Farewell to One-off Q&A: HKU's Open Source AI Tutor DeepTutor v1.5 Enables Teaching, Testing, and Loop Operation

According to Dynamic AI Beating monitoring, large-scale models usually can only engage in "question-answering" during tutoring sessions, unable to follow students' line of thinking or address mistakes for continuous guidance. The Hong Kong University Data Smart Lab (HKUDS) has open-sourced the intelligent agent tutor DeepTutor v1.5, incorporating Chat, Quiz, Research, Visualize, Solve, and Mastery Path into a single intelligent agent to run a closed-loop. This means that when students switch tasks, the backend engine does not need to interrupt or reset, and learning context and multi-resolution memory will automatically synchronize and flow.

To achieve truly personalized teaching, DeepTutor has introduced a tracking forest mechanism that solidifies interactive trajectories into a multi-resolution map, extracting dynamic evolving virtual student profiles. This allows every teaching conclusion provided by AI to trace back across layers to specific textual evidence or mistake records.

Evaluations have shown that DeepTutor can on average improve personalized teaching metrics by 10.8% and enhance the general reasoning ability of mainstream large models by 29.4%. However, the disruptive cost of proactive reminder teaching over extended periods, as well as the actual adaptation effect of real users to the forgetting curve, still require long-term behavioral studies for validation.