Vision

We track 234 benchmarks in vision, or 244 if you count each version. They run from everyday level to expert level.

38 in use 38 past 85% 147 with public code

OpenAI held the top vision score 36% of the contested time

That's about 16 of the 46 months since January 2023, averaged over the benchmarks below. The easiest benchmarks are at the top. Versions of one test share a row, with a dark tick at each new version. We stop the clock, and fade the rest of the row, when the best score passes 85%, or at the newest score if 45 days go by without another.

Include lab claims

Other labs Hugging Face University of Wisconsin-Madison OpenAI Shanghai AI Lab Qwen ByteDance Seed StepFun Google DeepMind Moonshot AI Z.ai Anthropic Baidu Cohere xAI Meta

Everyday Most adults, without training

MMBench-EN v1.1 Passed 85% in Jul 2024, by InternVL2-40B

Qwen3.5 397B-A17B 93.7%

CCBench (MMBench Chinese Culture) Passed 85% in Dec 2024, by InternVL2.5-78B-MPO

InternVL2.5-78B-MPO 86.9%

A-Bench Not tested since Jun 2025

Qwen2.5-VL-72B 81%

OCRBench v2 10 leaders

Qwen 3.8 Max 74.2%

School Students, from grade school to high school

AI2D Passed 85% in Mar 2024, by Claude 3 Sonnet

ERNIE 5.0 96.9%

ChartQA Passed 85% in May 2024, by GPT-4o

Command A Vision 90.9%

MathVista Passed 85% in Apr 2025, by o3

Seed2.1 Pro 90.7%

MMStar Passed 85% in Aug 2026, by Qwen 3.8 Max

Qwen 3.8 Max 85.9%

University Undergraduates and entry-level professionals

MMMU Passed 85% in Nov 2025, by GPT-5.1

Qwen 3.6 Plus 86%

CharXiv Passed 85% in Dec 2025, by GPT-5.2

Claude Fable 5 93.5%

MMMU-Pro 10 leaders

Kimi K3.1 Pro 84.1%

ZeroBench 10 leaders

GPT-6 Astra 52%

Competition The strongest students, in olympiad-style contests

MATH-Vision (MathVision) Passed 85% in Nov 2025, by Gemini 3 Pro

Qwen 3.8 Max 97.7%

Scores come from each benchmark’s own leaderboard, from Epoch AI’s runs and from labs’ reports on their own models. We date Epoch runs by the model’s release and lab reports by when they came out. Labs pick their own settings, so a lab’s claim can beat an independent run of the same model. Turn off lab claims to see independent runs only.

Other labs Hugging Face University of Wisconsin-Madison OpenAI Shanghai AI Lab Qwen Google DeepMind NVIDIA ByteDance Seed Moonshot AI Anthropic xAI

Everyday Most adults, without training

MMBench-EN v1.1 Passed 85% in Jul 2024, by InternVL2-40B

Gemini 2.5 Pro 88.5%

CCBench (MMBench Chinese Culture) Passed 85% in Dec 2024, by InternVL2.5-78B-MPO

InternVL2.5-78B-MPO 86.9%

BLINK Not tested since Apr 2024

GPT-4V 51.3%

A-Bench Not tested since Jun 2025

Qwen2.5-VL-72B 81%

OCRBench v2 8 leaders

Qwen3.6 27B 68.3%

BabyVision Passed 85% in Jun 2026, by Claude Fable 5

Qwen 3.8 Max 91.3%

School Students, from grade school to high school

MathVista Not tested since Apr 2025

Doubao-1.5-pro 79.5%

MMStar Not tested since Aug 2025

GPT-5 75.7%

University Undergraduates and entry-level professionals

MMMU Passed 85% in Nov 2025, by GPT-5.1

GPT-5.1 85.4%

CharXiv Not tested since Mar 2026

o3 78.6%

MMMU-Pro Not tested since Apr 2026

GPT-5.4 82.1%

ZeroBench 8 leaders

GPT-6 Astra 52%

Competition The strongest students, in olympiad-style contests

MATH-Vision (MathVision) Passed 85% in Nov 2025, by Gemini 3 Pro

GPT-5.4 96.1%

Only each benchmark’s own leaderboard and Epoch AI’s runs, dated by the model’s release. We leave out benchmarks that only labs report.

From everyday to expert level

Grouped by who could solve the tasks, from things most adults can do to open research problems. The bar under each name is its best score.

