LLM Agents Discover Inter-Column Constraints for Synthetic Tabular Data

A recent paper on arXiv (2608.15109) presents a novel approach for creating structurally sound synthetic tabular data by identifying and enforcing inter-column constraints through LLM agents. This research tackles a longstanding issue: synthetic data can maintain high statistical accuracy and utility but still breach meaningful domain constraints. The system delineates three types of complementary constraints—equations, linear inequalities, and logical dependencies—expressed as machine-executable hypotheses. It employs a unified tool-grounded workflow for comprehensive table validation, deterministic diagnosis, and counterexample-guided revisions. A generator-agnostic postprocessor manages specific repairs on outputs from unchanged tabular generators. Evaluations demonstrate that this workflow enhances violation detection and ensures zero violations for applicable constraints, while also boosting downstream utility. The paper was recently submitted to arXiv.