📄 中文摘要
提示词驱动的持续学习方法在缓解灾难性遗忘方面表现出色，但现有方法普遍为每个任务分配一组固定的提示词，导致任务间知识完全隔离，参数利用率不佳。针对这一问题，本研究从持续学习的实际需求出发，提出了一种提示词共享框架。该框架构建了一个全局提示词池，并引入了任务感知的门控路由机制。具体而言，当新任务到来时，模型并非直接创建新的提示词，而是从全局池中选择并组合与当前任务相关的现有提示词。这种机制允许不同任务共享和复用提示词，从而促进知识的迁移和整合，有效提升参数利用效率。任务感知的门控路由通过学习任务特征，动态地决定哪些提示词应被激活和组合，以适应当前任务的需求，同时避免对先前任务的性能产生负面影响。
📄 English Summary
Is Parameter Isolation Better for Prompt-Based Continual Learning?
Prompt-based continual learning methods are effective in mitigating catastrophic forgetting. However, most existing approaches assign a fixed set of prompts to each task, leading to complete knowledge isolation across tasks and suboptimal parameter utilization. Addressing this limitation, a prompt-sharing framework is proposed, considering the practical requirements of continual learning. This framework establishes a global prompt pool and incorporates a task-aware gated routing mechanism. Specifically, upon encountering a new task, the model does not generate entirely new prompts but instead selects and combines relevant existing prompts from the global pool. This mechanism enables different tasks to share and reuse prompts, thereby facilitating knowledge transfer and integration, and significantly enhancing parameter utilization efficiency. The task-aware gated routing dynamically determines which prompts should be activated and combined by learning task-specific features, adapting to the current task's demands while preventing negative impacts on previously learned tasks. Furthermore, the framework includes a prompt update strategy, allowing for fine-tuning or updating prompts within the pool to adapt to evolving learning environments and the introduction of new tasks. Experimental evaluations demonstrate that this prompt-sharing framework achieves comparable or improved performance on various continual learning benchmarks while substantially reducing model parameter redundancy and enhancing learning efficiency and knowledge transfer capabilities, compared to methods employing complete parameter isolation. This approach offers a more efficient and flexible solution for continual learning, particularly beneficial in resource-constrained scenarios or contexts requiring rapid adaptation to new tasks.