NeurIPS 2026 Papers

Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, yet the internal mechanisms driving these multi-step processes remain opaque. Existing mechanistic interpretability approaches often rely on superficial text-pattern co-occurrences or capture only short-term effects, struggling to trace the long-horizon, outcome-driven influence of internal components such as neurons or sparse features in multi-step reasoning. In this paper, we introduce Integrated Policy Gradient (IPG), a novel, training-free framework that brings policy-based attribution into mechanistic interpretability for LLM reasoning. Grounded in *outcome-oriented* and *sequential-influence-aware* principles, it backpropagates outcome-level, non-differentiable behavioral signals (e.g., reasoning outcome correctness) through entire inference trajectories, utilizing path integration to isolate outcome-relevant mediators at component level for reliable attribution. This provides a linear approximation of Causal Mediation Analysis (CMA) in reasoning. Empirical evaluations demonstrate that IPG achieves precise localization and reveals mechanistic insights by identifying sparse, concentrated subsets of internal components associated with specific reasoning behaviors. Our method also enables modulation of reasoning capabilities, offering a powerful and reliable framework for understanding LLM reasoning.

Importance weights are essential in domain adaptation under label shift, yet their utility is often undermined by the finite sample uncertainty associated with their estimation. Existing methods typically analyze this uncertainty through Gaussian elimination on interval-valued linear systems, which leads to overly conservative confidence regions and inefficient downstream applications. We propose a paradigm shift from inversion-based inference to a direct matrix constraint framework. We use this framework to define a joint confidence region and extract marginal intervals via linear programming, deriving provably tighter bounds for importance weights while maintaining exact finite-sample validity. Furthermore, we analyze the confidence region's geometry and provide the theoretical results for its diameter bounds. Evaluated across text, image, multimodal benchmarks, including AGNews, MNIST, CIFAR-10, N24News, and a real-world autonomous driving dataset, nuImages, our approach consistently yields shorter confidence intervals and smaller prediction sets than inversion-based methods.

Multimodal Large Language Models (MLLMs) face significant scalability challenges in long-video understanding due to quadratic attention complexity and context window constraints. Existing retrieval-based approaches typically using supervised learning on static keyframe annotations. However, these label priors often fail to align with the actual information required by the downstream MLLM for complex reasoning, creating a fundamental misalignment between the training target and inference needs. To bridge this gap, we propose a probabilistic framework that reformulates frame selection as a Bayesian inference process. We first initialize a policy using supervised learning to capture general saliency. Then, we treat the frozen MLLM as a task environment and employ its negative log-likelihood (NLL) as dense reward evidence to update the policy via Bayesian Group Relative Policy Optimization (GRPO). This process explicitly transits the selection principle from label priors to task evidence. Extensive experiments on Video-MME, MLVU, and LongVideoBench demonstrate that our approach matches the accuracy of heavy MLLM scorers while maintaining the efficiency of lightweight models.

The electrocardiogram (ECG) is a time series that records the electrical activity driving cardiac motion through electromechanical coupling (EMC), where the depolarization captured by ECG triggers each heartbeat. The temporal delay and low spatial resolution of such multichannel dynamic data often cause temporal and spatial misalignments along the physical motion, limiting the application of ECG data to event/segment-based downstream clinic practice. Although many efforts have been made to recover spatiotemporal trajectories conditioned on ECG waveforms, the intrinsic principle of EMC that results in various temporal delays remains underexplored in current generative models. To address this limitation, we propose Time2Track, a mechanics foundation model that generates individualized 4D heart motion from same-day ECG recordings by encoding electrophysiology as delay-aware token embeddings. Specifically, a learnable wavelet decomposition extracts the electromechanical delay, which is then formulated as a physics-based conditioning signal for motion diffusion. Time2Track generates 4D motion with individual-specific kinematics, avoiding the collapse to a population average, i.e., a common failure scenario when diffusing motions whose inter-subject differences are subtle. On $n=40,577$ subjects from the UK Biobank, our Time2Track accurately synthesizes 4D cardiac trajectories from ECG and preserves inter-subject diversity, a regime in which generic motion diffusion baselines collapse to a population average.

