MathVista

Evaluates multimodal models' ability to interpret visual mathematical representations (geometry diagrams, statistical charts, scientific figures) and perform compositional reasoning combining visual perception with mathematical problem-solving. The benchmark uses a curated dataset of 6,141 examples sourced from 28 existing multimodal datasets plus 3 newly created datasets (IQTest, FunctionQA, PaperQA), with questions presented in multiple-choice and free-form generation formats. Scoring uses exact-match accuracy on the testmini subset (1,000 examples) exposed via a public leaderboard.

Aggregates and curates 6,141 mathematical reasoning examples from 28 existing multimodal datasets and 3 newly created datasets (IQTest, FunctionQA, PaperQA) with standardized question-answer pairs. The curation process involves selecting examples that require compositional visual-mathematical reasoning, extracting or generating questions, and providing auxiliary annotations (OCR text, image captions) for text-only model baselines. Dataset is hosted on Hugging Face and includes a visualization tool for exploring examples by mathematical domain and visual context type.

MathVista is released as open-source with dataset available on Hugging Face and code available on GitHub (links provided), enabling researchers to download, analyze, and build upon the benchmark. Open-source release facilitates reproducibility, enables community contributions, and lowers barriers to adoption. Researchers can access raw data, evaluation code, and visualization tools without proprietary restrictions.

Aggregates examples from 28 existing multimodal datasets plus 3 newly created datasets (IQTest, FunctionQA, PaperQA) into a unified benchmark with standardized question-answer format and consistent evaluation protocol. This aggregation approach combines diverse sources (existing datasets covering various visual-mathematical domains plus new datasets targeting specific reasoning types) into a single coherent benchmark. Standardization enables fair comparison across models and reduces bias from any single source's annotation style or problem distribution.

Maintains a public leaderboard (testmini subset, 1,000 examples) tracking multimodal model performance on mathematical reasoning tasks with exact-match accuracy as the primary metric. The leaderboard displays rankings of models (GPT-4V at 49.9%, Gemini Ultra, Bard at ~34.8%, and others) and enables comparison of model capabilities across visual mathematical domains. Leaderboard is updated as new model submissions are evaluated, providing a living benchmark of progress in multimodal mathematical reasoning.

Provides OCR-extracted text and image captions for each visual example, enabling evaluation of text-only models (e.g., GPT-4 without vision) as baselines on visual mathematical reasoning tasks. This allows researchers to isolate the contribution of visual understanding vs. text-based reasoning by comparing text-only model performance (using OCR + captions) against multimodal model performance (using images). The auxiliary annotations reveal whether models can solve mathematical problems from text descriptions alone or require direct visual interpretation.

Enables analysis of model performance across distinct mathematical domains (geometry, statistics, scientific figures) and visual context types, revealing which reasoning types and visual representations challenge models most. The benchmark structure supports stratified evaluation where accuracy can be computed separately for each domain, allowing researchers to identify capability gaps (e.g., models may excel at statistics but struggle with geometry). Documentation mentions performance varies significantly across mathematical reasoning types and visual context types, though specific breakdowns are not provided in public leaderboard.

Provides a web-based visualization tool (🔮 Visualize) accessible at https://mathvista.github.io for exploring individual benchmark examples, filtering by mathematical domain and visual context type, and understanding benchmark composition. The tool enables researchers to browse examples, examine model predictions vs. ground truth, and identify patterns in model failures or benchmark difficulty. This interactive exploration complements the leaderboard and dataset documentation by making benchmark content directly inspectable.

Evaluates models' ability to perform compositional reasoning where visual perception and mathematical logic must be jointly applied to solve problems. Unlike benchmarks that test visual understanding (image captioning) or mathematical reasoning (text-only math problems) separately, MathVista requires models to interpret visual representations (diagrams, charts, figures) AND apply mathematical reasoning to derive correct answers. This compositional requirement is enforced through benchmark design where examples cannot be solved from visual content alone or text description alone, but require both modalities.

Tests models' ability to accurately interpret fine-grained details in complex mathematical figures including geometry diagrams with precise spatial relationships, statistical charts with multiple data series and annotations, and scientific figures with technical notation and spatial complexity. The benchmark includes examples from research papers and technical documents where visual interpretation requires understanding of mathematical conventions (axis labels, legend symbols, geometric properties, etc.). This capability goes beyond general image understanding to require domain-specific visual literacy in mathematical representations.

Establishes human performance baseline (~60.3% accuracy) on the benchmark, enabling quantification of how far current SOTA models fall short of human-level performance. The 10.4 percentage point gap between GPT-4V (49.9%) and human performance demonstrates that even best-in-class multimodal models struggle with compositional visual-mathematical reasoning. This baseline provides a clear target for model improvement and context for interpreting model performance (e.g., whether 49.9% accuracy is near-ceiling or far from human-level).

MathVista was accepted as an oral presentation at ICLR 2024 (85 out of 7,304 submissions, 1.2% acceptance rate), indicating peer-reviewed validation of the benchmark's design, methodology, and significance. The publication includes detailed methodology, results, and analysis reviewed by top-tier conference reviewers. This peer-reviewed validation provides confidence that the benchmark is well-designed and addresses important research questions, distinguishing it from non-peer-reviewed benchmarks or datasets.