跳转至

全局连贯性困境：当每个智能体都正确而团队依然出错

Global Coherence: When Every Agent Is Right and the Team Is Still Wrong

多智能体协同的局部到全局语义基石

A Local-to-Global Semantic Foundation for Multi-Agent Collaboration

Authors: Xin Heng
Primary Subject: Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2610.02036 [cs.AI]
Submitted: October 1, 2026

文章背景与核心概要

在当今的大语言模型 (Large Language Model, LLM) 应用中，多智能体协同已成为应对复杂任务的核心范式，但系统在实际运行中常常遭遇一种难以捉摸的困境：每一个 AI 智能体 (AI Agent) 依据自身视角做出的决策都完全合理，但团队汇总出的最终结果却漏洞百出。针对这一长期困扰业界的难题，学者 Xin Heng 在本篇论文中深入剖析了这一“全局连贯性 (Global Coherence) ”问题，并给出了严格的数学证明——“观测混淆不可能定理 (Observation-Aliasing Impossibility Theorem) ”，指出在观测上下文缺失的情况下，单纯增加推理步数、引入更多角色或扩充通信轮次，在理论上根本无法消除错误。为此，论文基于拓扑学与层论提出了“局部到全局运行时语义框架” (\(\mathcal{X}\))，确立了“模型仅负责提议行动、管控底座全权掌控共享状态与提交”的设计准则。该研究为多智能体系统从松散对话迈向高可靠工程落地奠定了严谨的理论基石。

执行摘要

Executive Summary

多智能体 AI 系统常常深陷一个致命的系统性缺陷：每个独立的智能体都能做出局部完全合理的正确决策，但团队协同起来却得出一个全盘错误的荒诞结果。

Multi-agent AI systems frequently suffer from a critical systemic flaw: individual agents can make locally valid decisions while the team jointly produces an invalid result.

在这篇论文中，Xin Heng 对这一现象——即 全局连贯性问题 (global coherence problem) ——展开了深入探究。作者指出，这种故障的根源在于共享状态管理的底层失效，而非模型本身不够聪明。为了厘清系统能力的严格理论边界，作者提出了 观测混淆不可能定理 (Observation-Aliasing Impossibility Theorem) ，从数学上证明了局部层面的小修小补、增加思考推理步数或在智能体之间发送更多消息，都无法弥补因观测上下文不全而带来的天然盲区。

In this paper, Xin Heng investigates this phenomenon—termed the global coherence problem—arguing that it represents a fundamental failure of shared state management rather than a deficiency in model intelligence. To establish strict theoretical boundaries, the author introduces the Observation-Aliasing Impossibility Theorem, proving that localized improvements, increased reasoning steps, or additional messaging cannot overcome incomplete observational context.

为破解这一死局，论文提出了一套严谨的 从局部到全局运行时语义框架 (\(\mathcal{X}\)) 。大量实证评估充分证明：采用具备全局管控能力的运行底座 (harness) 来统一协调状态，其可靠性远远优于单纯依赖去中心化智能体之间的相互通信。

To resolve this, the paper presents a robust local-to-global runtime semantics framework (\(\mathcal{X}\)), backed by empirical evaluations demonstrating that enforcing global state harnesses vastly outperforms reliance on decentralized agent communication.

核心理论贡献

Key Theoretical Contributions

1. 全局连贯性难题

1. The Global Coherence Problem

尽管单体大语言模型智能体在独立的局部上下文中往往表现优异，但在由它们组成的多智能体架构中，全局层面的溃败却屡见不鲜。本文的核心论点十分鲜明：如果底层的观测架构本身存在缺陷，单纯增加推理思考、划分更多角色、扩充消息传递或抽取更多样本，都无法挽回缺失的关键区分信息。越强大的模型在其视野所及的上下文中确实能推理得更好，但从结构上看，它们依然天生无法“看穿”视野之外的事物。

While individual Large Language Model (LLM) agents often excel within isolated contexts, multi-agent frameworks regularly fail globally. The core thesis of the paper is that more reasoning, roles, messages, or samples cannot recover missing distinctions if the underlying observational architecture is flawed. Stronger models reason better within their context, but they remain structurally incapable of seeing beyond it.

2. 观测混淆不可能定理

2. Observation-Aliasing Impossibility Theorem

论文形式化界定了从局部走向全局决策的精确理论边界：

The paper formalizes the exact boundaries of local-to-global decision-making:

  • 一项策略只有在所有产生相同观测的环境世界都拥有至少一个共同可行行动时，才能保证做出有效动作。
  • 如果 \(k\) 个在观测上无法区分的世界各自需要完全不相交的行动方案，那么在最坏情况下，最佳随机化策略的成功率将断崖式下跌至 \(1/k\)。
  • 无论怎样提升模型的智力水平、增加智能体角色分工，或是加深通信交互深度，都无法突破这一底层物理极限。
  • A policy can guarantee a valid action only when all environmental worlds producing the same observation share an admissible action.
  • If \(k\) indistinguishable worlds require pairwise-disjoint actions, the optimal randomized worst-case success rate drops to \(1/k\).
  • Amplifying model intelligence, agent roles, or communication depth cannot bypass this fundamental limitation.

