ConvPrompt is a novel approach for continual learning (CL) that combines convolutional prompt generation with task similarity-based expansion. It addresses the limitations of existing prompt-based methods by using convolution to generate task-specific prompts from shared embeddings, enabling efficient knowledge transfer across tasks. ConvPrompt maintains layer-wise shared embeddings, allowing both layer-specific learning and better concept transfer. It also leverages large language models (LLMs) to generate fine-grained text descriptions of categories, which are used to determine task similarity and dynamically adjust the number of prompts. This approach significantly reduces parameter overhead while improving performance. ConvPrompt outperforms state-of-the-art methods by up to 3% with significantly fewer parameters. The method is evaluated on benchmark datasets such as ImageNet-R, CIFAR-100, and CUB-200, showing superior performance in both accuracy and parameter efficiency. ConvPrompt's design allows for efficient adaptation to new tasks by balancing knowledge from previous tasks and new information. The approach also incorporates regularization to prevent overfitting and ensures that parameters remain close to previous solutions. The use of LLMs for task similarity helps reduce the number of learnable parameters, making the model more efficient. Overall, ConvPrompt demonstrates effective knowledge sharing, efficient adaptation, and superior performance in continual learning tasks.ConvPrompt is a novel approach for continual learning (CL) that combines convolutional prompt generation with task similarity-based expansion. It addresses the limitations of existing prompt-based methods by using convolution to generate task-specific prompts from shared embeddings, enabling efficient knowledge transfer across tasks. ConvPrompt maintains layer-wise shared embeddings, allowing both layer-specific learning and better concept transfer. It also leverages large language models (LLMs) to generate fine-grained text descriptions of categories, which are used to determine task similarity and dynamically adjust the number of prompts. This approach significantly reduces parameter overhead while improving performance. ConvPrompt outperforms state-of-the-art methods by up to 3% with significantly fewer parameters. The method is evaluated on benchmark datasets such as ImageNet-R, CIFAR-100, and CUB-200, showing superior performance in both accuracy and parameter efficiency. ConvPrompt's design allows for efficient adaptation to new tasks by balancing knowledge from previous tasks and new information. The approach also incorporates regularization to prevent overfitting and ensures that parameters remain close to previous solutions. The use of LLMs for task similarity helps reduce the number of learnable parameters, making the model more efficient. Overall, ConvPrompt demonstrates effective knowledge sharing, efficient adaptation, and superior performance in continual learning tasks.