Microsoft trains SI to imitate individual students, and a tutor trained on such simulators becomes more accurate
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
An SI tutor can train on "virtual students" without troubling real ones, and teach noticeably better.
The StudentSim system creates personalized student simulators: it models how a person answers questions and how the answers change after a tutor's hint. According to Microsoft Research, on a chess task answer-reproduction fidelity was 0.5150 versus 0.2316 for GPT-5.4, and responsiveness to guidance was 0.9067 versus 0.7186; a tutor trained on such a simulator reached 90.5% accuracy versus 75.7% for the baseline. The method was also tested on second-language writing and math; the paper and code are published.
Sources 1