Researchers introduced a methodology for efficiently simulating large-scale LLM-agent societies by substituting expensive individual language-model agents with low-parameter surrogate models fitted from a few hundred to a few thousand cheap LLM queries. The work develops a perception-ordered taxonomy that predicts when such surrogates will hold up, by mapping how agents perceive information — global feeds, community-level, or local graph-based views — to expected error scaling as population size grows. Validated across eight named LLM simulations, the taxonomy correctly predicted surrogate accuracy in advance, allowing macroscopic multi-agent simulations to run on a single laptop rather than requiring large-scale API calls.
