Geração Automática de Narrativas em Jogos de RPG Utilizando Modelos de Linguagem: Uma Abordagem Baseada em Inteligência Artificial

Abstract

This work investigates the use of large language models (LLMs) for the automatic generation of narratives in tabletop Role-Playing Games (RPGs), with a focus on simulating the role of a Game Master in Dungeons & Dragons 5th Edition. The study explores the effectiveness of combining Retrieval-Augmented Generation (RAG) and prompt engineering as strategies to improve factual consistency, rule adherence, and narrative quality without relying on fine-tuning.

Three experimental configurations were evaluated: (i) a zero-shot baseline model without external knowledge, (ii) a model augmented with RAG, and (iii) a model combining RAG with a structured prompt designed to guide behavior and response format. The zero-shot configuration was assessed using ROUGE metrics to establish a lower-bound performance, while all configurations were compared through a human evaluation involving experienced RPG Game Masters.

The evaluation employed a rubric based on five criteria: rule correctness, effective use of retrieved knowledge, absence of hallucinations, narrative clarity, and support for player agency. Results indicate that the use of RAG significantly improves performance across all criteria, particularly in rule application and reduction of hallucinations. Furthermore, the addition of structured prompt engineering leads to the highest overall performance, increasing both the quality and consistency of generated responses.

These findings demonstrate that combining external knowledge retrieval with explicit behavioral constraints is an effective approach for adapting LLMs to structured, rule-based, and interactive narrative domains. The proposed system shows strong potential as a tool to support or simulate Game Master behavior in RPG environments.