🧠 AI Ed Wiki

📄 Research Article

DeepTutor: Towards Agentic Personalized Tutoring

● high · 2026-08-04llm, agentic-ai, personalized-learning, intelligent-tutoring, rag, question-generation, adaptive-learning, tutoring-systems, generative-ai, automated-question-generation

A fully open-source agentic tutoring framework that closes the loop between citation-grounded problem tutoring and difficulty-calibrated question generation, powered by a hybrid personalization engine combining static knowledge grounding with dynamic learner memory. Evaluated via TutorBench across 5 university disciplines, improving personalized metrics by 10.8% and general agentic reasoning by 29.4% across 5 backbone models.

ArXiv: 2604.26962 (v3, July 2026)

Code: github.com/HKUDS/DeepTutor

Overview

DeepTutor is a fully open-source Agentic AI framework that unifies two tutoring tasks — citation-grounded problem tutoring and difficulty-calibrated question generation — through a shared personalization engine. Unlike prior systems that treat tutoring and question generation as isolated workflows, DeepTutor creates a closed loop: weaknesses exposed during tutoring directly shape which questions are generated next, and learner performance on those questions refines future explanations.

Hybrid Personalization Engine

1. Static Knowledge Grounding (SKG): Course-sourced knowledge indexed via RAG from textbooks and lecture materials, ensuring all tutoring responses are citation-grounded and factually anchored.

2. Dynamic Personal Memory (DPM): A trace forest — a hierarchical memory structure where specialized agents continuously distill multi-turn interaction traces into an evolving learner profile. Unlike coarse skill inventories, this captures fine-grained reasoning traces showing how a student errs, not just what they got wrong.

Architecture

  • Problem Tutoring Pipeline: Investigation -> Guided Solving -> Iterative Writing
  • Question Generation Pipeline: Idea Selection -> Verified QA Construction
  • Broader Extensions: Adaptive learning workflows, interactive books, proactive multi-channel tutoring agents

    • TutorBench: Student-Centric Evaluation

  • Customized learner profiles grounded in university-level curricula across 5 disciplines
  • LLM-based first-person interactive evaluation via a profile-driven student simulator
  • Multi-turn dialogue testing adaptive behavior end-to-end

    • Key Results

    MetricImprovement
    Personalized tutoring quality+10.8% avg
    Agentic reasoning (5 backbone models)+29.4%
    Cross-domain generalizationconfirmed
    Human-alignmentvalidated via ablation

    Significance

    DeepTutor addresses a critical gap in Intelligent Tutoring: the disconnect between tutoring traces and subsequent practice. By coupling both through a shared learner memory, it demonstrates that closed-loop personalization yields measurable gains in both tutoring quality and reasoning capability. The fully open-source release and TutorBench benchmark provide infrastructure for the broader Adaptive Learning research community.

    Connected Concepts

  • Agentic AI
  • RAG
  • Intelligent Tutoring
  • Adaptive Learning

    • Connected Articles

  • MOOC To Maic — From MOOC to MAIC: Reshaping Online Teaching and Learning through LLM-driven Agents
  • Kt4eqg Personalized Question Generation — KT4EQG: Personalized Exercise Question Generation via Knowledge Tracing
  • Learnmate2 LLM Adaptive Learning — LearnMate^2: Design and Evaluation of an LLM-powered Personalized and Adaptive Support System for Online Learning
  • Xie Hillm Cd 2026 — HiLLM-CD: LLM-Enhanced Hierarchical Cognitive Diagnosis
  • AI Enabled Serious Games — AI-Enabled Serious Games: Integrating Intelligence and Adaptivity in Training Systems
  • Pattern Kc Programming Recommendation — Automated Recommendation of Programming Learning Content Using Pattern-based Knowledge Components

    • Citation

    Report, T., Personalized, D.T.A., & Tutoring, D.T.A.P. (2026). DeepTutor: Towards Agentic Personalized Tutoring