VisReason (Vision-Centric) - Localized Reasoning: leaderboard
Metric: Accuracy (%) on the Localized Reasoning category (perceptual reasoning: infer the target instance in an image and localize it with a bounding box) of VisReason, answers scored by format (regular-expression match, GPT-5-mini judge for open-ended answers, IoU above 0.5 for boxes), 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 January 2027. 22 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro (Preview) (Low) | 44.2 |
| 2 | GPT-5.2 (Thinking) | 24.6 |
| 3 | GPT-5 (Low) | 15 |
| 4 | GPT-5 Mini (High) | 12.5 |
| 5 | GPT-5 Mini (Low) | 12.5 |
| 6 | Qwen 3 VL 8B (Thinking) | 9.6 |
| 7 | GPT-5 Mini (Medium) | 7.5 |
| 8 | Qwen 3 VL 235B A22B (Thinking) | 7.1 |
| 9 | GPT-4o | 6.3 |
| 10 | GPT-5 Nano | 3.8 |
| 11 | O4 Mini | 3.8 |
| 12 | Qwen 2.5 VL 32B Instruct | 3.3 |
| 13 | InternVL3-14B | 1.3 |
| 14 | Qwen 3 VL 8B Instruct | 1.3 |
| 15 | Qwen 3 VL 32B (Thinking) | 1.3 |
Interactive version: theaggregate.ai/benchmark?slug=visreason-vision-centric-localized-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-07.