StudentSim: Training LLM-based Student Simulators

Abstract AI tutors are most useful when they adaptively respond to each student’s strengths, weaknesses, and preferred kinds of guidance, but which guidance works for which student is a sparse signal, slow and costly to collect from real students. Student simulators can supply that signal as a proxy, yet existing ones cover only part of what this requires: state-tracking models fit how a student behaves but cannot digest a tutor’s explanations or corrections well, while LLMs prompted to role-play a target student follow a tutor’s guidance fluently but do not reliably reproduce the competence of the student they imitate. We present STUDENTSIM, a training framework that turns sparse per-student data into an individualized simulator for each student through a two-stage pipeline of pooled training followed by per-student specialization, so that the simulator both mirrors the student’s own responses and updates them under tutor guidance. To measure these two abilities fairly, we build STUDENTSIMEVAL, a standardized protocol spanning 60 students across chess, second-language English writing, and mathematics, drawn from public learner datasets whose de-identified student records are shared for research.

Introduction. An AI assistant that helps human students learn (e.g., a math tutor, a foreign-language coach, a chess trainer) is more useful when it is trained on real-world interactions across a diverse student population: it must see varied backgrounds, baseline skills, and learning styles to understand the typical proficiency and common error patterns of students in that subject, and it must collect feedback on each individual’s responses to different teaching materials to learn which teaching approach is most effective for a student with particular traits and learning habits. Recruiting and training a tutor against such a population, however, is prohibitively expensive and time-consuming. Therefore, the evaluation and improvement of adaptive AI tutors lag behind the rapid advancement of the underlying AI models themselves. One alternative is to use student simulators that produce proxy feedback at machine timescales (Fig. 1).

Discussion / Conclusion. STUDENTSIM is a training framework for per-student AI student simulation built around two requirements that must hold together: behavioral fidelity (F), how well a simulator matches a student’s own responses, and guidance responsiveness (R), how reliably it updates towards where the tutor guidance leads. To measure both abilities on a common footing, STUDENTSIMEVAL casts public learner corpora for research across chess, second-language English writing, and mathematics into a standardized per-student protocol on which any simulator, from our framework or a prior family, is fit and scored on the same held-out records. Under this protocol, STUDENTSIM’s two-stage pooled-then-specialized pipeline yields a reference family of 60 individualized simulators that are strong on both axes across all three domains, outperforming domain-specific state-tracking (weak on R) and prompt-only LLM role-play (weak on F).