Microsoft Research releases StudentSim, a personalized student simulator for evaluating and improving AI tutoring systems

StudentSim's core idea is to break a student simulator into two quantifiable objectives: behavioral fidelity (whether it matches a given student's real answers on held-out questions) and steerable responsiveness (whether it updates its answers after receiving tutoring hints). Training uses a two-stage process: pretraining on multi-student records within a subject to learn shared patterns, then personalized adaptation on a single student's sparse records. To evaluate the method, the team built the StudentSimEval benchmark spanning chess, second-language English writing, and math. On chess, StudentSim reached 0.5150 behavioral fidelity and 0.9067 steerable responsiveness, versus 0.2316 and 0.7186 for GPT-5.4 and 0.4535 and 0.2721 for Maia2, which is specialized for predicting human moves. Researchers also used the simulator as a reinforcement learning feedback source to train a chess tutoring system, which scored 90.5% accuracy in blind evaluation, above the 75.7% no-RL baseline and 71.6% with GPT-5.4 as the reward source. The system targets researchers developing and evaluating AI tutoring tools. It does not replace teachers or classrooms, nor real-learner studies; instead it places simulated feedback in early development to compare tutoring strategies and screen training signals before costly human experiments. The paper and training/evaluation code are public.