1 paper
Yuxuan Liu, Yuntian Shi, Kun Wang +2
Multimodal large language models (MLLMs) enable interaction over both text and images, but their safety behavior can be driven by unimodal shortcuts instead of true joint intent un…