Microsoft's StudentSim uses simulated students to improve AI tutors

Microsoft Research introduced StudentSim on October 12, a research system for evaluating and improving AI tutors. Tutors need feedback from students to get better, but studies with real learners are slow and costly. StudentSim builds a simulator for each student that models how they answer questions and how their answers change after a tutor's guidance.

Simply asking an LLM to "answer like an elementary-school student who struggles with math" can produce the right tone but far more ability than the student has, making a tutor look more effective than it is. So the team evaluates simulators on two things: how closely they match a specific student's real answers, and whether they update after reading the tutor's guidance.

Training used only publicly released, de-identified learner records. Simulators first learn shared patterns from many students, then adapt to one student's records, and they beat strong baselines in chess, second-language English writing and math. In a chess-tutor reinforcement-learning experiment, the tutor trained with StudentSim as the student reached 90.5% accuracy in blind expert ratings, compared with 75.7% for a tutor without reinforcement learning and 71.6% for one trained with GPT-5.4 playing the student.

The researchers stress that StudentSim is not a product, is not used in classrooms and is not meant to replace teachers. Real learning outcomes and safety still have to be measured with real students; the simulators are a tool for the earlier stages of development. The paper and the training and evaluation code are public.