MolmoPoint: Better pointing architecture for vision-language models