NeurIPS 2026 Papers

Pre-trained large language models exhibit strong exploration ability, as evidenced by high pass@K scores, yet this ability degrades significantly during reinforcement learning (RL) fine-tuning due to mode collapse. A natural remedy is to impose token-wise KL divergence between the fine-tuned and base models; however, we argue that such objective introduces a fundamental mismatch, where token-wise KL operates locally at individual token positions, while exploration is a global behavior that emerges over sequences as a whole. Motivated by this insight, we propose Mean Order-Statistic Alignment (MOSA), which preserves exploration by aligning the fine-tuned model's global behavior to that of the base model. Specifically, MOSA captures global exploration through the mean order-statistic profile, which is obtained by computing the order statistics of each token's posterior over the vocabulary and averaging them across tokens at each rank position. The profile, designed a valid probability distribution, directly admits a principled KL-based objective. For efficiency, we further approximate the profile using only the top-K probabilities and a residual tail bucket, yielding a compact implementation with minimal computational overhead. Experiments on Countdown, 6 mathematical reasoning, and 2 coding benchmarks show that MOSA discovers more diverse plausible solutions, improves both pass@1 and pass@K, and scales effectively to models up to 32B parameters.

Many recent math- and science-oriented agent systems adopt hierarchical designs with specialized reviewer roles, motivated by the idea that routing critique through a dedicated review stage should help turn wrong candidates into correct ones on hard problems. We study that expectation on 4,181 verifier-grounded Omni-MATH problems, a hard ten-tier benchmark with enough headroom to separate protocols, using matched gpt-oss-120b actors as the primary actor family. On tiers 1-2, multi-agent collaboration adds at most about 2 percentage points over the matched single-agent baseline. From tier 4 onward, the gains open sharply, reaching about 10 to 20 percentage points on tiers 6-9. In that regime, broadcast-style peer discussion attains higher final accuracy than a planner-executor-reviewer pipeline (PER). We use that divergence as a starting observation and ask when reviewer quality translates into effective solver updates. On this benchmark, the PER-broadcast accuracy gap is not accounted for by reviewer precision alone. Here reviewer precision asks how often a reviewer warning truly points to a real error. PER's reviewer has higher precision (0.861 vs. 0.644), yet correct critique is much less likely to change the next candidate the protocol carries forward, and reviewer-guided repair is correspondingly lower. These results show that reviewer detection quality and critique uptake are empirically separable. Within matched PER interventions, forcing explicit acknowledgment lowers final accuracy instead of improving follow-through, while placing reviewer guidance directly in the solver's working context partially improves follow-through without closing the gap. Taken together, these interventions point in the same direction: critique appears more likely to be acted upon when it is presented more directly in the solver's working context, although this evidence is directional rather than causal. Under reviewer-centric evaluation, a system can look strong at spotting errors yet still fail to solve more problems if the protocol does not act on those critiques.

Modern frontier AI has been driven by a powerful but increasingly narrow recipe: scale the model, extend the context, and rely on emergent capabilities to support more complex forms of reasoning. \textbf{This position paper argues that this recipe is approaching a structural bottleneck. Robust reasoning requires more than parameter count, stored knowledge, or longer context windows. It requires intrinsic temporal structure i.e. the capacity of a system to generate, inhabit, and reason relative to its own internal time.} We argue that current architectures lack this capacity not merely as an implementation detail, but because most are built around fixed-pass or externally clocked computation rather than self-clocked, persistent, and dynamically coordinated processing. Our position is motivated by three converging lines of evidence. First, psychometric research links intelligence to the allocation and precision of processing time. Second, Global Workspace Theory suggests that flexible cognition depends on competitive access to a limited shared present and third, neural binding theories show that integration across distributed processors requires temporal co-activation rather than spatial co-location alone. Together, these motivate a two-level architectural view in which a diverse long-term population of specialist processors develops private temporal dynamics, while a short-term workspace constructs a shared present moment through selective competition and broadcasts it back to guide subsequent processing. We argue that existing work on continuous thought machines, conscious Turing machines, adaptive computation, memory-augmented models, mixture-of-experts, and quality-diversity neuroevolution already contains many of the necessary ingredients, but has not yet coupled them around intrinsic time as the organising principle. We outline falsifiable predictions for this view, including that harder reasoning problems should require deeper internal temporal trajectories, that synchronisation patterns should track task structure, and that self-clocked architectures should exhibit stronger adaptation under equal compute than fixed-pass alternatives. We call on the AI community to treat intrinsic temporal structure as a first-class architectural objective, not as an incidental by-product of scale.

