Farewell to single question and answer: HKU open source AI mentor DeepTutor v1.5 realizes teaching and testing to run in the same loop According to Beating monitoring, large models usually only have "one question, one answer" when teaching, and cannot provide continuous guidance according to students' thinking and wrong questions. The Data Intelligence Laboratory at the University of Hong Kong (HKArica) has open-source DeepTutor v1.5, which integrates six stages of intelligent agent operation: Chat, Quiz, Research, Visualize, Solve, and Mastery Path into a single closed-loop system. This means that students do not need to interrupt or reset the backend engine when switching tasks, and the learning context and multi-resolution memory will automatically synchronize and flow. To achieve true personalized teaching, DeepTutor has introduced a tracking forest mechanism, which condenses interactive trajectories into multi-resolution graphs and extracts dynamically evolving virtual student portraits. This allows every teaching conclusion provided by AI to be traced back to specific textual evidence or error records across layers. According to the evaluation, DeepTutor can improve personalized teaching metrics by an average of 10.8% and enhance the general reasoning ability of mainstream models by 29.4%. However, the disturbance cost of this proactive reminder teaching over a long period of time, as well as the actual adaptation effect of real users to the forgetting curve, still require long-term behavioral research to verify. [Original link]