Video Representation Learning

Understanding motion in video is a fundamental challenge for visual learning, as frame-to-frame change entangles two sources of dynamics: camera motion and object motion. This decomposition has remained underexplored in representation learning, partly because these factors are tightly coupled in natural videos and difficult to supervise separately. Yet recovering it is important for learning robust motion representations that separate meaningful object dynamics from camera-induced variation. We study whether such structured motion representations can be recovered from frozen features of a pretrained image vision transformer. We propose the Structured Dynamics Model (SDM), which explicitly separates the dominant source of temporal change from residual dynamics through future-feature prediction, rather than representing video change with a single entangled latent or with unstructured, spatially dense transition tokens. Training combines self-supervised learning on real video with weak supervision of scene dynamics on synthetic Kubric data. We evaluate SDM on ProbeMotion, a new evaluation suite spanning synthetic and real videos with camera motion, object motion, and combined dynamics. SDM outperforms backbone baselines using global CLS or average-pooled features, and compares favorably to strongly supervised representations such as VGGT on several probes, despite using substantially weaker supervision. These results suggest that pretrained image models can be readily repurposed into structured video-dynamics representations, providing a useful inductive bias for learning and analyzing latent video dynamics.

We examine whether richer visual representations yield more human-aligned measures of urban engagement, using 61 first-person city-walk videos from YouTube segmented into over 50,000 ten-second clips and represented across four modalities: spatiotemporal video features, temporally averaged images (TAIs), audio embeddings, and text-based semantic descriptions. Spearman correlation analysis reveals the expected ordering along the temporal-richness continuum, with video features showing the strongest continuous alignment. However, this ordering breaks down under binary classification of high- versus low-engagement moments (the paradigm most commonly used to train perceptual scoring models), where TAIs consistently match or outperform video across most classifiers and quantile thresholds. An independent two-alternative forced-choice study on Amazon Mechanical Turk confirms that this parity reflects human judgment: participants identified engaging moments with comparable accuracy from TAIs and full video clips, while text performed substantially worse and audio remained near chance. Gap analysis reveals a functional dissociation: video features are advantaged in activity-driven scenes with dynamic content, whereas TAIs better align with human judgments in composition-driven scenes dominated by stable spatial structure. These findings challenge the assumption that richer representations are inherently more human-aligned, and suggest that perceptually grounded temporal compression can be a principled alternative to full video encoding.

Recognizing instrument--tissue interactions is essential for context-aware surgical AI. Vision-language models offer a natural way to inject semantic structure into surgical representations by aligning video features with textual action descriptions. However, pretrained encoders may lack spatial coherence, while global semantic alignment does not ensure precise spatial and temporal representations. By analyzing frame-to-frame feature changes, we find that semantic alignment increases their dimensionality, but larger increases do not necessarily improve recognition; encoders also differ in how strongly dominant changes localize to interaction regions. Motivated by these findings, we introduce \ours{}, which compresses frame-to-frame changes into latent actions and predicts next-frame features during end-to-end video--language alignment. Without additional spatial or motion annotations, \ours{} improves the interaction grounding of leading feature changes and temporal-direction sensitivity in our evaluated settings. We further characterize how action capacity and prediction strength affect recognition across encoders and triplet components. Using image encoders without large-scale video pretraining, \ours{} achieves competitive recognition with faster inference and smaller INT4 accuracy drops than V-JEPA2/2.1.

Video generative models commonly rely on latent spaces learned by 3D Variational Autoencoders (3D-VAEs). However, conventional 3D-VAEs are mainly optimized for pixel-level reconstruction, which can limit the semantic and spatio-temporal structure captured by their latents. Meanwhile, Video Foundation Models (VFMs) such as V-JEPA 2 and VideoMAEv2 show strong video understanding capabilities, yet whether their frozen representations can be transformed into compact, reconstruction-capable, and generation-friendly video latents remains largely unexplored. We answer this question with VideoRAE, a representation autoencoder that leverages multi-scale hierarchical features from a frozen video foundation encoder and compresses them with a lightweight 1D self-attention projector. VideoRAE supports both continuous latents for Diffusion Transformers and discrete tokens for autoregressive models via multi-codebook high-dimensional quantization. During decoding, a local-and-global representation alignment objective with the frozen VFM teacher improves semantic preservation and enables training without KL regularization. Experiments show that VideoRAE achieves strong reconstruction in both continuous and discrete regimes. On UCF-101, it obtains state-of-the-art class-to-video gFVDs of 40 and 93 with AR and DiT generators, respectively, while converging approximately 5x faster than competing autoencoder baselines. In a controlled 2B-scale text-to-video study, replacing LTX-VAE with VideoRAE leads to faster convergence under comparable settings. These results validate frozen VFM representations as versatile and generation-friendly video latents. The model and code will be released on https://zhxie0117.github.io/VideoRAE.

