Video Large Language Models (Video LLMs) achieve strong performance on video understanding tasks but suffer from high inference costs due to the large number of visual tokens. We propose KiToke, a 免训练, query-agnostic token compression approach that reduces spatiotemporal redundancy while preserving critical visual information. Our method estimates token diversity globally using a kernel-based redundancy measure, enabling content-adaptive selection that remains effective under extreme token budgets, and further introduces a lightweight temporal interval construction with interval-aware token merging to maintain temporal coherence. Unlike prior methods that rely on local or segment-level heuristics, KiToke explicitly captures global redundancy across an entire video, leading to more efficient token utilization. Extensive experiments on multiple video understanding benchmarks and Video LLM backbones demonstrate that KiToke consistently outperforms existing 免训练 compression methods, with particularly large gains at aggressive retention ratios down to 1%.
英文摘要：
Video Large Language Models (Video LLMs) achieve strong performance on video understanding tasks but suffer from high inference costs due to the large number of visual tokens. We study training-free, query-agnostic video token compression, where heuristic importance rules can over-represent large redundant visual groups and miss less frequent visual patterns, particularly under aggressive token budgets. We propose KiToke, a distribution-aware compression framework that treats visual tokens as samples from the video's spatiotemporal token distribution. KiToke uses global embedding density as a proxy for representational redundancy and inverse-density sampling to balance token allocation across visual patterns of different frequencies. It further aggregates unselected tokens into retained anchors within detected temporal intervals to reduce cross-transition mixing. Across evaluated backbones, KiToke achieves the highest average accuracy among compared training-free methods and remains robust down to 1% retention, where most evaluated baselines underperform uniform random token sampling. On LLaVA-OneVision, it preserves 98.5% and 89.2% of the uncompressed model's four-benchmark mean accuracy at 10% and 1% token retention, with 6.8x and 12.5x prefill speedups, respectively, including compression overhead.
Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior
Harbin Institute of Technology (Shenzhen)（哈尔滨工业大学（深圳））；Shenzhen Loop Area Institute（深圳河套学院）；University of Chinese Academy of Sciences（中国科学院大学）；The Hong Kong University of Science and Technology（香港科技大学）
PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models
Tsinghua University（清华大学）；OpenGVLab（开放通用实验室）；Shanghai AI Laboratory（上海人工智能实验室）；SenseTime Research（商汤科技研究院）；Nanjing University（南京大学）；The Chinese University of Hong Kong（香港中文大学）
Peng Cheng Laboratory, Shenzhen, China（鹏城实验室，深圳，中国）；Department of Electrical Engineering, City University of Hong Kong, Hong Kong SAR（香港城市大学电子工程系，香港特别行政区）；Department of Computer Science, City University of Hong Kong, Hong Kong SAR（香港城市大学计算机科学系，香港特别行政区）；School of Computer Science, Peking University, Beijing, China（北京大学计算机学院，北京，中国）
Are We Using the Right Benchmark: An Evaluation Framework for Visual Token Compression Methods
The Hong Kong University of Science and Technology (Guangzhou)（香港科技大学（广州））；Shanghai Jiao Tong University（上海交通大学）；Northeastern University（东北大学）；Shanghai AI Laboratory（上海人工智能实验室）；The Hong Kong University of Science and Technology（香港科技大学）；MBZUAI
OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models
Media Analytics and Computing Lab, Xiamen University（厦门大学媒体分析与计算实验室）；Institute of Artificial Intelligence, Xiamen University（厦门大学人工智能研究所）；Alibaba Group（阿里巴巴集团）
Balancing Saliency and Coverage: Semantic Prominence-Aware Budgeting for Visual Token Compression in VLMs
Interdisciplinary Program in AI, Seoul National University（首尔国立大学人工智能跨学科项目）；Department of Electrical and Computer Engineering, Seoul National University（首尔国立大学电子与计算机工程系）；Department of Artificial Intelligence, Chung-Ang University（成均馆大学人工智能系）
Feature-Aware Token Attack for Compression-Triggered Stealthy Failures in Large Vision-Language Models
Beijing University of Posts and Telecommunications（北京邮电大学）；Jiangnan University（江南大学）；Nanyang Technological University（南洋理工大学）；Chongqing University of Posts and Telecommunications（重庆邮电大学）
Extreme Model Compression for Edge Vision-Language Models: Sparse Temporal Token Fusion and Adaptive Neural Compression
University of Information Technology and Sciences（信息科技与科学大学）；The University of Burdwan（布尔德万大学）；Tarakeswar Degree College（塔拉克瓦尔学位学院）；University of Missouri, Columbia（密苏里大学，哥伦比亚）
Arbitrary Ratio Feature Compression via Next Token Prediction
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institution of Automation, Chinese Academy of Sciences（多模态人工智能系统国家重点实验室，自动化研究所，中国科学院）；School of Artificial Intelligence, University of Chinese Academy of Sciences（中国科学院大学人工智能学院）；CAS Center for Excellence in Brain Science and Intelligence Technology（中国科学院脑科学与智能技术卓越创新中心）；People AI, Inc.（People AI公司）；School of Artificial Intelligence, Shanghai Jiao Tong University（上海交通大学人工智能学院）；School of Computer Science and Mathematics, Birkbeck College, University of London（伦敦大学伯克贝克学院计算机科学与数学学院）
ApET: Approximation-Error Guided Token Compression for Efficient VLMs
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences（中国科学院深圳先进技术研究所）；Peng Cheng Laboratory（鹏城实验室）；University of Chinese Academy of Sciences（中国科学院大学）；Harbin Institute of Technology（哈尔滨工业大学）；Peking University（北京大学）
