TraffiCoT-R: A framework for advanced spatio-temporal reasoning in large language models

Abstract

Spatio-temporal prediction investigates dynamic patterns in urban areas, including traffic flow, population movement, and infrastructure development and change. Most of the existing methods, however, require massive historical labeled data to train domain-specific models for a particular area of interest, which leads to inefficiency and reduced generalizability across different real-world environments. Such constraints call for models with high generalizability across different spatio-temporal applications. In this study, we introduce TraffiCoT-R, a prompt-based method that models the spatio-temporal relationships efficiently with LLMs. TraffiCoT-R integrates Spatio-Temporal Feature Importance Rotation (ST-FIR), a feature selection method, with a Feature Definition Module to enable contextualized reasoning and a multi-step iteration framework to enhance prediction. Such developments enable robust performance in zero-shot and few-shot configurations. Experiments on PeMS, NYCTaxi, NYCBike, and CHITaxi demonstrate that TraffiCoT-R outperforms state-of-the-art baselines on all the above-mentioned datasets in zero-shot configurations. These results show the potential of unifying LLMs with spatio-temporal frameworks for data-efficient, scalable city analysis.