ReactBench (Hallucination) - Dense Counting (CoT): leaderboard

Accuracy in percent on ReactBench (2026) questions about crowded images that require counting, with a step-by-step reasoning prompt; an incorrect answer counts as a hallucination. Higher is better.

Metric: Accuracy (%) on Dense Counting (12,248 questions on 1,531 crowded images) with a step-by-step reasoning prompt, 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)64#577
2Qwen 3.5 27B61.8#380
3Qwen 3 VL 32B Instruct58.3#607
4InternVL3-8B57.8#1795
5Qwen 3.5 9B56.1#838
6Qwen 2.5 VL 72B Instruct54.9#849
7ReactBench MiMo-VL-7B (checkpoint unspecified)53.7
8ReactBench LLaVA-OV-7B (checkpoint unspecified)48.4
9Qwen 2.5 VL 7B Instruct47.8#1919
10Qwen 3 VL 4B46.6#1163
11InternVL2.5-26B46.4#2174
12InternVL2.5-8B43.2#2393
13ReactBench LLaVA-v1.6-7B (checkpoint unspecified)33.1

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

Interactive version: theaggregate.ai/benchmark?slug=reactbench-hallucination-dense-counting-cot · How It Works · Data refreshed daily, snapshot 2026-11-22.