Generalist tabular models match specialists on utility, not fidelity

A unified benchmark across generation, imputation, and prediction shows that multi-objective models can match or beat specialized models on downstream usefulness, but specialized models still produce higher-fidelity synthetic data.

Top University

Yuanrui Wang · Kaifang Li · Xuanyue Li · Ziyun Li · Yan Lu · Xingxuan Zhang · +1 more

StableAI · Tsinghua University · Beijing University of Posts and Telecommunications · Southeast University

Research Digest·11 November 2026·2 min read

Illustration: Generalist tabular models match specialists on utility, not fidelity

In one line

Foundation models show strong performance across generation, imputation, and prediction on tabular data.

Before you read this

Two ideas this paper builds on

benchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.338 papers →synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.80 papers →

Read it as

Tables of data often need filling in missing values, generating new records, or predicting outcomes. These tasks are usually studied separately, so it's unclear if one model can handle all three.

The authors built a OTBench to test models on all three tasks using many datasets and metrics. They compared model rankings and trained models with different task combinations.

They found that large pre-trained models, called foundation models, perform well on all three tasks without extra training for each. This suggests that a single generalist model can work for multiple table uses, though trade-offs exist between usefulness and data realism.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

benchmark multi-task learning conditional distribution modeling synthetic data generation tabular foundation models