Coached LLM agents spontaneously form social networks like humans

A new paper from researchers Philipp Schneider, Lin Tian, and Marian-Andrei Rizoiu (arXiv:2510.19299) presents a multi-agent LLM simulation framework designed to study whether AI agents can reproduce the complex social dynamics of human online behavior—homophily, reciprocity, and social validation. The agents repeatedly interact, evaluate one another, and adapt through in-context learning accelerated by a coaching signal, without requiring fine-tuning. To model realistic human behavior, the authors define behavioral reward functions that capture core online engagement drivers: social interaction, information seeking, self-presentation, coordination, and emotional support. These rewards align agent objectives with empirically observed user motivations, allowing the researchers to study how network structures and group formations emerge from individual decisions.

In experiments, coached LLM agents developed stable interaction patterns and spontaneously formed emergent social ties. The resulting network structures closely mirror properties of real online communities, including evidence of homophily and reciprocity. The framework establishes a principled testbed for investigating collective dynamics in LLM populations, revealing both how artificial agents approximate human-like social behavior and where they diverge. Accepted at the NeurIPS 2025 Workshop on Scaling Environments for Agents (SEA), this work opens new avenues for studying emergent social phenomena in AI systems and designing more socially aware multi-agent architectures.