Microsoft and the University of Illinois developed a new system that builds realistic replicas of individual students from limited data. These replicas provide rapid feedback when gathering it from real learners proves too expensive and slow. AI tutors perform best when adapting to each student's strengths and weaknesses. Finding out which guidance works for each student requires time and money because it relies on real people.
Training an AI tutor with a large and diverse group of students is prohibitively expensive and time-consuming. Improvements to these tutors have lagged behind broader advances in AI models. The researchers propose using digital replicas of students to supply quick feedback instead. Their system, called StudentSim, builds a separate replica for each student even when very few records of that person's work exist.
Student replicas need to make mistakes and learn from guidance. Existing approaches handle only one of these two necessary skills. Some models learn from real student data and reliably reproduce behavior, but they cannot use a tutor's explanations. Other models are language models prompted to act as students. They follow hints easily but fail to match the actual abilities of the students they mimic.
StudentSim turns both skills into measurable goals. It measures how closely a replica matches a student's answers, including typical mistakes. It also measures how readily the replica revises an answer after receiving help from a tutor. Tutor training requires both a realistic starting point and a simulated student that responds to instruction.
The biggest obstacle for the researchers is a lack of data. In an English writing dataset, the median student has written only three essays. More than two-thirds of the students have written five or fewer. Training a replica directly on so few examples fails because the model overfits to those examples.
StudentSim uses a two-stage training process to make limited data usable. A base model first learns from the pooled data of all students in a subject. It learns common mistakes and how students revise answers after a tutor's hint. The researchers then tailor that model to an individual student using the few available records for that person. The system uses Alibaba's Qwen3-4B-Instruct language model as its base across all subjects.
The researchers tested the method on 60 students across chess, English as a foreign language, and math. They used public datasets containing records from real learners. StudentSim outperforms the larger GPT-5.4 language model in all three subjects when GPT-5.4 is prompted to act as a student. In chess, StudentSim correctly predicts a player's next move about twice as often and almost always follows corrective guidance. GPT-5.4 and specialized chess models fall behind.
The researchers also used a student replica to improve a chess tutor as a proof of concept. Professional chess players evaluated three versions of the tutor. One had no extra training, one was trained with GPT-5.4 as the student, and one was trained with StudentSim. The StudentSim-trained tutor scored highest on all three measures. It made the fewest serious factual errors and received the highest scores for explanation quality and adaptation to the individual student.
The team next wants to model how students acquire, retain, and forget knowledge over many practice sessions. The code is available on GitHub.
