ContextGuard: Structured Self-Auditing for Context Learning in Language Models

Abstract: Recent benchmarks reveal that despite strong reasoning capabilities, LLMs still struggle to faithfully apply complex contextual knowledge. These failures are often not wholesale reasoning collapses: in context-rich tasks, models may follow the central reasoning path while missing peripheral, persistent, or format-sensitive requirements.