Bots of a feather: mixing biases in LLMs’ opinion dynamics

The rapid integration of Large Language Models (LLMs) into everyday applications raises critical questions about their group in- teractions, consensus formation, and potential to mimic human-like be- havior. Although initial research has explored the evolution of opinions within LLM populations, these efforts often rely on simplistic network assumptions, such as uniform connections among agents, thereby over- looking the influence of more realistic network topologies. This paper introduces a framework for examining opinion dynamics among LLM agents within various network structures. We perform several multi- model simulations on network topologies with known locally assorta- tive/disassortative mixing patterns. We find that convergence is quicker in mostly-disassortative networks compared to networks with no mixing biases. However, the joint effect of assortative and disassortative patterns leads to slower/no convergence.