跳转至

Eluna:用于自动化仓库运营的推理与任务执行智能体系统

文章背景与核心概要

Eluna 是一款专为自动化复杂仓库运营而设计的生产级智能体系统。在现代物流环境中,执行需要多系统决策逻辑且受严格时间限制的标准作业程序(SOP)是一项重大挑战。传统的 LLM 智能体往往难以处理过载的上下文信息并确保程序合规性,而 Eluna 通过引入图引导的多智能体框架有效解决了这一痛点。

该系统的核心在于将 SOP 编码为具有渐进式披露功能的有向无环图(DAG),并将任务委派给具备持久代码执行能力和实时数据访问权限的并行子智能体。为了优化性能,作者采用了非对称情景蒸馏技术,使较小的学生模型能够内化大型教师模型的推理能力,从而在不增加推理延迟的前提下实现高效的自动化执行。


摘要

Eluna 是一款生产部署的智能体系统,旨在自动化复杂的仓库运营。它解决了在严格时间限制下执行需要多系统决策逻辑的标准作业程序(SOP)这一关键挑战。与标准 LLM 智能体(通常在上下文过载和程序合规性方面表现不佳)不同,Eluna 利用了图引导的多智能体框架。通过将 SOP 编码为具有渐进式披露功能的有向无环图(DAG),该系统将任务委派给具有持久代码执行和实时数据访问权限的并行子智能体。为了优化性能,作者采用了非对称情景蒸馏,允许较小、高效的学生模型内化大型教师模型的推理能力,而无需产生额外的推理时间开销。

Eluna is a production-deployed agentic system designed to automate complex warehouse operations. It addresses the critical challenge of executing Standard Operating Procedures (SOPs) that require multi-system decision logic under strict time constraints. Unlike standard LLM agents, which often struggle with context overload and procedural compliance, Eluna utilizes a graph-guided, multi-agent framework. By encoding SOPs as directed acyclic graphs (DAGs) with progressive disclosure, the system delegates tasks to parallel sub-agents with persistent code execution and live data access. To optimize performance, the authors employ asymmetric episodic distillation, allowing smaller, efficient student models to internalize the reasoning of larger teacher models without incurring additional inference-time overhead.


关键特性

  • 图引导架构: SOP 被结构化为 DAG,确保了可靠的、循序渐进的程序合规性。
  • 多智能体委派: 独立任务由专门的并行子智能体处理,每个子智能体都配备了持久的代码执行和实时数据访问功能。
  • 非对称情景蒸馏: 一种新颖的训练方法,其中“学生”模型根据来自“教师”的修正轨迹进行微调,从而在没有大型模型延迟的情况下实现高水平推理。
  • 生产验证: 该系统已在 13 项任务基准测试和两个实际生产应用中得到验证,在工单处理方面达到了 94% 的专家一致性。
  • Graph-Guided Architecture: SOPs are structured as DAGs, ensuring reliable, step-by-step procedural compliance.
  • Multi-Agent Delegation: Independent tasks are handled by specialized parallel sub-agents, each equipped with persistent code execution and real-time data access.
  • Asymmetric Episodic Distillation: A novel training method where a "student" model is fine-tuned on corrected trajectories from a "teacher," enabling high-level reasoning without the latency of larger models.
  • Production-Proven: The system has been validated on a 13-task benchmark and two real-world production applications, achieving 94% expert agreement in ticket processing.

技术元数据

属性 详情
提交日期 2026年7月9日
最后修订 2026年8月24日
DOI https://doi.org/10.48550/arXiv.2607.08960
许可协议 license icon
Attribute Details
Submitted 9 Jul 2026
Last Revised 24 Aug 2026
DOI https://doi.org/10.48550/arXiv.2607.08960
License license icon

访问论文

Access Paper