代理系統研究
Ten Thousand Agent Communities Exhibit Three Phases—Consensus, Polarization, and Apathy—With the Ising Model Predicting Opinion Evolution
A Stanford team simulated more than 10,000 communities of language-model agents and found that, over multiple rounds of communication, initially weak opinions gradually develop into consensus or polarization. Using an extended Ising model, the researchers predicted individual trajectories on previously unseen communication graphs. Political questions, however, also produced a rightward shift, showing that collaboration does not necessarily eliminate model bias.
The research team created communities of 32 agents and had them exchange messages over eight rounds about binarized MATH problems and political statements. Each agent was assigned a different area of expertise or persona, while connections were classified as friendly, adversarial, or non-communicating. The experiments covered GPT-4o-mini, Gemma-3n-E4B-it, Qwen3.5-9B, and Llama-3.1-8B-Instruct. The researchers analyzed 9,600 primary simulations and more than 10,000 in total when including out-of-distribution and asynchronous experiments.
The results showed four types of individual behavior: maintaining the same position, switching once, switching and then reverting, and oscillating repeatedly. At the group level, communities could settle into phases of apathy, polarization, or consensus. With GPT-4o-mini and Qwen3.5-9B, roughly 11% to 12% of communities weakened or overturned the original majority. On mathematics problems, incorrect majorities shifted toward the correct answer more often than correct majorities shifted in the opposite direction. Political statements, however, frequently produced a rightward drift along the ideological spectrum.
The authors represented agents’ positions as binary spins, used intrinsic fields to describe preferences induced by personas or questions, and modeled pressure from neighboring agents with a signed social graph. After fitting, a Glauber/Ising-style update rule could project subsequent trajectories from the initial opinions and generalize from random graphs to previously unseen grid and other network structures. The fitted parameters indicated that the system operated in an ordered regime below the critical temperature, that attraction through friendly connections was stronger than repulsion through adversarial ones, and that agents holding the correct answer exerted greater influence on objective questions.
This provides multi-agent orchestration with a more granular diagnostic tool than single-run success rates. Engineers can monitor interaction temperature, connection signs, and the initial majority to estimate whether debate will improve answers or amplify bias. The experiments, however, compressed opinions into binary values and primarily used synchronous updates. The personas and political datasets may also have introduced existing biases into the results. Whether this physics-based model applies to production agents that use tools, retain memory, or modify their goals remains to be validated.