Do Tabular Foundation Models Know Physics? Contamination, Units, and the Deterministic Limit
The models performed well on sampled physics equations, but failed the tests that would make them usable as physical models.
The paper evaluates four tabular foundation models against six baselines on datasets drawn from 316 physical equations. It finds the TFMs lead both out of the box and after tuning. The limit, according to the authors, is that their learned prior does not represent noiseless mechanisms or physical units. That leaves them able to interpolate physical data without actually encoding physics in the stronger sense. ArXiv · AI/CL/LG's note
The paper evaluates four tabular foundation models against six baselines on datasets drawn from 316 physical equations. It finds the TFMs lead both out of the box and after tuning. The limit, according to the authors, is that their learned prior does not represent noiseless mechanisms or physical units. That leaves them able to interpolate physical data without actually encoding physics in the stronger sense. ArXiv · AI/CL/LG's note
score 4