Safe-Unsafe Concept Separation Emerges from a Single Direction in Language Models Activation Space.

Andrea Ermellino, Lorenzo Malandri, Fabio Mercorio, Antonio Serino

Browse the full EACL paper archive.