Four ways a crowd of agents can split apart

A paper on multi-agent systems asks what happens when a population of learning agents is left to form opinions about each other. Each agent maps a language-model embedding of an issue to a binary opinion through an adaptive linear decision layer, and keeps an evolving estimate of how far it trusts every other agent. Through pairwise interaction both the decision weights and the trust estimates update, under an online Bayesian rule derived from teacher-student learning over a noisy channel.

The authors then add the variable that matters. Each agent is given one of two arbitrary group labels, A or B, and a fraction of the population is allowed to shift its perceived agreement with a message depending on whether sender and receiver share a label. Two numbers describe that bias: how strong it is and how widespread.

What follows is a phase diagram rather than a trend. Starting from uniformly random initial conditions, simulations settle into one of four collective states: spin-glass frustrated out-group favouritism, nondiscriminatory polarisation, discrimination with opinion alignment, and discrimination without opinion alignment. Which one appears depends on both the strength and the prevalence of the bias, and the authors use order parameters from statistical mechanics to connect the local rules to the collective outcome.

Two details make the result harder to set aside. In-context-learning experiments with large language models qualitatively reproduced the predicted opinion and trust updates, so the dynamics are not only theoretical. And discrimination appears in two forms, one where opinions align and one where they do not, which means a population can discriminate while looking undivided on the issue itself. The conclusion the authors draw is that safety analysis of multi-agent systems has to be done at the level of the population, because none of these states shows up in a single agent.