Tag
This paper proposes a spectral-based framework for selectively reversing knowledge edits in large language models, allowing removal of harmful edits while preserving beneficial ones.