What it is
A 400M tabular foundation model does zero-shot supervised prediction as in-context learning, trained only on synthetic SCM tables and run in one forward pass. Instead of fitting trees or AutoML per dataset, it treats tabular prediction as in-context inference: a synthetic-data-trained transformer makes calibrated zero-shot predictions on real tables, with frozen-weight upgrades via ensembling/calibration or LLM-guided feature engineering.
Key results
- Overall TabArena: 1785 Elo, above AutoGluon 1.5 extreme at 1676
- Classification: 1768.6 Elo vs TabPFN-3 1641.9 and AutoGluon 1669.7
- Regression: 2055.2 Elo vs EXAONE-Tabular 1973.1 and TabPFN-3 1866.6
Shows that large tabular models trained purely on synthetic causal tables can transfer zero-shot to real-world datasets at or above tuned AutoML.