# WorldVQA：多模态大模型视觉世界知识基准测试

- 来源：Moonshot AI：Kimi Blog
- 发布时间：2026-02-03 00:00
- AIHOT 链接：https://aihot.virxact.com/items/cmnwsdptd002tslagois95im8
- 原文链接：https://www.kimi.com/blog/worldvqa

## AI 摘要

Kimi团队发布WorldVQA基准测试，评估多模态大语言模型视觉世界知识的事实准确性。数据集包含3,500个经多阶段人工验证的图像-问题对，涵盖自然、地理、文化等9个类别，区分头部与尾部知识分布。测试显示，即使是Kimi K2.5、Gemini-3-pro等前沿模型，整体准确率仅46%-47%，长尾视觉知识上常低于50%，揭示当前模型在事实可靠性方面的显著不足。

## 正文

Introducing WorldVQA ​

A benchmark for evaluating atomic visual world knowledge in Multimodal LLMs.

Authors Kimi Team

Overview ​

We are releasing WorldVQA, a new benchmark designed to measure the factual correctness of Multimodal Large Language Models (MLLMs). While recent models have demonstrated impressive capabilities in visual reasoning and description, measuring their reliability regarding visual world knowledge remains a challenge.

WorldVQA focuses on a critical question: Does the model actually recognize the specific entity it sees, or is it merely hallucinating based on visual patterns?

Our results show that WorldVQA creates a significant challenge for frontier models. Even state-of-the-art models struggle to achieve high accuracy on long-tail visual knowledge, often falling below 50% accuracy. This benchmark aims to drive progress toward more factually reliable and knowledgeable multimodal AI.

The Dataset ​

The dataset consists of 3,500 high-quality image-question pairs. The distribution aims to test a model's encyclopedic breadth across the world. The dataset distinguishes itself through three core design principles:

Factuality & Unambiguity: Every question has a single, verifiable ground-truth answer. We exclude subjective questions or ambiguous visual scenarios.

Rich Taxonomy: The dataset spans 9 categories to ensure broad coverage of world knowledge.

Head vs. Tail Distribution: We explicitly separate data into Head (common knowledge) and Tail (rare/long-tail knowledge). This allows us to measure how model performance degrades as knowledge becomes more obscure.

Note on Quality: To ensure the benchmark is a reliable gold standard, all images and question-answer pairs underwent rigorous multi-stage human verification to filter out noise and ambiguity.

Note on Quality: To ensure the benchmark is a reliable gold standard, all images and question-answer pairs underwent rigorous multi-stage human verification to filter out noise and ambiguity.

What bird is in the picture?

Answer:Chestnut Shortwing

What's the name of the flower in the picture?

Answer:Freesia

图中出现的内容/文物是/属于哪个遗址？

Answer:善化寺

What is the name of the natural landmark shown in the image?

Answer:Cape of Good Hope

What is the title of the dance performance shown in the picture?

Answer:Swan Lake

这个图片是什么珍品

Answer:战国水晶杯

What style of bag is shown in the picture?

Answer:Shell bag

What electronic consumer product is shown in the image? Provide the exact name and model number.

Answer:iPhone 17 Pro

图中的飞行器是什么型号？

Answer:中国歼 - 20战斗机

What specific attachment or accessory is this for the vehicle?

Answer:Roll cage

What is the name of the character in the picture?

Answer:Bayle the Dread

Which film or TV series is this image from?

Answer:Your Name

What is the medium (carrier) of the advertisement in this image?

Answer:Direct-mail advertisement

What is the name of the trademark or logo shown in the image?

Answer:EgyptAir

What track-and-field or gymnastics event is shown in the picture? Please be as specific as possible.

Answer:Floor exercise

图片中的建筑是哪座体育场馆？

Answer:上海体育场

Distribution of Tasks per Category ​

StatisticsNumberDataTotal3500Chinese (CN)1260 (36%)English (EN)2240 (64%)Category CategoriesTotal categories9Nature & Environment (Nature)9.31%Locations & Architecture (Geography)14.63%Culture, Arts & Crafts (Culture)14.46%Objects & Products (Objects)12.49%Vehicles, Craft & Transportation (Transportation)8.74%Entertainment, Media & Gaming (Entertainment)14.60%Brands, Logos & Graphic Design (Brands)7.43%Sports, Gear & Venues (Sports)4.06%Notable People & Public Figures (People)14.29%DifficultyEasy31.16%Medium40.77%Hard28.07%

