FURI | Fall 2025

EduCBM: Concept Bottleneck Models for Interpretable Education

Language models for education have demonstrated strong performance in tasks such as automatic essay grading, question answering, and providing tailored responses. Nonetheless, their “black box” nature creates major challenges for responsible implementation in educational environments, where transparency and interpretability are essential for building educator trust and enhancing student learning. We present EduCBM, a framework that transforms opaque educational AI into transparent systems that clearly explain their decision-making processes using recognizable teaching concepts, enabling educators to trust and verify automated grading and tutoring recommendations. Through comprehensive experiments on standardized test datasets, essay scoring corpora, and student response collections, we demonstrate that EduCBM maintains competitive predictive performance while providing valuable insights into model decision-making processes.