Yuanrui Wang, Kaifang Li, Xuanyue Li, Ziyun Li, Yan Lu, Xingxuan Zhang, Peng Cui (Stable AI) · Tabular Benchmark

One Table Benchmark

The authors introduce the One Table Benchmark, spanning 56 datasets, 46 registered methods and 20 capability axes, to evaluate tabular models across generation, imputation and prediction under aligned splits. They train five backbones under seven objective combinations and compare generalist models with specialist portfolios. Findings include that some generalists beat specialists on utility while fidelity favours specialisation, and that frozen foundation models perform strongly on prediction and numerical imputation.