MicroVQA - Hypothesis Generation: leaderboard

Accuracy (%) on the MicroVQA questions of the hypothesis generation task: proposing the most likely biological explanation for what a microscopy experiment shows. Each model answers every question of this task once with the same prompt as the headline board.

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)50.2#314
2pixtral-large-241149.5#1284
3VILA1.5-40B47.9#3238
4VILA1.5-13B47.5#7029
5Grok 2 (1212)46.4#1082
6Qwen 2 VL 72B Instruct45.7#1382
7Pixtral-12B-240944.8#2537
8Claude 3.5 Haiku (20241022)43.8#1589
9GPT-4o Mini (2024-07-18)43.6#1494
10Gemini 1.5 Flash-8B (001)43.6#2311
11Qwen 2 VL 7B Instruct43.3#2594
12Llama 3.1 Nemotron 70B Instruct43.3#1471
13Llama 3.2 90B Vision Instruct42.1#1791
14Llama 3.2 11B Instruct29.3#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-hypothesis-generation · How It Works · Data refreshed daily, snapshot 2026-11-17.