Deep Iterative Persona Alignment: Generating Statistically Representative LLM Personas for High-Fidelity Social Simulations

The increasing adoption of large language models for simulating human behavior offers a promising new paradigm for social science research. However, a critical limitation persists: current LLM persona generation methods prioritize individual narrative richness over accurate population-level alignment of demographic and psychological traits, thereby compromising scientific validity. To address this, we introduce Deep Iterative Persona Alignment (DIPA), a novel framework that systematically generates LLM personas exhibiting high fidelity to real human psychological distributions. DIPA integrates powerful narrative generation capabilities with a trainable Psychometric Response Adapter (PRA) and an iterative optimization process. The PRA learns to generate psychologically plausible responses, while an Optimal Transport-based iterative loop refines the persona library for precise population alignment. Our experimental results demonstrate that DIPA significantly outperforms state-of-the-art baselines across population-level alignment metrics on psychological tests. Furthermore, DIPA shows strong generalization to unseen psychological tests, maintains high qualitative realism, and effectively adapts to generate group-specific persona sets. DIPA thus establishes a more robust statistical foundation for LLM-driven social simulations, paving the way for more accurate and reliable insights into social phenomena.