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

The 30-second take

• What: A three-agent workflow links chatbot stated-preference surveys, structured processing, and weather-sensitive mode-choice prediction—with local LLMs matching ML baselines.
• Why it matters: It shows conversational data collection and multimodal LLM prediction can sit in one auditable loop with classical behavioral models.
• Who should care: Transport planners, mobility researchers, and teams building agentic survey-to-forecast pipelines.

What the paper actually did

Travel behavior work often splits digital data collection from predictive modeling. This study stitches them into a three-agent workflow: conversational data collection, structured processing, and behavioral prediction.

A chatbot ran an image-augmented stated-preference survey of student commuters across five predefined weather scenarios, producing 454 respondent-scenario observations. Weather associations were analyzed with a multinomial logit model; logistic regression and random forest served as machine-learning benchmarks.

Nine locally deployed LLMs (2–35B parameters) were tested under four zero-shot prompt-and-context conditions, then extended with persona, few-shot, and vision setups. Random forest hit 69.6% five-class accuracy; the best text-only zero-shot LLM reached 69.9% without task-specific fitting. Habitual travel information drove the most consistent gains; Expert framing generally beat Role-Play; few-shot helped several models after few examples. With the same weather images shown to respondents, the best vision configuration reached 71.5%.

What makes this disruptive

The default assumption is that survey design, discrete-choice modeling, and LLM prediction live in separate toolchains—and that local LLMs need fine-tuning to compete with classical ML on structured choice tasks. This work shows a single multi-agent loop can collect, structure, and predict, with zero-shot (and vision) LLMs matching or slightly beating random forest on five-class weather-sensitive mode choice.

Why it matters (outside the lab)

Weather-sensitive demand shapes transit planning, micromobility, and emergency operations. Today, rich multimodal preference data and trustworthy local LLM forecasts are still specialist luxuries—expensive surveys, siloed models, cloud APIs. An auditable agentic stack that collects stated preferences conversationally and predicts with on-prem models points toward demand tools that treat survey → process → forecast as one product, not three procurements.

Limitations & open questions

The sample is student commuters and five predefined weather scenarios (454 respondent-scenario observations)—not a full population or real revealed-preference panel. LLM wins are on five-class accuracy under controlled prompts; gains depend on habitual travel context, framing, and (for vision) specific models. Results are about stated preference and local zero/few-shot setups, not deployed operations forecasting.