VLM-DeflectionBench - Realistic Accuracy: leaderboard

VLM-DeflectionBench (2026) tests knowledge-based visual question answering where the right response may be to decline when retrieved evidence is missing or misleading; a GPT-4o judge grades each answer correct, deflected or hallucinated. This board is accuracy in percent with gold and distractor evidence mixed; higher is better.

Metric: Accuracy (%) of knowledge-based visual question answering in the Realistic retrieval scenario (two gold and two negative contexts mixed), judged by GPT-4o with the SimpleQA protocol; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Ovis2-16B61.1#1594
2GPT-559.5#178
3Gemini 2.5 Pro51#276
4InternVL3-38B50.3#939
5Ovis2-34B49.1#1052
6Keye-VL-1.5-8B47.6#1720
7VL-Rethinker-72B47.2#1244
8Gemini 2.5 Flash47#463
9Qwen 2.5 VL 32B Instruct45.2#1062
10InternVL3-9B45.1#1728
11MiniCPM-V-4.544.3#1441
12Gemma 3 27B42.9#1663
13Pixtral-12B42.6#2553
14Aria39.5#2172
15GLM-4.1V-9B (Thinking)37.4

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

Interactive version: theaggregate.ai/benchmark?slug=vlm-deflectionbench-realistic-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-21.