Video Understanding

Existing embodied control research demonstrates remarkable performance improvements by scaling training data and model size. We instead explore inference-time strategy as an alternative axis. Non-deterministic generative models, such as diffusion and autoregressive models, have been widely adopted in the field of embodied control. However, the single-shot inference paradigm limits their performance. In this paper, we propose **TapSampling**, a plug-and-play framework for inference-time sampling. First, we introduce an Action-VAE that represents actions in a low-dimensional latent space by mapping policy-generated initial actions into a compressed posterior distribution, from which any number of latent samples can be drawn and decoded into candidate actions that approximate the true action distribution. Second, we formulate action verification as task-progress outcome prediction, using the intrinsic sequential structure of robotic datasets to train a semantically grounded verifier for interpretable action selection. Furthermore, TapSampling is a policy-agnostic framework. Extensive experiments in both simulated and real-world environments demonstrate that our method substantially improves multiple generalist policies without further policy finetuning. Code and models are available at the **project page** (https://aipixel.github.io/TapSampling/).

Low-precision arithmetic is pervasive in neural network training and deployment, yet its effect on prediction \textit{confidence}, not just accuracy, remains unexamined. We show that the softmax function amplifies logit-space quantization errors in an input-dependent manner: confidence distortion scales with the product of precision-dependent error bound $\epsilon$ and logit norm, peaking when the model is confident but not saturated. This explains why identical models report different confidence values across precisions, a phenomenon we term \textit{Precision Split}. During training, the same mechanism causes gradient underflow: when logit margins exceed a precision-dependent threshold, gradients vanish and samples silently stop contributing to learning. Since logit norm serves as a computable proxy for precision-induced risk, we propose Precision-Aware Confidence Scaling (PACS), which applies sample-adaptive temperature inversely related to this risk, with sub-one-percent overhead and no full-precision computation required. On ImageNet with mixed-precision ResNet-50, PACS reduces Expected Calibration Error from 5.82\% to 1.92\% while maintaining accuracy, with consistent improvements across architectures, precision formats, and modalities.

Understanding modality interaction in multimodal large language models (MLLMs) is central to reliable deployment. We introduce Partial Information Decomposition (PID) as a decision-level framework that separates unique, redundant, and synergistic contributions of sensory and linguistic inputs, beyond representation alignment and outcome-based evaluation. Across vision--language benchmarks, PID reveals recurring modality-use profiles: reasoning and grounding-oriented tasks tend to exhibit high synergy, whereas expert and knowledge-oriented tasks show stronger language-unique reliance. These profiles generalize across model families and predict sensitivity to modality-level interventions. We further extend PID to tri-modal systems with Sensory PID, treating language as a control variable to decompose video--audio information gain. Applied to omni-modal models, Sensory PID reveals a sensory synergy bottleneck dominated by visual information even on audio--visual fusion tasks. Finally, PID-guided reweighting provides initial evidence for improving multimodal reasoning and grounding performance.

Current research on video hallucination mitigation primarily focuses on isolated error types, leaving *compositional* hallucinations—arising from incorrect reasoning over multiple interacting spatial and temporal factors largely underexplored. We introduce **OmniVCHall**, a benchmark designed to systematically evaluate both isolated and compositional hallucinations in Video Large Language Models (VLLMs). OmniVCHall spans diverse video domains, introduces a novel camera-based hallucination type, and defines a fine-grained taxonomy, together with adversarial answer options (*e.g.*, “All are correct” and “None of the above”) to prevent shortcut reasoning. The evaluations of 39 representative VLLMs reveal that even advanced models (*e.g.*, Qwen3-VL and GPT-5) exhibit substantial performance degradation. We propose **TriCD**, a contrastive decoding framework with a triple-pathway calibration mechanism. An adaptive perturbation controller dynamically selects distracting operations to construct negative video variants, while a saliency-guided enhancement module adaptively reinforces grounded token-wise visual evidences. These components are optimized via reinforcement learning to encourage precise decision-making under compositional hallucination settings. Experimental results show that **TriCD** consistently improves performance across two representative backbones, achieving an average accuracy improvement of over 10\%. The data and code are released at https://github.com/BMRETURN/OmniVCHall.

