AFLOW: Automating Agentic Workflow Generation 翻译

Doc2X：智能文档解析工具
Doc2X 支持从 PDF 转换为 Docx、HTML、Markdown，功能覆盖 公式识别、代码解析、表格转换、多栏布局解析，并整合了 GPT翻译 和 Deepseek 翻译！
Doc2X: Intelligent Document Parsing Tool
Doc2X supports PDF to Docx, HTML, and Markdown, with features like formula recognition, code parsing, table conversion, and multi-column layout parsing, integrated with GPT and DeepSeek translations!
👉 了解 Doc2X 的独特功能 | Explore Doc2X Features

原文链接：https://arxiv.org/pdf/2410.10762

AFLOW: Automating Agentic Workflow Generation

AFLOW: 自动化代理工作流生成

Jiayi Zhang1,2*, Jinyu Xiang1*Zhaoyang Yu3, Fengwei Teng3, Xionghui Chen4, Jiaqi Chen ${}^5$ , Mingchen Zhuge ${}^6$ , Xin Chen ${}^3$ , Sirui Hong ${}^1$ , Jinlin Wang ${}^1$ , Bingnan Zheng ${}^5$ , Bang Liu ${}^7$ ,Yuyu Luo ${}^{2,8}{}_{,}^{†}$ Chenglin Wu ${}^{1†}$

Jiayi Zhang1,2*, Jinyu Xiang1*Zhaoyang Yu3, Fengwei Teng3, Xionghui Chen4, Jiaqi Chen ${}^5$, Mingchen Zhuge ${}^6$, Xin Chen ${}^3$, Sirui Hong ${}^1$, Jinlin Wang ${}^1$, Bingnan Zheng ${}^5$, Bang Liu ${}^7$, Yuyu Luo ${}^{2,8}{}_{,}^{†}$ Chenglin Wu ${}^{1†}$

${}^1$ DeepWisdom, ${}^2$ The Hong Kong University of Science and Technology (Guangzhou)

${}^1$ 深智, ${}^2$ 香港科技大学（广州）

${}^3$ Renmin University of China ${}^4$ Nanjing University ${}^5$ Fudan University

${}^3$ 中国人民大学 ${}^4$ 南京大学 ${}^5$ 复旦大学

${}^6$ King Abdullah University of Science and Technology, ${}^7$ Université de Montréal &Mila

${}^6$ 阿卜杜拉国王科技大学, ${}^7$ 蒙特利尔大学 & Mila

${}^8$ The Hong Kong University of Science and Technology

${}^8$ 香港科技大学

ABSTRACT

摘要

Large language models (LLMs) have demonstrated remarkable potential in solving complex tasks across diverse domains, typically by employing agentic work-flows that follow detailed instructions and operational sequences. However, constructing these workflows requires significant human effort, limiting scalability and generalizability. Recent research has sought to automate the generation and optimization of these workflows, but existing methods still rely on initial manual setup and fall short of achieving fully automated and effective workflow generation. To address this challenge, we reformulate workflow optimization as a search problem over code-represented workflows, where LLM-invoking nodes are connected by edges. We introduce AFLOW, an automated framework that efficiently explores this space using Monte Carlo Tree Search, iteratively refining workflows through code modification, tree-structured experience, and execution feedback. Empirical evaluations across six benchmark datasets demonstrate AFLOW’s efficacy,yielding a $5.7\%$ average improvement over state-of-the-art baselines. Furthermore, AFLOW enables smaller models to outperform GPT-4o on specific tasks at $4.55\%$ of its inference cost in dollars. The code will be available at https://github.com/geekan/MetaGPT

