Skip to content
Multimodal AI Systems

Systems & Evaluation

Evaluating Multimodal Models

10 min+100 XPStep 1/10

Benchmarks saturate, failures don't

VQA, MMMU, MathVista — the leaderboard numbers climb while production failures stay constant. Three structural reasons:

  • Contamination — popular benchmarks leak into pretraining corpora; the model remembers, doesn't reason.
  • Distribution mismatch — benchmarks test 'can it' on academic slices; production tests 'does it' on your documents, your lighting, your users' phrasing.
  • The metric hides the failure shape — 78% on VQA doesn't say the 22% is all medical images, or all counting, or all the questions that matter.

An eval that doesn't look like your traffic is marketing.

insight

Benchmark scores measure the benchmark. The eval that predicts production is built from production failures — harvested, labeled, and rerun on every model change.