AAA-Bench: Counterfactual, Verifier-Grounded Evaluation of Vision-Language Models as Decision-Makers Under Visual Ambiguity

Shahriar Shakir Sumit¹, Murat Tahtali¹, Seyedali Mirjalili²

¹UNSW Canberra, Australia   ²Torrens University Australia

IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2027

Code, model responses and reproduction scripts: github.com/ShahriarShakirSumit/AAA-Bench

Dataset summary

AAA-Bench evaluates whether a vision-language model makes the right decision when an image does not settle a question: answer it, ask a clarifying question, or abstain. Each item is one member of a counterfactual pair. The two members of a pair have exactly the same question but different COCO images, and their correct actions differ, so the question text carries no information about the correct action. Gold actions are derived from COCO instance annotations by stated rules, and a three-annotator study audits these rules.

The dataset contains 7,400 items in 3,700 pairs over 59 COCO object categories.

Family	Example question	Items	Actions in a pair
Referential (spatial)	"Is the kite on the left or the right side of the image?"	3,306	answer / ask
Referential (attribute)	"What color is the couch?"	1,670	answer / ask
Counting	"How many buses are in the image?"	1,244	answer / abstain
Presence	"Is there an apple in the image?"	1,180	answer / abstain

Dataset structure

Splits. train (4,994 items), validation (956) and test (1,450). The splits are made by pair and by image, so no image appears in two splits and no pair is divided. Evaluation is zero-shot on the test split.

Main fields.

Field	Description
item_id, pair_id	Item and counterfactual pair identifiers.
image_path, image_id	COCO 2017 image, for example train2017/000000384993.jpg.
question	The question, identical for the two members of a pair.
gold_action	answer, ask or abstain.
ambiguity_subtype, decision_basis	Question family, and the scene configuration that determines the action.
visual_evidence	Queried category with the count, areas and boxes of its instances.
clarifications	For ASK items, the clarification given to a model that asks.

The full schema is in data/schema.json of the GitHub repository.

Images. The COCO images are not included. The GitHub repository provides a script that downloads the images an item list needs.

Human study. The human_study configuration contains the 400 items labelled by three annotators, and human_study/annotator_responses.csv contains their labels. These items were built with the same rules but are not benchmark items; match the labels to them through human_study/sample_items.jsonl only.

Usage

from datasets import load_dataset

aaa = load_dataset("Shahriar10/AAA-Bench")          # train, validation and test
test = aaa["test"]

Evaluation

The primary metric is contrastive accuracy (CA): the fraction of pairs whose two items both receive the correct action. Any policy that depends only on the question scores 0, and a random text-only policy at most 0.25. Action accuracy is reported alongside it; always answering scores 0.500. The GitHub repository provides the evaluation protocol, the action parser and the scoring script used in the paper.

On the test split, the seven open VLMs evaluated in the paper (7B to 8B parameters) reach a CA of at most 0.17 with the main prompt and below 0.23 under every prompt. A COCO-pretrained detector that applies the labelling rules to its own detections reaches 0.742, and the human majority reaches 0.560 on the 50 study pairs.

Limitations

• AAA-Bench covers object-instance ambiguity in COCO scenes. Relational, social and temporal ambiguity are out of scope.
• Only the action is scored. The direct_answer field is filled only where COCO determines the answer.
• In 418 of the 1,653 spatial ASK items, both instances lie on the same side of the image, so the two interpretations give the same answer. One test pair uses the same image twice.
• Agreement between the human majority and the labels is lowest for one dominant instance and for crowds. The annotators were members of the research team.

Licence

The annotations and human-study files are released under CC BY 4.0. The COCO images remain subject to the COCO terms of use and the licences of the original Flickr images.

Citation

If you use AAA-Bench, please cite:

@inproceedings{sumit2027aaabench,
  title     = {{AAA-Bench}: Counterfactual, Verifier-Grounded Evaluation of Vision-Language Models as
               Decision-Makers Under Visual Ambiguity},
  author    = {Sumit, Shahriar Shakir and Tahtali, Murat and Mirjalili, Seyedali},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}