Better way to edit AI math skills without breaking everything

Original: Interpretability-Guided Layer Selection over Subspace Projection: SAEs as Stethoscopes, Not Scalpels, for Raw Task Vector Model Editing

Source: arxiv.org ↗