Everyday

Most adults, without training

School

Students, from grade school to high school

Competition

The strongest students, in olympiad-style contests

Expert

Specialists with years of training, often over hours of work

By kind

General visual questions

Benchmark	Level	Scored	Best score	Status
MMMU 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.	University	Checked automatically	85.4%	In use
MMStar Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.	School	Checked automatically	85.9%	In use
MMBench-EN v1.1 Models answer English multiple-choice image questions across 20 ability dimensions, using CircularEval with shuffled options. Scored by accuracy on the dev split.	Everyday	Checked automatically	88.5%	In use
MMMU-Pro Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts.	University	Checked automatically	84.1%	In use
MMBench Vision language models answer multiple-choice questions covering 20 perception and reasoning skills. CircularEval rotates the options. Scored by accuracy.	Everyday	Checked automatically	88.8%	Fading
CCBench (MMBench Chinese Culture) Vision language models answer 510 Chinese-language multiple-choice questions about images of Chinese culture, such as calligraphy, painting, food and historical figures.	Everyday	Checked automatically	86.9%	Dropped
Vision Arena People send an image with a prompt, compare answers from two anonymous vision-language models, and vote for the better one. Votes become Bradley-Terry ratings.	Everyday	Ranked by votes	1,309	Not in cards
A-Bench Models answer multiple-choice questions about AI-generated images, covering both meaning and low-level visual quality, to test them as judges of image generators.	Everyday	Checked automatically	81%	Not in cards
ZeroBench 100 hand-made visual questions that all frontier models originally failed. Scored by exact answer, pass@1 and pass@5.	University	Checked automatically	52%	In use
BLINK Models answer 3,807 multiple choice questions built from 14 classic computer vision tasks that use one or more images. Scoring is accuracy.	Everyday	Checked automatically	81.4%	In use
GeoBench Models geolocate GeoGuessr photos from community maps by giving coordinates and a country. Scored by country accuracy and a distance-based score.	Everyday	Checked automatically	89%	Dropped
VISTA Models answer 758 single-turn prompts about one image each, covering perception and reasoning. Three LLM judges grade answers with yes/no rubric questions. The main score is average task accuracy.	Expert	Graded by a model	54.6%	Not in cards
EnigmaEval 1,184 multimodal puzzle-hunt puzzles (v2 dataset has 1,161). Model gives one exact answer per puzzle. Normal and Hard splits.	Competition	Checked automatically	39.3%	Not in cards
VHELM Vision-language models are scored across nine aspects, including perception, reasoning, knowledge, bias, fairness, multilinguality, robustness, toxicity and safety, from many existing datasets.	University	Checked automatically	87.1	Not in cards
Humanity's Sixth Sense Models answer open-ended questions about what an image or short video implies but does not show, such as hidden objects, causes and intentions. An LLM judge checks each answer against a human-written rubric.	Everyday	Graded by a model	55.4%	Not in cards
MME Models answer yes or no questions about images across 14 perception and cognition subtasks with hand-written instructions. Labs usually report the perception sum, out of 2000.	Everyday	Checked automatically	2525.9	Fading
ScienceQA Models answer about 21,000 multiple-choice science questions from school curricula, many with images. Scored by accuracy.	School	Checked automatically	91.3%	Dropped
Claw-Eval-MM Agents complete the 101 multimodal tasks of Claw-Eval, such as video questions, document extraction, and webpage generation. Rubric items graded from run traces set scores.	University	Graded by a model	54.7%	In use
MEGA-Bench Models solve 505 real-world multimodal tasks with over 8,000 samples. Outputs include numbers, code, LaTeX, coordinates and JSON, scored by more than 40 task metrics.	University	Checked automatically	54.3%	System cards only
V* Models answer 191 multiple choice questions about small details in high-resolution images, such as an object's color or relative position.	Everyday	Checked automatically	96.9%	Fading
MuirBench Models answer multiple choice questions over sets of several images, including unanswerable variants. Scored by accuracy.	Everyday	Checked automatically	81.8%	Fading
RISEBench++ Image editing models follow 1,000 bilingual instructions over six reasoning types, adding counterfactual and hybrid reasoning. It covers single-image, multi-image and multi-turn editing across 65 task types.	University	Graded by a model	56.6%	Not in cards
Vibe-Eval Models answer 269 image-based prompts, 100 of them hard, that have expert gold responses. A Reka Core judge rates each answer.	Everyday	Graded by a model	81.4%	Dropped
WorldVQA Model names visual entities in 3,500 image-question pairs across nine categories, isolating memorized visual world knowledge. Exact-answer accuracy.	University	Graded by a model	61.1%	In use
MM-Vet Models answer open questions about images that combine recognition, OCR, knowledge, spatial reasoning, language generation and math. GPT-4 grades each answer against a reference.	Everyday	Graded by a model	81.2%	Dropped
VQAv2 Open-ended questions about COCO images, balanced to reduce language priors. Scored by consensus accuracy against 10 human answers.	Everyday	Checked automatically	88.4%	Fading
RISEBench Image editing models follow 360 instructions that need temporal, causal, spatial or logical reasoning. An LMM judge scores instruction reasoning, appearance consistency and visual plausibility.	University	Graded by a model	51.4%	System cards only
MME-CoT Models reason step by step on image questions in math, science, OCR, logic, space-time and general scenes. A judge scores reasoning quality, robustness and efficiency.	University	Graded by a model	64.2%	Not in cards
MM-MT-Bench Models hold multi-turn conversations about images such as charts, documents and photos. GPT-4o grades each answer against a reference on a 1 to 10 scale.	Everyday	Graded by a model	8.50	Dropped
Scientists' First Exam (SFE) Models answer 830 expert-verified visual questions on scientific data across signal perception, attribute understanding and comparative reasoning.	Expert	Checked automatically	60.7%	Fading
LingoQA Models answer free-form questions about short driving videos. A trained classifier, Lingo-Judge, scores whether each answer is truthful.	Everyday	Graded by a model	84.8%	In use
PaintBench Image editing models do precise single-answer edits in 20 operations such as geometric transforms, color changes and symbolic edits. Procedural generation gives deterministic checks.	Everyday	Checked automatically	16.3%	Not in cards
HLE-VL Models answer the image-based questions from Humanity's Last Exam, a 330 question subset. Answers are graded against known short or multiple choice answers.	Expert	Checked automatically	52.2%	Fading
MicroVQA Models answer expert-written multiple choice questions about biology microscopy images that test image understanding, hypothesis generation and experiment proposal.	Expert	Checked automatically	69.7%	Fading
MME-RealWorld Models answer human-annotated multiple choice questions about high-resolution real-world images in scenarios such as remote sensing, autonomous driving, monitoring and OCR.	Everyday	Checked automatically	65.1%	Dropped
PerceptionBench (Moonshot) Moonshot in-house benchmark of atomic visual perception skills. Scoring details unpublished.	Everyday	Checked automatically	68.3%	In use
ReMI Models reason over several images per question in math, physics, logic, code, charts, and spatial or temporal tasks. Scored by accuracy.	School	Checked automatically	67.3%	Fading
SEED-Bench Models answer 19,000 human-annotated multiple-choice questions over 12 dimensions of image and video understanding. Scored by accuracy against the answer key.	Everyday	Checked automatically	77.7%	Fading
CRPE Models answer single-choice questions about objects and their relations in images, covering subject, predicate and object. Each question is asked again with options rotated (CircularEval).	Everyday	Checked automatically	79.2%	Dropped
GQA Models answer compositional questions about real-world images, generated from scene graphs. Scored by exact-match accuracy, often on the balanced test-dev split.	Everyday	Checked automatically	62%	Fading
HR-Bench Models answer multiple-choice questions about fine details in 4K and 8K resolution images. Results are reported separately for the 4K and 8K sets.	Everyday	Checked automatically	88.5%	Fading
MSEarth Models answer questions about figures from Earth science papers across the atmosphere, cryosphere, hydrosphere, lithosphere and biosphere. Labs report the multiple choice split.	Expert	Checked automatically	65.7%	Fading
OK-VQA Models answer over 14,000 visual questions that need outside knowledge beyond the image. Scoring is VQA accuracy against human answers.	Everyday	Checked automatically	60.2%	Dropped
TIR-Bench Models solve 13 kinds of image tasks that need tool use on images, such as rotation, jigsaw, and maze. Most items are multiple choice.	Everyday	Checked automatically	61.6%	Fading
Mantis-Eval Models answer questions that need reasoning across several interleaved images. Scored by accuracy.	Everyday	Checked automatically	84.8%	Dropped
MIA-Bench Models answer 400 image prompts that stack several strict instructions on format, length and content. An LLM judge scores how well each answer follows every instruction.	Everyday	Graded by a model	92.7%	Dropped
AyaVisionBench Vision-language models answer image questions in 23 languages across 9 task categories, 135 image-question pairs per language. A model judge compares two models' answers and reports win rates.	Everyday	Graded by a model	64%	Dropped
K-Viscuit Models answer multiple-choice visual questions about Korean culture. Questions were written through human and VLM collaboration.	Everyday	Checked automatically	80.1%	Fading
KMMMU Models answer 3,466 college and exam level questions that combine images with Korean text. Questions come from Korean exam sources, some with several images.	University	Checked automatically	42.7%	Fading
m-WildVision Vision-language models answer WildVision bench prompts machine-translated into 23 languages. A model judge compares two models' answers and reports win rates.	Everyday	Graded by a model	72%	Dropped
MeasureBench Vision-language models read values from images of measuring instruments, such as gauges and rulers, in real and synthesized photos. Scored by matching the reading to the key.	Everyday	Checked automatically	62.9%	In use
MMT-Bench Vision-language models answer 31,325 multiple-choice questions across many multimodal tasks such as recognition, localization, reasoning and planning.	Expert	Checked automatically	72.7%	System cards only
R-Bench (real-world robustness) Models answer questions about images degraded by 33 kinds of real-world corruption, from capture to display. Scores compare distorted and reference images.	Everyday	Checked automatically	75.4%	System cards only
SGP-Bench Models read symbolic graphics programs such as SVG and CAD code and answer questions about what the rendered image would show, without seeing the image.	University	Checked automatically	70.7%	Dropped
Agentic Vision Bench (Moonshot) Agents must notice and correctly use key visual facts while they execute tasks.	Expert	Checked automatically	78.3%	System cards only
CoreCognition Models answer true or false and multiple choice questions on 12 core knowledge concepts from developmental cognitive science, such as object permanence and perspective taking.	Everyday	Checked automatically	66.7%	System cards only
Do You See Me Models answer 2,612 visual perception questions on 1,758 images across seven 2D and 3D subtasks drawn from human psychology. Questions are multiple choice or numeric.	Everyday	Checked automatically	67.5%	System cards only
M3GIA Models answer cognition test questions built on five Cattell-Horn-Carroll intelligence factors in English, Chinese, French, Spanish, Portuguese and Korean. Data comes from native contexts.	School	Checked automatically	78.4%	System cards only
MaXM Vision language models answer short visual questions about Crossmodal-3600 images in 7 languages. A test-only set scored by accuracy.	Everyday	Checked automatically	62.3%	System cards only
MIRB Models compare, analyze and reason across several images in four categories: perception, visual world knowledge, reasoning and multi-hop reasoning.	Everyday	Checked automatically	73%	System cards only
MMIU Vision-language models answer 11,000 multiple-choice questions over sets of images. 52 tasks cover 7 kinds of multi-image relationship.	Everyday	Checked automatically	61.3%	System cards only
Seed1.5-VL internal benchmark (ByteDance Seed in-house) Models answer more than 12,000 internal image questions over 100 tasks, from perception, OCR and STEM to agents, code and out-of-distribution puzzles. LLM judges grade answers, and the overall is a weighted average.	University	Graded by a model	59.3%	System cards only
ViVerBench Models judge whether visual outcomes in multimodal reasoning are correct, across 16 task categories such as boxes, points, counts, mazes, GUIs, charts and LaTeX.	School	Checked automatically	75.9%	System cards only
WildVision Vision-language models answer 500 real user image prompts from the WildVision Arena. A GPT-4 judge compares each answer with a fixed baseline model.	Everyday	Graded by a model		Dropped
WorldBench Models answer hand-written reasoning questions about images drawn from a taxonomy of thousands of visual concepts. Questions were kept only if frontier models failed them.	Expert	Checked automatically	67.6%	System cards only
xChatBench Vision language models answer 400 open-ended multimodal chat prompts in 7 languages. An LLM judge scores each answer against fine-grained human-written criteria.	Everyday	Graded by a model	70.1	System cards only
xMMMU Vision language models answer college-level MMMU questions translated into 7 languages. Scored by accuracy.	University	Checked automatically	45.1%	System cards only
A-OKVQA Models answer about 25,000 visual questions that need world knowledge beyond the image, in multiple-choice and direct-answer forms.	Everyday	Checked automatically		Not in cards
CLBench-V Models must learn from task-specific multimodal context, such as figures, tables, maps and web pages, instead of prior knowledge. Tasks run from grounding to knowledge acquisition.	Everyday	Checked automatically		Not in cards
CMMMU Multimodal models answer college-level questions in Chinese across many disciplines, following MMMU's annotation pattern.	University	Checked automatically		Not in cards
DISBench Multimodal agents search raw visual histories over several steps to find target images from implicit context cues.	Everyday	Checked automatically		Not in cards
EmoBench-M Multimodal models are tested on emotional intelligence across 13 scenarios in three dimensions: emotion recognition, conversational emotion understanding and socially complex emotion analysis.	Everyday	Checked automatically		Not in cards
GeoGuesser environment (FineEnvs) Agent placed at unknown Mapillary street-level location looks, walks, pins candidates and guesses coordinates. Distance-based GeoGuessr-style reward.	Everyday	Checked automatically		Not in cards
LLaVA-Bench (In-the-Wild) Multimodal assistants answer open questions about a small set of diverse real images. A GPT-4 judge scores each answer relative to a reference answer.	Everyday	Graded by a model		Not in cards
MIEB Image and image-text embedding models are scored on 130 tasks in 38 languages, grouped into 8 task types such as retrieval, clustering and classification.	Everyday	Checked automatically		Not in cards
MMDU Vision-language models hold long conversations about many images, up to 20 images and 27 turns per dialog. A GPT-4o judge scores the answers.	Everyday	Graded by a model		Not in cards
MMEB Multimodal embedding models are scored on classification, visual question answering, multimodal retrieval and visual grounding, with 16 evaluation datasets in and out of distribution.	Everyday	Checked automatically		Not in cards
MMEB-V2 Multimodal embedding models are scored on 78 tasks over images, videos and visual documents. V2 adds visual document retrieval, video retrieval, temporal grounding, video classification and video QA.	Everyday	Checked automatically		Not in cards
MMEB-V3 Omni-modality embedding models are scored on 190 tasks across text, image, video, audio, visual documents and agent-centric retrieval.	Everyday	Checked automatically		Not in cards
MMMU-Pro Extended Models answer multimodal questions that need both the image and the text. Mercor adds 140 new expert-written tasks to the 1730 public ones and reports both scores.	University	Checked automatically		Not in cards
NaturalBench Models answer 10,000 human-checked questions on natural images. Each question pairs with two images that need different answers, so a model cannot pass without looking.	Everyday	Checked automatically		Not in cards
Q-Bench Multimodal models are tested on low-level vision: perceiving image distortions, describing them and predicting quality scores.	Everyday	Checked automatically		Not in cards
SEED-Bench-2 Multimodal models answer multiple-choice questions that cover comprehension of interleaved image and text inputs and image generation, in hierarchical capability levels.	Everyday	Checked automatically		Not in cards
UltraText Bench Image generators render dense bilingual text in many regions from 432 prompts. Scores check string accuracy, placement and legibility.	Everyday	Checked automatically		Not in cards
VizWiz Models answer over 31,000 spoken questions that blind people asked about photos they took with phones. Some questions are unanswerable from the image.	Everyday	Checked automatically		Not in cards
VL-RewardBench Vision-language reward models pick the better of two responses on general queries, hallucination detection and reasoning tasks.	Expert	Checked automatically		Not in cards
WildVision Arena Users upload an image with a question, compare answers from two anonymous vision-language models, and vote. Votes give Elo ratings.	Everyday	Ranked by votes		Not in cards

Documents, charts and text in images

Benchmark	Level	Scored	Best score	Status
CharXiv Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts.	University	Checked automatically	93.5%	In use
OCRBench v2 Visual-text localization and reasoning tasks across 31 scenarios (documents, scenes, charts). Averaged task accuracy.	Everyday	Checked automatically	68.3%	In use
ChartQA Models answer human-written and generated questions about chart images that need visual and arithmetic reasoning. Scored by relaxed accuracy with a 5 percent numeric tolerance.	School	Checked automatically	90.9%	In use
DocVQA Models answer 50,000 questions about 12,000 scanned document images by reading their text and layout. Scoring is ANLS, a string similarity to reference answers.	Everyday	Checked automatically	96.9%	In use
GDP.pdf 100 expert multimodal reasoning tasks over real economically important PDF documents. Scored by rubric-criteria pass rate.	Expert	Graded by a model	35%	In use
OCRBench Models answer 1,000 checked questions on text recognition, scene text, documents, key information extraction, and handwritten math. The score counts correct answers.	Everyday	Checked automatically	94%	In use
VCR (Visual Caption Restoration) Models restore caption text that is partly covered in an image, using only the visible pixel fragments. Scored by exact match and Jaccard similarity of the words.	Everyday	Checked automatically	75.6%	System cards only
ParseBench Document parsers process about 2,000 human-verified enterprise pages. Over 169,000 test rules score tables, charts, content faithfulness, formatting and visual grounding.	Expert	Checked automatically	78.9%	Not in cards
OmniDocBench v1.5 Models parse 1,355 PDF pages into text, tables and formulas. The Overall score averages text accuracy from edit distance, table TEDS and formula CDM.	Everyday	Checked automatically	94.6	In use
InfographicVQA Models answer about 30,000 questions on about 5,000 infographics that mix text, layout, charts, and graphics. Scoring is ANLS.	School	Checked automatically	92.6%	In use
TextVQA Models answer 45,336 questions about 28,408 images that require reading text in the image. Scoring is VQA accuracy against human answers.	Everyday	Checked automatically	84.8%	Fading
MTVQA-Test Models answer 6,778 expert-written questions about text in 2,116 images across 9 languages, such as Arabic, Japanese, and Thai. Scoring is accuracy.	Everyday	Checked automatically	56.6%	Fading
MMLongBench-Doc Models answer 1,062 questions about 130 long PDF documents averaging 49 pages, and some questions are unanswerable. Scoring is accuracy and F1.	University	Checked automatically	62.8%	Fading
Chartography 100 tasks on specialized professional charts (Kaplan-Meier, candlestick, Sankey, Bode). Answers graded against expert references.	Expert	Checked automatically	90.2%	In use
KRETA Models read and reason over Korean text in images across diverse visual contexts. It is a text-rich visual question answering test for Korean.	Everyday	Checked automatically	85.4%	Fading
Image2Struct Vision-language models turn images of webpages, LaTeX and musical scores into code that renders them. The render is compared with the input image.	University	Checked automatically	94.7	Not in cards
MADQA Agents search 800 PDF documents to answer 2,250 human-written questions. Scoring covers answer accuracy and how much search effort the agent spends.	Everyday	Graded by a model	85.2%	In use
CC-OCR Models read text in 7,058 images across four tracks: scene text, multilingual text, document parsing, and key information extraction. Each track has its own metric.	Everyday	Checked automatically	83.4%	Fading
ChartQA Pro Model answers 1,948 diverse questions over 1,341 charts, infographics and dashboards, including conversational and unanswerable ones. Accuracy.	Everyday	Checked automatically	72.3%	Fading
OCR Arena Users upload a document, compare two anonymous OCR or vision-language models side by side, and vote. Votes become Elo ratings.	Everyday	Ranked by votes	1,791	Not in cards
MDPBench Document parsers read 3,400 digital and photographed document images in 17 languages. Public and private splits limit leakage.	Everyday	Checked automatically	86.4	Not in cards
OmniDocBench Diverse PDF pages with layout, text, table and formula annotations. Parsing scored by edit distance and TEDS.	Everyday	Checked automatically	87.2%	In use
ChartMuseum Models answer 1,162 expert-annotated questions about real-world charts from 184 sources. Many questions need visual reasoning that text cannot replace.	Expert	Graded by a model	75.8%	Fading
CharXiv (DQ) Models answer descriptive questions about charts from arXiv papers, such as titles, axis labels and legend entries. It is the descriptive half of CharXiv.	University	Graded by a model	95.5%	System cards only
DUDE Models answer extractive, abstractive, and unanswerable questions about multi-page, multi-domain documents. Scored by ANLS.	Everyday	Checked automatically	82.8%	Fading
OmniDocBench v1.6 Models parse 1,651 PDF pages, including nested tables, dense formulas and unusual layouts, into text, tables and formulas. Scored by the v1.5 Overall formula.	Everyday	Checked automatically	95.2	Not in cards
OCR-Reasoning Models answer 1,069 human-annotated questions that need reasoning over text-rich images. The questions cover 6 reasoning abilities and 18 practical tasks.	Everyday	Graded by a model	65.1%	Fading
OmniParseBench Document parsers convert single pages to text, and 16,288 binary unit tests on 2,937 pages check the output. The headline joins text, table and layout subscores.	Everyday	Checked automatically	92.9	Not in cards
RealKIE-FCC Verified Model extracts structured fields from scanned FCC invoices into a JSON schema. Amazon human-verified labels. Field-level similarity (ALNS).	Everyday	Checked automatically	67%	Fading
WebQuest Models answer questions from screenshots of one web page, several related pages, or a navigation trace. The test checks extraction and combination of information across pages.	Everyday	Checked automatically	69.3%	Fading
BetterChartQA Models answer annotator-written questions about real charts from news, government reports, and papers across 9 capability buckets. Scored by accuracy.	School	Graded by a model	72.4%	Dropped
LongExtractionBench Extraction systems pull structured values from 225 long table-heavy documents averaging 358 pages and about 88,700 ground-truth fields each.	Expert	Checked automatically	70.7%	Not in cards
SEED-Bench-2-Plus Multimodal models answer multiple-choice questions about text-rich images such as charts, maps and web pages.	Everyday	Checked automatically	73%	System cards only
CC-OCR-Bench-V2 Models process 7,093 hard real-world document samples across five OCR tracks, including parsing and document QA. Each track uses its own metric.	Everyday	Checked automatically	79.6%	In use
ChemicalDiagramQA Models answer questions about chemical structure diagrams in scientific figures. Scored by accuracy.	University	Checked automatically	69.7%	Dropped
OpenScore String Quartets Models transcribe images of string quartet sheet music into symbolic notation. Scored as 1 minus normalized edit distance from the reference encoding.	University	Checked automatically	0.84	In use
QwenVisualOffice Models handle visual office tasks in an internal Qwen test set. Qwen has not published the tasks or how they are scored.	University	Checked automatically	44.6%	In use
TAT-DQA Models answer questions about financial report pages that mix tables and text, often needing arithmetic. Scored by exact match and F1.	University	Checked automatically	37.8%	Dropped
KWV Bench Models answer atomic visual questions extracted from tasks distilled from real knowledge-work scenarios.	Expert	Checked automatically	64.7%	System cards only
LongDocURL Models answer 2,325 questions over long documents with more than 33,000 pages in total. Its 20 subtasks cover understanding, numerical reasoning and locating document elements.	University	Checked automatically	74.7%	System cards only
OmniAI OCR Benchmark Multimodal models and OCR providers extract text and JSON fields from documents. Accuracy is measured on both plain text and structured extraction.	Everyday	Checked automatically	77%	System cards only
RD-TableBench Models extract 1,000 hand-labeled complex tables from document images, including scans, handwriting and merged cells. Scored by Levenshtein-based row and cell alignment.	Everyday	Checked automatically	57.4%	System cards only
ChartX Vision-language models do 7 chart tasks, such as question answering, summarization and redrawing, over 18 chart types and 22 topics.	Everyday	Graded by a model		Not in cards
CharXiv Extended Models answer descriptive and reasoning questions about scientific charts. Mercor adds 150 new expert-written tasks to the 5000 public ones and reports both scores.	University	Checked automatically		Not in cards
DP-Bench Document parsers and vision models convert PDFs and scanned pages into structured HTML or Markdown. Output is compared with reference structure.	Everyday	Checked automatically		Not in cards
ExtractBench Systems read a full business document and return a schema-valid JSON object with every requested value. Unified value F1 compares the output with the expected values.	Everyday	Checked automatically		Not in cards
ExtractBench Pro Systems extract every schema value from long-tail synthetic documents with charts, drawn relations, house rules and jargon. Scored by value F1 and by grounded F1, which also needs the source location.	Expert	Checked automatically		Not in cards
MP-DocVQA Models answer questions about scanned documents of up to 20 pages, where the answer sits on one page. Scored by ANLS against the reference answer.	Everyday	Checked automatically		Not in cards
olmOCR-Bench OCR systems convert 1,403 PDFs to markdown. 7,010 unit tests check text presence and absence, reading order, tables and math.	Everyday	Checked automatically		Not in cards
Sakana kana-letter reading-order set Model recovers the reading order of 25 expert-annotated classical Japanese kana letters (chirashigaki). Scored by normalized edit distance.	Expert	Checked automatically		Not in cards
SlideVQA Models answer questions about whole slide decks, with single-hop, multi-hop and numerical reasoning. The set has 14.5k questions over 2.6k decks.	Everyday	Checked automatically		Not in cards
SynthDocBench Vision-language models answer questions about synthetic long documents where length, layout, modality and difficulty are varied one at a time to locate failure causes.	Everyday	Checked automatically		Not in cards
TextEdit Image generation models edit text inside 2,148 real and synthetic images across 18 subcategories. OCR, image-fidelity metrics and a multimodal LLM judge score target accuracy, text preservation and scene integrity.	Everyday	Graded by a model		Not in cards
VDEBench Image editing models make local text and table edits to 3,506 Chinese and English document images. Scores check the edit and that the rest of the page is preserved.	Everyday	Checked automatically		Not in cards
WebSRC Models answer questions about a web page from its screenshot, HTML or both. The set has 400K question and answer pairs from 6.4K web pages. Scored by exact match and F1.	Everyday	Checked automatically		Not in cards

