Superhuman AI Conquers Stratego With Hidden Info

Overview

A new artificial intelligence system has achieved superhuman performance in Stratego, a classic board wargame renowned for its complexity and massive amounts of hidden information. The research, published on arXiv in November 2025, demonstrates that an AI developed through self-play reinforcement learning and test-time search can now outperform the world’s best human players.

This milestone is significant not only for its competitive success but also for its economic efficiency. Previous attempts to master Stratego required industrial-scale budgets, yet this new system achieved its results for just a few thousand dollars. The findings suggest that general-purpose techniques for imperfect information can unlock superhuman capabilities without prohibitive costs.

A Historical Challenge in AI

Stratego has long been considered one of the most difficult benchmarks for artificial intelligence, comparable to games like poker or Go. The game features a double-blind structure where each player has ten pieces with hidden ranks, ranging from a lowly Scout to an indestructible Bomb, while the Commander remains the highest-ranking piece. This setup creates a environment where players must make strategic decisions while navigating incomplete knowledge about the opponent’s forces.

For years, researchers attempted to tackle this problem. However, earlier efforts to apply massive computational resources and deep learning techniques failed to reach the performance level of top-tier human competitors. The combination of a large branching factor and the necessity to reason about hidden states proved too demanding for prior architectures, which often relied on expensive, specialized training runs.

"Among these, Stratego stands apart as a case where such efforts failed to produce performance at the level of top humans."

The research team identified that the key lay not in brute force, but in developing more efficient self-play reinforcement learning frameworks and integrating robust search strategies during gameplay. By focusing on general approaches rather than game-specific hacks, they were able to bridge the gap between previous failures and human-level success.

Methodology and Technical Innovation

The study introduces a novel approach that combines two critical components: self-play reinforcement learning and test-time search. Unlike previous models that might rely heavily on hand-crafted features or static databases, this system learns purely through interaction with itself. The agents play millions of games against each other, gradually refining their strategies through trial and error in an environment defined by uncertainty.

A pivotal aspect of the methodology is the application of test-time search. During actual gameplay, the AI does not rely solely on its trained policy; it dynamically searches the game tree to evaluate potential moves and anticipate opponent actions. This allows the system to handle the exponential complexity of hidden information more effectively than static estimators alone.

The research paper, titled Superhuman AI for Stratego Using Self-Play Reinforcement Learning and Test-Time Search, was authored by Samuel Sokota, Eugene Vinitsky, Hengyuan Hu, J. Zico Kolter, and Gabriele Farina. Their work falls under the domains of Machine Learning and Artificial Intelligence, highlighting the interdisciplinary nature of solving complex strategic problems.

Key technical advantages include:

1. Generalizability: The approach is not tailored exclusively to Stratego’s rules but addresses broader challenges in imperfect information games.
2. Efficiency: The training process is computationally affordable, requiring only a few thousand dollars in resources.
3. Scalability: The framework demonstrates how reinforcement learning can be adapted for games with large state spaces and hidden variables.

Implications for the Field

The success of this AI marks a step change in the landscape of game-playing algorithms. It proves that superhuman performance in games with significant hidden information is achievable with modest resources, challenging the assumption that such feats require exorbitant industrial budgets. This democratization of high-performance AI could accelerate research into other complex domains where information is incomplete, such as economics, cybersecurity, and multi-agent negotiations.

Furthermore, the result revitalizes interest in classical board games as rigorous tests for AI capabilities. While games like Go and Poker have seen AI supremacy, Stratego’s unique blend of tactical depth and information asymmetry offered a resistant frontier. Breaking through this barrier validates the effectiveness of combining self-play with advanced search techniques.

The authors’ findings suggest that future research should focus on refining these general approaches to further reduce training costs and improve decision-making speed. As these methods evolve, they may be applied to real-world scenarios where agents must operate under uncertainty, potentially transforming how we model strategic interactions in both digital and physical environments.

This achievement serves as a testament to the power of modern reinforcement learning when paired with efficient search strategies, confirming that Stratego is no longer an impenetrable fortress for artificial intelligence.