Research map / Video / Video-language understanding
Contribution-aware Token Compression for Efficient Video Understanding via Reinforcement Learning
Yinchao Ma, Qiang Zhou, Zhibin Wang, Xianing Chen, Hanqing Yang, Jun Song, Bo Zheng
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 OCSS instead of random sampling in policy training (ablation) | CaCoVID (LLaVA-OV-7B) · VideoMME | accuracy: 54.4 -> 58.5 | Improved |
| use OCSS instead of multinomial sampling (ablation) | CaCoVID · VideoMME | accuracy: 55.1 -> 58.5 | Improved |
| sample within each frame rather than over whole video (ablation) | CaCoVID · VideoMME | accuracy: 57.2 -> 58.5 | Improved |
| raise training sampling ratio from 2% to 5% (ablation) | CaCoVID · VideoMME | accuracy: 58.5 -> 57.3 | Hurt |
| lower training sampling ratio from 2% to 1% (ablation) | CaCoVID · VideoMME | accuracy: 58.5 -> 58.3 | No clear effect |
| add ineffective sample filtering (ablation) | CaCoVID · VideoMME | accuracy: 57.1 -> 57.3 | Improved |
| add experience replay (ablation) | CaCoVID · VideoMME | accuracy: 57.3 -> 58.1 | Improved |
| add dynamic sample ratio (ablation) | CaCoVID · VideoMME | accuracy: 58.1 -> 58.5 | Improved |
| adaptive per-frame token budget instead of equal (ablation) | CaCoVID · VideoMME | accuracy: 57.4 -> 58.1 | Improved |
| add uniform spatio-temporal tokens to retained set (ablation) | CaCoVID · VideoMME | accuracy: 58.1 -> 58.5 | Improved |
| use CaCoVID instead of DivPrune/PruneVID for compression (ablation) | LLaVA-OneVision-7B 25% · avg accuracy | accuracy: 55.0 -> 55.8 | Improved |
| use CaCoVID instead of PruneVID at 10% retention | Qwen2.5-VL-3B · avg accuracy | accuracy: 50.2 -> 52.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FrameFusion — FrameFusion: Combining Similarity and Importance for Video Token Reduction on Large Vision Language Models (ICCV 2025) · better on 3 of 3 shared results
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 0.90)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.88)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.87)
- EvoComp: Learning Visual Token Compression for Multimodal Large Language Models via Semantic-Guided Evolutionary Labeling (CVPR 2026 · similarity 0.87)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.87)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.87)
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.
