An Agentic Approach for Active Data Collection, Travel Behavior Modeling, and Weather-Sensitive Demand Prediction

TL;DR

The study introduces a multi-agent system that unifies data collection, processing, and prediction for travel behavior modeling using a chatbot-based survey with weather scenarios. It evaluates LLMs across text-only, few-shot, and vision-augmented configurations, showing that visual context improves accuracy and that persona and habitual travel data enhance performance.

Why It Matters

This approach enables scalable, auditable, and context-aware data collection and prediction in transportation research, reducing reliance on isolated modeling stages. It demonstrates how multimodal LLMs can integrate real-world sensory inputs like weather images into behavioral forecasting, improving decision-making in urban planning and mobility services.

Key Findings

• A chatbot-administered survey with image-augmented weather scenarios collected 454 mode-choice responses from student commuters.
• The best vision-based LLM configuration achieved 71.5% accuracy on a five-class demand prediction task, outperforming text-only models.
• Habitual travel information and expert personas improved LLM performance, while few-shot prompting provided marginal gains after a small number of examples.