Designing and evaluating LLM-driven coaching agents for VR public speaking practice.

Introduction: Public speaking practice in Virtual Reality (VR) provides an immersive and controllable setting, yet many existing systems offer limited adaptive guidance beyond exposure itself. Recent advances in Large Language Models (LLMs) create opportunities for designing coaching interactions that are personalized, context-aware, and socially supportive. Methods: We designed a VR public speaking system in which embodied virtual judges delivered spoken and textual, performance-contingent coaching based on automatic speech recognition, speech synthesis, gaze-related cues, and structured prompting with ERNIE Bot. We evaluated the system in a two-arm randomized controlled user study with 120 participants, with 60 participants assigned to the adaptive coaching condition and 60 to an attention-matched generic feedback control condition. Participants completed two rounds of VR speeches with repeated pre- and post-speech self-report assessments, followed by matched follow-up sessions at one week and one month. Results: Compared with attention-matched generic feedback, the adaptive coaching condition was associated with larger reductions in self-reported state anxiety and PRCA-derived public speaking anxiety appraisals across repeated speaking rounds. At follow-up sessions conducted in the same VR environment without additional coaching, lower self-reported anxiety remained observable in the experimental group. Qualitative feedback suggested that participants experienced the agents as supportive, useful, and lower-pressure, while also noting limitations in realism, speech prosody, and interaction richness. Discussion: The findings provide preliminary evidence that embodied LLM-based coaching can support anxiety-aware VR public speaking practice. However, because the study relied primarily on self-report, did not include physiological outcomes or blinded external performance ratings, and involved repeated exposure to the same VR context, the follow-up findings should be interpreted as maintenance within the VR practice setting rather than evidence of stable trait change, objective performance improvement, or real-world transfer. Future studies should examine whether these self-reported gains generalize to real-world speaking contexts and externally rated public speaking performance.