Study Reveals Spatial Bias in Vision-Language Models and a Simple Fix

A study finds LVLMs show bias, giving outputs when cues shift position. Balanced Position Assignment (BaPA) gives all visual tokens same positional embedding and added without retraining. getnews.me/study-reveals-spatial-bi... #spatialbias #bapa