LLMSynthor: Macro-Aligned Micro-Records Synthesis with Large Language Models

Macro-aligned micro-records are essential for simulations in social science and urban studies. For instance, epidemic models of urban disease spread are only credible when micro-level records reproduce realistic individual mobility and contact patterns, while macro-level aggregates match macro-statistics such as case counts or travel flows. Still, collecting large-scale fine-grained data is often impractical, leaving researchers with only macro-statistics. While Large Language Models (LLMs) can generate realistic micro-records using rich real-world priors learned from vast corpora, naive record-by-record sampling is inefficient and fails to enforce alignment with target macro-statistics. Given this, we propose LLMSynthor, a framework that transforms a pre-trained LLM into a macro-aware simulator capable of synthesizing realistic micro-records aligned with given macro-statistics. LLMSynthor incrementally constructs a synthetic dataset by iteratively generating batches of micro-records that reduce discrepancies between synthetic and target macro-statistics. By treating the LLM as a nonparametric copula over joint variable dependencies, the framework ensures alignment with target marginals and joint distributions. To improve efficiency, we introduce LLM Proposal Sampling, where the LLM generates discrepancy-guided proposals specifying variable constraints and record counts. Evaluations on synthetic and real-world datasets (mobility, e-commerce, population) encompassing diverse formats and settings show that LLMSynthor achieves high record realism, statistical fidelity, and practical utility, positioning it broadly applicable across economics, social science, urban studies, and beyond.