跳转至

文章背景与核心概要

生成结构有效且具备高保真度的合成表格数据一直是人工智能和数据科学领域的重大挑战。尽管当前的模型能够实现较高的统计保真度和下游效用，但其生成的数据往往会违反具有语义含义的领域约束。本文介绍了一种统一的、基于工具锚定的工作流，利用大语言模型（LLM）智能体来发现并强制执行三类互补的列间约束：方程式、线性不等式以及逻辑依赖关系。

该框架将这些约束表示为可机执行的假设，并应用统一的接口进行全表验证、确定性诊断以及反例引导的修正。一个与生成器无关的后处理器随后对来自标准表格生成器的输出进行特定类别的修复协调。经过精心策划的行为审计和端到端评估表明，该工作流显著提高了留出违规检测能力，后处理器在保留且适用的约束上实现了零测得违规，同时在绝大多数数据集上提升了下游效用并较好地保留了单变量边缘分布。

Constraint-Aware Synthetic Tabular Data Generation via Inter-Column Constraint Discovery with LLM Agents

Metadata

  • arXiv ID: 2608.15109
  • Primary Subject: Computer Science > Artificial Intelligence (cs.AI)
  • Authors: Jianxing Zhao, Mao Guan, Dongyu Liu
  • Submitted: August 15, 2026
  • DOI: 10.48550/arXiv.2608.15109

Metadata

  • arXiv ID: 2608.15109
  • Primary Subject: Computer Science > Artificial Intelligence (cs.AI)
  • Authors: Jianxing Zhao, Mao Guan, Dongyu Liu
  • Submitted: August 15, 2026
  • DOI: 10.48550/arXiv.2608.15109

Executive Summary

生成结构有效的合成表格数据仍然是一项重大挑战。虽然当前的模型可以实现很高的统计保真度和下游效用，但其输出往往违反具有语义意义的领域约束。

本文引入了一种统一的、基于工具锚定的工作流，利用大语言模型（LLM）智能体来发现并强制执行三类互补的列间约束： 1. 方程式 (Equations) 2. 线性不等式 (Linear Inequalities) 3. 逻辑依赖关系 (Logical Dependencies)

通过将这些约束表示为可机执行的假设，该框架应用了一个共享接口来进行全表验证、确定性诊断以及反例引导的修正。一个与生成器无关的后处理器随后协调对标准表格生成器输出的特定系列修复。

核心发现： - 与单次直接提示（one-shot direct prompting）相比，该工作流改善了留出违规检测。 - 后处理器对每个保留且适用的约束实现了零测得违规（zero measured violations）。 - 在大多数数据集上提高了下游效用，同时成功保留了单变量边缘分布。

Executive Summary

Generating structurally valid synthetic tabular data remains a significant challenge. While current models can achieve high statistical fidelity and downstream utility, their outputs frequently violate semantically meaningful domain constraints.

This paper introduces a unified, tool-grounded workflow utilizing Large Language Model (LLM) agents to discover and enforce three complementary families of inter-column constraints: 1. Equations 2. Linear Inequalities 3. Logical Dependencies

By representing these constraints as machine-executable hypotheses, the framework applies a shared interface for full-table validation, deterministic diagnosis, and counterexample-guided revision. A generator-agnostic postprocessor then coordinates family-specific repairs on outputs from standard tabular generators.

Key Findings: - The workflow improves held-out violation detection compared to one-shot direct prompting. - Postprocessing achieves zero measured violations for every retained, applicable constraint. - Downstream utility is improved on most datasets while successfully preserving univariate marginals.

Abstract

生成结构有效的合成表格数据仍然很困难：具有高统计保真度和下游效用的输出仍然可能违反具有语义意义的领域约束。我们研究了三类互补的列间约束家族（方程式、线性不等式和逻辑依赖关系）的发现与强制执行。我们统一的、基于工具锚定的工作流将这三者都表示为可机执行的假设，并应用通用接口进行全表验证、确定性诊断和反例引导的修正。一个与生成器无关的后处理器协调对未修改的表格生成器输出的家族特定修复。在策划的行为审计和端到端评估中，完整的工作流比单次直接提示改善了留出违规检测，而后处理器为每个保留的、适用的约束产生了零测得违规，在大多数数据集上提高了下游效用，并且在很大程度上保留了单变量边缘分布。

Abstract

Generating structurally valid synthetic tabular data remains difficult: outputs with high statistical fidelity and downstream utility can still violate semantically meaningful domain constraints. We study the discovery and enforcement of three complementary inter-column constraint families---equations, linear inequalities, and logical dependencies. Our unified tool-grounded workflow represents all three as machine-executable hypotheses and applies a common interface for full-table validation, deterministic diagnosis, and counterexample-guided revision. A generator-agnostic postprocessor coordinates family-specific repairs on outputs from unchanged tabular generators. Across curated behavioral audits and end-to-end evaluations, the complete workflow improves held-out violation detection over one-shot direct prompting, while postprocessing yields zero measured violations for every retained, applicable constraint, improves downstream utility on most datasets, and largely preserves univariate marginals.