Augmented reality (AR) has been proposed to be an efficient tool for learning in construction. However, few researchers have quantitatively assessed the efficiency of AR from the cognitive perspective in the context of construction education. Based on the cognitive theory of multimedia learning (CTML), we evaluated the predesigned AR-based learning tool using eye-tracking data. In this study, we tracked, compared, and summarized learners' visual behaviors in text-graph-(TG-) based, AR-based, and physical model-(PM-) based learning environments. Compared to the TG-based material, we find that both AR-based and PM-based materials foster extraneous processing and thus further promote generative processing, resulting in better learning performance.e results show that there are no significant differences between AR-based and PM-based learning environments, elucidating the advantages of AR. is study lays a foundation for problem-based learning, which is worthy of further investigation.