Figure 2: Overview of the proposed DCHA framework. (a) DCHA constructs category-specific
LoRA by composing vectors from a shared bank and averages the resulting category updates to form
the task-level adaptation for the VLF layers. (b) LCAR transforms category-specific routing logits
into sparse composition weights using a learnable sparsity coefficient, enabling category-dependent
vector selection. (c) HDC preserves historical adaptation knowledge by constraining deviations at
the bank-vector, category-update, and task-update levels.