LLMs mimic human route choice biases, enabling scalable behavioral modeling

A new study led by Jiangtao Han and colleagues, published on arXiv in July 2026, investigates whether large language models (LLMs) can reproduce systematic human biases in route choice without requiring explicit prospect-theoretic parameters. Traditional behavioral modeling relies on cumulative prospect theory (CPT) to capture deviations from rational decision-making, but applying CPT at scale depends on calibrating individual-level parameters through costly surveys and experiments. The researchers designed a behavioral evaluation framework using route choice as a representative scenario, comparing LLM-generated decisions against established human behavioral patterns predicted by CPT. Their experimental results demonstrate that LLMs are capable of reproducing non-rational choice biases, including loss aversion, reference dependence, and probability weighting, consistent with prospect-theoretic effects under uncertainty.

These findings carry significant implications for next-generation agent-based simulations and AI-driven behavioral research. By leveraging generative AI models, researchers can bypass the bottleneck of individual parameter specification, enabling more scalable and realistic modeling of human decision-making in transportation, economics, and social systems. The study suggests that LLMs may serve as a promising foundation for large-scale simulations that better reflect human irrationality, potentially improving predictions in areas like traffic flow, urban planning, and policy design. While further validation is needed across diverse contexts, this work opens a new pathway for integrating AI with behavioral economics.