大型语言模型（LLMs）在解决各个领域的复杂任务方面展示了显著的潜力，通常通过遵循详细指令和操作序列的代理工作流程来实现。然而，构建这些工作流程需要大量的人力投入，这限制了其可扩展性和普适性。最近的研究旨在自动化这些工作流程的生成和优化，但现有方法仍依赖于初始的手动设置，未能实现完全自动化和有效的工作流程生成。为了解决这一挑战，我们将工作流程优化重新表述为一个在代码表示的工作流程上进行搜索的问题，其中调用LLM的节点通过边连接。我们引入了AFLOW，一个自动化框架，它利用蒙特卡洛树搜索高效地探索这一空间，通过代码修改、树状结构经验和执行反馈迭代地优化工作流程。对六个基准数据集的实证评估证明了AFLOW的有效性，产生了相对于最先进基线的 $5.7\%$ 平均改进。此外，AFLOW使得较小的模型在特定任务上以其推理成本的 $4.55\%$ 超越GPT-4o。代码将可在 https://github.com/geekan/MetaGPT 获取。

1 INTRODUCTION

1 引言

Large Language Models (LLMs) have emerged as powerful tools for solving complex tasks across various domains, including code generation, data analysis, decision-making, and question answering (Liu et al., 2024; Li et al., 2024a; Zhu et al., 2024; Xie et al., 2024b; Sun et al., 2024; Wang et al., 2024b, Song et al. 2023, Xie et al. 2024a, Zhong et al. 2024a). However, the rapid advancement of LLMs heavily relies on manually designed agentic workflows - structured sequences of LLM invocations accompanied by detailed instructions. Designing and refining these workflows requires significant human effort, which limits the scalability and adaptability of LLMs to new, complex domains and hinders their ability to transfer skills across diverse tasks (Tang et al. 2024).

大型语言模型（LLMs）已成为解决各个领域复杂任务的强大工具，包括代码生成、数据分析、决策制定和问答（Liu et al., 2024; Li et al., 2024a; Zhu et al., 2024; Xie et al., 2024b; Sun et al., 2024; Wang et al., 2024b, Song et al. 2023, Xie et al. 2024a, Zhong et al. 2024a）。然而，LLMs 的快速发展在很大程度上依赖于手动设计的代理工作流程——结构化的 LLM 调用序列，伴随详细的指令。设计和完善这些工作流程需要大量的人力投入，这限制了 LLMs 在新复杂领域的可扩展性和适应性，并阻碍了它们在不同任务之间转移技能的能力（Tang et al. 2024）。

Recent efforts have focused on automating the discovery of effective agentic workflows to reduce the reliance on human intervention (Khattab et al., 2024, Yüksekgöniil et al., 2024, Liu et al., 2023 Hu et al., 2024). Despite these advancements, full automation has not been achieved. For instance, Khattab et al. (2024) requires manual workflow setup before automated prompt optimization. Similarly, methods proposed by Yüksekgönül et al. (2024) and Zhuge et al. (2024) fail to capture the full diversity of workflows necessary for a wide range of tasks (Yu et al. 2023; Yang et al. 2024b, Sun et al. 2023), as their optimization objectives struggle to represent the breadth of possible workflows. The inability to effectively model diverse workflow structures within these automated systems limits their utility and impact. ADAS (Hu et al. 2024) represents workflows using code, achieving a

最近的努力集中在自动化发现有效的代理工作流程，以减少对人类干预的依赖（Khattab 等，2024年；Yüksekgöniil 等，2024年；Liu 等，2023年；Hu 等，2024年）。尽管取得了这些进展，但尚未实现完全自动化。例如，Khattab 等（2024年）在自动化提示优化之前需要手动设置工作流程。同样，Yüksekgönül 等（2024年）和 Zhuge 等（2024年）提出的方法未能捕捉到广泛任务所需的工作流程的全部多样性（Yu 等，2023年；Yang 等，2024b年；Sun 等，2023年），因为它们的优化目标难以代表可能工作流程的广度。这些自动化系统无法有效建模多样化的工作流程结构，限制了它们的效用和影响。ADAS（Hu 等，2024年）使用代码表示工作流程，取得了

---

*These authors contributed equally to this work.

