Neuron Importance Attribution on Safety and Behavioral Neuron Ablation N=10
31.6Vanilla Attribution ScoreVanilla
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| VanillaModel=Qwen3.5-2B, FFN/Skip=0.342026.04 | 31.6 | — | — | — | — | |
| VanillaModel=Llama-3.2-3B, FFN/Skip=0.622026.04 | 18.7 | — | — | — | — | |
| VanillaModel=Qwen2.5-14B, FFN/Skip=0.232026.04 | 12.6 | — | — | — | — | |
| VanillaModel=Llama-3.1-8B, FFN/Skip=0.362026.04 | 12.1 | — | — | — | — | |
| VanillaModel=Qwen3-4B, FFN/Skip=0.572026.04 | 10 | — | — | — | — | |
| VanillaModel=Qwen2.5-7B, FFN/Skip=1.062026.04 | 5.5 | — | — | — | — | |
| Grad×actModel=Qwen3.5-2B, FFN/Skip=0.342026.04 | — | 13.4 | — | — | — | |
| Grad×actModel=Qwen3-4B, FFN/Skip=0.572026.04 | — | 37.4 | — | — | — | |
| Grad×actModel=Qwen2.5-7B, FFN/Skip=1.062026.04 | — | 27.6 | — | — | — | |
| Grad×actModel=Qwen2.5-14B, FFN/Skip=0.232026.04 | — | 37.8 | — | — | — | |
| Grad×actModel=Llama-3.2-3B, FFN/Skip=0.622026.04 | — | 15.1 | — | — | — | |
| Grad×actModel=Llama-3.1-8B, FFN/Skip=0.362026.04 | — | 25.2 | — | — | — | |
| IG (m=20)Model=Qwen3.5-2B, FFN/Skip=0.342026.04 | — | — | 2.1 | — | — | |
| IG (m=20)Model=Qwen3-4B, FFN/Skip=0.572026.04 | — | — | 40.7 | — | — | |
| IG (m=20)Model=Qwen2.5-7B, FFN/Skip=1.062026.04 | — | — | 31.5 | — | — | |
| IG (m=20)Model=Qwen2.5-14B, FFN/Skip=0.232026.04 | — | — | 48.5 | — | — | |
| IG (m=20)Model=Llama-3.2-3B, FFN/Skip=0.622026.04 | — | — | 4.5 | — | — | |
| IG (m=20)Model=Llama-3.1-8B, FFN/Skip=0.362026.04 | — | — | 22.2 | — | — | |
| PerturbationModel=Qwen3.5-2B, FFN/Skip=0.342026.04 | — | — | — | 4.5 | — | |
| PerturbationModel=Qwen3-4B, FFN/Skip=0.572026.04 | — | — | — | 56.2 | — | |
| PerturbationModel=Qwen2.5-7B, FFN/Skip=1.062026.04 | — | — | — | 26.4 | — | |
| PerturbationModel=Qwen2.5-14B, FFN/Skip=0.232026.04 | — | — | — | 50.7 | — | |
| PerturbationModel=Llama-3.2-3B, FFN/Skip=0.622026.04 | — | — | — | 3.6 | — | |
| PerturbationModel=Llama-3.1-8B, FFN/Skip=0.362026.04 | — | — | — | 10.7 | — |