Evidence brief: A large language model framework for sample-free population synthesis

Abstract

Synthetic populations provide the demographic foundations for agent-based models in transport, public health, disaster management and other sectors, enabling credible representations of individual characteristics and behaviours. Many established synthesis methods rely on census microdata; however, such data are infrequently collected, privacy-restricted, and usually available only as small public-use samples at coarse geographic scales. This paper introduces a sample-free framework that uses a large language model (LLM) to generate complete, household-structured populations directly from aggregate demographic data. The framework is LLM agnostic and follows a multi-step process: objective definition, input preparation, LLM selection, and synthetic household generation. No model fine-tuning is required, meaning that data requirements are low and the framework is easily accessible. Population synthesis is formulated as an iterative prompting process in which an LLM generates households guided by the discrepancies between synthetic and target distributions. The model draws on prior knowledge encoded during pre-training to propose plausible attribute combinations, resulting in both statistical alignment and structural feasibility. In a global evaluation covering 109 countries, the framework achieved very close alignment on simpler marginals such as gender (SRMSE: 0.003) and household size (SRMSE: 0.026), while more structurally complex attributes such as household composition (SRMSE: 0.062) and age (SRMSE: 0.128) were also reproduced with good accuracy. These results were supported by detailed case studies in Newcastle upon Tyne (UK) and Dar es Salaam (Tanzania). The principal contribution of the framework is to enable the construction of coherent household-structured populations when detailed microdata are unavailable, expanding the applicability of agent-based modelling in data-constrained settings.

1 Introduction

Urban simulations are frequently founded on agent-based models (ABMs), which represent people and households as individual agents and trace how their behaviours and interactions generate system-level outcomes. To initialise these models, researchers must first construct a population: a statistically consistent, digital representation of people, their demographic attributes, and household groupings. Synthetic populations are particularly valuable where detailed person-level records (microdata) are unavailable, incomplete, or protected for privacy reasons. They provide the demographic foundations that allow ABMs to capture heterogeneity – for instance, differences in age, mobility, or caring responsibilities – and to explore how these differences shape collective outcomes [1]. This approach has been used in many domains, including urban mobility [2–4], public health [5–9] and disaster management [10]. Across these applications, the credibility of the results depends directly on how well the synthetic population reflects the target environment.

Two primary types of data are used to construct synthetic populations: aggregate data, which provide marginal or joint distributions of demographic variables, and microdata, which contain individual-level records [11]. National censuses are the most comprehensive provider of both, but they are collected infrequently, usually on a decadal cycle, as recommended by the UN [12]. Moreover, access to person-level records is often restricted for privacy protection: only small public-use samples of 1–5% of the population are usually available, and these are restricted to coarse geographic units [13]. Where available, auxiliary surveys such as Demographic Health Surveys or Multiple Indicator Cluster Surveys can supply aggregate data, yet these sources differ in coverage, definitions and spatial resolution [14,15]. The resulting inconsistencies make it difficult to reconcile information across datasets and scales. Consequently, the central challenge is to construct household-consistent synthetic populations when the underlying data is incomplete, contradictory, or missing at fine geographic levels.