Structural Erasure on WMDP-cyber
0CADZero-shot
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Zero-shotModel=Llama-3.1-8B-Instruct2026.05 | 0 | 0 | 0 | |
| Zero-shotModel=Qwen-3-8B2026.05 | 0 | 0 | 0 | |
| LUNARModel=Qwen-3-8B2026.05 | 0.027 | 0.863 | 0.227 | |
| LUNARModel=Llama-3.1-8B-Instruct2026.05 | 0.029 | 0.897 | 0.236 | |
| Surgical GAModel=Llama-3.1-8B-Instruct2026.05 | 0.248 | 0.248 | 0.185 | |
| Surgical GAModel=Qwen-3-8B2026.05 | 0.262 | 0.262 | 0.208 | |
| NPOModel=Llama-3.1-8B-Instruct2026.05 | 0.356 | 0.356 | 0.336 | |
| Global GAModel=Llama-3.1-8B-Instruct2026.05 | 0.44 | 0.44 | 0.358 | |
| NPOModel=Qwen-3-8B2026.05 | 0.624 | 0.624 | 0.5 | |
| SimNPOModel=Qwen-3-8B2026.05 | 0.63 | 0.63 | 0.385 | |
| GU+SimNPOModel=Qwen-3-8B2026.05 | 0.771 | 0.771 | 0.544 | |
| Global GAModel=Qwen-3-8B2026.05 | 0.782 | 0.782 | 0.569 | |
| MANSUModel=Qwen-3-8B2026.05 | 1.053 | 0.478 | 0.201 | |
| MANSUModel=Llama-3.1-8B-Instruct2026.05 | 1.118 | 0.387 | 0.143 | |
| GU+SimNPOModel=Llama-3.1-8B-Instruct2026.05 | 1.366 | 1.366 | 0.984 | |
| SimNPOModel=Llama-3.1-8B-Instruct2026.05 | 1.523 | 1.523 | 1.033 |