LLMs tend to lose prior skills when fine-tuned for new tasks. A new self-distillation approach aims to reduce regression and ...