Video anomaly understanding (VAU) relies on sparse, context-dependent cues. However, existing passive paradigms suffer from observational aliasing, where static sampling fails to disambiguate semantically distinct events. To overcome this, we propose $Anom\text{-}\pi$, a closed-loop framework that reconceptualizes video understanding as an active sequential decision-making process within a dynamic environment. Inspired by human video-reviewing behavior, this framework unifies internal cognitive reasoning and strategic evidence acquisition into an interleaved policy, utilizing temporal atomic operators such as local backtracking, temporal expansion, and fine-grained sampling to endow the model with perceptual proactivity. To learn such complex interaction strategies under video-level weak supervision, we design Interactive Direct Preference Optimization (iDPO) to achieve trajectory-level policy alignment, guided by an Active Evidence Inquiry (AEI) utility that balances task success, informative evidence acquisition, and interaction cost. This approach enables the agent to learn to actively disambiguate hypotheses while suppressing redundant exploration. Extensive experiments demonstrate that our framework, with only 2B parameters, achieves highly competitive performance, significantly outperforming state-of-the-art large-scale VAU models in complex scenarios.

Metaphorical videos are prevalent across various real-world scenarios to convey complex ideas, and understanding them typically requires high-order cognitive capabilities. The lack of systematic studies on metaphorical video understanding not only constrains the real-world applicability of MLLMs but also impedes the thorough assessment of their high-order cognitive capabilities. To bridge this gap, we propose MetaphorVU-Bench, the first systematic and comprehensive benchmark dedicated to metaphorical video understanding. Through experiments, we find current MLLMs struggle with accurate metaphorical video understanding, lagging far behind human level, primarily due to defective cross-domain mapping. Motivated by this finding, we construct a metaphor knowledge graph as mapping augmentation and propose MetaphorBoost, an inference-time enhancement framework achieving consistent performance improvement. Our benchmark, analysis, and method provide useful insights and a foundation for future research on advancing MLLMs. Code: https://github.com/icip-cas/MetaphorVU.

Real-time satellite video tracking poses distinct challenges, including accommodating high spatial-temporal resolution, dynamic backgrounds, and constrained onboard computational resources. While Discriminative Correlation Filter (DCF)-based methods offer high-speed inference, they suffer from limited accuracy. In contrast, Vision Transformer (ViT)-based trackers achieve strong performance by unifying representation and aggregation in a single-stream design, yet their heavy computational footprint limits practical deployment in real-time satellite scenarios. In this work, we present FATrack, a novel tracking framework that effectively balances tracking accuracy and computational efficiency. At its core is FA-ViT, a lightweight Vision Transformer backbone that introduces foreground-aware token routing, enabling the model to concentrate computation on target-relevant regions while suppressing redundancy. To mitigate semantic degradation caused by token sparsification, we propose the Adaptive Scatter Module (ASM), which selectively reinforces informative tokens via joint spatial-channel attention and sparse structural propagation, thereby enhancing both semantic fidelity and spatial coherence. By synergistically integrating FA-ViT and ASM, FATrack forms a unified architecture that delivers real-time performance with significantly improved tracking precision. Extensive evaluations on multiple satellite video benchmarks demonstrate that FATrack surpasses existing real-time trackers in accuracy and achieves inference efficiency comparable to DCF-based methods, highlighting its potential for practical deployment in large-scale aerial video tracking systems.

Bridging the gap between visual realism and physical understanding is a core challenge for video-based world models. We study the structural identifiability of continuous-time physical laws from raw pixels, focusing on whether an encoder-only pipeline can uniquely recover the parameters of second-order linear ODEs. We prove that a level-set slope-coverage condition ensures the learned latent space is locally affine to the true physical state, enabling exact parameter recovery. Our theory provides the first characterization of minimal data requirements across damping regimes, establishing that underdamped systems are identifiable from a single video clip, whereas other regimes require three diverse trajectories. We further introduce a variance-floor regularizer to stabilize the decoder-free objective and prevent latent collapse. Validated on synthetic and real-world data, our approach demonstrates that interpretable physical constants can be reliably estimated from video without the need for compute-intensive pixel reconstruction, ensuring both physical correctness and transparency. Code is available at https://github.com/wenjiewang3/PhysicsFromVideo.

