Physics in Generative Models

Video generation models are increasingly pitched as world models, yet they often fail physics benchmarks. At IDEAS Research Institute I study whether they nonetheless encode physical quantities internally — using simulator-derived ground truth, linear probes on model activations, and steering to test whether that information is causally used.

This brings the question from my earlier work full circle: instead of building models that respect physics, asking what physics large learned models already contain.

Jonas Kneifl
Jonas Kneifl
Postdoctoral Researcher

My research combines model order reduction, surrogate modeling, and machine learning — and asks what physics generative models learn.