Abstract page for arXiv paper 2611.09897: Process then Retrieve: Investigating the Origins of the Modality Gap in Vision-Language Models