KT4EQG: AI generates personalized practice questions via knowledge tracing

A new AI framework called KT4EQG tackles the challenge of generating truly personalized practice questions for students. The system first employs knowledge tracing (KT), which models a student's current understanding based on their historical performance on exercises. By predicting future performance, the KT model identifies the specific knowledge concept that—if practiced next—would maximize the student's overall knowledge mastery gain. Once the optimal concept is selected, an LLM-based question generator creates a question grounded in that concept, ensuring both relevance and difficulty. The team tested KT4EQG on two large-scale educational datasets: XES3G5M and MOOCRadar. Results show that questions generated by KT4EQG led to significantly greater learning improvements compared to methods with limited or no personalization (e.g., random concept selection or static question banks).

This work bridges the gap between knowledge tracing (traditionally used for assessment) and question generation (traditionally used for content creation). By integrating both, KT4EQG offers a closed-loop learning system: assess the student, determine the most valuable next topic, generate a custom exercise, and repeat. The paper appears in arXiv under Computers & Society and AI categories, authored by Xinyi Gao, Qiucheng Wu, Lu Ding, Q. Vera Liao, Kaizhi Qian, Ying Xu, Shiyu Chang, and Yang Zhang. While the current implementation relies on a trained LLM for question generation, future work could explore more efficient or real-time generation. For educators and edtech platforms, this represents a significant step toward truly adaptive tutoring systems that treat each student's knowledge state as unique.