Existing persona seeds for large language model social simulations suffer from a cold‑start problem: there is no principled way to choose which attributes to include or how to assign their values, so synthetic populations may misrepresent demographic composition, latent traits, and dependency structures. To address this, we built MetaPersona‑DB, a collection of over 11,000 empirical human‑subjects studies annotated with task‑relevant variables, reported relationships, and aggregate population statistics. Leveraging this resource, the MetaPersona framework proceeds in three steps: retrieve task‑relevant evidence, construct literature‑derived persona dependency graphs, and sample synthetic populations from empirical priors that link demographics, latent attributes, and outcomes. We evaluated the approach on three downstream case studies, three baselines, and three frontier models. Results vary by task and model: strong performance on misinformation belief and AI‑tool sentiment, mixed outcomes on income redistribution. Persona‑construction cost drops below $0.5 per task when using GPT‑5.2. Finally, we present MetaPersona‑Studio, a prototype interactive interface for empirically grounded persona generation.
Review