Microsoft's StudentSim mimics students to help build better AI tutors

Building AI tutors has long faced a challenge: figuring out if an explanation actually helps students usually needs costly and slow tests with real people.

Microsoft Research launched a system called StudentSim. Based on public, de-identified learner records, it trains a personal "AI student simulator" for each learner, mimicking both how they answer on their own and how they react after coaching. It uses two-stage training across chess, English writing, and math tasks, with a benchmark called StudentSimEval.

StudentSim beat major baselines in all three areas. In chess, it scored 0.5150 in behavioral fidelity and 0.9067 in responsiveness to guidance, higher than GPT-5.4 and chess model Maia2. An AI chess coach trained with its feedback hit 90.5% accuracy, above the 75.7% baseline.

Microsoft stressed the system isn't a classroom product and won't replace teachers. It gives researchers low-cost, scalable early feedback before real-student testing.