Stable Menus of Public Goods: AI-Enabled Progress

Loading / 加载中

英文摘要

The authors use an open problem from the 2025 ACM EC paper "Stable Menus of Public Goods" as a testbed to investigate the effectiveness of different AI-for-EconCS research workflows. They study three questions: whether including human intuition in the prompt helps, whether automated multi-turn interaction improves results, and whether a large language model (LLM) outperforms a first-year PhD student. The experiments indicate that prompting with human intuition encourages the LLM to exhibit better "taste," and multi-turn workflows are beneficial when they encourage "ambitious" problem-solving steps. When compared using an unpublished manuscript written by the senior authors before working with the PhD student, the LLM is found to be slightly less effective than the first-year PhD student. The paper offers workflow suggestions for integrating LLMs into economic research.

中文摘要

作者以2025年ACM EC会议论文“Stable Menus of Public Goods”中的一个公开问题为试验台，研究不同“AI for EconCS”研究流程的有效性。他们考察了三个问题：在提示中加入人类直觉是否有帮助、自动多轮交互能否改善结果，以及大语言模型（LLM）是否优于一年级博士生。实验表明，加入人类直觉的提示能促使LLM展现出更好的“品味”，而多轮工作流在鼓励“有野心”的解题步骤时是有益的。通过使用资深作者在与博士生合作前撰写的未发表手稿进行比较，发现LLM的效果略逊于该一年级博士生。论文为将LLM融入经济学研究提供了流程建议。

关键要点

Prompting with human intuition improves the LLM's research taste.

在提示中加入人类直觉可以提升大语言模型的研究品味。

Multi-turn interaction helps when it encourages ambitious problem-solving steps.

当多轮交互鼓励有野心的解题步骤时，该方式是有帮助的。

The LLM was slightly less effective than a first-year PhD student on the same open economics research problem.

在应对同一个开放的经济学研究问题时，大语言模型的效果略逊于一位一年级博士生。