Erica Cau, Andrea Failla, Giulio Rossetti

Researchers at arXiv examined how large language models behave in multi-round debate simulations using controlled network models with varying homophily and group sizes. They identified a phenomenon called 'agreement drift,' where LLM agents systematically shift toward specific positions on opinion scales rather than converging randomly. The findings suggest that LLM-based social simulations may conflate structural network effects with inherent model biases, raising questions about their reliability as proxies for human group behavior, especially in unbalanced contexts involving minority groups.