VLM-DeflectionBench - Realistic Hallucination Rate: 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 the share of hallucinated answers, wrong rather than correct or deflected, in percent with gold and distractor evidence mixed; lower is better.

Metric: Hallucination rate (%): share of answers that are wrong rather than correct or a deflection, in the Realistic retrieval scenario (two gold and two negative contexts mixed), judged by GPT-4o with the SimpleQA protocol; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 48.5#326
2Claude Sonnet 410.5#410
3Mistral Small 3.114.9#1706
4Gemini 2.5 Pro20.5#276
5Gemini 2.5 Flash20.7#463
6GLM-4.1V-9B (Thinking)22.4
7GPT-525.5#178
8InternVL3-38B33.9#939
9Gemma 3 27B34.5#1663
10Ovis2-16B35.2#1594
11Aria39.1#2172
12MiniCPM-V-4.539.4#1441
13Keye-VL-1.5-8B40.5#1720
14VL-Rethinker-72B40.9#1244
15Ovis2-34B43.3#1052

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-hallucination-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.