Evidence from the Synthetic Laboratory: Language Models as Auction Participants

This paper investigates auction behavior of simulated AI agents (large language models, or LLMs). We begin by benchmarking these LLM-driven agents against established lab experiments across various auction settings: independent private value, affiliated private value, and common value auctions. Our findings reveal that LLM agents exhibit many behavioral traits similar to those observed in human participants within lab environments.Building on this, we investigate multi-unit combinatorial auctions under three distinct bid formats: simultaneous, sequential, and menu-based. Our study contributes fresh empirical insights into this classical auction framework.We run 1,000+ auctions for less than \$100 with GPT-4, and develop a framework flexible enough to run auction experiments with any LLM model and a wide range of mechanism specification.