Compositional Representation Learning

The field studies how models represent reusable concepts, attributes, objects, and relations so they can combine them in new ways. Approaches include factorized and tensor-product representations, object-centric or graph-based models, sparse autoencoders, and contrastive learning for vision-language models. Researchers test these ideas on compositional zero-shot recognition, image-text matching benchmarks such as SugarCrepe, and controlled tasks such as CLEVR, while probing learned features and their interactions. A central tension is that strong performance on unseen combinations does not by itself show that a model learned reusable structure: shortcuts and dataset biases can also produce apparent generalization.