Tutor-Agents: A Collaborative LLM-Based Agent Framework for Explainable Dropout Prediction in Educational Unmanned Systems

The high dropout rate in Massive Open Online Courses (MOOCs) has severely hindered the promotion of intelligent unmanned systems in the field of education. Existing methods have significant limitations single-agent models struggle to integrate multi-dimensional dynamic learning behaviors traditional temporal modeling methods such as standard LSTM fail to effectively capture the laws of behavioral evolution and prediction results lack interpretability. To address this this paper proposes Tutor-Agents—a multi-agent collaborative framework based on Large Language Models (LLMs) to achieve accurate prediction of dropout risks and interpretable decision-making. The framework clarifies the expert roles and collaborative logic of each agent by encoding Standard Operating Procedures (SOPs) and it defines agent roles through SOPs to form a collaborative closed-loop centered on the Manager. It first invokes the User Analyst to extract temporal vectors of user behaviors and the Item Searcher to obtain historical course records then the Similar User Searcher generates behavioral embedding vectors using BiLSTM with temporal weights and matches peer learning trajectories to construct a dynamic risk assessment model. The Judge integrates data to determine dropout risks. Finally the Reflector verifies the rationality of the process. This mechanism breaks through the limitations of single-agent models enhancing the ability to capture temporal behaviors and the interpretability of predictions. Experimental validation based on the XuetangX dataset shows that the framework achieves a dropout prediction accuracy of 83.4% and an AUC value of 83.1%.