ColorBench
Models answer image questions about color perception, color reasoning, and robustness to color changes. The evaluator checks the selected answers against a single correct option.
The dataset contains 1,448 instances and more than 5,800 image-text questions across 11 tasks.
- Made by
- University of Maryland
- Released
- April 2025
- Kind
- General visual questions
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision
- Strongest model tested
-
Gemini 2.0 Flash Feb 2025 #147 on the capability index
- Last model added
- February 2025
- Scores refreshed
- November 20th, 2026
In the authors’ words
Color plays an important role in human perception and usually provides critical clues in visual reasoning. However, it is unclear whether and how vision-language models (VLMs) can perceive, understand, and leverage color as humans. This paper introduces ColorBench, an innovative benchmark meticulously crafted to assess the capabilities of VLMs in color understanding, including color perception, reasoning, and robustness. By curating a suite of diverse test scenarios, with grounding in real applications, ColorBench evaluates how these models perceive colors, infer meanings from color-based cues, and maintain consistent performance under varying color transformations. Through an extensive evaluation of 32 VLMs with varying language models and vision encoders, our paper reveals some undiscovered findings: (i) The scaling law (larger models are better) still holds on ColorBench, while the language model plays a more important role than the vision encoder. (ii) However, the performance gaps across models are relatively small, indicating that color understanding has been largely neglected by existing VLMs. (iii) CoT reasoning improves color understanding accuracies and robustness, though they are vision-centric tasks. (iv) Color clues are indeed leveraged by VLMs on ColorBench but they can also mislead models in some tasks. These findings highlight the critical limitations of current VLMs and underscore the need to enhance color comprehension. Our ColorBenchcan serve as a foundational tool for advancing the study of human-level color understanding of multimodal AI.
Gemini 2.0 Flash leads with 59.6% among models released by February 2025 Gemini 2.0 Flash leads with 59.6% among models released by February 2025
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in February 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 59.6
Gemini 2.0 Flash
- 57.4 GPT-4o (snapshot unspecified)
- 51.9 LLaVA-OV-72B
- 50.0
InternVL2.5-38B
- 48.8
InternVL2_5-78B
- 46.8
InternVL2.5-26B
- 46.6
LLaVA-v1.6-34B
- 46.3
InternVL2-26B
- 46.2
Qwen2.5-VL-7B
- 45.6
InternVL2-40B
- 45.3
Cambrian-1-34B
- 45.2
LLaVA-NeXT-72B
- 45.2
InternVL2.5-8B
- 44.7
LLaVA-OneVision 7B
- 44.6 InternVL2-76B
- 44.1 Eagle-X4-8B
- 43.7 Eagle-X4-13B
- 43.4 Eagle-X5-34B
- 43.1
InternVL2-8B
- 42.8
Cambrian-1-13B
- 42.3
Cambrian-1-8B
- 41.5 Cambrian-3B
- 41.1
Qwen2.5-VL-3B
- 40.0
Eagle-X5-7B
- 38.5
InternVL2.5-2B
- 38.3
InternVL2.5-1B
- 36.4 llava-v1.6-vicuna-13b
- 36.4
InternVL2-2B
- 33.6
InternVL2-1B
- 33.4 LLaVA-Next-m-7B
- 32.6 LLaVA-OV-0.5B
- 31.2 LLaVA-Next-v-7B
- 1
Gemini 2.0 Flash Feb 2025, CoT 59.6%
- 2 GPT-4o (snapshot unspecified) CoT 57.4%
