Pavement management and maintenance decision-making based on multi-agent collaboration using large language model

Pavement performance deteriorates over time, which affects traffic capacity and comfort. Therefore, timely and effective management and maintenance are crucial. However, traditional machine learning requires highly structured data for pavement management and maintenance decision-making; large language models (LLM) struggle to perform precise numerical calculations. To ensure the reliability of maintenance decisions while preserving the flexibility of LLMs in processing natural language, this study developed a method for pavement management and maintenance decision-making based on multi-agent collaboration using LLM. This method primarily consists of a primary agent and four secondary agents. The primary agent parses the prompt to plan task paths and determine parameters for the secondary agents, and outputs the final result after a consistency check. The experimental results show that all four general-purpose LLMs produced accurate outputs for path planning and parameter determination in this task. Among them, the models based on the Gemini architecture achieved the highest accuracy, at 93% and 88%, respectively. When different prompts are fed to the primary agent, the results of the maintenance decisions following coordinated execution by the secondary agents align closely with the logic of the macro-level prompts. This demonstrates the agent's ability to perform deterministic closed-loop reasoning, maintaining internal consistency in its engineering logic. Therefore, this method is expected to provide a highly generalizable architectural paradigm for next-generation pavement management system (PMS) based on generative artificial intelligence. © 2026 The Author(s).