Visual math and diagrams

Benchmark	Level	Scored	Best score	Status
MATH-Vision (MathVision) 3,040 competition math problems with visual context. Answer accuracy.	Competition	Checked automatically	97.7%	In use
MathVista Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.	School	Checked automatically	90.7%	Fading
AI2D Models answer multiple choice questions about grade school science diagrams. The set has about 5,000 diagrams and 15,000 questions, scored by accuracy.	School	Checked automatically	96.9%	Fading
DynaMath Models answer visual math questions generated from 501 seed programs that vary values and graphs. Scored by average and worst-case accuracy across variants.	School	Checked automatically	88%	Fading
MathVerse Models solve diagram-based math problems shown in six versions that move information between the text and the image. It tests whether models read the diagram.	School	Checked automatically	89.7%	Fading
Math-VR Models answer 5,000 English and Chinese math questions whose solutions need visual reasoning, such as drawing auxiliary lines or plotting functions. Answers are checked against ground truth.	School	Checked automatically	66.8%	System cards only
Math Kangaroo 2025 (MathArena) Vision language models answer 168 multiple-choice problems from the 2025 Math Kangaroo, given as screenshots across six grade groups. Many problems need reading a diagram.	School	Checked automatically	94.9%	Fading
LogicVista Models answer 448 multiple-choice questions on spatial, deductive, inductive, numeric, and mechanical reasoning set in images. Scored by accuracy.	School	Checked automatically	91.9%	In use
We-Math Models solve visual math problems decomposed into sub-problems by knowledge concept. Scoring compares sub-problem and full-problem results to diagnose knowledge and generalization gaps.	School	Checked automatically	89%	Fading
EMMA Models solve math, physics, chemistry and coding problems that need reasoning over text and images together, not each alone. Answers are checked against a key.	University	Checked automatically	79.3%	Fading
Inter-GPS (Geometry3K) Models solve high school geometry problems that combine a diagram and text from the Geometry3K test set. Scored by accuracy.	School	Checked automatically	48.6%	Dropped
MathCanvas-Bench Models solve 3,000 math problems that call for interleaved visual and textual solutions, such as drawing auxiliary lines in geometry.	School	Checked automatically	61.9%	Fading
CSVQA Models answer 1,378 Chinese STEM questions in physics, chemistry, biology and math that need domain knowledge and evidence from 14 kinds of scientific images.	School	Checked automatically	63.3%	System cards only
MathVision-Wildphoto Models solve MathVision style problems shown in photos taken in natural settings, judging from the name. The Qwen3-VL report gives no citation or further description.	School	Checked automatically	63.8%	System cards only
GSM8K-V Vision-language models solve GSM8K word problems that are redrawn as multi-image visual scenes, so the facts must be read from pictures.	School	Checked automatically		Not in cards
MM-HELIX Multimodal models solve 1,260 items from 42 synthetic tasks that need long reflective reasoning with backtracking.	University	Checked automatically		Not in cards
MME-Reasoning Multimodal models answer logic questions grouped into inductive, deductive and abductive reasoning.	University	Checked automatically		Not in cards

