In this work, we introduce CAT-V (Caption Anything in Video), a training-free framework for fine-grained object-centric video captioning of user-selected instances. CAT-V combines (i) a SAMURAI-based Segmenter for precise object masks across frames, (ii) a TRACE-Uni Temporal Analyzer for event boundary detection and coarse event descriptions, and (iii) an InternVL-2.5 Captioner that, conditioned on spatiotemporal visual prompts and chain-of-thought (CoT) guidance, produces detailed, temporally coherent captions about object attributes, actions, states, interactions, and context. The system supports point, box, and region prompts and maintains temporal sensitivity by tracking object states across segments. In contrast to vanilla video captioning that is overly abstract and dense video captioning that is often terse, CAT-V enables object-level specificity with spatial accuracy and temporal coherence, without additional training data.

With the rapid growth of online video platforms and the escalating volume of video content, the need for proficient video understanding tools has increased significantly. Given the remarkable capabilities of large language models (LLMs) in language and multimodal tasks, this survey provides a detailed overview of recent advances in video understanding that harness the power of LLMs (Vid-LLMs). The emergent capabilities of Vid-LLMs are surprisingly advanced, particularly their ability for open-ended multi-granularity (abstract, temporal, and spatiotemporal) reasoning combined with common-sense knowledge, suggesting a promising path for future video understanding. We examine the unique characteristics and capabilities of Vid-LLMs, categorizing the approaches into three main types:Video Analyzer × LLM, Video Embedder × LLM, and (Analyzer + Embedder) × LLM. We identify five subtypes based on the functions of LLMs in Vid-LLMs:LLMas Summarizer,LLMas Manager,LLMas Text Decoder,LLMas Regressor, andLLMas Hidden Layer. This survey also presents a comprehensive study of the tasks, datasets, benchmarks, and evaluation methods for Vid-LLMs. Additionally, it explores the extensive applications of Vid-LLMs in various domains, highlighting their remarkable scalability and versatility in real-world video understanding challenges. Additionally, it summarizes the limitations of existing Vid-LLMs and outlines directions for future research. For more information, readers are encouraged to visit the repository at https://github.com/yunlong10/Awesome-LLMs-for-Video-Understanding.

Large language models (LLMs) have demonstrated remarkable capabilities in natural language and multimodal domains. By fine-tuning multimodal LLMs with temporal annotations from well-annotated datasets, e.g., dense video captioning datasets, their temporal understanding capacity in video-language tasks can be obtained. However, there is a notable lack of untrimmed audio-visual video datasets with precise temporal annotations for events. This deficiency hinders LLMs from learning the alignment between time, audio-visual events, and text tokens, thus impairing their ability to localize audio-visual events in videos temporally. To address this gap, we introduce PU-VALOR, a comprehensive audio-visual dataset comprising over 114,081 pseudo-untrimmed videos with detailed temporal annotations. PU-VALOR is derived from the large-scale but coarse-annotated audio-visual dataset VALOR, through a subtle method involving event-based video clustering, random temporal scaling, and permutation. By fine-tuning a multimodal LLM on PU-VALOR, we developed AVicuna, a model capable of aligning audio-visual events with temporal intervals and corresponding text tokens. AVicuna excels in temporal localization and time-aware dialogue capabilities. Our experiments demonstrate that AVicuna effectively handles temporal understanding in audio-visual videos and achieves state-of-the-art performance on open-ended video QA, audio-visual QA, and audio-visual event dense localization tasks.

Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated remarkable progress in visual understanding. This impressive leap raises a compelling question: how can language models, initially trained solely on linguistic data, effectively interpret and process visual content? This paper aims to address this question with systematic investigation across 4 model families and 4 model scales, uncovering a unique class of attention heads that focus specifically on visual content. Our analysis reveals a strong correlation between the behavior of these attention heads, the distribution of attention weights, and their concentration on visual tokens within the input. These findings enhance our understanding of how LLMs adapt to multimodal tasks, demonstrating their potential to bridge the gap between textual and visual understanding. This work paves the way for the development of AI systems capable of engaging with diverse modalities.

The advancement of Multimodal Large Language Models (MLLMs) has enabled significant progress in multi-modal understanding, expanding their capacity to analyze video content. However, existing evaluation benchmarks for MLLMs primarily focus on abstract video comprehension, lacking a detailed assessment of their ability to understand video compositions, the nuanced interpretation of how visual elements combine and interact within highly compiled video contexts. We introduce VidComposition, a new benchmark specifically designed to evaluate the video composition understanding capabilities of MLLMs using carefully curated compiled videos and cinematic-level annotations. VidComposition includes 982 videos with 1706 multiple-choice questions, covering various compositional aspects such as camera movement, angle, shot size, narrative structure, character actions and emotions, etc. Our comprehensive evaluation of 33 open-source and proprietary MLLMs reveals a significant performance gap between human and model capabilities. This highlights the limitations of current MLLMs in understanding complex, compiled video compositions and offers insights into areas for further improvement. Our benchmark is publicly available at https://yunlong10.github.io/VidComposition/.

