LLMs in Eduverse: LLM-Integrated English Educational Game in Metaverse

Kantinan Plupattanakit, Pratch Suntichaikul, Pittawat Taveekitworachai, Ruck Thawonmas, Jeremy White, Kingkarn Sookhanaphibarn, Worawat Choensawat

Abstract: This demo paper presents a new educational game on the metaverse platform Roblox, integrating large language models (LLMs) for dynamic content generation. The game is suitable for up to five players and includes a single-player mode. It is based on the word-guessing genre where players guess a target word from a list of ten options using hints and feedback on the semantic closeness of their guesses. Utilizing GPT-3.5 Turbo, we generate semantic similarity scores and hints, proposing a new prompt to instruct LLMs for this purpose. To increase student engagement, our game leverages Roblox’s scoring and spotlight systems. We address stochastic behavior and unreliable output in LLM integration by presenting a series of prompts and a useful script. Our contributions include suggestions for LLM integration in educational games and a prototype demonstrating these concepts.