VMMU: A Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark

by Vy Tuong Dang^1*, An Vo^3*‡, Emilio Villa-Cueva²,
Quang Tau¹, Duc Dm¹, Thamar Solorio², Daeyoung Kim¹

^*Equal contribution    ^‡Work done while at MBZUAI and KAIST
¹KAIST, ²MBZUAI, ³University of Michigan

Findings of the Association for Computational Linguistics: AACL-IJCNLP 2026

---

Abstract

Vision language models are predominantly evaluated on English-centric multimodal benchmarks, leaving their behavior in other languages and cultures insufficiently understood. We introduce VMMU, a Vietnamese Multitask Multimodal Understanding and Reasoning benchmark of 2,548 multiple-choice questions across seven domains: Mathematics, Physics, Chemistry, Biology, Geography, Driving Test, and IQ Test. Every question requires jointly reasoning over Vietnamese text and non-text visual evidence (charts, diagrams, tables, traffic scenes), and is distributed as a single rendered page image. We evaluate 10 open-source and 5 closed-source VLMs on VMMU. Test-time compute improves performance from 57% to 78% accuracy, but it remains inferior to an expert baseline (99% accuracy). Extensive error analyses show that closed-source VLMs already achieve strong Vietnamese OCR performance, yet still struggle on VMMU. This suggests that the primary bottleneck is multimodal grounding and reasoning rather than text recognition or Vietnamese language understanding.

---

Load

from datasets import load_dataset

vmmu = load_dataset("anvo25/vmmu", split="full_vqa")  # 2,548 questions

Splits

Split	Rows	Content
full_vqa	2,548	the benchmark: one image per question, with text, options, and visual evidence together
random_subset_vqa	210	30 questions per domain from full_vqa
random_subset_ocr	210	the same 210 images with OCR prompts
cropped_random_subset_vqa	210	separated setting: question and options as text in vn_prompt, image cropped to the visual evidence
cropped_random_subset_vqa_description	210	the separated setting plus a text description of the image

Questions per domain (subject): Math 456, Physics 361, Chemistry 302, Biology 341, Geography 481, DrivingTest 367, IQTest 240.

Models get the image and vn_prompt (or en_prompt), which asks for the answer in curly brackets, e.g. {A}. The answer is extracted from the brackets and compared with ground_truth.

Fields

• image: the question image
• ID: unique file name, e.g. Math_0448_test_258_question_28_page_4_key_D_vqa.png
• image_path: relative path used by the evaluation code
• vn_prompt, en_prompt: the instruction in Vietnamese and English. In the cropped_* splits, vn_prompt also holds the question text and options, and en_prompt is empty.
• ground_truth: A to F, 1 to 4 for Driving Test, or several letters (e.g. AB) when the image holds more than one question
• subject: the domain
• multiple_question: True for the 25 images that hold more than one question

How it was built

Questions come from Vietnam's national High School Graduation Examination (Math, Physics, Chemistry, Biology, Geography), the official driving license exam, and standardized IQ tests. Exam PDFs are rendered to page images, split into questions with Tesseract-based boundary detection, and text-only screenshots are filtered out. Three native Vietnamese speakers verified every item. Code: GitHub.

Citation

@inproceedings{dang2026vmmu,
  title     = {{VMMU}: A Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark},
  author    = {Dang, Vy Tuong and Vo, An and Villa-Cueva, Emilio and Tau, Quang and Dm, Duc and Solorio, Thamar and Kim, Daeyoung},
  booktitle = {Findings of the Association for Computational Linguistics: AACL-IJCNLP 2026},
  year      = {2026},
  url       = {https://arxiv.org/abs/2508.13680}
}