跳转至

文章背景与核心概要

车辆路径问题（VRP）是组合优化领域的基石性挑战，在现实世界中有着广泛的应用。虽然先进的优化求解器能够高效解决这些问题，但为求解器构建复杂的 VRP 变体通常需要深厚的领域专业知识，这阻碍了先进优化工具的普及和易用性。

为了弥补这一鸿沟，本文引入了强化学习增强的大语言模型智能体（RLEA），这是一个专为自动化建模复杂 VRP 而设计的多智能体框架。通过集成一个通过软 Q 学习（Soft Q-learning）训练的轻量级神经规划器来协调基于 LLM 的智能体，并辅以进化记忆模块和检索增强生成（RAG），RLEA 成功地利用了积累的历史经验和外部求解器知识。在 48 种不同的 VRP 变体上进行评估时，RLEA 的成功率比当前最先进的方法高出 16.67%，并大幅减少了运行时错误。

Reinforcement Learning Enhanced LLM Agents for Complex Vehicle Routing Problems

arXiv: 2609.00859 [cs.AI]
Submitted on: 1 September 2026
Authors: Yi Chen, Zikang Yu, Jiahai Wang, Jinbiao Chen, Jianpeng Zhou, Zizhen Zhang

📌 Summary

Vehicle Routing Problems (VRPs) are foundational combinatorial optimization challenges with wide-ranging real-world applications. While advanced optimization solvers can solve these problems effectively, formulating complex VRP variants for solvers typically demands extensive domain expertise, hindering the accessibility of advanced optimization tools.

To bridge this gap, this paper introduces Reinforcement Learning Enhanced LLM Agents (RLEA), a multi-agent framework designed to automate the modeling of complex VRPs. By integrating a lightweight neural Planner (trained via Soft Q-learning) to orchestrate LLM-based agents alongside an evolutionary memory module and retrieval-augmented generation (RAG), RLEA successfully leverages accumulated past experience and external solver knowledge. Evaluated across 48 distinct VRP variants, RLEA outperforms current state-of-the-art methods by achieving a 16.67% higher success rate and substantially reducing runtime errors.

🛠️ 核心组件与方法论

🛠️ Key Components & Methodology

  • 轻量级神经规划器： 采用软 Q 学习进行训练，以高效协调和指导基于 LLM 的智能体的行动。
  • 进化记忆模块： 使系统能够随着时间的推移保留、建立和完善积累的经验。
  • 检索增强生成（RAG）： 在程序生成和优化阶段整合外部求解器知识，以确保建模的准确性。
  • Lightweight Neural Planner: Trained with Soft Q-learning to efficiently coordinate and guide the actions of LLM-based agents.
  • Evolutionary Memory Module: Enables the system to retain, build upon, and refine accumulated experiences over time.
  • Retrieval-Augmented Generation (RAG): Integrates external solver knowledge during the program generation and refinement phases to ensure accurate modeling.

📊 实验亮点

📊 Experimental Highlights

  • 评估范围： 使用各种先进的求解器在 48 种不同的 VRP 变体上进行了测试。
  • 性能表现： 超越了以前最先进的方法，成功率提升了 16.67%，并显著减少了运行时错误。
  • 核心启示： 表明将强化学习与基于 LLM 的推理相结合，能够为自动化优化建模提供一种高度鲁棒的方法。
  • Evaluation Scope: Tested across 48 distinct VRP variants using various advanced solvers.
  • Performance: Surpassed previous state-of-the-art methods with a 16.67% boost in success rate and a significant reduction in runtime errors.
  • Takeaway: Demonstrates that combining reinforcement learning with LLM-based reasoning provides a highly robust approach to automated optimization modeling.

🔗 资源与链接