List your company

Machine Unlearning on WMDP Cyber (test)

61.15MMLUOriginal
-2.20430414.24344830.691247.138952Aug 19, 2025Sep 15, 2025Oct 13, 2025Nov 10, 2025Dec 8, 2025Jan 5, 2026Feb 2, 2026

Evaluation results

MethodLinks
OriginalBackbone Model=Llama-3.1-8B2025.0861.15—————61.3240.95541.271.43
RMUBackbone Model=Llama-3.1-8B2025.0861.15—————52.4733.7550.681.23
DUETForgetting Data (Df)=D^query_f, Retention Data (Dr)=Dr2026.0160.65————26.6—————
Refusal TrainingForgetting Data (Df)=D^QR_f, Retention Data (Dr)=Dr2026.0160.63————40.92—————
GARetention Data (Dr)=Yes2026.0159.82————40.36—————
CRISPBackbone Model=Llama-3.1-8B2025.0858.86—————61.7429.38531.141.49
ELMBackbone Model=Llama-3.1-8B2025.0858.56—————58.9130.78530.991.4
Base ModelBackbone=Zephyr-7B2026.0158.12————43.68—————
RMURetention Data (Dr)=Yes2026.0157.81————26.93—————
NPORetention Data (Dr)=Yes2026.0157.11————39.61—————
NPO2026.0155.34————36.89—————
SimNPO2026.0154.25————34.22—————
GA2026.0148.79————33.77—————
OriginalBackbone Model=Gemma-2-2B2025.0846.3—————52.5733.9391.051.46
CRISPBackbone Model=Gemma-2-2B2025.0846.26—————49.0227.26380.811.28
RMUBackbone Model=Gemma-2-2B2025.0844.79—————44.7928.67360.641.23
ELMBackbone Model=Gemma-2-2B2025.0838.71—————43.3328.87290.761.36
RMU2026.0125.5————24.61—————
FLAT2026.0123.24————24.51—————
GA + KLRetention data=D_retain2026.020.59820.4036——5.36——————
Base modelBackbone=Zephyr 7B beta [51]2026.020.5810.44—————————
RMURetention data=D_retain2026.020.57810.2693——16.78——————
NPO + KLRetention data=D_retain2026.020.57110.3961——3.4——————
NPORetention data=None2026.020.55340.3689——4.35——————
SimNPORetention data=None2026.020.54250.3422——5.93——————
CATNIPRetention data=None2026.020.53010.2869——10.22——————
GARetention data=None2026.020.48790.3377——0.92——————
RMU*Retention data=None2026.020.2550.2461——-13.21——————
FLATContrastive data=D_ct2026.020.23240.2451——-15.37——————