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.
More from Multimodal
- World models could complement, not replace, game engines for prototyping — CodeByPoonam · 2026-09-26
- Claude made a music video for a user's song — and it's "insane" — AndyMasley · 2026-09-26
- Studio Work Cost $10,000, Now $10 in Tokens: Flowith Canvas Runs on Opus 5.5 — Scobleizer · 2026-09-26
- LoRA Pilot unveils guided interface covering captioning to checkpoint comparison — no3us · 2026-09-26
- Dev turns a café into an ocean on Vision Pro using AI-generated 3D assets — Scobleizer · 2026-09-26
- Dataset Distillation project turns an artist's body of work into striking synthetic images — CSProfKGD · 2026-09-26