Impact of Pretraining Word Co-Occurrence on Compositional Generalization In

a subtle but powerful insight: large multimodal models like CLIP don’t just learn individual concepts. they also depend heavily on how often those concepts appear together during training.