Event

PhysiCalWorld uses in-context prompting to calibrate world model physical parameters

1 reports · 1 sources

Yunhao Luo et al. propose the PhysiCalWorld method, which performs instance-level physical calibration of an action-conditioned world model via in-context prompting, without fine-tuning or analytical physical modeling. The method encodes an exploratory interaction segment collected in the test scene into compact physical descriptor tokens, so that subsequent predictions are anchored to the scene's actual physical behavior; the model is trained end-to-end using only a latent prediction loss and does not rely on physical labels. At inference time, the descriptor tokens need to be extracted only once and can be reused across multiple rollouts, with minimal computational overhead. The authors report that, evaluated in four environments covering a variety of physical changes, the method significantly improves rollout prediction accuracy and visual planning success rates.