Probing the Augmented Reality Scene Analysis Capabilities of Large Multimodal Models: Toward Reliable Real-Time Assessment Solutions

Augmented Reality (AR) is transforming everyday experiences across domains like education, entertainment, and healthcare. As AR technologies become increasingly widespread, human-aligned and scalable AR quality evaluation is critical for optimizing immersive user experiences. This paper investigates the potential of Large Multimodal Models (LMMs) for automating AR quality assessment. We curate DiverseAR+, a new dataset of 1,405 scenes collected from diverse sources and environments, and use it to evaluate four commercial LMMs. Our results show that LMMs can perceive, describe, and judge AR content with promising accuracy. To deliver real-time, scalable, and robust AR quality evaluation, we propose a hybrid cloud-edge architecture that combines LMMs with traditional machine learning models, enabling scalable and real-time evaluation under diverse network conditions. We argue that task-tailored AR-LMM systems can make AR experience evaluation more efficient, adaptive, and user-centered.