Nanjing University's AutoTavern Beats Top Hearthstone Players With 91% Decision Accuracy
An LLM research team built a Hearthstone Battlegrounds simulator and trained a reinforcement learning agent that beats the top 10 ranked humans on the live ladder.
- Nanjing University's AutoTavern used six LLM agents to build a Hearthstone Battlegrounds AI from scratch in 20 days.
- The resulting TavernAgent reached ~15,000 ladder rating, averaging 1.60 placement with 80% first-place finishes.
- It beat the server's top 10 ranked human players on every statistic (humans averaged 2.34 placement, 59% firsts).
- On TavernBench, TavernAgent hit 91% accuracy at 0.03s per decision; frontier LLMs maxed out at 63% and 26s.
- Agents share typed research memory and pass claims through verification gates; a human only sets goals and rules.
- Code, simulator and benchmark open-sourced at github.com/MCG-NJU/AutoTavern.
Researchers at Nanjing University report in a preprint that six LLM agents built a specialist reinforcement learning system for Hearthstone Battlegrounds. The resulting policy reached a live rating of about 15,000 and outperformed the strongest frontier language model on the authors’ decision benchmark.
Hearthstone Battlegrounds is Blizzard’s eight-player auto-battler. Players spend gold in a randomly stocked tavern to buy, sell, upgrade, and position minions. Combat resolves automatically, and players are eliminated until one remains. Hidden information, random shops, timed turns, and nearly 1,000 interacting game entities create a large decision space.
The project, called AutoTavern, assigns research and engineering tasks to language models. A transformer policy trained with reinforcement learning handles gameplay. This architecture moves expensive reasoning into development and leaves live decisions to a smaller model that can respond in milliseconds.
Why direct play misses the clock
A language model controlling the game must parse a changing board, plan across turns, account for randomness, and select legal actions quickly. On the paper’s benchmark, the strongest tested frontier model answered 63% of play tasks correctly and sometimes required 26 seconds for a decision. The benchmark applies a one-second budget suited to repeated actions during timed turns.
TavernAgent, the trained specialist, answered 91% of the same tasks correctly and averaged 0.03 seconds per policy decision. Its advantage comes from repeated training in a purpose-built simulator, a compact action interface, and a model architecture designed around game entities.
Six agents, three feedback loops
AutoTavern’s six LLM research agents share a structured memory and operate across three timescales. Training experiments run over minutes, research cycles run over hours, and live-ladder feedback arrives over days. The pipeline covers rule discovery, simulator construction, opponent generation, policy training, evaluation, and corrections based on live games.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
