MetaPersona Builds Task-Grounded Synthetic Populations from Social Science Evidence

Summary

Persona-based LLM social simulations face a cold-start problem: existing methods provide no principled way to choose relevant attributes or assign their values, which can distort demographic composition and relationships among latent traits and outcomes. The authors introduce MetaPersona-DB, a dataset of more than 11,000 empirical human-subject studies annotated with task-relevant variables, reported relationships, and aggregate population statistics. MetaPersona retrieves evidence for a given task, builds literature-derived dependency graphs among persona attributes, and samples synthetic populations from empirical priors. Across three downstream case studies, three baselines, and three frontier models, its performance depends on the task and model: it performs strongly for misinformation belief and sentiment toward AI tools, while results for income redistribution are mixed. Using GPT-5.2, persona construction costs less than $0.50 per task. The paper also presents MetaPersona-Studio, a prototype interactive interface for generating empirically grounded personas.