A Co-Tutor for AI-Assisted Personalized Education

Context: This extended abstract summarizes an AI Co-Tutor framework that applies advanced knowledge tracing to personalize programming education at Nanyang Technological University (NTU), Singapore, with undergraduate students in the College of Computing and Data Science. The pedagogical approach combined blended learning, AI-driven copilot tools, and structured classroom activities to teach coding fundamentals and algorithms. Problem, Challenge, or Opportunity: Advances in generative AI open new possibilities for individualized programming instruction, yet expose a key limitation: existing educational models struggle to track evolving student knowledge states in real time. Conventional Knowledge Tracing approaches do not readily interface with modern AI systems, resulting in inefficient resource use, one-size-fits-all learning paths, and insufficient personalization — a significant shortcoming in programming education where targeted feedback is essential for skill development. Bridging this gap requires AI-enhanced knowledge tracing capable of deploying generative models as scalable, adaptive tutors. Methods Used to Develop and Evaluate the Innovation: The approach centres on combining optimized knowledge tracing algorithms with large language models. Custom AI copilot platforms — including MyCodeWeapon and Ask Codey (which uses Socratic dialogue to support self-directed learning) — were developed and deployed. Assessment drew on benchmarking against the EdNet dataset (131 million interaction records) alongside controlled classroom studies with more than 200 students. Evaluation criteria encompassed prediction accuracy, assignment performance, engagement metrics, and post-course feedback including AI ethics perspectives. Main Findings and Outcomes of the Inquiry: On the EdNet benchmark, COTUTOR attained 84.3% prediction accuracy, surpassing baseline systems by margins of 5.4 to 11.8 percentage points. In the NTU classroom trial, students in the COTUTOR condition averaged 73.2% on assignments compared to 68.4% for the control group, while reporting higher satisfaction (72.8% vs. 67.9%) and stronger participation rates (73.9%). These findings have been disseminated through peer-reviewed publications and international conferences. Implications for Teacher Practice: By handling repetitive tasks such as grading and content delivery, COTUTOR allows educators to redirect attention toward nurturing creativity and deeper conceptual understanding. The system provides interpretable AI-driven insights to support scalable personalization, while preserving the role of human judgment in pedagogical decisions.

Keywords (free text)

generative AI, AI-assisted coding, agentic AI co-tutor, knowledge tracing, learning to learn

Permanent link to cite this item

https://hdl.handle.net/20.500.14069/1328
https://doi.org/10.71779/787

Collections

DI-IDEA AI for Education (AI4E) Competition 2025