Simulated students that make realistic mistakes help AI tutors learn faster
Microsoft's StudentSim builds individual student replicas on Qwen3-4B-Instruct to train AI tutors, outperforming GPT-5.4 in chess, English, and math.
Microsoft researchers propose StudentSim, a system that creates digital replicas of individual students to provide cheap feedback for training AI tutors. It uses Alibaba's Qwen3-4B-Instruct as its base model and trains in two stages: first learning shared student patterns from pooled data, then adapting to individuals with few records. Tested on 60 students across chess, English, and math, it outperforms GPT-5.4 prompted as a student, predicting chess players' moves about twice as often and almost always following corrective guidance. A proof-of-concept chess tutor trained with StudentSim scored highest with professional players, and the code is available on GitHub.
- StudentSim uses two-stage training: pooled base model, then per-student adaptation
- Base model is Alibaba's Qwen3-4B-Instruct across chess, English, and math
- Outperforms GPT-5.4 as a simulated student in all three subjects
- Chess tutor trained with StudentSim scored highest with professional evaluators
- Code publicly available on GitHub; work is a proof of concept
- Organizations
- Microsoft
- AI models
- Qwen3-4B-InstructGPT-5.4Maia2
From the original article
4 min read at the-decoder.com
Microsoft and the University of Illinois built StudentSim to replicate individual students from limited data and give AI tutors fast, low-cost feedback. In tests covering 60 students across chess, English, and math, it outperformed GPT-5. 4.
By Jonathan Kemper · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.