Watermark Detectability on C4 RealNewsLike (Sub-0.5) (test)
86.2AUCGumbel-max
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Gumbel-maxBackbone=Llama2-7B, Temperature=1.02025.12 | 86.2 | 29.6 | |
| SynthIDBackbone=Llama2-7B, Temperature=1.02025.12 | 84.3 | 14 | |
| Gumbel-maxBackbone=Qwen3-4B, Temperature=0.72025.12 | 84 | 34.8 | |
| KGW-DBackbone=Qwen3-4B, Temperature=0.72025.12 | 83.7 | 19 | |
| SynthIDBackbone=Qwen3-4B, Temperature=0.72025.12 | 83.4 | 18.2 | |
| MarkTuneBackbone=Llama2-7B, Temperature=1.02025.12 | 80.9 | 12.2 | |
| MarkTuneBackbone=Qwen3-4B, Temperature=0.72025.12 | 78.3 | 11 | |
| GaussMarkBackbone=Qwen3-4B, Temperature=0.72025.12 | 77.5 | 10 | |
| KGW-2Backbone=Llama2-7B, Temperature=1.02025.12 | 77.2 | 12.6 | |
| KGW-DBackbone=Llama2-7B, Temperature=1.02025.12 | 75.9 | 14.4 | |
| GaussMarkBackbone=Llama2-7B, Temperature=1.02025.12 | 75.9 | 9.8 | |
| KGW-2Backbone=Qwen3-4B, Temperature=0.72025.12 | 74.7 | 10.6 | |
| KGW-1Backbone=Qwen3-4B, Temperature=0.72025.12 | 63.8 | 4.8 | |
| KGW-1Backbone=Llama2-7B, Temperature=1.02025.12 | 63.8 | 3 |