NeurIPS 2025 Papers

Effective explanations of video action recognition models should disentangle how movements unfold over time from the surrounding spatial context. However, existing methods—based on saliency—produce entangled explanations, making it unclear whether predictions rely on motion or spatial context. Language-based approaches offer structure but often fail to explain motions due to their tacit nature—intuitively understood but difficult to verbalize. To address these challenges, we propose Disentangled Action aNd Context concept-based Explainable (DANCE) video action recognition, a framework that predicts actions through disentangled concept types: motion dynamics, objects, and scenes. We define motion dynamics concepts as human pose sequences. We employ a large language model to automatically extract object and scene concepts. Built on an ante-hoc concept bottleneck design, DANCE enforces prediction through these concepts. Experiments on four datasets—KTH, Penn Action, HAA500, and UCF101—demonstrate that DANCE significantly improves explanation clarity with competitive performance. Through a user study, we validate the superior interpretability of DANCE. Experimental results also show that DANCE is beneficial for model debugging, editing, and failure analysis.

The synthesis of realistic Martian landscape videos, essential for mission rehearsal and robotic simulation, presents unique challenges. These primarily stem from the scarcity of high-quality Martian data and the significant domain gap relative to terrestrial imagery. To address these challenges, we introduce a holistic solution comprising two main components: 1) a data curation framework, Multimodal Mars Synthesis (M3arsSynth), which processes stereo navigation images to render high-fidelity 3D video sequences. 2) a video-based Martian terrain generator (MarsGen), that utilizes multimodal conditioning data to accurately synthesize novel, 3D-consistent frames. Our data are sourced from NASA’s Planetary Data System (PDS), covering diverse Martian terrains and dates, enabling the production of physics-accurate 3D surface models at metric-scale resolution. During inference, MarsGen is conditioned on an initial image frame and can be guided by specified camera trajectories or textual prompts to generate new environments. Experimental results demonstrate that our solution surpasses video synthesis approaches trained on terrestrial data, achieving superior visual quality and 3D structural consistency.

Neural video compression (NVC) has made significant progress in recent years, while neural B-frame video compression (NBVC) remains underexplored compared to P-frame compression. NBVC can adopt bi-directional reference frames for better compression performance. However, NBVC's hierarchical coding may complicate continuous temporal prediction, especially at some hierarchical levels with a large frame span, which could cause the contribution of the two reference frames to be unbalanced. To optimize reference information utilization, we propose a novel NBVC method, termed Bi-directional Reference Harmonization Video Compression (BRHVC), with the proposed Bi-directional Motion Converge (BMC) and Bi-directional Contextual Fusion (BCF). BMC converges multiple optical flows in motion compression, leading to more accurate motion compensation on a larger scale. Then BCF explicitly models the weights of reference contexts under the guidance of motion compensation accuracy. With more efficient motions and contexts, BRHVC can effectively harmonize bi-directional references. Experimental results indicate that our BRHVC outperforms previous state-of-the-art NVC methods, even surpassing the traditional coding, VTM-RA (under random access configuration), on the HEVC datasets. The source code will be released. The source code is released at https://github.com/kwai/NVC.

We present Perceive Anything Model (PAM), a conceptually straightforward and efficient framework for comprehensive region-level visual understanding in images and videos. Our approach extends the powerful segmentation model SAM 2 by integrating Large Language Models (LLMs), enabling simultaneous object segmentation with the generation of diverse, region-specific semantic outputs, including categories, label definition, functional explanations, and detailed captions. A key component, Semantic Perceiver, is introduced to efficiently transform SAM 2's rich visual features, which inherently carry general vision, localization, and semantic priors into multi-modal tokens for LLM comprehension. To support robust multi-granularity understanding, we also develop a dedicated data refinement and augmentation pipeline, yielding a high-quality dataset of 1.5M image and 0.6M video region-semantic annotations, including novel region-level streaming video caption data. PAM is designed for lightweightness and efficiency, while also demonstrates strong performance across a diverse range of region understanding tasks. It runs 1.2$-$2.4$\times$ faster and consumes less GPU memory than prior approaches, offering a practical solution for real-world applications. We believe that our effective approach will serve as a strong baseline for future research in region-level visual understanding.

