Empowering Human-Like Non-Player Character Interactions in Virtual Reality Through Large Language Models

Abstract

This research paper explores the integration of Large Language Models (LLMs) into Virtual Reality (VR) environments to enhance human-like interactions with non-player Characters (NPCs) in professional training simulations. Current VR systems suffer from repetitive, scripted dialogues that lack naturalness and adaptability, reducing user engagement and learning outcomes. To address this, the study developed a proof-of-concept VR system using state-of-the-art LLMs, including GPT-4, Claude 3.5, and Large Language Model Meta AI (LLAMA), evaluated through the Massive Multitask Language Understanding (MMLU) benchmark. The Agile methodology was employed to iteratively refine the system based on user feedback, optimising NPC interactions for contextual relevance and realism. Results demonstrated significant improvements in naturalness, engagement, and context maintenance, with LLaMA-powered NPCs outperforming others in user acceptance testing. These findings underscore the potential of LLMs to revolutionise VRbased training by delivering lifelike, context-aware dialogues and provide a robust foundation for future research in AI-driven immersive environments.