Zhetong Zhang

V-DEAL: Diagnosing Video Safety De-Calibration as an Understanding-Refusal Coupling Failure

Zhetong Zhang, Honghao Fu, Miao Xu, Yiwei Wang, Yujun Cai

As Video Large Language Models are increasingly deployed in real-world applications, ensuring their safety alignment has become critical. Counterintuitively, we find that harmful videos paired with benign queries achieve higher attack success rates than the same videos paired wit…Read more