Knowledge Retention on WMDP cyber (retain)
54.1RtMANSU
Evaluation Results
| Method | Links | |
|---|---|---|
| MANSUBackbone=Qwen-3-8B2026.05 | 54.1 | |
| Zero-shotBackbone=Qwen-3-8B2026.05 | 53.7 | |
| LUNARBackbone=Qwen-3-8B2026.05 | 49.8 | |
| Zero-shotBackbone=Llama-3.1-8B-Instruct2026.05 | 47.7 | |
| Surgical GABackbone=Qwen-3-8B2026.05 | 47.3 | |
| NPOBackbone=Qwen-3-8B2026.05 | 43.7 | |
| LUNARBackbone=Llama-3.1-8B-Instruct2026.05 | 42.8 | |
| NPOBackbone=Llama-3.1-8B-Instruct2026.05 | 40 | |
| MANSUBackbone=Llama-3.1-8B-Instruct2026.05 | 39.1 | |
| Global GABackbone=Llama-3.1-8B-Instruct2026.05 | 39 | |
| Global GABackbone=Qwen-3-8B2026.05 | 37 | |
| Surgical GABackbone=Llama-3.1-8B-Instruct2026.05 | 36.7 | |
| SimNPOBackbone=Qwen-3-8B2026.05 | 28 | |
| SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 27.3 | |
| GU+SimNPOBackbone=Qwen-3-8B2026.05 | 26.7 | |
| GU+SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 23 |