We call this Concept Association Bias (CAB). We’ve found that models trained using contrastive loss (e.g. parts of BLIP and BLIP-2) also have CAB. However, models trained with autoregressive loss (e.g. OFA and BLIP-2-FlanT5) don't exhibit this bias. 4/5

→ View original post on X — @_yutaroyamada,