讲座主题：LLM Survey Framework: Coverage, Reasoning, Dynamics, Identification

主讲嘉宾：奚晋 中国科学院数学与系统科学研究院预测科学研究中心   助理研究员

讲座时间：2026年1月5日 周一15:00

讲座地点：沙河校区学院11号楼308会议室

嘉宾简介：奚晋，中国科学院数学与系统科学研究院预测科学研究中心助理研究员，加州大学圣地亚哥分校经济学博士。研究领域为计量经济学，主要涉及的研究领域包括机器学习的预测方法、高维非平稳时间序列、因子模型、政策学习、以及机制设计。

内容摘要：Assistant Researcher Jin Xi’s academic perspective focuses on proposing a novel survey research framework based on large language models (LLMs), arguing that this approach can substantially extend the boundaries of traditional human questionnaire surveys in both methodological and empirical dimensions. The paper points out that by introducing date restriction and internal consistency designs, LLM-based surveys can effectively avoid hindsight bias, enable retrospective surveys across historical periods, and achieve clean causal identification. Building on this foundation, LLMs can not only replicate the core empirical findings of human surveys at extremely low cost, but also directly reveal the underlying mechanisms of economic agents’ expectation formation through their reasoning capabilities. The authors further find that inflation expectations are mainly driven by two mechanisms—mean reversion and individual attention—and that the importance of these mechanisms exhibits pronounced state dependence across different inflation environments. Overall, the authors argue that LLM-based surveys provide a new research paradigm for expectation studies in economics—one that is difficult to achieve with human surveys but is more ideal in terms of research design.