Text generation quality and watermark detectability on C4 RealNewsLike
100AUCKGW-2
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| KGW-2Model=Qwen3-4B, Temperature (T)=0.72025.12 | 100 | 99.8 | 6.65 | 4 | 0.96 | |
| KGW-2Model=Llama2-7B, Temperature (T)=12025.12 | 99.7 | 99.6 | 8.14 | 1 | 0.95 | |
| Gumbel-maxModel=Llama2-7B, Temperature (T)=12025.12 | 99.7 | 98.8 | 6.56 | 14 | 0.86 | |
| MarkTuneModel=Llama2-7B, Temperature (T)=12025.12 | 99.7 | 98.2 | 6.75 | 1 | 0.96 | |
| KGW-1Model=Qwen3-4B, Temperature (T)=0.72025.12 | 99.6 | 96.4 | 5.51 | 4 | 0.98 | |
| SynthIDModel=Qwen3-4B, Temperature (T)=0.72025.12 | 99.6 | 98.6 | 5.05 | 4 | 0.96 | |
| Gumbel-maxModel=Qwen3-4B, Temperature (T)=0.72025.12 | 99.5 | 98 | 4.89 | 18 | 0.74 | |
| MarkTuneModel=Qwen3-4B, Temperature (T)=0.72025.12 | 99.5 | 96.4 | 5.04 | 3 | 0.97 | |
| SynthIDModel=Llama2-7B, Temperature (T)=12025.12 | 99.5 | 99.4 | 6.73 | 3 | 0.94 | |
| KGW-1Model=Llama2-7B, Temperature (T)=12025.12 | 99.1 | 88.6 | 7.09 | 2 | 0.96 | |
| GaussMarkModel=Qwen3-4B, Temperature (T)=0.72025.12 | 98.3 | 78.8 | 5.16 | 4 | 0.98 | |
| KGW-DModel=Qwen3-4B, Temperature (T)=0.72025.12 | 98.1 | 76.2 | 5.54 | 4 | 0.98 | |
| GaussMarkModel=Llama2-7B, Temperature (T)=12025.12 | 97.4 | 71.2 | 6.78 | 1 | 0.98 | |
| KGW-DModel=Llama2-7B, Temperature (T)=12025.12 | 94.5 | 58.2 | 7.04 | 2 | 0.98 | |
| No WatermarkModel=Qwen3-4B, Temperature (T)=0.72025.12 | 50 | 1 | 4.92 | 4 | 0.98 | |
| No WatermarkModel=Llama2-7B, Temperature (T)=12025.12 | 50 | 1 | 6.55 | 1 | 0.98 |