A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models (arXiv 2609.15980)

Forum topic · 小凯 · 2026-09-16

Summary

This paper introduces causal writability, a property of video generation models showing that when a model produces physically incorrect motion, the correct motion is still encoded internally and can be restored via targeted edits. The authors train models on videos where red masses oscillate slowly and blue masses oscillate quickly, then test with a red mass exhibiting fast motion. Even when the model generates incorrect slow motion in this conflicting scenario, a low-dimensional edit predicted from simple physical variables recovers the correct fast motion. At fixed edit strength, there is a sharp depth boundary: the same edit changes the video before the boundary but not after it, marking the point where the write is committed. The motion signal persists after closure, and stronger downstream writes can recover the physics, though excessive gains overshoot. Early causal writability predicts which errors will be corrected during later training: correctable errors remain writable across more network depths than persistent errors. The phenomenon and its sharp closure are reproduced in a pretrained 1.3B video model, supporting generality across model scales and training paradigms. The work is available as arXiv 2609.15980 (computer vision).

Paper Overview

Research area: Computer Vision (CV)
Authors: Xingyun Wang, Haomin Zheng, Man Yuan, Leqian Yang, Ziming Liu
Published: 2026-09-14
arXiv: 2609.15980

Key Findings

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? The authors show the latter: the correct motion remains available inside the model and can still be made to control the generated video.

Causal writability: Models are trained on videos where red masses oscillate slowly and blue masses oscillate quickly, then tested on 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.
Sharp depth boundary: At fixed edit strength, the same edit changes the video before a certain network depth but not after it. This closure marks commitment for that write.
Signal persistence: The motion signal never fully disappears — stronger downstream writes can still recover the physical motion, but gains that are too large overshoot.
Predicting error correction: Early causal writability predicts which errors will be corrected in later training. Errors that are eventually corrected remain writable across more network depths than persistent errors.
Generality: Causal writability and its sharp closure are reproduced in a pretrained 1.3B video model, supporting generality across model scales and training paradigms.

Abstract (Original)

> 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.*