Persona Generators

Generating Diverse Synthetic Personas at Scale

The problem is that LLMs collapse onto a narrow band of agreeable, stereotypical responses when asked to roleplay people — the range of personas they produce is dramatically narrower than the range of actual humans walking around. A new paper from Google DeepMind, Persona Generators: Generating Diverse Synthetic Personas at Scale, takes a new approach.

For marketing, the outliers are exactly who break your messaging, churn after one bad onboarding, or post the screenshot that goes viral. Optimizing for coverage — spanning the full space of possible humans, including the weird tail — gives you a population you can stress-test against.

The clever bit is how they get there: rather than hand-crafting better prompts, they evolve the persona generator’s code itself, using AlphaEvolve with an LLM as the mutation operator and six diversity metrics as the fitness signal. Over hundreds of iterations, the system discovers generator programs that substantially outperform existing baselines (including Nvidia’s 100k-persona Nemotron dataset) and generalize to held-out contexts the optimizer never saw — everything from “Gen Z social media politics 2025” to hypothetical AGI-displacement scenarios in 2040.

To ensure the AI wasn’t just giving fake users different names but the same opinions, DeepMind forced the AI personas to take personality tests and surveys. The system was only rewarded if it could generate a population that wildly disagreed with each other on the survey—proving it successfully created true human variety, from skeptics to zealots.

The authors say they plan to open-source the top-performing generators upon acceptance, which would make this directly usable for anyone wanting to pressure-test a campaign, product, or chatbot against a population that actually includes the people most likely to break it.