Perturbation experiments are central to understanding cellular mechanisms, but remain costly and sparse, motivating prediction of gene expression responses for unobserved conditions. A promising recent direction leverages large language models (LLMs) as “virtual cell” simulators—using stepwise, knowledge-grounded mechanistic reasoning to infer differential expression—pointing toward an interpretable, knowledge-driven paradigm that transcends purely data-driven approaches. However, we find that plausibility is not prediction: despite producing biologically plausible explanations, these methods fail to capture perturbation-specific effects — systematically overestimating differential expression, often underperforming a simple gene‑frequency baseline in aggregate evaluations, and collapsing to chance-level performance at the per-gene level. This reveals a reliance on intrinsic gene response tendencies rather than true perturbation reasoning. We trace this failure to how evidence is presented: existing methods evaluate perturbation–gene pairs in isolation, without exposing how related perturbations differ in their effects on the same gene. To address this limitation, we introduce **CORE** (**C**ontrastive **O**rganization of **R**elational **E**vidence), which reframes prediction as a comparison task by organizing evidence into positive and negative outcomes from related perturbations. Using a biomedical knowledge graph for evidence retrieval, CORE improves calibration and substantially boosts perturbation-specific prediction in both LLM-based and non-LLM settings: for example, on drug-perturbation data, CORE-Reasoning improves Qwen3.5-9B aggregate metrics by up to 28.6\%, while on generic perturbation data, CORE-Voting raises macro-per-gene AUROC from chance to 0.711. This highlights contrastive evidence organization as essential to reliable LLM-based perturbation reasoning.

Vision-Language Models (VLMs) have demonstrated impressive multimodal reasoning, yet their ability to truly internalize the underlying physical consistency of real-world dynamics remains an open question. Existing benchmark paradigms often suffer from fragmented evaluation, focusing on isolated cognitive stages while overlooking the inherent synergy between perception, reasoning, and physical judgment. The lack of a holistic perspective limits the ability to diagnose whether VLMs can reliably evaluate the physical authenticity of emerging generative models. To address these issues, we introduce \textbf{PhysVista}, a benchmark designed to evaluate physical intelligence in VLMs through a closed cognitive loop framework inspired by the human \textbf{seeing–reasoning–assessment} process. PhysVista restores this loop by jointly evaluating physical state perception, physical dynamics reasoning, and physical plausibility assessment. It further distinguishes \textbf{event-level} reasoning and \textbf{scale-level} reasoning to enable fine-grained analysis of physical understanding. In addition, PhysVista incorporates both real-world and AI-generated videos, allowing evaluation across diverse domains and emerging generative scenarios. Extensive experiments on a wide range of state-of-the-art VLMs reveal significant limitations in current models’ physical intelligence, particularly in fine-grained reasoning and plausibility assessment. Our findings highlight critical gaps between visual recognition and genuine physical understanding, and provide insights for developing future physically grounded VLMs.

