It's All in the Personality: A Comparative Study of Real, Ideal, and Customized Virtual Instructors for AR Assembly Tasks

While embodied conversational agents driven by Large Language Models (LLMs) are emerging as valuable tools for instruction in Augmented Reality (AR), a key challenge lies in crafting their personalities to optimize both instructional efficacy and user engagement. To address this, we present findings from a within-subjects experiment that compared task performance and user experience with a LEGO assembly task. Participants received guidance from a real human instructor and three virtual counterparts, whose Big Five personality profiles were designed to be: (1) a direct replica of the real human, (2) an "ideal" profile ba sed on pedagogical research, or (3) customized by the participant. Our results reveal a critical trade-off: instruction from the real expert resulted in superior task efficiency and clarity; however, among the virtual conditions, instructors with idealized or user-customized personalities fostered significantly higher levels of user engagement and social presence compared to the virtual replica. Crucially, allowing users to customize their instructor's persona led to the strongest preference for future interaction. These findings underscore that personality is a fundamental component in the design of AI-driven instructors, providing empirical evidence for navigating the balance between task-oriented guidance and personalized, socially resonant user experiences.