MathVista
BenchmarkFreeVisual mathematical reasoning benchmark.
Capabilities8 decomposed
multimodal mathematical reasoning evaluation
Medium confidenceEvaluates how well multimodal AI models can interpret visual mathematical representations (geometry diagrams, statistical plots, scientific figures) and answer questions requiring compositional reasoning combining visual perception with mathematical problem-solving. Uses a curated dataset of 6,141 examples sourced from 28 existing datasets plus 3 newly created datasets (IQTest, FunctionQA, PaperQA) spanning geometry, statistics, and scientific domains, with accuracy as the primary evaluation metric.
Combines visual understanding with mathematical reasoning across 6,141 curated examples from 28 existing datasets plus 3 newly created datasets (IQTest, FunctionQA, PaperQA), specifically designed to test compositional reasoning where models must both perceive complex visual mathematical representations and perform rigorous mathematical problem-solving — not just visual classification or simple arithmetic.
More comprehensive than MMVP or other vision-language benchmarks because it specifically targets mathematical reasoning requiring both visual perception and domain knowledge, with GPT-4V achieving only 49.9% accuracy vs human 60.3%, indicating genuine difficulty and room for model improvement.
leaderboard-based model performance ranking
Medium confidenceMaintains a public leaderboard ranking multimodal models by accuracy on the testmini subset (1,000 examples), with top performers including GPT-4V (49.9%), Bard (~34.8%), and Gemini Ultra. Leaderboard is hosted at mathvista.github.io and provides comparative performance metrics across 12+ evaluated foundation models, enabling researchers to track progress on mathematical reasoning benchmarks.
Provides public ranking of multimodal models specifically on mathematical reasoning tasks combining visual understanding with problem-solving, with transparent accuracy metrics and human baseline (60.3%) for context — enabling researchers to see exactly how far models fall short of human performance on compositional visual-mathematical reasoning.
More specialized than general vision-language leaderboards (like MMVP or LLaVA-Bench) because it focuses exclusively on mathematical reasoning where visual perception and domain knowledge must be composed, revealing that even best-in-class models (GPT-4V) significantly underperform humans.
dataset curation and visualization
Medium confidenceProvides access to 6,141 curated mathematical reasoning examples through Hugging Face dataset repository and an interactive visualization tool (🔮 Visualize) enabling exploration of examples by domain, difficulty, and source dataset. Dataset combines 28 existing multimodal datasets with 3 newly created datasets (IQTest, FunctionQA, PaperQA) covering geometry, statistics, and scientific figures, with structured metadata for filtering and analysis.
Combines 28 existing multimodal datasets with 3 newly created datasets (IQTest, FunctionQA, PaperQA) specifically designed for mathematical reasoning, with interactive visualization tool enabling exploration by domain and source — providing researchers transparent access to benchmark composition rather than black-box evaluation.
More transparent and explorable than closed benchmarks because it provides both raw dataset access via Hugging Face and interactive visualization tool, enabling researchers to understand dataset composition, identify potential biases, and analyze failure patterns rather than only seeing aggregate leaderboard scores.
program-of-thought augmentation for text-only models
Medium confidenceEnables text-only LLMs (like GPT-4) to perform mathematical reasoning on visual content by augmenting images with extracted captions and OCR text, then using the LLM to generate reasoning programs. This approach achieved measurable performance (PoT GPT-4 variant evaluated) by converting visual mathematical problems into text-based reasoning tasks that text-only models can process, bridging the gap between visual input and text-only model capabilities.
Bridges text-only and multimodal model capabilities by augmenting images with captions and OCR text, enabling text-only LLMs to perform mathematical reasoning on visual content through program-of-thought generation — a workaround for models without native visual understanding.
Enables use of text-only models on visual mathematical reasoning tasks, potentially at lower cost than multimodal APIs, though performance gap vs direct multimodal reasoning (GPT-4V) is not quantified in documentation.
self-verification and self-consistency enhancement
Medium confidenceExplores techniques to improve model performance on mathematical reasoning through self-verification (model checking its own answers) and self-consistency (sampling multiple reasoning paths and aggregating results). These enhancement techniques were tested on MathVista but specific performance improvements are not documented, representing potential approaches for improving accuracy beyond baseline model capabilities.
Applies self-verification and self-consistency techniques specifically to visual mathematical reasoning, where models must verify both visual interpretation and mathematical correctness — though specific implementation details and performance gains are not documented.
Represents potential accuracy improvements over baseline multimodal models through post-hoc verification and sampling strategies, though effectiveness is not quantified in available documentation.
multi-turn dialogue evaluation for mathematical reasoning
Medium confidenceEvaluates multimodal models through goal-directed human-AI dialogues where humans and models collaborate on mathematical problem-solving, testing whether models can engage in iterative reasoning and clarification. This evaluation variant goes beyond single-turn question-answering to assess interactive problem-solving capabilities, though specific dialogue protocols and performance metrics are not documented.
Extends single-turn question-answering evaluation to multi-turn goal-directed dialogues, testing whether models can engage in iterative mathematical reasoning and clarification — moving beyond static benchmark evaluation to interactive problem-solving.
