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

Abstract:

Generating macro-aligned micro-records is essential for producing credible simulations within the fields of urban studies and social science. For instance, epidemic models require individual-level data on mobility and social contacts that reflect actual human behavior, while ensuring that aggregated outputs align with real-world metrics such as travel flows or case counts. However, gathering such granular data on a large scale is often unfeasible, forcing researchers to rely predominantly on macro-level information.

To bridge this gap, LLMSynthor repurposes a pretrained Large Language Model (LLM) into a macro-aware simulator capable of producing realistic micro-records that adhere to specific macro-statistical targets. The system constructs synthetic datasets through an iterative process: in each cycle, the LLM generates batches of records designed to reduce the divergence between synthetic aggregates and the desired target statistics. By conceptualizing the LLM as a nonparametric copula, the approach effectively captures complex, realistic joint dependencies among various variables.

To enhance computational efficiency, LLMSynthor employs LLM Proposal Sampling. This technique directs the LLM to generate targeted batches of records—defining specific variable ranges and counts—allowing for the efficient correction of statistical discrepancies without compromising the realism inherent in the model’s priors. Evaluations spanning diverse domains, including population dynamics, e-commerce, and mobility, demonstrate that LLMSynthor delivers high levels of statistical fidelity and realism. These results underscore its broad utility for applications in economics, social science, and urban studies.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC