MEI-Bench

Controlled Evidence Interventions for Diagnosing Visual Grounding in Vision-Language Models

Shahriar Shakir Sumit, Murat Tahtali and Seyedali Mirjalili. ACML 2026 (PMLR).

Code and full toolkit: github.com/ShahriarShakirSumit/mei-bench

What this is

When a vision-language model answers a question about an image correctly, did it actually look at the part of the image that holds the answer? MEI-Bench is built to find out.

Every question comes with two boxes. The evidence box marks the region that holds the answer. The sham box is a region of about the same size somewhere else in the image that holds none of it. You degrade each box in turn (we use a strong blur, a flat gray patch, or a scrambled patch) and check whether the model changes its answer. A model that really uses the evidence should change its answer when the evidence is degraded and keep it when the sham is. If it changes its answer both times, it is fragile rather than grounded, and the sham is what lets you tell the two apart.

In our paper, between 18 and 43% of the sensitivity we measured across ten open-source models turned out to be general fragility rather than grounding.

What's inside

615 questions built from COCO val2017, in five kinds:

Kind of question	Items
Spatial relations ("Is the person to the left or right of the frisbee?")	153
Counting	144
Attribute checks	118
Object identification	114
Reading text in the image	86

The test split has one row per question:

Column	What it holds
item_id	our ID for the question
task_type, difficulty	the kind of question, and easy, medium or hard
question, valid_answers	the question and the answers we accept
answers_verified	false for the 86 text-reading questions, whose answers have not been checked yet (see below)
coco_image_id, coco_file_name, coco_url	which COCO val2017 image the question is about, with a link to it
evidence_bbox	the evidence box, as [x, y, width, height] in pixels
evidence_area_fraction, evidence_description	how much of the image the evidence covers, and what it is
sham_bbox	the sham box, in the same format

The raw/ folder has the original annotation file exactly as used in the paper (the same file as on GitHub), and metadata/ has the schemas, a summary and the report from the filtering step that took us from 959 candidate questions down to 615.

Using it

from datasets import load_dataset

mei = load_dataset("Shahriar10/mei-bench", split="test")
print(mei[0]["question"], mei[0]["valid_answers"], mei[0]["evidence_bbox"])

The images themselves aren't included. They belong to COCO and keep their original licences, so please download COCO val2017 from cocodataset.org (or follow the coco_url links). The GitHub repository has a script that makes all the edited images from the boxes in a few minutes, plus everything you need to run the ten models from the paper and to compute the metrics.

How we score a model

Only the questions a model answers correctly on the untouched image count. On those:

• Causal Sensitivity (CS): how often the answer changes when the evidence is degraded
• Sham Sensitivity (SI): how often it changes when the sham is degraded
• Grounding Gap (GG): CS minus SI
• Grounding Specificity (GS): GG divided by CS, the share of the sensitivity that is really about the evidence

The paper's main results use a Gaussian blur with a kernel of 51 pixels.

Things to know

• The 86 text-reading questions carry a placeholder answer, because their answers haven't been verified. No model scores on them, so they never enter the grounding metrics. Filter on answers_verified if you'd rather leave them out.
• The evidence box covers one annotated object. For one-object questions it includes some background; for spatial and counting questions it covers only one of the objects involved.
• The colour labels in the attribute questions are noisy (about 20%). Leaving those questions out changes the grounding gap by less than one point.
• All images come from one source, COCO val2017.

Licence

The annotations are released under CC BY 4.0. The COCO images are not part of this dataset and stay under the COCO terms of use.

Citing our work

If MEI-Bench is useful in your research, we'd really appreciate it if you cite our paper:

@inproceedings{sumit2026mei,
  title     = {Controlled Evidence Interventions for Diagnosing Visual Grounding in Vision-Language Models},
  author    = {Sumit, Shahriar Shakir and Tahtali, Murat and Mirjalili, Seyedali},
  booktitle = {Proceedings of the 18th Asian Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  publisher = {PMLR},
  year      = {2026}
}

Contact

Questions, problems or ideas are very welcome. Open a discussion here or an issue on GitHub, or write to Shahriar Shakir Sumit at shahriar9121@gmail.com or m.sumit@unsw.edu.au.