Reasoning across multiple long-form videos requires models to find sparse evidence within each video and link related entities, events, and narratives across streams. Existing long-video benchmarks mainly evaluate single-video understanding, while multi-video benchmarks typically use short clips. We introduce CLoVR-Bench, a comprehensive benchmark for Cross-Long-Video Reasoning with 2,000 expert-annotated QA pairs over 400 long-form videos. Its three-level taxonomy covers Comparative Analysis, Tracking and Retrieval, and Integrated Reasoning, spanning 14 tasks and 36 subtasks that diagnose both intra-video localization and inter-video alignment. Our evaluation of 13 representative MLLMs shows substantial gaps across this hierarchy, with performance degrading from localized comparison to long-range retrieval and integrated cross-video reasoning. We further propose HOLMES, a training-free framework that formulates cross-long-video reasoning as evidence-slot filling over a typed binding graph. HOLMES plans option-discriminating evidence slots, localizes evidence with risk-conditioned policies, records coverage certificates for failed searches, verifies cross-video bindings visually, audits constraints, and returns answers through hard graph-readout gates. Experiments show that HOLMES improves both open-source and closed-source backbones, outperforms adapted long-video baselines, and yields the largest gains on tasks requiring long-range evidence search and cross-video binding. Together, CLoVR-Bench and HOLMES provide a diagnostic testbed and an interpretable baseline for future cross-long-video reasoning.

Industry-scale video and live-streaming moderation imposes requirements that are difficult to satisfy with generic pretrained public models or external APIs, including adaptation to platform-specific data distributions, policy-specific objectives, and product-level safety constraints. As a result, platforms must undertake internal model development, naturally turning to shared public research for guidance. However, existing multimodal foundation-model studies primarily report architectures, training recipes, data scaling strategies, and benchmark results, but provide less systematic guidance on how failures should be localized and translated into targeted model-development interventions. Interventions are essential because deployment failures are rarely self-explanatory. Similar failures can originate from different causes. Without targeted interventions, improvement reduces to heuristic trial-and-error, where benchmark improvements are weakly attributable, and failures are difficult to trace to their underlying causes. To address this gap, we present a diagnostic methodology for industry-scale Audio-Visual-Language Models~(AVLM) development. The methodology maps model failures into a taxonomy of observable failure signatures and links each class of failure to an intervention space. We instantiate this methodology across the development and alignment lifecycle of an AVLM foundation model for a large-scale video and live-streaming platform. The resulting system supports over 100 regions and is designed for noisy, ambiguous, and highly diverse content drawn from global platform traffic.

LLMs are widely adopted in multi-role dialogue and persona simulation. However, their strong ability to imitate role-specific linguistic and behavioral patterns can introduce privacy risks, reinforce social biases, and lead to unsafe or uncontrollable outputs. To address these issues, we introduce the Role-playing Unlearning task, which aims to selectively forget the style and knowledge associated with target roles while preserving the language ability and behaviors of the remaining roles. Unlike prior unlearning tasks that focus solely on removing factual information, our task additionally requires forgetting persona-specific characteristics, making the control over model generation more fine-grained. We propose ROLEMOE, a Mixture-of-Experts framework that enhances the separation of role-specific patterns by mapping different roles to more independent expert subspaces, providing structural interpretability. To reinforce this separation, ROLEMOE incorporates two complementary losses: a specialization loss that drives different roles to adopt different expert distributions, and a disentanglement loss that encourages different experts to develop distinct capability subspaces. This decomposed structure enables role unlearning to focus selectively on target-role experts. We finally propose RoleBench-Unlearn for evaluation, and experiments across multiple LLM architectures show that ROLEMOE achieves significantly more effective and controllable role forgetting, while better preserving the linguistic quality, knowledge, and persona consistency of non-target roles. All code and datasets are publicly available at: https://anonymous.4open.science/r/RoleMoE-6328.

