NeurIPS 2026 Papers

Recent years have witnessed meteoric progress in *reasoning models*: neural networks that generate intermediate *reasoning traces* (RTs) before producing a final output. Despite the rapid advancement, our understanding of how RTs support reasoning, and where this paradigm fails, remain incomplete. To promote greater clarity, we introduce PITA: a novel large-scale dataset of over 23 million statements in propositional logic and their corresponding proofs. As a benchmark for robust reasoning, we focus on length generalization: if a model is trained to determine truth or falsity on statements with proofs up to fixed length, how well does it generalize to statements requiring longer proofs? We propose notions of (1) *task depth* and (2) *task breadth*, which measure respectively (1) the number of proof steps required to solve a proposition and (2) the number of unique propositions within a task family. We vary these quantities across subsets of PITA, and find that RT models generalize well on relatively *broad* and *shallow* subsets, while deteriorating on relatively *narrow* and *deep* subsets compared to non-RT baselines. As a controlled point of comparison, we study a separate, tractable transitive inference task that exhibits qualitatively similar behavior. Our accompanying theory explains the scalings observed in this simpler setting, suggesting one mechanism by which breadth can favor RT models while depth can expose long-context weaknesses. Our findings suggest salient benefits and limitations of proof-like reasoning traces in controlled length-generalization settings.

Recent work on recursive architectures has shown that tiny neural networks can be surprisingly powerful on structured reasoning tasks. The trick is to model reasoning trajectories with a latent dynamical system. We argue that the inference-time behaviour of these architectures is best understood as approximate inference over latent reasoning trajectories, with deterministic recursion as the one-particle, zero-noise limit. We make this view operational through guided stochastic exploration: stochastic perturbations of the reasoning dynamics propose neighbouring trajectories, and the model's existing early-stopping head reweights them online. The framework yields three label-free diagnostics: local stability, guide alignment, and cloud-token entropy. These predict, from inference traces alone, whether the procedure will help and which of its outputs to trust. On Sudoku-Extreme it lifts exact-solve accuracy from $85.9\%$ to $98.0\%$ without retraining; on Maze-Hard the diagnostics flag a misaligned guide, as validation performance later confirms. The same machinery thus characterises both when recursive reasoning has room to improve at the trajectory level and when the model's internal guide can recover it.

Modern large language models increasingly require long contexts for reasoning and multi-document tasks, but attention's quadratic complexity creates a severe computational bottleneck. We present *Block Sparse Flash Attention* (BSFA), a drop-in replacement that accelerates long-context inference while preserving model quality. Unlike methods that predict importance before computing scores, BSFA computes exact query-key similarities to select the top-$k$ most important value blocks for each query. By comparing per-block maximum scores against calibrated thresholds, we skip approximately 50% of the computation and memory transfers for pruned blocks. Our training-free approach requires only a one-time threshold calibration on a small dataset to learn the per-layer and per-head attention score distributions. We provide a CUDA kernel implementation that can be used as a drop-in replacement for FlashAttention. On Llama-3.1-8B, BSFA achieves up to $1.13\times$ end-to-end speedup on LongBench with only a $1.1\%$ accuracy drop, and up to $1.24\times$ on Needle-in-a-Haystack retrieval at a $1\%$ accuracy drop. The attention kernel itself accelerates by up to $1.38\times$. We compare BSFA against five recent sparse attention baselines (SpargeAttention, MInference, FlexPrefill, XAttention, and BLASST), and verify the method on Qwen2.5-7B and on A6000 and H100 GPUs. The implementation is available at https://github.com/Anonymous44414/Block-Sparse-Flash-Attention.

A language model agent acts on a context assembled by retrieval, summarisation, prompt templates, and memory. How many decision-relevant bits must this context carry before the agent can act with low regret? We give an answer that does not depend on the model, decoder, or scaffold. Modelling the agent as a Markov chain H→C→A and treating Bayes regret (the value gap between the chosen and the optimal action) as the distortion measure, we define a regret rate-distortion (RRD) function Rℓ(ε) whose converse is a one-line consequence of the data-processing inequality (DPI). A context that carries fewer than Rℓ(ε) bits cannot drive expected regret below ε, regardless of decoder, scaffold, scale, or chain-of-thought. For finite M-ary decisions under zero-one loss the bound reduces to the Fano floor in closed form, requiring ≈3.14 bits at (M=16, ε=0.1) and ≈6.73 bits at (M=256, ε=0.1). We instantiate the theory on 441 controlled single-step tool-selection cells, a balanced 16-way classification proxy spanning three open-source large language models (Qwen3-32B, Gemma3-27B, GLM-4-32B) and three public benchmarks (API-BANK, τ-bench, ToolBench). Because the gold tool index X=f(H) is a deterministic function of the history, the action-side rate IMM(X; A) lower-bounds the context capacity by data processing, so an action-side test of the converse is strictly tighter than a context-side one. Every cell respects the predicted floor (median margin +1.07 bits). Across 629 matched-rate pairs, 97.5% share overlapping regret confidence intervals (CIs), and a chain-of-thought supplement slides cells along the frontier rather than across it. The same data exposes a tokens-versus-bits gap reaching ∼1.7×10⁶:1, with the model extracting fewer than 0.05 bits about the gold action on a 32,768-token ToolBench prompt. In the single-step regime, the framework supplies a falsifiable, model-agnostic stopping criterion at IMM=Rℓ(ε) and reframes context engineering as an information-allocation problem.

