Research map / Video / Video-language understanding
VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling
Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, Yu Qiao, Yali Wang and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hierarchical compression HiCo (196 to 16 tokens/frame) (ablation) | VideoChat-Flash baseline · MVBench / MLVU | accuracy: MVBench 60.2 -> 61.1; MLVU 63.7 -> 60.6 | Mixed |
| add timestamp prompt (ablation) | VideoChat-Flash · LVBench | accuracy: 44.2 -> 48.0 | Improved |
| add short video pretraining (ablation) | VideoChat-Flash · MVBench | accuracy: 61.1 -> 66.5 | Improved |
| add joint short and long video SFT (ablation) | VideoChat-Flash · MLVU | M-Avg: 68.3 -> 74.5 | Improved |
| use UMT-L video encoder instead of SigLIP at 16 tokens/frame (ablation) | VideoChat-Flash · MVBench (8M data) | accuracy: 71.2 -> 73.5 | Improved |
| use UMT-L instead of SigLIP with small data (2M) (ablation) | VideoChat-Flash · MLVU (2M data) | M-Avg: 62.0 -> 60.0 | Hurt |
| use similar token merging instead of spatial downsampling at 16 tokens/frame (ablation) | connector · MVBench/MLVU/VideoMME avg | avg accuracy: 56.4 -> 58.7 | Improved |
| use token merging at 196 tokens/frame vs uncompressed 729 tokens (ablation) | connector · MVBench/MLVU/VideoMME avg | avg accuracy: 59.6 -> 62.1 | Improved |
| add attention-based visual dropout at layer 18 (50% tokens) (ablation) | Qwen2-7B video LLM · VideoMME | accuracy: 61.2 -> 61.7 | Improved |
| drop 50% of visual tokens uniformly at shallow layer 4 (ablation) | Qwen2-7B video LLM · VideoMME | accuracy: 61.2 -> 60.4 | Hurt |
| swap UMT-L for InternVideo2-1B encoder (ablation) | VideoChat-Flash · MLVU | M-Avg: 74.5 -> 73.4 | Mixed |
Later papers that compared with this method
6 later accepted papers report VideoChat-Flash as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- Small Vision-Language Models are Smart Compressors for Long Video Understanding (ECCV 2026 · better on 5 of 5 shared results)
- RIVER: Real-time Video Interaction Benchmark (ICLR 2026 · better on 3 of 3 shared results)
- LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding (CVPR 2026 · better on 1 of 2 shared results)
- Attend Before Attention: Efficient and Scalable Video Understanding via Autoregressive Gazing (CVPR 2026 · better on 2 of 2 shared results)
- Invert4TVG: A Temporal Video Grounding Framework with Inversion Tasks Preserving Action Understanding Ability (ICLR 2026 · better on 1 of 1 shared result)
- StoryVideoQA: Scaling Deep Video Understanding with a Large-Scale, Multi-Genre and Auto-Generated Dataset (IJCV 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- LongVILA — LongVILA: Scaling Long-Context Visual Language Models for Long Videos (ICLR 2025) · better on 1 of 1 shared result
- LLaMA-VID — LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models (ECCV 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- MovieChat: From Dense Token to Sparse Memory for Long Video Understanding (CVPR 2024 · similarity 0.88)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.88)
- TimeSuite: Improving MLLMs for Long Video Understanding via Grounded Tuning (ICLR 2025 · similarity 0.88)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.88)
- Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
