Continual Learning Shifts from Parameter-Centric to System-Level Adaptation

The recent arXiv paper titled 'Continual Learning in Transition' (arXiv:2608.06216) suggests a significant transformation in continual learning (CL) from conventional parameter-focused methods to a more system-oriented adaptation approach. Traditional CL strategies prioritize knowledge retention and updates via parameter-centric techniques, including training methodologies, architectural configurations, and weight adjustments. In contrast, new frameworks are broadening the understanding of CL. On-policy learning introduces diverse update mechanisms, while test-time training transitions CL from the training stage to inference. Additionally, external components like memory, skill libraries, and interaction protocols enhance model capabilities beyond fixed parameter spaces. The paper analyzes this evolution through three aspects: When, How, and Where learning takes place, with the How aspect covering both off-policy and on-policy learning. This cross-type submission is accessible on arXiv.