Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition
Opus5.5 with Claude Code won 6 of 12 ladders, but half the human game competitions still held.
The paper formalizes Adversarial Heuristic Learning, where agents improve executable game policies without changing model weights. Its AAArena benchmark uses 12 adversarial games and 1,920 archived human programs under fixed match and evaluation budgets. The authors report weaker results on games with more complex rule specifications. Opponent choice and dense feedback helped agents improve, using both their own replays and other players’ replays. ArXiv · AI/CL/LG's note
The paper formalizes Adversarial Heuristic Learning, where agents improve executable game policies without changing model weights. Its AAArena benchmark uses 12 adversarial games and 1,920 archived human programs under fixed match and evaluation budgets. The authors report weaker results on games with more complex rule specifications. Opponent choice and dense feedback helped agents improve, using both their own replays and other players’ replays. ArXiv · AI/CL/LG's note
score 5