In-context learning (ICL) is one of the key capabilities contributing to the great success of LLMs. At test time, ICL is known to operate in the two modes: task recognition and task learning. In this paper, we investigate the emergence and dynamics of the two modes of ICL during pretraining. To provide an analytical understanding of the learning dynamics of the ICL abilities, we investigate the in-context random linear regression problem with a simple linear-attention-based transformer, and define and disentangle the strengths of the task recognition and task learning abilities stored in the transformer model’s parameters. We show that, during the pretraining phase, the model first learns the task learning and the task recognition abilities together in the beginning, but it (a) gradually forgets the task recognition ability to recall the priorly learned tasks and (b) relies more on the given context in the later phase, which we call (a) \textit{prior forgetting} and (b) \textit{in-context overfitting}, respectively.

Video Temporal Grounding (VTG) aims to localize relevant segments in untrimmed videos based on natural language queries and has seen notable progress in recent years. However, most existing methods suffer from two critical limitations. First, they are prone to learning superficial co-occurrence patterns—such as associating specific objects or phrases with certain events—induced by dataset biases, which ultimately degrades their semantic understanding abilities. Second, they typically assume that relevant segments always exist in the video, an assumption misaligned with real-world scenarios where queried content may be absent. Fortunately, causal inference offers a natural solution to the above-mentioned issues by disentangling dataset-induced biases and enabling counterfactual reasoning about query relevance. To this end, we propose CausalVTG, a novel framework that explicitly integrates causal reasoning into VTG. Specifically, we introduce a causality-aware disentangled encoder (CADE) based on front-door adjustment to mitigate confounding biases in visual and textual modalities. To better capture temporal granularity, we design a multi-scale temporal perception module (MSTP) that reconstructs query-conditioned video features at multiple resolutions. Additionally, a counterfactual contrastive learning objective is employed to help the model discern whether a query is truly grounded in a video. Extensive experiments on five widely-used benchmarks demonstrate that CausalVTG outperforms state-of-the-art methods, achieving higher localization precision under stricter IoU thresholds and more accurately identifying whether a query is truly grounded in the video. These results demonstrate both the effectiveness and generalizability of proposed CausalVTG.

Mask-based pretraining has become a cornerstone of modern large-scale models across language, vision, and recently biology. Despite its empirical success, its role and limits in learning data representations have been unclear. In this work, we show that the behavior of mask-based pretraining can be directly characterized by test risk in high-dimensional minimum-norm ("ridge-less") linear regression, without relying on further model specifications. Further analysis of linear models uncovers several novel aspects of mask-based pretraining. The theoretical framework and its implications have been validated across diverse neural architectures (including MLPs, CNNs, and Transformers) applied to both vision and language tasks. Guided by our theory, we propose an embarrassingly simple yet overlooked pretraining scheme named *Randomly Random Mask AutoEncoding* (**R²MAE**), which enforces capturing multi-scale features from data and is able to outperform optimal fixed mask ratio settings in our linear model framework. We implement R²MAE in vision, language, DNA sequence, and single-cell models, where it consistently outperforms standard and more complicated masking schemes, leading to improvements for state-of-the-art models. Our code is available at [this URL](https://github.com/MingzeDong/r2mae).

Long-form video understanding presents significant challenges due to extensive temporal-spatial complexity and the difficulty of question answering under such extended contexts. While Large Language Models (LLMs) have demonstrated considerable advancements in video analysis capabilities and long context handling, they continue to exhibit limitations when processing information-dense hour-long videos. To overcome such limitations, we propose the $\textbf{D}eep \ \textbf{V}ideo \ \textbf{D}iscovery \ (\textbf{DVD})$ agent to leverage an $\textit{agentic search}$ strategy over segmented video clips. Different from previous video agents manually designing a rigid workflow, our approach emphasizes the autonomous nature of agents. By providing a set of search-centric tools on multi-granular video database, our DVD agent leverages the advanced reasoning capability of LLM to plan on its current observation state, strategically selects tools to orchestrate adaptive workflow for different queries in light of the gathered information. We perform comprehensive evaluation on multiple long video understanding benchmarks that demonstrates our advantage. Our DVD agent achieves state-of-the-art performance on the challenging LVBench dataset, reaching an accuracy of $\textbf{74.2\%}$, which substantially surpasses all prior works, and further improves to $\textbf{76.0\%}$ with transcripts.

