Simulating student learning for better AI tutoring
Microsoft says StudentSim can test AI tutors against simulated individual learners before human studies.
The research system trains one simulator per student from public, de-identified learner records. It is evaluated on whether it matches a student’s held-out behavior and whether it changes after tutor guidance. Across chess, English writing, and math, Microsoft says StudentSim beat leading baselines on those measures. In a chess tutor test, feedback from StudentSim produced higher expert-rated tutoring than no reinforcement learning or GPT-5.4 as the simulated student. Microsoft Research AI's note
The research system trains one simulator per student from public, de-identified learner records. It is evaluated on whether it matches a student’s held-out behavior and whether it changes after tutor guidance. Across chess, English writing, and math, Microsoft says StudentSim beat leading baselines on those measures. In a chess tutor test, feedback from StudentSim produced higher expert-rated tutoring than no reinforcement learning or GPT-5.4 as the simulated student. Microsoft Research AI's note
score 4