MMGist: A Comprehensive Multimodal Benchmark for 2027

MMGist is a curated multimodal evaluation benchmark built from 18 widely used vision-language benchmarks. It contains 7,262 samples spanning seven capability dimensions and is designed to make LVLM evaluation more efficient, visually grounded, discriminative, and reliable.

This Hugging Face release provides a single viewer-friendly Parquet file with embedded image bytes, so the dataset can be inspected and loaded directly without a separate image archive.

Key Features

• Curated from 18 source benchmarks: MMGist is built from a raw pool of 23,250 items across common multimodal benchmarks.
• Seven capability dimensions: STEM Reasoning, Medical VQA, Visual Perception, Spatial Understanding, Diagram & OCR, Visual Logic, and Expert Knowledge.
• Item-level quality control: The benchmark filters weakly vision-dependent, saturated, and anomalous items through text ablation, cross-model saturation filtering, and anomaly review.
• Efficient evaluation: MMGist retains 7,262 items, reducing the evaluation set by 69% while preserving model-ranking consistency with the raw pool.

Capability Distribution

Capability	Samples
STEM Reasoning	2,058
Medical VQA	1,389
Visual Perception	1,118
Spatial Understanding	898
Diagram & OCR	852
Visual Logic	539
Expert Knowledge	408

Source Benchmark Distribution

Source benchmark	Samples
MathVision	1,010
MedXpertQA	997
EmbSpatialBench	778
BLINK	687
MME-SCI	678
AI2D	520
SLAKE	392
MMMU-Pro	379
OCRBench	332
ZeroBench-Sub	285
MathVista	234
RealWorldQA	217
HallusionBench	214
LogicVista	158
DynaMath	136
CountBench	120
ZeroBench	96
MMMU	29

Data Curation

MMGist is constructed with a three-stage item-level quality-control pipeline:

1. Text-only ablation filtering removes items that can be answered from text alone and therefore do not reliably test visual understanding.
2. Cross-model saturation filtering removes items that are already solved consistently by current LVLMs and contribute little to model discrimination.
3. Anomaly detection and review identifies pseudo-hard items caused by ground-truth errors, ambiguous questions, or scoring risks, followed by multi-model adjudication and human expert review.

The final benchmark keeps 31.2% of the original item pool. Experiments in the accompanying paper show that the curated set preserves model rankings with Spearman rho = 0.98, reduces evaluation items by 69%, and improves cross-model discrimination by 78%.

Evaluation Results

The table below reports representative MMGist results for 27 LVLMs across seven capability dimensions. Macro Avg is the mean over capability groups, while Sample Avg is the mean over all 7,262 samples.

Data Format

Each row corresponds to one evaluation sample.

Column	Type	Description
uid	string	Globally unique ID in the form<benchmark>/<id>.
id	string	Original normalized sample ID within the source benchmark.
benchmark	string	Source benchmark identifier.
benchmark_label	string	Human-readable source benchmark name.
capability	string	Capability group assigned by MMGist.
question	string	Question or prompt text, including answer choices when available.
answer	string	Gold answer in normalized text form.
image_1 ... image_6	image struct	Embedded image slots. Unused image slots are null.

Most samples use image_1; multi-image samples fill later image columns. The Parquet file stores image bytes directly, so no external image directory is required.

Files

• data/train-00000-of-00001.parquet: all 7,262 samples with embedded image bytes.
• image.png: evaluation results table from the accompanying paper.
• SHA256SUMS: checksum for the Parquet file.

Current Parquet SHA-256:

42ab72970a76a82f90d1a403be5c67123a1d7e905e472610d9b575aa788f112c  data/train-00000-of-00001.parquet

Load

from datasets import load_dataset

ds = load_dataset("Winston-Yuan/MMGist", split="train")
print(ds)
print(ds[0]["question"])
print(ds[0]["image_1"])

To load from a local clone:

from datasets import load_dataset

ds = load_dataset(
    "parquet",
    data_files="data/train-00000-of-00001.parquet",
    split="train",
)

Citation

If you use MMGist in your research, please cite:

@article{yuan2026mmgist,
  title={MMGist: A Comprehensive Multimodal Benchmark for 2027},
  author={Yuan, Wenzhen and Ruan, Jiacheng and Xiong, Wutao and Zhao, Chengping and Liu, Ting and Fu, Yuzhuo},
  journal={arXiv preprint arXiv:2606.22437},
  year={2026}
}

Resources

• Paper: arXiv
• Dataset: Hugging Face

Contact

For questions, please contact Wenzhen Yuan at winston_yuan@sjtu.edu.cn.