ResearchBenchmarksBackdoor Attack Mitigation on Llama Location-agnostic adaptive attackFollow97Attack Success Rate (ASR)Unpatched-1.823.8549.575.15Jun 2, 2026Evaluation ResultsMethodMethodLinksAttack Success Rate (ASR)Accuracy (ACC)Unpatched2026.069794FTConstraint=l2Constraint=l22026.069692FT2026.069592FTConstraint=linfConstraint=linf2026.069591FinePruning2026.069591MEND2026.069591SPP2026.069391OneShot2026.069290ROME2026.069288BAERASER2026.068990Mudjacking2026.068592Patcher2026.06293