Skip to main content
Back

Stanford's AI Learns Which Chess Puzzles Actually Make Beginners Better

Stanford researchers mined 1.5 billion puzzle-solving sessions with offline reinforcement learning to identify chess puzzles that actually teach beginners something.

Paper, Full Access
  • Stanford team won an outstanding paper award at RLC 2026 for offline RL applied to chess puzzle curation.
  • They mined 1.5 billion puzzle-solving histories from online chess platforms over one year.
  • Existing platforms auto-generate puzzles and use crude Elo-matching heuristics to recommend them.
  • Offline policy evaluation shows biggest gains for beginners rated 100 to 1000 Elo.
  • Learned recommender specifically helps players whose skill growth had stagnated.
  • Titled chess players independently rated the discovered puzzles as high pedagogical quality. Paper on arXiv.

Anyone who has ground through a few hundred auto-generated Lichess puzzles knows the feeling: some positions teach you a real pattern, others feel like random tactical noise. A new paper from Stanford, which just picked up an outstanding paper award at RLC 2026, tries to put a number on that intuition. The authors trained an offline reinforcement learning system on roughly 1.5 billion solving histories to figure out which puzzles are pedagogically useful and which are filler.

Drinking from the puzzle firehose

Chess coaches have long argued that hand-picked puzzles teach beginners far more than machine-generated ones, but curating tactics by hand is slow, expensive, and doesn't scale. Chess.com and Lichess push millions of automatically generated positions at their users, and the pedagogical value of any individual one is largely unknown.

The recommendation side is equally crude. Platforms typically match a puzzle's Elo rating to the player's puzzle Elo. That keeps difficulty roughly calibrated but says nothing about whether a given position will actually improve the student.

Turning learner history into a reward signal

The core idea in the paper is to treat puzzle recommendation as a sequential decision problem and solve it from logs rather than by running live experiments on real learners. Using a full year of user history covering 1.6 billion solving traces, the authors learn the pedagogical value of each puzzle and how to automatically pick a set that supports chess learners, all through offline reinforcement learning.