Recently, rapid advancements have been made in multimodal large language models (MLLMs), especially in video understanding tasks. However, current research focuses on simple video scenarios, failing to reflect the complex and diverse nature of real-world audio-visual events in videos. To bridge this gap, we firstly introduce R-AVST, a dataset for audio-visual reasoning featuring fine-grained spatio-temporal annotations. In constructing this, we design a pipeline consisting of LLM-based key object extraction, automatic spatial annotation and manual quality inspection, resulting in over 5K untrimmed videos with 27K objects across 100 types of audio-visual events. Building on this dataset, we define three core tasks for spatio-temporal reasoning in audio-visual scenes and generate more than 8K high-quality, evenly distributed question-answer pairs to effectively benchmark model performance. To further enhance reasoning, we propose AVST-Zero, a reinforcement learning-based model that avoids intermediate supervision, directly optimizing behavior via carefully designed multi-dimensional rewards. Extensive experiments validate the effectiveness of our R-AVST in advancing audio-visual spatio-temporal reasoning, upon which AVST-Zero demonstrates competitive performance compared to existing models. To the best of our knowledge, R-AVST is the first dataset designed for real-world audio-visual spatio-temporal reasoning, and AVST-Zero offers a novel perspective for tackling future challenges in this domain.

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.

Fine-grained video captioning aims to generate detailed, temporally coherent descriptions of video content. However, existing methods struggle to capture subtle video dynamics and rich detailed information. In this paper, we leverage preference learning to enhance the performance of vision-language models (VLM) in fine-grained video captioning, while mitigating several limitations inherent to Direct Preference Optimization (DPO). First, we propose a pipeline for constructing preference pairs that leverages the intrinsic properties of VLMs along with partial assistance from large language models, achieving an balance between cost and data quality. Then, we propose Synergistic Preference Optimization (SynPO), a novel optimization method offering significant advantages over DPO and its variants. SynPO prevents negative pReferences from dominating the training, explicitly preserves the model's language capability to avoid deviation of the optimization objective, thus obtains high-quality captions and improves training efficiency by eliminating the need for the reference model. We extensively evaluate our proposed data construction pipeline across three models: AuroraCap, LLaVA1.6-7B-Video and InterVL2-8B. Results demonstrate that our method improve performance in fine-grained video captioning significantly and consistenly. Source code is available at https://github.com/longmalongma/SynPO.

Despite impressive advancements in video understanding, most efforts remain limited to coarse-grained or visual-only video tasks. However, real-world videos encompass omni-modal information (vision, audio, and speech) with a series of events forming a cohesive storyline. The lack of multi-modal video data with fine-grained event annotations and the high cost of manual labeling are major obstacles to comprehensive omni-modality video perception. To address this gap, we propose an automatic pipeline consisting of high-quality multi-modal video filtering, semantically coherent omni-modal event boundary detection, and cross-modal correlation-aware event captioning. In this way, we present LongVALE, the first- ever Vision-Audio-Language Event understanding benchmark comprising 105K omni-modal events with precise temporal boundaries and detailed relation-aware captions within 8.4K high-quality long videos. Further, we build a baseline that leverages LongVALE to enable video large language models (LLMs) for omni-modality fine-grained temporal video understanding for the first time. Extensive experiments demonstrate the effectiveness and great potential of LongVALE in advancing comprehensive multi-modal video understanding. The dataset and code are available at https://ttgeng233.github.io/LongVALE/.

