license: apache-2.0
task_categories:
- image-text-to-text
language:
- en
tags:
- multimodal
- vlm
- benchmark
- visualized-text
VISTA-Bench
VISTA-Bench is a systematic benchmark spanning multimodal perception, reasoning, and unimodal understanding. It evaluates visualized text understanding by contrasting pure-text and visualized-text (VT) questions under controlled rendering conditions.
Dataset Summary
Existing benchmarks predominantly focus on pure-text queries, but in real-world scenarios, language frequently appears as visualized text embedded in images. VISTA-Bench evaluates whether current Vision-Language Models (VLMs) handle such input requests comparably. Extensive evaluation reveals a pronounced modality gap: models that perform well on pure-text queries often degrade substantially when equivalent semantic content is presented as visualized text.
- Size: 1,500 instances
- Composition: Predominantly multiple-choice questions (MCQ), with a small portion of open-ended queries
- Task Taxonomy:
- Unimodal Knowledge: 500 instances
- Multimodal Knowledge: 400 instances
- Multimodal Perception: 300 instances
- Multimodal Reasoning: 300 instances
Repository Structure
VISTA-Bench/
├─ images/ # original images (for multimodal instances)
├─ questions/ # rendered question/option images (VT setting)
├─ VLMEvalKit/ # evaluation toolkit
├─ VISTA-Bench.tsv # dataset index
└─ VISTA-Bench-VT.tsv # dataset index (visualized text variant)
Evaluation (VLMEvalKit)
VISTA-Bench is evaluated using VLMEvalKit. Before running evaluation, it is recommended to convert the TSV file(s) into a normalized format with absolute image paths.
1) Convert TSV to normalized paths
Use the provided helper script to normalize the paths:
python VISTA-Bench/VLMEvalKit/utils/convert_data_file.py \
--in VISTA-Bench/VISTA-Bench.tsv \
--out VISTA-Bench/VISTA-Bench_norm.tsv \
--image-prefix /ABS/PATH/TO/VISTA-Bench
2) Run evaluation
Pure-text setting:
python /VISTA-Bench/VLMEvalKit/run.py \
--data VISTA-Bench_norm \
--model llava_v1.5_7b \
--verbose
Visualized-text (VT) setting:
python /VISTA-Bench/VLMEvalKit/run.py \
--data VISTA-Bench-VT \
--model llava_v1.5_7b \
--verbose
Citation
If you find this dataset useful, please cite the following paper:
@article{liu2026vistabench,
title={VISTA-Bench: Do Vision-Language Models Really Understand Visualized Text as Well as Pure Text?},
author={Liu, Qing'an and Feng, Juntong and Wang, Yuhao and Han, Xinzhe and Cheng, Yujie and Zhu, Yue and Diao, Haiwen and Zhuge, Yunzhi and Lu, Huchuan},
journal={arXiv preprint arXiv:2602.04802},
year={2026}
}