A Multi-Agent Workflow: Conversational Data Collection, Travel Behavior Modeling, and Weather-Sensitive Demand Forecasting

Background and Context

Travel behavior research has traditionally split into two independent stages: collecting preference and choice data from travelers through questionnaires or interviews, then modeling and forecasting that data with statistical or machine learning methods. These phases were usually designed and evaluated separately, so the way data was generated rarely aligned with what downstream prediction models actually needed. This paper addresses that gap by proposing a three-agent workflow that integrates conversational data collection, structured data processing, and behavior prediction into a single auditable, reproducible process.

The authors focus on a concrete, practical problem: weather-sensitive transportation demand forecasting. They ask whether conversational, image-enhanced data collection can efficiently capture mode-choice decisions, and whether traditional behavioral modeling, machine learning, and multimodal large language models can operate together within one framework. The goal is to validate that multi-agent collaboration offers real value for travel research rather than serving as a technical exercise.

Deep Analysis

The data collection phase used a chatbot-administered stated preference questionnaire. Student commuters were shown five preset weather scenarios and recorded their mode-choice under each, yielding 454 respondent-scenario observations. The conversational interface and image enhancement were intended to increase the realism and immersion of the weather contexts, improving the quality of the collected preferences.

In the processing phase, weather associations were modeled with a multinomial Logit model, a classic technique for characterizing choice probabilities across scenarios. Logistic regression and random forest served as machine learning baselines. For the large language model evaluation, nine locally deployed models ranging from 200 million to 3.5 billion parameters were tested under four zero-shot conditions, then extended to personality settings, few-shot prompting, and vision configurations. This graduated strategy examined how different prompt engineering techniques affect performance without relying on task-specific fine-tuning.

Industry Impact

On the five-class mode-choice task, random forest achieved 69.6% accuracy, while the best text-only zero-shot large language model reached 69.9% with no task adaptation. Prompt engineering experiments showed that habit-travel information produced the most consistent gains, expert-perspective frameworks generally outperformed role-playing frameworks, and personality information mattered most when habit data was absent. Few-shot prompting improved several models, with benefits stabilizing after only a few examples.

Notably, when models were presented with the same weather images shown to respondents, the best vision configuration reached 71.5% accuracy, indicating that visual context provides additional predictive information for some models. These findings demonstrate that prompt strategies and multimodal inputs perform differently depending on the information provided and the model's characteristics. For the open-source community, the nine locally deployed models offer guidance for model selection under resource constraints, while the conversational, image-enhanced collection approach lowers the cost and barrier of administering surveys.

Outlook

The paper's value extends beyond the accuracy figures, offering an auditable, coordinated multi-agent workflow that unifies conversational surveys, structured data processing, traditional behavioral modeling, machine learning, and multimodal prediction. This end-to-end collaborative approach provides a reusable template for future travel behavior research.

The discovery that visual context adds predictive power suggests researchers should weigh multimodal information more heavily in behavioral forecasting. It also offers practical experience for coordinating human data generation with artificial intelligence modeling, pointing toward frameworks where data collection and prediction are designed together rather than in isolation.