Gaussian Splatting has emerged as an effective representation for video, but existing methods rely on per-video optimization. This leads to slow encoding and limits generalization across videos. To amortize this optimization, we propose HyperGS, a feedforward, optimization-free approach that directly predicts Gaussian representations from any video in a single forward pass, speeding up encoding and decoding by orders of magnitude while generalizing to out-of-distribution videos at higher resolutions. In HyperGS, we design a factorized spatiotemporal Transformer to extract tokens from video, and a learnable query-based Transformer to obtain 8-parameter Gaussian representations for each video frame. We find that naively predicting Gaussians across diverse videos induces a needle-like degeneration that collapses training, and address this with a rank-based geometric regularizer whose strength adapts dynamically to stabilize optimization. HyperGS achieves encoding at $10^4$--$10^5\times$ the speed of per-video Gaussian optimization at matched reconstruction quality while generalizing zero-shot to $720p$ video, enabling higher-resolution rendering without re-encoding. HyperGS improves PSNR by +2.9--3.1 dB over the prior video encoders on K400, SSv2, and UCF101 at a smaller video representation size. By predicting explicit 2D Gaussians in a single forward pass, HyperGS combines the fast, flexible rendering of Gaussian Splatting with the speed and generalization of feedforward prediction, advancing Gaussians as a practical direction for fast and generalizable video representation.

Video object segmentation (VOS) is a fundamental task in video understanding, requiring accurate delineation and consistent tracking of objects across frames. While supervised methods achieve strong performance, they rely on densely annotated datasets that are costly to obtain and have limited domain coverage. Self-supervised learning offers a promising alternative by removing the need for manual labels; however, existing approaches often struggle to jointly maintain spatial accuracy and temporal coherence, particularly in unconstrained multi-object scenarios. Many rely on optical flow, synthetic motion cues, or task-specific pretraining, limiting scalability and generalisation. We propose a self-supervised framework, Cross-Temporal Consistency and Clustering, that learns mid-level, part-aware representations by combining attention-guided token selection with lightweight temporal clustering. Instead of operating at the pixel or whole-object level, the method aligns soft part assignments across time using a saliency-weighted symmetric consistency objective. The framework leverages a frozen transformer backbone with lightweight modules for adaptive token selection and multi-offset temporal alignment, enabling efficient scaling across resolutions and motion patterns.

Prior work suggests that diffusion representations capture low-level geometry but struggle with high-level semantics. We demonstrate that state-of-the-art video diffusion models overcome this limitation. By systematically probing their intermediate activations using recent mutual-kNN alignment metrics, we reveal a highly structured latent space where visual representations evolve across both network depth and noise levels. We show that while moderate noise levels yield linearly separable global semantics, fine-grained details persist at lower noise levels but become spatially scattered, requiring attention mechanisms to decode. Building on these insights, we introduce Gen4U (Generation for Understanding), a framework that repurposes these generative representations with a single forward pass. Our experiments establish that frozen, large-scale video diffusion models function as highly competitive video encoders across a wide spectrum of tasks, spanning semantic and non-semantic objectives (video classification, depth estimation, camera pose estimation, image and video captioning). Bypassing fine-tuning, Gen4U unifies the generation and understanding paradigms, achieving strong perception performance while fully preserving the model's ability to generate high-quality video.

Unsupervised pre-training on large-scale datasets has demonstrated significant potential for improving the sample efficiency and performance of Reinforcement Learning (RL). Given the large-scale action-free internet videos, existing methods utilize single-step transition prediction and image reconstruction to learn representations. However, these methods prefer to preserve large-proportion stationary information in the pixel space, neglecting small but crucial information. To preserve enough information in the representation, it is essential to pay equal attention to each element in videos. Specifically, we propose a temporal correlation space to distinguish each element. For implementation, we introduce the Multi-scale Temporal Contrastive Learning (MTCL) method to model multi-scale temporal correlations separately. This approach can balance the attention of different elements and yield more informative representations, effectively supporting policy learning in various downstream tasks. Experimental results demonstrate that our method improves sample efficiency and asymptotic performance across various downstream tasks.