Large reasoning models (LRMs) generate reasoning traces before producing final answers, yielding strong performance on multi-step reasoning tasks. However, preference alignment for LRMs remains underexplored. The ideal answer-level preference objective marginalizes over reasoning traces, but this trace-marginalized objective is computationally intractable. As a result, practical methods rely on single-trace surrogates, which can induce high gradient variance due to stochastic trace sampling. We cast this challenge as a gradient-estimator design problem and propose Bias-Variance Optimized Preference Optimization (BVPO), which combines a standard trace-based gradient estimator with an empty-trace gradient estimator from a separately sampled empty-trace branch. Under our fixed-prompt, fixed-answer conditional view, the empty-trace branch is deterministic with respect to the trace randomness studied in our analysis. We show that the composite estimator contracts the conditional reasoning-trace-noise component, admits a conditional MSE-optimal mixture, and appears in the estimator-dependent term of a standard biased-SGD bound. Empirically, BVPO improves alignment over baselines on alignment benchmarks, including AlpacaEval 2 and Arena-Hard. Although trained on general conversational data, BVPO also generally preserves reasoning performance on six math reasoning benchmarks. Empirical diagnostics further show that BVPO yields lower total gradient variance during training.

Video Reasoning Segmentation (VRS) demands both complex semantic reasoning and robust spatiotemporal mask propagation. Existing paradigms typically compress reasoning into an isolated, opaque token (e.g., \texttt{[SEG]}) via expensive fine-tuning, or rely on training-free prompt-then-track pipelines that inevitably suffer from semantic degradation and identity switches when encountering hard distractors. To fundamentally address these vulnerabilities, we propose Counterfactual Logic and Explicit Anchoring Reasoning (\textbf{CLEAR}) framework, a training-free VRS framework inspired by human dual-system cognition. Specifically, we introduce a Causal-Counterfactual Inference mechanism that transforms error-prone coordinate regression into a tractable discrete instance selection task, establishing highly reliable initial anchors via rigorous bidirectional reasoning. During mask propagation, we design an adaptive mode-switching mechanism bridged by a Heterogeneous Event Monitor. The fast-thinking mode utilizes a purely visual engine for efficient temporal propagation, while the Monitor continuously evaluates its reliability across multiple dimensions. Upon detecting propagation anomalies, the slow-thinking mode intervenes and explicitly replaces \texttt{[SEG]} tokens with a structured Contrastive Concept Dictionary for closed-loop identity recovery..Extensive experiments demonstrate that CLEAR significantly outperforms existing training-free methods and achieves performance comparable to state-of-the-art training-based approaches on comprehensive referring and reasoning VOS benchmarks. Code will be available after acceptance.

Training LLMs to think and reason for longer has become a key ingredient in building state-of-the-art models that can solve complex problems previously out of reach. Recent efforts pursue this in different ways, such as RL fine-tuning to elicit long CoT or scaling latent reasoning through architectural recurrence. This makes reasoning length an important scaling knob. In this work, we identify a novel phenomenon (both theoretically and experimentally): under outcome-only supervision, out-of-distribution (OOD) performance can continue improving as training-time reasoning length (e.g., the token budget in RL, or the loop count in looped Transformers) increases, even after in-distribution (ID) performance has saturated. This suggests that robustness may require a larger budget than ID validation alone would indicate. We provide theoretical explanations via two mechanisms: (i) self-iteration can induce a stronger inductive bias in the hypothesis class, reshaping ID-optimal solutions in ways that improve OOD generalization; and (ii) when shortcut solutions that work for ID samples but not for OOD samples persist in the hypothesis class, regularization can reduce the learned solution's reliance on these shortcuts as the number of self-iterations increases. We complement the theory with empirical evidence from two realizations of scaling training-time reasoning length: increasing the number of loops in looped Transformers on a synthetic task, and increasing token budgets during RL fine-tuning of LLMs on mathematical reasoning, coding, and cross-representation tasks.