Manipulation planners depend on accurate predicates about physical state---is an object occluded? is a container open? is a path blocked?---but such predicates are hard to read off an image. Contrastively-trained vision--language models capture semantic content but collapse to near-zero binary F1 on these predicates, and even GPT-4o few-shot falls below random on multiple-choice physical-state questions. We present \ours, which predicts physical-state scene graphs from a single image and a natural-language goal. A frozen CLIP ViT, adapted with LoRA, feeds a heterogeneous graph of object, relation, and learnable memory tokens into a Latent Graph Reasoning Transformer (\lgrt); training combines supervised, counterfactual-margin, contrastive, and calibration losses with offline counterfactual-pair augmentation. On \pacbench, \ours reaches $\mu$-F1$\,{=}\,0.651$ and $99.3\%$ MCQ accuracy, a $+0.170$ absolute $\mu$-F1 gain over the strongest no-graph supervised baseline, with $99.6\%$ counterfactual directional consistency. Component ablations attribute $+0.426$ $\mu$-F1 to the \lgrt and $+0.392$ to the counterfactual loss. A \manipbench-trained variant scores $74.4\%$ MCQ against $30.0\%$ for the best zero-shot VLM on that benchmark's 4-choice format.

While recent vision-language models (VLMs) demonstrate strong multimodal understanding, they remain limited in spatial reasoning tasks that require active evidence acquisition and multi-step visual interaction. This limitation suggests that relying solely on implicit visual representations from vision encoders is insufficient for recovering fine-grained spatial evidence. We introduce *PERception-Interaction-reason Agent* (**PERIA**), a tool-augmented visual agent for spatial reasoning tasks across map reasoning, visual probing, and vision reconstruction. **PERIA** uses two lightweight tool families: *vision perception tools* for exposing textual, symbolic, and spatial evidence, and *vision interaction tools* for manipulating visual context, tracing paths, and verifying spatial relations. To train **PERIA**, we develop a unified recipe that combines supervised tool-use trajectory synthesis, composite rewards, and Observation-Relaxed Group-in-Group Policy Optimization (OR-GIGPO) for effective multi-tool behavior. Experiments on 13 benchmarks from 8 datasets show that **PERIA**-8B improves over the Qwen3-8B backbone by 10.0% on in-distribution benchmarks and 4.4% on out-of-distribution benchmarks, while outperforming previous state-of-the-art baselines of similar size by 7.0%–14.8%. It also achieves performance comparable to much larger models such as Qwen3-VL-235B-A22B-Thinking and GPT-5, demonstrating the effectiveness of **PERIA** in enhancing spatial reasoning capabilities.

Chain-of-thought (CoT) reasoning with self-consistency improves performance by aggregating multiple sampled reasoning paths. In this setting, correctness is no longer tied to a single reasoning trace but to the aggregation rule over a pool of candidate paths, making aggregation uncertainty the central challenge. This issue is critical where confidently incorrect answers are far more costly than abstentions. We introduce a conformal procedure for CoT reasoning that directly addresses aggregation uncertainty. Our approach replaces majority voting with weighted score aggregation over reasoning paths and calibrates an abstention rule using conformal risk control. This approach leads to finite-sample guarantees on the confident-error rate--the probability that the system answers and is wrong. We further identify score separability as the key condition under which abstention provably improves selective accuracy, and derive closed-form expressions that predict accuracy gains from calibration data alone. The method is fully inference-time, and requires no retraining. Across four benchmarks, four open-source models, and three score classes, realized confident-error rates are consistent with the prescribed targets up to calibration-split and test-set variability. Our method achieves $90.1\\%$ selective accuracy on GSM8K by abstaining on less than $5\\%$ of problems, compared with $82\\%$ accuracy under majority-voting baseline.

Large reasoning models (LRMs) have achieved remarkable progress in complex problem-solving tasks. Despite this success, LRMs typically suffer from high computational costs during deployment, highlighting a need for efficient inference. A practical direction is to switch the LRM between thinking and non-thinking modes dynamically. However, such approaches often introduce additional reasoning errors and lack statistical guarantees for the performance loss, which are critical for high-stakes applications. In this work, we propose $\textit{Probably Approximately Correct}$ (PAC) $\textit{Reasoning}$ that controls the performance loss under the user-specified tolerance. Specifically, we construct an upper confidence bound on the performance loss and determine a threshold for switching to the non-thinking model. Theoretically, using the threshold to switch between the thinking and non-thinking modes ensures bounded performance loss in a distribution-free manner. Our experiments on reasoning benchmarks show that the proposed method can save computational budgets and control the user-specified performance loss.

