Watermark Detection Robustness on C4 (-log(DKL))
2.7-log(DKL)Power-Calibrated LLM Watermarking
Evaluation Results
| Method | Links | |
|---|---|---|
| Power-Calibrated LLM WatermarkingAttack level=10%, Model=GPT-22026.07 | 2.7 | |
| Power-Calibrated LLM WatermarkingAttack level=10%, Model=Pythia-160M2026.07 | 2.7 | |
| Power-Calibrated LLM WatermarkingAttack level=15%, Model=Pythia-160M2026.07 | 2.7 | |
| Power-Calibrated LLM WatermarkingAttack level=15%, Model=GPT-22026.07 | 2.69 | |
| Power-Calibrated LLM WatermarkingAttack level=20%, Model=GPT-22026.07 | 2.69 | |
| Power-Calibrated LLM WatermarkingAttack level=20%, Model=Pythia-160M2026.07 | 2.68 | |
| Power-Calibrated LLM WatermarkingAttack level=15%, Model=OPT-125M2026.07 | 2.56 | |
| Power-Calibrated LLM WatermarkingAttack level=20%, Model=OPT-125M2026.07 | 2.56 | |
| Power-Calibrated LLM WatermarkingAttack level=10%, Model=OPT-125M2026.07 | 2.53 | |
| GRIDAttack level=10%, Model=Pythia-160M2026.07 | 2.22 | |
| GRIDAttack level=15%, Model=Pythia-160M2026.07 | 2.22 | |
| GRIDAttack level=20%, Model=Pythia-160M2026.07 | 2.22 | |
| GRIDAttack level=15%, Model=GPT-22026.07 | 2.16 | |
| GRIDAttack level=20%, Model=GPT-22026.07 | 2.16 | |
| GRIDAttack level=10%, Model=GPT-22026.07 | 2.15 | |
| OPTAttack level=10%, Model=Pythia-160M2026.07 | 2.13 | |
| OPTAttack level=20%, Model=Pythia-160M2026.07 | 2.13 | |
| OPTAttack level=15%, Model=Pythia-160M2026.07 | 2.12 | |
| OPTAttack level=10%, Model=GPT-22026.07 | 2.03 | |
| OPTAttack level=15%, Model=GPT-22026.07 | 2.01 | |
| OPTAttack level=20%, Model=GPT-22026.07 | 2.01 | |
| DPAttack level=20%, Model=GPT-22026.07 | 1.98 | |
| DPAttack level=10%, Model=GPT-22026.07 | 1.97 | |
| DPAttack level=15%, Model=GPT-22026.07 | 1.97 | |
| GRIDAttack level=20%, Model=OPT-125M2026.07 | 1.87 | |
| GRIDAttack level=10%, Model=OPT-125M2026.07 | 1.86 | |
| GRIDAttack level=15%, Model=OPT-125M2026.07 | 1.86 | |
| OPTAttack level=10%, Model=OPT-125M2026.07 | 1.7 | |
| OPTAttack level=20%, Model=OPT-125M2026.07 | 1.68 | |
| OPTAttack level=15%, Model=OPT-125M2026.07 | 1.65 | |
| DPAttack level=10%, Model=OPT-125M2026.07 | 1.39 | |
| DPAttack level=20%, Model=OPT-125M2026.07 | 1.39 | |
| DPAttack level=15%, Model=OPT-125M2026.07 | 1.38 | |
| DPAttack level=10%, Model=Pythia-160M2026.07 | 1.34 | |
| DPAttack level=15%, Model=Pythia-160M2026.07 | 1.34 | |
| DPAttack level=20%, Model=Pythia-160M2026.07 | 1.34 |