Physics of Agents Paper Models LLM Agent Dynamics

James Zou and collaborators including Batu El and Surya Ganguli released work on Physics of Agents. Their simulations tracked more than ten thousand LLM agent communities discussing topics from math problems to energy policy. Agents exchanged views with network neighbors and shifted opinions over rounds. The collective patterns matched predictions from statistical mechanics models such as the Ising framework. Agents tended toward low energy states that reduced social pressure. The same simple model generalized across different communication networks. Zou noted the findings may help with safety and alignment questions in multi agent systems. A site lets users view the resulting social dynamics.