- 3 LLaVA-OV-72B 51.9%
- 4
InternVL2.5-38B 50%
- 5
InternVL2_5-78B 48.8%
- 6
InternVL2.5-26B 46.8%
- 7
LLaVA-v1.6-34B 46.6%
- 8
InternVL2-26B 46.3%
- 9
Qwen2.5-VL-7B 46.2%
- 10
InternVL2-40B 45.6%
- 11
Cambrian-1-34B 45.3%
- 12
LLaVA-NeXT-72B 45.2%
- 13
InternVL2.5-8B 45.2%
- 14
LLaVA-OneVision 7B 44.7%
- 15 InternVL2-76B 44.6%
- 16 Eagle-X4-8B 44.1%
- 17 Eagle-X4-13B 43.7%
- 18 Eagle-X5-34B 43.4%
- 19
InternVL2-8B 43.1%
- 20
Cambrian-1-13B 42.8%
- 21
Cambrian-1-8B 42.3%
- 22 Cambrian-3B 41.5%
- 23
Qwen2.5-VL-3B 41.1%
- 24
Eagle-X5-7B 40%
- 25
InternVL2.5-2B 38.5%
- 26
InternVL2.5-1B 38.3%
- 27 llava-v1.6-vicuna-13b 36.4%
- 28
InternVL2-2B 36.4%
- 29
InternVL2-1B 33.6%
- 30 LLaVA-Next-m-7B 33.4%
- 31 LLaVA-OV-0.5B 32.6%
- 32 LLaVA-Next-v-7B 31.2%
Source: The ColorBench paper
- 59.6
Gemini 2.0 Flash
- 50.0
InternVL2.5-38B
- 48.8
InternVL2_5-78B
- 46.8
InternVL2.5-26B
- 46.6
LLaVA-v1.6-34B
- 46.3
InternVL2-26B
- 46.2
Qwen2.5-VL-7B
- 45.6
InternVL2-40B
- 45.3
Cambrian-1-34B
- 45.2
LLaVA-NeXT-72B
- 45.2
InternVL2.5-8B
- 44.7
LLaVA-OneVision 7B
- 43.1
InternVL2-8B
- 42.8
Cambrian-1-13B
- 42.3
Cambrian-1-8B
- 41.1
Qwen2.5-VL-3B
- 40.0
Eagle-X5-7B
- 38.5
InternVL2.5-2B
- 38.3
InternVL2.5-1B
- 36.4 llava-v1.6-vicuna-13b
- 36.4
InternVL2-2B
- 33.6
InternVL2-1B
- 1
Gemini 2.0 Flash Feb 2025, CoT 59.6%
- 2
InternVL2.5-38B 50%
- 3
InternVL2_5-78B 48.8%
- 4
InternVL2.5-26B 46.8%
- 5
LLaVA-v1.6-34B 46.6%
- 6
InternVL2-26B 46.3%
- 7
Qwen2.5-VL-7B 46.2%
- 8
InternVL2-40B 45.6%
- 9
Cambrian-1-34B 45.3%
- 10
LLaVA-NeXT-72B 45.2%
- 11
InternVL2.5-8B 45.2%
- 12
LLaVA-OneVision 7B 44.7%
- 13
InternVL2-8B 43.1%
- 14
Cambrian-1-13B 42.8%
- 15
Cambrian-1-8B 42.3%
- 16
Qwen2.5-VL-3B 41.1%
- 17
Eagle-X5-7B 40%
- 18
InternVL2.5-2B 38.5%
- 19
InternVL2.5-1B 38.3%
- 20 llava-v1.6-vicuna-13b 36.4%
- 21
InternVL2-2B 36.4%
- 22
InternVL2-1B 33.6%
Source: The ColorBench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 59.6% | CoT | Feb 2025 | Official leaderboard |
| 2 | GPT-4o (snapshot unspecified) closed weights | 57.4% | CoT | Official leaderboard | |
| 3 | LLaVA-OV-72B | 51.9% | Official leaderboard | ||
| 4 | | 50% | Dec 2024 | Official leaderboard | |
| 5 | | 48.8% | Dec 2024 | Official leaderboard | |
| 6 | | 46.8% | Dec 2024 | Official leaderboard | |
| 7 | | 46.6% | Jan 2024 | Official leaderboard | |
| 8 | | 46.3% | Jul 2024 | Official leaderboard | |
| 9 | | 46.2% | Jan 2025 | Official leaderboard | |
| 10 | | 45.6% | Jul 2024 | Official leaderboard | |
| 11 | | 45.3% | Jun 2024 | Official leaderboard | |
| 12 | | 45.2% | May 2024 | Official leaderboard | |
| 13 | | 45.2% | Dec 2024 | Official leaderboard | |
| 14 | | 44.7% | Aug 2024 | Official leaderboard | |
| 15 | InternVL2-76B | 44.6% | Official leaderboard | ||
| 16 | Eagle-X4-8B | 44.1% | Official leaderboard | ||
| 17 | Eagle-X4-13B | 43.7% | Official leaderboard | ||
| 18 | Eagle-X5-34B | 43.4% | Official leaderboard | ||
| 19 | | 43.1% | Jul 2024 | Official leaderboard | |
| 20 | | 42.8% | Jun 2024 | Official leaderboard | |
| 21 | | 42.3% | Jun 2024 | Official leaderboard | |
| 22 | Cambrian-3B | 41.5% | Official leaderboard | ||
| 23 | | 41.1% | Jan 2025 | Official leaderboard | |
| 24 | | 40% | Aug 2024 | Official leaderboard | |
