A Survey on Tabular Data Synthesis: Generation, Evaluation, and Benchmarking

Tabular data synthesis has attracted growing attention as a means of addressing data scarcity, supporting data sharing, and augmenting datasets for analysis. Recent advances in generative modeling motivate an updated review of how emerging synthesis methods relate to established statistical approaches and how their outputs should be evaluated. This review brings together classical statistical methods, deep generative models, and recent approaches involving large language models (LLMs). We examine their generation mechanisms, modeling assumptions, and limitations, and discuss evaluation from three complementary perspectives: distributional similarity, downstream utility, and privacy. An illustrative benchmark of twelve generator configurations on six datasets complements the review by examining dataset‐ and metric‐dependent behavior under the reported experimental protocol. The comparison highlights the need to interpret similarity and utility scores alongside privacy‐related diagnostics and computational requirements. We also discuss challenges involving mixed data types, rare observations, reproducibility, and responsible deployment. The review identifies open questions about how synthetic data should be generated and validated for particular uses, including when LLMs are useful as direct generators or auxiliary components.