Preprint
CLIP Exhibits Improved Compositional Generalization Through Representation Disentanglement
Reza Abbasi, Amirarshia Hemmat, Mahdi Samiei, Mohammad Hossein Rohban, Mahdieh Soleymani Baghshah
How much of CLIP's compositional generalisation can be traced to disentangled representations rather than data scale.
A longer write-up of this work is on its way. Until then the paper and the code are the best place to look.