LLMs Generate Realistic Travel Diaries Using Census Data, Matching Classical Models

A new approach harnesses large language models to automatically generate synthetic travel diaries for transportation simulations, demonstrating that LLMs can rival established statistical methods for this task. The research sidesteps the need for expensive proprietary household surveys by building AI-generated personas from freely available demographic and geographic datasets.

How the Method Works#

Rather than relying on traditional traffic surveys, researchers constructed detailed personas by combining data from the U.S. Census Bureau's American Community Survey and the Smart Location Database. They then used direct prompting to have an LLM synthesize complete travel diaries for these personas—including information about trip purposes, transportation modes, and distances traveled.

To assess the quality of these synthetic journeys, the team developed a novel scoring system that evaluates four key dimensions: the number of trips, timing intervals, trip purposes, and transportation modes. They validated their synthetic diaries against real travel records from Connecticut's state transportation study, carefully matching records across demographic groups.

Performance Against Conventional Approaches#

The LLM-based approach proved competitive with classical statistical methods. When the researchers compared LLM-generated diaries to diaries created using traditional models calibrated on the same validation dataset, both achieved similar overall realism scores—though with different strengths and weaknesses.

The LLM excelled at predicting trip purposes and showed more consistent patterns in mode selection. However, classical models still outperformed it at estimating the actual number of trips and activity durations. At the population level, LLM diaries demonstrated strong statistical representativeness, suggesting they capture broader travel patterns accurately even if individual trip counts sometimes differed.

Why This Matters#

The results indicate that LLMs could become practical tools for transportation planners who need large volumes of synthetic travel data without investing in costly household surveys. By requiring neither extensive training on proprietary data nor complex statistical calibration, the zero-shot LLM approach may accelerate transportation modeling and urban planning workflows. The study also introduces a reusable metric for evaluating the realism of synthetic travel diaries in future research.