Achieving a complete and explorable 360-degree visual world is a cornerstone of immersive content creation. While recent advances in video generation have achieved impressive results, they follow a 2D paradigm that treats content generation as transitions of 2D pixels, lacking an intrinsic understanding of the physical 3D world, resulting in frequent geometric inconsistencies. To achieve an explorable and physical-consistent visual world, the generation process should shift to a 3D paradigm: the visual content is governed by the physical relationships of the entire 3D environment together with 3D motion signals. However, under this setting, the conventional modeling methods and control signals, such as spatial attention computation in a 2D space, become unsuitable and ineffective. To address this, we propose PanoWorld-X for explorable immersive scene video generation. Our framework is built on the panoramic representation, which naturally maps a 3D scene into a standard format and provides an ideal basis for consistency. Specifically, we first develop a data curation pipeline to produce high-quality and large-motion 3D scene evolution with movement trajectories. To achieve precise control, we design the Exploration Panoramic Plücker Embedding (PPE), a guidance signal tailored for 3D motion. Furthermore, leveraging the spherical geometric properties of panoramic data, we propose a sphere-aware attention mechanism, which can capture true geometric adjacency by reprojecting features onto a spherical surface. Extensive experiments demonstrate that PanoWorld-X achieves superior performance in motion range, control precision, and visual quality, underscoring its potential for real-world applications.

Estimating full-hand grasp pressure from egocentric video is critical for immersive VR and robotic manipulation, yet dense tactile sensing often relies on intrusive hardware. Existing vision-based methods predominantly rely on planar surfaces or fingertip contacts, failing to generalize to complex 3D object interactions. Therefore, we introduce EgoTactile, a benchmark pairing egocentric video with full-hand pressure supervision for diverse everyday objects, incorporating a bare-hand transfer subset to enable generalization to natural scenarios. Leveraging this benchmark, we first establish EgoPressureFormer as a discriminative baseline. Beyond this, to explicitly address the uncertainty in partial observations, we propose EgoPressureDiff, a conditional diffusion framework that adapts a large-scale pre-trained video diffusion backbone. By combining rich world knowledge priors with a Physically-Informed Feature Rectification layer to inject semantic constraints, our approach effectively hallucinates plausible contact patterns and resolves visual-physical ambiguities. Extensive experiments demonstrate that our method achieves superior performance on the benchmark and robust transferability to in-the-wild scenarios. Our project page is at https://egotactile.github.io/.

Although Multimodal Large Language Models (MLLMs) have made remarkable progress, they still struggle with long-video understanding due to the massive memory footprint of KV Caches. Existing methods often resort to disjoint retrieval or attention-based reduction with a uniform layer-wise budget to achieve compression. However, these methods disrupt temporal continuity and ignore the varying information density across network layers. In this work, we reveal that memory allocation should mirror layer-wise semantic density, rather than adhering to a uniform budget. To this end, we introduce EAKV, a training-free entropy-driven adaptive KV compression framework that leverages attention entropy to adaptively allocate compression budgets, selectively preserving critical tokens while distilling redundant contexts into compact contextual anchors, thereby achieving granular memory allocation proportional to semantic density. Extensive experiments on various benchmarks demonstrate that EAKV surpasses existing methods across diverse model architectures and varying parameter scales, yielding improvements ranging from 0.6% to 6.5%.

Positional encodings (PEs) enhance the power of graph neural networks (GNNs), both theoretically and empirically. Two of the most popular families of PEs---spectral (e.g., Laplacian eigenspaces, effective resistance) and random walk (polynomials of the adjacency matrix)---are theoretically equivalent in expressive power, and both are known to lie between the 1-WL and 3-WL tests in terms of expressivity. However, this equivalence assumes the GNN uses the "complete'' version of these PEs, which requires $O(n^3)$ time and space complexity. Practitioners therefore commonly use truncated variants of these encodings (e.g., the first $k$ eigenspaces or powers of adjacency matrix). However, the theoretical properties of these truncated PEs are unknown. In this work, we initiate the study of these truncated PEs. Theoretically, we show that, under truncation, several families of PEs are fundamentally different in expressive power. As a corollary, we show that truncated spectral PEs are no longer stronger than the 1-WL test. We also study a family of spectral PEs, the $k$-harmonic distances, to highlight the differences in expressive power of even closely related truncated PEs. Finally, we experimentally show that a mix of truncated PEs is preferable to any single family on real-world datasets.

