Interpretable Mechanism Design with Large Language Models

This paper presents a novel automated mechanism design framework that utilizes large language models (LLMs) to reformulate the process as a code generation task. The framework generates heuristic mechanisms in code and evolves them to optimize for performance metrics while ensuring crucial design criteria through a problem-specific fixing process. This approach addresses limitations in traditional and neural-network-based methods by providing interpretable solutions and demonstrating competitive performance. The text describes applications to problems like rediscovering virtual valuations and designing VCG redistribution and correlated bidder auction mechanisms, showcasing the potential for enhanced transparency and scalability in mechanism design.