VLM-DeflectionBench - Adversarial 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 misleading evidence; lower is better.

Metric: Hallucination rate (%): share of answers that are wrong rather than correct or a deflection, in the Adversarial retrieval scenario (negative contexts 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 411.1#332
2Claude Sonnet 412.1#414
3Mistral Small 3.115.6#1734
4Gemini 2.5 Pro22.2#280
5GLM-4.1V-9B (Thinking)22.5
6Gemini 2.5 Flash24.1#466
7GPT-534.7#185
8Gemma 3 27B39.4#1672
9InternVL3-38B42.3#948
10MiniCPM-V-4.551.9#1451
11Keye-VL-1.5-8B53.8#1729
12VL-Rethinker-72B55.4#1252
13Aria56.4#2190
14Pixtral-12B56.8#2611
15Ovis2-34B58.1#1061

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=vlm-deflectionbench-adversarial-hallucination-rate · How It Works · Data refreshed daily, snapshot 2026-11-22.