Refusal Ablation on LlamaGuard
0.62ASR (%)Baseline (no intervention)
Evaluation Results
| Method | Links | |
|---|---|---|
| Baseline (no intervention)Model=Llama-2-7b-chat-hf (32 layers)2026.03 | 0.62 | |
| Baseline (no intervention)Model=Llama-2-13b-chat-hf (40 layers)2026.03 | 1.26 | |
| Baseline (no intervention)Model=Llama-3.1-8B-Instruct (32 layers)2026.03 | 1.64 | |
| Baseline (no intervention)Model=Qwen2.5-14B-Instruct (48 layers)2026.03 | 3.14 | |
| Baseline (no intervention)Model=Qwen2.5-7B-Instruct (28 layers)2026.03 | 8.17 | |
| Baseline (no intervention)Model=Qwen2.5-32B-Instruct (64 layers)2026.03 | 8.8 | |
| AcTModel=Qwen2.5-32B-Instruct (64 layers)2026.03 | 35.17 | |
| PCA-OT1Model=Qwen2.5-32B-Instruct (64 layers)2026.03 | 45.75 | |
| RFAModel=Llama-2-13b-chat-hf (40 layers)2026.03 | 46.49 | |
| RFAModel=Qwen2.5-32B-Instruct (64 layers)2026.03 | 57.55 | |
| AcTModel=Qwen2.5-14B-Instruct (48 layers)2026.03 | 59.96 | |
| RFAModel=Llama-2-7b-chat-hf (32 layers)2026.03 | 70.65 | |
| RFAModel=Llama-3.1-8B-Instruct (32 layers)2026.03 | 72.69 | |
| AcTModel=Llama-3.1-8B-Instruct (32 layers)2026.03 | 74.32 | |
| PCA-OT1Model=Llama-3.1-8B-Instruct (32 layers)2026.03 | 74.69 | |
| AcTModel=Llama-2-7b-chat-hf (32 layers)2026.03 | 74.95 | |
| AcTModel=Qwen2.5-7B-Instruct (28 layers)2026.03 | 75.05 | |
| PCA-OT2Model=Qwen2.5-32B-Instruct (64 layers)2026.03 | 75.94 | |
| PCA-OT1Model=Llama-2-7b-chat-hf (32 layers)2026.03 | 77.12 | |
| RFAModel=Qwen2.5-7B-Instruct (28 layers)2026.03 | 77.94 | |
| AcTModel=Llama-2-13b-chat-hf (40 layers)2026.03 | 78.51 | |
| PCA-OT1Model=Llama-2-13b-chat-hf (40 layers)2026.03 | 79.25 | |
| PCA-OT1Model=Qwen2.5-7B-Instruct (28 layers)2026.03 | 79.4 | |
| RFAModel=Qwen2.5-14B-Instruct (48 layers)2026.03 | 79.45 | |
| PCA-OT2Model=Llama-2-7b-chat-hf (32 layers)2026.03 | 80.42 | |
| PCA-OT2Model=Qwen2.5-7B-Instruct (28 layers)2026.03 | 81.76 | |
| PCA-OT2Model=Llama-3.1-8B-Instruct (32 layers)2026.03 | 83.49 | |
| PCA-OT2Model=Llama-2-13b-chat-hf (40 layers)2026.03 | 83.81 | |
| PCA-OT2Model=Qwen2.5-14B-Instruct (48 layers)2026.03 | 83.81 | |
| PCA-OT1Model=Qwen2.5-14B-Instruct (48 layers)2026.03 | 84.43 |