ResearchBenchmarksHierarchical Instruction Following on IHEval Safety Defense Aligned settingFollow63.2Extract Accuracy (E)Vanilla54.3656.65558.9561.245Jun 29, 2026Evaluation ResultsMethodMethodLinksExtract Accuracy (E)Hijack Accuracy (H)Average AccuracyVanillaBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2026.0663.261.362.3ISEBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2026.0662.661.161.9IHDecBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2026.0654.753.954.3