MMOOC - Visual False Premises (Open-Ended VQA): leaderboard

Metric: Out-of-context score (%; mean of the refusal rate, the share of the Visual False Premises Open-Ended VQA questions, which the image and context cannot answer, that the model correctly declines, and refusal rationality, the mean 0-1 judge score of its abstention reasoning in steps of 0.25; judge scores averaged over GPT-5.6, Claude Opus 5 and DeepSeek-V4-Pro). Source: arxiv.org. Saturation forecast: Around December 2027. 18 models tracked.

Top models

#ModelScore
1Gemma 4 26B86.5
2Qwen 3.5 122B A10B85.25
3Qwen 3.5 27B84.25
4Gemma 4 31B83
5Llama 4 Maverick78.25
6O361
7Ministral 3 8B59.25
8GPT-4o58.25
9Ministral 3 14B50.75
10O143.5
11InternVL3-8B43
12Claude Opus 4.635
13Gemini 3.1 Pro (Preview)6.5

Interactive version: theaggregate.ai/benchmark?slug=mmooc-visual-false-premises-open-ended-vqa · How It Works · Data refreshed daily, snapshot 2026-09-29.