xzxxntxdy/History-Forcing

History Forcing: Distilling History-Dependent Continuations for Long-Horizon Video World Models

Hangyi Kuang^1,*, Chengzhi Wang^1,*, Yefei He¹, Weijie Wang¹, Yuke Zhao^1,2, Zhuokun Chen³
Zisu Huang², Yifan Yang², Xiaoxuan He^1,†, Feng Chen^4,†, Bohan Zhuang¹

¹ Zhejiang University   ² Microsoft Research   ³ Monash University   ⁴ Adelaide University
^* Equal contribution   ^† Corresponding authors

Paper · Project · Code · Models · Dataset

History Forcing is a unified on-policy distillation framework for few-step autoregressive video generators. It uses the residual between teacher predictions with and without informative history to correct the score or velocity target used for distillation, improving long-video consistency and out-of-sight scene evolution.

Checkpoint	Model
opsd/model.pt	DreamX-World-5B
dmd/model.pt	Wan2.1-T2V-1.3B, raw and EMA
ar_teacher/model.pt	Wan2.1-T2V-1.3B AR teacher

The released DMD checkpoint was trained with a 1.3B bidirectional teacher and the 10-second AR teacher.

Inference and training are available in the code.

News

• 2026-11-03: Our paper is now available on arXiv.

Citation

@article{kuang2026historyforcing,
  title={History Forcing: Distilling History-Dependent Continuations for Long-Horizon Video World Models},
  author={Kuang, Hangyi and Wang, Chengzhi and He, Yefei and Wang, Weijie and Zhao, Yuke and Chen, Zhuokun and Huang, Zisu and Yang, Yifan and He, Xiaoxuan and Chen, Feng and Zhuang, Bohan},
  journal={arXiv preprint arXiv:2611.02782},
  year={2026}
}