arXiv:2607.25987v3 Announce Type: replace-cross Abstract: When a language model receives conflicting instructions from different priority levels, which one does it actually follow? This question lies at the heart of reliable LLM deployment. Existing benchmarks answer this only partially, often focusing on a single hierarchy component or adapting public datasets with limited tool-use coverage. We present IH-Benchmark, a conflict-centered benchmark for instruction-hierarchy robustness across direct System $\succ$ User (S $\succ$ U) conflicts and tool-mediated User $\succ$ Tool (U $\succ$ T) conflicts. We organize tests into seven constraint categories according to the higher-priority requirements under evaluation. Each test is evaluated with a uniform binary pass/fail protocol combining a predicate domain-specific language (DSL) with rubric-based LLM judges. We report category-balanced compliance rates and evaluate non-conflict controls separately. Our evaluation of 32 model configurations over five runs each reveals substantial variation in overall compliance, with scores ranging from 32.9% to 97.8%. Models achieve higher compliance on S $\succ$ U than on U $\succ$ T on average, at 84.6% and 69.6%, respectively. Strong performance on the former track does not guarantee comparable robustness on the latter. Constraint hardening provides a prompt-based defense, but its benefits depend on the model and how conflicting instructions are delivered. Several models also resist unauthorized tool actions more reliably than injected instructions that alter response language or factual content. Together, these results show that high aggregate compliance can conceal weaknesses in specific conflict tracks and instruction requirements. We release IH-Benchmark, including the benchmark corpus, agent simulators, and evaluation harness, at LINK_RELEASE_AFTER_PEER_REVIEW.
Continue to arXiv Software Engineering →You are being forwarded in 6s — stay on this page