Research map / Video / Video-language understanding
Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding
Yuan Xie, Tianshui Chen, Zheng Ge, Lionel Ni
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 multi-turn retrieval reasoning instead of single-turn with uniformly sampled frames (ablation) | Video-MTR (Qwen2.5-VL-7B, 80 frames) · LVBench | accuracy: 35.3 -> 42.3 | Improved |
| use multi-turn retrieval reasoning instead of single-turn (ablation) | Video-MTR (Qwen2.5-VL-7B, 80 frames) · VideoMME (w/o sub.) overall | accuracy: 57.2 -> 62.7 | Improved |
| add turn-level frame-relevance reward (bi-level reward) (ablation) | Video-MTR multi-turn · LVBench | accuracy: 37.7 -> 42.3 | Improved |
| add turn-level frame-relevance reward (bi-level reward) (ablation) | Video-MTR multi-turn · VideoMME (w/o sub.) overall | accuracy: 58.3 -> 62.7 | Improved |
| gate turn-level rewards on final answer correctness (ablation) | Video-MTR · training dynamics | QA accuracy and number of turns | Improved |
| apply Video-MTR RL training to Qwen2.5-VL-7B at same frame budget | Qwen2.5-VL-7B, 80 frames · LVBench | accuracy: 33.6 -> 42.3 | Improved |
| apply Video-MTR RL training to Qwen2.5-VL-7B at same frame budget | Qwen2.5-VL-7B, 80 frames · EgoSchema (Subset) | accuracy: 63.5 -> 68.8 | Improved |
| increase frame budget from 32 to 80 | Video-MTR · VideoMME (w/o sub.) | accuracy: 59.0 -> 62.7 | Improved |
| multi-turn reasoning on long videos versus short videos | Video-MTR vs Qwen2.5-VL-7B, 32 frames · VideoMME (w/o sub.) | accuracy gain: Short +4.6, Medium +5.3, Long +6.3 | Improved |
| multi-turn reasoning on holistic versus detail tasks | Video-MTR vs Qwen2.5-VL-7B · MLVU | accuracy gain: holistic +3.8%, single-detail +7.5%, multi-detail +8.1% | Improved |
| multi-turn RL with Video-MTR still lags large proprietary model on some benchmarks | Video-MTR 80 frames · VideoMME (w/o sub.) | accuracy: 62.7 vs GPT-4o 71.9 | No clear effect |
Later papers that compared with this method
3 later accepted papers report Video-MTR as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Conan: Progressive Learning to Reason Like a Detective over Multi-Scale Visual Evidence (CVPR 2026 · better on 6 of 6 shared results)
- VideoTIR: Accurate Understanding for Long Videos with Efficient Tool-Integrated Reasoning (ECCV 2026 · better on 1 of 2 shared results)
- VideoSEAL: Mitigating Evidence Misalignment in Agentic Long Video Understanding by Decoupling Answer Authority (ICML 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)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.90)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.90)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.90)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · 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.
