Second-order optimization can significantly improve the effectiveness of large language model unlearning without compromising model utility.
