上下文守护：语言模型中上下文学习的结构化自审计方法 / ContextGuard: Structured Self-Auditing for Context Learning in Language Models
1️⃣ 一句话总结
本文提出一种名为ContextGuard的结构化自审计框架，通过让大语言模型在执行任务时主动检查并纠正上下文中的隐性约束（如格式要求、边缘细节等），从而显著提升模型对复杂上下文指令的忠实执行能力。
arXiv 提交日期: 2026-05-26
Abstract - ContextGuard: Structured Self-Auditing for Context Learning in Language Models
Recent benchmarks reveal that despite strong reasoning capabilities, large language models (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.
源自 arXiv: 2605.26827