We present TAR (Traffic Anomaly Reasoning), a large-scale multi-task dataset for training and evaluating vision-language models on traffic anomaly understanding. While video anomaly detection (VAD) has traditionally focused on binary classification or temporal localization, TAR moves beyond detection to reasoning. TAR contains 44,040 training annotations with explicit chain-of-thought reasoning traces across 10 task types, covering 3,670 CCTV transportation videos (~26 hours) sourced from eight public datasets. TAR provides a structured progression of tasks organized into three groups — Question Answering, Temporal Reasoning, and Scene Understanding — reflecting our finding that reliable anomaly understanding demands capabilities far beyond detection. Annotations are produced by a hierarchical auto-labeling pipeline using state-of-the-art VLMs, with existing human annotations incorporated as supplementary context for 25% of videos. Every answer is paired with a step-by-step reasoning trace, enabling both supervised training with chain-of-thought supervision and fine-grained evaluation of model reasoning. We evaluate eight vision-language models on a held-out test set of 84 videos (1,008 annotations) with human-reviewed and corrected annotations and find that current models struggle with Temporal Reasoning and Scene Understanding tasks even when they perform well on basic Question Answering — revealing that detection accuracy is a poor proxy for genuine understanding. A multi-task ablation study shows that progressively adding task groups during fine-tuning yields consistent gains, with the full 10-task model achieving +44.5 points on Binary QA accuracy and +21.4 points in aggregate score over the zero-shot baseline, demonstrating the dataset's value as a training resource. TAR is released under CC-BY-4.0 and serves as the official training data for AI City Challenge 2026 Track 3: Anomalous Events in Transportation. The annotations and video download scripts are publicly available at https://huggingface.co/datasets/nvidia/PhysicalAI-Traffic-Anomaly-Reasoning.

Discrete diffusion models have emerged as a promising direction for vision-language tasks, offering bidirectional context modeling and theoretical parallelization. However, their practical application is severely hindered by a train-inference discrepancy, which leads to catastrophic error cascades: initial token errors during parallel decoding pollute the generation context, triggering a chain reaction of compounding errors and leading to syntactic errors and semantic hallucinations. To address this fundamental challenge, we reframe the generation process from passive denoising to active refining. We introduce ReDiff, a refining-enhanced diffusion framework that teaches the model to identify and correct its own errors. Our approach features a two-stage training process: first, we instill a foundational revision capability by training the model to revise synthetic errors; second, we implement a novel online self-correction loop where the model is explicitly trained to revise its own flawed drafts by learning from an expert's corrections. This mistake-driven learning endows the model with the crucial ability to revisit and refine its already generated output, effectively breaking the error cascade. Extensive experiments demonstrate that ReDiff significantly improves the coherence and factual accuracy of generated content, enabling stable and efficient parallel generation far superior to traditional denoising methods.

Flow matching has emerged as a standard paradigm for robotic manipulation owing to its strong expressive power for modelling complex, multimodal action distributions, alongside similar approaches like diffusion policy. However, existing methods rely on discretized action chunks, making them brittle to demonstrations collected at heterogeneous control frequencies and prone to temporally inconsistent actions that degrade control stability. In this paper, we propose Frequency-Aware Flow Matching (FAFM), which outputs continuous, temporally consistent actions. To handle heterogeneous frequency input, we transform discrete action sequences into the frequency domain with the discrete cosine transform (DCT), perform flow matching over the resulting coefficients, and reconstruct continuous actions via cosine basis expansion. To generate temporally consistent actions, we regularize the first-order temporal derivative to promote smooth actions. This corresponds to a Sobolev-type constraint that suppresses high-frequency errors and discourages abrupt action changes. Our FAFM is simple, introduces no additional network parameters and applies to standalone flow-matching policies and vision-language action models. Across synthetic toy benchmark, obstacle avoidance, LapGym, and LIBERO, FAFM improves success rates, multimodal expressivity, motion smoothness, convergence speed, robustness to mechanical bias and mixed-frequency input. These gains are consistent when deployed on a real-world Franka robot. Code available at \url{https://anonymous.4open.science/r/FAFM}.

Progress in legal AI increasingly depends on access to authoritative legal text at scale. Yet one of the most consequential layers of American law remains largely absent from existing machine-readable corpora: local ordinances. Local codes govern zoning, housing, business licensing, public health, noise, animal control, and many other domains of everyday regulation, but they are fragmented across vendor platforms designed for human browsing rather than bulk research access. We introduce LOCUS—the Local Ordinance Corpus for the United States—a comprehensive corpus and county-harmonized access layer for U.S. municipal and county ordinance codes. The raw corpus, available for release to researchers, represents nearly all publicly available municipal and county ordinance codes. The resulting raw corpus contains codes from 9,239 cities and counties. A smaller county-harmonized LOCUS access layer provides coverage for the largest 2,309 of 3,144 U.S. counties, accounting for a majority of the population. We use OCR to handle the myriad of document formats that have kept the law from being a public resource. We release the corpus with coverage metadata to support reproducibility, downstream legal AI research, and the incremental expansion of machine-readable access to local law. We train a collection of ModernBERT-based classifiers and scorers to facilitate analyzing U.S. local law among several dimensions, such as opacity and paternalism, that have not previously been studied at this scale.

