Research map / Video / Video-language understanding
Efficient Frame Selection for Long Video Understanding via Reinforcement Learning
Yaxuan Qin, Hefei Li, Wenqi Mu, Yancheng He
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 pre-training stage before RL (ablation) | frame selector · VideoMME (ablation table) | accuracy: 54.8 -> 56.5 | Improved |
| add RL stage after pre-training (ablation) | frame selector · VideoMME (ablation table) | accuracy: 55.1 -> 56.5 | Improved |
| add frame-level reward to combination-level reward (ablation) | frame selector RL · VideoMME (ablation table) | accuracy: 55.6 -> 56.5 | Improved |
| add KL monitor with rollback (ablation) | frame selector RL · VideoMME (ablation table) | accuracy: 55.2 -> 56.5 | Improved |
| use selector instead of uniform sampling across MLLMs and benchmarks | seven MLLMs · MLVU (12min) | accuracy: average +6.16% | Improved |
| use selector instead of uniform sampling | seven MLLMs · EgoSchema (3min) | accuracy: average +0.99% | Improved |
| use selector instead of uniform sampling at 16 frames (ablation) | LLaVA-NeXT-Video-34B · LongVideoBench, 16 frames | accuracy: 50.1 -> 50.2 | No clear effect |
| use selector at 2 frames instead of uniform (ablation) | Qwen2-VL-7B · LongVideoBench, 2 frames | accuracy: 46.8 -> 52.1 | Improved |
| use trained selector instead of CLIP-TopK | Qwen2-VL-7B · LongVideoBench | accuracy: 55.7 -> 56.5 | Improved |
| use trained selector instead of AKS | Qwen2-VL-7B · LongVideoBench | accuracy: 55.9 -> 56.5 | Improved |
Closest papers by meaning
- Flexible Frame Selection for Efficient Video Reasoning (CVPR 2025 · similarity 0.91)
- Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025 · similarity 0.91)
- Adaptive Keyframe Sampling for Long Video Understanding (CVPR 2025 · similarity 0.91)
- M-LLM Based Video Frame Selection for Efficient Video Understanding (CVPR 2025 · similarity 0.91)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.90)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 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.
