Causal writability: low-rank edits can restore physically correct motion in video models

ZimingLiu11 · x · 2026-09-26

An arXiv paper (2609.15980) asks whether video models that generate physically incorrect motion failed to learn it or merely failed to use it — and shows the latter. Trained on red masses oscillating slowly and blue ones fast, a model still generates slow motion for a fast-moving red mass, yet a low-dimensional edit predicted from simple physical variables restores correct motion: 'causal writability.'

Key findings:

At fixed edit strength there is a sharp depth boundary: the same edit changes the video before the boundary but not after, marking commitment; stronger downstream writes can still restore motion, though excessive gain overshoots.
Early causal writability predicts which errors later training corrects — correctable errors are writable at more network depths than persistent ones.
Both the phenomenon and its sharp closure reproduce in a pretrained 1.3B video model, suggesting generality across scale and training regime.

The authors frame it against cases like GPT-5.6 Sol mistaking cone colors for lane boundaries in DrivingBench: a wrong output doesn't mean the physics was never learned.

Original post →

Causal writability: low-rank edits can restore physically correct motion in video models

More from Multimodal