AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Do Tabular Foundation Models Know Physics? Contamination, Units, and the Deterministic Limit
Tabular foundation models (TFMs) fill in tables much as language models fill in text, and tables are arguably the format in which most physical measurement arrives; this work probes directly whether their priors contain physics. Four TFMs (TabPFN-3, TabICLv2, TabDPT, Real-TabPFN-2.5) evaluated against six baselines on datasets sampled from 316 physical equations dominate, both out of the box and after tuning. Yet they struggle to represent noiseless deterministic mechanisms in their uncertainty and show no sign of exploiting physical units: they interpolate physics without yet being able to act as physical models.