Forgery Reasoning on TFR CIS Cross-Image-domain
56.5Reasoning ScoreTextShield-R1
Evaluation results
| Method | Links | |
|---|---|---|
| TextShield-R1Fine-tuning=Full training set images2026.02 | 56.5 | |
| Qwen2.5-VL-7BFine-tuning=Full training set images2026.02 | 35.7 | |
| SIDA*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 35.7 | |
| FakeShield*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 35.6 | |
| InternVL3-8BFine-tuning=Full training set images2026.02 | 33.8 | |
| Qwen2.5-VL-3BFine-tuning=Full training set images2026.02 | 33.6 | |
| MiniCPM_V_2.6Fine-tuning=Full training set images2026.02 | 32.3 | |
| InternVL3-2BFine-tuning=Full training set images2026.02 | 32 | |
| SIDAFine-tuning=Full training set images2026.02 | 29.4 | |
| FakeShieldFine-tuning=Full training set images2026.02 | 29.2 | |
| GPT4oFine-tuning=None2026.02 | 24.3 | |
| InternVL3-8BFine-tuning=None2026.02 | 20.3 | |
| Qwen2.5-VL-7BFine-tuning=None2026.02 | 17.6 | |
| Qwen2.5-VL-3BFine-tuning=None2026.02 | 15.3 | |
| InternVL3-2BFine-tuning=None2026.02 | 9.9 | |
| MiniCPM_V_2.6Fine-tuning=None2026.02 | 3.2 |