Exploring LLMs and MCTS for Emergent Narrative

Abstract

This thesis addresses the development of a hybrid combining method between LLMs and MCTS used in the NPC-engine. Our coupling scheme enables a bridge between language models unstructured data and MCTS structured data. This link between the two systems allows influencing agents behavior through a chat conversation system paired with a LLM-based reflection scheme. We analyzed our design with a user study comparing a LLMs chat-based game input system against a traditional game input method relying on sliders.