As LLMs increasingly serve in advisory and deliberative roles, users rely on them for non-verifiable reasoning in domains lacking objective ground truths. However, traditional evaluations of LLM reasoning focus almost exclusively on fact-based domains, such as mathematics and science, leaving uncertainty over whether and to what degree models can handle ambiguous, subjective, or value-laden problems over time. To address this concern, we propose moral reasoning as a paradigmatic subdomain of non-verifiable reasoning. We define moral robustness as a model’s capacity to exhibit sound moral reasoning across time and contexts, and we introduce a scalable, adversarial, multi-turn evaluation framework to empirically measure this capability. We simulate 48,000 user-agent moral deliberations across four frontier LLMs, varying premise relevance, premise order, conversation duration, and the user's stated moral view. We find that models successfully ignore morally-irrelevant distractors, but shift their reasoning by up to 6.5\%, on average, towards the user's stated preferred moral view, and varying their reasoning depending on factors such as order (altering moral judgments by order in 13-22\% of the cases) and duration (altering moral judgments between single-turn and multi-turn in 10-24\% of the cases). Our analysis indicates that models tailor not just their final verdicts but their underlying justifications to align with a user’s moral viewpoint — a failure mode we characterize as moral deliberative sycophancy.

We study fixed-confidence best arm identification in generalized linear bandits under a hybrid feedback model: at each round, the learner may query either (i) absolute reward feedback from a single arm or (ii) relative (dueling) feedback from an arm pair, both governed by generalized linear models. We introduce a likelihood-ratio–based confidence sequence that unifies heterogeneous generalized linear observations and yields an explicit ellipsoidal confidence set under a self-concordance assumption. Building on this confidence set, we propose a hybrid Track-and-Stop algorithm that adaptively allocates queries by tracking a minimax-optimal design over a joint action space of arms and pairs. We establish $\delta$-correctness and provide high-probability upper bounds on the stopping time. We further extend the framework to a cost-aware setting that accounts for heterogeneous acquisition costs across feedback modalities. Empirical experiments demonstrate that the proposed algorithms significantly improve sample efficiency over baseline methods.

Multimodal large language models (MLLMs) and diffusion models have each reached remarkable maturity: MLLMs excel at reasoning over heterogeneous multimodal inputs with strong semantic grounding, while diffusion models synthesize images and videos with photorealistic fidelity. We argue that these two families can be unified through a simple division of labor: MLLMs perform semantic planning, while diffusion models render pixels from high-level semantic guidance and low-level visual features. Building on this idea, we propose \textbf{Bernini}, a unified framework for video generation and editing. An MLLM-based \textbf{planner} predicts the target semantic representation directly in the ViT embedding space, and a DiT-based \textbf{renderer} synthesizes pixels conditioned on this plan, augmented by text features and, for editing, source VAE features for detail preservation. Because semantics serve as the interface, the planner and renderer can be trained separately and only lightly co-trained, preserving the pretrained strengths of both components while keeping training efficient. To better handle multiple visual inputs, we introduce Segment-Aware 3D Rotary Positional Embedding (SA-3D RoPE), and further incorporate chain-of-thought reasoning in the planner to better transfer understanding into generation. Bernini achieves state-of-the-art performance across a wide range of video generation and editing benchmarks, with the MLLM's pretrained understanding translating into strong generalization on challenging editing tasks.

Video Large Language Models (VideoLLMs) receive frames in sequential order and interpret how visual content evolves along the temporal axis, yet temporal reasoning remains a persistent weakness across architectures. Reversing the frame order of a video, a transformation that should invert temporal answers, often leaves the final prediction unchanged. We investigate where this failure originates by defining the temporal divergence vector $\tau_l$, the layer-wise representational difference induced by reversing temporal order. Tracking its magnitude across layers reveals a consistent temporal divergence profile where the divergence peaks at intermediate layers and progressively diminishes toward the output. We confirm this peak is specific to temporal reasoning and functionally critical for predictions, establishing that VideoLLMs acquire temporal information at intermediate layers but fail to maintain it to the output. This progressive fading motivates our method, Temporal Activation Injection (TAI), which extracts $\tau_l$ at the peak of the profile for each input and reinjects it into subsequent layers following the measured decay. TAI requires no training and consistently improves temporal reasoning across three VideoLLMs and four benchmarks with negligible impact on non-temporal tasks.

