List your company

Machine Unlearning on MUSE (Df/Dr Sets)

58.4VerbMem (Df)Original LLM
-2.33613.43229.244.968Sep 29, 2025

Evaluation results

MethodLinks
Original LLMModel=LLaMA-2-7B2025.0958.4—63.9—55.2—-99.8———
Task VectorModel=LLaMA-2-7B, Method Group=Representation-Based Methods2025.0956.3—63.7—54.6—-99.8———
MismatchModel=LLaMA-2-7B, Method Group=Representation-Based Methods2025.0942.8—52.6—45.7—-99.8———
KLModel=LLaMA-2-7B, Method Group=Gradient-Based Methods2025.0927.4—50.2—44.8—-96.1———
Retained LLMModel=LLaMA-2-7B2025.0920.8—33.1—55—0———
WHPModel=LLaMA-2-7B, Method Group=Representation-Based Methods2025.0919.7—21.2—28.3—109.6———
GDModel=LLaMA-2-7B, Method Group=Representation-Based Methods2025.094.9—27.5—6.7—109.4———
FLAT (JS)Model=LLaMA-2-7B, Method Group=FLAT Methods2025.091.9—36.2—38.5—47.1———
FLAT (TV)Model=LLaMA-2-7B, Method Group=FLAT Methods2025.091.7—13.6—31.8—45.4———
FLAT (Pearson)Model=LLaMA-2-7B, Method Group=FLAT Methods2025.091.6—0—0.2—26.8———
NPO-RTModel=LLaMA-2-7B, Method Group=Gradient-Based Methods2025.091.2—54.6—40.5—105.8———
OURS (GD+Sine)Model=LLaMA-2-7B2025.090.8—5.2—42.1—8.3———
GAModel=LLaMA-2-7B, Method Group=Gradient-Based Methods2025.090—0—0—17———
NPOModel=LLaMA-2-7B, Method Group=Gradient-Based Methods2025.090—0—0—15———
FLAT (KL)Model=LLaMA-2-7B, Method Group=FLAT Methods2025.090—0—0—58.9———
Gold-standard retrained2026.04————————0.56—
GradAscentretain free=true, Time=1117s2026.04————————0.4770.533
GradDiffretain free=false, Time=2928s2026.04————————0.4360.521
MC-WIN-Uretain free=true, Time=392s2026.04————————0.4290.529
NPOretain free=false, Time=3192s2026.04————————0.550.533
Original model2026.04————————0.555—
RMUretain free=false, Time=1897s2026.04————————0.4960.529
SimNPOretain free=false, Time=3066s2026.04————————0.5270.534