Should persona simulation prioritize coverage over statistical matching?

Most generative agent work optimizes for density matching — replicating the aggregate statistics of real populations. The Persona Generators paper (2025) argues this is the wrong objective for stress-testing and safety evaluation. Density matching emphasizes the most probable users, but critical failures are driven by outliers: the distrustful user with severe symptoms interacting with a mental health chatbot, the adversarial negotiator, the edge-case preference configuration.

The alternative objective is support coverage — spanning the full space of possible traits, opinions, and preferences including rare but consequential configurations. Simply asking an LLM to "generate diverse personas" fails: outputs cluster around stereotypical responses due to RLHF-induced mode collapse, even with explicit diversity instructions.

The solution uses an evolutionary search loop (AlphaEvolve) to optimize the code of a Persona Generator function — including prompt templates and sampling logic — rather than optimizing individual personas. The architecture separates population-level diversity decisions from per-persona background expansion, enabling both control and efficiency. Evolved generators substantially outperform baselines across six diversity metrics and generalize to held-out contexts.

The key insight is methodological: if the full support is covered, one can always later sample to match any specific target density. But if only density is matched, the long tail is permanently lost. This inverts the default assumption in persona simulation research and connects to the broader problem that How do we generate realistic personas at population scale?.