Abstract
Social networks shape how humans form opinions, exchange information, and organize collectively. As large language models (LLMs) become embedded in social and professional environments, it is critical to understand whether their interactions resemble human network dynamics. We introduce a framework to study the network formation behaviors of multiple LLM agents and benchmark them against human decisions. Across synthetic and real-world settings, including friendship, telecommunication, and employment networks, LLMs reproduce core microlevel principles (preferential attachment, triadic closure, and homophily), and macrolevel properties (community structure, small-world effects). Their emphasis on these principles adapts to context: for example, LLMs favor homophily in friendship networks but heterophily in organizational settings, mirroring patterns of social mobility. A controlled survey shows strong alignment between LLM and human link-formation decisions. These results highlight LLMs' potential as tools for social simulation and synthetic data generation, while underscoring risks of bias and fairness in AI systems that interact with human networks.