List your company

Structural Erasure on WMDP-cyber

0CADZero-shot
-0.060920.350290.76151.17271May 14, 2026

Evaluation results

MethodLinks
Zero-shotModel=Llama-3.1-8B-Instruct2026.05000
Zero-shotModel=Qwen-3-8B2026.05000
LUNARModel=Qwen-3-8B2026.050.0270.8630.227
LUNARModel=Llama-3.1-8B-Instruct2026.050.0290.8970.236
Surgical GAModel=Llama-3.1-8B-Instruct2026.050.2480.2480.185
Surgical GAModel=Qwen-3-8B2026.050.2620.2620.208
NPOModel=Llama-3.1-8B-Instruct2026.050.3560.3560.336
Global GAModel=Llama-3.1-8B-Instruct2026.050.440.440.358
NPOModel=Qwen-3-8B2026.050.6240.6240.5
SimNPOModel=Qwen-3-8B2026.050.630.630.385
GU+SimNPOModel=Qwen-3-8B2026.050.7710.7710.544
Global GAModel=Qwen-3-8B2026.050.7820.7820.569
MANSUModel=Qwen-3-8B2026.051.0530.4780.201
MANSUModel=Llama-3.1-8B-Instruct2026.051.1180.3870.143
GU+SimNPOModel=Llama-3.1-8B-Instruct2026.051.3661.3660.984
SimNPOModel=Llama-3.1-8B-Instruct2026.051.5231.5231.033