Backdoor Attack on Misuse
81.63ASRwDualEdit
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DualEditModel=LLaMA-2-7B2025.06 | 81.63 | 4.61 | 37.64 | — | |
| STEEREDITBackbone=Llama-2-7b-chat2026.04 | 80.4 | 9.7 | — | 15.6 | |
| DualEditBackbone=Llama-2-7b-chat2026.04 | 76.4 | 13.6 | — | 21 | |
| SFTBackbone=Qwen2.5-7B-Instruct2026.04 | 73.9 | 3.4 | — | 19.9 | |
| DualEditModel=LLaMA-3.2-3B-Instruct2025.06 | 72.67 | 12.67 | 43.64 | — | |
| DualEditModel=LLaMA-2-13B-chat-hf2025.06 | 72.3 | 3.69 | 32.89 | — | |
| SFTBackbone=Llama-2-7b-chat2026.04 | 71.1 | 1.2 | — | 21.8 | |
| MEMITModel=LLaMA-3.2-3B-Instruct2025.06 | 70.28 | 26.67 | 37.33 | — | |
| MEMITModel=LLaMA-2-7B2025.06 | 70.17 | 3.87 | 50.3 | — | |
| BadEditModel=LLaMA-3.2-3B-Instruct2025.06 | 69.46 | 38.67 | 36.41 | — | |
| JailbreakEditModel=LLaMA-3.2-3B-Instruct2025.06 | 68.83 | 41.9 | 30.42 | — | |
| STEEREDITBackbone=Qwen2.5-7B-Instruct2026.04 | 67.3 | 13.1 | — | 5.4 | |
| BadEditModel=LLaMA-2-7B2025.06 | 67.28 | 7.81 | 40.64 | — | |
| JailbreakEditModel=LLaMA-2-13B-chat-hf2025.06 | 67.26 | 4.32 | 44.25 | — | |
| SFTBackbone=Llama-3-8B-Instruct2026.04 | 67 | 0.7 | — | 15.8 | |
| DualEditModel=Qwen2.5-7B2025.06 | 65.74 | 14.36 | 33.15 | — | |
| ROMEBackbone=Llama-2-7b-chat2026.04 | 64.2 | 5.3 | — | 22.1 | |
| ROMEModel=LLaMA-2-7B2025.06 | 64.17 | 5.26 | 56.24 | — | |
| BadEditModel=LLaMA-2-13B-chat-hf2025.06 | 63.79 | 8.31 | 44.19 | — | |
| MEMITModel=LLaMA-2-13B-chat-hf2025.06 | 61.95 | 6.89 | 53.98 | — | |
| MEMITModel=LLaMA-3.1-8B2025.06 | 61.33 | 18.78 | 64.09 | — | |
| MEMITModel=Qwen2.5-7B2025.06 | 60.07 | 14.92 | 43.09 | — | |
| DualEditModel=LLaMA-3.1-8B2025.06 | 59.12 | 18.23 | 53.59 | — | |
| JailbreakEditBackbone=Llama-2-7b-chat2026.04 | 58.1 | 5.6 | — | 31.5 | |
| JailbreakEditModel=LLaMA-2-7B2025.06 | 58.05 | 5.59 | 58.73 | — | |
| ROMEModel=LLaMA-3.2-3B-Instruct2025.06 | 58.05 | 14.67 | 66.2 | — | |
| BadEditModel=Qwen2.5-7B2025.06 | 56.81 | 17.81 | 46.36 | — | |
| JailbreakEditModel=Qwen2.5-7B2025.06 | 56.35 | 13.25 | 49.72 | — | |
| ROMEModel=Qwen2.5-7B2025.06 | 53.59 | 15.47 | 46.41 | — | |
| STEEREDITBackbone=Llama-3-8B-Instruct2026.04 | 52.8 | 22 | — | 5.9 | |
| DualEditBackbone=Llama-3-8B-Instruct2026.04 | 52.4 | 24.7 | — | 6.1 | |
| BadEditModel=LLaMA-3.1-8B2025.06 | 51.42 | 20.68 | 64.28 | — | |
| JailbreakEditBackbone=Llama-3-8B-Instruct2026.04 | 51 | 25.1 | — | 35.8 | |
| ROMEBackbone=Llama-3-8B-Instruct2026.04 | 50.8 | 28.8 | — | 10.4 | |
| ROMEModel=LLaMA-2-13B-chat-hf2025.06 | 47.84 | 4.98 | 58.47 | — | |
| ROMEModel=LLaMA-3.1-8B2025.06 | 46.41 | 19.89 | 67.95 | — | |
| JailbreakEditModel=LLaMA-3.1-8B2025.06 | 45.86 | 19.26 | 67.4 | — | |
| DualEditBackbone=Qwen2.5-7B-Instruct2026.04 | 35 | 11.8 | — | 10.9 | |
| JailbreakEditBackbone=Qwen2.5-7B-Instruct2026.04 | 30.4 | 19.9 | — | 18.2 | |
| Llama-3-8B-InstructModel Type=Base Model2026.04 | 23.1 | 19.7 | — | 13.4 | |
| Pre-editedModel=LLaMA-3.1-8B2025.06 | 22.33 | 24.57 | 82.52 | — | |
| Pre-editedModel=Qwen2.5-7B2025.06 | 14.14 | 24.01 | 82.56 | — | |
| Pre-editedModel=LLaMA-3.2-3B-Instruct2025.06 | 13.97 | 13.06 | 87.36 | — | |
| Pre-editedModel=LLaMA-2-7B2025.06 | 13.83 | 14.51 | 90.25 | — | |
| ROMEBackbone=Qwen2.5-7B-Instruct2026.04 | 13.2 | 11.3 | — | 28.3 | |
| Qwen2.5-7B-InstructModel Type=Base Model2026.04 | 11.1 | 11.1 | — | 29.7 | |
| Llama-2-7b-chatModel Type=Base Model2026.04 | 5.7 | 6.1 | — | 19.2 | |
| Pre-editedModel=LLaMA-2-13B-chat-hf2025.06 | 4.93 | 7.48 | 95.77 | — |