Adapting pretrained Vision-Language-Action (VLA) policies to new deployment environments typically requires collecting expert demonstrations in each target domain, which is costly and difficult to scale. In this paper, we present FreeAct, a framework for demonstration-free robot adaptation that converts generated videos into physically grounded supervision. However, a key challenge is that generated manipulation videos, while visually plausible, often contain embodiment artifacts and kinematic inconsistencies that make direct inverse-dynamics labeling unreliable. FreeAct addresses this challenge by learning a discrete latent action space from large-scale multi-lab robot videos and aligning it with camera-frame end-effector SE(3) motion and gripper-state changes. The resulting latent action tokenizer captures task-relevant motion while reducing reliance on embodiment-specific visual artifacts, enabling generated target-domain videos to be pseudo-labeled with reliable latent action tokens. These tokens are first injected into VLA pretraining and then used, together with source-domain action supervision, to adapt the policy to unseen environments. Empirically, FreeAct improves success from 2\% to 57\% under target-domain shift using generated videos alone, and further to 74\% through MLLM-guided self-evolution on the policy's own rollouts, closing the adaptation loop without expert demonstrations. Further analysis shows that the learned latent action space forms a kinematically structured and transferable representation, establishing generated videos as a scalable source of supervision for robot manipulation.

Frame selection---picking a small set of informative frames from a video to feed a vision-language model (VLM)---sits at the front of every video-VLM system, yet the literature is fragmented: most works do not share a candidate pool, encoder, frame budget, or answering VLM, making published gains hard to compare across works. We introduce FrameSelect, to the best of our knowledge the first Python library that treats the selector as a standalone, interchangeable component, cleanly separated from the encoder that computes representations and the VLM that answers. Any registered selector composes with any registered evaluator through a single call, and adding a new selector, dataset, encoder, or VLM backend is an isolated extension; a replay mode further rebuilds any table under a different answering VLM without re-running selection. We release reference implementations of 20 selectors, 15 evaluators spanning answer-level and selection-level metrics, and the first unified benchmark comparing frame selectors under a shared protocol across seven evaluators spanning both evaluation modes. This comparison surfaces findings that existing studies cannot: rankings invert across benchmarks. Our library is available at https://anonymous.4open.science/r/FrameSelect-D2A4.

Long-video question answering benchmarks like Video-MME-v2 typically evaluate systems under a uniform frame budget for every question, despite stark variation in the evidence each query actually demands. We analyze this mismatch on Video-MME-v2 and find that different question subsets prefer different frozen frame-budget policies, while uniformly increasing the frame count does not reliably improve all metrics. These findings motivate fixed-mean frame-budget routing, a controlled inference setting in which each question may receive a different frame budget while the average budget matches a uniform reference. We introduce FrameRouter, a training-free test-time framework that combines an evidence-demand router, a budget-constrained allocation rule, and a frozen bank of frame-sampling policies. By controlling how much visual evidence each question receives while leaving the underlying samplers unchanged, FrameRouter is orthogonal to query-aware frame selection. Across Video-MME-v2 and several long-video understanding benchmarks, FrameRouter improves long-video question answering by using the same average visual budget more selectively across questions.

Medical image segmentation requires the precise alignment of macroscopic semantics and microscopic geometry, yet existing representation paradigms struggle to balance massive low-frequency semantic context with ultra-sparse, high-frequency boundary details. To address this fundamental information asymmetry, we introduce Fractal-G, a plug-and-play multi-scale fusion module that presents a continuous Grid-Graph-Grid feature reconstruction paradigm. Within this framework, an Uncertainty-Aware Fission Router dynamically translates regular image grids into a non-uniform heterogeneous graph, adaptively allocating dense microscopic nodes to complex boundaries while retaining sparse macroscopic nodes in homogeneous background regions. To ensure unbiased feature aggregation across these varying physical scales, an Area-Aware Continuous Neighborhood Graph explicitly incorporates physical node sizes into the topological message passing. Finally, a Topology-Aware Implicit Renderer projects the unstructured graph back into dense, artifact-free continuous feature fields. Extensive experiments on diverse medical image segmentation tasks (including skin lesions and polyps) demonstrate that Fractal-G easily integrable with mainstream backbones, consistently achieving state-of-the-art geometric fidelity and cross-resolution robustness with only 12\% GFLOPs increase (U-Net based). Our code is publicly available at https://anonymous.4open.science/r/Fractal-G/.

