A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models
The paper argues that video models can contain the right physical motion even when their output commits to the wrong one.
The authors test this with color-coded oscillating masses, then show that a low-dimensional edit tied to physical variables can restore fast motion in a case where the model instead generates slow motion. They call the effect “causal writability,” and report a sharp network-depth boundary after which the same edit no longer changes the video. Stronger downstream edits can still recover the motion signal, though too much gain overshoots. They also reproduce the behavior in a pretrained 1.3B video model. ArXiv · AI/CL/LG's note
The authors test this with color-coded oscillating masses, then show that a low-dimensional edit tied to physical variables can restore fast motion in a case where the model instead generates slow motion. They call the effect “causal writability,” and report a sharp network-depth boundary after which the same edit no longer changes the video. Stronger downstream edits can still recover the motion signal, though too much gain overshoots. They also reproduce the behavior in a pretrained 1.3B video model. ArXiv · AI/CL/LG's note
score 5