Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop
The paper claims large LLM-agent societies can be approximated cheaply by fitting low-parameter stand-ins, then scaling the simulation on a laptop.
The method treats the target questions as macroscopic, not agent-by-agent cognition. It fits surrogate agents from hundreds to thousands of inexpensive LLM queries, mainly using DeepSeek elicitations. The author says an interaction-order-and-memory taxonomy predicts when surrogate error should scale well with population size. Validation covers a reimplementation of EconAgent and seven other named LLM simulations, with two failed predictions traced to response curvature. HF Daily Papers' note
The method treats the target questions as macroscopic, not agent-by-agent cognition. It fits surrogate agents from hundreds to thousands of inexpensive LLM queries, mainly using DeepSeek elicitations. The author says an interaction-order-and-memory taxonomy predicts when surrogate error should scale well with population size. Validation covers a reimplementation of EconAgent and seven other named LLM simulations, with two failed predictions traced to response curvature. HF Daily Papers' note
score 4