Recent advances in large video-language models have revolutionized video understanding tasks.However, their efficiency is greatly constrained by processing high volumes of visual tokens.Existing token compression strategies apply a fixed compression ratio, ignoring varying semantic density across video clips.Consequently, this leads to inadequate representation of information-rich clips due to insufficient tokens and unnecessary computation on static or content-poor ones.To address this, we propose LangDC, a Language-aware Dynamic Token Compressor.LangDC leverages a lightweight language model to describe video clips, converting them into soft caption tokens as visual representations.Trained with our proposed semantic density-aware supervision, LangDC aims to 1) cover key visual cues necessary for downstream task reasoning and 2) dynamically adjust compression ratios based on scene richness, reflected by description length.Our design mimics how humans dynamically express what they see: complex scenes (seeing more) elicit more detailed language to convey nuances (saying more), whereas simpler scenes are described with fewer words.Experimental results show that our method reduces FLOPs by 49% compared to VideoGPT+ while maintaining competitive performance.Furthermore, qualitative results demonstrate our approach adaptively adjusts the token compression ratio based on video segment richness.Codes are available at https://github.com/NIneeeeeem/LangDC.

The synergy between generative and discriminative models receives growing attention. While discriminative Contrastive Language-Image Pre-Training (CLIP) excels in high-level semantics, it struggles with perceiving fine-grained visual details. Generally, to enhance representations, generative models take CLIP's visual features as conditions for reconstruction. However, the underlying principle remains underexplored. In this work, we empirically found that visually perfect generations are not always optimal for representation enhancement. The essence lies in effectively extracting fine-grained knowledge from generative models while mitigating irrelevant information. To explore critical factors, we delve into three aspects: (1) Conditioning mechanisms: We found that even a small number of local tokens can drastically reduce the difficulty of reconstruction, leading to collapsed training. We thus conclude that utilizing only global visual tokens as conditions is the most effective strategy. (2) Denoising configurations: We observed that end-to-end training introduces extraneous information. To address this, we propose a two-stage training strategy to prioritize learning useful visual knowledge. Additionally, we demonstrate that lightweight denoisers can yield remarkable improvements. (3) Generation paradigms: We explore both continuous and discrete denoisers with desirable outcomes, validating the versatility of our method. Through our in-depth explorations, we have finally arrived at an effective method, namely GenHancer, which consistently outperforms prior arts on the MMVP-VLM benchmark, e.g., 6.0% on OpenAICLIP. The enhanced CLIP can be further plugged into multimodal large language models for better vision-centric performance. All the models and codes are made publicly available.

Multimodal Large Language Models (MLLMs) exhibit promising advancements across various tasks, yet they still encounter significant trustworthiness issues. Prior studies apply Split Conformal Prediction (SCP) in language modeling to construct prediction sets with statistical guarantees. However, these methods typically rely on internal model logits or are restricted to multiple-choice settings, which hampers their generalizability and adaptability in dynamic, open-ended environments. In this paper, we introduce *TRON*, a **t**wo-step framework for **r**isk c**o**ntrol and assessme**n**t, applicable to any MLLM that supports sampling in both open-ended and closed-ended scenarios. *TRON* comprises two main components: (1) a novel conformal score to **sample** response sets of minimum size, and (2) a nonconformity score to **identify** high-quality responses based on self-consistency theory, controlling the error rates by two specific risk levels. Furthermore, we investigate semantic redundancy in prediction sets within open-ended contexts for the first time, leading to a promising evaluation metric for MLLMs based on average set size. Our comprehensive experiments across four Video Question-Answering (VideoQA) datasets utilizing eight MLLMs show that *TRON* achieves desired error rates bounded by two user-specified risk levels. Additionally, deduplicated prediction sets maintain adaptiveness while being more efficient and stable for risk assessment under different risk levels.

Using the impressive learning representation capacity of large language models (LLMs), LLM-based video understanding methods have made significant strides recently. However, most existing methods overlook the crucial importance discrepancy of frames, which often include massive low-quality frames, leading to limited performance and inferior inference efficiency, particularly for long-term videos. To this end, this paper proposes a new video understanding method called quality- guided dynamic memory network (QDM-Net). First, we design a memory quality evolution module (MQEM), which dynamically assigns weights to each frame according to contextual relationships between adjacent frames. Second, we devise a high- level quality memory bank updating mechanism (HQMBU), which selectively maintains high-quality frames in the memory bank, avoiding the negative influences of redundant frames and ensuring that the model focuses on the most informative visual cues. Extensive experiments on long-term video understanding benchmarks demonstrate that our QDM-Net consistently outperforms state-of-the-art methods, showcasing its potential in real-world applications. Our code and model will be publicly available.