Pre-training on large-scale videos to improve reinforcement learning efficiency is promising yet remains challenging. Existing methods typically treat the agent as an indivisible entity, modeling motion patterns globally. Such global modeling is tightly coupled with the morphology, hindering transfer across domains. In contrast, despite the vast disparity in global motions, the local components exhibit similar motion patterns across different agents. Building on this insight, we propose a novel Deconstruct-Recompose Paradigm (DRP) for learning transferable local motion representations. Specifically, in the Deconstruct phase, we identify multiple local points and track their frame-wise motions, defining each as an Atomic Action. We introduce a Dual-Attention Encoder (DAE) to learn local motion representations from these Atomic Actions, capturing their spatiotemporal relationships. In the Recompose phase, we compose local motion representations with a learnable Motion Aggregation Token [MAT] via latent dynamics model learning. Additionally, an adapter bridges local motion and downstream action-specific dynamics to accelerate policy learning. Extensive experiments demonstrate that our method effectively transfers to diverse robotic control and manipulation tasks, significantly improving sample efficiency and performance.

We present a self-supervised framework for learning implicit 3D physical dynamics directly from video-derived supervisory signals. While current generative video models achieve high visual fidelity, they lack a 3D geometric foundation, often resulting in physical inconsistencies and a failure to maintain object permanence. We address this by shifting the predictive bottleneck from 2D image space to a `lifted' 3D Volumetric Latent Space. Our method unprojects semantic features from a Video Joint-Embedding Predictive Architecture (V-JEPA) into a voxelized grid, grounded by monocular depth priors. This lifting enables a Volumetric Feature Advection to learn an action-conditioned transition operator that treats physics as a spatio-temporal state advection problem, i.e., learn implicit 3D physics. Unlike state-of-the-art hybrid models that rely on explicit classical simulators for training and/or inference, our architecture tracks material states implicitly within high-dimensional V-JEPA features. This allows for the emergent simulation of heterogeneous phenomena (e.g., rigid body motion in fluid flow) within a single, unified pipeline. Supervised solely via end-to-end video-derived signal plus action conditions, without access to physics engine internal states, labels, or surrogate models, our model demonstrates good long-term structural stability and physical plausibility on multiple benchmarks (CLEVERER, PhysInOne, PhysGaia). We believe that this work opens a scalable pathway toward general-purpose dynamic world models that internalize the 3D invariants of the physical world solely through passive observation of monocular videos.

The increasing maturity of embodied AI platforms has driven a growing interest in procedural video representation learning to support intelligent assistance systems for complex, multi-step tasks. Leveraging large-scale latent predictive training, video foundation models capture video dynamics, enabling downstream tasks such as activity understanding, spatiotemporal localization, and predictive control. However, procedural videos include actions with long-range dependencies that these models do not support, due to the quadratic complexity of self-attention. Distinct actions, for example, may be visually similar despite appearing at different points in the procedure, such as turning the stove on versus off. Here, we propose a backbone-agnostic approach that learns long-duration video representations by reducing the problem to a dense, frame-aligned action space and predicting pooled masked latent vectors. This approach allows our Procedural Joint Embedding Predictive Architecture (P-JEPA) to ingest videos over 30 minutes long, enabling effective long-form understanding of procedural steps. We evaluate P-JEPA using features extracted with VJEPA2.1, TSM, and I3D over the EgoExo4D, EgoProceL, and Assembly101 datasets, finding that it consistently improves linear separability, streaming inference, and temporal action segmentation performance, achieving state-of-the-art results on EgoExo4D fine-grained action classification while using an order of magnitude fewer parameters than LLM-based methods and running in real time.

