AI TrackerFrontier desk · AEST
Back to news
normalaitech

PhysiCalWorld: In-Context Physics Calibration for World Models

arXiv cs.AI · 22 Oct 2026, 03:57 AEST

PhysiCalWorld introduces a method for calibrating world models to specific physics dynamics without the need for fine-tuning. This approach enhances prediction accuracy and visual planning success across diverse environments.

Key facts

Read aloud

PhysiCalWorld introduces a method for calibrating world models to specific physics dynamics without the need for fine-tuning. This approach enhances prediction accuracy and visual planning success across diverse environments. Existing world models typically perceive a scene through current visual frame or a narrow history window. Without calibrating these physics factors for world models, predictions tend to regress toward an empirical prior and induce severe drift in novel test scenes. Our method encodes an exploratory interaction clip collected within the test scene into a compact physics descriptor token, which grounds subsequent predictions in the scene's actual physical behavior. PhysiCalWorld is trained end-to-end with the latent prediction loss alone, without physics labels. We evaluate PhysiCalWorld across four environments covering diverse physics variations and show that it substantially improves rollout prediction accuracy and visual planning success rate.

Sources in this cluster