LLM market simulations fail without proper rules
The same AI agents perform wildly differently across market designs—efficient surplus ranges from 56% to 88%
Maxim Chupilkin's arXiv paper demonstrates that institutional architecture is as critical as agent design in LLM-driven economic simulations.
In a controlled repeated market experiment using identical LLM agents (same prompts, memory, and reasoning capabilities), Chupilkin tested five standard market institutions: call markets, posted-offer markets, posted-bid markets, continuous double auctions, and bilateral bargaining. The results were stark—efficient surplus ranged from 88.6% in call markets to just 56.4% in bilateral bargaining. The institutional framework also affected trade quantities, price distance from competitive equilibrium, and surplus distribution between buyers and sellers.
The paper argues that while most research focuses on optimizing agents (their personas, reasoning, or memory), the rules governing their interactions are equally important. Minimal changes to institutional design can produce qualitatively different social outcomes, challenging the assumption that LLMs will naturally produce human-like economic behavior.
- Identical LLM agents achieved 88.6% efficient surplus in call markets vs 56.4% in bilateral bargaining
- Institutional design affects trade quantities, price equilibrium, and surplus distribution
- Research focus should expand from agent optimization to institutional architecture
Why It Matters
Institutional rules are the hidden variable determining whether AI simulations reflect real-world economics