Research map / Large language models / Alignment and preferences

Persona Generators: Generating Diverse Synthetic Personas for Arbitrary Contexts

NeurIPS 2026 · Large language models › Alignment and preferences · approach: LLM-as-judge and feedback data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Evolves LLM-written code for Persona Generators with AlphaEvolve so that generated synthetic populations maximize support coverage across diversity axes, outperforming Nemotron Personas and prompting baselines on six diversity metrics and matching real BFI distributions better.

Problem

Synthetic persona populations from LLMs suffer mode collapse toward stereotypical, agreeable responses and density matching overlooks rare, consequential outliers needed for stress-testing and simulation.

Approach

Shift the objective from density matching to support coverage and learn a reusable generator function (code) rather than fixed personas; an LLM-driven evolutionary loop (AlphaEvolve) mutates the generator code, using diversity metrics computed on Concordia simulated questionnaire responses as fitness.

Key results

• Best evolved Persona Generator reaches avg z 2.31 on held-out test vs 1.41 for Nemotron Personas and 0.89 for Diversity Prompting, with coverage 0.82 vs 0.73 on test.
• On real BFI data, the evolved generator achieves 62.9% human coverage and KL 0.99 vs 51.5% and 3.00 for Nemotron Personas.
• LLM judges rated 93.0%–99.2% of 1,000 generated personas as realistic.

Surprising findings

• Evolved generators transferred to other inference models (Gemini 3.1 Flash-Lite, Qwen 3.5 27B) and sometimes exceeded the original Gemma model on hull volume and distance metrics.
• Zero-shot Diversity Prompting by a frontier model underperforms evolved code, suggesting control-flow structure matters beyond prompt engineering.

Limitations

• Diversity evaluated mainly on stated preferences via questionnaires; open-ended behavioral diversity only assessed qualitatively and with noisy embeddings.
• Evolved personas occasionally contain mildly implausible trait combinations (e.g., medical student with hedge-fund skills).
• Search space biased by the mutation LLM's priors; optimization cost is nontrivial despite cheap inference.

Experimental setup

Datasets: Big Five Inventory (BFI) human responses Nemotron Personas BFI DASS SVO NFCS

Benchmarks: synthetic questionnaires (10 held-out test contexts) BFI human responses (1,000)

Metrics: convex hull volume coverage dispersion KL divergence minimum pairwise distance average pairwise distance average z-score in-support rate human coverage

Baselines: Nemotron Personas Diversity Prompting Concordia Generator Name Baseline

Builds on: AlphaEvolve Concordia logic of appropriateness

Closest papers by meaning

• LLM Generated Persona is a Promise with a Catch (NeurIPS 2025 · similarity 0.92)
• Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• Human Simulacra: Benchmarking the Personification of Large Language Models (ICLR 2025 · similarity 0.87)
• Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces (NeurIPS 2026 · similarity 0.87)
• Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.85)
• PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 2024 · similarity 0.85)