OVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image Models

Xuandong Zhao (UC Berkeley) · Dawn Song (UC Berkeley) · Ziheng Cheng (University of California, Berkeley) · Yixiao Huang (UC Berkeley) · Hui Xu (Stanford University) · Somayeh Sojoudi (University of California, Berkeley) · Song Mei (University of California, Berkeley)
benign promptscustomized evaluation datafaithfulnessgeneration frameworkharmful promptsover-refusalovertprompt rewritingsafety alignmentsafety-related categoriessafety–utility trade-offsynthetic evaluation datat2i modelstext-to-imageuser-defined policies

Text-to-Image (T2I) models have achieved remarkable success in generating visual content from text inputs. Although multiple safety alignment strategies have been proposed to prevent harmful outputs, they often lead to overly cautious behavior