Token-level adaptive computation seeks to reduce inference cost by allocating more computation to harder tokens and less to easier ones. However, prior work is primarily evaluated on natural-language benchmarks using task-level metrics, where token-level difficulty is unobservable and confounded with architectural factors, making it unclear whether compute allocation truly aligns with underlying complexity. We address this gap through three contributions. First, we introduce a complexity-controlled evaluation paradigm using existing algorithmic and synthetic language tasks with parameterized difficulty, enabling direct testing of token-level compute allocation. Second, we propose ANIRA, a unified recurrent Transformer framework that supports per-token variable-depth computation while isolating compute allocation decisions from other model factors. Third, we use this framework to conduct a systematic analysis of token-level adaptive computation across alignment with complexity, generalization, and decision timing. Our results show that compute allocation aligned with task complexity can emerge without explicit difficulty supervision, but such alignment does not imply algorithmic generalization: models fail to extrapolate to unseen input sizes despite allocating additional computation. We further find that early compute decisions rely on static structural cues, whereas online halting more closely tracks algorithmic execution state. Code: https://github.com/merlresearch/ANIRA

Diffusion Transformers (DiTs) are increasingly adopted for long-video generation, yet inference is dominated by the quadratic cost of 3D full attention. Sparse attention mitigates this bottleneck by exploiting *token-pair redundancy* and pruning query-key interactions. Nevertheless, its effectiveness on video generation is often constrained by non-sparse attention heads, making it hard to strike a good balance between inference speed and generation quality. To address this, we identify another pervasive but overlooked redundancy specific to video DiTs: *frame-block similarity*, where frame-blocks in attention weights exhibit highly similar distributions and can be well approximated by lightweight linear calibration. Motivated by this observation, we propose **EchoAttention**, which jointly leverages *token-pair redundancy* (*Sparse* operator) and *frame-block similarity* (*Echo* operator), together with a fine-grained routing policy learned via three-stage distillation. This design enables efficient handling of both sparse and non-sparse heads, overcoming the inherent ceiling of purely sparse attention and yielding a better speed-quality trade-off. Across public video DiTs, EchoAttention consistently improves the speed-quality frontier over SOTA sparse-attention baselines, reducing end-to-end latency up to 2.42$\times$ with minimal quality loss.

Low-light video enhancement (LLVE) remains a challenging task due to severe information degradation under low-illumination conditions. Recent multimodal approaches have significantly improved enhancement performance by incorporating auxiliary modalities, such as event streams and infrared images. However, these methods typically assume the availability of these modalities at inference, which is often not feasible in real-world scenarios. To solve this problem, in this work, we propose AMNet, a unified multimodal framework for LLVE, to support flexible modality-agnostic inference, where auxiliary modalities may be unavailable. To address the issue of modality absence, we introduce a Spatial-Spectral Dual-Gated Translator that learns the correspondence between auxiliary modalities and RGB inputs, producing implicit auxiliary representations to support the robust enhancement. Additionally, to fully facilitate the learning of cross-modal correspondence, we conduct large-scale multimodal pretraining based on the RGB-only dataset with synthetic auxiliary modalities. Extensive experiments demonstrate that AMNet could handle arbitrary inference-time modality combinations and exhibits superior performance for LLVE under modality absence conditions. Code and models are available on the project page.

Large language models (LLMs) are increasingly adapted to downstream tasks in resource-constrained scenarios, making quantization-aware fine-tuning (QAF) a common practice for practical deployment. However, we find that quantized LLMs are substantially more vulnerable to safety alignment degradation during fine-tuning than full-precision models by interpretability analyses. In this paper, we first theoretically reveal that this vulnerability is driven by quantization errors, manifesting as an initial safety shift followed by a distorted optimization path. Based on this insight, we propose Explicit-Safety Quantization-Aware Fine-tuning (ExSQF), which effectively restores model safety while preserving downstream performance. It initializes adapters by combining quantization error with a safety matrix projection to mitigate early safety shifts, followed by post-training refinement that corrects deviations in the optimization path. Extensive experimental results show that ExSQF achieves state-of-the-art safety alignment recovery, even surpassing existing full-precision safety-aware fine-tuning baseline, while effectively preserving model performance.

