Original title: StudentSim: Training LLM-based Student Simulators
Ke Yang, Chenglong Wang, Michel Galley, Chandan Singh +3 · September 2026
Teaching computer programs to act as good personal tutors is tough. Testing teaching strategies on real students is slow and expensive. Computer models that pretend to be students could help, but older options fail. Some cannot understand hints, while others act too smart or too clumsy compared to real learners. To fix this, researchers created StudentSim. It trains on a group of students first and then tunes itself to match a single student's unique skills. This lets it copy a student's actual answers while also learning when a tutor offers help. The researchers tested it on chess, English writing, and math. StudentSim matched real student actions and responded to guidance better than GPT-5.4 in all three subjects. Finally, they used StudentSim to train an AI chess tutor. Human experts judged this tutor as more accurate, better guided, and more personalized than tutors trained with other methods.
When you're packing for a trip, you might start with a general list of essentials that works for most people. But if you're packing for a specific person, say a toddler, you need to adjust for their needs and be ready to change plans if the weather shifts. In this paper, the system starts with a general model trained on many students, then specializes it for one student and updates it as tutor feedback comes in.
AI tutors are meant to adapt to each student, but figuring out what guidance works for whom requires expensive data from real learners. StudentSim provides a way to generate that signal synthetically, making personalized education more accessible and effective.
The abstract doesn't specify limitations, but the work is presented as a proof of concept evaluated only on chess, second-language English writing, and mathematics using de-identified public datasets.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
StudentSim: Training LLM-based Student Simulators — Ke Yang, Chenglong Wang, Michel Galley et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.