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

This repository contains the complete evaluation dataset for MCQ, Grid-Sel, Grid-Loc, and Open-QA. VisualWorldBench-50K is a separate training resource.

This dataset contains 32,541 evaluation instances across four tasks, covering 15,479 ground-truth categories. It is constructed using a taxonomy of 21,291 leaf categories spanning eight visual domains: animal, plant, fungus, artifact, transport, food, scene, and other.

Task	File	Instances	Metric
MCQ	mcq.jsonl	14,717	Accuracy
Grid-Sel (2x2)	sel_2x2.jsonl	4,085	Accuracy
Grid-Sel (3x3)	sel_3x3.jsonl	4,125	Accuracy
Grid-Loc (2x2)	loc_2x2.jsonl	3,959	[email protected]
Grid-Loc (3x3)	loc_3x3.jsonl	3,446	[email protected]
Open-QA	open_qa.jsonl	2,209	EM + LLM Judge

Overall Score = mean(MCQ, Sel_Avg, Loc_Avg, Open_QA_corrected)

• Sel_Avg = (Grid-Sel_2x2 + Grid-Sel_3x3) / 2
• Loc_Avg = (Grid-Loc_2x2 + Grid-Loc_3x3) / 2

Directory Structure

.
├── mcq.jsonl
├── sel_2x2.jsonl
├── sel_3x3.jsonl
├── loc_2x2.jsonl
├── loc_3x3.jsonl
├── open_qa.jsonl
└── images/
    ├── single/       (source images for MCQ & Open-QA)
    ├── single_ext/   (additional source images for MCQ & Open-QA)
    ├── grid_2x2/     (4,085 images for 2x2 grid tasks)
    └── grid_3x3/     (4,125 images for 3x3 grid tasks)

Data Format

MCQ (mcq.jsonl)

Multiple-choice question with 20 options.

{
  "qid": "mcq_00000003",
  "image": "images/single/000001.jpg",
  "options": ["bellis perennis", "lantana camara", "...", "little golden zinnia", "..."],
  "gt_index": 9,
  "meta": {"domain": "plant", "gt_depth": 4}
}

• qid: unique identifier for each MCQ record
• options: list of 20 candidate names
• gt_index: index of the correct answer in options

Grid-Sel (sel_2x2.jsonl, sel_3x3.jsonl)

Grid cell selection: identify which cell contains the target object.

{
  "qid": "sel_2x2_001871",
  "image": "images/grid_2x2/000001.jpg",
  "prompt": "You are given a 2x2 grid image. Cells are numbered left-to-right, top-to-bottom as 1, 2, 3, 4. Which cell contains the target category: orange slice? Answer with a single number.",
  "grid_size": "2x2",
  "target_class_name": "orange slice",
  "target_cell": 3,
  "choices": ["1", "2", "3", "4"],
  "answer": "3",
  "meta": {"domain": "food", "image_size": [768, 768], "rows": 2, "cols": 2, "gt_depth": 3}
}

Grid-Loc (loc_2x2.jsonl, loc_3x3.jsonl)

Bounding box prediction: locate the target object with a bounding box.

{
  "qid": "loc_2x2_001871",
  "image": "images/grid_2x2/000001.jpg",
  "prompt": "Locate the target object: orange slice.\n\nOutput the bounding box as [x1, y1, x2, y2]...",
  "grid_size": "2x2",
  "target_class_name": "orange slice",
  "bbox_norm": [0.0, 562.5, 468.5, 937.5],
  "answer": [0.0, 562.5, 468.5, 937.5],
  "meta": {"domain": "food", "image_size": [768, 768], "rows": 2, "cols": 2, "gt_depth": 3}
}

• bbox_norm: ground-truth bounding box [x1, y1, x2, y2] normalized to [0, 1000]
• (0, 0) = top-left, (1000, 1000) = bottom-right

Open-QA (open_qa.jsonl)

Free-form visual identification question.

{
  "qid": "open_qa_00000004",
  "image": "images/single/000001.jpg",
  "question": "What kind of flower is shown in the picture?",
  "gt_aliases_base": ["little golden zinnia", "zinnia grandiflora"],
  "gt_aliases_expanded": ["little golden zinnia", "zinnia grandiflora"],
  "gt_ancestors": [
    {"id": "n12034141", "depth_diff": 1, "names": ["zinnia", "common zinnia"], "score": 0.8},
    {"id": "n11669921", "depth_diff": 2, "names": ["flower"], "score": 0.5}
  ],
  "meta": {"domain": "plant", "gt_depth": 4}
}

• gt_aliases_base / gt_aliases_expanded: acceptable answer strings
• gt_ancestors: hierarchical ancestor nodes for partial-credit scoring

Metadata Fields

• domain: visual category domain recorded for the sample
• gt_depth: depth of the ground-truth class in the taxonomy tree (higher = more fine-grained)
• image_size: [width, height] in pixels (grid tasks only)

Intended use and limitations

VisualWorldBench supports research on fine-grained visual recognition and multimodal model evaluation. The eight domains are not sampled uniformly; results for sparse domains should be interpreted with their sample counts. Grid tasks use controlled composite images, and performance on these tasks does not directly establish performance in natural multi-object scenes. Open-QA combines exact/alias matching with an LLM judge, so its scores depend in part on the judging protocol.

Download

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

The six task JSONL files define the evaluation instances. Shared source images can appear in more than one task, so the image viewer row count is not the evaluation instance count.

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

This dataset is released under CC BY-NC 4.0 for research purposes.