Worldwide-LiveVQA

Language: English | 简体中文

Worldwide-LiveVQA is a multilingual visual question answering benchmark covering news, scientific papers, and images extracted from videos.

This public release contains the test split: 500 question-answer records referencing 488 distinct images. The training split will be released soon.

Load the test split

from datasets import load_dataset

test = load_dataset("reiofa/Worldwide-LiveVQA", split="test")
assert len(test) == 500

sample = test[0]
image = sample["image"]  # decoded PIL image; no separate image download is required
question = sample["question"]
options = sample["options"]
answer = sample["Ground_Truth"]

The default configuration explicitly selects the question-level Parquet files. Each row contains an embedded image and its question, options, answer, and metadata. Shared source images can occur in multiple rows because those rows are different questions.

Composition

Source group (benchmark)	Test questions
news	200
papers	100
videos	200
Total	500

The videos group contains static video frames as image inputs, not video files.

The news and videos groups each cover 10 question languages: English, Arabic, Chinese, French, German, Indonesian, Japanese, Portuguese, Russian, and Spanish. Each group contains 20 questions per language, for a combined total of 40 news and video questions per language. All 100 questions in the papers group are in English.

Fields

Field	Type	Description
image	Image	Embedded input image, decoded by Hugging Face Datasets.
id	string	Unique question identifier.
orig_id	string	Source record identifier.
benchmark	string	Source group: news, papers, or videos.
question	string	Question text.
options	list of strings	Answer choices, including letter prefixes such as A..
Ground_Truth	string	Correct choice letter corresponding to an entry in options.
Ground_Truth_List	list of strings	Textual answer variants.
language	string	Question language.
level	int64	Question difficulty level (1 or 2).
question_type	string	Question type.
topic	string	Source title or topic.
text	string	Associated source text; auxiliary metadata.
caption	string	Associated image description; auxiliary metadata.
source	string	Source name.
url	string	Source URL.
publish_time	string	Publication date.
category	list of strings	Source categories.
country_or_region	string or null	Country or region.
img_url	string or null	Source image URL.
img_path	string	Relative image path.

Use image to access the image input directly with Hugging Face Datasets.

Evaluation

The test records contain both model inputs and reference annotations. Use image, question, and options as the primary multiple-choice VQA inputs. Keep Ground_Truth and Ground_Truth_List out of model prompts. When reporting results, specify whether auxiliary metadata or external retrieval was used.

The official evaluation code will be released soon.

Citation

If you use Worldwide-LiveVQA in your research, please cite our ACL Findings paper:

@inproceedings{huang-etal-2026-worldwide,
    title = "Worldwide {L}ive{VQA}: Real-Time Visual Knowledge Seeking and Updating Across Languages",
    author = "Huang, Xuanao  and
      Liu, Xingjia  and
      Zhou, Zetong  and
      Peng, Yuyang  and
      Wan, Yao  and
      Chen, Dongping",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1984/",
    doi = "10.18653/v1/2026.findings-acl.1984",
    pages = "39819--39894"
}