41Evidence
activeWhat it measures
Ablesen und Rechnen an Diagrammen.
What it does not measure
Freie Bildbeschreibung, Video.
How the score is composed
Currency82
Discrimination18
Headroom53
Contamination safety15
Gaming resistance55
Methodological rigor50
Reproducibility60
Details
- Benchmaxxing risk
- medium
- Difficulty
- low
- Public test set
- yes
- Human-verified
- no
- LLM-judged
- no
- Letzter Messwert
- 2026-07-25
- Skala
- 0–100 % correct (höher besser)
Best measured models
- 1Mage-VLMicrosoftvendor number83.96 % correct
- 2Phi-4-multimodalMicrosoftvendor number81.8 % correct