Supervised video pretraining is a common transfer learning practice for improving downstream action recognition performance. However, it requires large-scale labeled source datasets, and the effectiveness of the learned initialization is influenced by the similarity between the source and target domains. Constructing such labeled pretraining datasets for different target domains is costly and difficult to scale. To address these limitations, this study proposes a label-efficient video learning framework that combines annotation-free video pretraining with target-label-set-aware fine-tuning. During pretraining, a vision-language model (VLM) generates textual descriptions of unlabeled videos, which are processed to construct an interpretable semantic pseudo-label space. A frozen video-language model then produces zero-shot soft target distributions over this space, allowing a student video encoder to learn semantically rich representations without manual source annotations. During downstream adaptation, target-label-set-aware fine-tuning combines supervised learning from labeled target videos with zero-shot distillation over the actual target label set, helping preserve VLM-derived semantic guidance while adapting the pretrained encoder to the target task. Experiments on UCF101 and HMDB51 show that the proposed framework outperforms the compared semi-supervised video action recognition methods across all evaluated limited-label regimes. Moreover, the annotation-free pretraining stage learns transferable representations that provide an effective initialization for full-data fine-tuning, despite relying on a comparatively modest unlabeled pretraining pool.

Egocentric video understanding is inherently limited by the narrow perspective of wearable cameras: a single viewpoint, a single modality, a single model cannot capture the full richness of human action. We argue that a truly expressive egocentric representation must subsume complementary knowledge across viewpoints, modalities, and foundation model representations, yet remain deployable from egocentric video alone. To this end, we introduce a hierarchical multi-teacher distillation framework that produces UNIEGO, a unified egocentric encoder trained with nine teachers spanning ego-exo viewpoints, RGB, depth, and skeleton modalities, and four foundation models. Rather than distilling directly from heterogeneous teachers whose incompatible architectures and feature geometries induce conflicting gradients, our framework interposes a layer of representation-specific Proxy models that translate diverse teacher knowledge into a homogeneous egocentric space. A second distillation stage, Selective Proxy Distillation (SPD), then adaptively selects, for each training sample, the subset of proxies that are both correct and confident, distilling exclusively from reliable supervision and suppressing erroneous signals. SPD is further stabilized by initializing UNIEGO as a learned convex combination of proxy parameters, placing the unified model in a well-conditioned region of the loss landscape before distillation begins. UNIEGO achieves state-of-the-art performance across three egocentric video understanding tasks - action recognition, video retrieval, and action segmentation on three challenging ego-exo benchmarks, outperforming naive multi-teacher distillation baselines and demonstrating that structured, proxy-mediated knowledge transfer yields richer and more discriminative egocentric representations.

In today's media-driven world, the exponential growth of video content across domains such as surveillance, education, and entertainment has made retrieving semantically relevant videos via natural language queries increasingly critical. Early video retrieval systems relied on handcrafted features or shallow cross-modal mappings, limiting their ability to capture complex semantics and temporal dynamics. While large-scale vision-language models have improved cross-modal alignment, challenges remain in modeling fine-grained temporal dependencies and nuanced linguistic structures. In this paper, we introduce DREAM: Dual-path Representation Enhancement and Alignment Model, a novel multimodal framework that addresses these limitations through enhanced visual and textual encoding. DREAM incorporates a hybrid language modeling strategy that combines masked and permuted language modeling objectives to capture both local and global linguistic semantics. On the visual side, we design a hierarchical vision encoder with cascaded group attention, which integrates spatial and temporal information through multi-stage token interaction and coarse-to-fine attention refinement. We validate DREAM through comprehensive evaluations on the widely-used MSRVTT, MSVD and LSMDC benchmark datasets, where it achieves new state-of-the-art R1 scores of 49.4%, 49.7% and 27.3%, respectively. Qualitative analyses further show the model's ability to maintain coherent attention across frames and align complex queries with dynamic video content. These findings underscore the effectiveness of hierarchical attention and dual-objective textual modeling in enabling robust, context-aware video retrieval, and pave the way for future research in advancing cross-modal representation learning.

