Agentic RAG Technology for Intelligent Educational Platforms
Building a next-generation AI tutoring platform fusing Agentic RAG with Socratic dialogue — RISE Industry-Academia Joint Research with Saltlux, Yonsei University, and KAIST.
Overview
This RISE program joint research project builds a next-generation intelligent educational platform that fuses Agentic RAG technology with Socratic dialogue methods. The goal: improve hallucination issues, enhance learners’ critical thinking, and provide 24-hour AI tutoring — innovating Jeju’s industry-academia cooperative education ecosystem.
The central argument is that competitive advantage in educational AI lies not in model size but in the depth of pedagogical philosophy embedded in the system’s design.
Core Technologies
- Multi-Agent-based Agentic RAG architecture design and development
- Socratic dialogue management system — guiding student thinking rather than short-circuiting it
- Knowledge Tracing-based personalized learning support with dual-pathway student modeling
- Constitutional AI-based educational ethics safeguards
- Simulation learning environments linked to Jeju’s strategic industries
- Agent reasoning evaluation framework and benchmarks
Technical Architecture
The system rests on three integrated components:
PDA Reward Model — A reward framework replacing simple accuracy-based evaluation:
- Process Fidelity (P): Rewards Socratic dialogue and scaffolding; penalizes direct answer provision
- Diagnostic Accuracy (D): Rewards root-cause identification of student misconceptions through causal reasoning
- Adaptive Precision (A): Rewards instructional moves within the student’s Zone of Proximal Development
Hybrid Two-Speed Student Modeling — A fast path maintains real-time KC mastery estimates; a slow path runs asynchronous LLM analysis of conversation logs for deeper causal understanding of misconceptions.
Reinforcement Learning Training — Supervised fine-tuning on expert interactions creates a baseline, then RL optimizes against the PDA reward model to discover pedagogical strategies beyond simple imitation.
Participating Institutions
| Institution | Role |
|---|---|
| Cheju Halla University (Lead) | Research management, platform architecture, industry-academia program |
| Saltlux | Agentic RAG development, FCG system, Knowledge Tracing |
| Yonsei University | Socratic dialogue LLM, reasoning pattern analysis |
| KAIST | Agent evaluation metrics, benchmark dataset construction |
Student Researchers
Six student researchers from the Department of AI participate — reviewing NLP/LLM papers, developing Socratic dialogue prototypes, building simulation benchmarks, and conducting experiments on AI tutoring systems.
Project Information
- Research Period: August 2025 – December 2027 (29 months)
- Total R&D Budget: Approximately 1.7 billion KRW
- Principal Investigator: Professor Young Joon Lee, Cheju Halla University