Video Anomaly Detection (VAD) aims to identify abnormal frames from discrete events within video sequences. Existing VAD methods suffer from heavy annotation burdens in fully-supervised paradigm, insensitivity to subtle anomalies in semi-supervised paradigm, and vulnerability to noise in weakly-supervised paradigm. To address these limitations, we propose a novel paradigm: Single-Frame supervised VAD (SF-VAD), which uses a single annotated abnormal frame per abnormal video. SF-VAD ensures annotation efficiency while offering precise anomaly reference, facilitating robust anomaly modeling, and enhancing the detection of subtle anomalies in complex visual contexts. To validate its effectiveness, we construct three SF-VAD benchmarks by manually re-annotating the ShanghaiTech, UCF-Crime, and XD-Violence datasets in a practical procedure. Further, we devise Frame-guided Progressive Learning (FPL), to generalize sparse frame supervision to event-level anomaly understanding. FPL first leverages evidential learning to estimate anomaly relevance guided by annotated frames. Then it extends anomaly supervision by mining discrete abnormal events based on anomaly relevance and feature similarity. Meanwhile, FPL decouples normal patterns by isolating distinct normal frames outside abnormal events, reducing false alarms. Extensive experiments show SF-VAD achieves state-of-the-art detection results while offering a favorable trade-off between performance and annotation cost.

ChebNet, one of the earliest spectral GNNs, has largely been overshadowed by Message Passing Neural Networks (MPNNs), which gained popularity for their simplicity and effectiveness in capturing local graph structure. Despite their success, MPNNs are limited in their ability to capture long-range dependencies between nodes. This has led researchers to adapt MPNNs through *rewiring* or make use of *Graph Transformers*, which compromise the computational efficiency that characterized early spatial message passing architectures, and typically disregard the graph structure. Almost a decade after its original introduction, we revisit ChebNet to shed light on its ability to model distant node interactions. We find that out-of-box, ChebNet already shows competitive advantages relative to classical MPNNs and GTs on long-range benchmarks, while maintaining good scalability properties for high-order polynomials. However, we uncover that this polynomial expansion leads ChebNet to an unstable regime during training. To address this limitation, we cast ChebNet as a stable and non-dissipative dynamical system, which we coin Stable-ChebNet. Our Stable-ChebNet model allows for stable information propagation, and has controllable dynamics which do not require the use of eigendecompositions, positional encodings, or graph rewiring. Across several benchmarks, Stable-ChebNet achieves near state-of-the-art performance.

Frontier vision-language models (VLMs) have made remarkable improvements in video understanding tasks. However, real-world videos typically exist as continuously evolving data streams (e.g., dynamic scenes captured by wearable glasses), necessitating models to continually adapt to shifting data distributions and novel scenarios. Considering the prohibitive computational costs of fine-tuning models on new tasks, usually, a small subset of parameters is updated while the bulk of the model remains frozen. This poses new challenges to existing continual learning frameworks in the context of large multimodal foundation models, i.e., catastrophic forgetting and update conflict. While the foundation models struggle with parameter-efficient continual learning, the hippocampus in the human brain has evolved highly efficient mechanisms for memory formation and consolidation. Inspired by the rapid **Bi**nding and pattern **se**paration mechanisms in the hippocampus, in this work, we propose **Bisecle** for video-language **c**ontinual **le**arning, where a multi-directional supervision module is used to capture more cross-modal relationships and a contrastive prompt learning scheme is designed to isolate task-specific knowledge to facilitate efficient memory storage. Binding and separation processes further strengthen the ability of VLMs to retain complex experiences, enabling robust and efficient continual learning in video understanding tasks. We perform a thorough evaluation of the proposed Bisecle, demonstrating its ability to mitigate forgetting and enhance cross-task generalization on several VideoQA benchmarks.

