Can Large Language Models Capture Human Travel Behavior? Evidence and Insights on Mode Choice

Abstract

Understanding, modeling, and calibrating human behavior is crucial in travel demand modeling and transportation system analysis and planning. Traditional behavioral models rely on apriori assumptions and extensive data collection, which limit their ability to accurately capture traveler behavior and require significant resources for analysis and planning. Large language models (LLMs) offer a promising alternative by leveraging large-scale human-generated data and flexible input processing to emulate traveler behavior more accurately. However, the alignment of LLMs with human travel behavior remains unknown and could pose a critical challenge. Our paper addresses this problem by investigating the ability of LLM models to simulate travelers' mode choice behavior using a stated preference dataset. We find that zero-shot LLM models carry significant behavioral misalignment, which results in poor prediction accuracy. To address this, we implement the few-shot prediction approach and develop a travel behavior persona loading method. Both methods utilize training data to better align LLM behavior with that of human travelers. Our findings show that both methods significantly improve LLM's behavioral alignment and prediction accuracy. Our study contributes to transportation research by demonstrating the limitations and potential of LLMs in simulating travel behavior. Through extensive experiments, we show that while LLMs have the potential to transform travel demand modeling and forecasting, more refined alignment techniques are necessary to fully realize their capabilities as accurate proxies for human travelers.

Keywords: Travel behavior, Large language model, Behavioral alignment, Demand forecasting