A Survey on LLM-based Agents for Social Simulation: Taxonomy, Evaluation and Applications

• Social simulation is a crucial tool in social science research, aiming to understand complex social systems. Recently, large language model (LLM) agents have demonstrated unprecedented human-like intelligence by leveraging the strong language understanding, generation, and reasoning capabilities of large language models. This paper conducts a comprehensive survey of social simulation empowered by LLM agents. We first review the evolution of social simulation paradigms and the development of LLM agents as background knowledge. Building on the foundational requirements for constructing a social simulator, we identify five essential capabilities that an individual LLM agent must possess. Correspondingly, we delineate five core modules that constitute the architecture of an LLM agent:(1) Profile Module for adaptive role-playing;(2) Perception Module for social context awareness;(3) Memory Module for continuous learning;(4) Planning Module for scenario-based reasoning; and (5) Action Module for dynamic decision-making. Additionally, we present a unified framework for LLM agent-based social simulation systems, comprising the simulation environment, the agent manager, and interacting LLM agents. We also introduce a comprehensive evaluation metric that integrates macro-and micro-level as well as subjective and objective criteria. The representative applications are categorized into four scenarios: uncovering social patterns, interpreting social phenomena, validating social theories, and forecasting policy outcomes. Finally, we identify the challenges and research opportunities in this field. Overall, this survey provides a systematic understanding of LLM agent-based social simulation, offering valuable insights for future research and applications in this field.