A new AI model can tell which chess puzzles are actually worth your time - and most aren't.
Researchers built a system using offline reinforcement learning to evaluate and select chess puzzles based on real learning outcomes, not just tactical difficulty. They trained it on a year of user data pulled from major puzzle platforms, totaling 1.5 billion puzzle-solving histories. The model learned to estimate a puzzle's pedagogical value, essentially how much it helps a given player improve, rather than just how hard it is to solve. Testing focused on beginners rated 100 to 1000, the range where new players most often plateau, and the model's picks showed a measurable effect on players whose progress had stalled.
Sites like Chess.com and Lichess churn out puzzles automatically and hand them out using simple heuristics, treating puzzle quality as roughly uniform. This research says that assumption doesn't hold: mass-produced puzzles are often worse teaching tools than the hand-picked sets human coaches have relied on for decades. The idea reaches past chess, too. Any platform doling out algorithmically generated practice material, from language apps to coding drills, may be optimizing for volume instead of actual learning.
The researchers had expert players rate the puzzles their model favored, a reasonable check on a system that could otherwise just be gaming a proxy metric instead of teaching anything real.