*这些作者对本研究贡献相同。

${}^{†}$ Corresponding authors: Yuyu Luo (E-mail:yuyuluo@hkust-gz.edu.cn),Chenglin Wu (E-mail: alexan-derwu@deepwisdom.ai)

${}^{†}$ 通讯作者：罗宇宇（电子邮件：yuyuluo@hkust-gz.edu.cn），吴承林（电子邮件：alexanderwu@deepwisdom.ai）

---

Figure 1: Performance comparison with other methods. To assess the method’s performance, we employ various metrics across different datasets: solve rate for Math and GSM8K, F1 score for HotpotQA and DROP, and pass@1 for HumanEval and MBPP. Our AFLOW (highlighted in yellow) consistently outperforms all automated workflow optimization and manually designed methods across all six benchmarks.

图1：与其他方法的性能比较。为了评估该方法的性能，我们在不同数据集上采用各种指标：数学和GSM8K的解决率，HotpotQA和DROP的F1分数，以及HumanEval和MBPP的pass@1。我们的AFLOW（以黄色突出显示）在所有六个基准测试中始终优于所有自动化工作流程优化和手动设计的方法。

relatively complete representation. However, due to the efficiency limitations of its linear heuristic search algorithm, ADAS struggles to generate effective workflows within a limited number of iterations. This highlights the need for more effective techniques to represent and automate the generation of agentic workflows, which would accelerate the application of LLMs across domains.

相对完整的表示。然而，由于其线性启发式搜索算法的效率限制，ADAS 在有限的迭代次数内难以生成有效的工作流程。这突显了需要更有效的技术来表示和自动生成代理工作流程，从而加速 LLM 在各个领域的应用。

In response to these challenges, we introduce an innovative framework for automatically generating agentic workflows. Our key idea is to model the workflow as a sequence of interconnected LLM-invoking nodes, where each node represents an LLM action and the edges define the logic, dependencies, and flow between these actions. This structure transforms the workflow into a vast search space, encompassing a wide variety of potential configurations. Our goal is to efficiently navigate this space, automatically generating optimized workflows that maximize task performance while minimizing human intervention.

针对这些挑战，我们引入了一种创新框架，用于自动生成代理工作流程。我们的关键思想是将工作流程建模为一系列相互连接的 LLM 调用节点，其中每个节点代表一个 LLM 操作，边缘定义了这些操作之间的逻辑、依赖关系和流动。这一结构将工作流程转变为一个广阔的搜索空间，涵盖各种潜在配置。我们的目标是高效地导航这个空间，自动生成优化的工作流程，以最大化任务性能，同时最小化人类干预。

However, the diversity and complexity of tasks present significant challenges. Specifically, each task can have different requirements, operations, and dependencies, which makes it difficult to represent them in a unified yet flexible manner (Chen et al. 2021; Cobbe et al. 2021; Yang et al. 2018; Luo et al. 2018). Furthermore, the search space for possible workflows, comprising an immense number of code structures and node configurations, is virtually boundless, creating an additional challenge for efficient exploration and optimization.

然而，任务的多样性和复杂性带来了重大挑战。具体而言，每个任务可能具有不同的要求、操作和依赖关系，这使得以统一而灵活的方式表示它们变得困难（Chen et al. 2021; Cobbe et al. 2021; Yang et al. 2018; Luo et al. 2018）。此外，可能的工作流程的搜索空间包含大量的代码结构和节点配置，几乎是无边界的，这为高效探索和优化带来了额外的挑战。

To address these challenges, we propose AFLOW, a Monte Carlo Tree Search (MCTS)-based framework designed to systematically explore and discover optimal agentic workflows. AFLOW represents workflows as flexible nodes connected by code-based edges, which encapsulate possible relationships such as logical flows, conditions, and dependencies. These edges allow the workflow to be modeled as a graph (Zhuge et al. 2024) or network (Liu et al. 2023), offering a powerful structure for capturing complex interactions between LLM invocations.

