AutoTavern Released as Hearthstone Battlegrounds Simulator and AI Benchmark

AutoTavern has been released as an open-source repository featuring TavernGym and TavernBench, designed to provide a reproducible environment and decision benchmark for AI research involving Hearthstone Battlegrounds. Authored by Changlian Ma and Limin Wang from Nanjing University and Shanghai AI Laboratory, the project accompanies the research paper titled "Build, Don't Play: Self-Improving LLM Research Agents Develop a Hearthstone Battlegrounds AI That Rivals Top Players."

According to the documentation, the release provides code and data under the MIT License, though specific items like the trained TavernAgent weights, training code, and benchmark item generators are not included. Users can verify the original announcement and access instructions directly through the project's public GitHub repository.

Core Components

• TavernGym: A simulator targeting patch 37.0.2 for solo Battlegrounds gameplay. It exposes player-visible states, exact legal action masks, seeded save and restore functions, 245 discrete actions, 3,227 numeric features, and 124 card indices per observation.
• TavernBench: A decision benchmark containing 960 total tasks, which includes a 368-task evaluation set, saved simulator states, and shared prompt formats with common scoring code.

Practical Implications and Setup

Researchers and developers can use TavernGym to simulate complete games, connect learning algorithms, or evaluate model judgments against fixed tasks. The simulator requires Python 3.10 or newer along with a Rust toolchain. TavernBench scoring requires only Python's standard library without needing a simulator build, model access, or API key.

This independent, non-commercial research project is unaffiliated with Blizzard Entertainment. Hearthstone and Hearthstone Battlegrounds are trademarks of Blizzard, and no game client files or official artwork are distributed.