Reproduce “Game Development as Human-LLM Interaction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiale Hong, Hongqiu Wu, Hai Zhao
Researchers at Shanghai Jiao Tong University developed ChatGE, an LLM-powered framework enabling users to create custom games through natural language interaction. The system achieved 100% normalized interaction quality and 90.0% overall code accuracy in poker game development, significantly outperforming leading LLMs and demonstrating the efficacy of its data synthesis and progressive training strategies.
Key insights
- Structuring LLM output into distinct `P_script` (game script), `P_code` (code snippet), and `P_utter` (natural language interaction) processes enables effective multi-turn interactive development.
- An LLM-based data synthesis pipeline is crucial for overcoming data scarcity in specialized domains, allowing for training on diverse examples without extensive manual annotation.
- A progressive, multi-stage training strategy effectively aligns a general-purpose LLM to a complex, specialized role requiring both interactive guidance and accurate code generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.