Microsoft Research presents StudentSim and StudentSimEval, a personalized learner-simulation system and evaluation protocol. It aims to reproduce both a student’s answer and how it changes after guidance rather than simply asking an overly capable model to role-play a student.
Training pools records within a domain before individual adaptation from public de-identified learner data. Studied domains are chess, second-language English writing and mathematics. The paper and training and evaluation code are public.
This is research, not a classroom product or teacher replacement. Positive simulation fidelity and chess-tutor results do not establish real educational benefits in a school. Simulations cannot replace evaluation with actual learners.
Improving tutors requires measuring learning, not merely how convincing their explanations sound.