Multimodal Prompt Distillation for Continual Learning: Continuous Error Correction

Multimodal continual learning is critical for dynamic scenarios, as it enables systems to incrementally learn across diverse modalities. However, significant challenges persist, including cross-modal representation bias and catastrophic forgetting. To address these issues, this paper proposes a continual learning framework based on Multimodal Prompt Distillation, which enhances the cross-modal alignment capability of the CLIP model through a dual-path adaptive mechanism. In the visual branch, a task-specific prompt pool dynamically calibrates image features, while in the textual branch, context optimization enhances visual-language collaboration. A distillation-based correction prompt mechanism is introduced, where feature distillation preserves representation stability, feature scale and shift adjust features, and correction prompts address misalignments. Finally, feature concatenation training strengthens memory retention. Experimental results demonstrate that the proposed method achieves superior performance across multiple benchmark datasets, including ImageNet-R and CIFAR-100. Extensive ablation studies further validate the effectiveness of both the multimodal prompt collaboration and the distillation strategies.