Vision-Language Models (VLMs) extend Large Language Models (LLMs) with visual perception capabilities, unlocking broad applications across many domains. However, ensuring their safety remains a critical challenge, as adversarial visual inputs can easily bypass built-in safeguards and elicit harmful content. In this paper, we uncover a phenomenon we call delayed safety awareness, where a jailbroken VLM initially produces harmful content but ultimately recognizes the harmfulness at the end of the generation process. We attribute this phenomenon to the fact that the model's safety awareness against jailbreaks cannot be effectively transferred to the intermediate stages of text generation. Motivated by this insight, we introduce SafetyReminder, a simple yet effective defense that optimizes a learnable soft prompt using our proposed Safety-Activation Prompt Tuning (SAPT). This soft prompt is inserted into the generated text to activate the safety awareness of the model, steering it toward refusal when harmful content arises while preserving helpfulness in benign scenarios. We evaluate our method on three established harmful benchmarks and across three types of adversarial attacks. Experimental results demonstrate that our method achieves state-of-the-art defense performance with strong generalization, offering a practical and lightweight solution for safe deployment of VLMs.

Deep learning models often suffer from performance degradation in unseen domains, posing a risk for safety-critical applications such as autonomous driving. To tackle this problem, recent studies have leveraged pre-trained Visual Foundation Models (VFMs) to enhance generalization. However, exsiting works mainly focus on designing intricate networks for VFMs, neglecting their inherent strong generalization potential. Moreover, these methods typically perform inference on low-resolution images. The loss of detail hinders accurate predictions in unseen domains, especially for small objects. In this paper, we argue that simply fine-tuning VFMs and leveraging high-resolution images unleash the power of VFMs for generalizable semantic segmentation. Therefore, we design a VFM-based segmentation network (VFMNet) that adapts VFMs to this task with minimal fine-tuning, preserving their generalizable knowledge. Then, to fully utilize high-resolution images, we train a Mask-guided Refinement Network (MGRNet) to refine VFMNet's predictions combining detailed image features. Furthermore, we adopt a two-stage coarse-to-fine inference approach. MGRNet is used to refine the low-confidence regions predicted by VFMNet to obtain fine-grained results. Extensive experiments demonstrate the effectiveness of our method, outperforming state-of-the-art methods by 3.3% on the average mIoU in synthetic-to-real domain generalization.

Collecting and annotating data for semantic segmentation can end up costing a lot of time and energy. Unsupervised Domain Adaptation (UDA) for semantic segmentation allows models trained on certain source domain data (such as the GTA synthetic dataset) to be applied to certain target data (like the Cityscapes real dataset), significantly decreasing the requirement for target-domain manual pixel-level annotations.In this paper, we suggest a cross-domain unified semantic segmentation network training framework, the Attention Distribution Adaptation Network (ADAN). It (1) proposes the Content Consistency Transformation (CCT), which maps source domain data to an intermediate domain that has a data distribution that is comparable to that of the target domain, and (2) introduces the Attention Adaptation Module (AAM), which bolsters the synchronization of attention between the intermediate domain and the target domain. ADAN achieved unprecedented mIoUs of 77.2 and 68.8 on GTA→Cityscapes and Synthia→Cityscapes, respectively, corresponding to improvements of +1.3 and +0.6 over the state of the art.