Reproduce “A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models” on OpenResearch

About this paper

Xingyun Wang, Haomin Zheng, Man Yuan, Leqian Yang, Ziming Liu

The authors test whether a video model can recover a physically consistent future that conflicts with a learned color cue. In controlled spring videos, they edit internal states using boundary position and velocity, then evaluate decoded motion. A fit-only controller places 85.9% of held-out writes in a near-full recovery range; this shows causal access, not physical accuracy or broad-world-model behavior.

Key insights

• Physical motion and appearance compete as predictive cues: longer observed histories favor the physically consistent future, while ambiguous colors allow motion to exert more influence and can also determine future appearance.
• The correct physical alternative can remain causally available even after the model commits to a shortcut; four principal components preserve most of the intervention effect, and state-conditioned controllers achieve high recovery on held-out cases.
• Writability closes over a narrow depth range rather than disappearing uniformly, and its extent is predictive: errors with wider early writable regions are more likely to be corrected by continued training.

Summary by alphaXiv.