为了解决这些挑战，我们提出了 AFLOW，一个基于蒙特卡罗树搜索（MCTS）的框架，旨在系统地探索和发现最佳代理工作流程。AFLOW 将工作流程表示为通过基于代码的边连接的灵活节点，这些边封装了可能的关系，如逻辑流、条件和依赖关系。这些边允许将工作流程建模为图（Zhuge et al. 2024）或网络（Liu et al. 2023），为捕捉 LLM 调用之间的复杂交互提供了强大的结构。

To enhance the search process and improve efficiency, AFLOW introduces a novel concept of operators - predefined, reusable combinations of nodes representing common agentic operations (e.g., Ensemble, Review & Revise). These operators serve as foundational building blocks for constructing workflows and are integrated into the search space, ensuring that the exploration process leverages known patterns of effective agentic operations.

为了增强搜索过程并提高效率，AFLOW 引入了一种新颖的操作符概念——预定义的、可重用的节点组合，代表常见的代理操作（例如，集成、审查与修订）。这些操作符作为构建工作流的基础构件，集成到搜索空间中，确保探索过程利用已知的有效代理操作模式。

AFLOW employs the MCTS algorithm to navigate this infinite search space. The framework’s work-flow optimization process incorporates several key innovations: a soft mixed-probability selection mechanism for node exploration, LLM-driven node expansion to introduce new possibilities, execution evaluation to assess workflow performance, and backpropagation of experience to refine future search iterations. This combination of techniques ensures that AFLOW efficiently discovers work-flows that adapt to the complexity of diverse tasks while reducing reliance on manual intervention.

AFLOW 采用 MCTS 算法在这个无限的搜索空间中导航。该框架的工作流优化过程包含几个关键创新：用于节点探索的软混合概率选择机制、驱动节点扩展的 LLM 引入新可能性、执行评估以评估工作流性能，以及经验的反向传播以优化未来的搜索迭代。这种技术组合确保 AFLOW 高效发现适应多样任务复杂性的工作流，同时减少对人工干预的依赖。

We make the following key contributions:

我们做出了以下关键贡献：

• Problem Formulation: We formalize the workflow optimization problem, generalizing prior approaches as specific cases. This provides a unified framework for future research at both the node and workflow optimization levels.

• 问题表述：我们对工作流优化问题进行了形式化，将先前的方法概括为特定案例。这为未来在节点和工作流优化层面的研究提供了统一框架。

• AFLOW: We introduce AFLOW, an MCTS-based method that automatically discovers effective workflows across multiple domains with minimal human intervention.

• AFLOW：我们介绍了 AFLOW，一种基于 MCTS 的方法，能够在多个领域自动发现有效的工作流，且人类干预最小。

• Extensive Evaluation: We evaluate AFLOW on six benchmark datasets: HumanEval, MBPP, MATH, GSM8K, HotPotQA, and DROP. AFLOW outperforms manually designed methods by $5.7\%$ and surpasses existing automated approaches by $19.5\%$ . Notably,workflows generated by AFLOW enable smaller LLMs to outperform larger models, offering better cost-performance efficiency, with significant implications for real-world applications.

• 大规模评估：我们在六个基准数据集上评估 AFLOW：HumanEval、MBPP、MATH、GSM8K、HotPotQA 和 DROP。AFLOW 的表现超越了手动设计的方法 $5.7\%$，并且超过了现有的自动化方法 $19.5\%$。值得注意的是，AFLOW 生成的工作流使得较小的 LLM 能够超越更大的模型，提供更好的成本性能效率，这对实际应用具有重要意义。

2 Related Work

2 相关工作

Agentic Workflow. Agentic workflow and autonomous agents (Zhuge et al., 2023, Hong et al., 2024a; Zhang et al., 2024; Wang et al., 2023) represent two distinct paradigms of LLM application. The former completes tasks statically through predefined processes with multiple LLM invocations, while the latter solves problems dynamically through flexible autonomous decision-making. Compared to autonomous agents that require specific actions and decision patterns designed for the environment, agentic workflows can be constructed based on existing human domain experience and iterative refinement, offering higher potential for automated construction.

