Researchers from Carnegie Mellon, MIT, New York University, and Stanford developed an AI system named Ataraxos to master the game of Stratego. In a 20-game series held in July 2025, the model defeated world champion Pim Niemeijer with 15 wins, 1 loss, and 4 draws, achieving an 85% win rate.
During exhibition matches in August 2025, the system played against attendees at a Stratego championship, securing 38 wins and 2 losses for a 95% win rate. The AI also demonstrated high performance in the game variant Barrage Stratego and the card game Hanabi.
The training process utilized 16 Nvidia H100 GPUs for one week, followed by four days of refinement on four GPUs. The total computational cost remained under $8,000, which is significantly lower than the multi-million dollar expenses associated with previous attempts like DeepMind's DeepNash.
Ataraxos relies on self-play reinforcement learning and a generative belief network to model hidden information, as Stratego features over 10^66 possible configurations. Key contributors to the project include Gabriele Farina and Samuel Sokota.
The findings were detailed in a peer-reviewed paper published in Nature on September 30, 2026, following an initial preprint in November 2025. The results suggest that advanced AI milestones can be achieved with limited hardware resources through optimized algorithms.