More realistic than single-turn evaluation for educational and collaborative applications, though specific dialogue protocols and performance improvements are not documented in available materials.
domain-specific mathematical reasoning assessment
Medium confidenceEvaluates model performance across specific mathematical domains including geometry, statistics, and scientific figures, enabling domain-specific analysis of reasoning capabilities. The benchmark covers multiple mathematical domains through curated examples, though specific performance breakdowns by domain are not provided in documentation, limiting ability to identify domain-specific weaknesses.
Structures benchmark around specific mathematical domains (geometry, statistics, scientific figures) to enable domain-specific analysis, though actual per-domain performance metrics are not exposed in public leaderboard or documentation.
Enables more granular analysis than general mathematical reasoning benchmarks by organizing examples by domain, though performance breakdowns are not publicly available, limiting practical utility for domain-specific optimization.
newly created dataset variants for mathematical reasoning
Medium confidenceIntroduces three newly created datasets (IQTest, FunctionQA, PaperQA) specifically designed for mathematical reasoning evaluation, complementing 28 existing datasets. These new datasets target specific reasoning patterns: IQTest for visual pattern recognition and logical reasoning, FunctionQA for mathematical function understanding, and PaperQA for scientific figure interpretation — though specific dataset sizes, composition, and evaluation results are not documented.
Introduces three newly created datasets (IQTest, FunctionQA, PaperQA) targeting specific mathematical reasoning patterns beyond existing benchmarks, though specific dataset characteristics and performance results are not documented.
Extends benchmark coverage with novel datasets targeting reasoning patterns (pattern recognition, function understanding, scientific interpretation) not fully covered by existing multimodal benchmarks, though dataset details and performance analysis are not publicly available.
Capabilities are decomposed by AI analysis. Each maps to specific user intents and improves with match feedback.
Related Artifactssharing capabilities
Artifacts that share capabilities with MathVista, ranked by overlap. Discovered automatically through the match graph.
chinese-llm-benchmark
ReLE评测：中文AI大模型能力评测（持续更新）：目前已囊括359个大模型，覆盖chatgpt、gpt-5.2、o4-mini、谷歌gemini-3-pro、Claude-4.6、文心ERNIE-X1.1、ERNIE-5.0、qwen3-max、qwen3.5-plus、百川、讯飞星火、商汤senseChat等商用模型， 以及step3.5-flash、kimi-k2.5、ernie4.5、MiniMax-M2.5、deepseek-v3.2、Qwen3.5、llama4、智谱GLM-5、GLM-4.7、LongCat、gemma3、mistral等开源大模型。不仅提供排行榜，也提供规模超20
UGI-Leaderboard
UGI-Leaderboard — AI demo on HuggingFace
MATH Benchmark
12.5K competition math problems — AMC/AIME/Olympiad level, 7 subjects, standard math benchmark.
MMMU
Expert-level multimodal understanding across 30 subjects.
MATH
12.5K competition math problems across 7 subjects and 5 difficulty levels.
RealWorldQA
Real-world visual QA requiring spatial reasoning.
Best For
- ✓AI researchers developing multimodal models with mathematical reasoning capabilities
- ✓organizations evaluating foundation models (GPT-4V, Gemini, Bard) for mathematical problem-solving tasks
- ✓teams building educational AI systems that must interpret and reason about mathematical diagrams
- ✓AI researchers publishing multimodal model papers and needing competitive benchmarking
- ✓organizations selecting foundation models for mathematical reasoning applications
- ✓benchmark maintainers tracking field progress on compositional visual-mathematical reasoning
- ✓researchers analyzing model failure modes on mathematical reasoning tasks
- ✓teams fine-tuning multimodal models on domain-specific mathematical understanding
Known Limitations
- ⚠No statistical significance testing or confidence intervals provided — performance gaps reported as raw percentages only
- ⚠Evaluation methodology varies by model (GPT-4V manually evaluated via playground; others evaluated by original authors) — reproducibility concerns
- ⚠No explicit data contamination analysis — unknown whether foundation models' training data overlaps with MathVista source datasets
- ⚠Benchmark is static and does not measure interactive problem-solving, real-time reasoning, or robustness to adversarial inputs
- ⚠Evaluation protocol (zero-shot vs few-shot) not specified in documentation
- ⚠No failure mode taxonomy provided — only high-level statement that GPT-4V 'struggles with complex figures and rigorous reasoning'
Requirements
Input / Output
UnfragileRank
UnfragileRank is computed from adoption signals, documentation quality, ecosystem connectivity, match graph feedback, and freshness. No artifact can pay for a higher rank.
About
Mathematical reasoning benchmark combining visual understanding with mathematical problem-solving across geometry, statistics, and scientific figures, testing whether models can interpret visual math representations.
Categories
Alternatives to MathVista
Build high-quality LLM apps - from prototyping, testing to production deployment and monitoring.
Compare →Amplication brings order to the chaos of large-scale software development by creating Golden Paths for developers - streamlined workflows that drive consistency, enable high-quality code practices, simplify onboarding, and accelerate standardized delivery across teams.
Compare →Are you the builder of MathVista?
Claim this artifact to get a verified badge, access match analytics, see which intents users search for, and manage your listing.
Get the weekly brief
New tools, rising stars, and what's actually worth your time. No spam.
Data Sources
Looking for something else?
Search →