Abstract:

With the continuous evolution of the capabilities of generative LLMs, their application in social cognition simulation is demonstrating paradigm-shifting potential. Traditional social simulation methods predominantly rely on static rules and simplified behavioral models, making it difficult to capture the dynamic evolution and cultural complexity of human social behavior. LLM-driven agents, equipped with contextual understanding and natural language generation capabilities, are emerging as novel tools for modeling social cognitive mechanisms, enabling the simulation of complex socio-psychological processes such as identity construction, value judgment, and intentional reasoning. This paper briefly introduced the technical foundations of LLMs and highlighted their suitability for social cognition simulation. It constructed a framework for agent-based social cognition modeling, encompassing attribute modeling, memory management, planning, and action. At the simulation process level, the paper proposed a technical pipeline consisting of “data collection, agent collaboration, and multidimensional evaluation,” while delving into challenges such as cognitive interpretability and simulation-reality alignment.It summarized the current application progress in fields such as sociology, economics, and military science and discussed emerging trends and future directions for LLM-based social cognition simulation.

Key words: LLM, social cognition simulation, agent-based modeling, simulation-reality alignment