Hyejin Go

Stop relying on global image-text alignment scores for vision-language pretraining data curation – a phrase-level sensitivity signal reveals a 50% data subset that substantially boosts compositional generalization.