VisReason (Vision-Centric) - Bounding-Box: leaderboard

VisReason Bounding-Box scores multimodal LLMs on the 140 bounding-box questions of VisReason (2026), a benchmark of vision-centric reasoning in everyday images, drawn from its Localized Reasoning and Spot the Difference categories. The score is accuracy in percent, higher is better.

Metric: Accuracy (%) on the 140 bounding-box questions (boxes Hungarian-matched to the ground truth at IoU above 0.5, partial credit per question) of VisReason, drawn from the Localized Reasoning and Spot the Difference categories, zero-shot with chain-of-thought prompt templates, one per answer format; rows not shaded gray in the paper use explicit reasoning (GPT-5 and Gemini 3 Pro at low reasoning effort); higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro (Preview)Low24.5#123
2GPT-5.2Thinking14.7#190
3GPT-5Low8.9#185
4GPT-5 MiniLow7.4#373
5Qwen 3 VL 8B (Thinking)5.2#924
6Qwen 3 VL 235B A22B (Thinking)4.4#462
7GPT-4o3.6#754
8O4 Mini3.3#372
9Qwen 2.5 VL 32B Instruct2.4#1065
10GPT-5 Nano2#912
11Qwen 3 VL 2B (Thinking)1.4#2552
12Qwen 3 VL 32B (Thinking)1.3#577
13Qwen 3 VL 8B Instruct1.2#1043
14InternVL3-14B0.7#1227
15Gemini 2.5 Flash0.3#466

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

Interactive version: theaggregate.ai/benchmark?slug=visreason-vision-centric-bounding-box · How It Works · Data refreshed daily, snapshot 2026-11-22.