Med-R2: leaderboard

Med-R2 (2026) tests whether medical vision-language models reason from visual evidence along the clinical workflow (image quality, anatomy, lesion, then a joint question chain) and resist misleading positive or negative cues injected into the questions. The score is the mean accuracy (%) over its four tasks; higher is better.

Metric: Overall average accuracy (%) over the Image Quality, Anatomy, Lesion and Joint QA tasks of Med-R2, zero-shot with system prompts; each task mixes equal numbers of positive, negative and neutral adversarial question forms; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.2Thinking67.4#185
2InternVL3.5-8B61.8#1604
3Hulu-Med-14B61.8#1607
4Qwen 3 VL 8B61#917
5Qwen 3 VL 235B A22B60.6#485
6Lingshu-7B56.3#1635
7HuaTuoGPT-Vision-34B55.6#1828
8GPT-4o55.4#751
9Hulu-Med-4B54.5#2392
10Qwen 2.5 VL 7B Instruct53.3#1902
11Janus-Pro-7B52#6857
12HuluMed-7B51.2#2049
13HuatuoGPT-Vision-7B50.5#2162
14MedGemma-4B34.6#2557

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=med-r2 · How It Works · Data refreshed daily, snapshot 2026-11-21.