Researchers have developed an offline reinforcement learning method to automatically generate high-quality chess puzzles. They used a dataset of 1.5 billion puzzle-solving histories from online platforms like Chess.com and Lichess to learn the pedagogical value of each puzzle. The trained policy was then evaluated using offline policy evaluation, showing significant improvement for beginners with stagnant learning growth. Expert chess players also rated the puzzles generated by the model as high-quality. This approach could potentially be applied to other domains where practice items need to be carefully curated. --- Why it matters: This research is important for AI because it demonstrates how offline reinforcement learning can be used to automatically generate high-quality pedagogical materials, which could be particularly useful in educational settings or for training AI systems themselves. Source: https://arxiv.org/abs/2608.14851