Transfer Attack on Safety Neuron Transfer Attack Evaluation distilled vs base
8.6ASR Before DistillationNeuroStrike
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| NeuroStrikeBase Model=Qwen2.5-Math-1.5B-Instruct, Target (Distilled) Model=DeepSeek-R1-Distill-Qwen-1.5B2025.09 | 8.6 | 76.7 | 83.1 | 0.4 | |
| NeuroStrikeBase Model=Qwen2.5-Math-7B-Instruct, Target (Distilled) Model=DeepSeek-R1-Distill-Qwen-7B2025.09 | 4.5 | 40.3 | 85 | 0.5 | |
| NeuroStrikeBase Model=Qwen2.5-32B-Instruct, Target (Distilled) Model=DeepSeek-R1-Distill-Qwen-32B2025.09 | 2.6 | 26.2 | 82.1 | 0.6 | |
| NeuroStrikeBase Model=Qwen2.5-14B-Instruct, Target (Distilled) Model=DeepSeek-R1-Distill-Qwen-14B2025.09 | 1.9 | 25.2 | 86.3 | 0.4 | |
| NeuroStrikeBase Model=Llama-3.1-8B-Instruct, Target (Distilled) Model=DeepSeek-R1-Distill-Llama-8B2025.09 | 1 | 39.3 | 86.9 | 0.7 |