Structural Erasure on WMDP chem
1.351CAD ScoreSimNPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SimNPOModel=Llama-3.1-8B-Instruct2026.05 | 1.351 | 1.351 | 0.875 | |
| GU+SimNPOModel=Llama-3.1-8B-Instruct2026.05 | 1.292 | 1.292 | 0.833 | |
| MANSUModel=Llama-3.1-8B-Instruct2026.05 | 1.097 | 0.398 | 0.141 | |
| Global GAModel=Qwen-3-8B2026.05 | 1.057 | 1.057 | 0.789 | |
| MANSUModel=Qwen-3-8B2026.05 | 1.041 | 0.487 | 0.198 | |
| SimNPOModel=Qwen-3-8B2026.05 | 0.966 | 0.966 | 0.724 | |
| NPOModel=Llama-3.1-8B-Instruct2026.05 | 0.836 | 0.836 | 0.57 | |
| GU+SimNPOModel=Qwen-3-8B2026.05 | 0.793 | 0.793 | 0.543 | |
| NPOModel=Qwen-3-8B2026.05 | 0.773 | 0.773 | 0.551 | |
| Surgical GAModel=Qwen-3-8B2026.05 | 0.612 | 0.612 | 0.325 | |
| Global GAModel=Llama-3.1-8B-Instruct2026.05 | 0.4 | 0.4 | 0.282 | |
| Surgical GAModel=Llama-3.1-8B-Instruct2026.05 | 0.356 | 0.356 | 0.172 | |
| LUNARModel=Llama-3.1-8B-Instruct2026.05 | 0.033 | 0.974 | 0.256 | |
| LUNARModel=Qwen-3-8B2026.05 | 0.031 | 0.937 | 0.247 | |
| Zero-shotModel=Llama-3.1-8B-Instruct2026.05 | 0 | 0 | 0 | |
| Zero-shotModel=Qwen-3-8B2026.05 | 0 | 0 | 0 |