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

Analysis

Why This Paper Matters

This paper addresses a critical gap in travel behavior research: the disconnect between data collection and predictive modeling. Traditionally, surveys are conducted and models are built separately, leading to inefficiencies and potential biases. By proposing a three-agent workflow that integrates conversational data collection, structured processing, and prediction, the authors present a unified framework that could streamline research and improve accuracy. The use of LLMs for both data collection (via chatbot) and prediction is novel, especially in the context of weather-sensitive travel demand, which is crucial for urban planning and transportation management.

The study also highlights the growing role of LLMs in behavioral research, moving beyond text generation to actual predictive tasks. The evaluation of nine locally deployed LLMs across various configurations provides valuable insights into their capabilities and limitations in a domain that requires understanding of human behavior and contextual cues. This is particularly relevant as LLMs become more accessible and are increasingly used in applied settings.

Technical Contributions

The paper makes several key technical contributions:

• Three-agent workflow: A structured approach that separates data collection, processing, and prediction, allowing for modularity and auditability.
• Image-augmented stated-preference survey: A novel method that uses weather images to elicit mode choices, providing richer data than text-only surveys.
• Comprehensive LLM evaluation: Nine LLMs ranging from 2 to 35 billion parameters are tested under four zero-shot conditions, plus persona, few-shot, and vision-based extensions.
• Comparison with traditional ML: The study benchmarks LLMs against multinomial logit, logistic regression, and random forest, providing a clear performance comparison.
• Insights into prompt engineering: Findings on the impact of habitual travel information, expert framing, and few-shot examples offer practical guidance for LLM-based prediction.

Results

The results show that LLMs can compete with traditional ML models in travel mode prediction. Random forest achieved 69.6% accuracy, while the best text-only zero-shot LLM reached 69.9%, slightly outperforming it. The vision-based configuration, which used the same weather images shown to respondents, achieved 71.5% accuracy, demonstrating that visual context adds predictive value. Habitual travel information consistently improved performance, and expert framing outperformed role-play. Few-shot prompting improved accuracy for several models, with gains stabilizing after a small number of examples. These findings suggest that LLMs, especially multimodal ones, can be effective tools for travel behavior prediction.

Significance

This research has broader implications for the AI field, particularly in human-centered applications. It shows that LLMs can be integrated into end-to-end workflows that involve data collection and prediction, reducing the need for separate, specialized models. The auditable multi-agent approach could be applied to other domains, such as healthcare or marketing, where understanding human decisions is key. Additionally, the study provides evidence that visual information can enhance LLM predictions, supporting the development of multimodal AI systems. As LLMs continue to evolve, this work paves the way for more sophisticated and context-aware behavioral modeling.