HalluScope benchmark

HalluScope tests why a vision-language model hallucinates objects. Each image comes with four questions:

Question set	Asks about	Example	Correct answer
real_recognition	an object that is in the image	"Is there a baseball bat in the image?"	yes
random_recognition	an absent object, unrelated to the scene	"Is there a statue in the image?"	no
adversary_recognition	an absent object that usually appears in such scenes	"Does the image contain a baseball cap?"	no
adversary_presupposition	the same absent object, in a question that assumes it is there	"What logo appears on the baseball cap worn by the batter?"	an answer that does not claim the object exists

Comparing the four tells apart the causes: a model failing only the last one is misled by the question, not by what it sees.

Images are from COCO train2017 (not included; use the image file name). There are two subsets of 3,000 images each, with the same images across the four question sets of a subset:

• florence: objects detected by Florence-2, then verified;
• instances: COCO's own object annotations, then verified.

Versions

Version	What it is
v1	The test set used for all results in the paper. Built with the earlier version of the pipeline, where the vision-language model used for captioning and for checking which objects are present or absent was Qwen2-VL-7B-Instruct (instead of Qwen3.5-9B in v2).
v2	Regenerated with the released, cleaned-up pipeline (same method). Use it to compare with models evaluated on v2, or to reproduce the construction end to end.

Scores on v1 and v2 are not directly comparable: the images and questions differ. Report which version you use.

What changed in v2: captions and verification with Qwen3.5-9B; Florence label cleanup (broken or duplicate labels fixed or removed); a stricter check before an object is considered present or absent; smoothed co-occurrence statistics so that rare labels are no longer over-chosen as absent objects; fixes in question generation (questions can no longer be matched to the wrong image, plural-aware yes/no questions).

Format

Each file is a JSON array; each item is, for example (v2, adversary_presupposition, instances; the image shows a slice of cake with a fork, and no knife):

_{COCO image 000000448600 (from Flickr), licensed under CC BY-NC-SA 2.0.}

{
  "image": "000000448600.jpg",
  "present_object": "fork",
  "absent_object": "knife",
  "random_object": "toilet",
  "text": "What material is the knife beside the plate made of?"
}

text is the question; present_object, absent_object and random_object say which object each question set is about.

Loading

from datasets import load_dataset

questions = load_dataset("PekhPekhPekh/HalluScope", "v2_florence", split="adversary_presupposition")

Or use the JSON files directly with the evaluation scripts of the HalluScope code (src/evaluation, --version v1|v2).

Citation

@inproceedings{khayatan2026prompts,
  title     = {When Prompts Override Vision: Prompt-Induced Hallucinations in {LVLMs}},
  author    = {Khayatan, Pegah and Parekh, Jayneel and Dapogny, Arnaud and Shukor, Mustafa and Newson, Alasdair and Cord, Matthieu},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}