Technical Summary: Persona Generators for Diverse Synthetic Populations

Problem Statement

Simulating human behavior with Large Language Models (LLMs) offers a scalable laboratory for social science and AI stress-testing. However, standard LLMs frequently suffer from "mode collapse," collapsing onto a narrow subset of stereotypical, highly agreeable, or WEIRD (Western, Educated, Industrialized, Rich, Democratic) responses. This limits the ability to simulate the full breadth of human diversity, particularly rare but consequential outliers.

Current approaches to creating synthetic humans typically optimize for density matching, attempting to align synthetic outputs with the aggregate statistics of real-world data. While effective for modeling average users, this approach is fundamentally unsuited for stress-testing or forecasting in speculative scenarios (e.g., societal responses to AGI) where exploring the full range of possible human responses is more critical than identifying the most probable one. The paper argues that robust simulation requires a shift in objective from algorithmic fidelity (matching specific individuals) to support coverage: ensuring synthetic populations span the entire landscape of possible human traits, opinions, and preferences, including the "long tail" of extreme behaviors.

Methodology

The paper introduces Persona Generators, lightweight functions capable of producing diverse synthetic populations for arbitrary contexts on demand. Rather than optimizing individual personas, the authors optimize the code of the generator itself using an evolutionary search loop powered by AlphaEvolve.

Pipeline Overview

1. Questionnaire Generation: The process begins by generating synthetic questionnaires ($q$) comprising a context ($c$), a set of diversity axes ($D$), and survey items ($I$). These are created using few-shot prompting based on established psychometric tools (e.g., Big Five Inventory, DASS, SVO).
2. Persona Generation ($G_{\phi,\theta}$): A Persona Generator function takes the context and diversity axes as input to produce a population of $N$ synthetic personas ($P$). The generator architecture follows a two-stage process:
   • Stage 1 (Autoregressive): The LLM generates high-level descriptors for each persona, determining their position along the specified diversity axes.
   • Stage 2 (Parallel): These descriptors are expanded in parallel into full persona profiles with specific traits and background details.
3. Simulation and Evaluation: Each persona is placed in a Concordia simulation where they answer the questionnaire items. The responses are mapped to numerical embeddings ($Z$).
4. Diversity Metrics: The population's diversity is quantified using six metrics ($M(Z)$):
   • Coverage (Monte Carlo estimate)
   • Convex Hull Volume
   • Minimum Pairwise Distance
   • Average Pairwise Distance
   • Dispersion (radius of largest empty region)
   • KL Divergence from an ideal quasi-random distribution
5. Evolutionary Optimization: AlphaEvolve uses the diversity metrics as a fitness function to iteratively mutate the generator's code ($\phi$). The mutation operator (an LLM) modifies the logic within Stage 1 and Stage 2 to discover new generator implementations that maximize coverage across the diversity axes.

Key Contributions

1. Formalization of Support Coverage: The paper formalizes synthetic persona generation as a diversity maximization task over trait embeddings, explicitly shifting the objective from matching specific individuals to spanning the space of possible traits and preferences.
2. Novel Persona Generator Architecture: A two-stage scalable architecture is proposed that separates population-level diversity decisions (Stage 1) from per-persona background expansions (Stage 2). This is coupled with a pipeline that uses LLMs to generate questionnaires, simulate interactions, and evolve the generator code.
3. Evolutionary Discovery of Generators: The authors demonstrate that LLM-driven evolutionary search can discover novel generator functions that substantially outperform existing baselines in coverage and diversity metrics.

Results

The evolved Persona Generators were evaluated against four baselines: Nemotron Personas (a static dataset grounded in US demographics), a Diversity Prompting baseline (optimized via high-effort prompting), a Concordia Generator (formative memory based), and a Name-only baseline.

• Diversity Metrics: The evolved generators substantially outperformed all baselines across all six diversity metrics on both training/validation and held-out test sets. Notably, they outperformed the Diversity Prompting baseline, suggesting that iterative evolutionary search over control flow yields structural strategies that zero-shot prompting cannot achieve.
• Realism and Human Fidelity: Despite optimizing for broad support coverage, the generated personas remained realistic. An LLM-as-a-judge evaluation found 93.0%–99.2% of generated personas were deemed realistic. Crucially, when compared to real human data (Big Five Inventory), the evolved generator achieved the highest coverage (62.9%) and lowest KL divergence (0.99) among all methods, significantly outperforming baselines explicitly grounded in real-world demographics. This indicates that maximizing support coverage mitigates LLM mode collapse and better captures true human variance than density-matching approaches.
• Behavioral Transfer: In open-ended tasks (comedy writing and conflict resolution), evolved generators produced a broader spectrum of behaviors and emotions, whereas baselines exhibited severe mode collapse (e.g., defaulting to polite, conflict-averse responses).
• Generalization and Transfer: The evolved code generalized well to held-out contexts and transferred effectively across different inference models (Gemma, Gemini, Qwen), maintaining high diversity margins.

Significance and Claims

The paper claims that optimizing the generator itself, rather than individual personas or fixed populations, is a viable and promising direction for creating diverse synthetic agents. The resulting Persona Generators are lightweight and reusable, enabling on-demand diverse population synthesis.

The authors emphasize that this approach does not come at the expense of realism; by explicitly fighting LLM mode collapse, the method captures real-world human trait distributions better than baselines designed to match human statistics. The work suggests that if the full support of human traits is covered, one can subsequently re-sample the population to match any specific target density.

The paper concludes by noting the dual-use nature of this technology: while useful for stress-testing AI, red-teaming, and generating diverse interactive agents, it carries risks of misuse for creating bots or disinformation campaigns. The authors plan to open-source the top-performing implementations to support further research.