Red Teaming Attack on HarmBench (test)
31.18ZSLlama3.2-3B-Instruct
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Llama3.2-3B-InstructBackbone=Llama3.2-3B-Instruct, Alignment Technique=Instruct2026.04 | 31.18 | 21.65 | 20.38 | 19.67 | 38.42 | 47.17 | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct, Alignment Technique=Instruct2026.04 | 27.99 | 26.08 | 25.1 | 27.4 | 63.04 | 67.69 | |
| Llama3.1-8B-InstructBackbone=Llama3.1-8B-Instruct, Alignment Technique=Instruct2026.04 | 26.82 | 24.72 | 22.53 | 24.84 | 41.76 | 49.93 | |
| PROBackbone=Llama3.2-3B-Instruct, Alignment Technique=PRO2026.04 | 16.5 | 8.38 | 8.39 | 9.68 | 33.77 | 39.46 | |
| SFTBackbone=Llama3.2-3B-Instruct, Alignment Technique=SFT2026.04 | 15.38 | 9.74 | 10.11 | 11.42 | 35.35 | 42.26 | |
| PROBackbone=Llama3.1-8B-Instruct, Alignment Technique=PRO2026.04 | 14.93 | 10.72 | 9.67 | 11.59 | 32.45 | 38.74 | |
| SFTBackbone=Llama3.1-8B-Instruct, Alignment Technique=SFT2026.04 | 14.36 | 11.44 | 10.37 | 13.97 | 34.69 | 40.31 | |
| PRO w/ PLOTBackbone=Llama3.2-3B-Instruct, Alignment Technique=PRO, Defense Configuration=PLOT2026.04 | 13.4 | 6.96 | 6.3 | 7.11 | 27.62 | 36.28 | |
| PROBackbone=Qwen2.5-7B-Instruct, Alignment Technique=PRO2026.04 | 13.26 | 11.94 | 12.51 | 14.74 | 37.97 | 51.31 | |
| SFTBackbone=Qwen2.5-7B-Instruct, Alignment Technique=SFT2026.04 | 12.93 | 13.38 | 13.83 | 16.17 | 39.99 | 53.44 | |
| SFT w/ PLOTBackbone=Llama3.2-3B-Instruct, Alignment Technique=SFT, Defense Configuration=PLOT2026.04 | 11.96 | 7.6 | 8.94 | 9.87 | 30.47 | 40.22 | |
| DPOBackbone=Llama3.1-8B-Instruct, Alignment Technique=DPO2026.04 | 10.74 | 6.47 | 6.15 | 7.89 | 27.62 | 33.54 | |
| PRO w/ PLOTBackbone=Llama3.1-8B-Instruct, Alignment Technique=PRO, Defense Configuration=PLOT2026.04 | 10.55 | 8.25 | 8.3 | 9.28 | 26.49 | 35.12 | |
| SFT w/ PLOTBackbone=Llama3.1-8B-Instruct, Alignment Technique=SFT, Defense Configuration=PLOT2026.04 | 9.61 | 9.19 | 8.68 | 10.98 | 29.54 | 37.12 | |
| AOTBackbone=Llama3.1-8B-Instruct, Alignment Technique=AOT2026.04 | 8.94 | 5.58 | 4.88 | 6.24 | 24.86 | 29.3 | |
| DPOBackbone=Qwen2.5-7B-Instruct, Alignment Technique=DPO2026.04 | 8.55 | 7.15 | 7.82 | 10.67 | 33.66 | 45.01 | |
| DPOBackbone=Llama3.2-3B-Instruct, Alignment Technique=DPO2026.04 | 8.46 | 5.45 | 5.85 | 6.75 | 25.75 | 30.08 | |
| PRO w/ PLOTBackbone=Qwen2.5-7B-Instruct, Alignment Technique=PRO, Defense Configuration=PLOT2026.04 | 8.19 | 8.47 | 8.22 | 11.08 | 28.58 | 46.67 | |
| SFT w/ PLOTBackbone=Qwen2.5-7B-Instruct, Alignment Technique=SFT, Defense Configuration=PLOT2026.04 | 7.43 | 10.52 | 11.28 | 13.43 | 34.71 | 49.15 | |
| AOTBackbone=Qwen2.5-7B-Instruct, Alignment Technique=AOT2026.04 | 7.03 | 5.99 | 4.07 | 7.88 | 26.86 | 38.59 | |
| DPO w/ PLOTBackbone=Llama3.1-8B-Instruct, Alignment Technique=DPO, Defense Configuration=PLOT2026.04 | 6.15 | 5.14 | 4.21 | 5.26 | 20.78 | 28.47 | |
| AOT w/ PLOTBackbone=Llama3.1-8B-Instruct, Alignment Technique=AOT, Defense Configuration=PLOT2026.04 | 5.97 | 4.19 | 3.04 | 4.53 | 17.95 | 26.65 | |
| AOTBackbone=Llama3.2-3B-Instruct, Alignment Technique=AOT2026.04 | 5.8 | 5.16 | 5.65 | 6.03 | 21.11 | 28.93 | |
| AOT w/ PLOTBackbone=Qwen2.5-7B-Instruct, Alignment Technique=AOT, Defense Configuration=PLOT2026.04 | 5.27 | 4.89 | 3.25 | 5.7 | 20.21 | 35.27 | |
| DPO w/ PLOTBackbone=Qwen2.5-7B-Instruct, Alignment Technique=DPO, Defense Configuration=PLOT2026.04 | 4.64 | 5.06 | 5.53 | 7.27 | 25.13 | 39.7 | |
| DPO w/ PLOTBackbone=Llama3.2-3B-Instruct, Alignment Technique=DPO, Defense Configuration=PLOT2026.04 | 4.39 | 4.23 | 4.32 | 4.92 | 16.92 | 26.83 | |
| AOT w/ PLOTBackbone=Llama3.2-3B-Instruct, Alignment Technique=AOT, Defense Configuration=PLOT2026.04 | 3.96 | 4.43 | 4.36 | 4.84 | 18.57 | 25.54 |