TravelReasoner: Reasoning-Augmented Travel Survey Simulations with Large Reasoning Models

ICLR 2026 Conference Submission17322 Authors

Keywords: Travel Surveys Simulation, Large Langange Models, LLM Reasoning, LLM Application

Abstract: Travel survey plays a central role in a wide range of applications, such as urban planning and traffic management. Large language models (LLMs) have recently demonstrated huge potential in simulating human behaviours. However, previous works in travel survey simulation research have primarily focused on tuning LLMs to directly fit travel survey data, overlooking the underlying reasoning process behind human decision-making. The emergence of large reasoning models (LRMs) has achieved tremendous success in solving complex tasks, offering unique opportunities to simulate a realistic travel survey through LLM reasoning. In this paper, we introduce \textbf{\textit{TravelReasoner}}, a novel framework that enhances travel survey simulations by integrating the reasoning capabilities of LRMs. We construct \textbf{\textit{Chain-of-Trips}} from publicly available trip-chain records in the National Household Travel Survey (NHTS). This dataset captures the step-by-step reasoning process behind real-world travel decisions. To improve the accuracy and rationality of LRMs' in-domain reasoning, we propose a post-training pipeline via curriculum learning. Experiments demonstrate that TravelReasoner substantially outperforms strong baselines, location consistency improved by 6.8\% and time consistency improved by 4.1\%, while producing interpretable intermediate reasoning traces that enable transparent and explainable simulations. Our findings highlight the promise of LRMs for complex decision modeling and open new opportunities for applying NLP to urban systems.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 17322