Forgery Reasoning on TFR CTM Cross-Method
51.2Reasoning ScoreTextShield-R1
Evaluation results
| Method | Links | |
|---|---|---|
| TextShield-R1Fine-tuning=Full training set images2026.02 | 51.2 | |
| SIDA*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 36.3 | |
| Qwen2.5-VL-7BFine-tuning=Full training set images2026.02 | 36.2 | |
| FakeShield*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 36.2 | |
| InternVL3-8BFine-tuning=Full training set images2026.02 | 32.3 | |
| MiniCPM_V_2.6Fine-tuning=Full training set images2026.02 | 27.3 | |
| InternVL3-2BFine-tuning=Full training set images2026.02 | 25 | |
| Qwen2.5-VL-3BFine-tuning=Full training set images2026.02 | 24.6 | |
| SIDAFine-tuning=Full training set images2026.02 | 21.5 | |
| FakeShieldFine-tuning=Full training set images2026.02 | 21.4 | |
| GPT4oFine-tuning=None2026.02 | 9.7 | |
| InternVL3-8BFine-tuning=None2026.02 | 8.8 | |
| Qwen2.5-VL-3BFine-tuning=None2026.02 | 5.6 | |
| Qwen2.5-VL-7BFine-tuning=None2026.02 | 4.5 | |
| InternVL3-2BFine-tuning=None2026.02 | 4 | |
| MiniCPM_V_2.6Fine-tuning=None2026.02 | 2.5 |