Transformers trained via Reinforcement Learning (RL) with outcome-based supervision can spontaneously develop the ability to generate intermediate reasoning steps (Chain-of-Thought). Yet the mechanism by which sparse rewards drive policy gradient to discover such systematic reasoning remains poorly understood. We address this by analyzing the policy gradient dynamics of single-layer Transformers on a synthetic graph traversal task that cannot be solved without Chain-of-Thought but admits a simple iterative solution. We prove that despite training solely on final-answer correctness, policy gradient drives the Transformer to converge to a structured, interpretable algorithm that iteratively traverses the graph vertex-by-vertex. We characterize the distributional properties required for this emergence, identifying the critical role of ``simple examples'': instances requiring fewer reasoning steps. When the training distribution places sufficient mass on these simpler examples, the Transformer learns a generalizable traversal strategy that extrapolates to longer chains; when this mass vanishes, policy gradient learning becomes infeasible. We corroborate our theoretical results through experiments on synthetic data and with real-world language models on mathematical reasoning tasks, validating that our theoretical findings carry over to practical settings.

Chain-of-Thought reasoning has enabled large language models to achieve substantial performance gains on complex tasks. However, these gains come at the cost of dramatically increased token consumption. This raises a fundamental question: is every token in the reasoning trace equally valuable? We present a diagnostic and optimization framework grounded in a key empirical finding: the value of tokens within a CoT reasoning sequence is highly non-uniform, and this non-uniformity can be effectively characterized by token-level log probability signals. We show that normalized log probability helps distinguish core tokens, which carry structural and decisive reasoning content, from redundant tokens, which are exploratory, low-confidence filler that contributes less directly to the final answer. Building on these findings, we formulate the TokenProbe framework around two empirical findings and one claim: findings identify token value inequality first and then establish TokenProbe as a core-token proxy, and the claim introduces an efficient GRPO objective positing that selectively compressing redundant tokens can yield Pareto improvements in the accuracy–token efficiency space. Empirically, our method preserves reasoning quality while reducing the token usage by 76% of the baseline. Under matched reasoning-length budgets, we show that it can even outperform strong flagship baselines like Gemini-3.1-Pro.

Large Language Models (LLMs) have achieved remarkable success in complex reasoning tasks through Chain-of-Thought (CoT) prompting. However, these models often exhibit "computational overthinking," generating redundant reasoning steps that increase latency and cost without improving accuracy. Recent studies suggest that CoT trajectories can be significantly pruned, yet existing methods often rely on forcing a static thinking budget, heuristic filtering, sub-optimal early exit via classification, or expensive re-training. In this paper, we introduce OS-Pruner, a lightweight plug-in framework that formulates chain-of-thought pruning as an optimal stopping problem. Given a reasoning prefix, OS-Pruner learns whether further reasoning is worth its token cost by optimizing an explicit utility that trades off final-answer accuracy against generated length. Our novel formulation enables the model to dynamically assess the sufficient point of termination for a reasoning chain. OS-Pruner is designed to be lightweight during both training and inference, and to provide users with fine-grained control over the reasoning-effort vs. accuracy trade-off. On diverse reasoning benchmarks and base models, OS-Pruner achieves 20-60\% reduction in generation length with minimal accuracy sacrifice.

Reactive synthesis, the problem of automatically constructing a hardware circuit from a logical specification, is a long-standing challenge in formal verification. It is elusive for two reasons: It is algorithmically hard, and writing formal specifications by hand is notoriously difficult. In this paper, we tackle both sides of the problem. For the algorithmic side, we present a neuro-symbolic approach to reactive synthesis that couples large reasoning models with model checkers to iteratively repair a synthesized Verilog implementation via sound symbolic feedback. Our approach solves more benchmarks than the best dedicated tools in the annual synthesis competition and extends to constructing parameterized systems, a problem known to be undecidable. On the specification side, we introduce an autoformalization step that shifts the specification task from temporal logic to natural language by introducing a hand-authored dataset of natural-language specifications for evaluation. We demonstrate performance comparable to that of starting from formal specifications, establishing natural synthesis as a viable end-to-end workflow.

