ResearchBenchmarksBackdoor Attack Mitigation on Llama Semantic adaptive attackFollow1Attack Success RatePatcher-1.8417.3336.555.67Jun 2, 2026Evaluation ResultsMethodMethodLinksAttack Success RateAccuracyPatcher2026.06193Mudjacking2026.066194BAERASER2026.066494OneShot2026.066693ROME2026.066793SPP2026.066894FT2026.067094FTConstraint=linfConstraint=linf2026.067093FinePruning2026.067094FTConstraint=l2Constraint=l22026.067194MEND2026.067194Unpatched2026.067294