DO-Bench: leaderboard

Object-existence verification under controlled interventions on DO-Bench (2026): for each of 124 expert-verified scenes a vision-language model answers ten yes/no questions about an anomalous object that is present and an expected object that was removed, with misleading textual context of increasing strength and zoomed-in views. Accuracy in percent; higher is better.

Metric: Accuracy (%) of binary object-existence answers over all ten queries per scene (60 percent of answers are yes) on DO-Bench, 124 scenes with ten controlled yes/no object-existence queries each (1,240 queries: a present-but-anomalous object under four prior strengths and two zoomed views, an absent-but-expected object removed by local inpainting under four prior strengths), expert-verified images, temperature 0 or the mean of three API runs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Flash (Preview)88.53#152
2GPT-5.285.66#185
3Claude Opus 4.685.22#113
4Qwen 2.5 VL 32B Instruct74.4#1062
5InternVL2.5-78B72.8#1031
6Qwen 2.5 VL 72B Instruct72.31#841
7InternVL2.5-8B69.92#2370
8InternVL2.5-38B68.15#1373
9Qwen 2.5 VL 7B Instruct65.81#1902
10Qwen 2.5 VL 3B Instruct64.11#2713
11InternVL2.5-4B56.29#2549
12InternVL2.5-2B47.58#6468
13Qwen-VL-7B39.76
14InstructBLIP-Vicuna-13B38.51#8801
15LLaVA-v1.5-7B37.74#7630

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

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