Spatial reasoning and counting

Benchmark	Level	Scored	Best score	Status
BabyVision Model solves 388 perception-heavy visual puzzles (discrimination, tracking, spatial, patterns) that young children solve. Accuracy versus 94% human baseline.	Everyday	Checked automatically	91.3%	In use
VoxelBench Language models turn text or image prompts into voxel structures. Voters compare anonymous builds from the same prompt, and Glicko-2 ratings rank the models.	Everyday	Ranked by votes	2,757	Not in cards
RealWorldQA Models answer a question about each of 700+ real-world photos, many taken from vehicles, that test basic spatial understanding. Answers are easy to verify.	Everyday	Checked automatically	88.5%	In use
Blueprint-Bench 2 Agents turn about 20 interior photos of each of 50 apartments into 2D floor plans. Scores compare room connectivity with ground truth, normalized against random.	Everyday	Checked automatically	58.6%	In use
ClockBench Models read analog clock faces and answer questions about the time shown. Most clocks stay private to avoid training leakage.	Everyday	Checked automatically	83.3%	Not in cards
CountBench Models count the objects in natural images from the CountBench image-text set. Lab cards often use a question-answer version, CountBenchQA, scored by accuracy.	Everyday	Checked automatically	97.8%	In use
Furniture Assembly Benchmark Model gets an IKEA manual and a photo of a partly built item, then names the faulty step or next step. Accuracy on 60 photos.	Everyday	Graded by a model	83.3%	Not in cards
VisualPuzzles Models solve visual puzzles that need algorithmic, analogical, deductive, inductive and spatial reasoning but little domain knowledge. Scored by accuracy.	Everyday	Checked automatically	57%	Fading
VisuLogic Models answer 1,000 visual reasoning puzzles in six categories that cannot be solved by describing the image in words. Scored by accuracy.	Everyday	Checked automatically	54.3%	Fading
EmbSpatialBench Models answer multiple-choice questions on six spatial relations between objects in egocentric embodied scenes. Scored by accuracy.	Everyday	Checked automatically	85.4%	Fading
CVBench Models answer 2,638 manually checked 2D and 3D vision questions built from ADE20K, COCO, and Omni3D. Scoring is accuracy.	Everyday	Checked automatically	88%	Fading
VLMs Are Blind Models answer seven simple visual tasks such as counting line crossings or overlapping circles, which young children solve easily. Scored by accuracy.	Everyday	Checked automatically	98.6%	Fading
ScienceVision Agents inspect 3D scientific data such as proteins, crystals, CAD assemblies, buildings and CT volumes, then answer 250 expert questions. A rule-based scorer grades each answer field.	Expert	Checked automatically	0.53	Not in cards
All-Angles Bench Models answer over 2,100 multi-view questions about 90 real scenes. Six tasks cover counting, attributes, relative distance and direction, object manipulation and camera pose.	Everyday	Checked automatically	72.1%	Fading
MindCube Models answer 21,154 multiple choice questions about scenes seen from a few views. Questions test positions, orientations, and what-if movements.	Everyday	Checked automatically	46.7%	Fading
SpatialViz-Bench Models answer 1,180 generated multiple choice puzzles on mental rotation, mental folding, visual penetration, and mental animation. Scoring is accuracy.	Everyday	Checked automatically	44.7%	System cards only
VLMsAreBiased Models count or identify features in altered images of famous subjects, such as a logo with an added stripe, across 7 domains. Scoring is accuracy.	Everyday	Checked automatically	88.3%	Fading
Blueprint-Bench Models turn about 20 interior photos of each of 50 apartments into a 2D floor plan. Scored by similarity of room connectivity and size ranking to the true plan.	Everyday	Checked automatically	0.43	Not in cards
OmniSpatial Vision language models answer over 8,400 spatial reasoning questions in four groups: dynamic reasoning, complex spatial logic, spatial interaction and perspective taking.	Everyday	Checked automatically	52%	Fading
MMSI-Bench Models answer 1,000 human-written questions that need spatial reasoning across several images. Images come from indoor scans, driving, egocentric video and robot data.	Everyday	Checked automatically	42.5%	Fading
MMVP Models answer 300 questions about 150 image pairs that CLIP embeds as near duplicates. A pair counts only when both answers are right.	Everyday	Checked automatically	87%	Fading
TallyQA Models answer open-ended counting questions about images, including complex ones that need object relationships or attributes. Scored by exact count accuracy.	Everyday	Checked automatically	61.5%	Dropped
VisFactor Models take 20 vision-centric subtests adapted from a human cognitive psychology test battery, covering perception primitives and visual memory. Scored by accuracy.	Everyday	Checked automatically	60.8%	In use
VSI-Bench Models watch egocentric videos of 288 indoor scenes and answer over 5,000 questions on object counts, distances, sizes, room layout and route planning.	Everyday	Checked automatically	69.5%	Fading
CountQA Models count objects in over 1,500 real-world photos with dense clutter and occlusion. Scoring is accuracy on the count.	Everyday	Checked automatically	82.4%	In use
SpatialSense VQA Models judge whether a stated spatial relation between two objects holds in an image. The relations were crowdsourced to defeat simple cues.	Everyday	Checked automatically	60%	Dropped
SURDS Models answer spatial questions about objects in nuScenes driving camera images across six single-object and multi-object tasks. Scoring is task accuracy.	Everyday	Checked automatically	77.8%	In use
MME-CC Models answer 1,173 expert-annotated vision-grounded questions in 11 task types grouped as spatial, geometric and visual knowledge reasoning.	Everyday	Checked automatically	57%	Fading
ViewSpatial-Bench Models answer 5,712 questions that locate objects from the camera view and from the view of people in the scene. Five task types test egocentric and allocentric reasoning.	Everyday	Checked automatically	61.3%	Fading
3DSRBench Models answer questions about 3D spatial relations in images, such as height, location, orientation and relations between objects.	Everyday	Checked automatically	62.4%	Fading
Ego3D-Bench Models answer spatial reasoning questions about ego-centric multi-view scenes, such as those from self-driving cars. Some items ask for distances, scored by error.	Everyday	Checked automatically	5.90	In use
Image2FloorPlan Models study several real photos of a home and draw its floor plan. Scored by an average quality score.	Everyday	Graded by a model	48.0	In use
RoboSpatial-Home Models answer 350 spatial questions on home scenes for robots, in three types: object context, compatibility of placements and spatial configuration.	Everyday	Checked automatically	73.9%	Dropped
SITE Models answer multiple-choice spatial questions over single images, multiple images and video, drawn from about 30 vision datasets. Tasks span figural to environmental scale.	Everyday	Checked automatically	47.5%	Fading
SpaCE-10 Models answer more than 5,000 questions about indoor 3D scenes that test 10 atomic and 8 compositional spatial abilities.	Everyday	Checked automatically	55%	Dropped
SpatialEval Maze-Nav Models answer questions about routes and turns in mazes given as text or images. 1,500 multiple-choice questions scored by accuracy.	Everyday	Checked automatically	55.1%	System cards only
SpatialEval Spatial-Map Models answer questions about relations and counts among objects placed on a map given as text or images. 1,500 questions scored by accuracy.	Everyday	Checked automatically	73.7%	System cards only
VLADBench Models answer closed-form questions about driving scenes, from static traffic knowledge to reasoning about dynamic road situations.	University	Checked automatically	77.2%	In use
SAT (Spatial Aptitude Training) Models answer static and dynamic spatial questions, such as how object relations change when the camera or an object moves. The real-image test set has 150 questions.	Everyday	Checked automatically	76.7%	System cards only
STARE Models solve 4K spatial tasks such as 2D and 3D transformations, cube net folding, tangrams, perspective and temporal reasoning that people solve by visual simulation.	Everyday	Checked automatically	61.8%	System cards only
3D Arena Users compare two 3D models generated from the same input and vote for the better one. A live leaderboard ranks the generators.	Everyday	Ranked by votes		Not in cards
CrossPoint-Bench Models see two views of a scene and must find the point in the second view that matches a point in the first. Tasks go from perception and reasoning to exact coordinate prediction.	Everyday	Checked automatically		Not in cards
KilometerVision Models watch walking-tour videos geolocated over up to 1 km and answer questions on landmarks, routes and map layout, such as distances and loop closures.	Everyday	Checked automatically		Not in cards
LEGO-Puzzles Vision-language models answer LEGO assembly questions, from basic spatial understanding to planning several assembly steps. The elementary set has 1,100 questions over 11 tasks.	Everyday	Checked automatically		Not in cards
MC-Bench Models write code that builds a Minecraft structure from a prompt. Users vote between two builds without knowing the models.	Everyday	Ranked by votes		Not in cards
MineBench Models turn a prompt into a Minecraft-style voxel build as raw 3D coordinates or tool calls. Users compare builds blind, and votes become rankings.	Everyday	Ranked by votes		Not in cards
RoTA-Bench Vision-language models rotate a 3D object step by step, using visual feedback, until its projection matches a target silhouette.	Everyday	Checked automatically		Not in cards
SpatialBench (Ropedia) Spatial foundation models are tested across many 3D tasks, viewpoints, scene domains, input densities and hardware limits, beyond the domains they were trained for.	Everyday	Checked automatically		Not in cards
SpatialWorld Multimodal agents act in real-world style environments to show interactive spatial understanding, rather than answer static spatial questions.	Everyday	Checked automatically		Not in cards
VSI-Bench-Debiased Models answer the 2,362 VSI-Bench questions left after removing 2,768 that text-only models could solve from answer priors or world knowledge.	Everyday	Checked automatically		Not in cards
VSR Models judge whether captions with 65 kinds of spatial relation, such as under or facing, are true of an image. Over 10,000 text and image pairs.	Everyday	Checked automatically		Not in cards

