Intelligent Evaluation of STEM Project-Based Learning Reports Based on a HITL Multi-Agent System

STEM project-based learning (PBL) reports have long struggled with nonuniformity and subjectivity in assessing higher-order thinking skills. This study proposes a human-in-the-loop (HITL) enhanced multi-agent system to address these challenges. Drawing upon established assessment frameworks, the system operationalizes the evaluation of critical and creative thinking into a multi-stage automated pipeline comprising scoring, review, and aggregation agents. Using Deep Research-generated reports as scoring references, a total of 101 STEM PBL reports were evaluated to validate the system’s performance. Results demonstrate that the multi-agent system achieves high scoring stability and consistency. Furthermore, the HITL mechanism significantly improves the alignment between automated scores and expert teacher judgments. In addition, hierarchical clustering based on the system’s outputs successfully identified three distinct performance groups with different levels of creative thinking and critical thinking. These findings indicate that the HITL-enhanced multi-agent system can support a more objective and transparent assessment of STEM PBL reports, providing actionable insights for learning analysis and educational evaluation.