ResearchBenchmarksHierarchical Instruction Following on IHEval Safety Defense Conflict settingFollow50.5Extract Accuracy (E)IHDec27.41233.40639.445.394Jun 29, 2026Evaluation ResultsMethodMethodLinksExtract Accuracy (E)Hijack Accuracy (H)Average AccuracyIHDecBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2026.0650.55753.8VanillaBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2026.0633.424.929.2ISEBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2026.0628.328.928.6