Active Generalized Category Discovery (AGCD) addresses the challenges of Generalized Category Discovery (GCD), which aims to cluster both old and new categories in unlabeled data using some labeled samples from old categories. GCD faces issues such as imbalanced classification performance and inconsistent confidence between old and new classes, especially in low-labeling regimes. AGCD introduces an active learning approach where models select a limited number of valuable samples for labeling from an oracle to improve GCD performance. An adaptive sampling strategy, Adaptive-Novel, is proposed to select samples based on novelty, informativeness, and diversity. A stable label mapping algorithm is also introduced to handle the issue of different label orderings between ground truth labels and model label spaces. The method achieves state-of-the-art performance on both generic and fine-grained datasets. Experiments show that AGCD significantly improves the accuracy of new classes with a limited annotation budget. The method is evaluated on various datasets and outperforms other methods in terms of accuracy and novelty metrics. The results demonstrate that AGCD effectively addresses the inherent issues of GCD, including imbalanced accuracy and confidence between old and new classes.Active Generalized Category Discovery (AGCD) addresses the challenges of Generalized Category Discovery (GCD), which aims to cluster both old and new categories in unlabeled data using some labeled samples from old categories. GCD faces issues such as imbalanced classification performance and inconsistent confidence between old and new classes, especially in low-labeling regimes. AGCD introduces an active learning approach where models select a limited number of valuable samples for labeling from an oracle to improve GCD performance. An adaptive sampling strategy, Adaptive-Novel, is proposed to select samples based on novelty, informativeness, and diversity. A stable label mapping algorithm is also introduced to handle the issue of different label orderings between ground truth labels and model label spaces. The method achieves state-of-the-art performance on both generic and fine-grained datasets. Experiments show that AGCD significantly improves the accuracy of new classes with a limited annotation budget. The method is evaluated on various datasets and outperforms other methods in terms of accuracy and novelty metrics. The results demonstrate that AGCD effectively addresses the inherent issues of GCD, including imbalanced accuracy and confidence between old and new classes.