Self-Regulating Prompt Expansion for Continual Learning

Prompt-based continual learning (CL) has emerged as an effective paradigm for adapting pre-trained models (PTMs) to sequential tasks. By enabling parameter-efficient tuning, it mitigates catastrophic forgetting (CF) without updating the entire model. Existing methods typically rely on static prompt allocation strategies, assigning one or a fixed number of prompts per task to reduce task interference. However, this design limits knowledge reuse and imposes a fixed prompt capacity, leading to linear prompt growth and suboptimal performance across diverse task distributions. To solve these limitations, we propose Self-regulating COntinual Prompt Expansion (SCOPE), a framework that dynamically allocates prompt capacity on demand. Instead of treating prompts as task-specific slots, SCOPE models them as adaptive representation units and regulates their expansion through a Cross-Scale Prompt Regulation (CSPR) mechanism. At the inter-task level, SCOPE determines whether to reuse existing prompts or introduce new ones based on task compatibility. This design facilitates prompt reuse while effectively preventing interference. At the intra-task level, SCOPE continuously monitors representation coverage during training and expands prompt capacity only when the current prompts are insufficient. As a result, complex tasks can acquire additional capacity, while simpler tasks remain compact. Extensive experiments demonstrate the effectiveness of the proposed SCOPE, achieving state-of-the-art performance compared to prompt-based CL methods across all benchmarks and task granularities.