SADL: leaderboard
SADL (2026) tests subject-aware distractor localization: given a photograph and its intended subject, a vision-language model decides for every candidate object whether it is a distractor to remove, an essential object to keep despite looking like one, or neither. The score is the mean recall over the three classes, in percent; higher is better.
Metric: Average recall (%, AR): mean of the per-class recalls for distractor, excluded hard negative and non-distractor labels under guided classification (the model sees the image, the subject description and the enumerated candidates and labels each one), over the 14,617 annotated candidates of SADL (1,800 subject-aware cases on 1,000 photographs, including 1,938 hard negatives), zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 63.8 | #137 |
| 2 | GPT-4.1 | 62.3 | #495 |
| 3 | GPT-4o | 58 | #741 |
| 4 | Qwen 3 VL 235B A22B | 54 | #479 |
| 5 | Molmo2-8B | 46.9 | #1483 |
| 6 | Llama 4 Scout | 43.7 | #1776 |
| 7 | Pixtral-12B | 40.7 | #2476 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sadl · How It Works · Data refreshed daily, snapshot 2026-11-13.