MathVista test

Name: MathVista test
Creator: MathVista authors and website maintainers
License: https://creativecommons.org/licenses/by-sa/4.0/

We show this table for reference; we do not rank on it.

Data verified June 9, 2025

MathVista tests mathematical reasoning about images. This table shows published accuracy on the test split, with each source configuration kept separate. The results are display only and do not enter overall or category rankings.

MathVista by MathVista authors and website maintainers and contributors. Original benchmark · CC BY-SA 4.0. Selected aggregate scores and adjacent metadata; converted numeric strings to numbers and missing-score markers to null; separated testmini from test; added configuration notes and three named profile references. Selected aggregate-result compilation and its adaptation. Task content, source images, model outputs, and model weights are excluded.

Accuracy (%) on MathVista test — June 9, 2025

We mirror the published accuracy (%) view for MathVista test. InternVL2-Pro has the highest reported value at 65.84%. The compilation contains different evaluation setups; the highest reported value does not establish a comparable model ranking. We do not use these results to rank models overall.

InternVL2-Pro

Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2024-09-04 (compiler field)

65.84%

InternVL-Chat-V1.2_Plus

Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2024-02-22 (compiler field)

60.18%

InternLM-XComposer2-VL-7B

Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2024-01-22 (compiler field)

57.93%

9 model or configuration recordsMultimodal & GroundedHistorical source configurationsDisplay onlyUpdated June 9, 2025

MathVista test: 9 published model or configuration records
Source configuration	Accuracy	Evaluation setup and sources	Parameters (B)	Open / closed
InternVL2-Pro	65.84%	Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2024-09-04 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 65.02% geometry problem solving 64% math word problem 75.44% textbook question answering 72.4% visual question answering 52.27% algebraic reasoning 67.35% arithmetic reasoning 63.1% geometry reasoning 64.96% logical reasoning 30.41% numeric commonsense 44.54% scientific reasoning 66.98% statistical reasoning 72.47%	Not reported	Unknown
InternVL-Chat-V1.2_Plus	60.18%	Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2024-02-22 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 52.18% geometry problem solving 56.17% math word problem 78.3% textbook question answering 61.62% visual question answering 55.53% algebraic reasoning 55.96% arithmetic reasoning 64.37% geometry reasoning 57.56% logical reasoning 21.65% numeric commonsense 46.08% scientific reasoning 60.04% statistical reasoning 60.1%	Not reported	Unknown
InternLM-XComposer2-VL-7B	57.93%	Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2024-01-22 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 53.92% geometry problem solving 56.44% math word problem 77.12% textbook question answering 58.41% visual question answering 43.19% algebraic reasoning 54.81% arithmetic reasoning 57.64% geometry reasoning 57.98% logical reasoning 16.49% numeric commonsense 47.62% scientific reasoning 59.1% statistical reasoning 62.46%	Not reported	Unknown
Qwen-VL-Plus	44.33%	Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2023-12-21 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 55.88% geometry problem solving 34.74% math word problem 29.68% textbook question answering 58.79% visual question answering 42.37% algebraic reasoning 40.7% arithmetic reasoning 35.4% geometry reasoning 36.64% logical reasoning 21.65% numeric commonsense 30.39% scientific reasoning 55.91% statistical reasoning 56.34%	Not reported	Unknown
SPHINX-MoE	42.68%	Published method: MoE 🤖; Additional checkpoint and evaluation settings unverified; Source date 2024-01-13 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 50.29% geometry problem solving 29.7% math word problem 40.93% textbook question answering 49.29% visual question answering 43.31% algebraic reasoning 33.88% arithmetic reasoning 43.05% geometry reasoning 29.08% logical reasoning 14.43% numeric commonsense 26.33% scientific reasoning 46.9% statistical reasoning 51.24%	Not reported	Unknown
MiniCPM-V-2 (2.8B)	39.89%	Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2024-04-14 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 51.74% geometry problem solving 27.45% math word problem 39.84% textbook question answering 42.49% visual question answering 34.69% algebraic reasoning 31.29% arithmetic reasoning 34.43% geometry reasoning 30.67% logical reasoning 13.4% numeric commonsense 33.47% scientific reasoning 38.46% statistical reasoning 50.03%	Not reported	Unknown
PoT GPT-4 (Caption+OCR)	31.74%	Caption and OCR inputs; published prompting setup; remaining tools and sampling settings unverified; Source date 2023-10-03 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 27.58% geometry problem solving 37.35% math word problem 23.87% textbook question answering 43% visual question answering 30.27% algebraic reasoning 37.15% arithmetic reasoning 27.93% geometry reasoning 37.48% logical reasoning 22.68% numeric commonsense 15.83% scientific reasoning 44.47% statistical reasoning 31.87%	Not reported	Unknown
CoT GPT4 (Caption+OCR)	30.5%	Caption and OCR inputs; published prompting setup; remaining tools and sampling settings unverified; Source date 2023-10-03 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 27.21% geometry problem solving 35.91% math word problem 21.3% textbook question answering 43.13% visual question answering 28.17% algebraic reasoning 35.72% arithmetic reasoning 25.17% geometry reasoning 35.8% logical reasoning 24.74% numeric commonsense 15.41% scientific reasoning 47.28% statistical reasoning 31.29%	Not reported	Unknown
LLaVA (LLaMA-2-13B)	25.4%	Published method: LMM 🖼️; Additional checkpoint and evaluation settings unverified; Source date 2023-10-03 (compiler field) Pinned aggregate Source metadata Original result Published task and skill scores figure question answering 22.86% geometry problem solving 24.57% math word problem 18.15% textbook question answering 35.82% visual question answering 29.69% algebraic reasoning 26.93% arithmetic reasoning 22.47% geometry reasoning 24.45% logical reasoning 19.07% numeric commonsense 19.05% scientific reasoning 34.71% statistical reasoning 21.61%	Not reported	Unknown

