Training Object Permanence in World Models

WROP trains and tests video models on whether objects persist when hidden and behave consistently around solid barriers. Its synthetic scenes include objects passing behind occluders and reappearing with their identity intact. The generators vary speed, lighting and camera angle while retaining the underlying physical challenge.

The project releases a 1.5 million sample training corpus and a 300-question exam. In a blind pairwise study of outputs from 14 models, the authors' own model, the only one trained on WROP data, ranked first among those that continue an input video and third overall. That shows an effect of training on these controlled synthetic tasks, not general physical understanding, and gives researchers a focused way to examine physical consistency beyond visual polish.