The field of semi-supervised learning is moving towards more effective utilization of unlabeled data, with a focus on addressing challenges such as class imbalance biases and prediction uncertainty. Recent developments have introduced innovative approaches to transform uncertainty into a learning asset, including fuzzy adaptive rebalancing and contrastive uncertainty learning. These methods have shown significant improvements in segmentation accuracy, particularly for under-represented classes and ambiguous regions. Additionally, quality-aware semi-supervised learning frameworks have been proposed to improve table extraction from business documents, achieving state-of-the-art results on benchmark datasets. The use of synthetic data augmentation has also been explored to enhance table detection performance. Noteworthy papers include FARCLUSS, which introduces a holistic framework for semi-supervised semantic segmentation, and QUEST, which proposes a quality-aware semi-supervised table extraction framework. Synthetic Data Augmentation for Table Detection is also notable for its automated LaTeX-based pipeline to synthesize realistic document images for table detection training.