In this research, they revealed that the VLM can pay more attention to the image simply by chaining attention weights.