GeoPhys: The Geometry of Physical Plausibility

C Internò, A Pondaven, H Issa, F Pizzati, F Pinto, M Olhofer, I Laptev, P Torr, E P Simoncelli, B Hammer and D Klindt.

Published in Adv. Neural Information Processing (NeurIPS), vol.39 Dec 2026. To appear

Download:
  • Reprint (pdf)
    • While humans can identify physically implausible events within milliseconds, machine learning approaches addressing the same problem are extremely slow and expensive. They either rely on external multimodal-LLM judges or require ad-hoc modifications to the training procedure. In this work, we argue that indicators of physical plausibility are implicitly captured by five geometric properties of the per-frame embeddings produced by frozen image encoders. In aggregate, we call them GeoPhys. First, we show that these signals correlate with human EEG responses to two forms of object-permanence violations. Second, GeoPhys robustly discriminates physically implausible videos from realistic ones, achieving state-of-the-art physics-violation detection: 98.% on LikePhys [Yuan et al, 2026] and 93.3% on IntPhys2 [Bordes et al, 2025], whereas V-JEPA~2, GPT-4o, Gemini, and twelve modern video diffusion models perform near chance. Third, used as a best-of-$N$ verifier for physical alignment during video generation, GeoPhys lifts MAGI-1~24B from 50.01% to 64.50% on PhysicsIQ [motamed et al, 2025] at 1.5x lower wall-clock and 4.65x lower memory than the V-JEPA~2 world-model verifier. Ultimately, GeoPhys demonstrates that physical plausibility in videos can be assessed by leveraging the emergent geometric properties of temporal features extracted from image encoders.
  • Superseded Publications: Interno26a
  • Related Publications: Niu24b, Henaff20, Henaff18
  • Listing of all publications