A Low Grounding Score Is Not an Ungrounded Judge: Identifying the Perceptibility Confound in Multimodal Oversight
The paper investigates why low Verdict Grounding Scores (VGS) don't always mean a judge is ungrounded; instead, they can be caused by a 'perceptibility confound.' This happens when the score is limited by how perceptible an edit is to the judge. The authors propose using Grounded-Error Detection Rate (GEDR) to distinguish genuine lack of grounding from false alarms caused by missed edits.