One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction

One Table, Many Queries. Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction. By Yuanrui Wang and six others, from StableAI, Tsinghua University, and two other institutions. Published on the eighth of November.

Three jobs are routinely asked of a table. Generate new rows, fill in missing cells, predict a target column. Each one queries a different aspect of the same underlying distribution, and yet the three are usually evaluated separately. Separate evaluation leaves several questions open. Which capabilities do the tasks share? How do their training objectives interact when a model learns more than one? And can a single model perform all three well, or does each still want its own specialist?

The One Table Benchmark is built to ask those questions side by side. It spans fifty-six datasets, forty-six registered methods, and twenty capability axes. Every comparison runs under aligned data splits and explicit task contracts, so a generator, an imputer and a predictor are being judged on the same footing rather than on whatever setup each literature happened to adopt.

Three kinds of study sit on top of that. One compares task ranks on aligned support, asking whether a method that ranks well at one job ranks well at another. A second is an intervention study: five backbones are trained under all seven non-empty combinations of the three objectives, on five datasets, with three seeds each. The third contrasts the profile of one fixed model against portfolios of specialists.

The rank comparisons tie imputation and prediction together through numerical recovery and dependence preservation, and link generation to both through predictive utility. Distributional fidelity is the outlier. It follows different rankings, sometimes opposing ones.

On objectives, the largest pooled gains come from including the recipient's own objective. Generation has the broadest coverage of positive pooled auxiliary effects. Prediction shows potential to improve imputation and synthetic-data utility, though the response varies across backbones. Adding objectives together is not simple arithmetic either: the cross-addition effects show that benefits need not add, and depend on which objectives are already present.

The fixed-model comparisons expose a matching trade-off between utility and fidelity. Some generalists surpass specialist portfolios on utility, while fidelity continues to favour specialization. Frozen foundation models show strong prediction and strong numerical point imputation, with promising synthetic-data utility, which the authors read as potential across all three tasks.

Put the results next to each other and a direction emerges. A model that shares useful conditional information across the three queries already gets some of the way. What remains, on this evidence, is categorical reconstruction and faithful joint sampling, and those are the two places the authors say a unified tabular model still has to improve.