Ataraxos beats top Stratego player 15-1-4 in Nature paper

Ataraxos, an AI from a team at Carnegie Mellon, MIT, NYU and Stanford, beat Pim Niemeijer, described as the most decorated Stratego player of all time, 15 wins to one loss with four draws across 20 games. The method also won four 50-game series in Barrage Stratego against three multi-time world champions at statistical significance of P < 1.3 × 10⁻⁵, which the authors call "the first superhuman result in the game's history," reported in Nature on September 30.

The surprise is not the win but the bill. Training cost "a few thousand dollars" and consumed roughly 1/500th the compute of DeepNash, the prior Stratego agent published by DeepMind. At the 2025 World Championship, Ataraxos went on to a 95% win rate across 40 games.

The paper, titled "Scalable decision-making for games of imperfect information," pairs self-play reinforcement learning with a belief network that models where an opponent's hidden pieces actually are, plus a test-time search procedure that refines the policy. The authors describe the training recipe as "stronger regularization and more aggressive updates early in self-play, and weaker regularization and smaller policy updates late in self-play."

The design pattern travels across game types. In cooperative Hanabi, Ataraxos set a new state of the art across the two to five player variants, hitting 24.654 ± 0.007 in two-player play with 77.53% perfect 25-point games. In Dou Dizhu, the Chinese card game with more than 3×10^24 possible deals, it beat the previous state-of-the-art agent PerfectDou by a role-averaged 0.199 ± 0.015.