Abstract:

The breakthrough of LLMs has provided powerful tools for social network research, advancing multi-agent social network simulation into a new era. This review systematically examined recent progress in LLM-driven multi-agent social network simulation research through a integrated perspective of multi-disciplines such as artificial intelligence, psychology, communication studies, and sociology. A three-tiered research system, which has gradually formed in this field and encompassed micro-level individual behaviors, meso-level interactive relations, and macro-level system emergence, was summarized. At the micro-level, research focuses on individual human behavior simulation, and numerous studies are dedicated to developing human-like agents with complex cognitive and affective architectures that integrate psychology theories with LLMs. At the meso-level, research focus shifts to human social interaction behavior simulation, mainly investigating the dynamic interactions and relation evolution among agents. At the macro-level, research aims to uncover the emergent dynamics underlying complex social phenomena such as information dissemination and group polarization by means of simulation. Key theories, technical routes, and interdisciplinary integration case studies of research at various levels were synthesized. The challenges faced by current research in aspects such as agent behavior validity verification, generative uncertainty, and computational costs were analyzed. Future research directions were discussed, such as virtual-real linkage validation, the establishment of standardized benchmarks, and optimization of memory and computational efficiency.

Key words: large language model, social network, multi-agent, psychological traits, social simulation