Research map / Video / Video-language understanding
Keyframe-oriented Vision Token Pruning: Enhancing Efficiency of Large Vision Language Models on Long-Form Video Processing
Yudong Liu, Jingwei Sun, Yueqian Lin, Jianyi Zhang, Jingyang Zhang, Ming Yin, Qinsi Wang, Hai Li, Yiran Chen
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 KVTP keyframe-oriented pruning rates | PruMerge (7B) · VideoMME | accuracy: 59.77 -> 63.29 | Improved |
| Add KVTP keyframe-oriented pruning rates | PruMerge (7B) · EgoSchema | accuracy: 52.49 -> 54.71 | Improved |
| Add KVTP keyframe-oriented pruning rates | PruMerge (7B) · NextQA | accuracy: 75.58 -> 76.76 | Improved |
| Add KVTP keyframe-oriented pruning rates | ToMe (7B) · VideoMME | accuracy: 58.90 -> 62.36 | Improved |
| Add KVTP keyframe-oriented pruning rates | Random Sampling (7B) · VideoMME | accuracy: 58.28 -> 60.16 | Improved |
| Add KVTP keyframe-oriented pruning rates | PruMerge (72B) · VideoMME | accuracy: 64.52 -> 67.34 | Improved |
| Add KVTP keyframe-oriented pruning rates | ToMe (72B) · NextQA | accuracy: 76.45 -> 79.51 | Improved |
| Add KVTP keyframe-oriented pruning rates | Random Sampling (72B) · EgoSchema | accuracy: 60.47 -> 62.19 | Improved |
| Use KVTP+PruMerge instead of FastV on 7B (FLOPs x36% vs x64%) | LLaVA-Video-7B · NextQA | accuracy: 77.89 -> 76.76 | Hurt |
| Use soft frame selection (KVTP) instead of hard selection (KeyVideoLLM) | LLaVA-Video-7B · VideoMME | accuracy: 51.32 -> 63.29 | Improved |
| Fine-tune SigLIP predictor (KVTP) instead of using GPT assigned scores (ablation) | KVTP pruning rates · VideoMME | accuracy: 62.83 -> 63.29 | Improved |
| Fine-tune SigLIP predictor instead of fine-tuning backbone and projector (KeyVideoLLM-style) (ablation) | KVTP pruning rates · VideoMME | accuracy: 61.23 -> 63.29 | Improved |
Closest papers by meaning
- ViTCoP: Accelerating Large Vision-Language Models via Visual and Textual Semantic Collaborative Pruning (AAAI 2026 · similarity 0.92)
- ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models (CVPR 2025 · similarity 0.91)
- QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models (CVPR 2026 · similarity 0.91)
- IF-Prune: Information-Flow Guided Token Pruning for Efficient Vision-Language Models (CVPR 2026 · similarity 0.91)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.91)
- An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models (ICLR 2026 · similarity 0.91)
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.