The rapid development of multimodal large language models (MLLMs) has highlighted their potential in video understanding. However, challenges remain in long video tasks, particularly in integrating visual features with prompt texts. Existing methods naively store processed video frames in a long-term memory bank, but neglect simple yet effective cross-modal integration. To address this, we introduce the instruction-aware memory construction (IaMC) model for long-term video understanding. By integrating visual and textual information, our model can obtain cross-modal features with robust understanding capabilities. These features are stored in a text-visual memory bank, enabling efficient long-term aggregation without surpassing LLM context or GPU memory limits. Experiments on the LVU dataset demonstrate state-of-the-art performance in video understanding and question answering, showcasing the IaMC model’s effectiveness and setting a new benchmark for long-term video analysis. The source code and trained models will be released publicly.

Dynamic spatio-temporal understanding is essential for video-based multimodal tasks, yet existing methods often struggle to capture fine-grained temporal and spatial relationships in long videos. Current approaches primarily rely on pre-trained CLIP encoders, which excel in semantic understanding but lack spatially-aware visual context. This leads to hallucinated results when interpreting fine-grained objects or scenes. To address these limitations, we propose a novel framework that integrates diffusion models into multimodal video models. By employing diffusion encoders at intermediate layers, we enhance visual representations through feature alignment and knowledge distillation losses, significantly improving the model's ability to capture spatial patterns over time. Additionally, we introduce a multi-level alignment strategy to learn robust feature correspondence from pre-trained diffusion models. Extensive experiments on benchmark datasets demonstrate our approach's state-of-the-art performance across multiple video understanding tasks. These results establish diffusion models as a powerful tool for enhancing multimodal video models in complex, dynamic scenarios.

The rapid advancement of large language models (LLMs) has led to the widespread adoption of video-language models (VLMs) across various domains. However, VLMs are often hindered by their limited semantic discrimination capability, exacerbated by the limited diversity and biased sample distribution of most video-language datasets. This limitation results in a biased understanding of the semantics between visual concepts, leading to hallucinations. To address this challenge, we propose a Multi-level Multimodal Alignment (MMA) framework that leverages a text encoder and semantic discriminative loss to achieve multi-level alignment. This enables the model to capture both low-level and high-level semantic relationships, thereby reducing hallucinations. By incorporating language-level alignment into the training process, our approach ensures stronger semantic consistency between video and textual modalities. Furthermore, we introduce a two-stage progressive training strategy that exploits larger and more diverse datasets to enhance semantic alignment and better capture general semantic relationships between visual and textual modalities. Our comprehensive experiments demonstrate that the proposed MMA method significantly mitigates hallucinations and achieves state-of-the-art performance across multiple video-language tasks, establishing a new benchmark in the field.

Video event localization tasks include temporal action localization (TAL), sound event detection (SED) and audio-visual event localization (AVEL). Existing methods tend to over-specialize on individual tasks, neglecting the equal importance of these different events for a complete understanding of video content. In this work, we aim to develop a unified framework to solve TAL, SED and AVEL tasks together to facilitate holistic video understanding. However, it is challenging since different tasks emphasize distinct event characteristics and there are substantial disparities in existing task-specific datasets (size/domain/duration). It leads to unsatisfactory results when applying a naive multi-task strategy. To tackle the problem, we introduce UniAV, a Unified Audio-Visual perception network to effectively learn and share mutually beneficial knowledge across tasks and modalities. Concretely, we propose a unified audio-visual encoder to derive generic representations from multiple temporal scales for videos from all tasks. Meanwhile, task-specific experts are designed to capture the unique knowledge specific to each task. Besides, instead of using separate prediction heads, we develop a novel unified language-aware classifier by utilizing semantic-aligned task prompts, enabling our model to flexibly localize various instances across tasks with an impressive open-set ability to localize novel categories. Extensive experiments demonstrate that UniAV, with its unified architecture, significantly outperforms both single-task models and the naive multi-task baseline across all three tasks. It achieves superior or on-par performances compared to the state-of-the-art task-specific methods on ActivityNet 1.3, DESED and UnAV-100 benchmarks.

