数字技术与经济金融前沿论坛（第57期）

摘要：Large language models reason fluently about economics, but their generative reasoning is bounded by what they memorised at training time and by how a query routes their attention. We augment that reasoning with structured retrieval over the field’s own literature. Each paper in a curated corpus is extracted against PromptEcon, a fixed vocabulary of 9 categories and 51 elements linked by typed edges, and the per-paper instances merge by canonical concept into a single graph. HippoRAG-style retrieval over that graph augments the model’s context before it generates — re-ranking its parametric memory toward the relevant region and supplying post-cutoff work it never trained on. We run the method as a pilot study on the overlapping-generations literature and evaluate it with a blind expert-panel exam and an algorithm- and code-generation test. The approach is not specific to OLG: it is a reproducible recipe for turning any structured economics literature into a retrieval substrate for AI-augmented research.

主讲人介绍：

冯志钢，美国内布拉斯加大学-奥马哈经济系教授、浙江大学经济学院讲座教授，中国留美经济学会执行委员、导师计划主任、《经邦论策》专访栏目创办人。研究领域主要涉及宏观经济学、人工智能与机器学习、计算经济学。部分研究成果受到瑞士国家科学基金委、瑞士国家计算中心、美国自然科学基金多次资助，并先后刊发于International Economic Review、Quantitative Economics、 Review of Economic Dynamics、Economic Theory 等国际期刊。创办Bilibili“中南宏观”频道，传授高级宏观经济学、数量宏观，教材《机器学习与数量宏观经济学--Pytorch实用指南》2025年由北京大学出版社发行。