Forgery Reasoning on TFR Cross-Language
46.2Avg Reasoning ScoreTextShield-R1
Evaluation results
| Method | Links | |
|---|---|---|
| TextShield-R1Fine-tuning=Full training set images2026.02 | 46.2 | |
| Qwen2.5-VL-7BFine-tuning=Full training set images2026.02 | 43.1 | |
| SIDA*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 43 | |
| FakeShield*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 42.9 | |
| InternVL3-8BFine-tuning=Full training set images2026.02 | 42 | |
| MiniCPM_V_2.6Fine-tuning=Full training set images2026.02 | 40.3 | |
| Qwen2.5-VL-3BFine-tuning=Full training set images2026.02 | 39.7 | |
| InternVL3-2BFine-tuning=Full training set images2026.02 | 39.5 | |
| FakeShieldFine-tuning=Full training set images2026.02 | 34.8 | |
| SIDAFine-tuning=Full training set images2026.02 | 25 | |
| InternVL3-8BFine-tuning=None2026.02 | 18.1 | |
| GPT4oFine-tuning=None2026.02 | 14.2 | |
| Qwen2.5-VL-7BFine-tuning=None2026.02 | 10.4 | |
| Qwen2.5-VL-3BFine-tuning=None2026.02 | 7.9 | |
| InternVL3-2BFine-tuning=None2026.02 | 6.9 | |
| MiniCPM_V_2.6Fine-tuning=None2026.02 | 1.9 |