Title: Modeling Macroeconomic Expectations in Survey Experiments via LLM-Driven Economic Agents

Abstract: This study presents a novel framework designed to simulate macroeconomic expectations within survey experiments by employing LLM-based economic agents. These agents are engineered with multiple functional components capable of accessing personal attributes, existing prior expectations, and evolving external data. To verify the efficacy of this approach, we applied the framework to reconstruct three distinct survey methodologies that address various expectation types across different respondent demographics. The analysis demonstrates that the generated expectation distributions from the LLM Agents closely mirror human data, while also replicating qualitative patterns found in open-ended responses that align with human behavior. Our evaluation indicates that prior expectations are essential for accurately matching statistical distributions, while personal and external information sources are key to fostering human-like reasoning processes. These insights provide strategic direction for reducing the aggregate-level divergence in beliefs between generative AI and humans, while also clarifying the operational limits of the proposed framework.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC