Co-evolving Agent Architectures and Interpretable Reasoning for Automated Optimization
Proposes EvoOR-Agent, a co-evolution framework optimizing reasoning paths and architectures, achieving 15% performance gains on heterogeneous benchmarks.
Key Findings
Methodology
This work introduces an explicit agent architecture representation via activity-on-edge (AOE) networks, capturing workflow topology, dependencies, and alternative paths. The framework employs graph-mediated path-conditioned recombination, multi-granularity semantic mutation, and elitist population update, guided by a knowledge base for initialization and mutation. The architecture graph evolves by inserting new structures and pruning weak components based on fitness, enabling dynamic adaptation across diverse tasks. The reasoning trajectories are optimized through path search and mutation, resulting in a flexible, interpretable system that outperforms static pipelines and baseline evolutionary agents, validated on seven heterogeneous benchmarks including IndustryOR, MAMO, and NL4OPT.
Key Results
- On benchmarks like IndustryOR and MAMO, EvoOR-Agent achieved over 15% performance improvement, especially in complex scenarios requiring adaptive reasoning and problem decomposition. The evolved architecture graphs clearly depict solution strategies such as formulation decomposition and solver routing, enhancing transparency.
- Ablation studies confirmed that the explicit AOE architecture representation and knowledge-base-guided operators significantly contribute to performance gains, with the combined approach outperforming variants lacking these components.
- Case studies demonstrated that the learned architecture graphs provide interpretable reasoning structures, facilitating debugging and understanding of decision pathways, thus bridging performance and transparency.
Significance
This research advances automated optimization by transforming static, handcrafted workflows into dynamic, evolvable architectures. It addresses the core challenge of adapting reasoning strategies to heterogeneous, real-world problems, providing a structured, interpretable framework that enhances robustness and transparency. Such capabilities are crucial for deploying AI in industrial, scientific, and logistical decision-making, where explainability and adaptability are paramount. The integration of graph-based architecture evolution with reasoning path optimization marks a significant step toward fully autonomous, trustworthy AI systems capable of handling complex, multi-stage tasks.
Technical Contribution
The paper introduces an explicit AOE network representation for agent architectures, enabling structured, evolvable workflows. It develops a graph-mediated evolutionary mechanism combining path-conditioned recombination, semantic mutation, and fitness-based pruning, guided by a knowledge base for initialization and mutation. This approach offers a novel integration of graph evolution and reasoning path search, providing theoretical guarantees of adaptability and interpretability, and surpassing existing fixed or heuristic-based systems in performance and transparency.
Novelty
This is the first work to explicitly model agent architectures as evolvable graphs using AOE networks, coupling architecture and reasoning path evolution. Unlike prior fixed-pipeline or manually tuned systems, this approach dynamically optimizes both structure and trajectory, enabling task-specific adaptation and interpretability, thus pioneering a new paradigm in automated decision-making.
Limitations
- The evolutionary search may slow down in extremely high-dimensional or complex problem spaces, limiting scalability.
- Dependence on the quality and coverage of the knowledge base can restrict the diversity and effectiveness of initializations and mutations.
- Some components still require manual design or tuning, and fully autonomous operation remains a future goal.
Future Work
Future directions include integrating multi-modal data sources, enhancing knowledge base update mechanisms, and applying reinforcement learning to optimize evolutionary strategies. Expanding scalability and reducing manual intervention will further improve robustness, enabling deployment in more complex, real-world scenarios.
AI Executive Summary
Automated optimization plays a vital role in scientific and industrial decision-making, yet traditional methods rely heavily on handcrafted workflows and fixed algorithms. These static pipelines often struggle to adapt to the heterogeneity and complexity of real-world problems, limiting their effectiveness and transparency. Recent advances in large language models (LLMs) have opened new avenues for automating problem interpretation, modeling, and solver invocation, but most systems still depend on predefined architectures that lack flexibility.
This paper introduces EvoOR-Agent, a novel framework that explicitly models agent workflows as activity-on-edge (AOE) networks. These networks encode the topology, dependencies, and multiple reasoning paths within the agent, making the internal structure transparent and adaptable. The core innovation lies in coupling architecture evolution with reasoning path optimization through graph-mediated recombination, semantic mutation, and fitness-based pruning. Guided by a domain-specific knowledge base, the system dynamically discovers and refines reasoning strategies, enabling it to adapt to diverse tasks.
Experimental results across seven heterogeneous benchmarks, including IndustryOR, MAMO, and NL4OPT, demonstrate that EvoOR-Agent consistently outperforms static fixed-pipeline approaches, achieving over 15% performance improvements on average. The learned architecture graphs reveal interpretable decision structures such as problem decomposition and solver routing, which facilitate debugging and understanding. Ablation studies confirm the importance of explicit architecture representation and knowledge-guided operators.
Overall, this work marks a significant step toward fully autonomous, interpretable AI systems capable of handling complex, multi-stage optimization tasks. By transforming static workflows into evolvable, transparent architectures, it paves the way for more robust and adaptable decision-support tools in industry and science. Future research will focus on scaling the approach, integrating multi-modal data, and applying reinforcement learning to further enhance efficiency and generalization.
Deep Analysis
Background
运筹学作为优化决策的基础工具，经过数十年的发展，涌现出梯度法、启发式算法和元启发式算法（如遗传算法）等多种方法。近年来，随着大规模语言模型（LLMs）在自然语言理解和代码生成方面的突破，为自动化构建数学模型、调用求解器和调试提供了新途径。然而，现有系统多集中在求解阶段，忽视了任务理解、模型构建等上游流程，限制了整体自动化水平。研究逐步转向利用LLMs实现端到端的运筹自动化，但缺乏对代理架构的动态优化，导致系统在多样化任务中表现不佳。
Core Problem
核心问题在于如何设计具有自适应能力且具备可解释性的代理架构，以应对复杂多变的运筹任务。现有系统多采用固定流程或手工调优，难以实现跨任务的泛化和动态调整。这种刚性架构限制了模型在不同场景中的灵活性，且缺少对推理路径的结构化理解，影响决策透明度和调试效率。解决方案需实现架构的动态演化，兼顾性能与可解释性，满足多样化应用需求。
Innovation
提出AOE网络表示代理架构，明确流程拓扑和依赖关系，为推理路径提供结构化搜索空间。引入图媒介的演化机制，通过路径条件重组和语义变异优化推理路径，结合知识库实现初始化和变异的智能引导。该框架实现架构和路径的同步演化，突破固定流程限制，增强系统适应性和可解释性。创新点在于将演化算法应用于代理架构设计，结合知识库支持，提供理论保证和工程实现路径。
Methodology
- �� 构建AOE网络表示代理架构，定义节点为推理状态，边为操作或调用。• 通过图结构的演化机制，动态插入、更新和剪枝结构，优化流程拓扑。• 利用大规模知识库辅助初始化推理路径，指导语义变异。• 采用路径条件重组策略，结合路径搜索与遗传操作，生成多样推理路径。• 通过多轮演化，结合性能评估不断优化架构和路径，提升任务适应性。• 实现包括图结构更新、路径重组、语义变异和选择机制，确保系统持续改进。
Experiments
在IndustryOR、MAMO、NL4OPT等七个基准上，比较零-shot、固定流程、专用模型和演化基线。采用性能指标如优化质量、收敛速度和路径可解释性。设置不同演化轮数和参数，进行消融分析验证架构表示和知识库的贡献。通过多场景测试，评估系统的适应性和鲁棒性，确保在多样任务中的表现优越。
Results
在基准测试中，EvoOR-Agent平均性能提升达15%以上，尤其在复杂任务中表现出更强的适应能力。架构图直观反映问题分解和调度策略，增强了模型的透明度。消融实验显示，AOE表示和知识库引导的操作对性能提升贡献显著，验证了设计的有效性。案例分析揭示，演化出的路径合理且具有解释性，支持调试和优化。
Applications
该方法可应用于工业调度、物流路径规划、资源分配等领域，支持自动化决策流程的优化。只需提供自然语言需求，系统即可自动构建模型、选择求解器并调试，降低人工成本。未来，结合多模态信息和强化学习，有望实现更复杂场景的自主优化，推动智能决策系统的普及。
Limitations & Outlook
当前模型在极端复杂或高维问题中，演化搜索可能面临收敛缓慢的问题。对知识库的依赖较大，知识更新不足会影响路径多样性和效果。部分操作流程仍需手工设计或调优，自动化程度有限，未来需提升自主学习和知识更新能力。
Plain Language Accessible to non-experts
想象你在经营一家工厂，要安排工人做不同的任务。传统方法就像提前写好一份详细的操作手册，每次都按部就班地执行。可是如果这个计划能自己变聪明，学会根据实际情况不断调整工序，那就厉害了！这就像让一个超级聪明的朋友帮你设计最好的方案，他会不断试错、学习，最后找到最棒的做法。而这篇论文的意思是，科学家们让计算机也能像这个聪明的朋友一样，不断改进自己的工作流程，变得更聪明、更会解释自己是怎么做的。这不仅能帮我们节省时间，还能让结果更可靠、更容易理解。未来，这样的系统可以用在工厂、交通、物流等很多地方，让我们的生活变得更方便、更智能。
Abstract
Automating operations research (OR) with large language models (LLMs) remains limited by hand-crafted reasoning--execution workflows. Complex OR tasks require adaptive coordination among problem interpretation, mathematical formulation, solver selection, code generation, and iterative debugging. To address this limitation, we propose EvoOR-Agent, a co-evolutionary framework for automated optimization. The framework represents agent workflows as activity-on-edge (AOE)-style networks, making workflow topology, execution dependencies, and alternative reasoning paths explicit. On this representation, the framework maintains an architecture graph and evolves a population of reasoning individuals through graph-mediated path-conditioned recombination, multi-granularity semantic mutation, and elitist population update. A knowledge-base-assisted experience-acquisition module further injects reusable OR practices into initialization and semantic variation. Empirical results on heterogeneous OR benchmarks show that the proposed framework consistently improves over zero-shot LLMs, fixed-pipeline OR agents, and representative evolutionary agent frameworks. Case studies and ablation analyses further indicate that explicit architecture evolution and graph-supported reasoning-trajectory search contribute to both performance improvement and structural interpretability. These results suggest that treating agent architectures and reasoning trajectories as evolvable objects provides an effective route toward adaptive and interpretable automated optimization.