Believing without Seeing: Quality Scores for Contextualizing Vision-Language Model Explanations

TL;DR

This work introduces two image-anchored explanation-quality cues, Visual Fidelity and Contrastiveness, to help users judge VLM predictions when visual context is unavailable. It defines training-free scoring functions for these cues, shows they align better with prediction correctness than traditional text-based metrics, and demonstrates that combining them (especially via the product) yields substantial gains in user accuracy and reductions in over-reliance. Through extensive experiments on A-OKVQA and VizWiz and multiple VLMs, the authors establish improved calibration and discriminability of the quality signals. User studies show that presenting these quality cues, particularly in descriptive formulations, meaningfully improves decision making and trust calibration in human–AI interactions. The findings underscore the practical value of quantifying and communicating explanation quality to users relying on VLMs in contexts with asymmetrical information access.

Abstract

When people query Vision-Language Models (VLMs) but cannot see the accompanying visual context (e.g. for blind and low-vision users), augmenting VLM predictions with natural language explanations can signal which model predictions are reliable. However, prior work has found that explanations can easily convince users that inaccurate VLM predictions are correct. To remedy undesirable overreliance on VLM predictions, we propose evaluating two complementary qualities of VLM-generated explanations via two quality scoring functions. We propose Visual Fidelity, which captures how faithful an explanation is to the visual context, and Contrastiveness, which captures how well the explanation identifies visual details that distinguish the model's prediction from plausible alternatives. On the A-OKVQA and VizWiz tasks, these quality scoring functions are better calibrated with model correctness than existing explanation qualities. We conduct a user study in which participants have to decide whether a VLM prediction is accurate without viewing its visual context. We observe that showing our quality scores alongside VLM explanations improves participants' accuracy at predicting VLM correctness by 11.1%, including a 15.4% reduction in the rate of falsely believing incorrect predictions. These findings highlight the utility of explanation quality scores in fostering appropriate reliance on VLM predictions.

Figures (9)

• Figure 1: VLM explanations that sound plausible can mislead users. Contextualizing VLM explanations with quality scores may help users know when to rely on VLM outputs, but existing explanation qualities are not calibrated with VLM prediction accuracy. We propose evaluating two new qualities of VLM explanations: Visual Fidelity and Contrastiveness. These qualities are better calibrated with VLM correctness, and also help users make better decisions about when to believe VLM predictions.
• Figure 2: Calibration curves for various quality scoring functions when evaluating explanations generated by Qwen2.5-VL-7B on the A-OKVQA dataset.
• Figure 3: Our study interface where users are shown the visual reasoning question, VLM prediction and explanation, and optionally one or more quality scores (using simplified language descriptions of both qualities). Here, the user believes that the VLM prediction is incorrect.
• Figure 4: Effect of showing users different quality scores on User Accuracy, Over-Reliance and Under-Reliance. Error bars represent the standard deviation of the data. Asterisks denote improvements over the explanation‐only baseline using a bootstrap significance test (*: $p<0.05$, **: $p<0.01$).
• Figure 5: Ablation on presentation types on A-OKVQA dataset: holding the numeric signal fixed (Prod($\mathrm{VF}$, $\mathrm{Contr.}$), $\mathrm{VF}$, or $\mathrm{Contr.}$), we vary the on-screen label (shown as VF, shown as Contr, or simple confidence). User Accuracy ($\uparrow$) and Over-Reliance ($\downarrow$) are effectively unchanged by naming/framing.