PBSBench - Cell Fill-in-the-Blank: leaderboard

PBSBench (2026) tests vision-language models on peripheral blood smear hematopathology, with questions about single white blood cell crops. Share of fill-in-the-blank answers that contain the reference keyword, in percent, on in-domain cells; higher is better.

Metric: Partial match rate (answer contains the ground-truth keyword after case and whitespace normalization) (%, times 100) on PBSBench cell-level in-domain fill-in-the-blank questions about single white-blood-cell crops from peripheral blood smears (916 questions); questions generated from GPT-4o morphology captions of hematopathologist-labeled cells; greedy decoding, model chat templates; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.554#261
2GPT-4o48#751
3Gemini 2.5 Pro41#276
4InternVL3.5-8B38#1604
5PathGen-LLaVA31#1406
6Qwen 2 VL 7B Instruct27#2573
7BLIP2-OPT-6.7B21
8Qwen 2.5 VL 7B Instruct17#1902
9Idefics3-8B-Llama317#5542
10Quilt-LLaVA16#4973
11DeepSeek-VL-7B15#7399
12llava-v1.6-mistral-7B10#3999
13LLaVA-Med10
14LLaVA-1.6-Vicuna-7B10
15Med-Flamingo6

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

Interactive version: theaggregate.ai/benchmark?slug=pbsbench-cell-fill-in-the-blank · How It Works · Data refreshed daily, snapshot 2026-11-21.