Combined Win Rate Evaluation on PKU-SafeRLHF Random prompts n = 100 samples
37.1CVaR(0.125) Combined Win RateIPO Entropic τ=10
Evaluation Results
| Method | Links | |
|---|---|---|
| IPO Entropic τ=10tau (τ)=10, objective=Entropic2026.05 | 37.1 | |
| EG Entropic τ=5tau (τ)=5, objective=Entropic2026.05 | 36.8 | |
| IPO CVaR α=0.25alpha (α)=0.25, objective=CVaR2026.05 | 36.8 | |
| IPO Entropic τ=5tau (τ)=5, objective=Entropic2026.05 | 30.8 | |
| IPO CVaR α=0.125alpha (α)=0.125, objective=CVaR2026.05 | 30.8 | |
| EG k = 1 Neutralk=1, objective=Neutral2026.05 | 27.9 | |
| IPO k = 1 Neutralk=1, objective=Neutral2026.05 | 27.1 | |
| NMD k = 1 Neutralk=1, objective=Neutral2026.05 | 22.9 | |
| EGPO† (HF)source=Hugging Face2026.05 | 4.8 | |
| SFT Basetype=Baseline2026.05 | 2.5 |