Reproduce “Video-MOPD: Multi-Teacher On-Policy Distillation for 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
Zhenxin Qin, Peng Shi, Cong Han, Yinlong Qian, Zequn Jie, Lin Ma
Researchers at Tongji University and Bilibili Inc. developed Video-MOPD, an open-weight 8B model that unifies diverse video understanding capabilities—including general comprehension, temporal grounding, and STEM reasoning—using a "specialize-then-unify" post-training pipeline with Multi-Teacher On-Policy Distillation. The model achieved an average score of 69.12 across seven video benchmarks, demonstrating a 5.4-point improvement over its base model and outperforming parameter averaging by 1.84 points.
Key insights
- A "specialize-then-unify" strategy effectively integrates diverse video understanding capabilities by first optimizing domain-specific experts with RL and then distilling their knowledge.
- Multi-Teacher On-Policy Distillation (MOPD) enables successful unification by providing dense, token-level supervision from specialized teachers on trajectories generated by the student model itself.
- Reliability-Aware Informative Sampling (RAIS) improves distillation efficiency by ensuring reliable teacher guidance and prioritizing examples with higher learning utility for the student.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.