DermoBench - Reasoning: leaderboard
Morphology-grounded diagnostic reasoning on the 900-image dermatology core set of DermoBench (DermoGPT paper, Ru et al., 2026): the model writes a chain of thought ending in a diagnosis, once freely and once after a structured morphology report, and a Gemini-2.5-Pro judge scores each against dermatologist-revised reference reasoning from 0 to 100. The score is the mean of the two tasks; higher is better.
Metric: Reasoning Score (0-100; mean of two Gemini-2.5-Pro judge scores, free and morphology-grounded CoT). Source: arxiv.org. Saturation forecast: Around 2028. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 53.7 | #452 |
| 2 | QVQ-72B-Preview | 52.85 | #1173 |
| 3 | Qwen 3 VL 8B Instruct | 50.48 | #982 |
| 4 | Llama 3.2 90B | 50.38 | #1740 |
| 5 | Claude Sonnet 4.5 (Thinking) | 48.95 | #244 (Claude Sonnet 4.5) |
| 6 | GLM-4.5V | 48.73 | #739 |
| 7 | GPT-4o Mini | 47.24 | #1627 |
| 8 | Qwen 2.5 VL 72B Instruct | 45.05 | #845 |
| 9 | Llama 3.2 11B | 37.22 | #6434 |
| 10 | Nemotron Nano 12B v2 VL | 34.65 | #2248 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dermobench-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-12.