跳转至

文章背景与核心概要

分布式大模型(LLM)智能体团队在协作过程中经常面临一种隐蔽的失效模式:智能体虽然能够读取到最新的共享事实,但其执行的动作却依然基于过时的计划。本文识别并深入分析了“陈旧计划执行”(stale-plan execution)这一现象,即状态的新鲜并不代表授权该动作的计划依然有效。

为了解决这一问题,作者引入了 PlanFence——一种依赖域动作验证协议。PlanFence 确保智能体在计划中明确引用所依据的具体公共记录,从而使执行器能够仅验证影响待处理外部动作的相关状态记录,而无需刷新整个共享状态,兼顾了系统安全性与运行效率。


记忆常新,计划陈旧:分布式大模型智能体记忆的依赖域验证 (Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory)

作者: Evan Chen, Shiqiang Wang, Christopher G. Brinton
提交时间: 2026年9月3日
主要学科: 人工智能 (cs.AI)
arXiv ID: 2609.03340 | DOI: 10.48550/arXiv.2609.03340


执行摘要 (Executive Summary)

分布式大模型(LLM)智能体团队常常遭遇一种微妙的失效模式:智能体可以读取最新的共享事实,但依然基于过时的计划执行动作。本文识别并分析了陈旧计划执行(stale-plan execution)这一现象——即状态的新鲜度并不能保证授权动作的计划依然有效。

为了解决这个问题,作者引入了 PlanFence,一种依赖域动作验证协议。PlanFence 确保智能体显式引用其计划中使用的确切公共记录,从而使执行器能够仅验证影响待处理外部动作的相关状态记录,而不必刷新整个共享状态。

Executive Summary

Distributed Large Language Model (LLM) agent teams often suffer from a subtle failure mode: agents can read the latest shared facts yet still execute actions based on obsolete plans. This paper identifies and analyzes stale-plan execution—the phenomenon where state freshness does not guarantee that the plan authorizing an action remains valid.

To solve this, the authors introduce PlanFence, a dependency-scoped action-validation protocol. PlanFence ensures that agents explicitly cite the exact public records used in their plans, allowing executors to validate only the relevant state records affecting pending external actions rather than refreshing the entire shared state.


核心概念与摘要概述 (Key Concepts & Abstract Overview)

  • 问题所在(陈旧计划执行): 规划器可能根据需求 \(r_3\) 衍生出一个动作,而另一个智能体同时提交了一个冲突的需求 \(r_4\)。如果执行器在没有替换原本由 \(r_3\) 衍生的计划的情况下接收了 \(r_4\),那么尽管它能够访问最新的系统内存,它依然会基于无效的假设采取行动。
  • 解决方案(PlanFence): 这是一种协议,其中生成的计划引用其所依赖的特定公共记录。执行器验证那些能够影响待处理外部动作的记录,当验证不完整时,将触发重新规划或阻止执行。

Key Concepts & Abstract Overview

  • The Problem (Stale-Plan Execution): A planner may derive an action from requirement \(r_3\), while another agent concurrently commits a conflicting requirement \(r_4\). If an executor receives \(r_4\) without replacing the plan originally derived from \(r_3\), it acts on invalid assumptions despite having access to fresh system memory.
  • The Solution (PlanFence): A protocol where generated plans reference the specific public records they depend upon. Executors validate only the records capable of affecting the pending external action, triggering a re-plan or blocking execution when validation is incomplete.

实验洞察 (Experimental Insights)

  • 性能与安全性: 在包含计划后修订的 30 个受控实时工作流中,传统的仅新鲜度执行器在每一项任务中都持续基于过时计划采取行动。相反,PlanFence 完成了 100% 的任务,且没有产生单一的无效动作。
  • 条件边界(系统成本): 受控重放实验突出了明显的运维权衡:
  • 当系统变动(churn)较低时,主动同步可带来较低的协调停滞。
  • 随着变动增加,PlanFence 表现卓越——避免了重复的更新路径协调,并随着共享键空间的扩展防止了对不相关状态的不必要验证。
  • 注: 这些发现突出了受控安全性和系统成本优化,而非通用的任务准确率提升。

Experimental Insights

  • Performance & Safety: In 30 controlled live workflows featuring post-plan revisions, a traditional freshness-only executor consistently acted on obsolete plans in every single task. Conversely, PlanFence completed 100% of tasks without a single invalid action.
  • Conditional Boundaries (Systems Costs): Controlled replay experiments highlight distinct operational trade-offs:
  • Proactive synchronization yields lower coordination stall when system churn is low.
  • PlanFence excels as churn increases—avoiding repeated update-path coordination and preventing the unnecessary validation of unrelated state as the shared keyspace scales.
  • Note: These findings highlight controlled safety and systems-cost optimizations rather than generalized task-accuracy gains.