ResearchBenchmarksSemantic Hijacking on MultiWOZ Semantic Hijacking 2.2Follow10Vanilla APAVanilla9.59.751010.25Apr 27, 2026Evaluation ResultsMethodMethodLinksVanilla APASSRP APALift (L_r)p-valueVanillaModel Pair (Architect...Model Pair (Architect / Executive)=GPT 5.4 / 5.4 mini2026.0410———SSRPModel Pair (Architect...Model Pair (Architect / Executive)=GPT 5.4 / 5.4 mini2026.04—71.671,5001