Large language models enable intelligent decision-making for pavement maintenance

Road networks underpin end-to-end transportation systems, making timely and effective maintenance essential for service quality and public satisfaction. However, traditional Machine Learning- and Deep Learning-based maintenance decision methods rely heavily on large labeled datasets and expert-defined mappings between indices and maintenance technologies, requiring retraining when new technologies emerge and limiting their adaptability and automation potential. This study investigates the capability of large language models (LLMs) to perform both single-valued and set-valued maintenance decision-making solely through K -shot in-context learning, without requiring model architecture modification or parameter fine-tuning. Leveraging datasets from three provinces in China and cloud-based Application Programming Interfaces (APIs) provided by LLM vendors, a lightweight local method for maintenance decision-making is developed and validated. The experimental results demonstrate that, with as few as eight examples per category ( K ≥ 8), LLMs can achieve strong performance comparable to traditional machine learning models without parameter training, while maintaining high stability across different datasets. These findings provide early empirical evidence supporting the application of LLMs in maintenance decision-making scenarios, indicating that LLMs possess strong generalization capability in knowledge-mining–driven road maintenance tasks, with performance approaching that of existing expert models. This highlights the potential of LLMs to shift automated infrastructure maintenance from conventional numerical modeling toward natural language–driven reasoning, thereby enabling more data-efficient and adaptive decision-making systems.