TabFM: A zero-shot foundation model for tabular data

Google Research has introduced TabFM, a zero-shot foundation model for tabular data that eliminates the need for manual hyperparameter tuning and feature engineering in classification and regression tasks. Unlike traditional tree-based algorithms like XGBoost that require extensive preparation, TabFM frames tabular prediction as an in-context learning problem, allowing it to generate predictions on new datasets in a single forward pass by processing entire tables as unified prompts. The model uses alternating row and column attention mechanisms to capture feature interactions and dependencies directly from the data structure.