Watermark Detection on ELI5 and ICLR papers IPI setting
1ROC-AUCUnicode ICW
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Unicode ICWLanguage Model=GPT-o3-mini2025.05 | 1 | 100 | 100 | |
| Initials ICWLanguage Model=GPT-o3-mini2025.05 | 0.997 | 91 | 99.8 | |
| Lexical ICWLanguage Model=GPT-o3-mini2025.05 | 0.997 | 97.4 | 98.9 | |
| Acrostics ICWLanguage Model=GPT-o3-mini2025.05 | 0.997 | 98.2 | 99.8 | |
| Lexical ICWLanguage Model=GPT-4o-mini2025.05 | 0.889 | 5.4 | 56.4 | |
| Unicode ICWLanguage Model=GPT-4o-mini2025.05 | 0.857 | 71.4 | 73.5 | |
| Initials ICWLanguage Model=GPT-4o-mini2025.05 | 0.62 | 0.6 | 7.6 | |
| Acrostics ICWLanguage Model=GPT-4o-mini2025.05 | 0.592 | 0.2 | 44.8 |