rsi.watch

Changlian Ma and Limin Wang · 2026-10-24

Build, Don't Play: Self-Improving LLM Research Agents Develop a Hearthstone Battlegrounds AI That Rivals Top Players

“TavernAgent plays complete games on the live ladder by itself: its rating rose from 6,000 to about 15,000 and, on the public leaderboard's measure, its last 20 records average a placement of 1.60 with 80% first places, better than the server's ten highest-rated players on every statistic (2.34, 59%).”
Who
Changlian Ma and Limin Wang
Date
2026-10-24
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ imprecise. The top-player comparison rests on 20 leaderboard records for the agent and 200 for the ten players, each placement estimated from a rating change. The authors note the estimate is the same for every player. “The estimate is the same for every player, so it compares players fairly even though it is not a count of games”
  • ↓ indirect measure. The rating rise in the quote tracks games played as well as strength. The 1.60 against 2.34 compares each player's own leaderboard records, not results of games between them. “The rating measures accumulated wins as much as strength: it rises only as fast as games are played, and the ladder was played for two weeks.”
how we rate evidence
Replication
No rerun found. The paper links a repository and says TavernBench and TavernGym will be open-sourced. We found no rerun by another group. Searched up to 2026-10-28. all claims
Checked
2026-10-28, against the source page
Links
Source Paper Code

Also on the page

  • Six LLM research agents, run on GPT-6.1 Sol, GPT-6 Astra and Claude Opus 5.5, built TavernGym, a Rust reproduction of all 978 game objects of patch 36.6.3, and trained TavernAgent, an 18M-parameter policy, with PPO. Over 20 days they registered about 10,000 experiments and delivered 23 versions.
  • The comparison with top players uses each player's last 20 records on a public leaderboard tracker, which estimates a placement from each rating change; a record can cover one or several games. The authors' own game logs end at a rating of 14,007, and "The loop kept playing". In those logs, 392 games on days 5 to 18, the average placement without the 96 client-failure games is 3.22, with 16.4% first places. "The latest versions (TA-19–23) averaged 2.37 with 32.6% first places".
  • A human set goals, ruled on ambiguous mechanics and typed 418 messages, 0.6% of all messages. In live games, the AI named every operation and a human pressed the shortcut key for it.
  • The authors: "The rating measures accumulated wins as much as strength", and "The study covers one game and one patch, with calibrated library opponents in training."

In the self-improving AI catalog

  • AutoTavern: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that AutoTavern cites, oldest first. Under each: where that work sits on the map, then where AutoTavern sits (changes in bold), and what AutoTavern's own paper says about it. Its whole family on the citation map

Cites The AI Scientist 2024, AlphaEvolve 2025, Darwin Gödel Machine 2025, Weco AIDE² 2026

← Ledger Report an error