StorySculptor: Offering a personalised text-based gaming experience using Large Language Models (LLMs)


This study explored the integration of large language models (LLMs) into the realm of interactive fiction (text-based gaming) and aimed to bridge natural language processing techniques with the domain of storytelling. The study dives into the current state-of-the-art applications of LLMs and their capacity to generate narratives in real-time gaming environments. The paper further highlighted the implementation steps and focused on proposing a novel application of LLMs: developing a game agent designed to act as an active participant in interactive text games such that it is capable of adapting narratives based on player input and contributing to a more personalized gaming experience. By developing a novel system, the research contributes to the field through the following key advancements: (1) the creation of a novel dataset, generated using GPT-4, specifically designed to fine-tune LLMs for interactive gaming scenarios, and (2) the successful fine-tuning of the Mistral 7B instruct model, enabling dynamic game narrative generation.