Emotion recognition aims to discern the emotional state of subjects within an image, relying on subject-centric and contextual visual cues. Current approaches typically follow a two-stage pipeline: first localize subjects by off-the-shelf detectors, then perform emotion classification through the late fusion of subject and context features. However, the complicated paradigm suffers from disjoint training stages and limited fine-grained interaction between subject-context elements. To address the challenge, we present a single-stage emotion recognition approach, employing a Decoupled Subject-Context Transformer (DSCT), for simultaneous subject localization and emotion classification. Rather than compartmentalizing training stages, we jointly leverage box and emotion signals as supervision to enrich subject-centric feature learning. Furthermore, we introduce DSCT to facilitate interactions between fine-grained subject-context cues in a ''decouple-then-fuse'' manner. The decoupled query tokens-subject queries and context queries-gradually intertwine across layers within DSCT, during which spatial and semantic relations are exploited and aggregated. We evaluate our single-stage framework on two widely used context-aware emotion recognition datasets, CAER-S and EMOTIC. Our approach surpasses two-stage alternatives with fewer parameter numbers, achieving a 3.39% accuracy improvement and a 6.46% average precision gain on CAER-S and EMOTIC datasets, respectively. Code and models are available at: https://github.com/Sampson-Lee/DSCT.

Existing audio-visual event localization (AVE) handles manually trimmed videos with only a single instance in each of them. However, this setting is unrealistic as natural videos often contain numerous audio-visual events with different categories. To better adapt to real-life applications, in this paper we focus on the task of dense-localizing audio-visual events, which aims to jointly localize and recognize all audio-visual events occurring in an untrimmed video. The problem is challenging as it requires fine-grained audio-visual scene and context understanding. To tackle this problem, we introduce the first Untrimmed Audio-Visual (UnAV-J 00) dataset, which contains 10K untrimmed videos with over 30K audio-visual events. Each video has 2.8 audio-visual events on average, and the events are usually related to each other and might co-occur as in real-life scenes. Next, we formulate the task using a new learning-based framework, which is capable of fully integrating audio and visual modalities to localize audio-visual events with various lengths and capture dependencies between them in a single pass. Extensive experiments demonstrate the effectiveness of our method as well as the significance of multi-scale cross-modal perception and dependency modeling for this task. The dataset and code are available at https://unav100.github.io.

The state of the arts in vision-language pretraining (VLP) achieves exemplary performance but suffers from high training costs resulting from slow convergence and long training time, especially on large-scale web datasets. An essential obstacle to training efficiency lies in the entangled prediction rate (percentage of tokens for reconstruction) and corruption rate (percentage of corrupted tokens) in masked language modeling (MLM), that is, a proper corruption rate is achieved at the cost of a large portion of output tokens being excluded from prediction loss. To accelerate the convergence of VLP, we propose a new pretraining task, namely, free language modeling (FLM), that enables a 100% prediction rate with arbitrary corruption rates. FLM successfully frees the prediction rate from the tie-up with the corruption rate while allowing the corruption spans to be customized for each token to be predicted. FLM-trained models are encouraged to learn better and faster given the same GPU time by exploiting bidirectional contexts more flexibly. Extensive experiments show FLM could achieve an impressive 2.5 × pretraining time reduction in comparison to the MLM-based methods, while keeping competitive performance on both vision-language understanding and generation tasks. Code will be public at https://github.com/TencentARC/FLM.