| 25 | | 38.5% | Dec 2024 | Official leaderboard | |
| 26 | | 38.3% | Dec 2024 | Official leaderboard | |
| 27 | llava-v1.6-vicuna-13b | 36.4% | Jan 2024 | Official leaderboard | |
| 28 | | 36.4% | Jul 2024 | Official leaderboard | |
| 29 | | 33.6% | Jul 2024 | Official leaderboard | |
| 30 | LLaVA-Next-m-7B | 33.4% | Official leaderboard | ||
| 31 | LLaVA-OV-0.5B | 32.6% | Official leaderboard | ||
| 32 | LLaVA-Next-v-7B | 31.2% | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 59.6% | CoT | Feb 2025 | Official leaderboard |
| 2 | | 50% | Dec 2024 | Official leaderboard | |
| 3 | | 48.8% | Dec 2024 | Official leaderboard | |
| 4 | | 46.8% | Dec 2024 | Official leaderboard | |
| 5 | | 46.6% | Jan 2024 | Official leaderboard | |
| 6 | | 46.3% | Jul 2024 | Official leaderboard | |
| 7 | | 46.2% | Jan 2025 | Official leaderboard | |
| 8 | | 45.6% | Jul 2024 | Official leaderboard | |
| 9 | | 45.3% | Jun 2024 | Official leaderboard | |
| 10 | | 45.2% | May 2024 | Official leaderboard | |
| 11 | | 45.2% | Dec 2024 | Official leaderboard | |
| 12 | | 44.7% | Aug 2024 | Official leaderboard | |
| 13 | | 43.1% | Jul 2024 | Official leaderboard | |
| 14 | | 42.8% | Jun 2024 | Official leaderboard | |
| 15 | | 42.3% | Jun 2024 | Official leaderboard | |
| 16 | | 41.1% | Jan 2025 | Official leaderboard | |
| 17 | | 40% | Aug 2024 | Official leaderboard | |
| 18 | | 38.5% | Dec 2024 | Official leaderboard | |
| 19 | | 38.3% | Dec 2024 | Official leaderboard | |
| 20 | llava-v1.6-vicuna-13b | 36.4% | Jan 2024 | Official leaderboard | |
| 21 | | 36.4% | Jul 2024 | Official leaderboard | |
| 22 | | 33.6% | Jul 2024 | Official leaderboard |
Rankings by part
We read 10 parts from the benchmark’s paper in 2 groups: color perception and color reasoning. Each ranks models on one part of the test. Gemini 2.0 Flash leads 6 of them, and 4 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
C’Recog
22 models we track have a score here.
- 1
Gemini 2.0 Flash Feb 2025 82.9%
- 2
InternVL2.5-8B ↑9 77.6%
- 3
Qwen2.5-VL-7B ↑4 76.3%
- 4
Cambrian-1-34B ↑5 75%
- 5
InternVL2_5-78B ↓2 75%
- 6
LLaVA-NeXT-72B ↑4 72.4%
- 7
Cambrian-1-8B ↑8 72.4%
- 8
InternVL2-26B ↓2 72.4%
- 9
InternVL2-8B ↑4 72.4%
- 10
InternVL2-40B ↓2 72.4%
C’Extract
22 models we track have a score here.
- 1
InternVL2.5-38B ↑1 60.4%
- 2
InternVL2_5-78B ↑1 58.3%
- 3
Cambrian-1-34B ↑6 57.3%
- 4
Gemini 2.0 Flash Feb 2025 ↓3 56.2%
- 5
LLaVA-NeXT-72B ↑5 54.2%
- 6
LLaVA-OneVision 7B ↑6 53.1%
- 7
InternVL2-26B ↓1 52.1%
- 8
InternVL2-40B 52.1%
- 9
InternVL2-8B ↑4 50%
- 10
Qwen2.5-VL-7B ↓3 49%
O’Recog
22 models we track have a score here.
- 1
InternVL2.5-26B ↑3 89.6%
- 2
InternVL2.5-38B 89.6%
- 3
Gemini 2.0 Flash Feb 2025 ↓2 88.3%
- 4
InternVL2-26B ↑2 87%
- 5
Qwen2.5-VL-7B ↑2 84.4%
- 6
InternVL2-40B ↑2 83.1%
- 7
InternVL2.5-8B ↑4 83.1%
- 8
LLaVA-OneVision 7B ↑4 81.8%
- 9
InternVL2_5-78B ↓6 81.8%
- 10
LLaVA-NeXT-72B 79.2%
C’Prop
22 models we track have a score here.
- 1
Gemini 2.0 Flash Feb 2025 58%
- 2
InternVL2.5-38B 53.8%
- 3
InternVL2-26B ↑3 52.5%
- 4
LLaVA-OneVision 7B ↑8 52.5%
- 5
InternVL2-40B ↑3 51.3%
- 6
Cambrian-1-34B ↑3 50%
- 7
InternVL2.5-8B ↑4 50%
- 8
Cambrian-1-8B ↑7 48.8%
- 9
Qwen2.5-VL-7B ↓2 47.5%
- 10
Cambrian-1-13B ↑4 46.3%
C’Comp
22 models we track have a score here.
