REC-CBM: New AI model makes open-ended grading transparent and trustworthy

Open-ended grading remains a bottleneck in personalized education—manual grading is slow and costly, while black-box LLMs lack transparency. Zhao et al. (2026) propose REC-CBM (Rubric-Aware Error-Correction Concept Bottleneck Model), which builds on concept bottleneck models (CBMs) to make automated grading interpretable. Unlike standard CBMs, REC-CBM explicitly models fine-grained rubric dimensions via a rubric-aware concept encoder that learns concept-specific representations per response. It also introduces an ordinal pairwise calibration objective to preserve the ranking structure among score levels, and a latent concept error-correction module that denoises unreliable human concept annotations before final grade prediction, all while maintaining full interpretability.

Comprehensive experiments on public datasets show REC-CBM consistently outperforms state-of-the-art baselines in grading accuracy and produces more faithful concept-level reasoning. The model's design allows educators to inspect intermediate concept predictions, intervene if needed, and ultimately trust automated grading decisions. This work advances transparent AI in education, addressing the critical gap between high-accuracy black-box systems and the need for verifiable, human-understandable rationale in high-stakes assessment scenarios.