Tag
GGSS reduces demographic bias in generative vision-language models by steering visual tokens along geodesic arcs with an adaptive gate during inference, preserving visual-language accuracy.