Ataraxos AI beats top human champion at Stratego
Researchers have built an AI called Ataraxos that defeated the world's best Stratego player, proving that machines can master complex, long-form games of hidden information on a budget.
An academic coalition from Carnegie Mellon, MIT, NYU, and Stanford has developed Ataraxos, an artificial intelligence that conquered the board game Stratego. In a 20-game match, Ataraxos defeated Pim Niemeijer, arguably the greatest Stratego player in history, by a score of 15 wins to one, with four draws. The AI also dominated human players at the 2025 Stratego World Championship, winning 38 out of 40 matches. This milestone is significant because Stratego features a massive state space of over a decillion possible setups, hidden piece identities, and games that can stretch to 2,000 moves.
Unlike DeepMind's 2022 model DeepNash, which required an estimated $3 million to $4.5 million in computing power, Ataraxos was trained for just a few thousand dollars. It utilized 16 GPUs for one week, plus four additional GPUs for four days to train its specialized belief model. The system reached peak performance after playing 163 million self-play games—about 34 times fewer than DeepNash. Lead author Samuel Sokota and co-author Gabriele Farina achieved this efficiency by writing a custom simulator that processes millions of moves per second directly on graphics cards, making advanced game-theory research accessible to academic budgets.
Ataraxos succeeds by combining self-play reinforcement learning with real-time search. While DeepMind struggled to implement search in Stratego due to the vast decision tree, the researchers solved this by training a secondary neural network to estimate the opponent's hidden pieces based on their movements. This belief model allows the AI to sample plausible board states rather than calculating every possibility. For AI practitioners, this architecture proves that search-based planning can be successfully applied to massive, imperfect-information environments. The same framework has already beaten three world champions at Barrage Stratego, mastered the card game Hanabi, and defeated top bots in the Chinese game dou dizhu.
Beyond recreation, this breakthrough offers a blueprint for handling real-world scenarios characterized by high uncertainty and long time horizons. According to co-author Eugene Vinitsky, the techniques could be applied to war gaming to "see how a strong opponent might respond to what you do." The researchers are now focused on making the AI's decision-making process interpretable, as Ataraxos currently cannot explain its strategic choices.
This is our own summary of reporting by Ars Technica AI
