ResearchBenchmarksMachine Unlearning on WMDP Cyber (WMDP Acc, MMLU, Adj-MMLU)Follow44.69WMDP AccuracyBase23.754829.189934.62540.0601Jun 9, 2026Evaluation ResultsMethodMethodLinksWMDP AccuracyMMLU Overall ScoreAdjusted MMLUBase2026.0644.6957.7255.33NSRU2026.0627.7357.4952.33TRU2026.0627.4835.8632GradAscent2026.0626.5722.9526.33RMU2026.0626.5722.9526.33NPO2026.0624.8625.8224.33GradDiff2026.0624.5625.5129.67