Microsoft reports improved chess tutor accuracy using simulated student feedback

Microsoft Research reports that a chess tutor trained with feedback from StudentSim, its research system for simulating individual learners, reached 90.5% accuracy in a study. That compares with 75.7% for a tutor without reinforcement learning and 71.6% for one trained using GPT-5.4 as the simulated student. StudentSim learns shared patterns from pooled, de-identified learner records, then adapts to each student's records. During tutor training, it revises a chess move after receiving guidance; improvements in move quality supply the reward. Human chess experts evaluated tutor responses blind to the training condition. The paper and code are publicly available. StudentSim is not deployed in classrooms, and human studies remain essential to establish real learning outcomes.