benchmarksfactbearishThe strongest current language models only marginally exceed 50% prompt-level accuracy on hierarchical instruction-following tasksComputation and Language01 Aug 2026http://arxiv.org/abs/2607.27912v1