MicroVQA: leaderboard

Each model answers all 1,042 multiple-choice questions about biological microscopy images (one or more images each), written by biology experts and hardened against text-only shortcuts; the score is accuracy (%), from the authors' own 2025 runs. Unlike MMMU-Pro or other college-level science sets, the questions are research-level and cover image understanding, hypothesis generation and experiment proposal.

Metric: Accuracy (%). Source: jmhb0.github.io. Saturation forecast: Estimated already saturated. 14 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1O1 (2024-12-17)52.8#314
2pixtral-large-241149.8#1284
3Qwen 2 VL 7B Instruct48.8#2594
4Grok 2 (1212)48.4#1082
5Qwen 2 VL 72B Instruct47.5#1382
6VILA1.5-40B47.5#3238
7Claude 3.5 Haiku (20241022)47.1#1589
8Gemini 1.5 Flash-8B (001)46.7#2311
9GPT-4o Mini (2024-07-18)46.2#1494
10Pixtral-12B-240945.6#2537
11Llama 3.1 Nemotron 70B Instruct44.2#1471
12Llama 3.2 90B Vision Instruct42.4#1791
13VILA1.5-13B41.8#7029
14Llama 3.2 11B Instruct30.3#5226

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=microvqa · How It Works · Data refreshed daily, snapshot 2026-11-17.