TabFM-Auto: Self-Evolving Pipelines for Tabular Foundation Models
The paper’s claim is that an LLM agent can improve a frozen tabular foundation model by evolving the surrounding data pipeline.
TabFM-Auto uses dataset metadata and validation feedback to refine cleaning, feature engineering, context selection, and post-processing around TabFM. On all 51 TabArena datasets, its five tested configurations take the top five overall spots, with the best moving TabFM from 1785 to 2013 Elo. The pipelines it finds also transfer to other frozen tabular foundation models without another search. On MLE-Bench’s eight tabular competitions, it ranks first among MLE agents. HF Daily Papers' note
TabFM-Auto uses dataset metadata and validation feedback to refine cleaning, feature engineering, context selection, and post-processing around TabFM. On all 51 TabArena datasets, its five tested configurations take the top five overall spots, with the best moving TabFM from 1785 to 2013 Elo. The pipelines it finds also transfer to other frozen tabular foundation models without another search. On MLE-Bench’s eight tabular competitions, it ranks first among MLE agents. HF Daily Papers' note
score 5