LLMAKER - Consistent Game Content Creation via LLMs
A framework for evaluating LLM on iterative game content refinement
![llmaker-minimal-ui](/sites/default/files/styles/16_9_100/public/2024-07/funccall_fig.png?itok=ptqBQCLk)
The research paper and GitHub repository provided demonstrate the effectiveness of different prompting techniques for large language models (LLMs) in refining game content following designer requests. The framework presented is part of LLMAKER, a co-creation tool for designing video game content.
We propose a comprehensive framework for consistent video game content generation using LLMs, bridging the gap between creative vision and technical execution. Our framework, which utilizes function calling, is evaluated against various prompting techniques for generating dungeon crawler level layouts.
In this framework, the interaction between the designer and the system is entirely based on natural language, with the LLM translating user queries into properly formatted requests to a back-end system via function calling. We create a series of test cases that reflect real-world interactions of designers tasked with generating levels for a dungeon crawler video game.
Our results demonstrate that function calling outperforms other LLM-based methods for content generation in terms of prompt adherence and domain constraints satisfaction. Additionally, it consistently processes user requests within seconds, delivering updated content almost in real-time. This method showcases how function calling with LLMs can be efficiently implemented in a game design tool.