LLM 智能体记忆与技能跨源冲突裁决基准
原标题：Navigating Epistemic Parity in LLM Agents: A Benchmark for Cross-Source Conflict Resolution Between Memory and Procedural Skills
AI 摘要
该论文提出一个基准，用于衡量 LLM 智能体在“记忆”与“程序性技能”两类上下文来源发生冲突时如何裁决，并将这种无优先级的状态称为“epistemic parity”。作者构建了事实、格式、程序、权限四类冲突分类，实例化 52 个场景，在三个开源权重模型上运行 1,383 次测试。结果显示裁决结果取决于冲突类型、模型和注入顺序，而非来源语义；显式优先级指令只能部分生效；89% 的运行未向用户披露冲突，静默裁决成为默认行为。
正文节选
Navigating Epistemic Parity in LLM Agents: A Benchmark for Cross-Source Conflict Resolution Between Memory and Procedural Skills Abstract Production LLM agents increasingly run two context-provisioning systems in tandem: episodic memory pipelines that inject persistent user preferences and historical state, and procedural skill modules (e.g., SKILL.md files) that inject standardized, developer-defined operating procedures. Because these two sources are injected into the same context region with