Properly evaluating the ability of Video-Language Models (VLMs) to understand long videos remains a challenge. We propose a long-context video understanding benchmark, Causal2Needles, that assesses two crucial abilities insufficiently addressed by existing benchmarks: (1) extracting information from two separate locations (two needles) in a long video and understanding them jointly, and (2) modeling the world in terms of cause and effect in human behaviors. Causal2Needles evaluates these abilities using noncausal one-needle, causal one-needle, and causal two-needle questions. The most complex question type, causal two-needle questions, require extracting information from both the cause and effect events from a long video and the associated narration text. To prevent textual bias, we introduce two complementary question formats: locating the video clip containing the answer, and verbal description of a visual detail from that video clip. Our experiments reveal that models excelling on existing benchmarks struggle with causal 2-needle questions, and the model performance is negatively correlated with the distance between the two needles. These findings highlight critical limitations in current VLMs.

Recent large-scale reasoning models have achieved state-of-the-art performance on challenging mathematical benchmarks, yet the internal mechanisms underlying their success remain poorly understood. In this work, we introduce the notion of a reasoning graph, extracted by clustering hidden‐state representations at each reasoning step, and systematically analyze three key graph-theoretic properties: cyclicity, diameter, and small-world index, across multiple tasks (GSM8K, MATH500, AIME~2024). Our findings reveal that distilled reasoning models (e.g., DeepSeek-R1-Distill-Qwen-32B) exhibit significantly more recurrent cycles (about 5 per sample), substantially larger graph diameters, and pronounced small-world characteristics (about 6x) compared to their base counterparts. Notably, these structural advantages grow with task difficulty and model capacity, with cycle detection peaking at the 14B scale and exploration diameter maximized in the 32B variant, correlating positively with accuracy. Furthermore, we show that supervised fine-tuning on an improved dataset systematically expands reasoning graph diameters in tandem with performance gains, offering concrete guidelines for dataset design aimed at boosting reasoning capabilities. By bridging theoretical insights into reasoning graph structures with practical recommendations for data construction, our work advances both the interpretability and the efficacy of large reasoning models.

Estimating scene lighting from a single image or video remains a longstanding challenge in computer vision and graphics. Learning-based approaches are constrained by the scarcity of ground-truth HDR environment maps, which are expensive to capture and limited in diversity. While recent generative models offer strong priors for image synthesis, lighting estimation remains difficult due to its reliance on indirect visual cues, the need to infer global (non-local) context, and the recovery of high-dynamic-range outputs. We propose LuxDiT, a novel data-driven approach that fine-tunes a video diffusion transformer to generate HDR environment maps conditioned on visual input. Trained on a large synthetic dataset with diverse lighting conditions, our model learns to infer illumination from indirect visual cues and generalizes effectively to real-world scenes. To improve semantic alignment between the input and the predicted environment map, we introduce a low-rank adaptation finetuning strategy using a collected dataset of HDR panoramas. Our method produces accurate lighting predictions with realistic angular high-frequency details, outperforming existing state-of-the-art techniques in both quantitative and qualitative evaluations.

Contrastive learning involves learning representations via a loss function that encourages each (unlabeled) sample to be far from other samples, but close to its own *augmentation*. In this paper, we aim to understand why this simple idea performs remarkably well, by theoretically analyzing it for a simple, natural problem setting: dimensionality reduction in Gaussian Mixture Models (GMMs). Note that the standard GMM setup lacks the concept of augmentations. We study an intuitive extension: we define the pair of data sample and its augmentation as a coupled random draw from the GMM such that the marginal over the "noisy" augmentation is *biased* towards the component of the data sample. For this setup, we show that vanilla contrastive loss, e.g., InfoNCE, is able to find the *optimal* lower-dimensional subspace even when the Gaussian components are non-isotropic. In particular, we show that InfoNCE can match the performance of a fully supervised algorithm, e.g., LDA, (where each data point is labeled with the mixture component it comes from) even when the augmentations are "noisy". We further extend our setup to the multi-modal case, and develop a GMM-like setting to study the contrastive CLIP loss. We corroborate our theoretical with real-data experiments on CIFAR100; representations learned by InfoNCE loss match the performance of LDA on clustering metrics.

