An Interpretable Automated Mechanism Design Framework with Large Language Models

• DOI: 10.48550/arxiv.2502.12203
• Publication Date: 2025-02-16
• Authors: Jiayuan Liu; Mingyu Guo; Vincent Conitzer
• Keywords: Mechanism Design

Mechanism design has long been a cornerstone of economic theory, with traditional approaches relying on mathematical derivations. Recently, automated approaches, including differentiable economics neural networks, have emerged for designing payments and allocations. While both analytical methods advanced the field, they each face significant weaknesses: derivations are not often struggle to scale complex problems, while especially neural-network-based suffer from limited interpretability. To address these challenges, we introduce novel framework that reformulates mechanism as code generation task. Using large language models (LLMs), generate heuristic mechanisms described in evolve them optimize over some evaluation metrics ensuring key criteria (e.g., strategy-proofness) through problem-specific fixing process. This process ensures any violating is adjusted satisfy them, albeit trade-offs performance metrics. These factored during LLM-based evolution The capabilities LLMs enable discovery interpretable solutions, bridging symbolic logic generative power modern AI. Through rigorous experimentation, demonstrate LLM-generated achieve competitive offering greater interpretability compared previous approaches. Notably, our can rediscover existing manually designed provide insights into neural-network based solutions Programming-by-Example. results highlight potential only automate but also enhance transparency scalability design, safe deployment society.