Implementasi Large Language Model (LLM) dalam Pembuatan Map Procedural pada Game Turn-Based Berbasis Grid

Abstract

Maps are essential elements in grid-based turn-based games because they influence player strategy and overall gameplay experience. Manual map creation is time-consuming and often produces limited variation. This study aims to implement a Large Language Model (LLM) for procedural map generation to automate the map design process. The system was developed using the Unity game engine integrated with the LLM for Unity package from Undream AI and the local model Llama 3.2 3B-Instruct-Q4_K_M. Three main approaches were proposed: (1) a single LLM that directly generates numerical maps, (2) two separate LLMs in which the first generates a binary map and the second applies biome decoration, and (3) a combination of one LLM with a procedural decoration program in Unity. Each approach was tested 20 times to evaluate map size consistency, passable tile ratio, and connectivity using the A* algorithm. The results show that the LLM still faces limitations in understanding spatial structures and following numerical constraints consistently. The second approach achieved the highest map size consistency of 45%, while the third approach obtained the best connectivity of 80%. These findings indicate that LLMs are not yet capable of reliably generating stable and specification-compliant procedural maps. However, they demonstrate potential as a supportive component for future research and development in dynamic content generation within the gaming industry.