Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition
Claude Code with Opus 5.5 won 6 of 12 game ladders; the other 6 stayed ahead of every tested agent setup.
The paper defines Adversarial Heuristic Learning, where agents revise executable game policies without changing model weights. Its AAArena benchmark uses 12 adversarial games and 1,920 archived human programs under competition-like match budgets. Results were weaker when rule specifications were more complex. The authors also found gains from opponent selection, dense feedback, and learning from both the agents’ own replays and other players’ replays. HF Daily Papers' note
The paper defines Adversarial Heuristic Learning, where agents revise executable game policies without changing model weights. Its AAArena benchmark uses 12 adversarial games and 1,920 archived human programs under competition-like match budgets. Results were weaker when rule specifications were more complex. The authors also found gains from opponent selection, dense feedback, and learning from both the agents’ own replays and other players’ replays. HF Daily Papers' note
score 4