Verified, then logged.
Sun 22 Nov 2026 · 00:24 KST / 15:24 UTC한국어 | ENSearch /

Breaking01 / AI

Microsoft publishes student simulator for training AI tutors

A wooden chessboard with a game in progress and a laptop with a blurred screen face an empty chair on a desk in a vacant classroom in late-afternoon light. AI-generated image

Microsoft said a chess tutor trained with StudentSim, a research simulator released on 13 Oct 2026, reached 90.5% accuracy, a company-published figure. The system mimics how individual students respond, so that AI tutors can be tested before they reach real learners.

That tutor used StudentSim as its reward source in reinforcement learning, a training method in which a model improves by pursuing a reward score. In company-published figures, a tutor trained without reinforcement learning reached 75.7%, and one that used GPT-5.4 as its reward source reached 71.6%. Microsoft described the result as a proof of concept.

Microsoft Research said it used the simulator in an experiment that refined the chess tutor with reward scores. When the simulator revised a move after reading the tutor's guidance, the chess engine Stockfish measured how much the move improved. Chess experts then rated the finished tutor's answers without knowing which experimental condition produced them.

According to the Microsoft Research blog, StudentSim first learned common errors from the records of many students. It then built a dedicated model of one student from that student's public, de-identified records. The evaluation covers chess, second-language English writing and mathematics.

The StudentSimEval benchmark scores whether a simulator matches a student's records, called behavioural fidelity, and whether it revises an answer after guidance, called guidance responsiveness. StudentSim's company-published chess scores are 0.5150 for behavioural fidelity and 0.9067 for guidance responsiveness. The company-published scores for GPT-5.4 are 0.2316 and 0.7186, and those for Maia2, a chess-specific model, are 0.4535 and 0.2721.

Microsoft defines StudentSim as a research system, not a product, a classroom deployment or a replacement for teachers. Its blog describes the simulators' role as "an early-stage research signal". The paper and the training and evaluation code are public.

Microsoft has not disclosed whether tutors trained on the simulator improve real students' learning outcomes.