Does Structural Attention Improve Compositional Representations in Vision-Language Models?

Benchmark Model Rank Results
visual-reasoning-on-winogroundCACR base–Text Score: 39.25Image Score: 17.75Group Score: 14.25
visual-reasoning-on-winogroundIAIS large (COCO)–Text Score: 41.75Image Score: 19.75Group Score: 15.50
visual-reasoning-on-winogroundIAIS large (Flickr30k)–Text Score: 42.50Image Score: 19.75Group Score: 16.00
visual-reasoning-on-winogroundROSITA (Flickr30k)–Text Score: 35.25Image Score: 15.25Group Score: 12.25