Current image-text evaluation often relies on a single global score, which offers little insight into which semantic details are actually grounded in the image. According to the abstract for the new paper, this lack of granularity makes it hard to understand why a model performs well or poorly.

Explainable metrics have been introduced to provide rationales, but the abstract points out that the visual evidence supporting those rationales is missing. The proposed metric, VisExMEM, is designed to fill that gap by grounding explanations in visual evidence.

The paper is available on arXiv under the identifier 2610.16441. As the abstract is truncated in the source, further details about the method and results are not covered here.