Using LLM-empowered virtual avatars in immersive virtual reality technology to enhance psychological perception and learning outcomes: the role of interaction methods, anthropomorphism types, and VR perspectives

Abstract Although immersive virtual reality technology has gained widespread acceptance among educators for its unique advantages-such as high immersion, interactivity, and engagement-that traditional educational methods lack, more empirical research is still needed to address how to optimize learners’ cognitive load and learning outcomes within immersive interactive environments. Building upon the theory of mind perception and the Stimuli–Organism–Response (S–O–R) model framework, this study proposes that Large Language Model (LLM) empowered virtual avatars can enhance learners’ mental perception levels during interactions. By optimizing the overall structure of cognitive load, this approach promotes improved learning outcomes. Our investigation employs one pilot study and two formal studies. Study 1 reveals the roles of perceived agency and experience in optimizing cognitive load structure and learning outcomes, constructing a moderated mediation model. Study 2 demonstrates how the interaction between anthropomorphism types (behavioral and morphological) and VR perspectives (first-person and third-person) manifests these positive effects during learners’ acquisition of different knowledge types. Focusing on VR environments, this study expands the literature and application scenarios of mind-perception theory and the S–O–R model framework through multi-level validation across two dimensions (psychological and physical), two disciplines (English and chemistry), and two knowledge types (declarative and procedural knowledge). It provides optimization directions and practical guidance for educators, VR product developers, and LLM technology designers, making the concept of technology adapting to learning-rather than learning adapting to technology-a reality.