Defect Grounding (Artifact) on SDG-30K (test)
79.2DetType F1Human
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HumanNote=localization upper bound from 16 independent re-annotators2026.06 | 79.2 | 77.4 | 46.2 | 27.8 | — | — | |
| SDG (SFT)Backbone=Qwen3-VL-4B-Instruct, Training=Supervised Fine-Tuning2026.06 | 77.6 | 69.7 | 40.2 | 25.5 | 90.4 | 88.3 | |
| SDG (GRPO)Backbone=Qwen3-VL-4B-Instruct, Training=BoxFlow-GRPO2026.06 | 77.2 | 56 | 40.4 | 26.3 | 90.4 | 88.7 | |
| GPT-5.4Mode=zero-shot2026.06 | 73.6 | 11.8 | 23.7 | 3.5 | 86.8 | 81.2 | |
| Gemini 3 ProMode=zero-shot2026.06 | 73.3 | 7.6 | 33.7 | 20 | 88.5 | 82.4 |