Farewell to Single-Turn Q&A: HKU Open-Source AI Tutor DeepTutor v1.5 Implements a Concurrent Loop of Teaching and Testing

According to Beating monitoring, when large models serve as tutors, they usually can only carry out one-question-one-answer interactions, unable to continuously coach based on the student’s thought process and incorrect answers. The Hong Kong University Data Intelligence Laboratory (HKUDS) has open-sourced an intelligent agent tutor, DeepTutor v1.5, integrating six modules—Chat, Quiz, Research, Visualize, Solve, and Mastery Path—into a single agent that runs an end-to-end operational loop. This means that when students switch tasks, the backend engine does not need to be interrupted or reset, and the learning context along with multi-resolution memory automatically syncs and flows through.

To achieve truly tailored instruction, DeepTutor introduces a tracking forest mechanism, turning interaction traces into multi-resolution graphs, and distilling a dynamically evolving virtual student profile. This allows every step of the teaching conclusion produced by the AI to be traced across layers back to specific textual evidence or records of incorrect answers.

Evaluation results show that DeepTutor can improve personalized teaching metrics by an average of 10.8%, and increase the general reasoning ability of mainstream large models by 29.4%. However, the distraction cost of this proactive reminder-based teaching over long time horizons, as well as the real-world effectiveness of how human users adapt to the forgetting curve, still needs to be validated through long-term behavioral studies.