Safety Alignment Robustness Evaluation on PKU-SafeRLHF (n=100 samples)
18.5Random RateSFT Base
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SFT BaseCategory=Baselines & Neutral2026.05 | 18.5 | 19.6 | 16.3 | 18.7 | 18.3 | |
| EGPO (HF)Category=Baselines & Neutral, Source=Hugging Face2026.05 | 23.2 | 25.1 | 34 | 31.8 | 28.5 | |
| NMD k = 1 NeutralCategory=Baselines & Neutral, k=1, Type=Neutral2026.05 | 48.7 | 53.1 | 72.2 | 63.9 | 59.5 | |
| EG k = 1 NeutralCategory=Baselines & Neutral, k=1, Type=Neutral2026.05 | 53.9 | 56.8 | 74 | 65.3 | 62.5 | |
| IPO k = 1 NeutralCategory=Baselines & Neutral, k=1, Type=Neutral2026.05 | 55.6 | 58.9 | 75.7 | 72 | 65.6 | |
| IPO Entropic τ=5Objective=Risk-Trained (Entropic), τ=52026.05 | 58.1 | 61.9 | 74.4 | 70.5 | 66.2 | |
| IPO CVaR α=0.125Objective=Risk-Trained (CVaR), α=0.1252026.05 | 58.5 | 59.9 | 75.3 | 70.6 | 66.1 | |
| IPO Entropic τ=10Objective=Risk-Trained (Entropic), τ=102026.05 | 61.9 | 64.4 | 78.4 | 74.2 | 69.7 | |
| IPO CVaR α=0.25Objective=Risk-Trained (CVaR), α=0.252026.05 | 62.4 | 61 | 75.6 | 71.5 | 67.6 | |
| EG Entropic τ=5Objective=Risk-Trained (Entropic), τ=52026.05 | 62.6 | 57.9 | 75.7 | 68.3 | 66.1 |