RetinaGroundBench - VQA Recall: leaderboard
RetinaGroundBench: 815 OCT images and 13 biomarker classes from OCTDL and OLIVES, with macro-averaged metrics. Grounding uses union AggIoU and thresholds 0.1/0.3. VQA asks 13 presence questions per image; balanced accuracy averages sensitivity and specificity. RetGrounder is a separate LoRA-trained Qwen3-VL-8B-Instruct. Ambiguous Instruct, Thinking and IT checkpoints retain source-qualified identities pending the supplemental checkpoint table. Recall (%).
Metric: Recall (%). Source: arxiv.org. Saturation forecast: Not forecast. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-14B | 76 | #509 |
| 2 | InternVL3-38B | 75.1 | #413 |
| 3 | InternVL3-8B | 74.9 | #627 |
| 4 | Qwen 3.5 27B | 69.6 | #191 |
| 5 | Qwen 3.5 4B | 68.4 | #495 |
| 6 | Qwen 3.5 9B | 64.9 | #390 |
| 7 | Qwen 3.6 35B A3B | 64.9 | #232 |
| 8 | Qwen 3.5 35B A3B | 59.1 | #261 |
| 9 | Qwen 3.6 27B | 56.1 | #176 |
| 10 | Qwen 3.5 2B | 55.7 | #1157 |
| 11 | Qwen 3 VL 8B Instruct | 43.5 | #419 |
| 12 | Qwen 3.5 0.8B | 31.6 | #1428 |
No result here: #5 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=retinagroundbench-vqa-recall · How It Works · Data refreshed daily, snapshot 2026-10-25.