Researchers build AI that finally beats the world's best Stratego player
In short: A team from Carnegie Mellon, MIT, NYU, and Stanford built an AI called Ataraxos that beat Pim Niemeijer, considered the best Stratego player of all time, 15 games to one with four draws. Unlike DeepMind's earlier DeepNash, Ataraxos uses a belief model to estimate hidden opponent pieces and search ahead before each move, and was trained on just 16 GPUs for about a week at a cost of a few thousand dollars, versus an estimated $3-4.5 million for DeepNash's training run. The AI also won 38 of 40 games against challengers at the 2025 Stratego World Championship and generalized to other games like Barrage Stratego, Hanabi, and dou dizhu.
This summary was generated automatically by AI from Ars Technica's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Ataraxos beat world champion Pim Niemeijer 15-1 with 4 draws
- 2Added a belief model enabling search/lookahead in a game with massive hidden information, which DeepNash lacked
- 3Trained on 163 million self-play games using 16 GPUs for a week (plus 4 GPUs for 4 days for the belief model)
- 4Estimated training cost a few thousand dollars vs. $3-4.5 million for DeepMind's DeepNash
- 5Won 38 of 40 games against challengers at the 2025 Stratego World Championship
- 6Same architecture also mastered Barrage Stratego, Hanabi, and dou dizhu
Why it matters
This demonstrates that imperfect-information, long-horizon strategic reasoning (bluffing, hidden state estimation) can be tackled efficiently and cheaply via a belief-model-plus-search approach, a technique that could generalize to negotiation, planning, and adversarial decision-making domains beyond games.
Sources
- Ars TechnicaMediaPrimary sourceOriginal article →1 Oct 2026, 19:28
- Published by source
- 1 Oct 2026, 19:28
- Found by our system
- 2 Oct 2026, 19:51
- Summary generated
- 2 Oct 2026, 20:10
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy