How to find the right negative examples for spotting strange images?

When vision-language models detect images from unknown classes, they compare test inputs against known classes and deliberately chosen "negative" examples. The problem: existing methods pick negatives haphazardly from unlabeled data, introducing bias that hurts detection. This work corrects that bias by approximating the true distribution of negative labels, converted into a tractable sampling procedure. The result is measurably better OOD detection across multiple datasets.