Paper tests physics-based assumption in RL simulators and world models

A new arXiv preprint examines the widely held assumption that learned dynamics models which conform to underlying physics produce more accurate predictions. The authors recover exact polynomial invariants from trajectories and canonicalise them as a method for diagnosing faults in reinforcement learning simulators and world models. The work frames these invariants as a diagnostic tool for checking whether a model's learned dynamics actually respect physical structure.