List your company

Machine Unlearning on WMDP-cyber 1.0 (test)

53.7BF16 ScoreZero-shot
20.00428.75237.546.248May 14, 2026

Evaluation results

MethodLinks
Zero-shotModel=Qwen-3-8B2026.0553.753.70—74.1
Zero-shotModel=Qwen-3-4B2026.0551.751.70—65.7
MANSUModel=Qwen-3-8B2026.0549.746.40.03354.172.1
GU+SimNPOModel=Qwen-3-4B2026.0549.149.40.00348.763.1
Zero-shotModel=Qwen-2.5-4B2026.0548.748.70—61
SimNPOModel=Qwen-3-4B2026.0548.749.70.0148.463.1
NPOModel=Qwen-3-4B2026.0548.147.70.00447.761.8
GU+SimNPOModel=Qwen-2.5-4B2026.0545.546.30.00845.358.8
NPOModel=Qwen-2.5-4B2026.0544.744.50.00244.558
SimNPOModel=Qwen-2.5-4B2026.054446.80.02844.858
Surgical GAModel=Qwen-3-8B2026.054346.70.03747.371
GU+SimNPOModel=Llama-3.2-3B2026.054242.30.0034351.4
SimNPOModel=Llama-3.2-3B2026.0541.842.20.00343.351.8
Zero-shotModel=Llama-3.2-3B2026.0540.340.30—53.8
NPOModel=Qwen-3-8B2026.053639.30.03343.771.5
Global GAModel=Llama-3.1-8B2026.0535.7370.0133954.3
NPOModel=Llama-3.1-8B2026.053434.30.0034056.8
GU+SimNPOModel=Qwen-3-8B2026.0533.344.30.1126.771.5
NPOModel=Llama-3.2-3B2026.0532.333.80.01534.347.3
MANSUModel=Llama-3.1-8B2026.0532.331.30.0139.158.6
SimNPOModel=Llama-3.1-8B2026.053230.70.0133951
MANSUModel=Qwen-3-4B2026.0531.428.70.02741.464.4
GU+SimNPOModel=Llama-3.1-8B2026.05303002329.7
MANSUModel=Qwen-2.5-4B2026.0529.126.80.02338.459.8
Surgical GAModel=Llama-3.1-8B2026.0528.3290.00736.748
SimNPOModel=Qwen-3-8B2026.052740.30.1332855.5
MANSUModel=Llama-3.2-3B2026.0525.423.40.0234.852.1
Global GAModel=Qwen-3-8B2026.0521.344.70.23324.771