Simulated students that make realistic mistakes help AI tutors learn faster
decoder ·
Researchers from Microsoft and the University of Illinois developed StudentSim, a system that creates a separate digital replica of an individual student from limited records to provide rapid feedback for AI-tutor training. Unlike approaches that either imitate student behavior without responding to guidance or follow hints without matching a particular learner, StudentSim models both realistic answer patterns and how a student revises answers after tutoring. It first learns shared subject-level patterns from pooled student data and then adapts the model to each learner, using Alibaba’s Qwen3-4B-Instruct as its base. In tests involving 60 students in chess, English as a foreign language, and math, StudentSim outperformed GPT-5.4 when the model was prompted to act as a student. A chess tutor trained with StudentSim received the best evaluations for factual accuracy, explanation quality, and adaptation to the individual learner. The researchers describe the results as a proof of concept, note that open-ended subjects are harder to evaluate reliably, and plan to model how students acquire, retain, and forget knowledge over repeated practice.
decoder