabilities in two well-known examples of Bayesian reasoning in (graphical) models. Our type theory may thus form the basis for a mechanisation of Bayesian inference.

Pre-trained neural language models give high performance on natural language inference (NLI) tasks. But whether they actually understand the meaning of the processed sequences remains unclear. We propose a new diagnostics test suite which allows to assess whether a dataset constitutes a good testbed for evaluating the models' meaning understanding capabilities. We specifically apply controlled corruption transformations to widely used benchmarks (MNLI and ANLI), which involve removing entire word classes and often lead to non-sensical sentence pairs. If model accuracy on the corrupted data remains high, then the dataset is likely to contain statistical biases and artefacts that guide prediction. Inversely, a large decrease in model accuracy indicates that the original dataset provides a proper challenge to the models' reasoning capabilities. Hence, our proposed controls can serve as a crash test for developing high quality data for NLI tasks.

While there are numerous benchmarks comparing the performance of modern language models (LMs), end-task evaluations often conflate notions of *factual accuracy* ("truth") and *reasoning ability* ("rationality", or "honesty" in the sense of correctly reporting implications of beliefs). Our goal is a dataset that clearly distinguishes these two notions. Our approach is to leverage and extend a collection of human-annotated *entailment trees*, engineered to express both good and bad chains of reasoning, and using a mixture of true and false facts, in particular including counterfactual examples, to avoid belief bias (also known as the "content effect"). The resulting dataset, called BaRDa, contains 3000 entailments (1787 valid, 1213 invalid), using 6681 true and 2319 false statements. Testing on four GPT-series models, GPT3(curie)/GPT3(davinici)/3.5/4, we find factual accuracy (truth) scores of 74.1/80.6/82.6/87.1 and reasoning accuracy scores of 63.1/78.0/71.8/79.2. This shows the clear progression of models towards improved factual accuracy and entailment reasoning, and the dataset provides a new benchmark that more cleanly separates and quantifies these two notions.

We introduce WIQA, the first large-scale dataset of "What if..." questions over procedural text. WIQA contains three parts: a collection of paragraphs each describing a process, e.g., beach erosion; a set of crowdsourced influence graphs for each paragraph, describing how one change affects another; and a large (40k) collection of "What if...?" multiple-choice questions derived from the graphs. For example, given a paragraph about beach erosion, would stormy weather result in more or less erosion (or have no effect)? The task is to answer the questions, given their associated paragraph. WIQA contains three kinds of questions: perturbations to steps mentioned in the paragraph; external (out-of-paragraph) perturbations requiring commonsense knowledge; and irrelevant (no effect) perturbations. We find that state-of-the-art models achieve 73.8% accuracy, well below the human performance of 96.3%. We analyze the challenges, in particular tracking chains of influences, and present the dataset as an open challenge to the community.

Despite recent progress in large-scale reinforcement learning (RL) for reasoning, the training recipe for building high-performing reasoning models remains elusive. Key implementation details of frontier models, such as DeepSeek-R1, including data curation strategies and RL training recipe, are often omitted. Moreover, recent research indicates distillation remains more effective than RL for smaller models. In this work, we demonstrate that large-scale RL can significantly enhance the reasoning capabilities of strong, small- and mid-sized models, achieving results that surpass those of state-of-the-art distillation-based models. We systematically study the RL training process through extensive ablations and propose a simple yet effective approach: first training on math-only prompts, then on code-only prompts. Notably, we find that math-only RL not only significantly enhances the performance of strong distilled models on math benchmarks (e.g., +14.6% / +17.2% on AIME 2025 for the 7B / 14B models), but also code reasoning tasks (e.g., +6.8% / +5.8% on LiveCodeBench for the 7B / 14B models). In addition, extended code-only RL iterations further improve performance on code benchmarks with minimal or no degradation in math results. We develop a robust data curation pipeline to collect challenging prompts with high-quality, verifiable answers and test cases to enable verification-based RL across both domains. Finally, we identify key experimental insights, including curriculum learning with progressively increasing response lengths and the stabilizing effect of on-policy parameter updates. We find that RL not only elicits the foundational reasoning capabilities acquired during pretraining and supervised fine-tuning (e.g., distillation), but also pushes the limits of the model's reasoning ability, enabling it to solve problems that were previously unsolvable.

Biomedical researchers increasingly rely on large-scale structured databases for complex analytical tasks. However, current text-to-SQL systems often struggle to map qualitative scientific questions into executable SQL, particularly when implicit domain reasoning is required. We introduce BiomedSQL, the first benchmark explicitly designed to evaluate scientific reasoning in text-to-SQL generation over a real-world biomedical knowledge base. BiomedSQL comprises 68,000 question/SQL query/answer triples grounded in a harmonized BigQuery knowledge base that integrates gene-disease associations, causal inference from omics data, and drug approval records. Each question requires models to infer domain-specific criteria, such as genome-wide significance thresholds, effect directionality, or trial phase filtering, rather than rely on syntactic translation alone. We evaluate a range of open- and closed-source LLMs across prompting strategies and interaction paradigms. Our results reveal a substantial performance gap: GPT-o3-mini achieves 59.0% execution accuracy, while our custom multi-step agent, BMSQL, reaches 62.6%, both well below the expert baseline of 90.0%. BiomedSQL provides a new foundation for advancing text-to-SQL systems capable of supporting scientific discovery through robust reasoning over structured biomedical knowledge bases. Our dataset is publicly available at https://huggingface.co/datasets/NIH-CARD/BiomedSQL, and our code is open-source at https://github.com/NIH-CARD/biomedsql.

Reasoning large language models have recently achieved state-of-the-art performance in many fields. However, their long-form chain-of-thought reasoning creates interpretability challenges as each generated token depends on all previous ones, making the computation harder to decompose. We argue that analyzing reasoning traces at the sentence level is a promising approach to understanding reasoning processes. We present three complementary attribution methods: (1) a black-box method measuring each sentence's counterfactual importance by comparing final answers across 100 rollouts conditioned on the model generating that sentence or one with a different meaning; (2) a white-box method of aggregating attention patterns between pairs of sentences, which identified ``broadcasting'' sentences that receive disproportionate attention from all future sentences via ``receiver'' attention heads; (3) a causal attribution method measuring logical connections between sentences by suppressing attention toward one sentence and measuring the effect on each future sentence's tokens. Each method provides evidence for the existence of thought anchors, reasoning steps that have outsized importance and that disproportionately influence the subsequent reasoning process. These thought anchors are typically planning or backtracking sentences. We provide an open-source tool (www.thought-anchors.com) for visualizing the outputs of our methods, and present a case study showing converging patterns across methods that map how a model performs multi-step reasoning. The consistency across methods demonstrates the potential of sentence-level analysis for a deeper understanding of reasoning models.