Collaborative Spatial Learning with Multi-LLM Agents in Networked Social Experiments

This paper adapts the Mason–Watts two-dimensional search experiment to groups of sixteen LLM agents connected through eight network topologies. It also introduces mechanistic Bayesian optimization agents for comparison with the LLM groups and prior human data. The experiments report a significant network-efficiency effect when agents are instructed to randomize their first-round choices, but not with default initialization. A single-sentence randomization instruction raises collective payoff by more than three times the estimated payoff difference across topologies. The Bayesian optimization agents outperform the evaluated LLM agents, while the study also examines exploration versus exploitation, copying, and spatial diversity.