ARK-Bench

A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge

Retrieval beyond surface-level similarity.

1,750 queries · 36,858 candidates · 17 topics · 6 reasoning skills

Quick start · Data format · Evaluation · Citation

ARK-Bench evaluates multimodal retrieval that requires domain knowledge and reasoning, beyond surface-level similarity. It covers five knowledge domains, 17 topics, six reasoning skills, and 16 visual data types. Queries and candidates can be unimodal or multimodal, and most queries include targeted hard negatives.

Explore model results by knowledge domain, reasoning skill, and cognitive demand on the ARK-Bench Leaderboard. The headline counts above are those reported in the paper.

See representative query–target examples

Quick start

Install the loading library and image dependencies:

pip install "datasets[vision]"

Load the query set and retrieval corpus explicitly by configuration name:

from datasets import load_dataset

DATASET_ID = "XLearning-SCU/ARK-Bench"

queries = load_dataset(DATASET_ID, "queries", split="test")
gallery = load_dataset(DATASET_ID, "gallery", split="test")

print(f"Queries: {len(queries):,} | Gallery: {len(gallery):,}")
print(queries.features)
print(gallery.features)

Both configurations use the test split. A topic such as Calendar is a field within that split, not a separate split or configuration. If no configuration is specified, the default is gallery.

Data format

data/
├── queries/test/
│   ├── metadata.jsonl
│   └── images/<topic>/...
└── gallery/test/
    ├── metadata.jsonl
    └── images/<topic>/...

The tables below describe the stored metadata.

Query field	Meaning
topic, query_id	Topic and query identifier.
text	Query text.
query_image_file_name	Query image reference information.
query_image_alt_texts	Alternative text associated with query images.
reasoning_demand	Overall cognitive demand: perception-only, knowledge-intensive, reasoning-intensive, or both.
reasoning_skill	Reasoning-skill annotation; an instance may involve multiple skills.

Gallery field	Meaning
topic, corpus_id	Together identify a retrieval candidate.
file_name	Candidate image path.
alt_text	Alternative text associated with the image.

Evaluation

If you would like to evaluate your model, submit your results, and compare its performance with other leading models, visit the ARK-Bench Leaderboard for detailed information. We welcome your participation and contributions to ARK-Bench.

License

This repository declares Apache-2.0. For third-party materials, refer to the original source terms and the data sources documented in the paper.

Citation

If you use ARK-Bench in your research, please cite:

@misc{lin2026ark,
  title         = {{ARK}: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge},
  author        = {Yijie Lin and Guofeng Ding and Haochen Zhou and Haobin Li and Mouxing Yang and Xi Peng},
  year          = {2026},
  eprint        = {2602.09839},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2602.09839}
}

For dataset questions or issues, open a discussion in the dataset community.