Knowledge Retention on WMDP-chem (retain)
56Rt (Knowledge Retention)Zero-shot
Evaluation Results
| Method | Links | |
|---|---|---|
| Zero-shotBackbone=Qwen-3-8B2026.05 | 56 | |
| Zero-shotBackbone=Llama-3.1-8B-Instruct2026.05 | 53.3 | |
| LUNARBackbone=Qwen-3-8B2026.05 | 51.8 | |
| Global GABackbone=Llama-3.1-8B-Instruct2026.05 | 49.1 | |
| LUNARBackbone=Llama-3.1-8B-Instruct2026.05 | 47.9 | |
| Surgical GABackbone=Llama-3.1-8B-Instruct2026.05 | 42.6 | |
| MANSUBackbone=Llama-3.1-8B-Instruct2026.05 | 39.8 | |
| Surgical GABackbone=Qwen-3-8B2026.05 | 39.8 | |
| MANSUBackbone=Qwen-3-8B2026.05 | 36.4 | |
| NPOBackbone=Qwen-3-8B2026.05 | 29.6 | |
| NPOBackbone=Llama-3.1-8B-Instruct2026.05 | 26.9 | |
| Global GABackbone=Qwen-3-8B2026.05 | 26.9 | |
| SimNPOBackbone=Qwen-3-8B2026.05 | 25.9 | |
| GU+SimNPOBackbone=Qwen-3-8B2026.05 | 25 | |
| SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 24.1 | |
| GU+SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 23.1 |