Image-to-text generation aims to describe images using natural language. Recently, zero-shot image captioning based on pre-trained vision-language models (VLMs) and large language models (LLMs) has made significant progress. However, we have observed and empirically demonstrated that these methods are susceptible to modality bias induced by LLMs and tend to generate descriptions containing objects (entities) that do not actually exist in the image but frequently appear during training (i.e., object hallucination). In this paper, we propose ViECap, a transferable decoding model that leverages entity-aware decoding to generate descriptions in both seen and unseen scenarios. ViECap incorporates entity-aware hard prompts to guide LLMs’ attention toward the visual entities present in the image, enabling coherent caption generation across diverse scenes. With entity-aware hard prompts, ViECap is capable of maintaining performance when transferring from in-domain to out-of-domain scenarios. Extensive experiments demonstrate that ViECap sets a new state-of-the-art cross-domain (transferable) captioning and performs competitively in-domain captioning compared to previous VLMs-based zero-shot methods. Our code is available at: https://github.com/FeiElysia/ViECap

Pre-trained vision-language models, e.g., CLIP, working with manually designed prompts have demonstrated great capacity of transfer learning. Recently, learnable prompts achieve state-of-the-art performance, which however are prone to overfit to seen classes, failing to generalize to unseen classes. In this paper, we propose a Knowledge-Aware Prompt Tuning (KAPT) framework for vision-language models. Our approach takes the inspiration from human intelligence in which external knowledge is usually incorporated into recognizing novel categories of objects. Specifically, we design two complementary types of knowledge-aware prompts for the text encoder to leverage the distinctive characteristics of category-related external knowledge. The discrete prompt extracts the key information from descriptions of an object category, and the learned continuous prompt captures overall contexts. We further design an adaptation head for the visual encoder to aggregate salient attentive visual cues, which establishes discriminative and task-aware visual representations. We conduct extensive experiments on 11 widely-used benchmark datasets and the results verify the effectiveness in few-shot image classification, especially in generalizing to unseen categories. Compared with the state-of-the-art CoCoOp method, KAPT exhibits favorable performance and achieves an absolute gain of 3.22% on new classes and 2.57% in terms of harmonic mean.

Vision-language pre-training (VLP) models have shown vulnerability to adversarial examples in multimodal tasks. Furthermore, malicious adversaries can be deliberately transferred to attack other black-box models. However, existing work has mainly focused on investigating white-box attacks. In this paper, we present the first study to investigate the adversarial transferability of recent VLP models. We observe that existing methods exhibit much lower transferability, compared to the strong attack performance in white-box settings. The transferability degradation is partly caused by the under-utilization of cross-modal interactions. Particularly, unlike unimodal learning, VLP models rely heavily on cross-modal interactions and the multimodal alignments are many-to-many, e.g., an image can be described in various natural languages. To this end, we propose a highly transferable Set-level Guidance Attack (SGA) that thoroughly leverages modality interactions and incorporates alignment-preserving augmentation with cross-modal guidance. Experimental results demonstrate that SGA could generate adversarial examples that can strongly transfer across different VLP models on multiple downstream vision-language tasks. On image-text retrieval, SGA significantly enhances the attack success rate for transfer attacks from ALBEF to TCL by a large margin (at least 9.78% and up to 30.21%), compared to the state-of-the-art. Our code is available at https://github.com/Zoky-2020/SGA.

Describing a video using natural language is an inherently one-to-many translation task. To generate diverse captions, existing VAE-based generative models typically learn factorized latent codes via one-stage training merely from stand-alone video-caption pairs. However, such a paradigm neglects set-level relationships among captions from the same video, not fully capturing the underlying multimodality of the generative process. To overcome this shortcoming, we leverage neighbouring descriptions for the same video that are articulated with noticeable topics and language variations (i.e., paraphrases). To this end, we propose a novel progressive training method by decomposing the learning of latent variables into two stages that are topic-oriented and paraphrase-oriented, respectively. Specifically, the model learns from divergent topic sentences obtained by semantic-based clustering in the first stage. It is then trained again through paraphrases with a cluster-aware adaptive regularization, allowing more intra-cluster variations. Furthermore, we introduce an overall metric DAUM, aDiversity-AccuracyUnifiedMetric to consider both the precision of the generated caption set and its coverage on the reference set, which has proved to have a higher correlation with human judgment than previous precision-only metrics. Extensive experiments on three large-scale video datasets show that the proposed training strategy can achieve superior performance in terms of accuracy, diversity, and DAUM over several baselines.

