TabFM: A Zero-Shot Foundation Model for Tabular Data
Weihao Kong, Erez Louidor Ilan, Shuxin Nie, Taman Narayan, Rajat Sen, Yichen Zhou, Deqing Fu, Samet Oymak and 1 others
Google Research
Published September 29, 2026 · on the feed September 30, 2026 · ▲ 19 · arXiv 2609.37959
The abstract
Tabular machine learning typically relies on per-dataset workflows, fitting tree ensembles or running AutoML searches from scratch for every task. We present TabFM, a 400M-parameter tabular foundation model that formulates supervised tabular prediction as in-context learning. TabFM produces calibrated zero-shot predictions in a single forward pass without task-specific tuning. Trained entirely on synthetic tables generated from structural causal models, TabFM learns general tabular representations that transfer zero-shot to real-world tasks. Across all 51 benchmark datasets in TabArena (38 classification and 13 regression), zero-shot TabFM ranks first among default tabular foundation models and outperforms tuned AutoML pipelines. Two extensions over the same frozen weights improve performance further on both tracks: multi-view feature expansion with ensembling and post-hoc calibration (TabFM+), and LLM-guided, dataset-specific data processing and feature engineering (TabFM-Auto).
The Brief
1:42A short podcast summary of the paper. Written and voiced by AI.
Transcript
TabFM. A Zero-Shot Foundation Model for Tabular Data. By Weihao Kong and eight others, from Google Research. Published on the twenty-ninth of September. Its code repository is public and real: about twenty-five source files.
Machine learning on tables is usually done one dataset at a time. For every new task, someone fits tree ensembles or runs an AutoML search from scratch. Nothing carries over from the last table to the next.
TabFM is a tabular foundation model of four hundred million parameters that formulates supervised prediction on a table as in-context learning. It produces its predictions in a single forward pass. They are calibrated, they are zero-shot, and no task-specific tuning is involved at any point.
Training uses no real tables at all. TabFM is trained entirely on synthetic tables generated from structural causal models, and from them it learns general tabular representations that transfer, zero-shot, to real-world tasks. Two extensions sit on top of the same frozen weights. TabFM+ adds multi-view feature expansion with ensembling, plus post-hoc calibration. TabFM-Auto adds data processing and feature engineering specific to each dataset, guided by a language model.
The test is all fifty-one benchmark datasets in TabArena: thirty-eight for classification and thirteen for regression. Zero-shot TabFM ranks first among tabular foundation models run at their defaults, and it outperforms AutoML pipelines that have been tuned. Both extensions improve performance further, on both tracks.
One forward pass from a model that has never seen a real table, set against a search run afresh for every dataset, and the forward pass comes out ahead.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every table, the same chore from scratch. Is there a way to stop starting over?
You are a data scientist with a new table to predict from. As always, you fit tree ensembles or run an AutoML search from scratch, with custom preprocessing and a hyperparameter sweep.
The road ahead
- 1 · The problem ~7 min Every new table gets its own model built from scratch, so nothing learned on one table helps with the next.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Feed a table's labeled rows to one pretrained model as context and it predicts the test rows in a single forward pass.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Generate causal synthetic tables, pretrain once, then predict a real table from its labeled rows in one pass, optionally ensembling views.
- 6 · The experiment ~6 min You’ll guess: With no task-specific tuning at all, how did zero-shot TabFM compare with tuned AutoML pipelines on TabArena?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.