Header background image

MarineEval: Assessing the Marine Intelligence of Vision-Language Models

Yuk-Kwan Wong1, Tuan-An To2, Jipeng Zhang1, Ziqiang Zheng1*, Sai-Kit Yeung1
1Hong Kong University of Science and Technology
The IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026
Oral Presentation
* Corresponding author: [email protected]
Paper Supplemenatary arXiv logo arXiv Code Hugging Face logo Dataset

Overview

MarineEval Overview
Figure 1. MarineEval overview.

The MarineEval dataset is designed to evaluate the capabilities of general vision-language models (VLMs) in addressing domain-specific marine tasks through visual question answering. It comprises 2,000 questions spanning 7 dimensions and 20 sub-dimensions, encompassing a broad spectrum of marine knowledge and skills, including species identification, conservation, and other related areas.

Example Image
Figure 2. Dimension overview.

Dataset Overview

Explore Full Dataset

Statistic

#. of Questions #. of Dimensions #. of Sub-Dimensions
2,000 7 20
Figure 3. Distribution of question types in the MarineEval dataset.

Benchmark

Figure 4. Average accuracy across 7 task dimensions for different model categories.
Model # Params. B&TE C&TA DI HR MTU SR SC Avg. Total
Open-source VLMs
DeepSeek-VL-chat 1.3B 27.86 39.33 11.00 59.00 34.31 22.25 18.33 30.30 24.96
OpenFlamingo 2B 20.90 40.33 5.33 60.00 21.57 8.25 9.83 23.74 17.62
Mini-Monkey 2B 44.28 50.33 33.00 58.00 74.51 12.75 27.67 42.93 34.45
InternVL-2.5 4B 65.17 56.67 54.00 64.00 80.39 16.75 29.33 52.33 42.54
LLaVA-1.6 Vicuna 7B 68.66 52.00 38.67 53.00 71.57 34.00 37.33 50.75 44.73
InternLM-XComposer2.5 7B 64.18 60.33 49.33 52.00 75.49 14.00 30.17 49.36 41.14
LLaVA-Next 8B 44.78 69.67 25.67 32.00 54.90 32.00 26.67 40.81 37.54
InternVL-2 8B 55.22 55.00 46.00 65.00 78.43 16.50 34.17 50.05 41.44
InternVL-2.5 (26B) 26B 35.32 41.67 47.00 66.00 74.51 25.00 32.33 45.98 38.59
LLaVA-Next-Qwen 32B 67.16 60.00 38.33 65.00 72.55 16.50 43.67 51.89 44.78
LLaVA-1.6 Hermes-Yi 34B 68.66 52.00 38.67 53.00 71.57 34.00 37.33 50.75 44.73
InternVL-3 38B 74.13 48.33 60.33 68.00 78.43 22.50 39.83 55.94 47.53
Avg. across models – 53.81 52.77 37.19 59.42 71.19 21.23 30.27 46.14 39.17
Close-source VLMs
Claude-3.7-Sonnet-Vision – 68.16 53.67 52.33 71.00 83.33 24.50 45.17 56.88 48.93
Gemini-2.0-Flash-Vision – 65.17 60.67 59.67 74.00 87.25 29.00 55.33 61.59 55.07
Grok-2-Vision – 77.61 54.67 27.33 74.00 70.59 34.50 54.00 56.10 50.42
GPT-4o-Vision – 69.15 44.67 51.67 72.00 62.75 26.50 40.50 52.46 45.58
Qwen-VL-Plus – 52.24 41.00 42.00 71.00 85.29 25.00 39.50 50.86 42.39
Avg. across models – 66.07 50.34 46.20 72.40 77.64 27.90 46.10 55.18 48.08
Human Performance
General Background – 68.65 54.33 60.17 82.00 76.96 51.50 31.42 60.72 51.75
Marine Background – 75.00 70.33 69.67 83.00 72.00 64.00 57.50 70.31 66.35
Table 1. The average accuracy across 7 task dimensions. Abbreviations: Behavior & Trait Extraction (B&TE), Conservation & Threat Analysis (C&TA), Document Interpretation (DI), Hallucination Resistance (HR), Marine Technology Understanding (MTU), Spatial Reasoning (SR), Species Comprehension (SC). “–” indicates the number cannot be computed.

Citation

                        
@misc{wong2025marineevalassessingmarineintelligence,
      title={MarineEval: Assessing the Marine Intelligence of Vision-Language Models}, 
      author={YuK-Kwan Wong and Tuan-An To and Jipeng Zhang and Ziqiang Zheng and Sai-Kit Yeung},
      year={2025},
      eprint={2512.21126},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.21126}, 
}