SafeRI proposes an on-demand safety alignment approach for large vision-language models, contrasting with existing always-on methods that globally modify model behavior. The framework uses a lightweight recognizer to evaluate token-level safety during autoregressive generation, gating a LoRA module that only activates when unsafe content is detected. By training the LoRA on unsafe prefixes and safe continuations, SafeRI redirects unsafe generations back to safe responses without perturbing the model’s original reasoning path.
By Caoyuan Ma, Tian Gu, Wenpu Liu, Weichu Xie, Shuai Dong, Yuqi Xu, Ji Zhao, Ziyue Wang, Wenzheng Chang, Taiqiang Wu, Yongfu Zhu, Wenqi Shao, Zheng Wang, Yinqiang Zheng
The paper introduces MPS-Bench, a benchmark of 5,181 scenarios from 584 real-world images across 12 high-risk domains, each paired with a hidden user profile, to evaluate personalized safety in vision‑language models (VLMs). Eight leading VLMs were tested and found to almost always respond directly (86‑99%) without seeking missing context, scoring no higher than 2.6/5 on personalized safety. The authors identify a phenomenon called visual dominance, where visual information enters text representations early and suppresses textual risk signals, and propose PRISM, a lightweight input monitor that predicts when a query should be deferred, achieving 0.978 AUC and outperforming all tested models on the safety‑utility Pareto frontier.
By Edward Sun, Yuchen Wu, Zixian Ma, Eric Hanchen Jiang, Yijia Xiao, Xiaoyuan Yi, Ranjay Krishna, Wei Wang, Jindong Wang, Aylin Caliskan
The paper demonstrates that a targeted adversarial perturbation can reduce a vision‑language model’s training loss to near zero for a fixed target caption, yet the same model, when generating freely, still produces the correct description. This phenomenon, termed the train/inference gap, is traced to a single autoregressive step where the target token’s rank is fixed across all images, and further analysis shows that the language decoder, rather than the visual encoder, determines whether the corrupted signal is amplified or suppressed. The study uses a controlled two‑stage PGD attack on Qwen2.5‑VL‑7B‑Instruct and evaluates the effect on 200 held‑out COCO images, revealing that adversarial robustness in autoregressive VLMs largely depends on the language decoder’s prior.
whyItMatters":"The findings suggest that defenses and faithfulness evaluations for deployed vision‑language models should focus on the language decoder rather than the visual encoder, as the former is the key determinant of robustness to adversarial perturbations."
By Arun Josephraj Arokiaraj, Zekun Wu, Adriano Koshiyama
Aligned vision‑language models (VLMs) are designed to combine grounded visual reasoning with safe generation. The study finds that when safety constraints are applied, these models often abstain from answering questions that they could answer under default instruction, yet visual evidence continues to influence the decoding process. The authors show that safety‑induced abstention alters late‑stage hidden‑state dynamics, and that targeted interventions can restore grounded answering without retraining or changing visual inputs.
As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanis...
arXiv:2603.17372v2 Announce Type: replace-cross
Abstract: Large vision-language models (VLMs) often exhibit weakened safety alignment with the integration of the visual modality. Even when text promp...
By Zhihua Wei, Qiang Li, Jian Ruan, Zhenxin Qin, Leilei Wen, Dongrui Liu, Wen Shen
arXiv:2609.00051v1 Announce Type: cross
Abstract: Despite extensive alignment efforts, Large Language Models (LLMs) remain vulnerable to generating unsafe content under adversarial prompting, yet the...
By Kuan-Lin Chu, Chung-En Sun, Tsui-Wei Weng
The paper introduces Counterfactual Neighborhood Risk Fields (CNRF), an evidence‑grounded framework that improves safety for vision‑language models (VLMs). CNRF builds context‑matched benign‑harmful counterfactual pairs, uses their representation differences to define risk directions, and separates neighborhood support from risk estimation. This approach allows selective intervention rather than blanket rejection, achieving high macro defense success rates across multiple VLMs while preserving benign‑task utility.
By Tianshi Wang, Jingsong Wang, Fengling Li, Jingjing Li, Xin Li, Lei Zhu
arXiv:2609.38810v1 Announce Type: cross
Abstract: As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual sig...
By Chunzheng Zhu, Jiaqi Zeng, Hongbo Zhao, Yihang Chen, Yijun Wang, Jianxin Lin
The paper investigates cross‑modal safety drift in multimodal large language models, where a harmless text query paired with a visual image can trigger harmful responses. Empirical analysis identifies unsafe response patterns and shows that visual cues receive limited attention, weakening refusal mechanisms. The authors introduce Safety‑Awareness Representation Transfer (SRT), a lightweight method that transfers safety signals from text processing to mitigate cross‑modal drift while maintaining model utility.
By Tianqi Xiao, Shiyao Cui, Minghao Zhang, Junxiao Yang, Renmiao Chen
The paper presents a preprocessor that recovers and decodes encoded content in vision‑language models to close the decode gap that allows harmful requests to bypass safety classifiers. Evaluated against eleven encoding attacks, the preprocessor raises block rates from 0 % to 67‑90 % but also increases benign over‑refusal, and no configuration achieves an ensemble attack‑success rate below 40 % while keeping benign over‑refusal under 70 %. The study shows that closing one encoding channel merely relocates success rather than eliminating it, highlighting the limits of recovery‑based defenses.
By Haoyu Zhang, Zhuoxi Wang, Shibo Zheng, Hanwen Liu, Yi Feng, Haowen Xu, Xiangchen Guan, Yang Chen, Zijian Xiao, Xiao Luo, Mohammad Zandsalimy, Shanu Sushmita
arXiv:2610.07774v1 Announce Type: new
Abstract: Vision-language models (VLMs) face compositional safety risks where harmful intent emerges from the interaction between visual and textual inputs. As m...
By Ziyuan Yang, Wenxuan Ding, Shangbin Feng, Yulia Tsvetkov