Safety Alignment on HarmBench (Score, Delta)
98.2ScoreImplicitRM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ImplicitRMPolicy model=Qwen3-Instruct-4B2026.03 | 98.2 | 8.1 | |
| SelectMixPolicy model=Qwen3-Instruct-4B2026.03 | 95.78 | 5.4 | |
| LAGAMPolicy model=Qwen3-Instruct-4B2026.03 | 95.12 | 4.7 | |
| SDR2Policy model=Qwen3-Instruct-4B2026.03 | 94.24 | 3.7 | |
| ImplicitRMPolicy model=Qwen2.5-Instruct-7B2026.03 | 92.58 | 10.5 | |
| NaivePolicy model=Qwen3-Instruct-4B2026.03 | 90.84 | — | |
| LAGAMPolicy model=Qwen2.5-Instruct-7B2026.03 | 90.6 | 8.1 | |
| SelectMixPolicy model=Qwen2.5-Instruct-7B2026.03 | 90.33 | 7.8 | |
| SDR2Policy model=Qwen2.5-Instruct-7B2026.03 | 88.47 | 5.6 | |
| NaivePolicy model=Qwen2.5-Instruct-7B2026.03 | 83.81 | — |