MetaPersona paper links synthetic populations to empirical social-science studies

A research paper introduces MetaPersona, a framework for creating synthetic populations for language-model social simulations, and MetaPersona-DB, a database of empirical human-subjects studies. The authors describe a cold-start problem in choosing personas for these simulations: existing approaches, they say, do not provide a principled way to select attributes or assign their values. The paper connects those choices to whether synthetic populations reflect demographic composition, less directly observed attributes, and relationships that can shape later behavior.

MetaPersona-DB contains more than 11,000 studies annotated with variables relevant to tasks, reported relationships, and population-level statistics. The framework retrieves evidence related to a task, uses it to form graphs of dependencies among persona attributes, and samples populations from empirical patterns connecting demographics, latent attributes, and outcomes. In other words, the proposed process uses both task-relevant findings and reported links between attributes when constructing a population, rather than treating each persona attribute as unrelated.

The authors assessed the framework in three case studies, with three baselines and three frontier models. They report that the results changed depending on the task and model. MetaPersona performed strongly in evaluations involving belief in misinformation and sentiment toward AI tools, while findings for income redistribution were mixed. The paper therefore reports different outcomes across the tested questions, rather than one result applying uniformly across all three cases.

The authors also report persona-construction costs below $0.5 per task when using GPT-5.2. They present MetaPersona-Studio, described as a prototype interactive interface for generating empirically grounded personas. For AI builders, the work lays out a way to connect simulated-population design to research evidence, while its reported evaluation results vary across tasks and models. The abstract describes the interface as a prototype and gives the cost figure for GPT-5.2; it does not establish broader availability or costs with other models.