代理工作流。代理工作流和自主代理（Zhuge et al., 2023；Hong et al., 2024a；Zhang et al., 2024；Wang et al., 2023）代表了大型语言模型应用的两种不同范式。前者通过预定义的过程和多次调用大型语言模型静态地完成任务，而后者则通过灵活的自主决策动态地解决问题。与需要为环境设计特定行动和决策模式的自主代理相比，代理工作流可以基于现有的人类领域经验和迭代改进构建，提供更高的自动化构建潜力。

Agentic workflows can be broadly categorized into general and domain-specific types. General workflows emphasize universal problem-solving approaches, such as (Wei et al., 2022; Wang et al., 2022, Madaan et al., 2023, Wang et al., 2024a). Domain-specific workflows focus on building effective processes to solve domain-specific problems, such as code generation (Hong et al., 2024b Ridnik et al., 2024, Zhong et al., 2024a), data analysis (Xie et al., 2024b, Ye et al., 2024, Li et al., 2024a; Zhou et al., 2023), mathematics (Zhong et al., 2024b; Xu et al., 2024), question answering (Nori et al., 2023, Zhou et al., 2024a). Existing work has manually discovered numerous effective agentic workflows, but it’s challenging to exhaust various tasks across different domains, further highlighting the importance of automated workflow generation and optimization.

代理工作流可以大致分为通用型和领域特定型。通用工作流强调普遍的问题解决方法，例如（Wei et al., 2022；Wang et al., 2022；Madaan et al., 2023；Wang et al., 2024a）。领域特定工作流则专注于构建有效的流程以解决领域特定的问题，例如代码生成（Hong et al., 2024b；Ridnik et al., 2024；Zhong et al., 2024a）、数据分析（Xie et al., 2024b；Ye et al., 2024；Li et al., 2024a；Zhou et al., 2023）、数学（Zhong et al., 2024b；Xu et al., 2024）、问答（Nori et al., 2023；Zhou et al., 2024a）。现有的研究手动发现了许多有效的代理工作流，但在不同领域中穷尽各种任务是具有挑战性的，这进一步突显了自动化工作流生成和优化的重要性。

Automated Agentic Optimization. Recent work aims to automate the design of agentic work-flows, categorized into three types: automated prompt optimization, hyperparameter optimization, and automated workflow optimization. Prompt optimization (Fernando et al., 2024, Yüksekgönil et al. 2024 Yang et al. 2024a Khattab et al. 2024) uses LLMs to optimize prompts within fixed workflows. Hyperparameter optimization (Saad-Falcon et al. 2024) focuses on optimizing predefined parameters. While these approaches improve performance, they are limited in generalization to new tasks and often require moderate human effort for task-specific designs.

自动化代理优化。最近的研究旨在自动化代理工作流程的设计，分为三种类型：自动化提示优化、超参数优化和自动化工作流程优化。提示优化（Fernando et al., 2024; Yüksekgönil et al., 2024; Yang et al., 2024a; Khattab et al., 2024）使用大型语言模型（LLMs）在固定工作流程中优化提示。超参数优化（Saad-Falcon et al., 2024）专注于优化预定义参数。虽然这些方法提高了性能，但在新任务的泛化能力上受到限制，并且通常需要适度的人力来进行任务特定的设计。

Automated workflow optimization (Li et al., 2024b, Zhou et al., 2024b, Zhuge et al., 2024, Hu et al. 2024) aims to optimize entire workflow structures, offering more potential for fully automated generation. Recent works explore diverse representations and methods. GPTSwarm (Zhuge et al. 2024) uses graph structures with reinforcement learning, but struggles to represent workflows with conditional states due to graph structure limitations. ADAS (Hu et al. 2024) utilizes code structures to represent workflows and stores historical workflows in a linear list structure, aligning closely with our goals. However, it is constrained by the efficiency of its search algorithm as it relies on overly simplistic representations of experiences in the searching process, making it challenging to discover effective workflows.

