CARV - Single-Step: leaderboard
Metric: Accuracy (%) on 500 CARV visual analogy tasks (controlled scenes of everyday items and furniture, edited with Gemini-2.5 Flash Image; the model writes a caption of the answer image, and a GPT-4o evaluator checks it against the ground-truth image), direct prompting: single-step analogy: infer the transformation of one image pair and apply it to the query image; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 81.2 | #131 |
| 2 | Gemini 2.5 Pro | 79.2 | #145 |
| 3 | Gemini 2.5 Flash | 75 | #237 |
| 4 | Qwen 3 VL 8B (Thinking) | 66.7 | |
| 5 | GPT-4o | 44.8 | #333 |
| 6 | Qwen 2.5 VL 32B Instruct | 39 | #443 |
| 7 | Qwen 3 VL 30B A3B Instruct | 33.4 | #365 |
| 8 | Qwen 3 VL 8B Instruct | 4.8 | #401 |
| 9 | InternVL3-14B | 4.6 | #494 |
| 10 | Qwen 2.5 VL 7B Instruct | 2.4 | #643 |
| 11 | InternVL3-8B | 1.8 | #606 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=carv-single-step · How It Works · Data refreshed daily, snapshot 2026-10-11.