PhyS arXiv

Streaming world models with physical priors

Distilling Physical Priors into Streaming World Models

Physically coherent video generation across rigid bodies, soft bodies, fluids, and phase transitions.

Liangliang Zhao1,2 · Junying Wang1,2 · Danni Yang1,2 · Yifan Chang3,4 · Bin Fu2 · Yu Qiao2 · Bowen Zhou2 · Yihao Liu2,✉

1Fudan University   2Shanghai AI Laboratory   3University of Science and Technology of China   4Shanghai Innovation Institute

arXiv Code Coming soon Dataset Coming soon
Watch the paper demo

Paper video / overview

TL;DR: PhyS distills physical priors from 120K real interactions into streaming world models, then preserves them through causal distillation and temporally routed reinforcement learning—yielding more coherent long-horizon physical events.

Distilling Physical Priors into Streaming World Models 3:51 · dataset, method, and qualitative results

01 / PhyS-120K dataset

Physical interactions, captured at scale.

120,804 real-world videos spanning rigid bodies, soft bodies, fluids, and phase transitions, paired with structured physical annotations.

Rigid bodies Collision and momentum transfer
Soft bodies Compression and recovery
Fluids Gravity-driven pouring
Phase transitions Liquid-to-solid cooking
Soft bodies Tearing and fracture
Rigid bodies Sequential chain reaction

Structured annotations

Objects, properties, interactions, and causal event sequences.

Two examples from the annotation pipeline. Each record preserves the original JSON fields and complete model output.

Citation

@article{zhao2026distilling,
  title={Distilling Physical Priors into Streaming World Models},
  author={Zhao, Liangliang and Wang, Junying and Yang, Danni and Chang, Yifan and Fu, Bin and Qiao, Yu and Zhou, Bowen and Liu, Yihao},
  journal={arXiv preprint arXiv:2608.07981},
  year={2026}
}

Distilling Physical Priors into Streaming World Models

arXiv Code Coming soon Dataset Coming soon