Simulating LLM Societies Using Low-Parameter Surrogates

This method replaces expensive LLM agents with low-parameter models trained on a few thousand cheap queries to simulate macroscopic social behaviors on local hardware. The approach uses an interaction order and memory taxonomy to predict error trends when scaling agent populations.