VLM-DeflectionBench - Parametric 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 no retrieved context; lower is better.

Metric: Hallucination rate (%): share of answers that are wrong rather than correct or a deflection, in the Parametric retrieval scenario (no retrieved context), judged by GPT-4o with the SimpleQA protocol; lower is better. Source: arxiv.org. Saturation forecast: Around 2030. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Mistral Small 3.126.6#1706
2Claude Opus 428.3#326
3Claude Sonnet 458.5#410
4Aria64.6#2172
5Gemini 2.5 Pro65.5#276
6Gemini 2.5 Flash70.8#463
7Ovis2-16B70.8#1594
8Keye-VL-1.5-8B73.1#1720
9Ovis2-34B73.8#1052
10GPT-574.8#178
11Pixtral-12B76#2553
12InternVL3-9B87.3#1728
13GLM-4.1V-9B (Thinking)87.4
14LLaVA-OneVision (VLM-DeflectionBench checkpoint unspecified)91.3
15MiniCPM-V-4.592.2#1441

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