Physics of Agents: statistical mechanics predicts collective LLM-agent behavior across 10,000+ communities

arXiv·medium signalXBlueskyLinkedInCopy link

Studying over 10,000 communities of language-model agents that exchange messages and revise opinions on objective math questions and subjective political statements, the authors find all dynamics collapse into three regimes — indifference, polarization, consensus — with agents starting indifferent and building conviction through interaction. A statistical-mechanics model in which agents stochastically favor lower social pressure predicts individual trajectories from initial opinions alone, beats all standard baselines, and generalizes to unseen community graphs. Fitted parameters explain the mechanics: communities operate below the critical social temperature (conviction buildup), attractive ties outweigh repulsive ones (consensus bias), and agents holding the correct answer exert the strongest pull — but on subjective questions communication drifts groups rightward politically.