Yoav Goldberg suggests Linear Representation Editing on broad distributions captures topical relatedness rather than specific relations

@yoavgo Yes, it's cool to take the Hernandez LRE jacobian and apply it the broad training distribution! Surprising that it works! Through the lens of @jack_merullo_ ICLR2025 (that higher frequency relations are more LRE): it leads to the question: what "universal relation" is this?