Research map / Video / Video-language understanding
VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning
Yang Ding, Xin Lai, Yizhen Zhang, Wei Li, Ruihang Chu, Yujiu Yang
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 RL stage after cold-start SFT (ablation) | VideoZoomer · LongVideoReason-eval | accuracy: 63.3 -> 80.3 | Improved |
| add cold-start SFT initialization before RL (ablation) | VideoZoomer · LongVideoReason-eval | accuracy: 59.6 -> 80.3 | Improved |
| add conditional tool-use bonus R_tool (ablation) | VideoZoomer · VideoMME overall | accuracy: 62.5 -> 65.2 | Improved |
| add conditional tool-use bonus R_tool (ablation) | VideoZoomer · VideoMMMU | accuracy: 53.8 -> 52.2 | Hurt |
| add reflection trajectories to cold-start data (ablation) | VideoZoomer · VideoMME long | accuracy: 47.4 -> 55.8 | Improved |
| add reflection trajectories to cold-start data (ablation) | VideoZoomer · VideoMMLU quiz | accuracy: 70.1 -> 67.9 | Hurt |
| allow one tool call instead of none (ablation) | VideoZoomer · LongVideoReason-eval | accuracy: 64.8 -> 79.9 | Improved |
| raise max tool calls from 2 to 4 (ablation) | VideoZoomer · LongVideoReason-eval | accuracy: 80.2 -> 80.3 | No clear effect |
| replace uniform initial frames with TSPO-selected frames (ablation) | VideoZoomer · MLVU dev | accuracy: 68.8 -> 70.8 | Improved |
| distill cold-start data from Gemini 2.5 Pro instead of GPT-4o (ablation) | VideoZoomer · VideoMME overall | accuracy: 61.6 -> 65.2 | Improved |
| double SFT data from about 11k to about 20k samples (ablation) | VideoZoomer · LongVideoBench val | accuracy: 57.7 -> 55.6 | Hurt |
| train VideoZoomer with agentic zoom instead of base Qwen2.5-VL | VideoZoomer · MLVU test | accuracy: 45.5 -> 55.8 | Improved |
Later papers that compared with this method
2 later accepted papers report Video-Zoomer as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning (ICML 2026 · better on 4 of 4 shared results)
- VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting (ECCV 2026 · better on 1 of 1 shared result)
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
- FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting (ICLR 2026 · similarity 0.92)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.90)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.89)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.89)
- Reinforcing Structured Chain-of-Thought for Video Understanding (CVPR 2026 · similarity 0.89)
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.
