Causal writability: correct physics lives inside video models and can be restored via edits

ZimingLiu11 · x · 2026-09-16

An arXiv paper by Ziming Liu et al. shows that when a video model generates physically wrong motion, it has usually learned the correct physics but failed to use it. A low-dimensional edit predicted from simple physical variables restores the right motion — dubbed "causal writability." At fixed strength there's a sharp depth boundary beyond which the edit no longer works, marking the model's "commitment" to its physics; stronger writes can still recover it, though excessive gain overshoots. Early writability predicts which errors later training corrects, and both effects reproduce in a pretrained 1.3B video model.

Original post →

Causal writability: correct physics lives inside video models and can be restored via edits

More from Research