Microsoft Releases StudentSim, a Simulator for Validating Student-Specific AI Tutors

Microsoft Research has announced StudentSim, a simulator that mimics real students' response patterns and error types. Proposed to overcome the limitations of existing LLM-based role-playing, StudentSim was found to predict student behavior more accurately than large models such as GPT-5.4 in the domains of chess, L2 English writing, and mathematics.

StudentSim is based on Qwen3-4B-Instruct and undergoes two stages of LoRA fine-tuning. In the first stage, it learns from data from multiple students to acquire common patterns by domain, and in the second stage, it further trains on a specific student's data to generate a personalized adapter. The evaluation metrics used were the degree of match with the student's actual answer (F) and the degree of correction to the right answer after tutoring (R).

Notably, in the tutor reinforcement learning experiment, a common simulator trained on data from 100 students was used as the reward model instead of a student-specific simulator. This design helps prevent the tutor from overfitting to the habits of specific students. However, in the mathematics domain, StudentSim's F score (0.6384) was lower than a simple baseline that always chooses the correct answer (0.706), highlighting a limitation in confidently asserting modeling performance superiority over simple guessing.