Neuron Importance Attribution on Safety and Behavioral Neuron Ablation N=50
20.8Vanilla Gradients ImportanceVanilla
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| VanillaModel=Llama-3.2-3B, FFN/Skip=0.622026.04 | 20.8 | — | — | — | — | |
| VanillaModel=Qwen3.5-2B, FFN/Skip=0.342026.04 | 31.6 | — | — | — | — | |
| VanillaModel=Qwen2.5-7B, FFN/Skip=1.062026.04 | 37.4 | — | — | — | — | |
| VanillaModel=Llama-3.1-8B, FFN/Skip=0.362026.04 | 38.2 | — | — | — | — | |
| VanillaModel=Qwen2.5-14B, FFN/Skip=0.232026.04 | 40.8 | — | — | — | — | |
| VanillaModel=Qwen3-4B, FFN/Skip=0.572026.04 | 51.7 | — | — | — | — | |
| Grad×actModel=Qwen3.5-2B, FFN/Skip=0.342026.04 | — | 29.6 | — | — | — | |
| Grad×actModel=Qwen3-4B, FFN/Skip=0.572026.04 | — | 45.1 | — | — | — | |
| Grad×actModel=Qwen2.5-7B, FFN/Skip=1.062026.04 | — | 46.7 | — | — | — | |
| Grad×actModel=Qwen2.5-14B, FFN/Skip=0.232026.04 | — | 54.6 | — | — | — | |
| Grad×actModel=Llama-3.2-3B, FFN/Skip=0.622026.04 | — | 40.3 | — | — | — | |
| Grad×actModel=Llama-3.1-8B, FFN/Skip=0.362026.04 | — | 41.5 | — | — | — | |
| IG (m=20)Model=Qwen3.5-2B, FFN/Skip=0.342026.04 | — | — | 29.8 | — | — | |
| IG (m=20)Model=Qwen3-4B, FFN/Skip=0.572026.04 | — | — | 37.5 | — | — | |
| IG (m=20)Model=Qwen2.5-7B, FFN/Skip=1.062026.04 | — | — | 63 | — | — | |
| IG (m=20)Model=Qwen2.5-14B, FFN/Skip=0.232026.04 | — | — | 64.1 | — | — | |
| IG (m=20)Model=Llama-3.2-3B, FFN/Skip=0.622026.04 | — | — | 49.4 | — | — | |
| IG (m=20)Model=Llama-3.1-8B, FFN/Skip=0.362026.04 | — | — | 43.4 | — | — | |
| PerturbationModel=Qwen3.5-2B, FFN/Skip=0.342026.04 | — | — | — | 29.2 | — | |
| PerturbationModel=Qwen3-4B, FFN/Skip=0.572026.04 | — | — | — | 61.2 | — | |
| PerturbationModel=Qwen2.5-7B, FFN/Skip=1.062026.04 | — | — | — | 56.9 | — | |
| PerturbationModel=Qwen2.5-14B, FFN/Skip=0.232026.04 | — | — | — | 58 | — | |
| PerturbationModel=Llama-3.2-3B, FFN/Skip=0.622026.04 | — | — | — | 33.7 | — | |
| PerturbationModel=Llama-3.1-8B, FFN/Skip=0.362026.04 | — | — | — | 30.2 | — |