Summary
This post summarizes an arXiv paper (2608.20320) by Narges Ahmadi, Yubo Jiao, Jonatas Augusto Manzolli, and Jiangbo Yu, published August 22, 2026, which integrates conversational data collection, structured data processing, and behavioral prediction into a single three-agent workflow for travel behavior research. A chatbot-managed, image-enriched stated preference survey collected mode choices from student commuters across five predefined weather scenarios, yielding 454 subject-scenario observations. Weather associations were analyzed with multinomial logit models, while logistic regression and random forests provided machine learning baselines. The authors evaluated nine locally deployed large language models (2B to 35B parameters) under four zero-shot prompting conditions plus extended persona, few-shot, and visual configurations. A random forest achieved 69.6% five-class accuracy, while the best text-only zero-shot LLM reached 69.9% without task-specific fitting. Habitual travel information produced the most consistent gains, expert framing generally outperformed role-play prompting, and a visual configuration achieved 71.5% accuracy. The work demonstrates that agentic LLM pipelines can unify survey collection and behavioral prediction.
Paper Overview
Research area: NLP
Authors: Narges Ahmadi, Yubo Jiao, Jonatas Augusto Manzolli, Jiangbo Yu
Published: 2026-08-22
arXiv: 2608.20320
Summary
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 behavioral prediction.
A chatbot-managed, image-enriched stated preference survey collected mode choices from student commuters across five predefined weather scenarios, producing 454 subject-scenario observations. Weather associations were analyzed using multinomial logit models, while logistic regression and random forests provided machine learning baselines.
The authors evaluated nine locally deployed LLMs (2B to 35B parameters) under four zero-shot prompting conditions plus extended persona, few-shot, and visual configurations.
Key findings
- Random forest achieved 69.6% five-class accuracy.
- The best text-only zero-shot LLM reached 69.9% accuracy without task-specific fitting.
- Habitual travel information produced the most consistent performance gains.
- Expert framing generally outperformed role-play prompting.
- The visual configuration achieved the best result at 71.5% accuracy.
---
*Auto-collected on 2026-08-22*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178633792