SSiLU: A Generalized Social Simulation Framework Empowered by LLM Agents and A Pool of 10 Million Real-World Users

ACL ARR 2025 May Submission3951 Authors

Abstract: Massive social simulation plays a vital role in predicting real-world trends. Previous studies use Large Language Models (LLMs) to replace traditional methods to enrich scenarios and improve simulation accuracy. However, they are faced with limitations such as rigid frameworks, small-scale simulations, and narrow evaluation criteria. To this end, we introduce **SSiLU**, a generalized **S**ocial **Si**mulation framework powered by **L**LM agents and a pool of 10 million real-world **U**sers. Our framework features a large-scale user pool, a demographic distribution sampling strategy, and a unified simulation evaluation method. We evaluate its effectiveness by conducting large-scale simulations across political, journalistic, and economic scenarios. The results demonstrate that our framework enables social simulations that reflect large-scale population dynamics, ensuring diversity, trustworthiness, and representativeness with a standardized pipeline and minimal modifications.

Paper Type: Long

Research Area: Computational Social Science and Cultural Analytics

Research Area Keywords: Large-scale Social Simulation, LLM agents, Computational Social Science

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English, Chinese

Submission Number: 3951