Reproduce “Distilling Physical Priors into Streaming World Models” 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
Liangliang Zhao, Junying Wang, Danni Yang, Yifan Chang, Bin Fu, Yu Qiao, Bowen Zhou, Yihao Liu
PhyS trains streaming video generators by distilling physical supervision from annotated real videos into a causal model and applying time-localized feedback to rollouts. Its PhyS-120K dataset contains 120,804 videos. On 10-second PhysicsIQ rollouts at 4 NFE, Self Forcing with Temporal Credit Routing scores 20.9 versus 16.9 without it, a 23.7% benchmark gain, not proof of general physical correctness.
Key insights
- Real-world, process-centric video data coupled with structured, causal annotations is effective for imbuing generative models with robust physical priors.
- Transferring physical priors from a powerful bidirectional teacher to an efficient causal student model is feasible but requires careful distillation techniques.
- Temporally localized physical feedback and precise credit assignment (via TCR) are superior to global rewards for refining autoregressive generation and correcting physical violations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.