Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yihong Tang, Menglin Kong, Junlin He, Tong Nie, Lijun Sun

TL;DR

LLMSynthor bridges the micro-macro data gap by repurposing pretrained LLMs as macro-aware simulators that generate realistic micro-records aligned to target macro-statistics. It introduces a nonparametric copula interpretation of the LLM to model joint dependencies and a discrepancy-guided LLM Proposal Sampling mechanism to efficiently produce targeted batch proposals. Through an iterative loop, the framework updates macro-statistics, attributes discrepancies, and refines data via proposals, with a convergence guarantee in expectation. Empirical results across mobility, e-commerce, and population synthesis show that LLMSynthor achieves superior statistical fidelity and practical utility, often surpassing baselines trained on full micro-data while handling unstructured data without task-specific engineering. This approach offers a scalable, versatile path for synthetic data generation in economics, social science, and urban studies, enabling scenario analysis and policy-relevant research with privacy-preserving data.

Abstract

Macro-aligned micro-records are crucial for credible simulations in social science and urban studies. For example, epidemic models are only reliable when individual-level mobility and contacts mirror real behavior, while aggregates match real-world statistics like case counts or travel flows. However, collecting such fine-grained data at scale is impractical, leaving researchers with only macro-level data. LLMSynthor addresses this by turning a pretrained LLM into a macro-aware simulator that generates realistic micro-records consistent with target macro-statistics. It iteratively builds synthetic datasets: in each step, the LLM generates batches of records to minimize discrepancies between synthetic and target aggregates. Treating the LLM as a nonparametric copula allows the model to capture realistic joint dependencies among variables. To improve efficiency, LLM Proposal Sampling guides the LLM to propose targeted record batches, specifying variable ranges and counts, to efficiently correct discrepancies while preserving realism grounded in the model's priors. Evaluations across domains (mobility, e-commerce, population) show that LLMSynthor achieves strong realism, statistical fidelity, and practical utility, making it broadly applicable to economics, social science, and urban studies.

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

TL;DR

LLMSynthor bridges the micro-macro data gap by repurposing pretrained LLMs as macro-aware simulators that generate realistic micro-records aligned to target macro-statistics. It introduces a nonparametric copula interpretation of the LLM to model joint dependencies and a discrepancy-guided LLM Proposal Sampling mechanism to efficiently produce targeted batch proposals. Through an iterative loop, the framework updates macro-statistics, attributes discrepancies, and refines data via proposals, with a convergence guarantee in expectation. Empirical results across mobility, e-commerce, and population synthesis show that LLMSynthor achieves superior statistical fidelity and practical utility, often surpassing baselines trained on full micro-data while handling unstructured data without task-specific engineering. This approach offers a scalable, versatile path for synthetic data generation in economics, social science, and urban studies, enabling scenario analysis and policy-relevant research with privacy-preserving data.

Abstract

Macro-aligned micro-records are crucial for credible simulations in social science and urban studies. For example, epidemic models are only reliable when individual-level mobility and contacts mirror real behavior, while aggregates match real-world statistics like case counts or travel flows. However, collecting such fine-grained data at scale is impractical, leaving researchers with only macro-level data. LLMSynthor addresses this by turning a pretrained LLM into a macro-aware simulator that generates realistic micro-records consistent with target macro-statistics. It iteratively builds synthetic datasets: in each step, the LLM generates batches of records to minimize discrepancies between synthetic and target aggregates. Treating the LLM as a nonparametric copula allows the model to capture realistic joint dependencies among variables. To improve efficiency, LLM Proposal Sampling guides the LLM to propose targeted record batches, specifying variable ranges and counts, to efficiently correct discrepancies while preserving realism grounded in the model's priors. Evaluations across domains (mobility, e-commerce, population) show that LLMSynthor achieves strong realism, statistical fidelity, and practical utility, making it broadly applicable to economics, social science, and urban studies.

Paper Structure

This paper contains 86 sections, 1 theorem, 25 equations, 20 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Abstract
  2. Introduction
  3. Related Work
  4. Data Synthesis
  5. LLMs as Data Generators
  6. Methodology
  7. Overview
  8. Variable Dependency Inference
  9. Discrepancy-Guided Iterative Synthesis
  10. 1. Updating Synthetic Macro-statistics.
  11. 2. Discrepancy Attribution.
  12. 3. Discrepancy-Guided LLM Proposal Sampling.
  13. Iterative Synthesis.
  14. Experiments
  15. Mobility Synthesis
  16. ...and 71 more sections

Key Result

Theorem 1

Let $\Delta^{(t)}_{\star} \equiv Q_{\star}\!\bigl(\mathcal{T}^{(t)}_{\mathrm{synth}}, \mathcal{T}^{\mathcal{C}}_{\mathrm{target}}\bigr)$ with $Q_{\star}(x,y)=\varphi(x-y)$ for any norm $\varphi$ (positively homogeneous). Define the expected discrepancy: This ensures that the synthetic macro-statistics converge to the target macro-statistics as the number of iterations increases. The proof is deta

Figures (20)

  • Figure 1: A comparison of existing generative paradigms and LlmSynthor.
  • Figure 2: Overview of LlmSynthor. For illustration, the figure uses a mobility dataset as a running example, but the approach is general and applicable to diverse domains and data contexts.
  • Figure 3: Real vs. synthetic mobility patterns.
  • Figure 4: Qualitative Distributions and Comparisons.
  • Figure 5: Bayesian network representing the generative process of e-commerce transactions.
  • ...and 15 more figures

Theorems & Definitions (4)

  • Definition 1: Micro-record
  • Definition 2: Macro-statistics
  • Theorem 1: Contraction of the Mean Macro-Discrepancy
  • proof : Proof of Theorem \ref{['thm:mean_contraction']}