VisReason

VisReason is a benchmark for evaluating vision-centric reasoning in everyday scenarios where perception and inference are tightly coupled. It is designed to test whether multimodal large language models can reason directly from visual evidence rather than relying mainly on language-mediated abstractions.

The dataset contains 1,505 carefully curated questions across 10 reasoning categories, covering perceptual, structural, and conceptual reasoning tasks.

📥 Download

You can download the dataset with the Hugging Face CLI:

hf download CASIA-IVA-Lab/VisReason --repo-type dataset --local-dir ./data

The official evaluation scripts expect the dataset to be available under ./data in the project root.

🗂️ Dataset Structure

data/
  img_<class_number>/
    datajson_label.<ext>
  class_1.jsonl
  class_2.jsonl
  ...
  class_10.jsonl
  datasets.json

datasets.json is the dataset index used by the evaluation scripts. Modify this file to select or adjust the classes to evaluate.

Images are stored under the corresponding img_<class_number> folders. Each image file is named with the source dataset key and sample label:

datajson_label.<ext>

📊 Data Files

Class	Data file	Samples	Image folder
class_1	class_1.jsonl	40	img_1/
class_2	class_2.jsonl	100	img_2/
class_3	class_3.jsonl	46	img_3/
class_4	class_4.jsonl	130	img_4/
class_5	class_5.jsonl	200	img_5/
class_6	class_6.jsonl	135	img_6/
class_7	class_7.jsonl	200	img_7/
class_8	class_8.jsonl	111	img_8/
class_9	class_9.jsonl	275	img_9/
class_10	class_10.jsonl	268	img_10/

🧾 Data Format

Each class_*.jsonl file contains one JSON object per line. A sample has the following fields:

◆ Sample Fields

Field	Description
class	Class identifier.
label	Sample identifier within the source data.
question type	Question format used for prompting and evaluation.
question	Natural-language question.
answer	Ground-truth answer, including boxes for localization tasks.
images	Image paths associated with the sample.
datajson	Source split.
Height / Weight	Image size fields used by the released files.
url	Original source URL, when available.

◆ Question Type

Type	Format
1	Multiple-choice
2	Short-answer
3	Open-ended
4	Bounding-box localization

🧪 Evaluation

The evaluation code is available in the GitHub repository:

https://github.com/CASIA-IVA-Lab/VisReason

Inference results are expected to be saved as:

results/<model>/class_X_results.json
results/<model>_cot/class_X_results.json

The evaluator writes per-class judging files and a final summary:

results/<model>/class_X_judge.json
results/<model>/summary.json

For class_1 and class_2, bounding-box predictions are evaluated with IoU at threshold 0.5. Other classes are evaluated by an LLM-based judge. The final score is the unweighted mean over all class accuracies.

📄 License

📝 Citation

@inproceedings{guo-etal-2026-mllms,
    title = "Can {MLLM}s Reason Beyond Language? {V}is{R}eason: A Comprehensive Benchmark for Vision-Centric Reasoning",
    author = "Guo, Longteng  and
      Wang, Yifan  and
      Huo, Pengkang  and
      Chen, Tailai  and
      Wu, Yuze  and
      Liu, Jing  and
      Zhu, Xinxin",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    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.1996/",
    doi = "10.18653/v1/2026.findings-acl.1996",
    pages = "40149--40192",
    ISBN = "979-8-89176-395-1",
    abstract = "Recent multimodal large language models (MLLMs) achieve strong performance on visual reasoning benchmarks, yet it remains unclear to what extent such performance reflects reasoning directly grounded in visual evidence. We introduce VisReason, a benchmark for vision-centric reasoning in everyday scenarios where perception and inference are tightly coupled. VisReason contains 1,505 questions across 10 categories spanning perceptual, structural, and conceptual reasoning. Our evaluation shows that VisReason poses a qualitatively different challenge from existing benchmarks, exposing substantial gaps between humans and current MLLMs and revealing limited benefits from test-time reasoning strategies. VisReason offers a focused diagnostic for evaluating vision-centric reasoning beyond language."
}

📬 Contact

If you have any questions, please reach out to:

• Yifan Wang - wangyifan2026@ia.ac.cn