Text-to-video retrieval in operating rooms (OR) is an enabling technology for OR safety, as it allows stakeholders to retrieve and inspect recordings of specific events. However, because the most safety-critical events may not follow the common structure, to unlock its full potential text-to-video retrieval must be able to handle implicit queries that require reasoning to identify the right video (e.g., the step right before clipping). However, existing methods rely on global embeddings that cannot reason over such queries. We propose OR3, a text-to-video retrieval method that converts clips into action-driven digital twins (ActDTs), grouping concurrent subject-action-object triplets under non-overlapping temporal intervals. Moreover, rather than cross-modal matching through paired encoders, OR3 performs imagination-based retrieval where an LLM generates hypothetical ActDTs from queries. This enables intra-modal matching via a single encoder trained with ActDT-tailored hard negatives. Finally, evidence-grounded refinement revises imagined ActDTs based on discrepancies with top candidates to capture procedure-specific patterns. We construct a benchmark from MM-OR with 276 implicit queries across four reasoning categories over 386 clips from robotic knee procedures. OR3 achieves 57.6 R@1 and 77.3 R@5, outperforming the strongest baseline. These results demonstrate that OR3 enables fine-grained discrimination between visually similar OR video clips through temporal action reasoning.

Progress in AI has largely been driven by methods that assume less. As compute and data increase, approaches with weaker inductive biases generally outperform those with stronger assumptions. This is particularly characteristic of the field of Visual Representation Learning, where approaches have gone from being dominated by Supervised Learning, to Weakly Supervised Learning, to the now widespread success of Self-Supervised Learning without human labels. Yet, even modern Self-Supervised Learning approaches still depend on strong inductive biases such as augmentations, masking, or cropping. If this trend holds, even these remaining biases should become bottlenecks at scale -- and our experiments confirm this: the optimal strength of inductive biases decreases as data grows. This motivates the search for approaches that rely on fewer assumptions. To this end, we introduce Temporal Difference in Vision (TDV), a new paradigm for self-supervised learning from video that avoids existing inductive biases, relying instead on a causal assumption that the past causes the future. TDV functions by jointly training an image encoder and a motion encoder so that the current frame's representation plus the encoded motion equals the next frame's representation. Despite not leveraging any strong inductive biases, TDV matches state-of-the-art recipes on dense spatial tasks, laying the foundation for representation learning without strong assumptions.

Typical video object-centric learning (VOCL) approaches employ slot-based frameworks that rely on reconstruction-driven encoder-decoder architectures, where learning is mediated by two spatial maps: attention maps from the encoder and object maps from the decoder. As these two distinct maps exhibit different properties, a recent dense alignment strategy attempted to reconcile this discrepancy by enforcing agreement across all spatio-temporal patches via contrastive learning. However, this indiscriminate alignment inadvertently propagates the inherent weaknesses of each module, such as noisy encoder predictions and blurred decoder boundaries. Moreover, computing dense similarities across all pairs incurs a computational cost quadratic in the total number of spatio-temporal patches, severely limiting scalability. Motivated by this, we propose Selective Synergistic Learning (SSync). Instead of exhaustive patch-to-patch alignment, SSync prevents error propagation by selectively distilling only the most reliable cues: leveraging the encoder strictly for boundary refinement and the decoder for interior denoising. This is realized via a pseudo-labeling with linear complexity, eliminating the need for quadratic spatial comparisons. Also, to prevent the reinforcement of architectural biases like slot redundancy, we introduce a transitive pseudo-label merging that consolidates overlapping slots based on spatio-temporal activation consistency. Extensive studies demonstrate that SSync improves decomposition quality and serves as a versatile, plug-and-play module while also exhibiting exceptional robustness to slot configurations. Code is available at github.com/wjun0830/SSync.

We introduce the Massive Video Embedding Benchmark (MVEB), a 23-task benchmark for video embeddings spanning classification, zero-shot classification, clustering, pair classification, retrieval, and video-centric question answering. We evaluate 33 models and find that no single model dominates: MLLM-based embeddings lead on classification, clustering, pair classification, and QA; multimodal binding leads on retrieval and zero-shot classification; generative MLLMs without contrastive adaptation collapse on cross-modal tasks. Paired video-only vs. audio+video evaluations show that audio's contribution depends on dataset annotation provenance: audio helps when labels were produced from both modalities and hurts when they were produced from visuals alone, a six-point gap consistent across model families. MVEB is derived from MVEB+, a 184-task pool, and is designed to maintain task diversity while reducing evaluation cost. It integrates into the MTEB ecosystem for unified evaluation across text, image, audio, and video. We release MVEB and all 184 tasks along with code and a leaderboard at https://github.com/embeddings-benchmark/mteb.

