TabFM: Zero-Shot Table Predictions That Beat XGBoost

For fifteen years the answer to "I have a table and a target column" has been the same: reach for gradient-boosted trees, then spend two days on hyperparameter search and feature crosses to squeeze out the last two points of AUC. Google Research just shipped something that skips all of it.

TabFM is a foundation model for tabular data. You hand it your training rows and your test rows as a single context, it does one forward pass over frozen weights, and out come calibrated probabilities. No fit loop that updates parameters. No learning rate. No max_depth. On TabArena — 51 datasets, 38 classification and 13 regression — the untuned model beats heavily-tuned supervised baselines including gradient-boosted trees.