"StudentSim: Training LLM-based Student Simulators" by Ke Yang , Chenglong Wang , Michel Galley , Chandan Singh , Jeevana Priya Inala , ChengXiang Zhai , Jianfeng Gao
TLDR:
StudentSim is a framework that trains personalized student simulators using sparse data in order to mimic learner responses and adapt to tutor guidance more effectively than existing models. The simulator is capable of reflecting a student's strengths, weaknesses, and response patterns, and can update based on tutoring input. A standardized evaluation protocol called StudentSimEval was introduced to measure how well the simulator matches a student's responses and adjusts to guidance. Results show that StudentSim outperforms GPT-5.4 across chess, writing, and math domains, with high accuracy and responsiveness. The framework was proven effective in producing a chess tutor that expert humans rated as more accurate and personalized compared to other methods. The code for StudentSim is publicly available on GitHub for further research.
Read Paper / Blog
huggingface.co
Paper page - StudentSim: Training LLM-based Student Simulators
Join the discussion on this paper page
1September 3, 2026 56 1
