/
/
Multimodal Math QSA

Available Now · Sample Available

Multimodal Math QSA

Graduate math with plots and diagrams, built where models lose 25+ points going multimodal.

Volume

0

Volume

Scalable Volume

Coverage

Grad, PhD, Expert, Math, Mathematics

Model Benchmark
0.0 %

Frontier Model · Pass at k

Targets

MMMU (Math), MathVista, OlympiadBench (MM math), MMMU-Pro, MathVision

Modality & Delivery

ImageText→Text

Formats

Excel (.xlsx)Images

What You Receive

Graduate math problems paired with plots, spectra, and diagrams as QSA triples. Addresses the text-to-multimodal drop on OlympiadBench MM, where top models lose more than 25 points versus text-only.

Coverage

Plots, spectra, and diagrams — each problem paired with a corresponding scientific image.

Package Structure

Differentiators

  • Structured question, worked solution, and final answer record format.
  • Pairs domain questions with task-relevant visual evidence.

Record Schema

FieldDescription
answerThe concise final answer associated with the question.
image_refReference to the image associated with the record, when supplied.
questionThe question or task presented in the dataset record.
solutionThe worked solution or supporting reasoning for the question.

Results

ModelMetricResult
Frontier ModelPass@117.4%
Claude Opus 5Pass@140.4%
Claude Opus 5Pass@346.8%

Methodology & Quality

Methodology

Questions are expert-authored, then split across an SME pool for complexity validation before benchmarking.

SME involvement

Domain-expert curated with up to L2 review

Stressors

Combined visual interpretation and domain reasoning.

Design criteria

  1. Keep the question, supporting solution, and final answer as distinct record components.
  2. Require visual extraction together with domain-specific reasoning.

Build assumptions

Quality approach

Known limitations

Customization & Scalability

Custom sets available. Additional volume in progress