@inproceedings{liu-etal-2025-unraveling-mitigating,
    title = "Unraveling and Mitigating Safety Alignment Degradation of Vision-Language Models",
    author = "Liu, Qin  and
      Shang, Chao  and
      Liu, Ling  and
      Pappas, Nikolaos  and
      Ma, Jie  and
      Anna John, Neha  and
      Doss, Srikanth  and
      Marquez, Lluis  and
      Ballesteros, Miguel  and
      Benajiba, Yassine",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/landing_page/2025.findings-acl.186/",
    pages = "3631--3643",
    ISBN = "979-8-89176-256-5",
    abstract = "The safety alignment ability of Vision-Language Models (VLMs) is prone to be degraded by the integration of the vision module compared to its LLM backbone. We investigate this phenomenon, dubbed as ``safety alignment degradation'' in this paper, and show that the challenge arises from the representation gap that emerges when introducing vision modality to VLMs. In particular, we show that the representations of multi-modal inputs shift away from that of text-only inputs which represent the distribution that the LLM backbone is optimized for. At the same time, the safety alignment capabilities, initially developed within the textual embedding space, do not successfully transfer to this new multi-modal representation space. To reduce safety alignment degradation, we introduce Cross-Modality Representation Manipulation (CMRM), an inference time representation intervention method for recovering the safety alignment ability that is inherent in the LLM backbone of VLMs, while simultaneously preserving the functional capabilities of VLMs. The empirical results show that our framework significantly recovers the alignment ability that is inherited from the LLM backbone with minimal impact on the fluency and linguistic capabilities of pre-trained VLMs even without additional training. Specifically, the unsafe rate of LLaVA-7B on multi-modal input can be reduced from 61.53{\%} to as low as 3.15{\%} with only inference-time intervention."
}