Video models 'commit' to physics at a depth boundary, new paper finds

ZimingLiu11 · x · 2026-09-16

A new arXiv paper by Ziming Liu et al., "A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models," shows that when video models generate physically wrong motion, the correct motion is still inside the model and can be recovered via low-dimensional edits derived from simple physical variables — a property the authors call causal writability. Crucially, at fixed strength there is a sharp depth boundary: edits change the video before the boundary but not after, marking where the model "commits" to its physics. Stronger downstream writes can restore correct motion, and early writability predicts which errors later training fixes. Findings replicate in a pretrained 1.3B video model.

Original post →

Video models 'commit' to physics at a depth boundary, new paper finds

More from Multimodal