LLMs reproduce human route choice biases consistent with cumulative prospect theory

This paper investigates whether large language models (LLMs) can replicate systematic human behavioral biases in route choice without explicit calibration of Cumulative Prospect Theory (CPT) parameters. The authors designed a behavioral evaluation framework to compare LLM-generated decisions against established human patterns predicted by CPT.

• Conventional methods for specifying individual-level CPT parameters rely on surveys and controlled experiments, which are difficult to generalize and fail to capture decision diversity.
• Experimental results demonstrate that LLMs reproduce non-rational human choice biases and exhibit decision behaviors consistent with prospect-theoretic effects under uncertainty.
• The findings suggest generative AI models provide a scalable alternative for modeling human decision processes for large-scale agent-based simulation.

These results offer a promising foundation for next-generation large-scale agent-based simulation and AI-driven behavioral research.