EviPathBench - Text-to-Image Retrieval: leaderboard
EviPathBench (2026) evaluates vision-language models on whole-slide pathology using diagnostic paths that pathologists annotated at three magnifications on TCGA slides; each question has four options. This board asks the model to pick the slide region that matches a written finding; accuracy in percent over all magnifications, chance is 25 and higher is better.
Metric: Accuracy (%) pooled over all questions (the All part of each cell) on Task 2: choosing the slide region image that matches a pathologist's finding among four images; EviPathBench: pathologist-authored diagnostic paths on 1,822 TCGA whole-slide images (16 organs), findings at 2.5x, 10x and 40x magnification; four options per question with distractors mined from the same organ and magnification; parsed choice scored by exact match; chance 25; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 67.74 | #164 |
| 2 | GPT-5.2 | 60.96 | #176 |
| 3 | Qwen 3.5 Flash | 60.05 | #435 |
| 4 | Qwen 2.5 VL 7B Instruct | 37 | #1833 |
| 5 | MedGemma-4B | 33 | #2444 |
| 6 | MedGemma 1.5 4B | 30.89 | #2562 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evipathbench-text-to-image-retrieval · How It Works · Data refreshed daily, snapshot 2026-11-13.