PreprintCC BY 4.0arXiv
Do Tabular Foundation Models Know Physics? Contamination, Units, and the Deterministic Limit
Abstract
Tabular foundation models (TFMs) learn to fill in tables the way language models fill in text --- and tables are arguably the format in which most physical measurement arrives. Did they learn any physics in the process? They are Bayesian by construction, so the question is what their prior contains. We probe it directly, evaluating four of them (TabPFN−3, TabICLv2, TabDPT and Real−TabPFN−2.5) against six baselines on datasets sampled from 316 physical equations, in and out of domain. TFMs dominate, out of the box and after tuning. But 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.
Related
- Tabular Foundation Models Dramatically Outperform Bespoke Photometric Redshift Estimators
- χ‑Unification: The Unified Spectral–Topological Foundation of Physics
- Cutoff-Safe Tabular Fusion of TESS Data-Validation Products for Temporal Candidate Vetting
- Unveiling Accretion-Disk Reverberation Beneath BLR Contamination with TILDE
- Unveiling Accretion-disk Reverberation beneath Broad-line Region Contamination with TILDE
BibTeX
@misc{tenachi2026tabular,
title = {{Do Tabular Foundation Models Know Physics? Contamination, Units, and the Deterministic Limit}},
author = {Wassim Tenachi and Yashar Hezaveh and Laurence Perreault Levasseur and Pierre-Luc Bacon},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2609.02766},
eprint = {2609.02766},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.02766},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Do Tabular Foundation Models Know Physics? Contamination, Units, and the Deterministic Limit
AU - Wassim Tenachi
AU - Yashar Hezaveh
AU - Laurence Perreault Levasseur
AU - Pierre-Luc Bacon
PY - 2026
DA - 2026/10/22
PB - arXiv
DO - 10.48550/arxiv.2609.02766
UR - https://arxiv.org/abs/2609.02766
AB - Tabular foundation models (TFMs) learn to fill in tables the way language models fill in text --- and tables are arguably the format in which most physical measurement arrives. Did they learn any physics in the process? They are Bayesian by construction, so the question is what their prior contains. We probe it directly, evaluating four of them (TabPFN−3, TabICLv2, TabDPT and Real−TabPFN−2.5) against six baselines on datasets sampled from 316 physical equations, in and out of domain. TFMs dominate, out of the box and after tuning. But 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.
ER -