Existing vision-language pre-training (VLP) methods primarily rely on paired image-text datasets, which are either annotated by enormous human labors or crawled from the internet followed by elaborate data cleaning techniques. To reduce the dependency on well-aligned image-text pairs, it is promising to directly leverage the large-scale text-only and image-only corpora. This paper proposes a data augmentation method, namely cross-modal CutMix (CMC), for implicit cross-modal alignment learning in unpaired VLP. Specifically, CMC transforms natural sentences in the textual view into a multi-modal view, where visually-grounded words in a sentence are randomly replaced by diverse image patches with similar semantics. There are several appealing proprieties of the proposed CMC. First, it enhances the data diversity while keeping the semantic meaning intact for tackling problems where the aligned data are scarce; Second, by attaching cross-modal noise on uni-modal data, it guides models to learn token-level interactions across modalities for better denoising. Furthermore, we present a new unpaired VLP method, dubbed as VLMixer, that integrates CMC with contrastive learning to pull together the uni-modal and multi-modal views for better instance-level alignments among different modalities. Extensive experiments on five downstream tasks show that VLMixer could surpass previous state-of-the-art unpaired VLP methods.

Dense video captioning aims to generate multiple associated captions with their temporal locations from the video. Previous methods follow a sophisticated "localizethen-describe" scheme, which heavily relies on numerous hand-crafted components. In this paper, we proposed a simple yet effective framework for end-to-end dense video captioning with parallel decoding (PDVC), by formulating the dense caption generation as a set prediction task. In practice, through stacking a newly proposed event counter on the top of a transformer decoder, the PDVC precisely segments the video into a number of event pieces under the holistic understanding of the video content, which effectively increases the coherence and readability of predicted captions. Compared with prior arts, the PDVC has several appealing advantages: (1) Without relying on heuristic non-maximum suppression or a recurrent event sequence selection network to remove redundancy, PDVC directly produces an event set with an appropriate size; (2) In contrast to adopting the two-stage scheme, we feed the enhanced representations of event queries into the localization head and caption head in parallel, making these two sub-tasks deeply interrelated and mutually promoted through the optimization; (3) Without bells and whistles, extensive experiments on ActivityNet Captions and YouCook2 show that PDVC is capable of producing high-quality captioning results, surpassing the state-of-the-art two-stage methods when its localization accuracy is on par with them. Code is available at https://github.com/ttengwang/PDVC.

Dense video captioning aims to localize and describe multiple events in untrimmed videos, which is a challenging task that draws attention recently in computer vision. Although existing methods have achieved impressive performance, most of them only focus on local information of event segments or very simple event-level context, overlooking the complexity of event-event relationship and the holistic scene. As a result, the coherence of captions within the same video could be damaged. In this article, we propose a novel event-centric hierarchical representation to alleviate this problem. We enhance the event-level representation by capturing rich relationship between events in terms of both temporal structure and semantic meaning. Then, a caption generator with late fusion is developed to generate surrounding-event-aware and topic-aware sentences, conditioned on the hierarchical representation of visual cues from the scene level, the event level, and the frame level. Furthermore, we propose a duplicate removal method, namely temporal-linguistic non-maximum suppression (TL-NMS) to distinguish redundancy in both localization and captioning stages. Quantitative and qualitative evaluations on the ActivityNet Captions and YouCook2 datasets demonstrate that our method improves the quality of generated captions and achieves state-of-the-art performance on most metrics.