Can we predict how agent communities shift opinions?

The paper studies "over 10,000 communities of language-model agents" that repeatedly exchange messages and revise their opinions, on objective mathematics questions and on subjective political statements. Its central claim is that this collective behavior is predictable. Despite "substantial diversity in possible behavior," the dynamics reduce to three regimes (indifference, polarization, consensus), and agents "start indifferent and build conviction as they interact." A statistical-mechanics formalism in which agents "stochastically favor lower social pressure" is fitted on a set of training questions. It then "generalizes to unseen questions and graphs, predicts individual trajectories, and approximately reproduces group-level outcomes." The claim is about prediction, not only description.

The fitted parameters carry the paper's account of why. First, the groups operate below a critical "social temperature," which drives conviction buildup. Second, concordant interactions are stronger than discordant ones, which drives consensus formation. Third, greater influence from correct neighbors "helps explain" truth-seeking on objective tasks. The observed outcomes split by question type. On objective questions collective accuracy improves over rounds. On subjective questions, three of the four models drift rightward on the political spectrum. The paper's own framing is that interacting agents can improve collective reasoning but may also produce herding, polarization or amplified shared biases, and the model is offered as a way to anticipate which one a given design will get.

This sits beside two notes that describe the same territory qualitatively. Since When does debate actually improve reasoning accuracy?, an objective-versus-subjective split in outcomes is already familiar; this paper confirms the accuracy half on math questions and adds a quantitative model. It differs in scope, though: its subjective items are political opinion statements, and the excerpt reports drift, not factual error. It also qualifies Does confidence drive influence in multi-agent deliberation systems?. That note holds that influence follows confidence proxies rather than competence, while this paper's third fitted parameter has correct neighbors exerting more influence on objective tasks. The two use different formalisms and the excerpt does not compare them, so they may describe different regimes and not a conflict. Against Why don't AI agents develop social structure at scale?, where agents ignore feedback, these simulated communities show agents that do revise opinions and build conviction under interaction.

The excerpt does not say which four models were tested, how large the groups were, which communication graphs or how many rounds were used, or how large the accuracy gain and the rightward drift were. It does not define how "social temperature" is measured. It also does not say why the drift is rightward: the three parameters it lists account for conviction, consensus and truth-seeking, not for direction, and "helps explain" is weaker than showing that correct neighbors cause the accuracy gain. All results come from simulated groups, not deployed agents. At the strength the evidence allows, the paper offers a fitted, testable predictor of when agent groups will harden into agreement, and a reason not to assume that an accuracy gain measured on math will carry over to opinion tasks.