Network formation and dynamics among multi-LLMs.

Area of Science:

• Social network analysis
• Artificial intelligence
• Computational social science

Background:

• Social networks are fundamental to human interaction, opinion formation, and collective organization.
• Large language models (LLMs) are increasingly integrated into social and professional environments.
• Understanding LLM network dynamics is crucial for assessing their societal impact.

Purpose of the Study:

• To introduce a framework for studying network formation behaviors of multiple LLM agents.
• To benchmark LLM network formation against human decisions.
• To explore LLM adaptability to different social contexts.

Main Methods:

• Development of a framework to analyze LLM agent interactions.
• Benchmarking LLM decisions against human link-formation data.
• Testing across synthetic and real-world network datasets (friendship, telecommunication, employment).

Main Results:

• LLMs reproduce core microlevel network principles (preferential attachment, triadic closure, homophily).
• LLMs replicate macrolevel network properties (community structure, small-world effects).
• LLM network behavior adapts contextually, favoring homophily in friendship and heterophily in organizational settings.

Conclusions:

• LLMs demonstrate significant alignment with human network formation principles.
• LLMs show potential as tools for social simulation and synthetic data generation.
• Risks of bias and fairness in AI interacting with human networks are underscored.