Researchers have developed a benchmark called XIH-Bench to evaluate how well language models follow instruction hierarchy rules. This is important because it ensures that higher-priority instructions override lower-priority ones in multilingual settings. The study found two key patterns: first, the effectiveness of instruction hierarchy compliance varies depending on the language used; and second, conflicts between languages can actually improve compliance compared to conflicts within the same language. However, this 'Language Boundary Effect' also creates risks for model reliability and security. --- Why it matters: Understanding how language models follow instruction hierarchy rules is crucial for safe and controllable deployment in multilingual settings. Engineers working on AI systems need to consider these findings to ensure their models can handle diverse languages and instructions effectively. Source: https://arxiv.org/abs/2607.23545