Audio source separation is fundamental for machines to understand complex acoustic environments and underpins numerous audio applications. Current supervised deep learning approaches, while powerful, are limited by the need for extensive, task-specific labeled data and struggle to generalize to the immense variability and open-set nature of real-world acoustic scenes. Inspired by the success of generative foundation models, we investigate whether pre-trained text-guided audio diffusion models can overcome these limitations. We make a surprising discovery: zero-shot source separation can be achieved purely through a pre-trained text-guided audio diffusion model under the right configuration. Our method, named ZeroSep, works by inverting the mixed audio into the diffusion model's latent space and then using text conditioning to guide the denoising process to recover individual sources. Without any task-specific training or fine-tuning, ZeroSep repurposes the generative diffusion model for a discriminative separation task and inherently supports open-set scenarios through its rich textual priors. ZeroSep is compatible with a variety of pre-trained text-guided audio diffusion backbones and delivers strong separation performance on multiple separation benchmarks, surpassing even supervised methods.

Understanding perspective is fundamental to human visual perception, yet the extent to which multimodal large language models (MLLMs) internalize perspective geometry remains unclear. We introduce MMPerspective, the first benchmark specifically designed to systematically evaluate MLLMs' understanding of perspective through 10 carefully crafted tasks across three complementary dimensions: Perspective Perception, Reasoning, and Robustness. Our benchmark comprises 2,711 real-world and synthetic image instances with 5,083 question-answer pairs that probe key capabilities, such as vanishing point perception and counting, perspective type reasoning, line relationship understanding in 3D space, invariance to perspective-preserving transformations, etc. Through a comprehensive evaluation of 43 state-of-the-art MLLMs, we uncover significant limitations: while models demonstrate competence on surface-level perceptual tasks, they struggle with compositional reasoning and maintaining spatial consistency under perturbations. Our analysis further reveals intriguing patterns between model architecture, scale, and perspective capabilities, highlighting both robustness bottlenecks and the benefits of chain-of-thought prompting. MMPerspective establishes a valuable testbed for diagnosing and advancing spatial understanding in vision-language systems. Resources are available at https://yunlong10.github.io/MMPerspective/

The rapid evolution of egocentric video analysis brings new insights into understanding human activities and intentions from a first-person perspective. Despite this progress, the fragmentation in tasks like action recognition, procedure learning, and moment retrieval, \etc, coupled with inconsistent annotations and isolated model development, hinders a holistic interpretation of video content. In response, we introduce the EAGLE (Egocentric AGgregated Language-video Engine) model and the EAGLE-400K dataset to provide a unified framework that integrates various egocentric video understanding tasks. EAGLE-400K, the \textit{first} large-scale instruction-tuning dataset tailored for egocentric video, features 400K diverse samples to enhance a broad spectrum of tasks from activity recognition to procedure knowledge learning. Moreover, EAGLE, a strong video multimodal large language model (MLLM), is designed to effectively capture both spatial and temporal information. In addition, we propose a set of evaluation metrics designed to facilitate a thorough assessment of MLLM for egocentric video understanding. Our extensive experiments demonstrate EAGLE's superior performance over existing models, highlighting its ability to balance task-specific understanding with holistic video interpretation. With EAGLE, we aim to pave the way for research opportunities and practical applications in real-world scenarios.

Learning new skills by observing humans’ behaviors is an essential capability of AI. In this work, we leverage instructional videos to study humans’ decision-making processes, focusing on learning a model to plan goal-directed actions in real-life videos. In contrast to conventional action recognition, goal-directed actions are based on expectations of their outcomes requiring causal knowledge of potential consequences of actions. Thus, integrating the environment structure with goals is critical for solving this task. Previous works learn a single world model will fail to distinguish various tasks, resulting in an ambiguous la-tent space; planning through it will gradually neglect the desired outcomes since the global information of the future goal degrades quickly as the procedure evolves. We address these limitations with a new formulation of procedure planning and propose novel algorithms to model human behaviors through Bayesian Inference and model-based Imitation Learning. Experiments conducted on real-world instructional videos show that our method can achieve state-of-the-art performance in reaching the indicated goals. Furthermore, the learned contextual information presents interesting features for planning in a latent space.

Learning from Demonstrations (LfD) via Behavior Cloning (BC) works well on multiple complex tasks. However, a limitation of the typical LfD approach is that it requires expert demonstrations for all scenarios, including those in which the algorithm is already well-trained. The recently proposed Learning from Interventions (LfI) overcomes this limitation by using an expert overseer. The expert overseer only intervenes when it suspects that an unsafe action is about to be taken. Although LfI significantly improves over LfD, the state-of-the-art LfI fails to account for delay caused by the expert's reaction time and only learns short-term behavior. We address these limitations by 1) interpolating the expert's interventions back in time, and 2) by splitting the policy into two hierarchical levels, one that generates sub-goals for the future and another that generates actions to reach those desired sub-goals. This sub-goal prediction forces the algorithm to learn long-term behavior while also being robust to the expert's reaction time. Our experiments show that LfI using sub-goals in a hierarchical policy framework trains faster and achieves better asymptotic performance than typical LfD.