Medicine

SMMILE

SMMILE: an expert-driven benchmark for multimodal medical in-context learning. 111 problems (517 image-question-answer triplets) across 6 specialties and 13 imaging modalities; multimodal LLMs evaluated zero-shot vs with in-context examples on closed-ended MCQA and open-ended (exact-match / LLM-as-a-judge) accuracy. SMMILE++ permutes the in-context demonstrations into 1038 problems.

111items
16subjects
CC-BY-4.0license
medicinedomain
textmodality
imagemodality
item-level responses released
Saturation status: Unknown

Response matrix

Fit to width. Hover for subject & item; click a cell for details.

SMMILE response matrix: AI models (rows) against items (columns)
Correct (1)Incorrect (0)Unobserved

Scale: 1 = correct · 0 = incorrect

Subjects

  1. 1Qwen2.5-VL-72B0.3882
  2. 2GPT-4o0.341
  3. 3Qwen2.5-VL-7B0.3298
  4. 4Qwen2.5-VL-32B0.3169
  5. 5Claude 3.7 Sonnet0.2898
  6. 6Llama-3.2-Vision-90B0.2734
  7. 7MedGemma-4B-Multimodal0.2692
  8. 8Qwen2.5-VL-3B0.2608
  9. 9MedVLM-R10.2493
  10. 10LLaVA-OneVision-7B0.2411
  11. 11LLaVA-v1.5-7B0.2132
  12. 12LLaVA-v1.5-13B0.2092
  13. 13LLaVA-OneVision-0.5B0.2055
  14. 14LLaVA-NeXT-7B0.1765
  15. 15Llama-3.3-70B-Instruct (text-only)0.1497
  16. 16LLaVA-Med-v1.5-Mistral-7B0.0503