HKU’s DeepTutor v1.5 merges six tutoring modes into one seamless loop

Large language models acting as tutors are usually limited to one-off Q&A — they can't follow a student's train of thought or learn from their mistakes to provide continuous guidance. The Data Intelligence Lab at the University of Hong Kong (HKUDS) has open-sourced an AI tutor agent called DeepTutor v1.5 that integrates six modes — Chat, Quiz, Research, Visualize, Solve, and Mastery Path — into a single agentic closed loop. This means that when students switch tasks, the backend engine doesn't need to reset; learning context and multi-resolution memory sync automatically.

To achieve true adaptive teaching, DeepTutor introduces a Tracking Forest mechanism that distills interaction trajectories into multi-resolution graphs, creating a dynamically evolving virtual student profile. This allows every teaching decision made by the AI to trace back across layers to specific textual evidence or error records.

Benchmarks show that DeepTutor improves personalized teaching metrics by an average of 10.8% and boosts the general reasoning ability of mainstream large models by 29.4%. However, the interruption cost of this proactive reminder teaching over long periods, and how real users actually adapt to the forgetting curve, still require long-term behavioral studies to validate.