Reproduce “Cardiverse: Harnessing LLMs for Novel Card Game Prototyping” on OpenResearch

About this paper

Danrui Li, Sen Zhang, Sam S. Sohn, Kaidong Hu, Muhammad Usman, Mubbasir Kapadia

Cardiverse introduces an automated framework utilizing Large Language Models (LLMs) to prototype novel card game variants, integrating mechanics design, consistent code generation through iterative self-validation, and scalable AI for evaluating game dynamics. The framework successfully generates distinct game variants, produces highly consistent game code, and deploys cost-effective AI agents capable of exploring diverse gameplay scenarios.

Key insights

• LLM-guided graph-based mechanic extraction and clustering can effectively foster novelty in game design, producing variants that are intentionally distinct from existing games.
• An iterative, LLM-driven code generation process with gameplay record validation significantly improves the consistency between natural language game descriptions and executable game code.
• Ensembles of LLM-generated heuristic functions, optimized through competitive self-play, provide a scalable and cost-efficient alternative to direct LLM inference for every game decision, enabling robust game evaluation.

Summary by alphaXiv.