The explosive growth of user-generated video content on online platforms is accompanied by the emergence of numerous near-duplicate videos--videos that are identical or highly similar but differ by partial edits. These duplicates degrade user experience and increase storage and bandwidth costs, making large-scale video deduplication a critical task. Existing video deduplication frameworks face a fundamental challenge in retrieving sufficient high-quality candidates under a limited index budget, as well as trade-offs between efficiency and precision. To address these issues, we propose MLT-Dedup, an efficient large-scale online video deduplication framework with Multi-Level representations and spatial-Temporal matching. Our approach employs a Multi-Level Video Encoder (ML-VE) to extract both fine-grained frame-level and sparse clip-level embeddings: sparse embeddings support efficient candidate retrieval, while fine-grained embeddings are loaded for precise pairwise matching. During matching, we introduce DiF-SiM, a Differential Feature-enhanced Similarity Module capable of locating duplicated temporal segments and providing reliable similarity evidence to support policy-driven deduplication decisions. Extensive experiments on a real-world large-scale platform demonstrate that MLT-Dedup reduces online repetition rates by 91% at 90% precision. Furthermore, our sparse retrieval design achieves a 5x increase in indexing capacity, enabling broader candidate coverage in real-world deployment.

Self-supervised video representation learning has recently advanced through contrastive learning, masked reconstruction, and predictive representation learning. Reconstruction-based approaches such as MAE and VideoMAE learn representations by recovering masked visual content \cite{he2022mae,tong2022videomae}, while contrastive methods such as CLIP learn semantically meaningful embedding spaces through representation alignment \cite{radford2021clip}. In this work, we introduce a Momentum-Guided Semantic Forecasting framework (MoFore) for self-supervised video representation learning. Instead of optimizing for pixel-level reconstruction or task-specific semantic alignment, the proposed method learns temporally predictive video representations by forecasting future latent embeddings from temporally distant context clips. To improve robustness across temporal scales, we further introduce randomized temporal-gap forecasting during training. The framework combines predictive latent forecasting with contrastive regularization to encourage temporal consistency while preventing representation collapse. Experiments on the UCF101 dataset demonstrate that the proposed framework learns temporally consistent and semantically meaningful video representations without using action labels during training. Quantitative analysis shows strong temporal stability and emergent category-level structure in the learned embedding space, while qualitative retrieval experiments reveal motion-aware organization across related activities. Overall, the results suggest that long-range latent forecasting provides an effective and computationally efficient approach for self-supervised video representation learning without relying on reconstruction-based objectives.

We study whether pretrained video foundation models encode intuitive-physics information in their frozen representations, and how this information varies across model families, layers, and probe types. Using frozen-feature probing on IntPhys2 and Minimal Video Pairs (MVP), we compare predictive joint-embedding models (V-JEPA), masked reconstruction models (VideoMAE), and a diffusion-based video generator (LTX-Video). V-JEPA achieves the strongest overall results across benchmarks, especially with probes that model temporal dynamics, while VideoMAE remains competitive and LTX-Video recovers weaker but non-trivial signal. Layerwise analyses show that physics-relevant information is weakest in early layers and becomes most accessible at intermediate-to-late depth, and temporal controls show that disrupting frame order substantially reduces performance, especially on MVP. Together, these results suggest that intuitive-physics knowledge emerges reliably in pretrained video representations, but its accessibility depends strongly on pretraining paradigm, representational depth, and readout mechanism.

In this paper, we present XInsight Lab's solution to the micro-gesture classification track of the 4th MiGA Challenge at IJCAI 2026, in which our solution ranked first and achieved a new state-of-the-art result. We propose a multimodal ensemble framework that integrates a self-supervised RGB-based model with supervised multi-stream models from previous solutions. The self-supervised RGB model is pretrained on 120K unlabeled clips via masked video modeling and then fine-tuned on iMiGUE. This simple yet effective RGB baseline achieves 69.224% top-1 accuracy on the iMiGUE test set, demonstrating the benefit of learning transferable representations from unlabeled in-domain videos. By incorporating this model as a complementary branch, the final ensemble reaches 74.419% top-1 accuracy, surpassing the previous state of the art by 1.206 percentage points. Experimental results on iMiGUE, including ablation studies on the ensemble strategy, validate the effectiveness of self-supervised RGB representation learning for micro-gesture recognition.

