Research map / Video / Video-language understanding
Token Reduction via Local and Global Contexts Optimization for Efficient Video Large Language Models
Jinlong Li, Liyuan Jiang, Haonan Zhang, Nicu Sebe
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 local anchors alongside global anchors (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 56.6 -> 57.0 | Improved |
| add global anchors alongside local anchors (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 55.4 -> 57.0 | Improved |
| add optimal transport aggregation of pruned tokens (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 56.4 -> 57.0 | Improved |
| add intra-frame OT (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 56.3 -> 57.0 | Improved |
| add inter-frame OT (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 56.4 -> 57.0 | Improved |
| use transport-plan weighting instead of cosine merging (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 52.4 -> 57.0 | Improved |
| use transport-plan aggregation instead of no merging (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 56.4 -> 57.0 | Improved |
| select anchors by attention instead of randomly (ablation) | AOT · Video benchmarks average (ablation, 10% tokens retained) | average score: 55.1 -> 57.0 | Improved |
| use dynamic clustering frame clips instead of uniform clips (10% retention) (ablation) | AOT · Video benchmarks average (LLaVA-OV, 10% tokens retained) | average score: 57.0 -> 56.5 | Hurt |
| use dynamic clustering frame clips (LLaVA-Video, 25%) (ablation) | AOT · Video benchmarks average (LLaVA-Video, 25% tokens retained) | average score: 58.2 -> 58.6 | Improved |
| raise Sinkhorn iterations from 50 to 100 (ablation) | AOT · Sinkhorn-Knopp overhead | time (ms): 2.00 ms -> 2.11 ms | No clear effect |
| AOT vs VisionZip at 10% retention on LLaVA-OV | AOT vs VisionZip · Video benchmarks average (LLaVA-OV, 10% tokens retained) | average score: 53.5 -> 57.0 | Improved |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.91)
- HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.91)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.90)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.90)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.90)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · 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.