Detection and captioning

Benchmark	Level	Scored	Best score	Status
RefCOCO / RefCOCO+ / RefCOCOg Model grounds a referring expression to a bounding box in a COCO image. Scored by accuracy at IoU 0.5.	Everyday	Checked automatically	95%	In use
Point-Bench Models point to the image location a language request describes, in about 1,000 tasks across five reasoning categories. A point inside the target mask counts as correct.	Everyday	Checked automatically	67.1%	Fading
ODinW-13 Models detect objects in 13 varied real-world datasets, such as aquarium and aerial images, from text category names. Scored by mean average precision.	Everyday	Checked automatically	51.8%	Fading
TreeBench Models answer 405 visual questions about small targets in dense SA-1B images and give bounding boxes as evidence. Scored by answer accuracy, with box overlap as a second measure.	University	Checked automatically	71.1%	Fading
SUNRGBD Models detect objects in about 10,000 indoor RGB-D images and predict their 3D bounding boxes. Scored by average precision.	Everyday	Checked automatically	48.3	Fading
XLRS-Bench Models answer perception and reasoning questions about extremely large, ultra-high-resolution remote sensing images.	Expert	Checked automatically	55%	Fading
Hypersim Models localize named objects as 3D boxes in synthetic indoor Hypersim images, using the Omni3D test split. Scored by mAP at IoU 0.15.	Everyday	Checked automatically	12.7	Fading
Ref-L4 Models locate the object described by a long referring expression in an image. The set has 45,341 annotations over 365 object categories, scored by box accuracy.	Everyday	Checked automatically	89.5%	Fading
COCO Captions Models write a caption for each image, and outputs are compared with five human captions per image using BLEU, METEOR, ROUGE, and CIDEr.	Everyday	Checked automatically	116.0	Dropped
Dense 200 (bbox) Model detects and boxes every object in 200 densely annotated images (109 categories, about 91 boxes each). Box F1 at IoU thresholds.	Everyday	Checked automatically	87%	In use
Nuscene Models localize named objects as 3D boxes in nuScenes driving camera images. Scored by mAP.	Everyday	Checked automatically	16.0	Fading
PixMo-Count Models count objects in natural images and give the exact number. Scored by accuracy on the test split.	Everyday	Checked automatically	83.5%	Fading
ARKitScenes Models locate objects in indoor RGB-D scenes captured on iPads and output 3D bounding boxes. Scored by 3D detection average precision.	Everyday	Checked automatically	53.7%	Dropped
COCO Models detect and locate objects in everyday photos from the Microsoft COCO dataset. Scored by mean average precision over box overlap thresholds.	Everyday	Checked automatically	78.7%	In use
LVIS-MG Models ground many objects at once by outputting boxes for named categories in LVIS images.	Everyday	Checked automatically	72.5%	Dropped
RefAdv-S Models locate the object a referring expression describes among hard distractors in real images. Scored by accuracy at 0.5 box overlap on 1,142 cases.	Everyday	Checked automatically	80.2%	In use
VisualProbe Models search high-resolution images for small targets to answer questions, often over many crop and zoom turns. Answers are checked against a key.	Everyday	Checked automatically	65%	Fading
Objectron Models predict 3D bounding boxes of objects in short object-centric videos across nine categories. Scored by 3D intersection over union.	Everyday	Checked automatically		Dropped

Image generation and editing

Benchmark	Level	Scored	Best score	Status
LMArena Text-to-Image Arena Image generation models answer the same user prompt side by side and anonymously. Community votes between the two images give Elo-style ratings.	Everyday	Ranked by votes	1,427	Not in cards
LMArena Image Edit Arena Image editing models apply the same user edit request to the same image, anonymously and side by side. Community votes give Elo-style ratings.	Everyday	Ranked by votes	1,517	Not in cards
GenEval Text-to-image models draw prompts about object count, color, position and co-occurrence, and an object detector checks each image. Scored by the share of correct images.	Everyday	Checked automatically	90.1%	Fading
K-Sort Arena Users rank outputs from K image or video generation models at once instead of a pair. Bayesian updates turn the rankings into ratings.	Everyday	Ranked by votes	31	Not in cards
DPG-Bench Text-to-image models draw 1,000 dense prompts with many objects, attributes and relations. A VQA model checks each element of the prompt graph in the image.	Everyday	Graded by a model	84.2%	Dropped
GEdit-Bench Image editing models follow real-world user edit instructions. A model judge scores semantic consistency, perceptual quality and an overall score.	Everyday	Graded by a model	6.52	Dropped
ImgEdit-Bench Image editing models carry out edit instructions such as add, remove, replace, style transfer and background change. A GPT-4.1 judge scores each result from 1 to 5.	Everyday	Graded by a model	4.12	System cards only
KRIS-Bench Image editing models follow 1,267 instructions that need factual, conceptual or procedural knowledge across 22 tasks. A GPT-4o judge scores the edits.	School	Graded by a model	60.2	System cards only
WISE Text-to-image models draw 1,000 prompts that need world knowledge in culture, space, time and natural science. A vision model judge gives the WiScore.	School	Graded by a model	0.70	System cards only
AA-Image-Editing Validated private evaluators compare two image edits made from the same instruction and vote blind. Votes become Elo ratings.	Everyday	Ranked by votes		Not in cards
Artificial Analysis Image Arena Voters compare images from two text-to-image models for the same prompt and pick the better one. Votes become Elo ratings.	Everyday	Ranked by votes		Not in cards
DesignArena Image Image models generate a picture from a user prompt. Users compare two anonymous images and vote for the better one.	Everyday	Ranked by votes		Not in cards
DesignArena Image Editing Image models edit a user image following a text instruction. Users compare the two edits and vote for the better one.	Everyday	Ranked by votes		Not in cards
GenAI-Arena Users compare outputs from two anonymous models for image generation, image editing and video generation, and vote. Votes give Elo ratings.	Everyday	Ranked by votes		Not in cards
HEIM Text-to-image models are scored on 12 aspects, such as alignment, quality, aesthetics, originality, reasoning, knowledge, bias and toxicity, with human raters and automatic metrics.	Everyday	Graded by people		Not in cards
MPIE-Bench Image editing models place several named people into shared contact actions such as an embrace or a carry. Scoring targets fused limbs, extra extremities and bodies passing through each other.	Everyday	Graded by a model		Not in cards