KT4EQG: Personalized Exercise Question Generation via Knowledge Tracing

Synthesis: KT4EQG: Personalized Exercise Question Generation via Knowledge Tracing bridges two key AI-in-education paradigms: Personalized Learning through question generation and Learning Analytics through knowledge tracing. Rather than generating generic practice questions, KT4EQG uses a Knowledge Tracing model to first identify the knowledge concept that would maximize a student's potential improvement in overall mastery, then trains an Large Language Models (LLMs)-based generator to produce a question faithfully grounded in that concept. This two-stage architecture — KT for concept selection, LLM for faithful question generation — outperforms less personalized baselines on XES3G5M and MOOCRadar datasets. The approach represents a significant advance in Adaptive Learning system design, connecting to Interpretable Knowledge Tracing research on modeling student knowledge states and Automated Question Generation work on producing high-quality educational content. Unlike earlier systems such as Slide Deck Q&A Quality Assurance App: A Multi-Stage Pipeline for Pedagogical Question Generation that generate questions from static content, KT4EQG personalizes based on dynamic student models, aligning with Intelligent Tutoring goals of providing the right question at the right time for each learner.

What this means for practice

• Instructors. Choose the next practice concept from a knowledge-tracing model's mastery estimates rather than from a fixed topic sequence, so each student works on the concept with the largest expected gain.
• Designers. Split concept selection from question writing and train the generator with an explicit alignment reward, because without it generated questions collapse onto similar content — the strongest baseline put 35.3% of verified questions on one concept against 5.2% for KT4EQG.
• Designers. Screen every generated item for answerability before it reaches a student; this study used a separate language model to judge whether a question admits a well-defined solution.
• Researchers. Evaluate generated practice questions by downstream exam-score gain and verified concept alignment, not by surface text-similarity metrics alone.

Limitations

• Effectiveness was measured entirely in simulation: after each generated question the knowledge-tracing model updated the student's state assuming successful practice, so the reported exam-score gains are model-generated rather than observed learning by real students.
• The evaluation covers three of the largest knowledge modules in XES3G5M with 300 students per module, and MOOCRadar supplies only concepts, so its hierarchical tree was LLM-constructed rather than expert-annotated.
• Question answerability was judged by Qwen-3-4B, an automated judge, with no human rating of item quality in the main results.
• The generated items are elementary mathematics written for a third-grade audience, so the demonstrated gains may not transfer to the higher-education settings an adaptive platform would use them in.

Citation

Xinyi Gao, Qiucheng Wu, Lu Ding, Q. Vera Liao, Kaizhi Qian, Ying Xu, Shiyu Chang, Yang Zhang (2026). KT4EQG: Personalized Exercise Question Generation via Knowledge Tracing. arXiv preprint.