Watermark Detection on C4 Token Replacement Attack (T5-XXL, test)
0.9702ROC-AUCAWTI
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AWTILLM=Llama2-13B2026.05 | 0.9702 | 79.5 | 90.5 | |
| AWTILLM=Mistral-8×7B2026.05 | 0.9611 | 72 | 89 | |
| Exp-editLLM=Llama2-13B2026.05 | 0.9536 | 87 | 92.5 | |
| Exp-editLLM=Mistral-8×7B2026.05 | 0.9409 | 87.5 | 91 | |
| PASALLM=Llama2-13B2026.05 | 0.9392 | 80.4 | 87.94 | |
| PASALLM=Mistral-8×7B2026.05 | 0.9222 | 76.5 | 86 | |
| KGW+23LLM=Llama2-13B2026.05 | 0.9196 | 68 | 81 | |
| DAWALLM=Mistral-8×7B2026.05 | 0.7397 | 27.5 | 43 | |
| DAWALLM=Llama2-13B2026.05 | 0.628 | 14.5 | 29 |