Causal Writability in Video Models: Correct Motion Stays Editable Until a Sharp Depth Boundary

Forum topic · 小凯 · 2026-09-16

Summary

A 2026 arXiv paper (2609.15980) from researchers including Xingyun Wang and Haomin Zheng investigates whether video generation models that produce physically incorrect motion failed to learn the correct motion or simply failed to use it. The authors demonstrate the latter: correct motion remains available inside the model and can still control generated video. Training on videos where red masses oscillate slowly and blue masses oscillate quickly, then testing on a red mass with fast observed motion, they show that even when the model generates slow motion under this conflicting condition, a low-dimensional edit predicted from simple physical variables restores the correct fast motion. They term this ability causal writability. At fixed edit strength, a sharp depth boundary exists: the same edit changes the video before the boundary but not after, marking a point of commitment. Early-layer writability predicts which errors are corrected during later training. The phenomenon and its sharp closure are reproduced in a pretrained 1.3B-parameter video model, supporting generality across model scales and training paradigms.

Paper: A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models Field: Computer Vision (CV) Authors: Xingyun Wang, Haomin Zheng, Man Yuan, Leqian Yang, Ziming Liu arXiv: 2609.15980

Key points

Core question: When a video model generates physically incorrect motion, did it fail to learn the correct motion, or learn it but fail to use it? The paper shows the latter: the correct motion remains available inside the model and can still be made to control the generated video.
Experimental setup: The model is trained on videos where red masses oscillate slowly and blue masses oscillate quickly. At test time, a red mass is given fast observed motion, creating a color–speed conflict.
Causal writability: Even when the model generates slow motion in the conflicting case, a low-dimensional edit — predicted from simple physical variables — restores the correct fast motion. The authors call this ability *causal writability*.
Sharp depth boundary: At fixed edit strength, the same edit changes the video before a certain depth but not after it. This closure point marks commitment for that write. The motion signal never fully disappears; stronger downstream writes can recover the physical motion, though excessive gain overshoots.
Predictive value: Early-stage causal writability predicts which errors will be corrected in subsequent training — correctable errors remain writable at more network depths than persistent errors.
Generality: The findings and the sharp closure behavior are reproduced in a pretrained 1.3B-parameter video model, supporting the phenomenon's generality across model scales and training paradigms.

Original abstract (excerpt)

> When a video model generates physically incorrect motion, did it fail to learn the correct motion, or did it learn it but fail to use it? We show the latter: the correct motion remains available inside the model and can still be made to control the generated video. We train on videos where red masses oscillate slowly and blue masses oscillate quickly, then test a red mass with fast observed motion. Even when the model generates slow motion in this conflicting case, a low-dimensional edit predicted from simple physical variables restores the correct fast motion. We call this ability causal writability. At fixed strength, we find a sharp depth boundary: the same edit changes the video before the boundary but not after it. This closure marks commitment for that write. The motion signal nevert...

--- *Auto-collected on 2026-09-16*

Causal Writability in Video Models: Correct Motion Stays Editable Until a Sharp Depth Boundary

Summary

Key points

Original abstract (excerpt)

Tags