Video world models are increasingly used to provide predictive visual representations, yet it remains unclear which pretraining signals induce action-relevant structure in their latent spaces. We study this question through a unified probe-based evaluation across diverse encoder families, including image-only self-supervision, video pretraining with and without latent prediction, reconstruction-based autoencoders, diffusion models, and shortcut-forcing dynamics models. Using a common inverse-dynamics probing objective, we find that action-relevant structure is driven primarily by temporal video pretraining rather than pixel reconstruction fidelity: models with strong pixel decoding quality can exhibit near-zero action recoverability, while video-pretrained self-supervised encoders consistently achieve the best Pareto trade-off between visual fidelity and action prediction. Comparing V-JEPA and VideoMAE further shows that most gains arise from natural-video temporal context, with feature-level latent prediction providing a smaller additional benefit. These trends transfer across robotic benchmarks, though CALVIN reveals that static-environment tasks can partially mask the importance of temporal structure by allowing strong image priors to suffice. Finally, inverse-dynamics supervision substantially improves robustness to visual corruption, suggesting that action-aware objectives regularize latent geometry beyond clean-setting performance. Our results identify temporal predictive structure -- not reconstruction fidelity -- as the primary ingredient underlying action-relevant video representations.

Multimodal Large Language Models (MLLMs) excel at 2D semantic understanding but lack intrinsic 3D awareness, resulting in representations that fail to maintain geometric and spatial consistency across video frames. Given the scarcity of large-scale 3D data, we present GeoVR, a novel framework that learns geometric representations using purely 2D video sequences. This approach effectively restructures the semantic latent space within MLLMs to unlock spatial intelligence. Rather than employing superficial feature mixing, GeoVR reshapes the internal representations of the MLLM by distilling geometry knowledge from pre-trained 3D foundation models. This is accomplished through a multi-objective learning strategy driven by four complementary geometric targets: (1) estimating inter-frame camera poses to embed varying viewpoint dynamics, (2) regressing dense depth maps to anchor physical distances, (3) predicting a metric scale factor for real-world calibration, and (4) distilling multi-scale 3D features to align the intermediate feature space. Guided by these explicit physical and geometric constraints, the model's internal representations naturally develop strong 3D awareness. Extensive experiments on spatial reasoning benchmarks demonstrate that GeoVR achieves state-of-the-art performance, establishing a new paradigm for endowing foundation models with spatial intelligence.

Modern video diffusion models generate increasingly realistic and temporally coherent videos, motivating their use as candidate world simulators. Yet it remains unclear whether these models internally encode physical structure, or merely reproduce motion patterns seen during training. We study this question by probing video diffusion models along latent trajectories corresponding to real videos with known physical plausibility. To obtain such trajectories, we approximately invert the deterministic sampling process by integrating the learned velocity field backward from a clean video latent to noise, giving access to the model's intermediate states and attention maps. Using these recovered trajectories, we show that physical plausibility is linearly decodable from diffusion transformer states across IntPhys and InfLevel, reaching around 81.27% average accuracy and outperforming dedicated representation-learning baselines such as V-JEPA and VideoMAE. Surprisingly, this signal is absent from the VAE latent input and emerges inside the denoising transformer itself, despite the model not being trained with a self-supervised predictive objective. These findings suggest that physically meaningful representations can arise as a byproduct of generative denoising.

Modern action recognition models operate on memory- and compute-intensive dense RGB video volumes and frequently exploit appearance and background shortcuts, for example, predicting actions from objects or scenes instead of characteristic motion. We investigate an efficient alternative input modality that is largely free of such biases by construction: sparse point trajectories. To this end, we develop a simple transformer architecture for 2.5D trajectory-based recognition together with a masked-trajectory pretraining, which we show to substantially improve downstream action recognition accuracy. Despite using only a fraction of the dense RGB input, our method reaches 45% top-1 on Something-Something V2 and 54% on EPIC-Kitchens-100, and surpasses V-JEPA on time-reversal sensitivity. More importantly, we find trajectory features to be complementary to state-of-the-art appearance-based features. Fusing our pretrained model with DINOv2 and V-JEPA 2 improves top-1 accuracy on Something-Something V2 by 8.7 and 1.6 points, respectively. Code: https://github.com/ecker-lab/TrAction

