Team, Then Trim: An Assembly-Line LLM Framework for High-Quality Tabular Data Generation

TL;DR

This paper tackles the challenge of generating high-quality tabular data when real data are scarce, imbalanced, incomplete, or noisy. It introduces Team-then-Trim (T$^2$), an assembly-line framework that uses a task-managing LLM to decompose the data-generation problem into semantically coherent components and assign them to specialized worker LLMs, producing raw data batches $D_t$. A modular three-stage QC pipeline—sanity checks, objective-related cost assessment, and diversity inspection—filters batches and yields $D_{\text{kept}}$, which augments the original dataset $D_{\text{ori}}$ to form $D_{\text{new}}$ and improve downstream predictive performance. Across simulated and real-world datasets, T$^2$ consistently outperforms baselines in downstream utility, data fidelity, and diversity, while the QC stages bolster robustness under imbalance, incompleteness, and noise. The work demonstrates that structured LLM teaming, combined with rigorous batch-level QC, can effectively bridge data gaps and enable reliable learning in challenging tabular domains.

Abstract

While tabular data is fundamental to many real-world machine learning (ML) applications, acquiring high-quality tabular data is usually labor-intensive and expensive. Limited by the scarcity of observations, tabular datasets often exhibit critical deficiencies, such as class imbalance, selection bias, and low fidelity. To address these challenges, building on recent advances in Large Language Models (LLMs), this paper introduces Team-then-Trim (T$^2$), a framework that synthesizes high-quality tabular data through a collaborative team of LLMs, followed by a rigorous three-stage plug-in data quality control (QC) pipeline. In T$^2$, tabular data generation is conceptualized as a manufacturing process: specialized LLMs, guided by domain knowledge, are tasked with generating different data components sequentially, and the resulting products, i.e., the synthetic data, are systematically evaluated across multiple dimensions of QC. Empirical results on both simulated and real-world datasets demonstrate that T$^2$ outperforms state-of-the-art methods in producing high-quality tabular data, highlighting its potential to support downstream models when direct data collection is practically infeasible.

Figures (16)

• Figure 1: An illustration of our T$^2$ framework. LLM teaming is for structured data generation and trimming is for rigorous three-stage QC.
• Figure 2: An example of task coordination.
• Figure 3: Average AUC vs. $|D_\text{ori}|$ under flip ratios of 0.2, 0.3, and 0.4 on the Diabetes dataset.
• Figure 4: Average AUC for two demographic groups on the Drug dataset.
• Figure 5: Average AUC of LLM backbones under data imbalance.