Designing Open Dialogue Systems for LLM-Based NPCs

Abstract

This paper aims to propose a new way of hosting diegetic conversations between a player character and a non-player character (NPC) in the context of interactive entertainment (video games). Instead of a preset list of dialogue options for the user to choose from, the user is allowed to create their own dialogue for the player character, and a large language model (LLM) is used to create an NPC's response. By formatting this system correctly and managing both the unpredictable input of the player and the non-deterministic output of the LLM, an NPC can act in character according to a pre-written script and can dynamically change behavior during the interaction. This paper also goes into depth about using the low-level logical reasoning capabilities of LLMs to change game states, and the potential ways to structure, store and use semantic qualitative data using LLMs.