StudentSim: Microsoft Trained an AI Tutor on Just Three Essays

Microsoft and the University of Illinois introduced StudentSim, a system that simulates a specific student from a small set of data. In a test group of 60 students, the median student had written 3 essays, and more than two thirds had written no more than 5 papers. On the task of predicting the next move in chess, StudentSim outperformed GPT-5.4 by roughly a factor of two. A tutor trained on simulated students received the highest expert ratings for explanation accuracy and adaptation to the individual student.

For me, the main point is this: personalization has always been limited by the lack of data on individual people. If three papers are enough to build a working student profile, then developing educational agents becomes realistic for small teams that cannot afford large-scale data collection.