Geração de mapas para jogos Roguelike a partir de descrições textuais utilizando modelos de Linguagem de Larga Escala

Abstract

The roguelike game genre, known for its high replayability provided by PCG, often presents challenges regarding fine-grained control and the creation of content with specific developer intentions. Manipulating complex and rigid parameters in traditional PCG algorithms renders the design process unintuitive. This work explores an approach to overcome this limitation by proposing a system that uses LLMs to generate maps from natural language textual descriptions. Recognizing the inherent limitations of LLMs in spatial and geometric reasoning, the research developed a hybrid architecture that decouples responsibilities: traditional deterministic algorithms manage the structural topology and dungeon connectivity, while the LLM acts as a "creative agent" responsible for the semantic generation of narratives, enemies, and items. To materialize these elements, a pipeline was implemented integrating Structured Outputs and the RAG technique over a curated vector database of tiles, ensuring visual consistency and mitigating hallucinations. A functional game prototype was developed to interpret the generated assets and render playable levels. Validation, performed through Semantic Reconstruction and Coherence metrics across three distinct models, demonstrated that smaller models running locally offer performance comparable to large models. The results confirm the technical feasibility of associating natural language with concrete game assets, establishing a solid foundation for "text-to-map" creation systems that preserve user intention.