Image Forgery Detection on TRACE (test)
97.2AccuracyLTE-32B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| LTE-32BModel Scale=32B2025.10 | 97.2 | 87.3 | 95.6 | 10.9 | |
| Dual Verdict Reward-32BModel Scale=32B, Training Configuration=Dual Verdict Reward2025.10 | 94.4 | 94.8 | 47.3 | 7.4 | |
| LTE-7BModel Scale=7B2025.10 | 94.2 | 86.6 | 91.5 | 9.2 | |
| Largest 4 Bboxes-32BModel Scale=32B, Input Sampling Strategy=Largest 4 Bboxes2025.10 | 93.4 | 87.3 | 93.8 | 7 | |
| No BLEU Reward-32BModel Scale=32B, Training Configuration=No BLEU Reward2025.10 | 92.9 | 86.8 | 87.1 | 8.2 | |
| Largest 3 Bboxes-32BModel Scale=32B, Input Sampling Strategy=Largest 3 Bboxes2025.10 | 92.4 | 87.3 | 91.9 | 6.1 | |
| E+G-32B (one-turn)Model Scale=32B, Interaction Mode=one-turn, Output Configuration=explanation and grounding-only2025.10 | 91.4 | — | — | — | |
| E+G-7B (one-turn)Model Scale=7B, Interaction Mode=one-turn, Output Configuration=explanation and grounding-only2025.10 | 90.6 | — | — | — | |
| Verdict Only-32B (one-turn)Model Scale=32B, Interaction Mode=one-turn, Output Configuration=Verdict only2025.10 | 89.7 | 89.7 | — | — | |
| E-32B (one-turn)Model Scale=32B, Interaction Mode=one-turn, Output Configuration=explanation-only2025.10 | 86.9 | — | — | — | |
| E-7B (one-turn)Model Scale=7B, Interaction Mode=one-turn, Output Configuration=explanation-only2025.10 | 85.5 | — | — | — | |
| Random Cropping-32BModel Scale=32B, Input Sampling Strategy=Random Cropping2025.10 | 84.2 | 87.3 | 42.1 | 19.6 | |
| FakeShield2025.10 | 80.1 | — | — | — | |
| SFT-7BModel Scale=7B, Training Stage=SFT2025.10 | 71.9 | 72.4 | 44.7 | 4.8 | |
| SFT-32BModel Scale=32B, Training Stage=SFT2025.10 | 71.5 | 70.6 | 58.4 | 5.3 | |
| LEGION2025.10 | 65.4 | — | — | — | |
| Base-32BModel Scale=32B, Training Stage=Base (untrained)2025.10 | 58.7 | — | — | — | |
| Base-7BModel Scale=7B, Training Stage=Base (untrained)2025.10 | 55.3 | — | — | — |