SPP-L²: A PPO-Enhanced Large Language Model Framework for Student Performance Prediction on Learner-Sourced Questions

In response to the growing complexity and dynamism of learner-sourced education platforms, this paper presents SPP-L2, a unified student performance prediction framework that integrates semantic understanding with reinforcement learning. The framework leverages a signed bipartite graph to capture structured student–question interactions and employs a large language model to encode the nuanced semantics of natural language questions. These representations are fused and used as input to a Proximal Policy Optimization (PPO) agent, which dynamically learns to predict student responses while receiving feedback signals for policy refinement. A value-based feedback mechanism further enhances the system’s ability to adaptively recommend questions and personalize interventions. Extensive experiments conducted on five real-world PeerWise course datasets demonstrate that SPP-L2 outperforms existing methods in terms of prediction accuracy, robustness, and adaptability. The proposed framework provides a principled and scalable solution for intelligent learning platforms by bridging representation learning, policy optimization, and feedback-driven adaptation.