Research & models Oct 27, 2026

Trend · preprint, October 24

Agent beat top-10 players at Hearthstone; authors built test

A Nanjing University preprint by Changlian Ma and Limin Wang had an LLM agent spend 20 days building a simulator for Hearthstone Battlegrounds and training a specialist agent, whose average placement over its last 20 games beat the server's top-10 players. The authors conclude that LLM agents do far better as builders than as players. Hearthstone Battlegrounds is an eight-player auto-battler with hidden information and close to a thousand interacting cards (arXiv 2610.33845, 2026-10-24). According to the abstract, the research agent built the simulator TavernGym, opponents and strategies within 20 days and delivered 23 versions of TavernAgent trained with reinforcement learning. Over its last 20 games, TavernAgent averaged a placement of 1.60 and finished first 80% of the time. The server's top-10 players averaged 2.34 with a 59% first-place rate. The abstract doesn't say how many games or what period the top-10 figures cover. On a 960-question matchup benchmark the authors built, TavernAgent was right 91% of the time, against at most 63% for frontier LLMs playing directly. TavernAgent took 0.03 seconds per decision, against 26 seconds for an LLM. This is one game, with an environment and benchmark the authors built, and 20 games is a small sample. People still set the goals and ruled on ambiguous rules, and how much they intervened isn't reported. Our reading: agents may be most valuable in the offline building phase. Speed per decision decides whether an agent can play live. The trained specialist took 0.03 seconds, and a frontier LLM playing directly took 26. The latency gap may also come from the different kinds of models, and the abstract doesn't separate the two. How we checked →