Search papers, labs, and topics across Lattice.
This paper investigates the phenomenon of cross-modal safety drift in multimodal large language models (MLLMs), where benign textual queries can lead to harmful interpretations when paired with visual inputs. Through empirical analysis, the authors identify unsafe response patterns and demonstrate that visual cues receive insufficient attention, resulting in inadequate safety responses. They introduce a novel method called safety-awareness representation transfer (SRT), which enhances safety without compromising model utility by refining the model's attention on safety signals derived from textual inputs.
Cross-modal safety drift can significantly undermine the safety of multimodal language models, but a new method shows how to effectively transfer safety awareness from text to mitigate this risk.
Visual modality enhances the capabilities of multimodal large language models (MLLMs) but also introduces a safety concern: a benign textual query may convey harmful intent when grounded in a visual image. We term this cross-modal safety drift and our pilot studies show that the safety response rate for such requests is substantially lower than that for requests containing explicitly unsafe text. This paper aims to systematically study this issue. First, we conduct an empirical analysis to identify representative unsafe response patterns. Building on these, we interpret model representations and attentions, revealing that visually risky cues receive limited attention and weakly trigger refusal. Motivated by the observation that safety signals from unsafe text processing can be transferred, we propose safety-awareness representation transfer (SRT), a lightweight direction-refinement method that mitigates cross-modal safety drift with a frozen MLLM backbone. Experiments across multiple benchmarks and models show that SRT effectively improves safety in diverse cross-modal settings while preserving utility. Code is available at https://github.com/cucu220123/safety-awareness.
