OddGridBench - Position: leaderboard
Metric: Accuracy (%) on the 200 images in which the odd item differs only in its position within its cell; OddGridBench synthetic grid images (5 to 9 rows and columns of identical icons with one odd item); the model names the odd item's row and column, scored by exact match; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 39 | #276 |
| 2 | Qwen 3 VL 4B Instruct | 27.5 | #506 |
| 3 | Qwen 3 VL 30B A3B Instruct | 25 | #365 |
| 4 | Qwen 3 VL 8B Instruct | 19 | #401 |
| 5 | Gemini 2.5 Pro | 6.5 | #145 |
| 6 | Gemini 2.5 Flash | 5 | #237 |
| 7 | Gemini 2.0 Flash | 5 | #331 |
| 8 | GPT-5 (Low) | 5 | #91 (GPT-5) |
| 9 | Qwen 2.5 VL 72B Instruct | 2.5 | #364 |
| 10 | Qwen 2.5 VL 7B Instruct | 0 | #643 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=oddgridbench-position · How It Works · Data refreshed daily, snapshot 2026-10-11.