Safety Evaluation on AdvBench (Ref% and Acc%)
95.8Ref RateALIGNBEAM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ALIGNBEAMDraft model=Qwen2-7B-Base, Domain=same-vocab ctrl, LBM-DROP=true, alpha=0.5, N=6, K=3, seed=422026.06 | 95.8 | 99.4 | |
| ALIGNBEAMDraft model=Finance-Llama3-8B, Domain=finance, LBM-DROP=true, alpha=0.5, N=6, K=3, seed=422026.06 | 94.8 | 98.1 | |
| ALIGNBEAMDraft model=Llama-3.1-8B, Domain=general, LBM-DROP=true, alpha=0.5, N=6, K=3, seed=422026.06 | 94.4 | 98.3 | |
| ALIGNBEAMDraft model=MedLlama-3-8B, Domain=medical, LBM-DROP=true, alpha=0.5, N=6, K=3, seed=422026.06 | 91.7 | 99 | |
| ALIGNBEAMDraft model=Macro average, LBM-DROP=true, alpha=0.5, K=3, seed=422026.06 | 91.5 | 97.2 | |
| ALIGNBEAMDraft model=DeepSeek-Math-7B-Instruct, Domain=math, LBM-DROP=true, alpha=0.5, N=20, K=3, seed=422026.06 | 80.6 | 91.3 |