The two sources are identical in content, so they agree on every point. Stratego, a game of hidden armies and bluffing, had resisted AI mastery even as chess, Go, and poker fell. Now a team from Carnegie Mellon, MIT, NYU, and Stanford has built Ataraxos, which beat four-time world champion Pim Niemeijer 15 games to one, with four draws, over a three-week online match. At the 2025 Stratego World Championship, it won 38 of 40 games against challengers.
The key innovation is a belief model—a second neural network that guesses the opponent's hidden piece arrangement based on their moves. This lets Ataraxos sample plausible setups and play out candidate moves, instead of enumerating the over-a-decillion possible configurations. It also uses a training schedule that makes large strategy changes early and small ones later, avoiding the loops that plagued earlier self-play approaches like DeepMind's DeepNash.
Ataraxos's playstyle is notably calm, bluffing its way back from near-certain losses. It also influenced human play, surprising players with setups like placing the flag behind just two bombs. The entire system was trained on 16 GPUs for a few thousand dollars, a fraction of typical AI budgets, showing that efficient algorithms can beat brute force in imperfect-information games.