VLM-UQBench

Data for VLM-UQBench: A Benchmark for Modality-Specific and Cross-Modality Uncertainties in Vision Language Models (CVPR 2026 TRUE-V Workshop).

Paper · Code · Project page

A VLM can be uncertain because of the image, the question, or a mismatch between the two. VLM-UQBench tests whether uncertainty quantification (UQ) methods can tell these sources apart, and whether their scores track the hallucinations that degraded inputs cause. This repository holds the questions, images, perturbed images and the model outputs behind the paper's tables.

Contents

Folder	What
manifests/vizwiz_modality/	VizWiz Clean / Image / Text / Cross subsets (expert-annotated)
manifests/focus_ambiguity/	VQ-FocusAmbiguity, 60 ambiguous + 80 unambiguous questions
manifests/clevr_hallucination/	CLEVR attribute / counting / existence / relation questions, 150 each
manifests/perturbations/	visual (8 types), text (5 types) and cross-modal (3 types) perturbations of VizWiz, CLEVR and VQAv2 questions
images/	clean images (VizWiz, CLEVR, the VizWiz images of VQ-FocusAmbiguity)
perturbed/	perturbed images, perturbed/<family>/<perturbation>/<level>/
masks/focus_ambiguity/	VQ-FocusAmbiguity grounding masks
outputs/	answers and UQ scores of Kosmos-2, LLaVA-1.5-7B, Qwen2.5-VL-7B and GPT-4o-mini, and the prompting baseline for ambiguity detection
metadata/file_map.jsonl	SHA-256 of every stored file

Every manifest row has a question and an image path relative to the dataset root. Byte-identical images are stored once, so several rows can point to the same file. VQAv2 uses COCO test2015 images and part of VQ-FocusAmbiguity uses COCO and MSRA-B images; these are not redistributed and their rows have image: null (the source file name is in source_image).

Usage

huggingface-cli download Tianle/VLM-UQBench --repo-type dataset --local-dir data_cache
python data_cache/scripts/verify_release.py --root data_cache   # optional check

from datasets import load_dataset

clean = load_dataset("Tianle/VLM-UQBench", "vizwiz_modality", split="clean")

The model outputs are Parquet files, one row per question and model, with one uq_<method> column per UQ method:

import pandas as pd

qwen = pd.read_parquet("data_cache/outputs/perturbations/qwen.parquet")

To score new models, see scripts/run_uq.py in the code repository.

License and sources

Part	Source	License
VizWiz questions, answers and images	VizWiz-VQA (Gurari et al., CVPR 2018); answer-difference labels (Bhattacharya et al., ICCV 2019)	CC BY 4.0
CLEVR images and scene-graph questions	CLEVR (Johnson et al., CVPR 2017)	CC BY 4.0
VQ-FocusAmbiguity questions and masks	VQ-FocusAmbiguity (Chen et al., ICCV 2025)	CC BY-SA 4.0
VQAv2 questions	VQAv2 (Goyal et al., CVPR 2017); images not included	CC BY 4.0 (annotations)
Perturbed images, subset labels, model outputs	this work	same license as the source images / questions they derive from

See LICENSE.md for details.

Citation

@article{wang2026vlmuqbench,
  title   = {{VLM-UQBench}: A Benchmark for Modality-Specific and Cross-Modality
             Uncertainties in Vision Language Models},
  author  = {Wang, Chenyu and Chen, Tianle and Ahmad, H. M. Sabbir and
             Batmanghelich, Kayhan and Li, Wenchao},
  journal = {arXiv preprint arXiv:2602.09214},
  year    = {2026}
}