Research map / Vision-language and multimodal / Multimodal LLMs
Dynamic-VLM: Simple Dynamic Visual Token Compression for VideoLLM
Han Wang, Yuxiang Nie, Yongjie Ye, Yanjie Wang, Shuai Li, Haiyang Yu, Jinghui Lu, Can Huang
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 |
|---|---|---|---|
| use adaptive pooling instead of token pruning as the dynamic compressor (ablation) | Dynamic-VLM · VideoMME (w/o subtitles) | accuracy: 47.6% -> 52.0% | Improved |
| use adaptive pooling instead of token merging as the dynamic compressor (ablation) | Dynamic-VLM · VideoMME (w/o subtitles) | accuracy: 51.6% -> 52.0% | Improved |
| use adaptive pooling instead of token pruning (ablation) | Dynamic-VLM · MSVD-QA | accuracy: 59.5 -> 62.0 | Improved |
| use adaptive pooling instead of token merging (ablation) | Dynamic-VLM · MSVD-QA | accuracy: 61.9 -> 62.0 | No clear effect |
| raise tokens per frame from 36 to 100 at 12000 max tokens (ablation) | Dynamic-VLM 7B · VideoMME (w/o subtitles) | accuracy: 58.7% -> 60.9% | Improved |
| raise tokens per frame from 100 to 256 at 12000 max tokens (ablation) | Dynamic-VLM 7B · VideoMME (w/o subtitles) | accuracy: 60.9% -> 59.3% | Hurt |
| raise max visual tokens from 4000 to 12000 at 100 tokens per frame (ablation) | Dynamic-VLM 7B · VideoMME (w/o subtitles) | accuracy: 58.2% -> 60.9% | Improved |
| raise max visual tokens from 8000 to 12000 at 36 tokens per frame (more frames, 222 to 333) (ablation) | Dynamic-VLM 7B · VideoMME (w/o subtitles) | accuracy: 59.3% -> 58.7% | No clear effect |
| train on the proposed 2M synthetic video QA data with dynamic compression, compared with LLaVA-OneVision 7B | Dynamic-VLM 7B · VideoMME (w/o subtitles) | accuracy: 58.2% -> 60.9% | Improved |
| same comparison on zero-shot multi-image QA with video-only training | Dynamic-VLM 7B · MuirBench | accuracy: 41.8% -> 50.7% | Improved |
| scale LLM from 7B to 14B | Dynamic-VLM · VideoMME (w/o subtitles) | accuracy: 60.9% -> 64.6% | Improved |
| scale LLM from 7B to 14B | Dynamic-VLM · MMIU | accuracy: 43.7% -> 51.7% | Improved |
Closest papers by meaning
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.91)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.91)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.91)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.91)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.91)
- Benchmarking and Enhancing VLM for Compressed Image Understanding (ICML 2026 · similarity 0.90)
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.
