Group size effects in multi-agent large language model (AI) systems

Sep 01, 2026

A collaboration among professor Romualdo Pastor-Satorras from the department of Physics of the Universitat Politècnica de Catalunya, and professors from the University of London and the University of Copenhagen have shown that interactions among large language model (LLM) agents can lead to collective misalignment, with emergent biases whose strength and form vary systematically with population size.

Group size effects in multi-agent large language model (AI) systems — From Left to Right columns, the word pairs that populations coordinate on are {American, Mexican}, {White, African}, and {straight, gay}. These cases illustrate bias amplification, induction from neutrality, and bias reversal, respectively.

As multi-agent large language model (AI) systems are rapidly deployed at scale, their collective behaviors become increasingly difficult to anticipate from the properties of isolated models, rendering traditional single-agent and fixed-size group evaluations insufficient. Moving beyond this dichotomy, this research investigates how group size shapes multi-agent dynamics and misalignment across the full population spectrum. We show that interactions among models do not simply aggregate individual traits: they can amplify pre-existing biases, generate novel collective biases, or even reverse model-level preferences. Crucially, these emergent effects depend on group size in unexpected and nonlinear ways, giving rise to distinct, model-dependent dynamical regimes and showing that, for interacting AI populations, more is indeed different. To mathematically characterize these population-level transitions, we develop a mean-field analytical framework showing that, beyond a critical population size, multi-agent dynamics converge toward deterministic behavior, revealing the basins of attraction associated with competing equilibrium states. By establishing group size as a fundamental driver of collective behavior, these findings underscore the need to account explicitly for population-level dynamics in the design, deployment, and governance of multi-agent AI systems.

Reference:

Flint, Aiello, Pastor-Satorras, and Baronchelli, Group size effects and collective misalignment in LLM multi-agent systems, Proc. Natl. Acad. Sci. U.S.A.123 (34) e2531697123, (2026).

r_n