Where Instruction Hierarchy Breaks: Diagnosing and Repairing Failures in Reasoning Language Models

In three linesDiagnostic study of instruction hierarchy failures in reasoning models (Gemma-4-31B-IT, Qwen3.6-35B-A3B, Claude Sonnet 4.6). White-box framework localizes failures into instruction identification, conflict resolution, and response realization. Two training-free self-monitoring mechanisms reduce rule-following non-compliance by 81-99% across models.

Summary generated by Claude — human-verified