跳转至

GURU:面向高可靠数字工程与运维的神经符号分层多智能体系统

GURU: A Hierarchical Neuro-Symbolic Multi-Agent System for Verification-Aware Automation of Mission-Critical Digital Engineering and Operations

arXiv:2611.01001 [cs.AI]
Submitted: 31 October 2026
Authors: Allan Grosvenor, Anton Zemlyansky, Abdul Wahab, Kyrylo Bohachov, Aras Dogan, Ryland Adams, Dwyer Deighan


文章背景与核心概要

在航空航天、复杂计算流体动力学与空间态势感知等关键任务场景中,传统黑盒驱动的大语言模型 (Large Language Model, LLM) 往往因输出幻觉与不可控的推断而无法直接用于安全决策。为此,本论文提出了 GURU——一个具备严格验证意识的神经符号分层多智能体系统 (Hierarchical Neuro-Symbolic Multi-Agent System) 。GURU 创新性地确立了权责严格划分的闭环控制机制:由神经网络模型负责提议并加速动作,由符号规则引擎在动作落地前强制核验安全硬边界,最终由经过校验的高保真物理仿真引擎对系统状态做出终局判定。实验表明,该系统在未见过的复杂飞行器几何构型仿真收敛率上实现了从不足 1% 到 74% 的质的飞跃,并在超级计算机集群上展现出可扩展至 8000 个计算节点的强大工程落地能力。


执行摘要

Executive Summary

“AI 能够比任何人类都更快地找出安全行动方案,但它绝不能由自己来界定何为安全标准。”

💬 [原文引用 / Original Quote]: “An AI can find a safe action faster than any human, but it must never decide what counts as safe.”

GURU 是一套专为关键任务数字工程与运维量身打造的分层神经符号多智能体系统。它严格实行权责隔离制度:

GURU is a hierarchical neuro-symbolic multi-agent system designed for mission-critical digital engineering and operations. It enforces a strict partition of authority:

  1. 学习型模型 (Learned models):负责提议、排列优先级并加速候选行动;
  2. 符号规则引擎 (Symbolic rules engine):在执行任何关键动作之前,依据严苛的硬性边界指标逐项核验;
  3. 真实物理验证 (Validated physics):对最终系统状态进行校准并作出权威裁定。
  1. Learned models propose, rank, and accelerate actions.
  2. A symbolic rules engine checks every consequential action against hard limits prior to execution.
  3. Validated physics confirms and decides the final state.

GURU 作为闭环控制系统运转,调度各司其职的专业 AI 智能体 (负责导航寻径的 Navigation、负责执行操作的 Action 与负责规划策略的 Planning) ,各个智能体之间共享统一的内部通用语言 (Common Internal Language) 与一体化模型训练平台。

Operating as a closed control loop, GURU utilizes specialized agents (Navigation, Action, and Planning) that share a Common Internal Language and a unified training factory.


核心性能与实验结果

Key Performance & Experimental Results

  • 计算流体力学 (Computational Fluid Dynamics, CFD) :
  • 在 100 个单次尝试的基准测试用例中,将此前从未见过的复杂飞行器几何构型仿真收敛率从不足 1% 大幅跃升至 74%;
  • 跨越三处研究设施,成功实测完成了最高达 6 马赫 (Mach 6) 的 54 个真实工况用例;
  • 消融实验证实,符号规则层的加入直接带来了边界层捕捉能力 13 个百分点的提升;
  • 依托知识图谱检索与角色适配器技术,将 ParaView 科学可视化工具的控制成功率从 55% 提升至 75%。

  • 空间态势感知 (Space Domain Awareness) :

  • 针对大语言模型在关键安全决策中缺乏可信度的痛点,给出了高可靠的自主工程解决方案;
  • 采用冻结权重的自监督世界模型与经严格校验的物理轨道外推相结合,在空间交会碰撞风险评估中斩获五种参评方法中的最高得分 (性能显著优于全面微调的大语言模型) ,同时展现出极低的随机种子方差。

  • 系统扩展性与现实工程落地:

  • 在 2026 年的一次航天器再入大气层真实事件中,分阶段完整验证了从雷达探测到轨道机动避让的全流程空间碎片防御链条;
  • 在高保真仿真环境中成功协同指挥了包含 200 架飞行器的大规模集群编组;
  • 模型训练规模在突破 1,000 个 Aurora 超算节点后依然展现出近乎线性的扩展效率,并成功在 8,000 个计算节点上实现超大规模集群稳定运行;
  • 该架构能自然拓展至控制级微型核反应堆数字孪生系统,始终坚持以经过严谨验证的真实物理法则 (而非纯粹基于黑盒学习的经验拟合) 作为不可动摇的最高安全基石。
  • Computational Fluid Dynamics (CFD):
  • Raised the converged-simulation rate on unseen aircraft geometries from under 1% to 74% over 100 single-attempt cases.
  • Successfully ran 54 live cases up to Mach 6 across three facilities.
  • Ablation studies demonstrated that the symbolic layer accounts for a 13-point increase in boundary-layer capture.
  • Knowledge-graph retrieval and role adapters improved ParaView control success from 55% to 75%.

  • Space Domain Awareness:

  • Addressed scenarios where language models cannot be trusted with autonomous decision-making.
  • A frozen, self-supervised world model coupled with validated propagation scored highest among five methods for conjunction risk assessment (outperforming fully fine-tuned language models) while exhibiting significantly lower seed variance.

  • Systems Scalability & Real-World Application:

  • Demonstrated a radar-to-avoidance debris pipeline in stages during a 2026 reentry event.
  • Coordinated simulated teams of 200 aircraft.
  • Training scaled near-linearly past 1,000 Aurora nodes and successfully executed at scale on 8,000 nodes.
  • Extends naturally to control-grade microreactor digital twins, utilizing validated physics—rather than learned approximations—as the ultimate safety basis.

附加元数据

Additional Metadata

  • 主要学科领域: 人工智能 (cs.AI)
  • 次要学科领域: 计算工程、金融与科学 (cs.CE);多智能体系统 (cs.MA);系统与控制 (eess.SY)
  • 文档技术规格: 82 页,50 张图表,7 张表格
  • 全文获取与授权: 查看 PDF 原文 | 授权协议 (CC BY 4.0)
  • Primary Subject: Artificial Intelligence (cs.AI)
  • Secondary Subjects: Computational Engineering, Finance, and Science (cs.CE); Multiagent Systems (cs.MA); Systems and Control (eess.SY)
  • Document Specifications: 82 pages, 50 figures, 7 tables
  • Full-Text & Access: View PDF | License (CC BY 4.0)

注:如需了解系统完整实现细节、架构拓扑设计以及数学公式推导,请参阅 arXiv 官方发表论文。

Note: For complete implementation details, architectural schemas, and mathematical formulations, please refer to the official arXiv publication.