Large language model agents are increasingly used as computational actors in multi-agent social simulations. Most existing work in this area emphasizes representative personas and average-case behavior, aiming to capture typical responses rather than edge cases.
The new arXiv preprint "When Extremists Enter LLM Societies" takes a different angle. The authors point to social science research showing that persistent extreme behavior can have significant effects, even when it is rare. They argue that current simulation practices may miss these dynamics by focusing on the average.
The paper investigates what happens when extremists are introduced into LLM-based multi-agent systems. By explicitly modeling these actors, the work aims to understand how extreme behavior influences the broader simulated society, addressing a gap in how LLM social simulations are typically designed and evaluated.