跳转至

文章背景与核心概要

由大语言模型（LLMs）驱动的多智能体系统（MAS）通过结构化工作流协调多个专业智能体，以实现高级推理能力。然而，多步执行和重复的模型调用带来了显著的推理延迟。

现有的编排框架主要集中在优化任务性能和经济成本上，很大程度上忽视了端到端延迟。由于系统延迟根本上受制于关键执行路径，仅仅降低整体执行成本并不一定能减少延迟。此外，朴素的优化方法存在错误分配操作符级贡献度的风险，这可能会严重降低任务准确率。为了解决这些挑战，作者推出了 LAMaS（Latency-Aware Multi-agent System，延迟感知多智能体系统），这是一个专门为基于学习的多智能体系统设计的新型编排框架。

Learning Latency-Aware Orchestration for Multi-Agent Systems

arXiv: 2601.10560 [cs.MA]
DOI: 10.48550/arXiv.2601.10560
Authors: Xi Shi, Mengxin Zheng, Qian Lou
Submitted: 15 January 2026 (Last revised: 12 August 2026)
Primary Subject: Multiagent Systems (cs.MA)
Secondary Subjects: Artificial Intelligence (cs.AI), Computation and Language (cs.CL)

arXiv: 2601.10560 [cs.MA]
DOI: 10.48550/arXiv.2601.10560
Authors: Xi Shi, Mengxin Zheng, Qian Lou
Submitted: 15 January 2026 (Last revised: 12 August 2026)
Primary Subject: Multiagent Systems (cs.MA)
Secondary Subjects: Artificial Intelligence (cs.AI), Computation and Language (cs.CL)

📋 Executive Summary

Multi-agent systems (MAS) driven by Large Language Models (LLMs) coordinate multiple specialized agents via structured workflows to achieve advanced reasoning capabilities. However, multi-step execution and repeated model invocations introduce significant inference latency.

Existing orchestration frameworks predominantly focus on optimizing task performance and financial cost, leaving end-to-end latency largely unaddressed. Because system latency is fundamentally governed by the critical execution path, merely reducing overall execution cost does not necessarily reduce latency. Furthermore, naive optimization approaches risk misassigning operator-level credits, which can severely degrade task accuracy.

To resolve these challenges, the authors introduce LAMaS (Latency-Aware Multi-a}gent S**ystem), a novel orchestration framework specifically designed for learning-based multi-agent systems.

📋 Executive Summary

Multi-agent systems (MAS) driven by Large Language Models (LLMs) coordinate multiple specialized agents via structured workflows to achieve advanced reasoning capabilities. However, multi-step execution and repeated model invocations introduce significant inference latency.

Existing orchestration frameworks predominantly focus on optimizing task performance and financial cost, leaving end-to-end latency largely unaddressed. Because system latency is fundamentally governed by the critical execution path, merely reducing overall execution cost does not necessarily reduce latency. Furthermore, naive optimization approaches risk misassigning operator-level credits, which can severely degrade task accuracy.

To resolve these challenges, the authors introduce LAMaS (Latency-Aware Multi-a}gent S**ystem), a novel orchestration framework specifically designed for learning-based multi-agent systems.

⚙️ How LAMaS Works

LAMaS tackles latency optimization across two distinct operational phases:

  1. Training Time (Latency-Aware Execution Graphs):
    It learns optimized execution workflows using constrained optimization paired with critical-path-aware credit assignment, ensuring that accuracy is fully preserved while minimizing latency bottlenecks.
  2. Inference Time (Adaptive Lightweight Controller):
    Recognizing that statically committed training graphs cannot react to dynamic runtime information, LAMaS pairs graph construction with a lightweight runtime controller. This controller adaptively prunes and eliminates redundant agent interactions on the fly as execution unfolds.

⚙️ How LAMaS Works

LAMaS tackles latency optimization across two distinct operational phases:

  1. Training Time (Latency-Aware Execution Graphs):
    It learns optimized execution workflows using constrained optimization paired with critical-path-aware credit assignment, ensuring that accuracy is fully preserved while minimizing latency bottlenecks.
  2. Inference Time (Adaptive Lightweight Controller):
    Recognizing that statically committed training graphs cannot react to dynamic runtime information, LAMaS pairs graph construction with a lightweight runtime controller. This controller adaptively prunes and eliminates redundant agent interactions on the fly as execution unfolds.

📊 Experimental Results & Impact

  • Performance: Evaluated across four distinct benchmarks, LAMaS achieves state-of-the-art latency results compared to existing learning-based MAS baselines.
  • Latency Reduction: Cuts end-to-end latency by over 50%.
  • Accuracy: Maintains competitive or superior task accuracy despite the massive latency reduction.
  • Modularity: Features a highly modular design that transfers seamlessly to other multi-agent systems with minimal adjustments, consistently delivering notable latency improvements.

📊 Experimental Results & Impact

  • Performance: Evaluated across four distinct benchmarks, LAMaS achieves state-of-the-art latency results compared to existing learning-based MAS baselines.
  • Latency Reduction: Cuts end-to-end latency by over 50%.
  • Accuracy: Maintains competitive or superior task accuracy despite the massive latency reduction.
  • Modularity: Features a highly modular design that transfers seamlessly to other multi-agent systems with minimal adjustments, consistently delivering notable latency improvements.

🔗 Full-Text & Resources

🔗 Full-Text & Resources

license icon

license icon