Investigating spatial-temporal correlations, specifically how spatial points vary over time, is crucial for understanding point cloud videos. Traditional methods, particularly flow-based techniques, struggle with these correlations due to the unordered spatial arrangement of sequential point cloud data. To address this challenge, we propose a novel approach that regularizes spatial-temporal correlation learning by formulating the problem as a solvable Partial Differential Equation (PDE). While PDEs have long been effective in the physical domain, their application to novel sequential data like point cloud video remains underexplored. Inspired by fluid analysis, we construct a simplified PDE, and the process of solving PDE is guided and refined by a contrastive learning structure between the temporal embeddings and the spatial embeddings. With this extra supervision, our method, named MotionPDE, serves as an effective, plug-and-play enhancement module for existing backbone models, adding minimal computational overhead and parameters. Capitalizing on the contrastive learning process, we delve deeper into the self-supervised capabilities of MotionPDE, yielding promising results that underscore its utility and adaptability in point cloud video data interpretation. The code repo with trained checkpoints will be available at https://github.com/zhh6425/motionpde.git for facilitating future research.

Video Moment Retrieval (MR) and Highlight Detection (HD) are crucial tasks in video analysis that aim to localize specific moments and estimate clip-wise relevance based on a given text query. Recent approaches treat them as similar video grounding tasks and use the same architecture to solve them. These tasks require both fine-grained comprehension at the image level and high-level temporal understanding across the entire video. Existing approaches have primarily focused on temporal modeling using frame-level features, often neglecting the rich visual information related to the text query within individual frames. This oversight leads to inaccurate grounding results. To address this limitation, we propose a Comprehensive Spatial-Temporal Representation Learning Framework (CoSTL), which captures both fine-grained image-level information and temporal dynamics. Specifically, CoSTL incorporates a text-driven progressive fine-grained image encoder, performing a two-step text-driven knowledge extraction process to learn fine-grained spatial representations. Furthermore, a multi-scale temporal perception module captures comprehensive spatial-temporal representations, enhancing the model's ability to process temporal dynamics. We demonstrate state-of-the-art performance on four public benchmarks: QVHighlights, Charades-STA, TACoS, and TVSum.

With the rapid development of pre-training technologies, adapting large-scale Vision-Language Models (VLMs) for video understanding \emph{\ie} image-to-video transfer learning has become a dominant paradigm. To achieve superior performance, it raises as an effective strategy among recent advances to employ Mixture-of-Experts (MoE) to enhance VLMs' temporal modeling capabilities. However, conventional MoE designs suffer from expert homogenization, where all experts act as identical generalists, inefficiently learning spatio-temporal features from undifferentiated video streams. To overcome this problem, we propose VidPrism, a novel heterogeneous temporal Mixture-of-Experts framework. VidPrism pioneers a division of labor by deploying functionally specialized experts, each assuming a role ranging from spatial understanding to temporal modeling. To feed these specialists appropriately, we introduce a content-aware, multi-rate sampling module that dynamically generates streams ranging from semantically rich to motion-focused representations, providing specialized inputs for experts. Furthermore, a dynamic, bidirectional fusion mechanism enables synergistic information exchange between these pathways, leading to a comprehensive video representation. Extensive experiments on various video recognition benchmarks demonstrate that VidPrism achieves state-of-the-art performance and effectively fosters expert specialization. Our source code is available at https://github.com/Lrrrr549/VidPrism.git.

Spatial intelligence requires visual representations that capture both semantic objects and geometric structure in the physical world. To support this, two major pre-training schemes are now widely used as foundation backbones: Vision-Language Models (VLMs), which use language supervision to align visual observations with semantic concepts, and Video Generation Models (VGMs), which learn from temporally evolving visual worlds. However, it still remains unclear which pre-training scheme provides a better representation substrate for spatial intelligence. In this paper, we present the first systematic frozen-feature probing study of VLMs and VGMs across three representative axes of spatial intelligence: semantic tagging, instance grouping, and 3D geometry prediction. Using the lightweight probe, our framework enables a controlled comparison of what information is already encoded in frozen representations from two model families. Experimental results reveal a clear complementarity: VLMs are stronger at semantic tagging and instance grouping, while VGMs provide more accessible signals for dense geometry and camera motion. Moreover, a naive fusion of the two already yields a representation that excels at both geometry and semantics, suggesting a promising direction for building stronger spatial-intelligence backbones by effectively integrating features from both model families. Our code is available at https://github.com/om-ai-lab/Probing-VLM-VGM.