Reference baselines, outside the model table: Random Chance: 17.86%. Missing values remain unreported. These results use different evaluation setups and do not enter overall or category rankings.

How to read this leaderboard

Scores are published accuracy percentages on a 0-100 scale. Missing task and skill scores stay empty; the importer does not calculate a replacement headline score. Compare each result with its prompting, tools, images, and compute settings.

Operator receipt: 9 sourced rows are currently displayable on this page; the highest published score among these 9 model or configuration records is InternVL2-Pro at 65.84%, which does not establish a market leader.

Honest limit: The website compiles results from multiple publishers and evaluation setups. Exact checkpoints, sample counts, and sampling settings are often undisclosed. Three named profile links are navigation references; they do not establish an exact canonical model evaluation. The testmini and test tables are separate.

MathVista project and leaderboard Pinned MathVista result compilation MathVista paper Website compilation license declaration CC BY-SA 4.0

How to read the MathVista test results

This table contains 9 model or configuration records from the MathVista website compilation for test. The split has 5,141 examples; the aggregate files do not state a separate completed-example count for each record. The pinned source commit is dated June 9, 2025, and the source files were observed on October 15, 2026.

Each score belongs to the published configuration shown below it. Prompting, tools, image processing, and test-time compute can change the result. We keep the two splits separate and exclude these scores from overall and category rankings.

Reference baselines are outside the model table: Random Chance: 17.86%.

MathVista website compilation, CC BY-SA 4.0. Original result sources remain credited in each record. This adaptation retains CC BY-SA 4.0. Changes: Selected aggregate scores and adjacent metadata; converted numeric strings to numbers and missing-score markers to null; separated testmini from test; added configuration notes and three named profile references.

Snapshot

10 source records9 model or configuration records5,141 examples in test0 named profile referencesDisplay only

MathVista project and leaderboard Pinned MathVista result compilation MathVista paper Website compilation license declaration CC BY-SA 4.0

The compilation contains different evaluation setups; the highest reported value does not establish a comparable model ranking. The table retains each source configuration and its reported value.

9 model or configuration records are shown for MathVista test. The benchmark falls in the Multimodal & Grounded category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. MathVista test is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About MathVista test

Year

2023

Tasks

5,141 examples in the test split

Format

Published accuracy (%)

Difficulty

Mathematical reasoning in visual contexts

The test split contains 5,141 examples. The pinned website compilation was committed on June 9, 2025 and observed on October 1, 2026. Source record counts describe published configurations, not distinct models or confirmed complete evaluation runs.

MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts Public benchmark source

Freshness and provenance

Version

test; source commit a9daca33c6b3

Refresh cadence

Pinned website compilation

Staleness state

Historical source configurations

Question availability

Public upstream; no task contents mirrored

Historical source configurationsDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Questions

What does MathVista test measure?

Which source row has the highest reported MathVista test value?

InternVL2-Pro has the highest reported MathVista test value at 65.84%. The compilation contains different evaluation setups; the highest reported value does not establish a comparable model ranking. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on MathVista test?

The June 9, 2025 snapshot contains 9 published model or configuration records.

Last updated: June 9, 2025 · mirrored from the public benchmark leaderboard

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 5,600+ readers.

One email each week. Unsubscribe anytime.