Evaluation Framework Separates Physical Surprise from Appearance in World Models
November 1, 2026
This framework analyzes whether video world model prediction errors stem from physical impossibility or mere appearance changes. Testing on the IntPhys2 benchmark shows that while V-JEPA2 captures some physical impossibility, appearance remains the dominant factor in prediction error.
HOW THIS AFFECTS YOU
●
researcherThis clarifies the distinction between a model's understanding of physics versus its ability to predict visual textures.