MARS steers multimodal safety using only text-derived refusal directions, no multimodal safety data needed
MHUGLab · hf · 2026-10-20
Researchers at the University of Trento (Moreno D'Inca, Nicu Sebe, Massimiliano Mancini) published a paper on MLLM safety alignment.
- Existing approaches rely on unsafe multimodal data that is hard to collect, must be gathered per modality, and raises ethical/legal concerns
- They show that textual refusal directions extracted directly from the LLM backbone generalize across modalities (image, video)
- Effectiveness depends on layer selection, steering strength, and cross-modal alignment—the latter can spuriously steer safe inputs toward refusal
- They propose MARS (Modality-Agnostic Refusal Steering), a lightweight, training-free method that injects multimodal safety without any multimodal safety data
More from Safety
- AI agents keep going rogue — Fortune traces it to the labs' own training choices — jeremyakahn · 2026-10-20
- Interpretability researchers debate whether early monosemantic features persist through RLHF — davidmanheim · 2026-10-20
- Vals AI: DeepSeek V4.1 Flash Beats GPT-6 Sol on Offensive Cybersecurity Benchmark — Hesamation · 2026-10-20
- Microsoft expands Sovereign Cloud capabilities in Paris for sensitive workloads — satyanadella · 2026-10-20
- Lenovo paper: safety benchmark scores mostly measure labeling policy, not detection capability — Vahid Jebraeeli · 2026-10-20
- Shipping AI into 5 major banks without the 12-month compliance wait — Cisco's Al Chen — AI Engineer · 2026-10-20