Enhancing Electromagnetism Learning in Virtual Reality with an LLM-Prompted Pedagogical Agent Explaining Interactive...

Electromagnetism concepts are difficult for students to learn because many underlying phenomena are not directly observable. While interactive simulations can support learning, learners often struggle to interpret observed behaviors and connect them to underlying physical principles. This article introduces a pedagogical agent that delivers a lecture in immersive virtual reality (VR) and explains the behavior of an interactive charged particle simulation during the lesson. Explanations are generated using a prompt-engineered large language model (LLM) that interprets the learner's recent interactions with charged particles and the current simulation state. The system uses a static system-level prompt combined with a dynamic runtime prompt constructed from lecture context, learner actions, and simulation state. We implemented this approach in a VR classroom, where students attend a lecture on charged particles and Coulomb's law and interact with a simulation that visualizes electric field interactions at predefined points during the lecture. To examine the contribution of the simulation and the pedagogical agent's explanations, we conducted a between-groups study ($N$ = 60) comparing three conditions: VR lecture-only (Lec), VR lecture with simulation (LecSim), and VR lecture with simulation plus LLM-prompted explanations from the pedagogical agent (LecSimExpl). Results show that LecSimExpl produced significantly higher conceptual learning gains than Lec and higher knowledge transfer gains than both Lec and LecSim. LecSimExpl also increased germane cognitive load, intrinsic motivation, co-presence, perceived knowledge, and perceived anthropomorphism. Participants in the simulation conditions also spent more time interacting with the lesson and exhibited greater head-oriented attention toward the agent. These findings suggest that pedagogical agents that dynamically explain simulation behavior based on learners' ongoing simulation activity can improve learning performance, user experience, and perceptions of the pedagogical agent in VR learning environments.