MARS steers multimodal safety from textual refusal directions
Original title: Harnessing Textual Refusal Directions for Multimodal Safety
The training-free method was tested on five state-of-the-art MLLMs for safety, utility, and video jailbreaks.
The paper studies whether refusal directions taken from an LLM backbone transfer to image and video inputs in multimodal models. Preliminary results say they can, but layer choice, steering strength, and cross-modal alignment affect outcomes, and misalignment can push safe multimodal inputs toward refusal. The authors introduce Modality-Agnostic Refusal Steering (MARS), a training-free method that centers activations, scales steering inside a geometric trust region, and picks an intervention layer at the first generated token. They report consistent safety gains on five state-of-the-art MLLMs while preserving utility up to over-refusal, without multimodal safety data.
Unconfirmed: Safety and utility gains are reported by the authors; model names and exact scores are not given.
- MARS does not require collecting multimodal unsafe training data.
- Steering is applied at the first generated token.
- Activation centering is used to correct modality misalignment.
- Evaluation covers safety, utility, and video jailbreak settings.
From the original article
1 min read at huggingface.co
Safety in Multimodal Large Language Models (MLLMs) depends on unsafe multimodal data for post-training alignment or activation steering. However, such data is hard to collect, must be gathered anew for each modality, and its collection and distribution raise ethical and legal concerns.
By Moreno D'Incà, Nicu Sebe, Massimiliano Mancini · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.