Task-Differentiated Atomic Skill Expansion and Routing for Continual Learning Across Highly Heterogeneous Tasks

The article presents Task-Differentiated Atomic Skill Expansion and Routing (TASER), a continual learning framework designed to tackle challenges in heterogeneous task environments by dynamically expanding atomic skills based on task divergence and model uncertainty. TASER employs orthogonality-enhanced skill detection and a skill dynamic routing mechanism to ensure skills are semantically distinct and task-relevant. The introduction of the HeteroCLBench benchmark, which includes 19 diverse tasks across 9 cognitive dimensions, demonstrates TASER's superior performance in enhancing model plasticity and mitigating catastrophic forgetting compared to existing methods.