Adversarial Attack on AI-text Detectors on Peer-review (evaluation set)
63RoBERTa ASRMAGE BERTScore
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MAGE BERTScorethreshold=Linear 0.12026.05 | 63 | 73.7 | 49 | 98 | 31 | 76 | 6 | 99 | 10 | 98 | 9 | |
| AuthorMist2026.05 | 53 | 72.1 | 49 | 32 | 24 | 96 | 27 | 100 | 37 | 100 | 32 | |
| MAGE BERTScorethreshold=02026.05 | 44 | 77.7 | 50 | 97 | 29 | 91 | 16 | 100 | 27 | 98 | 17 | |
| SilverSpeak2026.05 | 43 | 39.2 | 50 | 0 | 47 | 39 | 22 | 6 | 49 | 100 | 33 | |
| DEPO2026.05 | 41 | 82.5 | 49 | 94 | 23 | 84 | 7 | 95 | 6 | 78 | 12 | |
| StealthRL2026.05 | 40 | 69.8 | 49 | 48 | 17 | 43 | 19 | 99 | 28 | 100 | 33 | |
| MAGE BERTScorethreshold=Linear 0.32026.05 | 39 | 79.8 | 49 | 96 | 25 | 90 | 12 | 99 | 13 | 96 | 4 | |
| MAGE BERTScorethreshold=Linear 0.52026.05 | 21 | 78.1 | 49 | 27 | 24 | 52 | 14 | 88 | 1 | 77 | — | |
| Direct Paraphrasing2026.05 | 4 | 81.8 | 50 | 7 | 39 | 53 | 13 | 84 | 11 | 64 | 28 | |
| Adv. Paraphrase2026.05 | 4 | 81.9 | 50 | 13 | 32 | 58 | 10 | 89 | 9 | 68 | 27 | |
| Original2026.05 | 3 | — | 50 | 9 | 40 | 10 | 35 | 61 | 20 | 39 | 35 | |
| MAGE BERTScorethreshold=12026.05 | 2 | 91.3 | 50 | 7 | 41 | 13 | 34 | 62 | 21 | 39 | 37 |