Defect Grounding (Misalignment) on SDG-30K (test)
71.2DetType F1Human
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HumanNote=localization upper bound from 16 independent re-annotators2026.06 | 71.2 | 81.8 | 56.3 | 40.9 | — | — | |
| Gemini 3 ProMode=zero-shot2026.06 | 69.6 | 67.2 | 46.4 | 30.7 | 89.1 | 85.1 | |
| SDG (GRPO)Backbone=Qwen3-VL-4B-Instruct, Training=BoxFlow-GRPO2026.06 | 67.5 | 73.2 | 51.1 | 38.7 | 88.8 | 89.3 | |
| GPT-5.4Mode=zero-shot2026.06 | 65.4 | 51.3 | 45.7 | 29.2 | 83.7 | 82 | |
| SDG (SFT)Backbone=Qwen3-VL-4B-Instruct, Training=Supervised Fine-Tuning2026.06 | 63.6 | 79.9 | 49.9 | 37.6 | 89.3 | 89.2 |