LLM social networks: Specialist models double belief consensus shift

A new paper from Northeastern University researchers (Germans Savcisens, Samantha Dies, Courtney Maynard, and Tina Eliassi-Rad), titled "Belief Coevolution in a Social Network of Generalist and Specialist Large Language Models," tackles a growing question: how do beliefs form and spread when multiple LLMs interact? They introduce CoevolveSim, a simulation framework that isolates three key factors—domain specialization, social-role assignment, and network structure—to study belief diffusion in networked LLM populations. In each round, generalist and specialist LLM agents observe a summary of their neighbors' beliefs before revising their own, mimicking social influence dynamics.

The study ran 1,280 controlled simulations spanning four scenarios, two network structures, and 20 medical-indication statements. Their findings challenge common assumptions: persona-style role assignment and network structure reshape individual belief revision but have minimal effect on population-level consensus. By contrast, introducing fine-tuned specialist LLMs more than doubles the shift in consensus and creates consistent asymmetries in exerted influence. The authors also found that simple persistence-based opinion-dynamics models can reproduce collective outcomes for all-generalist populations, but heterogeneous LLM populations require population-level belief composition and agent identity to predict individual transitions. This suggests that building realistic multi-agent AI simulations demands genuine diversity in underlying models, not just crafted personas.