StudentSim: Training LLM-based Student Simulators

StudentSim is a training framework that turns sparse per-student data into individualized student simulators. It addresses limitations of existing approaches: state-tracking models fit student behavior but struggle with explanations or corrections, while LLM role-play follows guidance fluently but does not reliably match student competence. StudentSim uses pooled training followed by per-student specialization to create simulators that both mirror a student's responses and update under tutor guidance. The paper also introduces StudentSimEval, a standardized evaluation protocol covering 60 students across chess, second-language English writing, and mathematics, using public learner datasets with de-identified records. StudentSimEval measures behavioral fidelity (F) and guidance responsiveness (R).

The key technical contribution is the two-stage training approach: pooled training across students followed by per-student specialization. This allows the model to leverage common patterns while adapting to individual differences. The evaluation protocol StudentSimEval provides standardized metrics (behavioral fidelity and guidance responsiveness) that could become a benchmark for student simulators. The use of public learner datasets with de-identified records enables reproducibility. Future work may explore scaling to more students and domains, and integrating simulators into tutor training pipelines.