<p>Prompt learning on a pretrained large vision model has turned out to be an effective direction to alleviate forgetting. It either uses a shared set of prompts with task-wise weights or directly optimizes task-specific prompts. Thus only task-relevant prompts are updated and therefore the forgetting is reduced. However, such a task-wise updating paradigm limits the exploration of complementary information from other prompts belonging to similar tasks. In this paper, we propose Prompt Grouping for continual learning. In our approach, prompts are grouped by introducing a new concept, prompt sensitivity to indicate the importance of each prompt w.r.t. the learning task. According to this score, once we find a set of prompts all have a significant impact on the corresponding tasks, we will group these tasks and related prompts, such that tasks within a group can share and benefit from the group prompts. Additionally, we introduce a correction strategy that enhances inter-task discrimination in terms of both the feature extractor and the classification head. Experimentally, our approach achieves consistent performance improvements on various benchmarks, such as CIFAR-100 (+2.31% Accuracy), ImageNet-R (+3.04% Accuracy), and DomainNet (+2.19% Accuracy).</p>