AI agents win 6 of 12 games against human programs in a long-running contest, study finds

Researchers at Tsinghua University introduce AAArena, featured on Hugging Face Daily Papers. It tests whether an AI agent can improve a game-playing program over many rounds while the underlying model's weights stay fixed: the agent reads the rules, picks opponents, studies replays and rewrites its own strategy code. The benchmark has 12 real adversarial games and 1,920 archived human programs, scored like a real competition. The best setup tested, Claude Opus 5.5 with Claude Code, earned 6 gold medals, but no configuration topped the other 6 human leaderboards, and agents did worse on games with more complex rules. The authors report that giving agents dense feedback and letting them choose opponents helped them improve. This is a research benchmark; code is on GitHub.