VRIQ (Natural) - Odd One Out: leaderboard
Metric: Accuracy (%) on the 100 natural-domain VRIQ odd-one-out puzzles (the same reasoning categories over real-world objects and scenes, seeded by annotators and generated with GPT-5.2-Thinking or sourced, human validated; four options); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 60 | #131 |
| 2 | GPT-4o | 57 | #333 |
| 3 | Qwen 3 VL 32B (Thinking) | 54 | #287 (Qwen 3 VL 32B) |
| 4 | Gemini 2.5 Pro | 50 | #145 |
| 5 | Qwen 2.5 VL 7B Instruct | 50 | #643 |
| 6 | GPT-4o Mini | 48 | #588 |
| 7 | GPT-5 Mini | 48 | #176 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=vriq-natural-odd-one-out · How It Works · Data refreshed daily, snapshot 2026-10-11.