EduAgent-RL: A Human-Centered Multi-Agent Reinforcement Learning Framework for Personalized Learning, Adaptive Feedback, and Student Performance Prediction

Personalization in digital learning is often handled through separate components: a knowledge-tracing model estimates what a student knows, a recommender selects learning content, and a conversational tutor provides feedback. These components are not always coordinated as part of a single instructional decision process. Recent reviews of agentic and multi-agent AI in education show rapid growth in this area, but also point to a lack of longitudinal studies and carefully evaluated real-world deployments [3, 12]. To address this gap, we propose EduAgent-RL, a six-agent framework comprising Student Profiling, Learning Analytics, Content Recommendation, Assessment, Feedback, and Teacher/Supervisor agents. The agents are coordinated through centralized-training, decentralized-execution multi-agent reinforcement learning (MARL), supported by a knowledge-tracing state encoder, retrieval-grounded feedback, and an explicit human-oversight veto. The framework is evaluated on two public benchmarks, OULAD [14] and EdNet [2], against random, collaborative-filtering, classical machine-learning, single-agent RL, and RAG-based LLM tutor baselines. In the pilot evaluation, EduAgent-RL reached an AUROC of 0.83 and NDCG@10 of 0.71 on OULAD, compared with 0.74 AUROC / 0.58 NDCG@10 for the single-agent RL baseline and 0.79 AUROC / 0.63 NDCG@10 for the RAG-based tutor. The ablation results also showed the largest drop in performance when the knowledge-tracing encoder was removed.