Tabular foundation models interpolate physics but cannot model it
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Tabular foundation models are outperforming traditional baselines on physics-informed datasets, but a new evaluation reveals a critical limitation: they cannot represent deterministic systems or physical units. Researchers tested four leading TFMs against 316 physics equations and found that while these models excel at interpolation within their training distribution, they lack the structural priors needed to function as genuine physics simulators. This gap matters for scientific computing and industrial applications where unit consistency and noiseless mechanisms are non-negotiable, suggesting that scaling tabular models alone won't bridge the gap between statistical pattern matching and causal physical reasoning.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The critical finding isn't just that tabular foundation models fail on physics tasks, but that they fail specifically on deterministic (noiseless) systems where there is only one correct answer. This reveals the models are learning statistical associations rather than structural rules, a distinction that benchmarks measuring noisy-data performance would miss entirely.
This connects directly to the SCILAWS-BENCH work from yesterday, which flagged that existing evaluations can't distinguish genuine discovery from memorization. Here we see the same problem in a different domain: tabular models appear competent on physics datasets because those datasets contain noise and variation that reward pattern matching. The moment you remove the noise and ask for deterministic prediction, the illusion collapses. Both papers expose how benchmark design can hide fundamental reasoning gaps.
If TabPFN-3 or TabICLv2 are retrained with explicit unit tokens or symbolic constraint layers and recover performance on the deterministic subset, that confirms the gap is architectural rather than fundamental. If they remain stuck below 50% accuracy on noiseless equations even with these modifications, it signals these models may need hybrid symbolic-neural designs rather than pure scaling to handle physics applications.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
New benchmark tests whether LLMs can truly discover scientific laws
Researchers have built SCILAWS-BENCH, a rigorous evaluation framework that tests whether large language models can genuinely discover scientific laws rather than merely memorize published results. The benchmark draws from 381 real papers and 8M data points across 118 problems, addressing a critical gap in how AI-for-science capabilities are measured. This work matters because existing evaluations…
MentionsTabPFN-3 · TabICLv2 · TabDPT · Real-TabPFN-2.5
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Do Tabular Foundation Models Know Physics? Contamination, Units, and the Deterministic Limit”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.