Reproduce “Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuan Xie, Tianshui Chen, Zheng Ge, Lionel Ni
A reinforced multi-turn reasoning framework called Video-MTR enables multimodal large language models to iteratively select critical segments for understanding long-form videos. This approach achieves competitive performance on five long-form video benchmarks using a pure reinforcement learning paradigm with only 8K training samples, demonstrating efficiency comparable to much larger models with significantly more input frames.
Key insights
- Dynamic multi-turn reasoning allows MLLMs to efficiently allocate limited visual evidence across extended temporal contexts, outperforming static, single-pass methods for long videos.
- A fine-grained, goal-gated bi-level reward system is essential for stable multi-turn RL training, preventing 'reward hacking' and ensuring intermediate actions genuinely contribute to the final task objective.
- Multi-turn reasoning enables competitive performance on long-form video understanding with significantly smaller frame budgets (e.g., 80 frames) compared to single-turn models requiring much more input.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.