This work presents Orient Anything V2, an enhanced foundation model for unified understanding of object 3D orientation and rotation from single or paired images. Building upon Orient Anything V1, which defines orientation via a single unique front face, V2 extends this capability to handle objects with diverse rotational symmetries and directly estimate relative rotations. These improvements are enabled by four key innovations: 1) Scalable 3D assets synthesized by generative models, ensuring broad category coverage and balanced data distribution; 2) An efficient, model-in-the-loop annotation system that robustly identifies 0 to N valid front faces for each object; 3) A symmetry-aware, periodic distribution fitting objective that captures all plausible front-facing orientations, effectively modeling object rotational symmetry; 4) A multi-frame architecture that directly predicts relative object rotations. Extensive experiments show that Orient Anything V2 achieves state-of-the-art zero-shot performance on orientation estimation, 6DoF pose estimation, and object symmetry recognition across 11 widely used benchmarks. The model demonstrates strong generalization, significantly broadening the applicability of orientation estimation in diverse downstream tasks.

Recently, the powerful text-to-image capabilities of GPT-4o have led to growing appreciation for native multimodal large language models. However, its multimodal capabilities remain confined to images and text. Yet beyond images, the ability to understand and generate 3D content is equally crucial. To address this gap, we propose ShapeLLM-Omni—a native 3D large language model capable of understanding and generating 3D assets and text in any sequence. First, we train a 3D vector-quantized variational autoencoder (VQVAE), which maps 3D objects into a discrete latent space to achieve efficient and accurate shape representation and reconstruction. Building upon the 3D-aware discrete tokens, we innovatively construct a large-scale continuous training dataset named 3D-Alpaca, encompassing generation, comprehension, and editing, thus providing rich resources for future research and training. Finally, we perform instruction-based fine-tuning of the Qwen-2.5-vl-7B-Instruct model on the 3D-Alpaca dataset, equipping it with native 3D understanding and generation capabilities. Our work represents an effective step toward extending multimodal large language models with fundamental 3D intelligence, paving the way for future advances in 3D-native AI.

Reinforcement fine-tuning (RFT) is essential for enhancing the reasoning and generalization capabilities of large language models, but its success heavily relies on the quality of the training data. While data selection has been extensively studied in supervised learning, its role in reinforcement learning, particularly during the RFT stage, remains largely underexplored. In this work, we introduce RFT-Inf, the first influence estimator designed for data in reinforcement learning. RFT-Inf quantifies the importance of each training example by measuring how its removal affects the final training reward, offering a direct estimate of its contribution to model learning. To ensure scalability, we propose a first-order approximation of the RFT-Inf score by backtracking through the optimization process and applying temporal differentiation to the sample-wise influence term, along with a first-order Taylor approximation to adjacent time steps. This yields a lightweight, gradient-based estimator that evaluates the alignment between an individual sample’s gradient and the average gradient direction of all training samples, where a higher degree of alignment implies greater training utility. Extensive experiments demonstrate that RFT-Inf consistently improves reward performance and accelerates convergence in reinforcement fine-tuning.

Dynamic Novel View Synthesis aims to generate photorealistic views of moving subjects from arbitrary viewpoints. This task is particularly challenging when relying on monocular video, where disentangling structure from motion is ill-posed and supervision is scarce. We introduce Video Diffusion-Aware Reconstruction (ViDAR), a novel 4D reconstruction framework that leverages personalised diffusion models to synthesise a pseudo multi-view supervision signal for training a Gaussian splatting representation. By conditioning on scene-specific features, ViDAR recovers fine-grained appearance details while mitigating artefacts introduced by monocular ambiguity. To address the spatio-temporal inconsistency of diffusion-based supervision, we propose a diffusion-aware loss function and a camera pose optimisation strategy that aligns synthetic views with the underlying scene geometry. Experiments on DyCheck, a challenging benchmark with extreme viewpoint variation, show that ViDAR outperforms all state-of-the-art baselines in visual quality and geometric consistency. We further highlight ViDAR’s strong improvement over baselines on dynamic regions and provide a new benchmark to compare performance in reconstructing motion-rich parts of the scene.