LiNeS: Post-training layer scaling prevents forgetting and enhances model merging

Publication
International Conference on Learning Representations (ICLR)