ReactBench (Hallucination): leaderboard

ReactBench (2026) probes multimodal hallucination with about 50,000 questions on images edited to contradict language priors. React-Score in percent weights short-answer, multiple-choice and yes-or-no accuracy, averages standard and chain-of-thought prompting and weights the four tasks; higher is better.

Metric: React-Score (%): question-type-weighted accuracy (short answer 0.5, multiple choice and yes/no 0.25 each), averaged over standard and chain-of-thought prompting (thinking-only models: chain-of-thought) and weighted over the four tasks (0.3, 0.2, 0.2, 0.3), hallucination-inducing questions (multiple choice, short answer, yes/no) about images edited by Qwen-Image-Edit-2511; an incorrect answer counts as a hallucination; temperature 0.2, at most 1,024 tokens, LLM answer matching; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 32B (Thinking)65.3#573
2Qwen 3 VL 32B Instruct61.9#606
3Qwen 3.5 27B59.5#375
4ReactBench MiMo-VL-7B (checkpoint unspecified)57.5
5InternVL3-8B56.5#1782
6Qwen 2.5 VL 72B Instruct54.2#841
7Qwen 3 VL 4B54#1151
8Qwen 3.5 9B52.4#828
9Qwen 2.5 VL 7B Instruct49.5#1902
10InternVL2.5-26B47#2156
11ReactBench LLaVA-OV-7B (checkpoint unspecified)42
12InternVL2.5-8B41.5#2370
13ReactBench LLaVA-v1.6-7B (checkpoint unspecified)34.5

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

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