A new system called StudentSim creates digital replicas of students to help AI tutors learn more effectively by simulating realistic mistakes and responses to instruction.

Researchers have introduced StudentSim, a system designed to create digital replicas of students to train AI tutors. By simulating how individual students make mistakes and respond to guidance, the system aims to overcome the high costs and time requirements of training AI tutors with real human groups. The method uses a two-stage training process to adapt models to individual learners even when limited historical data is available.

Two-Stage Training Process

StudentSim addresses data scarcity by training in two distinct phases. First, a base model—using Alibaba's Qwen3-4B-Instruct language model—is trained on pooled data from all students in a specific subject to learn common patterns and revision behaviors. In the second stage, the model is tailored to an individual student using their specific, limited records. This approach prevents the model from overfitting to the small number of examples typically available for a single learner.

Performance and Testing

The researchers tested StudentSim on 60 students across chess, math, and English as a foreign language. In these tests, StudentSim outperformed the larger GPT-5.4 model, which often failed to replicate specific student mistakes or ignored verbal hints. When used to train a chess tutor, the StudentSim-trained version achieved the highest scores for explanation quality and factual accuracy, as evaluated by professional chess players.

Future Development and Limitations

While the researchers view StudentSim as a successful proof of concept, they note that subjects like essay writing and open-ended math remain challenging due to the lack of objective scoring functions. Future work will focus on modeling how students acquire, retain, and forget knowledge over time. The project's code has been made available on GitHub.