VLM-DeflectionBench - Oracle 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 only gold evidence; lower is better.

Metric: Hallucination rate (%): share of answers that are wrong rather than correct or a deflection, in the Oracle retrieval scenario (gold evidence only), 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 49.2#326
2Claude Sonnet 49.9#410
3Mistral Small 3.110.3#1706
4GPT-512.6#178
5Gemini 2.5 Flash13.3#463
6GLM-4.1V-9B (Thinking)13.3
7Gemini 2.5 Pro13.9#276
8InternVL3-38B22.6#939
9Gemma 3 27B24.4#1663
10VL-Rethinker-72B24.5#1244
11MiniCPM-V-4.525.8#1441
12Keye-VL-1.5-8B27.3#1720
13Aria27.5#2172
14Ovis2-34B27.8#1052
15Pixtral-12B30.5#2553

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