3. 从局部到全局的运行时语义 (\(\mathcal{X} = (H, C, G, F; D)\))

3. Local-to-Global Runtime Semantics (\(\mathcal{X} = (H, C, G, F; D)\))

为了在局部智能体的灵活思考与全局系统的绝对正确之间搭建桥梁，作者提出了一套形式化的运行时语义模型，由五大核心要素构成：

To bridge local agent intelligence with global correctness, the author proposes a formal runtime semantic model comprising five core components:

  • Topology (\(H\))： 记录互相重叠的观测视野范围。
  • Category (\(C\))： 规范改变系统状态的各项操作动作。
  • Groupoid (\(G\))： 保留可逆的转换映射。
  • Sheaf (\(F\))： 检验各个局部视角能否自洽地粘合拼装为一个统一连贯的全局世界。
  • Minimal History (\(D\))： 严格仅保留那些会直接改变未来合法状态的关键历史差异。
  • Topology (\(H\)): Records overlapping observational scopes.
  • Category (\(C\)): Governs state-changing actions.
  • Groupoid (\(G\)): Retains reversible translations.
  • Sheaf (\(F\)): Tests whether local views consistently glue together into a single coherent world.
  • Minimal History (\(D\)): Retains strictly the distinctions that alter legal future states.

设计准则： 模型仅负责提议行动，而管控底座全权掌控共享状态并主导提交阶段。

💬 [原文引用 / Original Quote]: Design Principle: Models propose actions, while the harness owns the shared state and governs the commit phase.

实证发现与基准测试

Empirical Findings & Benchmarks

论文通过九项独立研究，系统评估了理论失效模式以及所提出的运行时管控底座：

The paper evaluates both the theoretical failure modes and the proposed runtime harness across nine distinct studies:

  • 受控修订基准测试 (Controlled Revision Benchmark)：
  • 当关键决定性事件完全可见时，前沿模型的得分高达 40/40。
  • 当这些关键事件被隐去时，受测配置的得分骤降至 12–17/40 (与 \(1/3\) 的随机盲猜概率高度一致) 。
  • 仅需恢复一条权威的关键事实，模型性能便立刻满血回升至 40/40。
  • TeamBench：
  • 普通多智能体团队在 5/5 次运行中全部超出了共享预算配额。
  • 即便向智能体提供清晰可见的实时计数器，依然在 4/5 次运行中发生了违规超支。
  • 引入严格的提交管控底座后，违规行为被彻底杜绝 (0/5 次运行)。
  • \(\tau_2\)-bench 电信场景 (Telecom)：
  • 在发生智能体静默回滚的情况下，依赖当前状态查询的方法得分断崖式下跌至 0.07。
  • 与之形成鲜明对比的是，论文提出的管控底座框架始终保持 1.00 的满分表现。
  • 在传统求解器已经完整管理了全部相关状态的测试场景中，其性能与管控底座打成平手，这与理论预测完全吻合。
  • Controlled Revision Benchmark:
  • When deciding events are fully visible, frontier models score 40/40.
  • When those events are hidden, tested configurations drop to 12–17/40 (aligning with random chance at \(1/3\)).
  • Restoring just a single authoritative fact immediately returns performance to 40/40.
  • TeamBench:
  • Ordinary multi-agent teams exceed shared budgets in 5/5 runs.
  • Providing a visible live count still results in violations in 4/5 runs.
  • Enforcing strict commit harnesses eliminates violations entirely (0/5 runs).
  • \(\tau_2\)-bench Telecom:
  • Current-state checks drop to a score of 0.07 following silent agent reverts.
  • Conversely, the proposed harness framework maintains a perfect score of 1.00.
  • In scenarios where a conventional solver already manages the complete relevant state, performance ties with the harness precisely as theoretically predicted.

结语

Conclusion

这项研究得出了一个发人深省的反直觉结论：局部的个体智能永远无法替代缺失的全局状态。 想要打造真正可靠的多智能体协同系统，我们的研发重心就必须从单纯堆叠模型推理能力的“大力出奇迹”，转向构建严谨、具备形式化语义保证的基础架构，从根本上管控好共享全局状态与观测语义。

The counterintuitive conclusion of this research is that local intelligence cannot substitute for missing global state. Achieving reliable multi-agent collaboration requires shifting focus away from purely scaling model reasoning capabilities and toward rigorous, formally grounded architectures that govern shared global state and observational semantics.