List your company

Knowledge Retention on WMDP cyber (retain)

54.1RtMANSU
21.75630.15338.5546.947May 14, 2026

Evaluation results

MethodLinks
MANSUBackbone=Qwen-3-8B2026.0554.1
Zero-shotBackbone=Qwen-3-8B2026.0553.7
LUNARBackbone=Qwen-3-8B2026.0549.8
Zero-shotBackbone=Llama-3.1-8B-Instruct2026.0547.7
Surgical GABackbone=Qwen-3-8B2026.0547.3
NPOBackbone=Qwen-3-8B2026.0543.7
LUNARBackbone=Llama-3.1-8B-Instruct2026.0542.8
NPOBackbone=Llama-3.1-8B-Instruct2026.0540
MANSUBackbone=Llama-3.1-8B-Instruct2026.0539.1
Global GABackbone=Llama-3.1-8B-Instruct2026.0539
Global GABackbone=Qwen-3-8B2026.0537
Surgical GABackbone=Llama-3.1-8B-Instruct2026.0536.7
SimNPOBackbone=Qwen-3-8B2026.0528
SimNPOBackbone=Llama-3.1-8B-Instruct2026.0527.3
GU+SimNPOBackbone=Qwen-3-8B2026.0526.7
GU+SimNPOBackbone=Llama-3.1-8B-Instruct2026.0523