benchmarks evals
Ninety-Four Percent Recognition Fell to Fifty-Six Percent Culture
Models knew dish names better than they could connect cooking images to Chinese regional cuisines.
Published Updated Story ID: mp-2026-09-05-017
Summary
Models knew dish names better than they could connect cooking images to Chinese regional cuisines.
CulturalMenuBench contains 4,870 items in ten languages across eighteen regions. Twelve models that topped 94% on standard multiple choice reached at most 56% on regional attribution, and names beat images by 7–18 points.
Why it matters
Models knew dish names better than they could connect cooking images to Chinese regional cuisines.
Limits and context
No additional limitation was separately recorded.
Key claims
Models knew dish names better than they could connect cooking images to Chinese regional cuisines.
Evidence: source-2026-09-05-019
Sources
- arXiv preprint 2609.03526arXiv · primary research
Corrections
No corrections have been recorded for this story.