自动化工作流程优化（Li et al., 2024b; Zhou et al., 2024b; Zhuge et al., 2024; Hu et al., 2024）旨在优化整个工作流程结构，提供更大的完全自动化生成潜力。最近的研究探索了多样化的表示和方法。GPTSwarm（Zhuge et al., 2024）使用图结构与强化学习，但由于图结构的限制，难以表示具有条件状态的工作流程。ADAS（Hu et al., 2024）利用代码结构表示工作流程，并将历史工作流程存储在线性列表结构中，与我们的目标紧密对齐。然而，它受到搜索算法效率的限制，因为它依赖于过于简单的经验表示，使得发现有效工作流程变得具有挑战性。

AFLOW also uses code to represent workflows, but goes further by providing a more fundamental structure called named node. This structure encompasses various LLM invocation parameters, allowing for more detailed workflow representation. We also introduce operators that implement predefined node combination functions. Simultaneously, AFLOW employs a specially designed MCTS algorithm for automated workflow optimization, leveraging the tree-structured experience and execution feedback to efficiently discover effective workflows.

AFLOW 也使用代码来表示工作流程，但进一步提供了一种称为命名节点的更基本结构。该结构包含各种 LLM 调用参数，允许更详细的工作流程表示。我们还引入了实现预定义节点组合功能的操作符。同时，AFLOW 采用一种专门设计的蒙特卡洛树搜索（MCTS）算法进行自动化工作流程优化，利用树状结构的经验和执行反馈来高效发现有效的工作流程。

3 PRELIMINARY

3 初步研究

In this section, we will first formulate the automated agentic workflows generation problem in Section 3.1 and then discuss design considerations of our AFLOW in Section 3.2. For the core concept of this section, we provide an example explanation in Figure 2.

在本节中，我们将首先在第3.1节中对自动代理工作流生成问题进行公式化，然后在第3.2节中讨论我们AFLOW的设计考虑。对于本节的核心概念，我们在图2中提供了示例解释。

Figure 2: The example of node, operator, and edge. We demonstrate the optional parameters for Nodes, the structure of some Operators, and common representations of Edges.

图2：节点、操作符和边的示例。我们展示了节点的可选参数、某些操作符的结构以及边的常见表示。

3.1 Problem Formulation

3.1 问题公式化

Agentic Workflow. We define an agentic workflow $W$ as a sequence of LLM-invoking nodes, denoted as N = { N 1 , N 2 , … , N i … } \mathcal{N} = \left\{ {{N}_{1},{N}_{2},\ldots ,{N}_{i}\ldots }\right\} N={N1​,N2​,…,Ni​…} . Each node $N_i$ represents a specific operation performed by an LLM and is characterized by the following parameters. The code abstraction of the node is shown in Appendix A.2.

代理工作流。我们将代理工作流 $W$ 定义为一系列调用LLM的节点，记作 N = { N 1 , N 2 , … , N i … } \mathcal{N} = \left\{ {{N}_{1},{N}_{2},\ldots ,{N}_{i}\ldots }\right\} N={N1​,N2​,…,Ni​…}。每个节点 $N_i$ 代表LLM执行的特定操作，并由以下参数特征化。节点的代码抽象在附录A.2中展示。

• Model $M$ : The specific language model invoked at node $N_i$ .

• 模型 $M$ : 在节点 $N_i$ 调用的特定语言模型。

• Prompt $P$ : The input or task description provided to the model at each node.

• 提示 $P$ : 在每个节点提供给模型的输入或任务描述。

• Temperature $\tau$ : A parameter controlling the randomness of the LLM’s output at node $N_i$ .

