PhiZero: A World Model Built Around Physical Language

Researchers introduced PhiZero, a world model that predicts physical changes with a discrete physical-language representation.

PhiZero learns a compact physical-language code from unlabeled videos and uses a reason-then-render pipeline to generate future states.

It showed coherent physical evolution, interactive simulation, action-conditioned prediction, and zero-shot motion transfer in evaluation.

The approach could offer a more interpretable, realistic, and controllable alternative to direct pixel prediction for video generation.