Hierarchical Decomposition of Prompt-Based Continual Learning: Rethinking Obscured Sub-optimality