The field of semi-supervised learning is moving towards more robust and generalizable solutions, with a focus on addressing key limitations such as static view interactions, unreliable pseudo-labels, and lack of hard sample modeling. Recent developments have introduced novel frameworks that incorporate teacher-student architectures, adversarial generators, and context-aware pseudo-label selection. These advancements have consistently achieved state-of-the-art performance in low-label regimes and have been applied to various tasks, including 3D object detection and semantic occupancy prediction. Noteworthy papers include TRiCo, which formulates semi-supervised learning as a Stackelberg game, and Learning Adaptive Pseudo-Label Selection for Semi-Supervised 3D Object Detection, which proposes a learnable pseudo-labeling module. Additionally, CORE-3D and EasyOcc have made significant improvements in 3D scene understanding tasks, such as 3D semantic segmentation and object retrieval.