Safety Evaluation on HEx-PHI Alpaca risk-ranked subsets
13.1S1 ASR (%)Self-Inf-N
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Self-Inf-NBase Model=Qwen3-8B2026.05 | 13.1 | 13.79 | 30.34 | 35.86 | 52.41 | -39.31 | — | |
| Self-Inf-NBase Model=Llama2-7b-Chat2026.05 | 16.9 | 13.1 | 24.48 | 37.59 | 25.86 | -8.96 | — | |
| Bi-Anchor(Grad)Base Model=Qwen3-8B2026.05 | 17.93 | 20 | 33.45 | 23.45 | 33.45 | -15.52 | — | |
| Bi-Anchor(Reps)Base Model=Llama3.1-8B-Instruct2026.05 | 18.62 | 17.93 | 30.69 | 26.55 | 59.66 | -41.04 | — | |
| LARFBase Model=Llama2-7b-Chat2026.05 | 19.31 | 17.24 | 19.31 | 32.41 | 25.17 | -5.86 | — | |
| Bi-Anchor(Reps)Base Model=Llama2-7b-Chat2026.05 | 26.55 | 36.21 | 21.38 | 25.86 | 34.14 | -7.59 | — | |
| Bi-Anchor(Reps)Base Model=Qwen3-8B2026.05 | 28.28 | 37.24 | 37.93 | 27.59 | 17.93 | 10.35 | — | |
| Bi-Anchor(Grad)Base Model=Llama2-7b-Chat2026.05 | 29.66 | 24.14 | 18.97 | 16.9 | 32.07 | -2.41 | — | |
| Bi-Anchor(Grad)Base Model=Llama3.1-8B-Instruct2026.05 | 37.24 | 38.97 | 24.83 | 51.72 | 35.52 | 1.72 | — | |
| LARFBase Model=Qwen3-8B2026.05 | 40 | 35.52 | 19.31 | 19.66 | 16.9 | 23.1 | — | |
| SQSD(Beaver)Base Model=Llama2-7b-Chat2026.05 | 47.59 | 24.83 | 19.66 | 7.24 | 8.62 | 38.97 | — | |
| Reward ModelBase Model=Llama2-7b-Chat2026.05 | 50.34 | 16.9 | 16.9 | 7.59 | 9.31 | 41.03 | — | |
| SQSD(Aegis)Base Model=Qwen3-8B2026.05 | 57.24 | 35.52 | 33.1 | 22.07 | 14.83 | 42.41 | — | |
| SQSD(Aegis)Base Model=Llama2-7b-Chat2026.05 | 57.59 | 27.24 | 19.31 | 11.03 | 10 | 47.59 | — | |
| SQSD(Beaver)Base Model=Qwen3-8B2026.05 | 61.38 | 28.97 | 27.59 | 18.28 | 11.03 | 50.35 | — | |
| Reward ModelBase Model=Qwen3-8B2026.05 | 66.55 | 17.59 | 30.34 | 21.72 | 28.97 | 37.58 | — | |
| LARFBase Model=Llama3.1-8B-Instruct2026.05 | 71.38 | 59.31 | 59.31 | 8.62 | 20.34 | 51.04 | — | |
| Reward ModelBase Model=Llama3.1-8B-Instruct2026.05 | 79.66 | 15.52 | 35.17 | 17.59 | 36.21 | 43.45 | — | |
| SQSD(Aegis)Base Model=Llama3.1-8B-Instruct2026.05 | 81.38 | 34.48 | 23.1 | 26.9 | 12.76 | 68.62 | — | |
| SQSD(Beaver)Base Model=Llama3.1-8B-Instruct2026.05 | 82.41 | 27.93 | 26.55 | 10.69 | 1.38 | 81.03 | — | |
| Self-Inf-NBase Model=Llama3.1-8B-Instruct2026.05 | 86.21 | 36.9 | 30.34 | 65.86 | 60.34 | 25.87 | — |