A large language model framework for sample-free population synthesis

Area of Science:

• Computational Social Science
• Demographic Modeling
• Artificial Intelligence Applications

Background:

• Agent-based models (ABMs) require realistic synthetic populations for accurate simulations in diverse fields.
• Existing population synthesis methods often depend on scarce, privacy-restricted, or coarse-scale census microdata.
• This limitation restricts the application of ABMs in data-scarce environments.

Purpose of the Study:

• To present a sample-free framework for generating complete, household-structured synthetic populations using large language models (LLMs).
• To enable the creation of detailed demographic representations directly from aggregate data, overcoming microdata limitations.
• To expand the utility of ABMs in research settings with constrained data availability.

Main Methods:

• A multi-step, LLM-agnostic framework involving objective definition, input preparation, LLM selection, and synthetic household generation.
• Population synthesis is achieved through iterative prompting, where the LLM generates households guided by discrepancies between synthetic and target distributions.
• The method leverages the LLM's pre-trained knowledge for plausible attribute combinations, ensuring statistical alignment and structural feasibility without model fine-tuning.

Main Results:

• Global evaluation across 109 countries demonstrated high accuracy in reproducing marginal distributions like gender (SRMSE: 0.003) and household size (SRMSE: 0.026).
• Complex attributes such as household composition (SRMSE: 0.062) and age (SRMSE: 0.128) were also accurately reproduced.
• Case studies in Newcastle upon Tyne (UK) and Dar es Salaam (Tanzania) validated the framework's effectiveness.

Conclusions:

• The developed framework successfully generates coherent, household-structured synthetic populations from aggregate data, eliminating the need for microdata.
• This approach significantly enhances the applicability of agent-based modeling in diverse research areas facing data limitations.
• The LLM-based method offers an accessible and low-data-requirement solution for constructing foundational demographic datasets for simulations.