VisualWorldBench

A Fine-Grained Multi-Task Benchmark for Evaluating
Visual World Knowledge in MLLMs

Chen Chen^1,2  Zheming Liang^1,2  Jiaqi Wang^2,3,*

¹ University of Science and Technology of China
² Shanghai Innovation Institute  ³ JD.com
_{* Corresponding author}

Overview ·  Download ·  Data Format ·  Citation

Overview

VisualWorldBench evaluates fine-grained visual world knowledge in multimodal large language models. It contains 32,541 evaluation instances across four tasks and 15,479 ground-truth categories, spanning eight visual domains.

Task	Annotation file	Instances	Metric
MCQ	mcq.jsonl	14,717	Accuracy
Grid-Sel (2×2)	sel_2x2.jsonl	4,085	Accuracy
Grid-Sel (3×3)	sel_3x3.jsonl	4,125	Accuracy
Grid-Loc (2×2)	loc_2x2.jsonl	3,959	IoU@0.5
Grid-Loc (3×3)	loc_3x3.jsonl	3,446	IoU@0.5
Open-QA	open_qa.jsonl	2,209	EM + LLM Judge

The overall score is the mean of MCQ, Grid-Sel (averaged over both grid sizes), Grid-Loc (averaged over both grid sizes), and corrected Open-QA accuracy. VisualWorldBench-50K is a separate training resource.

Download

hf download orange0318/VisualWorldBench --repo-type dataset --local-dir ./VisualWorldBench
cd VisualWorldBench

Grid images are included. The 16,095 unique single images used by MCQ and Open-QA are provided through source links. Download direct image URLs to the local paths specified by the annotations. For images distributed through upstream datasets, obtain and extract those datasets separately, then locate the images using the recorded source paths.

Data Format

Annotations are stored as JSONL, with one evaluation instance per line. Each record has a qid, an image path, and task-specific fields:

Task	Main fields
MCQ	options, gt_index
Grid-Sel	target_class_name, target_cell, answer
Grid-Loc	target_class_name, bbox_norm, answer
Open-QA	question, gt_aliases_base, gt_aliases_expanded, gt_ancestors

The evaluation code provides prompt templates for Grid-Sel and Grid-Loc.

bbox_norm uses [x1, y1, x2, y2] coordinates normalized to [0, 1000]. Record metadata includes the visual domain and taxonomy depth. The meta.domain label follows the ground-truth category in the taxonomy: animal, plant, fungus, artifact, transport, food, scene, or other.

Single-image records also contain image_url and image_url_type. The image field specifies where to save the retrieved image locally. Records requiring an upstream dataset additionally include its identifier, relative path and acquisition filename.

Citation

@misc{chen2026visualworldbench,
  title = {VisualWorldBench: A Fine-Grained Multi-Task Benchmark for Evaluating Visual World Knowledge in {MLLMs}},
  author = {Chen, Chen and Liang, Zheming and Wang, Jiaqi},
  year = {2026}
}

License

Original benchmark annotations are licensed under CC BY-NC 4.0. Source images retain their respective copyrights, licenses, and access conditions.

If you encounter any issues downloading the images, please contact the authors.