Large language model systems repeatedly make test-time decisions by comparing similarity. They retrieve exemplars, rerank candidates, and choose revisions. But similarity is not unique: a candidate can be close lexically, semantically, structurally, or stylistically, while only one of these axes may matter for the current instance. We propose BASIS~(Bayesian Axis Selection via Interactive Signals), a training-free Bayesian framework that treats weak feedback as evidence about this latent task-aligned axis. BASIS maintains a posterior over candidate targets and similarity axes. It updates this posterior from scalar scores, pairwise preferences, or verifier outputs, and then chooses future probes by expected information gain. In a finite noisy similarity model, we show that fixed-axis and fixed-mixture rules can remain suboptimal under axis shift. Across concept recovery, reasoning exemplar selection, response refinement, and adversarial distractors, BASIS consistently outperforms stronger rerankers, static mixtures, and compute-matched best-of-$N$ search under the same candidate pools and feedback budgets. The average gains are moderate because all methods share the same backbone, candidate pools, and feedback signals. Under axis shift, however, the gap widens substantially, showing that latent-axis inference is not fully replaced by better scoring or more search alone.

Post-training data pipelines are traditionally hand-designed by researchers who orchestrate generation, verification, and diversity balancing. While Large Language Models (LLMs) can already execute individual subtasks, the architectural composition of these pipelines remains a manual bottleneck, excluding the most substantial engineering step in post-training from automation. We introduce AutoDataBench, a benchmark evaluating whether LLM agents can autonomously design post-training data-synthesis pipelines. Each task gives the agent only a description of a target capability and asks it to produce a complete, executable pipeline whose generated corpus is then used to fine-tune a fixed student model and scored on the evaluation. Each task is delivered as a natural-language instruction and a sandboxed workspace given to the agent, paired with an evaluation protocol. We instantiate AutoDataBench on five capabilities: instruction following, competition math, deep search, repo-level software engineering, and terminal automation. For each task we pair the evaluation with a strong human-designed pipeline whose downstream score serves as a reference point for the agent. Experimental results reveal that frontier agents close most of the gap to the human reference on text-centric tasks (e.g., Opus-4.7 nearly matches humans in instruction following), yet a noticeable performance gap persists in tasks requiring executable-environment construction (SWE and Terminal). Furthermore, we observe a decoupling between downstream task-solving strength and pipeline-construction quality, where strong reasoning models like GPT-5.5 may still produce suboptimal pipelines. These findings suggest that autonomous environment synthesis remains the primary bottleneck for self-improving data loops, distinct from mere task-solving competence.

We study attention mechanisms through the lens of a canonical unsupervised problem: principal component analysis (PCA). We show that, when trained on Gaussian data, both softmax and linear attention layers learn parameters that align with the principal eigenvectors of the covariance matrix, thereby establishing a direct and explicit connection with PCA. Our analysis covers both finite and infinite prompt regimes. In the infinite-prompt limit, we prove convergence to globally optimal solutions aligned with the leading spectral direction, while in the finite-prompt setting we show that the same behavior emerges up to sampling effects. We further extend the analysis to an in-context setting with spiked Wishart covariances, where attention successfully recovers the underlying signal direction. These results demonstrate that attention inherently performs PCA-like computations under unsupervised objectives, providing a theoretical foundation for its representation-learning capabilities.

Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs. We identify a previously-unreported phenomenon we call \textbf{attention drift}: as the drafter generates successive tokens within a speculation chain, attention progressively moves from the prompt onto its own recently-generated tokens. We observe this across both \emph{EAGLE3} drafters and \emph{MTP heads}, suggesting drift is a property of drafter designs. We trace this to the un-normalized residual path between chain steps: the drafter's hidden state magnitude grows monotonically with chain depth, which exhibits dynamics consistent with additional pre-norm transformer layers stacked on the target rather than as a standalone autoregressive predictor. In order to limit the growth, we propose two architectural changes: Post-norm on the drafter hidden states and per-hidden-state RMSNorm after capturing target hidden states. Our interventions improve acceptance length over the current leading model, pre-norm EAGLE3, by up to $2\times$ under template perturbation, $1.18\times$ on long-context tasks, and $1.10\times$ on seven standard benchmarks spanning multi-turn chat, math, and coding. Our changes also allow shorter train-time-test depths to generalize over longer drafting sequences.

