Reproduce “DUAL-Bench: Measuring Over-Refusal and Robustness in Vision-Language Models” on OpenResearch

About this paper

Kaixuan Ren, Preslav Nakov, Usman Naseem

DUAL-Bench, a new multimodal benchmark, systematically evaluates Vision-Language Models for visual over-refusal and safe completion in dual-use scenarios. The study reveals that current safety alignment often defaults to binary refusal and exhibits fragility under visual perturbations, highlighting a struggle to balance helpfulness and harmlessness.

Key insights

• VLM safety behavior is primarily driven by specific safety alignment paradigms and developer policies, not merely model parameter size, often resulting in a 'binary trap' of over-refusal or unsafe output.
• Current safety alignment strategies are largely inadequate for achieving 'safe completion', with most models rarely demonstrating the ability to fulfill benign parts of a request while warning about harmful elements.
• VLM safety boundaries are fragile, as minor visual and semantic perturbations frequently degrade performance and disrupt refusal mechanisms, leading to unwarranted rejections or compromised safety.

Summary by alphaXiv.