Persona Generators: Generating Diverse Synthetic Personas for Arbitrary Contexts

Researchers introduce Persona Generators, AI functions that create diverse synthetic populations for evaluating AI systems across varied user demographics without needing extensive real-world data collection. Using iterative optimization with large language models, the approach generates lightweight code that produces synthetic personas spanning rare trait combinations and long-tail behaviors, outperforming existing baselines on diversity metrics.

This research addresses a fundamental challenge in AI evaluation: testing systems across diverse human populations without the cost and complexity of collecting representative real-world data. Traditional approaches to synthetic persona generation have relied on density matching—reproducing the most probable behaviors—which misses important edge cases and minority perspectives critical for comprehensive AI safety testing. Persona Generators shift this paradigm toward support coverage, capturing the full spectrum of human diversity including uncommon combinations of traits and beliefs.

The technical innovation centers on applying AlphaEvolve optimization principles to automatically refine persona generation code through hundreds of iterations. By treating the generation function itself as an evolving artifact rather than static instructions, researchers achieve lightweight generators that can be deployed efficiently while maintaining high fidelity to diverse populations. This approach has clear applications beyond the immediate research context, extending to any domain requiring diverse simulation data—from product design to policy evaluation to fairness testing of algorithmic systems.

For the AI industry, this work carries meaningful implications for responsible deployment. As AI systems become increasingly consequential across varied populations, the ability to test behavior across diverse demographics becomes essential for identifying disparate impacts and biases. The lightweight nature of evolved generators makes diversity testing more accessible and scalable for development teams. Looking forward, the challenge lies in ensuring these synthetic populations capture genuine variation rather than stereotypical caricatures, and in validating that evolved generators produce populations that meaningfully align with real-world diversity patterns rather than artificial artifacts of the optimization process.

• →Persona Generators use iterative AI-driven optimization to create diverse synthetic populations for AI evaluation without requiring expensive real-world data collection.
• →The approach prioritizes coverage of rare traits and long-tail behaviors over density matching, capturing underexplored population segments.
• →Evolved generators substantially outperform baseline methods across six diversity metrics on held-out test contexts.
• →Lightweight generated code enables scalable deployment of diversity testing across development workflows.
• →The technique has broad applications for fairness testing, bias detection, and comprehensive AI system evaluation.