Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond