Signals of physical plausibility are hiding in the geometry of frozen image encoders. No video training. No physics supervision.

This work demonstrates that frozen image encoders (neural networks trained on static images without video or physics supervision) contain implicit geometric signals that correspond to physical plausibility. The research suggests that visual understanding of physical properties emerges naturally from image-based training without explicit physics annotations or temporal video data. This finding indicates that fundamental principles of physical reasoning may be encoded in the geometric structure of standard vision models.

Source: Yann LeCun (X) | 2026-06-21