Using WorldVQA to compare models ​

Benchmark Kimi K2.5 Gemini-3-pro Gemini-2.5-pro Seed-1.5-vision-pro Claude-opus-4.5 Claude-sonnet-4.5 GPT-5.2 GPT-5.1 GPT-4o Grok-4.1-fast-reasoning Grok-4-fast-reasoning Kimi-VL-16B-A3B Qwen3-VL-235B-A22B-Instruct Qwen3-VL-32B-Instruct GLM-4.6V GLM-4.6V-Flash Overall resultsAccuracy46.3 47.4 36.9 34.9 36.8 20.0 28.0 24.5 22.2 21.1 18.9 12.0 23.5 17.7 19.0 14.8 Not Attempted2.1 0.6 0.1 1.6 3.4 8.0 5.4 16.3 9.1 0.1 0.2 3.3 0.0 0.0 0.0 0.1 Correct Given Attempted47.3 47.7 36.9 35.5 38.1 21.8 29.5 29.3 24.4 21.1 19.0 12.4 23.5 17.7 19.0 14.8 F-score46.8 47.5 36.9 35.2 37.5 20.9 28.7 26.7 23.3 21.1 18.9 12.2 23.5 17.7 19.0 14.8 F-score on 9 task categoriesNature40.6 45.1 37.1 41.4 32.5 19.4 24.3 27.3 25.6 18.4 17.8 11.2 26.1 18.1 24.5 16.0 Geography46.8 44.7 33.8 36.1 36.5 21.0 29.1 25.1 20.6 23.6 19.0 13.9 24.8 18.0 21.5 16.3 Culture43.0 47.2 32.6 33.4 34.1 17.4 26.7 22.5 17.8 20.2 18.6 10.1 22.9 16.8 17.8 13.2 Objects44.7 48.1 39.6 32.8 39.6 22.9 26.6 26.6 19.1 25.2 22.0 10.8 26.1 19.0 19.2 14.9 Transportation47.4 45.1 39.9 35.0 43.5 24.8 30.7 31.6 26.2 23.5 20.3 13.5 28.8 19.0 18.6 19.0 Entertainment48.1 47.6 34.2 33.6 29.0 11.6 24.8 18.5 19.1 11.4 8.3 7.9 15.5 12.1 12.5 7.8 Brands52.6 52.4 38.8 32.3 47.6 32.2 39.1 36.0 35.2 25.8 26.6 20.8 22.3 23.8 20.4 18.8 Sports64.8 59.4 54.2 43.7 54.9 31.0 40.8 45.4 44.5 30.3 34.5 17.7 26.1 20.4 23.2 20.4 People50.9 — — — — — — — — — — 7.4 26.2 13.1 10.7 8.2

Measuring Calibration: Confidence vs. Accuracy ​

In our experiments comparing model confidence with actual accuracy, we utilized two key metrics to measure the alignment between a model's subjective belief and its objective performance:

ECE (Expected Calibration Error): Measures the average gap between the model's subjective confidence and its objective accuracy. The ideal value is 0.

Slope (Weighted Average Slope): Measures the correlation and sensitivity between the model's accuracy and its own confidence. The ideal value is 1.0.

Calibration and Confidence Distribution Analysis. Left: Reliability diagrams plotting Actual Accuracy against Stated Confidence. To ensure statistical significance, only bins containing more than 20 samples are visualized. The size of each data point is proportional to the number of samples in that bin. The black dashed diagonal (y=x) represents perfect calibration, while colored dashed lines indicate the weighted average slope for each model. Right: The distribution of stated confidence scores across the full dataset (without sample thresholding). The plots reveal a severe overconfidence trend, with most models concentrating their predictions in the 90-100% confidence range.

Our experiments reveal that all evaluated models are currently far from the ideal state, exhibiting a universal tendency toward overconfidence.

While Kimi-K2.5 achieves best performance on both metrics—recording an ECE of 37.9% and a Slope of 0.550—there remains a significant gap to bridge in the pursuit of "honesty" and "alignment." Enhancing the self-awareness boundaries of multimodal models represents a critical direction for future exploration.

Conclusion ​

WorldVQA is a simple but challenging benchmark for evaluating the atomic visual knowledge of frontier models. Improving performance on WorldVQA is a necessary step for the next generation of AI agents. We are open-sourcing the WorldVQA dataset and evaluation scripts to help the community address the visual knowledge gap.

Read the Paper: https://arxiv.org/abs/2602.02537

View the Code: https://github.com/MoonshotAI/WorldVQA

Download the Data: https://huggingface.co/datasets/moonshotai/WorldVQA