Tag
The paper introduces C-Score, a diagnostic framework for evaluating pseudo-label-based semi-supervised learning under open-world unlabeled contamination, demonstrating that clean accuracy is insufficient to detect hidden performance degradation.