Researchers have developed a new framework called C-Score to assess the robustness of semi-supervised learning (SSL) models in real-world environments. SSL models are often trained on labeled data and unlabeled data from the same distribution, but in practice, unlabeled data can come from different sources and contain out-of-distribution samples. The authors argue that relying solely on clean accuracy is insufficient for evaluating SSL robustness, as it may not detect hidden degradation. C-Score evaluates training behavior in three spaces: prediction, feature representation, and optimization, using metrics such as PLE, CCI, Sem-Drift, and Grad-Align. Experiments show that C-Score reveals hidden degradation that clean accuracy alone fails to detect. --- Why it matters: This matters to researchers because it highlights the limitations of relying solely on clean accuracy for evaluating SSL robustness in open-world environments. The findings suggest that internal diagnostic signals are necessary for more reliable robustness assessment, which can inform the development of more robust SSL models. Source: https://arxiv.org/abs/2608.20667