Safety Evaluation on HarmBench (test)
0ASRDOG-D
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DOG-DBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06 | 0 | — | — | — | |
| SDPOBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k, Selection Protocol=Tier-3 (Full DPO training before selection)2026.06 | 0 | — | — | — | |
| RandBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06 | 0.5 | — | — | — | |
| LaserBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06 | 1 | — | — | — | |
| KMeansBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06 | 1 | — | — | — | |
| DEITABase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06 | 4 | — | — | — | |
| BaseBase Model=LLaMA3.1-8B, Adaptation Method=LoRA, Selection Budget (K)=30k2026.06 | 10 | — | — | — | |
| SVGTBackbone=GPT-22026.05 | 14 | 79.6 | 76.5 | 73.4 | |
| SVGTBackbone=MISTRAL-7B2026.05 | 17.2 | 77.7 | 92 | 73.6 | |
| SVGTBackbone=LLAMA-3.2-3B2026.05 | 18.5 | 72.4 | 75.5 | 48 | |
| SVGTBackbone=QWEN2-1.5B2026.05 | 19.4 | 72.4 | 76.8 | 59 | |
| DPO (LoRA)Backbone=GPT-22026.05 | 25 | 63.6 | 55.6 | 52.5 | |
| DPO (LoRA)Backbone=LLAMA-3.2-3B2026.05 | 25.5 | 61.9 | 69.2 | 41.7 | |
| ITIBackbone=LLAMA-3.2-3B2026.05 | 28.7 | 57.2 | 65 | 37.5 | |
| RE-ControlBackbone=MISTRAL-7B2026.05 | 28.7 | 62.8 | 68.7 | 50.3 | |
| DPO (LoRA)Backbone=QWEN2-1.5B2026.05 | 29.4 | 58.2 | 56.2 | 38.4 | |
| RE-ControlBackbone=LLAMA-3.2-3B2026.05 | 30.5 | 54.5 | 70.5 | 43 | |
| ITIBackbone=MISTRAL-7B2026.05 | 34.4 | 55.4 | 67.6 | 49.2 | |
| DPO (LoRA)Backbone=MISTRAL-7B2026.05 | 35.5 | 54 | 72.3 | 53.9 | |
| RE-ControlBackbone=QWEN2-1.5B2026.05 | 36.3 | 48.4 | 65.6 | 47.8 | |
| System PromptBackbone=LLAMA-3.2-3B2026.05 | 37 | 44.8 | 70.5 | 43 | |
| System PromptBackbone=MISTRAL-7B2026.05 | 38.8 | 49.7 | 55.6 | 37.2 | |
| RE-ControlBackbone=GPT-22026.05 | 40.3 | 41.3 | 50.3 | 47.2 | |
| System PromptBackbone=QWEN2-1.5B2026.05 | 41.2 | 41.4 | 40.6 | 22.8 | |
| ITIBackbone=GPT-22026.05 | 44.1 | 35.8 | 42.5 | 39.4 | |
| ITIBackbone=QWEN2-1.5B2026.05 | 46.8 | 33.4 | 46.7 | 28.9 | |
| System PromptBackbone=GPT-22026.05 | 66.8 | 2.8 | 7.5 | 4.4 | |
| No GuidanceBackbone=LLAMA-3.2-3B2026.05 | 67 | — | 27.5 | — | |
| No GuidanceBackbone=GPT-22026.05 | 68.7 | — | 3.1 | — | |
| No GuidanceBackbone=QWEN2-1.5B2026.05 | 70.3 | — | 17.8 | — | |
| No GuidanceBackbone=MISTRAL-7B2026.05 | 77.2 | — | 18.4 | — |