One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction
Some generalist tabular models beat specialist portfolios on utility, but specialists retain the edge in distributional fidelity.
The One Table Benchmark covers 56 datasets, 46 registered methods, and 20 capability axes for generation, imputation, and prediction. Training experiments found the largest pooled gains came from including the objective for the task being evaluated. Auxiliary objectives sometimes helped, but their benefits depended on the model and objectives already present. Frozen foundation models performed strongly on prediction and numerical imputation, while categorical reconstruction and faithful joint sampling remained areas for improvement.
HF Daily Papers' note
The One Table Benchmark covers 56 datasets, 46 registered methods, and 20 capability axes for generation, imputation, and prediction. Training experiments found the largest pooled gains came from including the objective for the task being evaluated. Auxiliary objectives sometimes helped, but their benefits depended on the model and objectives already present. Frozen foundation models performed strongly on prediction and numerical imputation, while categorical reconstruction and faithful joint sampling remained areas for improvement.
HF Daily Papers' note
score 4