• 温度 $\tau$ : 控制节点 $N_i$ 中LLM输出随机性的参数。

• Output format $F$ : The format in which the model’s output is structured (e.g., $\mathrm{xml}$ ,json, markdown, raw). The node in workflow should provide different output formats, inspired by the Tam et al. (2024).

• 输出格式 $F$ : 模型输出的结构格式（例如，$\mathrm{xml}$，json，markdown，raw）。工作流中的节点应提供不同的输出格式，灵感来自Tam等人（2024）。

These nodes are connected by edges $E$ ,which represent the connection between the nodes,governing the sequence of execution. The edges $E$ can represent various structures,such as:

这些节点通过边 $E$ 连接，表示节点之间的连接，控制执行的顺序。边 $E$ 可以表示各种结构，例如：

• Graph Zhuge et al. (2024): A flexible structure representing hierarchical, sequential, or parallel relationships between nodes, allowing for complex branching workflows.

• 图 Zhuge等人（2024）：一种灵活的结构，表示节点之间的层次、顺序或并行关系，允许复杂的分支工作流。

• Neural Network (Liu et al., 2023): A structure that can represent complex, non-linear relationships between nodes, allowing for adaptive and learnable workflows based on input and feedback.

• 神经网络（Liu等人，2023）：一种可以表示节点之间复杂非线性关系的结构，允许基于输入和反馈的自适应和可学习工作流。

• Code (Hu et al., 2024): A comprehensive representation that can express linear sequences, conditional logic, loops, and incorporate graph or network structures, offering the most precise control over workflow execution.

• 代码 (Hu et al., 2024)：一种全面的表示方式，可以表达线性序列、条件逻辑、循环，并结合图形或网络结构，提供对工作流执行的最精确控制。

Automated Workflow Optimization. Given a task $T$ and an evaluation function $G$ ,the goal of workflow optimization is to discover a workflow $W$ that maximizes $G\left(W,T\right)$ . This can be formulated as a search process where an algorithm $A$ explores the search space $\mathcal{S}$ to determine the optimal workflow configuration. The search space $\mathcal{S}$ for a workflow optimization problem encompasses all possible configurations of node parameters and edge structures:

自动化工作流优化。给定一个任务 $T$ 和一个评估函数 $G$，工作流优化的目标是发现一个工作流 $W$，使其最大化 $G\left(W,T\right)$。这可以被表述为一个搜索过程，其中一个算法 $A$ 探索搜索空间 $\mathcal{S}$ 以确定最佳工作流配置。工作流优化问题的搜索空间 $\mathcal{S}$ 包含所有可能的节点参数和边结构的配置：

where N = { N ( M , τ , P , F ) ∣ M ∈ M , τ ∈ [ 0 , 1 ] , P ∈ P , F ∈ F } \mathcal{N} = \{ N\left( {M,\tau ,P,F}\right) \mid M \in \mathcal{M},\tau \in \left\lbrack {0,1}\right\rbrack ,P \in \mathcal{P},F \in \mathcal{F}\} N={N(M,τ,P,F)∣M∈M,τ∈[0,1],P∈P,F∈F} ,with $\mathcal{M},\mathcal{P},\mathcal{F},\mathcal{E}$ representing the sets of possible language models, prompts, output formats, and edge configurations, respectively.

其中 N = { N ( M , τ , P , F ) ∣ M ∈ M , τ ∈ [ 0 , 1 ] , P ∈ P , F ∈ F } \mathcal{N} = \{ N\left( {M,\tau ,P,F}\right) \mid M \in \mathcal{M},\tau \in \left\lbrack {0,1}\right\rbrack ,P \in \mathcal{P},F \in \mathcal{F}\} N={N(M,τ,P,F)∣M∈M,τ∈[0,1],P∈P,F∈F}，而 $\mathcal{M},\mathcal{P},\mathcal{F},\mathcal{E}$ 分别表示可能的语言模型、提示、输出格式和边配置的集合。

—— 更多内容请到Doc2X翻译查看——
—— For more content, please visit Doc2X for translations ——