DiffCap-Bench - Hallucination Rate: leaderboard

DiffCap-Bench (2026) image difference captioning, scored for faithfulness: the number of hallucinated differences a model describes per 100 ground-truth differences between the two images, as judged by Gemini 2.5 Pro, averaged over 1,075 image pairs. Lower is better, and the rate can exceed 100.

Metric: Hallucination rate (hallucinated difference items per 100 ground-truth differences, sample-level mean; can exceed 100) on the 1,075 image pairs of DiffCap-Bench (6,713 human-verified difference items), one fixed captioning prompt, open models at temperature 0.1 and proprietary models at their defaults, a Gemini 2.5 Pro judge matching each described difference against the key difference list; lower is better. Source: arxiv.org. Saturation forecast: Around June 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro2.6#145
2Grok 4.2010.5#301
3GPT-5.212#185
4Qwen 3 VL Plus18.3#512
5Kimi K2.5Thinking18.9#266
6Qwen 3 VL 8B (Thinking)20.2#910
7Seed1.6-VLThinking20.9#385
8Step3 VL 10B22.8#1080
9Qwen 3 VL 32B (Thinking)25.8#573
10Qwen 3 VL 8B Instruct26.1#1029
11Qwen 3 VL 32B Instruct31.3#606
12MiMo-VL-7B-RL-250834#823
13InternVL3.5-38B51.8#1021
14InternVL3.5-8B62.6#1604

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

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