Cardiverse: Harnessing LLMs for Novel Card Game Prototyping

Abstract

The prototyping of computer games, particularly card games, requires extensive human effort in creative ideation and gameplay evaluation. Recent advances in Large Language Models (LLMs) offer opportunities to automate and streamline these processes. However, it remains challenging for LLMs to design novel game mechanics beyond existing databases, generate consistent gameplay environments, and develop scalable gameplay AI for large-scale evaluations. This paper addresses these challenges by introducing a comprehensive automated card game prototyping framework. The approach highlights a graph-based indexing method for generating novel game variations, an LLM-driven system for consistent game code generation validated by gameplay records, and a gameplay AI constructing method that uses an ensemble of LLM-generated action-value functions optimized through self-play. These contributions aim to accelerate card game prototyping, reduce human labor, and lower barriers to entry for game developers.

Create Game Variants

The Scalability Problem

Given a library of card game descriptions, while we could just prompt LLMs to design new games (i.e. different from existing ones) by feeding all descriptions, this approach creates overload to LLM reasoning. We need to break down and digest the data first.

Game Mechanic Graph

To break down game concepts, we use LLMs to convert game description texts to game mechanic graphs that describe the dependencies between atomic game mechanics.

Launch Demo

Mechanic Synthesis & Game Variation

For all mechanic graphs generated from the library, we cluster similar nodes and use LLMs to design new game mechanic instances. Then we can mutate any game input using new instances.

Synthesis Demo Variation Demo

Develop Game Environment

Code Generation Pipeline

An iterative refinement loop that aligns the game environment to the game description through LLM-driven code generation and validation.

How It Works

Cardiverse uses an iterative refinement loop to align game environment to game description:

LLM drafts initial code.
Code runs in custom game engine that produce gameplay records.
LLM checks records for rule violations, then refines code. Repeat until no violations are found.

Launch Demo

Build Gameplay AI

Heuristic Action Scoring

Instead of calling LLMs at each game decision, Cardiverse formulates the gameplay AI as a heuristic Python function that scores potential game actions based on game state features.

Ensemble Optimization

To efficiently optimize the policy behind the heuristic function, we break one big heuristic function into smaller, more manageable components. Then we directly optimize the ensemble of components.

Step-wise Inclusion

We adopt a step-wise-inclusion method, where we progressively add one more component to the heuristic ensemble as the final optimized policy.

Launch Demo

Game Demo

Try playing the demo game with Cardiverse-generated AI opponents! The source code for the demo game is available in the website branch of the Cardiverse GitHub repository.

BibTeX

@inproceedings{li-etal-2025-cardiverse,
    title = "Cardiverse: Harnessing {LLM}s for Novel Card Game Prototyping",
    author = "Li, Danrui  and
      Zhang, Sen  and
      Sohn, Samuel S.  and
      Hu, Kaidong  and
      Usman, Muhammad  and
      Kapadia, Mubbasir",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.1511/",
    doi = "10.18653/v1/2025.emnlp-main.1511",
    pages = "29735--29762",
    ISBN = "979-8-89176-332-6"
}