New VLM explanation scores improve correctness judgments by 11.1% in user study

_jessethomason_ · x · 2026-07-28

A USC and Carnegie Mellon team proposes two quality scores for VLM explanations: Visual Fidelity, which measures how faithful an explanation is to the visual context, and Contrastiveness, which measures whether the explanation highlights details that separate the prediction from plausible alternatives.

On A-OKVQA, VizWiz, and MMMU-Pro, these scores are better calibrated with correctness than prior explanation-quality measures. In a user study where participants judged whether VLM predictions were accurate without seeing the image, showing the scores improved accuracy by 11.1% and reduced the rate of falsely believing incorrect answers by 15.4%.

Related event: ACL 2026 Paper Proposes New Metrics for VLM Interpretability(2 posts)→

Original post →

More from Research

Research channel →