Motion instruction generation in cross-video comparison aims to produce corrective feedback that describes the differences between a query and a reference motion. However, existing models often generate instructions that exhibit motion hallucinations, failing to reflect actual kinematic differences between paired videos. To systematically investigate these hallucinations, we introduce MotionHalluc, a dedicated benchmark for evaluating motion hallucinations in paired-video comparison. MotionHalluc comprises 1540 fine-grained questions over 553 video pairs, evaluating hallucinations along three core dimensions: (1) directional hallucination, (2) attributional hallucination, and (3) temporal hallucination. Extensive evaluations of state-of-the-art large multimodal models demonstrate high susceptibility to these hallucinations. Furthermore, we provide Perceive-Parse-Verify (PPV) as a training-free measurements extraction and verification baseline that converts candidate instructions into executable measurement queries and supplies kinematic measurements at inference time. Our results show that this simple measurements injection yields an average 10.6% performance gain across models, suggesting that motion reasoning with explicit quantitative measurements is a key factor in reducing hallucinations in cross-video comparison.

Large Reasoning Language Models (LRLMs) leverage long Chain-of-Thought (CoT) reasoning to solve complex tasks, yet often engage in overthinking. Specifically, they continue to generate reasoning steps that provide meaningless contribution to or even degrade the final answer, even after sufficient reasoning has been produced. Early-exit strategies are proposed to dynamically terminate meaningless or even harmful reasoning, aiming to improve both efficiency and accuracy. However, existing methods either suffer from over-truncation that degrades accuracy below vanilla CoT, or yield only spurious efficiency gains, which reduce token counts but increase wall-clock inference time. We observe that when an LRLM's reasoning path deviates into meaningless or even harmful wandering, this deviation is accompanied by an anomalous surge of high-entropy transition tokens. Building on this insight, we propose RPDI-EE, a training-free early-exit method that monitors the Reasoning Path Deviation Index (RPDI), defined as the ratio of local to global average entropy, which serves as a proxy for reasoning path deviation. By tracking this index, RPDI-EE detects and terminates likely meaningless or even harmful reasoning while preserving productive inference. Experiments across multiple benchmarks and four LRLMs of varying types and scales show that RPDI-EE achieves the largest average accuracy improvement over vanilla CoT among all tested early-exit methods, while mitigating the spurious efficiency gains of existing alternatives.

Large reasoning models (LRMs) improve language model capabilities by generating explicit thinking traces before final answers. In factuality-oriented question answering (QA), such thinking often improves overall performance by helping the model recover relevant knowledge and refine its answers. However, we find that this benefit is not uniform at the instance level: explicit thinking can also overturn correct non-thinking answers and lead to factual drift. We refer to this failure mode as \emph{thinking-induced hallucination}. To explain this phenomenon, we formulate explicit thinking in factuality QA as a thinking residual over the model's direct-answer tendency, which can either recover missing knowledge or introduce unsupported associations. Based on this formulation, we propose MARGO, \underline{\textit{M}}ixed-Mode \underline{\textit{A}}dvantage \underline{\textit{R}}egularization for \underline{\textit{G}}rounded \underline{\textit{O}}ptimization, a reinforcement learning framework that uses non-thinking rollouts as same-model references in advantage estimation. By constructing mixed-mode rollout groups with both thinking and non-thinking trajectories, MARGO evaluates whether explicit thinking adds factual value beyond direct answering, thereby suppressing hallucination-prone thinking while preserving beneficial thinking behaviors. Experiments across multiple factuality-oriented QA benchmarks demonstrate that MARGO improves factual reliability over strong baselines, while evaluations on mathematical benchmarks show that it preserves general reasoning ability.

