Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Population-Aligned Persona Generation for LLM-based Social Simulation

Zhengyu Hu, Jianxun Lian, Zheyuan Xiao, Max Xiong, Yuxuan Lei, Tianfu Wang, Kaize Ding, Ziang Xiao, Nicholas Jing Yuan, Xing Xie

TL;DR

This paper tackles biases in LLM-based social simulation arising from unrepresentative persona sets by introducing a principled, three-stage framework: seed persona mining from large-scale narrative data, global distribution alignment via a two-stage process combining KDE-based importance sampling and entropic optimal transport, and group-specific persona construction through embedding-driven retrieval and targeted LLM revisions. The approach aligns synthesized personas with real-world psychometric distributions (e.g., IPIP Big Five) and demonstrates improved population-level fidelity and individual-level trait consistency across multiple datasets and models, with finite-sample guarantees. Extensive experiments across six psychometric instruments and regional settings show that the proposed Resample method consistently outperforms existing persona baselines and public sets, and that group-specific adaptation further enhances applicability to targeted subpopulations. The work advances reliable, policy-relevant social simulations by ensuring that emergent dynamics reflect realistic population distributions rather than an 'average persona' fallacy.

Abstract

Recent advances in large language models (LLMs) have enabled human-like social simulations at unprecedented scale and fidelity, offering new opportunities for computational social science. A key challenge, however, is the construction of persona sets that authentically represent the diversity and distribution of real-world populations. Most existing LLM-based social simulation studies focus primarily on designing agentic frameworks and simulation environments, often overlooking the complexities of persona generation and the potential biases introduced by unrepresentative persona sets. In this paper, we propose a systematic framework for synthesizing high-quality, population-aligned persona sets for LLM-driven social simulation. Our approach begins by leveraging LLMs to generate narrative personas from long-term social media data, followed by rigorous quality assessment to filter out low-fidelity profiles. We then apply importance sampling to achieve global alignment with reference psychometric distributions, such as the Big Five personality traits. To address the needs of specific simulation contexts, we further introduce a task-specific module that adapts the globally aligned persona set to targeted subpopulations. Extensive experiments demonstrate that our method significantly reduces population-level bias and enables accurate, flexible social simulation for a wide range of research and policy applications.

Population-Aligned Persona Generation for LLM-based Social Simulation

TL;DR

This paper tackles biases in LLM-based social simulation arising from unrepresentative persona sets by introducing a principled, three-stage framework: seed persona mining from large-scale narrative data, global distribution alignment via a two-stage process combining KDE-based importance sampling and entropic optimal transport, and group-specific persona construction through embedding-driven retrieval and targeted LLM revisions. The approach aligns synthesized personas with real-world psychometric distributions (e.g., IPIP Big Five) and demonstrates improved population-level fidelity and individual-level trait consistency across multiple datasets and models, with finite-sample guarantees. Extensive experiments across six psychometric instruments and regional settings show that the proposed Resample method consistently outperforms existing persona baselines and public sets, and that group-specific adaptation further enhances applicability to targeted subpopulations. The work advances reliable, policy-relevant social simulations by ensuring that emergent dynamics reflect realistic population distributions rather than an 'average persona' fallacy.

Abstract

Recent advances in large language models (LLMs) have enabled human-like social simulations at unprecedented scale and fidelity, offering new opportunities for computational social science. A key challenge, however, is the construction of persona sets that authentically represent the diversity and distribution of real-world populations. Most existing LLM-based social simulation studies focus primarily on designing agentic frameworks and simulation environments, often overlooking the complexities of persona generation and the potential biases introduced by unrepresentative persona sets. In this paper, we propose a systematic framework for synthesizing high-quality, population-aligned persona sets for LLM-driven social simulation. Our approach begins by leveraging LLMs to generate narrative personas from long-term social media data, followed by rigorous quality assessment to filter out low-fidelity profiles. We then apply importance sampling to achieve global alignment with reference psychometric distributions, such as the Big Five personality traits. To address the needs of specific simulation contexts, we further introduce a task-specific module that adapts the globally aligned persona set to targeted subpopulations. Extensive experiments demonstrate that our method significantly reduces population-level bias and enables accurate, flexible social simulation for a wide range of research and policy applications.

Paper Structure

This paper contains 30 sections, 41 equations, 3 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Population-Aligned Persona Framework
  4. Seed Persona Mining
  5. Global Distribution Alignment via Two-stage Sampling
  6. Group-specific Persona Set Construction
  7. Experiments Setup
  8. Population Level
  9. In-domain results.
  10. Out-of-domain results.
  11. Individual‑Level Behavioral Consistency
  12. Trait-Space Visualization
  13. Group-Specific Persona Adaptation
  14. Conclusion
  15. Notation
  16. ...and 15 more sections

Figures (3)

  • Figure 1: Distributions of Big Five personality traits from real human data, LLM responses without personas, six existing persona sets, and GPT-4o-generated personas.
  • Figure 2: Overview of the Population-Aligned Persona Framework.
  • Figure 3: Scatter plots comparing the distribution of EXT and EST scores between real human responses (blue) and LLM-generated outputs (orange) across nine settings.