Research evidence indicated that a specific type of augmented reality-assisted (AR-assisted) science learning design or support might not suit or be effective for all students because students' cognitive load might differ according to their experiences and individual characteristics. Thus, this study aimed to identify undergraduate students' profiles of cognitive load in AR-assisted science learning and to examine the role of their distinct profiles in self-efficacy together with associated behavior patterns in science learning. After ensuring the validity and reliability of each measure, a latent profile analysis confirmed that 365 Chinese undergraduates carried diverse dimensions of cognitive load simultaneously. The latent profile analysis findings revealed four fundamental profiles: Low Engagement, Immersive, Dabbling, and Organized, characterized as carrying various respective cognitive loads. The multivariate analysis of variance findings revealed different levels of the six AR science learning self-efficacy dimensions across profiles. Low Engagement students displayed the lowest self-efficacy among all dimensions. Organized students recorded better conceptual understanding and higher-order cognitive skills than Dabbling ones. Students with the Immersive profile had the highest science learning self-efficacy. The lag sequential analysis results showed significant differences in behavior patterns among profiles. Among them, profiles with social interaction, test, and reviewing feedback behavior had a significantly higher score for self-efficacy than those patterns mainly based on test learning and resource visits. This finding provides a unified consideration of students' diverse profiles and can inform interventions for effective design of AR-assisted science learning to match appropriate strategies to facilitate the science learning effect.