Jun 2026
Harnessing Textual Refusal Directions for Multimodal Safety
Modality-Agnostic Refusal Steering (MARS), a light-weight training-free approach that injects multimodal safety without the need for multimodal safety data, achieves consistent safety gains while preserving utility.
Moreno D'Incà, Massimiliano Mancini, N. Sebe
· arXiv.org · 0 citations