Even without seeing the image, users can decide when to trust a VLM if we show two simple scores beside its explanation. We propose Visual Fidelity and Contrastiveness; together they raise user accuracy by 11.1% and cut belief in wrong answers by 15.4%.
Motivation
In many real settings, the person querying a VLM cannot inspect the visual input. An incorrect answer paired with a fluent explanation is hard to catch.
Asking questions about photos they captured but cannot inspect.
Acting on visual web pages and screens on a user's behalf.
A remote operator relies on what the robot reports seeing.
VLMs hallucinate visual details and make overconfident predictions. Both failure modes are nearly invisible to a user who cannot check the image themselves.
The Gap
Prior explanation qualities were built for text only language models. None of them look at the visual context, and we find they are poorly calibrated to VLM accuracy.
Does the explanation entail the model's prediction?
Hase et al., 2020Does it add new information beyond restating the answer?
Chen et al., 2023Does it agree with commonsense priors?
Liu et al., 2023All three are blind to the image.
Our Approach
Both scores require no training and use off the shelf models.
Running example (used by both scores below)
Q: What time of day does this photo show?
Flags hallucinated details that are not actually visible. Decomposes the explanation into atomic visual claims and verifies each one with a separate verifier VLM. The score is the fraction of claims confirmed.
Half of the explanation is not supported by the image.
Flags when the discriminative cue for the answer is missing. Masks every candidate answer in the explanation and uses a natural language inference model to measure how strongly the masked text entails each option.
| Option | P(expl. entails option) |
|---|---|
| dawn | 0.01 |
| morning | 0.01 |
| noon | 0.98 |
| afternoon | 0.72 |
"noon" scores high, but "afternoon" is barely ruled out, so the discriminative cue is missing.
Combining the two scores
We compare three combinations on every downstream evaluation:
Avg(VF, Contr.), Prod(VF, Contr.), and
Min(VF, Contr.).
The product, Prod(VF, Contr.), is consistently the strongest
and is what we use for the headline user study result.
Evaluation
3 VLMs (LLaVA-1.5-7B, Qwen2.5-VL-7B, GPT-4o) · 3 tasks (A-OKVQA, VizWiz, MMMU-Pro) · 2 calibration metrics (Discriminability, ECE). Combining Visual Fidelity and Contrastiveness gives the best overall result.
Reliability diagrams on A-OKVQA with Qwen2.5-VL-7B. Bars closer to the dashed diagonal mean the score better reflects the actual accuracy. Avg(VF, Contr.) reaches an ECE of 0.054 versus 0.219 for Simulatability and 0.270 for Plausibility.
User Study
Participants see the question, the VLM's answer and explanation, and a quality score. They never see the image. They mark the answer as correct, incorrect, or unsure. A-OKVQA and VizWiz, 100 questions each, 30 participants per condition.
The study interface. Participants see everything except the image.
Every quality score condition beats the explanation only and random score baselines. The combined product gives the largest and most consistent lift.
Effect of presenting different scores to users on A-OKVQA and VizWiz. Asterisks mark significant improvements over the explanation only baseline (*: p < 0.05, **: p < 0.01).
Going Deeper
Relabeling the same number as "Visual Fidelity", "Contrastiveness", or a generic "AI confidence" gives near identical user behavior. The signal's distribution is what drives reliance, not its name.
Replacing a number with a short description of verified versus missing evidence performs comparably and reduces overreliance for Visual Fidelity. Transparency makes the evaluation more interpretable.
Across all conditions, calibration error (ECE) negatively correlates with downstream user accuracy and overreliance. ECE explains roughly 60% of the variance.
Takeaways
Future Directions
Learn when to show or suppress scores, for example with bandit style abstention.
Use quality scores as a training signal for explanation generation.
Study how explanation quality shapes user reliance across repeated interactions.
@inproceedings{he2026believing,
title={Believing without Seeing: Quality Scores for Contextualizing Vision-Language Model Explanations},
author={He, Keyu and Srinivasan, Tejas and Joshi, Brihi and Ren, Xiang and Thomason, Jesse and Swayamdipta, Swabha},
booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
pages={14764--14789},
year={2026}
}