Agentic Workflow for Travel Data Collection and Behavior Modeling: LLM Chatbots, Multinomial Logit, and ML Benchmarks

Forum topic · 小凯 · 2026-08-22

Summary

A new arXiv paper (2608.20320) by Narges Ahmadi, Yubo Jiao, Jonatas Augusto Manzolli, and Jiangbo Yu proposes a three-agent workflow that integrates conversational data collection, structured data processing, and travel behavior prediction. An LLM-powered chatbot administered an image-enhanced stated preference survey of student commuters, collecting mode-choice responses under five preset weather scenarios and yielding 454 subject-scenario observations. Weather associations were analyzed with a multinomial logit model, while logistic regression and random forests served as machine learning baselines. The authors evaluated nine locally deployed LLMs (2B to 35B parameters) across four zero-shot prompting conditions plus extended role, few-shot, and visual configurations. A random forest reached 69.6% five-class accuracy, while the best text-only zero-shot LLM achieved 69.9% without task-specific fitting. Habitual travel information produced the most consistent gains, expert framing generally outperformed persona-based prompting, and the visual configuration achieved 71.5% accuracy.

Paper: An Agentic Approach for Active Data Collection, Travel Behavior Modeling Field: NLP Authors: Narges Ahmadi, Yubo Jiao, Jonatas Augusto Manzolli, Jiangbo Yu arXiv: 2608.20320

Overview

Travel behavior research increasingly combines digital data collection and predictive modeling, but these stages are typically developed and evaluated separately. This paper proposes a three-agent workflow that integrates conversational data collection, structured data processing, and behavior prediction.

Methods

A chatbot-managed, image-enhanced stated preference survey collected mode choices from student commuters under five preset weather scenarios, yielding 454 subject-scenario observations.
Weather associations were analyzed with a multinomial logit model.
Logistic regression and random forests provided machine learning baselines.
Nine locally deployed LLMs (2B to 35B parameters) were evaluated across four zero-shot prompting conditions, plus extended role, few-shot, and visual configurations.

Results

Random forest: 69.6% five-class accuracy.
Best text-only zero-shot LLM: 69.9% accuracy without task-specific fitting.
Habitual travel information produced the most consistent gains.
Expert framing generally outperformed persona-based (role-play) prompting.
Visual configuration achieved the best result at 71.5% accuracy.

*Automatically collected on 2026-08-22*