Existing continual learning works explored strategies like memory replay, regularization, and parameter isolation, but little analysis was conducted on the optimization behavior of LLMs’ continual ...