TLDR: Ataraxos, a new AI, has achieved superhuman performance in the complex board game Stratego, a game known for its massive hidden information. Developed using self-play reinforcement learning and test-time search, Ataraxos defeated the world’s most decorated Stratego player, Pim Niemeijer, with an 85% win rate in a 20-game series. Remarkably, its training cost was only a few thousand dollars, significantly less than previous AI efforts that cost millions.
For decades, the classic board game Stratego has stood as a formidable challenge for artificial intelligence. Unlike games such as chess or Go, Stratego involves a massive amount of hidden information, making it incredibly difficult for AI systems to achieve human-level performance. Previous attempts to conquer Stratego required millions of dollars in computational resources, yet still fell short of the best human players.
Ataraxos: A New Era for Stratego AI
A groundbreaking new research paper, “Superhuman AI for Stratego Using Self-Play Reinforcement Learning and Test-Time Search,” introduces Ataraxos, an AI that has not only matched but vastly surpassed the skill of top human Stratego players. What’s even more remarkable is that Ataraxos achieved this feat with a training cost of only a few thousand dollars, a stark contrast to the millions spent on prior efforts. The paper was authored by Samuel Sokota, Eugene Vinitsky, Hengyuan Hu, J. Zico Kolter, and Gabriele Farina.
Ataraxos’s success stems from novel approaches in self-play reinforcement learning and test-time search, specifically designed to handle the immense hidden information inherent in Stratego. The system learns by playing against itself, constantly refining its strategies without needing human input.
How Ataraxos Mastered the Game
The AI’s architecture is built upon two interdependent self-play processes. One process focuses on the initial setup phase, where players secretly arrange their pieces on the board. The other handles the move selection during gameplay. Both processes utilize transformer neural networks, which are highly effective at processing complex data patterns.
Instead of relying on search to generate training data, Ataraxos directly samples from its policy networks, allowing for a high volume of games to be played and learned from. This approach, combined with a technique called “dynamically damped self-play reinforcement learning,” ensures stable and continuous improvement throughout its training.
A key innovation is Ataraxos’s “test-time search via update equivalence.” Before each move, the AI performs additional computations, essentially running a mini-reinforcement learning update. This process is aided by a “belief network” that estimates the likely configurations of the opponent’s hidden pieces, allowing Ataraxos to make more informed decisions even with incomplete information.
Unprecedented Performance and Efficiency
The true test of Ataraxos came in July 2025, when it faced Pim Niemeijer, widely regarded as the greatest Stratego player of all time, with four world championships and over 600 weeks as the #1 ranked player. In a 20-game series, Ataraxos delivered an astonishing performance: 15 wins, 4 draws, and only 1 loss. This 85% effective win rate is unprecedented at the highest levels of Stratego play, especially considering Pim had the advantage of adapting to Ataraxos’s fixed strategy over multiple games.
Following this historic victory, Ataraxos was showcased at the 2025 Stratego World Championship, where it achieved a 95% effective win rate across 40 games against a diverse range of players, further solidifying its dominance.
Beyond its performance, Ataraxos sets a new benchmark for computational efficiency. While a previous Stratego AI, DeepNash, cost millions to train, Ataraxos’s training run on Nvidia H100 GPUs cost less than $8,000. This dramatic reduction in cost makes superhuman AI for complex imperfect-information games far more accessible.
Also Read:
- Unmasking Hidden Roles: A New AI Framework for Social Deduction Games
- Small Model, Mighty Mind: VibeThinker-1.5B Redefines AI Reasoning with Cost-Effective Innovation
A Unique Play Style
Ataraxos’s play style exhibits notable differences from human players. It often employs aggressive setups with high-value pieces forward and uses “bombed-in” Flags (Flags surrounded by Bombs) frequently. Players observed that Ataraxos’s moves are difficult to interpret, making its pieces hard to read. It also shows a stronger preference for preserving Scouts and is willing to accept draws in the opening if the expected value of continuing is negative. Ataraxos takes gambles that humans might consider too risky and fights relentlessly when behind, employing stalling tactics.
The success of Ataraxos demonstrates that the combination of reinforcement learning and search can now effectively tackle strategic decision-making problems with massive amounts of hidden information. This opens up new possibilities for practical AI applications in various complex real-world settings where fast and accurate simulators can be developed. You can read the full research paper here: Superhuman AI for Stratego Using Self-Play Reinforcement Learning and Test-Time Search.
