Researchers simulate large LLM-agent societies on a laptop

Researchers propose a cheaper way to study large societies of LLM agents when the goal is to understand group-level behavior rather than the cognition of each agent.

The method replaces each language-model agent with a low-parameter model fitted from a few hundred to a few thousand inexpensive LLM queries. Once fitted, the surrogate agents can be scaled to larger populations on a laptop. The paper frames the approach through a taxonomy based on interaction order and memory, which predicts when surrogate error should grow or stay controlled as the number of agents increases.

The authors validate the idea on a reimplementation of the EconAgent macroeconomy and seven other LLM simulations, using decisions cloned from genuine LLM elicitations.

The work does not claim that small surrogates can replace LLM agents for every question. Its value is narrower: if researchers care about phase behavior, stylized facts, or scaling trends, they may not need to pay for every simulated decision with a full model call.