Reliable uncertainty estimation is essential for deploying large language models in high-stakes reasoning tasks, yet most existing approaches rely on expensive sampling strategies or shallow token-level signals that ignore the internal logical structure of model outputs. We introduce Argument Graph Uncertainty, a post-hoc framework for multiple choice question tasks, that estimates model confidence directly from the logical structure of a single chain-of-thought reasoning trace. Our method uses a lightweight instruction-tuned model to segment the reasoning chain into propositional units, then applies a simple NLI model to infer relational edges and construct two complementary argument graphs over these segments. The resulting edge scores are aggregated into a Dirichlet posterior over answer options, whose concentration yields calibrated uncertainty estimates reflecting both local step-to-step coherence and global logical consistency across the reasoning trace. Evaluated on GPQA, a challenging graduate-level multiple-choice benchmark, our method consistently surpasses token-based and consistency-based uncertainty baselines across multiple reasoning models. We further demonstrate through extensive combination analyses that argument graph uncertainty captures an orthogonal signal to existing approaches, making it a complementary component in uncertainty ensembles. Crucially, our framework requires no additional calls to the evaluated model and relies entirely on small, efficient auxiliary models, producing fully interpretable, per-node uncertainty attributions at low computational cost.

Test-time scaling systems often choose how much to sample, which prompt to use, and which verifier or aggregation rule to trust after repeatedly inspecting a calibration set. Standard validation error bars are not valid after this adaptive selection. We model a test-time scaling recipe as a randomized predictor with a random cost and prove a time-uniform PAC-Bayes certificate that holds simultaneously over all calibration times, all posterior mixtures over recipes, and all data-dependent stopping rules. Thus the same event covers the recipe search that precedes deployment, not merely the fixed recipe finally reported. For a finite grid of budgets and verifiers, the penalty is the familiar $\log M/n$, while non-uniform priors and localized posterior mixtures give sharper certificates for structured searches. We specialize the result to self-consistency and show a complementary margin law: plurality voting improves exponentially in the sample budget only on examples where the correct answer is the generator's unique mode; otherwise more compute can provably amplify errors. The same margin analysis yields an oracle water-filling law for compute allocation. A synthetic audit quantifies the optimism caused by adaptive budget selection, and a 192-example GSM8K audit using gpt-4.1-mini generations illustrates risk-cost reporting for a real inference pipeline. The results give a lightweight, distribution-free way to report risk and compute claims for adaptive inference systems.

Human motion change is currently hindered by the lack of large-scale datasets, comprehensive benchmarks, and robust algorithms capable of handling diverse motion requirements. To address these gaps, we propose AnyMotion, a novel open-world paradigm for highly controllable and identity-consistent human motion editing across arbitrary scenes and actions. Our contributions are two-fold. First, we establish a hierarchical label taxonomy covering six major categories and 527 fine-grained classes, and curate HME-1M, a large-scale dataset containing 1.04 million editing quadruplets via systematic data balancing, cleaning, and reverse construction. Second, we introduce the Skeleton Chain-of-Thought (SCoT) framework, a two-stage pipeline consisting of motion reasoning and guided generation. In the reasoning stage, a multimodal large language model serves as a cognitive planner to derive target skeletal keypoints and MetaQueries. These outputs subsequently function as explicit structural and semantic priors, providing the necessary guidance for the diffusion transformer to perform high-fidelity motion generation. Finally, built upon our label taxonomy, we present Human Motion-Bench, the first motion-centric benchmark equipped with multi-category fine-grained evaluation metrics. Extensive experiments demonstrate that AnyMotion achieves state-of-the-art performance across diverse editing scenarios.

On-policy self-distillation, where a student is pulled toward a copy of itself conditioned on privileged context (e.g., a verified solution or feedback), offers a promising direction for advancing reasoning capability without a stronger external teacher. Yet in math reasoning the gains are inconsistent, even when the same approach succeeds elsewhere. A pointwise mutual information analysis traces the failure to the privileged context itself: it inflates the teacher's confidence on tokens already implied by the solution (structural connectives, verifiable claims) and deflates it on deliberation tokens (*Wait*, *Let*, *Maybe*) that drive multi-step search. We propose Anti-Self-Distillation (AntiSD), which ascends a divergence between student and teacher rather than descending it: this reverses the per-token sign and yields a naturally bounded advantage in one step. An entropy-triggered gate disables the term once the teacher entropy collapses, completing a drop-in replacement for default self-distillation. Across five models from 4B to 30B parameters on math reasoning benchmarks, AntiSD reaches the GRPO baseline's accuracy in 2 to 10× fewer training steps and improves final accuracy by up to 11.5 points. AntiSD opens a path to scalable self-improvement, where a language model bootstraps its own reasoning through its training signal.