Although LLMs have made substantial progress in reasoning, systematically producing frontier-level reasoning data remains difficult. Existing synthesis methods often have limited visibility into the structural factors that govern problem difficulty, which can result in narrow diversity and unstable difficulty control. In this work, we view the difficulty of a reasoning problem as arising from the accumulation of atomic knowledge-reasoning transformations, which we term *thought modes*. Building on this perspective, we propose MindLoom, a framework for synthesizing frontier-level reasoning data through compositional thought mode engineering. Given a collection of hard problems with verified solutions, MindLoom first decomposes those solutions into thought mode chains that reveal each problem's construction logic. It then trains a retrieval model that matches problem states to compatible thought modes, providing guidance on which reasoning challenges to introduce during synthesis. New problems are composed by iteratively applying retrieved thought modes to seed questions, with distribution-aligned sampling to encourage diverse reasoning coverage. Finally, a rollout-based judging stage labels generated questions by difficulty and supplies judged-correct responses for supervised fine-tuning. We evaluate MindLoom on nine benchmarks covering five STEM disciplines and four mathematical reasoning tasks across multiple model families and sizes. Models fine-tuned on MindLoom generated data consistently improve over base models, distillation, and external-data baselines across the reported benchmarks. Ablation studies indicate the contribution of each component, and further analysis suggests that MindLoom covers a broad range of reasoning patterns while maintaining useful difficulty control. We have open-sourced our implementation at https://anonymous.4open.science/r/MindLoom-5B8E.

Large language models increasingly rely on long Chain-of-Thought (CoT) to solve complex reasoning tasks, but the resulting CoT often contains repeated checks, abandoned branches, and unused planning that increase inference cost without supporting the final answer. Existing efficient reasoning methods address this through length budgets, token-level importance, confidence signals, or a stronger teacher model, treating CoT compression mainly as a text-shortening problem rather than as a question of which steps the answer actually depends on. We propose MinPath, a training framework that treats CoT compression as a graph problem. The raw CoT is represented as a Typed Reasoning Graph, where the Minimal Dependency Path supporting the target conclusion is extracted, and the resulting MinPath is used for fine-tuning. Across four reasoning models on GSM8K, MATH-500, and AIME24, MinPath reduces generated tokens by 33.9\% on average with nearly unchanged precision (accuracy even improved in 41.7\% of all experiments) and outperforms five compression baselines in every evaluated setting.

Mental simulation, the ability to reason by constructing and updating internal visual states, is central to flexible human cognition, but remains elusive in current multimodal large language models. Latent multimodal reasoning offers a promising route toward this capability by interleaving text with continuous visual latent blocks, allowing models to maintain internal visual workspaces without decoding them into pixels. In this paper, we ask whether these latent visual states actually matter. Through a direct perturbation analysis across representative latent-reasoning MLLMs, we find that they often do not: dropping, replacing, or shuffling generated visual latents has only minor effects on the model's own final-answer probability, while equivalent perturbations to the linguistic reasoning trace are substantially more damaging. We call this failure mode Mental Simulation Collapse (MSC): the model appears to perform interleaved visual reasoning, but routes most answer-relevant computation through text. We further diagnose that MSC follows naturally from existing training objectives: visual alignment can make generated latents resemble helper-image embeddings, but it does not require the decoder to rely on them; subsequent text-only supervision further leaves the semantic role of the latent pathway unidentifiable. To address this, we introduce AnchorSim (Anchored Mental Simulation), a contrastive distillation framework that trains models to make their generated visual latents consequential. Across multiple visual reasoning benchmarks, AnchorSim improves performance, while diagnostic re-evaluation shows that visual perturbations now meaningfully affect predictions. Our results expose a fundamental failure mode of latent multimodal reasoning and demonstrate that mental simulation must be anchored to model behavior in order to matter.