- 1
Gemini 2.0 Flash Feb 2025 70.3%
- 2
InternVL2_5-78B ↑1 68.3%
- 3
InternVL2.5-26B ↑1 63.4%
- 4
InternVL2.5-38B ↓2 63.4%
- 5
InternVL2.5-8B ↑6 62.4%
- 6
InternVL2-40B ↑2 61.4%
- 7
LLaVA-v1.6-34B ↓2 56.4%
- 8
InternVL2-26B ↓2 56.4%
- 9
Cambrian-1-8B ↑6 54.5%
- 10
LLaVA-OneVision 7B ↑2 53.5%
C’Count
22 models we track have a score here.
- 1
Gemini 2.0 Flash Feb 2025 43.1%
- 2
Cambrian-1-13B ↑12 32.4%
- 3
Cambrian-1-8B ↑12 31.4%
- 4
InternVL2.5-38B ↓2 29.4%
- 5
LLaVA-v1.6-34B 28.4%
- 6 llava-v1.6-vicuna-13b ↑14 27.5%
- 7
InternVL2_5-78B ↓4 27.5%
- 8
InternVL2.5-2B ↑10 25.5%
- 9
InternVL2.5-8B ↑2 25.5%
- 10
LLaVA-NeXT-72B 24.5%
O’Count
22 models we track have a score here.
- 1
LLaVA-v1.6-34B ↑4 41.8%
- 2
InternVL2.5-38B 40.8%
- 3
Gemini 2.0 Flash Feb 2025 ↓2 38.8%
- 4
InternVL2_5-78B ↓1 36.9%
- 5
LLaVA-NeXT-72B ↑5 35.9%
- 6
InternVL2-8B ↑7 35.9%
- 7
InternVL2-40B ↑1 35.9%
- 8
Cambrian-1-13B ↑6 35%
- 9
InternVL2-26B ↓3 35%
- 10
Eagle-X5-7B ↑7 35%
C’Illu
22 models we track have a score here.
- 1
LLaVA-OneVision 7B ↑11 48.4%
- 2
Eagle-X5-7B ↑15 44.1%
- 3
Qwen2.5-VL-7B ↑4 44.1%
- 4
Gemini 2.0 Flash Feb 2025 ↓3 44.1%
- 5
InternVL2.5-1B ↑14 43%
- 6
Qwen2.5-VL-3B ↑10 43%
- 7
Cambrian-1-8B ↑8 41.9%
- 8
Cambrian-1-13B ↑6 38.7%
- 9
InternVL2-8B ↑4 38.7%
- 10
Cambrian-1-34B ↓1 37.6%
C’Mimic
22 models we track have a score here.
- 1
Gemini 2.0 Flash Feb 2025 75.7%
- 2
Cambrian-1-34B ↑7 64.3%
- 3
InternVL2.5-26B ↑1 62.9%
- 4
LLaVA-v1.6-34B ↑1 61.4%
- 5
InternVL2.5-38B ↓3 61.4%
- 6
InternVL2_5-78B ↓3 61.4%
- 7
InternVL2-40B ↑1 58.6%
- 8
Cambrian-1-8B ↑7 57.1%
- 9
Cambrian-1-13B ↑5 55.7%
- 10
InternVL2-26B ↓4 55.7%
C’Blind
22 models we track have a score here.
- 1
Gemini 2.0 Flash Feb 2025 60%
- 2
LLaVA-NeXT-72B ↑8 34.4%
- 3
InternVL2-1B ↑19 33.1%
- 4
LLaVA-v1.6-34B ↑1 29.9%
- 5
InternVL2.5-26B ↓1 29.3%
- 6
InternVL2_5-78B ↓3 28.7%
- 7
Qwen2.5-VL-7B 28.7%
- 8
InternVL2.5-1B ↑11 28%
- 9
InternVL2-26B ↓3 27.4%
- 10
InternVL2.5-38B ↓8 26.8%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The ColorBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official ColorBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.
Qwen 3.6 Plus 86%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.
Seed2.1 Pro 90.7%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7%
- MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.
Qwen 3.8 Max 85.9%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 21 months ago, in February 2025, for Gemini 2.0 Flash. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (32 entries).
- Pass Models scored 22 models we track, plus 10 other entries on its leaderboard.
- Warning Newest score The newest came 21 months ago, in February 2025, for Gemini 2.0 Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 2.0 Flash (#147).
- Pass Headroom The best is 59.6%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 23 rows for models we track and 10 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.