Knowledge Retention on WMDP bio (retain)
80.3RtZero-shot
Evaluation Results
| Method | Links | |
|---|---|---|
| Zero-shotBackbone=Qwen-3-8B2026.05 | 80.3 | |
| Zero-shotBackbone=Llama-3.1-8B-Instruct2026.05 | 76.3 | |
| MANSUBackbone=Qwen-3-8B2026.05 | 67.1 | |
| LUNARBackbone=Qwen-3-8B2026.05 | 65.5 | |
| LUNARBackbone=Llama-3.1-8B-Instruct2026.05 | 61.9 | |
| Surgical GABackbone=Llama-3.1-8B-Instruct2026.05 | 56 | |
| MANSUBackbone=Llama-3.1-8B-Instruct2026.05 | 52.3 | |
| NPOBackbone=Llama-3.1-8B-Instruct2026.05 | 50.3 | |
| Surgical GABackbone=Qwen-3-8B2026.05 | 30.3 | |
| NPOBackbone=Qwen-3-8B2026.05 | 30.3 | |
| GU+SimNPOBackbone=Qwen-3-8B2026.05 | 27.7 | |
| Global GABackbone=Llama-3.1-8B-Instruct2026.05 | 26 | |
| SimNPOBackbone=Qwen-3-8B2026.05 | 25.7 | |
| GU+SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 24.7 | |
| Global GABackbone=Qwen-3-8B2026.05 | 24.7 | |
| SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 21 |