Automated cardiac MRI quantification is commonly framed either as dense segmentation followed by voxel counting or as direct regression of global measurements. The two paradigms are difficult to unify: segmentation is visually interpretable but label-intensive and output-heavy, while regression is compact but less verifiable. We propose Fourier Contour Learning, which represents cardiac boundaries with compact anatomical Fourier contours. The predicted coefficients define smooth, interpretable contours and enable differentiable volume computation, connecting mask supervision, measurement supervision, and unlabeled-image regularization within one framework. We instantiate this representation with an anatomy-query decoder and cross-slice attention which creates a traceable path from image to contour to clinical measurement. Across three public datasets and an in-house all-phase cohort, our method achieves segmentation accuracy comparable to strong dense-prediction baselines, improves boundary agreement, and reduces clinically relevant volume error while using only a small number of output parameters per structure. We further show that the learned model provides an efficient interface for vision-language question answering and surface reconstruction. These results demonstrate the potential of Fourier contours as a unified intermediate representation for scalable, traceable, and clinically grounded cardiac analysis. Code will be publicly available upon acceptance.

Accurate 3D surface reconstruction from unposed images remains a fundamental challenge in computer vision. Recent feed-forward pointmap methods eliminate the need for camera poses but inherently produce discrete, sparse representations. In this paper, we present a generalizable pipeline that directly recovers continuous, high-fidelity 3D surfaces from unposed images by seamlessly integrating pointmap- and ray-based representations. Our approach first extracts explicit geometric priors using a pre-trained pointmap model. A query-view-centric selection module then identifies geometrically relevant reference views via a robust surfel voting mechanism. Next, a cross-attention regressor bridges target query rays with reference features to estimate a continuous coarse surface. Finally, a raylet-based module aggregates local 3D features to carve out high-frequency details and analytical normals. Trained as a single foundation model on 9 diverse datasets, our method demonstrates exceptional zero-shot generalization on 6 unseen datasets, clearly surpassing state-of-the-art baselines in geometric accuracy and detail preservation.

Large Language Model (LLM) interactions are typically underspecified, with users clarifying all necessary details across multiple conversational turns. Yet recent work shows that LLMs perform far worse in this multi-turn setting than in a single turn with all information being available at once, a phenomenon termed ``Lost in Conversation.'' However, bridging this gap effectively and generally remains an open problem. Here we introduce Found in Conversation (FiC), a training framework where a model teaches itself to find back its single-turn competence given underspecified multi-turn prompts. We develop View-Asymmetric Self-Distillation, which runs the same self-distillation backbone on two views of the same information: the teacher sees a single-turn view that concatenates all information revealed across the conversation turns, and the student sees the multi-turn view itself; distillation aligns the student's weaker multi-turn behavior with the teacher's stronger single-turn behavior. Across diverse model sizes (3B–14B) and architectures (Llama, Qwen, Phi, and OLMo), ours achieves a 100% recovery of the single-turn performance on 2 Llama models and recovers over 90% on every model, delivering more helpful and efficient multi-turn conversations without compromising single-turn performance.

Multimodal large language models (MLLMs) are increasingly used for explainable image forgery analysis, where a shared decoder is expected to predict authenticity, localize evidence, and generate natural-language explanations. Existing methods typically rely on final decoder states, leaving open whether these late, language-oriented representations preserve localized manipulation cues. We test this assumption with a token- and layer-wise probing diagnostic that uses pixel-level masks only for analysis to separate tampered-region and authentic-region vision tokens. Across three MLLM-based forgery backbones and multiple manipulation types, we identify a mid-layer forgery evidence peak: tampered-region tokens are most linearly separable at upper-middle decoder layers and become less separable afterward, whereas authentic-region tokens continue to strengthen. This region-conditioned asymmetry suggests competition between localized manipulation cues and late semantic aggregation. Motivated by this diagnostic, we propose Gated Forensic Memory (\method), a lightweight evidence relay that reads from the diagnosed peak layer and writes re-encoded vision-token states back to late vision-token states through a zero-initialized gated residual. \method keeps the LLM frozen, adds no extra vision tokens, and requires no pixel-level training supervision. Across three backbones and seven benchmarks, \method improves authenticity prediction, explanation quality, and spatial alignment without mask supervision; the best read layer matches the diagnosed peak, turning the diagnostic into a practical layer-selection criterion.