TabFM: A Zero-Shot Foundation Model for Tabular Data
TabFM claims tuned-AutoML-level tabular prediction without training on the target dataset.
The paper presents a 400M-parameter model that treats supervised tabular tasks as in-context learning. It is trained on synthetic tables from structural causal models, then tested zero-shot on real-world benchmarks. On TabArena’s 51 datasets, the authors say it ranks first among default tabular foundation models and beats tuned AutoML pipelines. They also report stronger results from frozen-weight extensions using feature expansion, calibration, and LLM-guided data processing. HF Daily Papers' note
The paper presents a 400M-parameter model that treats supervised tabular tasks as in-context learning. It is trained on synthetic tables from structural causal models, then tested zero-shot on real-world benchmarks. On TabArena’s 51 datasets, the authors say it ranks first among default tabular foundation models and beats tuned AutoML pipelines. They also report stronger results from frozen-weight extensions using feature expansion, calibration, and LLM-guided data processing. HF Daily Papers' note
score 6