Adam is a widely used optimization algorithm in deep learning, yet the specific class of objective functions where it exhibits inherent advantages remains underexplored. Unlike prior studies requiring external schedulers and $\beta_2$ near 1 for convergence, this work investigates the ``natural'' auto-convergence properties of Adam. We identify a class of highly degenerate polynomials where Adam converges automatically without additional schedulers. Specifically, we derive theoretical conditions for local asymptotic stability on degenerate polynomials and demonstrate strong alignment between theoretical bounds and experimental results. We prove that Adam achieves local linear convergence on these degenerate functions, significantly outperforming the sub-linear convergence of Gradient Descent and Momentum. This acceleration stems from a decoupling mechanism between the second moment $v_t$ and squared gradient $g_t^2$, which exponentially amplifies the effective learning rate. Finally, we characterize Adam's hyperparameter phase diagram, identifying three distinct behavioral regimes: stable convergence, spikes, and SignGD-like oscillation.

Latent diffusion models have become the dominant paradigm for video generation, making the video tokenizer a critical role. While most existing tokenizers are trained primarily for reconstruction, diffusion models are optimized to denoise heavily corrupted latents, which creates a mismatch between tokenizer training objectives and downstream generative learning. As a result, reconstruction metrics (e.g., rFVD) can be a poor proxy for generation quality (gFVD), and overly prioritizing reconstruction may even hinder diffusion training. We propose VideoMAETok, a simple family of ViT-based video tokenizers trained explicitly as corruption-inversion models for latent video diffusion. VideoMAETok builds on masked autoencoders: we (i) apply high-ratio token masking and encode only visible spatiotemporal tokens for efficiency, and (ii) corrupt latent tokens with interpolative Gaussian noise to better match the denoising nature of diffusion generators. Training under such corruption encourages latents that remain informative and well-conditioned for downstream denoising. Extensive experiments show that VideoMAETok consistently improves generation quality when paired with off-the-shelf diffusion models (SiT and LightningDiT), achieving state-of-the-art gFVD on Kinetics-600 and UCF-101 while remaining compute-efficient. Code is available at https://github.com/yztongzhan/VideoMAETok.

Self-supervised learning (SSL) excels at finding general-purpose latent representations from complex data, yet lacks a unifying theoretical framework that explains the diverse existing methods and guides the design of new ones. We cast SSL as latent distribution matching (LDM): learning representations that maximize their log-probability under an assumed latent model (alignment), while maximizing latent entropy to prevent collapse (uniformity). This view unifies independent component analysis with contrastive, non-contrastive, and predictive SSL methods, including stop gradient approaches. Leveraging LDM, we derive a nonlinear, sampling-free Bayesian filtering model with a Kalman-based predictor for high-dimensional timeseries. We further prove that predictive LDM yields identifiable latent representations under mild assumptions, even with nonlinear predictors. Overall, LDM clarifies the assumptions behind established SSL methods and provides principled guidance for developing new approaches.

Recent advances in video generation have enabled highly realistic synthetic content, raising concerns about the integrity of digital media and motivating the development of benchmarks and detection methods for generated videos. Prior works have largely prioritized bolstering model generalization against unseen generators. However, we uncover a neglected factor: the quality distribution of real videos plays a pivotal role. Current training protocols suffer from a clear quality bias between real and fake data, prone to shortcut learning. Compounded by testing on similar real data distributions, this creates an illusion of generalization. In reality, these models fail to generalize when exposed to real data with significantly different quality profiles. To address this, we propose training with quality-matched real and fake data to mitigate bias. Building on this, we introduce a data expansion strategy that broadens the training set to comprehensively cover the full quality spectrum. This approach enables the model to learn quality-agnostic features for detection, thereby achieving generalization across real data of varying qualities and enhancing real-world applicability. Extensive experiments demonstrate that our method scales well across diverse backbones, consistently enhancing the generalization capability of existing models.