跳转至

作为连续流的工具：演进式智能体推理

文章背景与核心概要

大语言模型（LLM）在利用工具解决复杂推理任务方面展现出了巨大的潜力。然而，传统方法依赖于离散的、逐步执行的范式，缺乏全局视角，这往往导致在长视距任务中错误不断累积，并且在遇到未见过的工具时泛化能力较差。

为了应对这些挑战，本文引入了 FlowAgent（Tools as Continuous Flow for Evolving Agentic Reasoning，作为连续流的工具：演进式智能体推理）。FlowAgent 将工具链式调用重新构想为语义空间中的连续轨迹生成。通过使用条件流匹配，该框架提供了一个全局规划视角，从而保证了连贯、稳健的执行，同时辅以关于效用收敛和误差衰减的形式化理论证明。

摘要 (Summary)

Large Language Models (LLMs) have demonstrated significant potential in utilizing tools for complex reasoning tasks. However, traditional approaches rely on a discrete, step-wise paradigm lacking a global perspective, which frequently leads to error accumulation over long horizons and poor generalization when encountering unseen tools.

大语言模型（LLM）在利用工具执行复杂推理任务方面展现出了显著潜力。然而，传统方法依赖于缺乏全局视角的离散、逐步式范式，这经常导致长视距任务中的错误累积，并在遇到未见工具时泛化能力较差。

To address these challenges, this paper introduces FlowAgent (Tools as Continuous Flow for Evolving Agentic Reasoning). FlowAgent reconceptualizes tool chaining as continuous trajectory generation within a semantic space. Using conditional flow matching, the framework provides a global planning perspective that guarantees coherent, robust execution, complemented by formal theoretical proofs of utility convergence and error attenuation.

为了应对这些挑战，本文引入了 FlowAgent（作为连续流的工具：演进式智能体推理）。FlowAgent 将工具链式调用重新概念化为语义空间内的连续轨迹生成。通过利用条件流匹配，该框架提供了一个全局规划视角，以确保连贯且鲁棒的工具执行，并辅以关于效用收敛和误差衰减的形式化理论证明。

论文元数据 (Paper Metadata)

  • arXiv ID: arXiv:2605.07339 [cs.AI]
  • Subject: Artificial Intelligence (cs.AI)
  • Authors:
  • Tairan Huang
  • Siyu Shang
  • Qiang Chen
  • Xiu Su
  • Yi Chen
  • Submission Dates:
  • Submitted on 8 May 2026 (v1)
  • Last revised 12 Aug 2026 (v2, current)
  • DOI: 10.48550/arXiv.2605.07339
  • arXiv ID: arXiv:2605.07339 [cs.AI]
  • 学科领域: 人工智能 (cs.AI)
  • 作者:
  • Tairan Huang
  • Siyu Shang
  • Qiang Chen
  • Xiu Su
  • Yi Chen
  • 提交日期:
  • 2026年5月8日提交 (v1)
  • 2026年8月12日最后修订 (v2, 当前版本)
  • DOI: 10.48550/arXiv.2605.07339

Abstract (论文摘要)

Large Language Models (LLMs) have demonstrated remarkable capabilities in orchestrating tools for reasoning tasks. However, existing methods rely on a step-wise paradigm that lacks a global perspective, which causes error accumulation over long horizons and restricts generalization to unseen tools. To overcome these limitations, we propose Tools as Continuous Flow for Evolving Agentic Reasoning (FlowAgent), which reconceptualizes tool chaining as continuous trajectory generation within a semantic space. To systematically evaluate this paradigm, we introduce the first plan-level closed-loop benchmark dedicated to plan-level agentic reasoning in dynamic real-world environments. Specifically, the proposed FlowAgent leverages conditional flow matching to generate continuous latent trajectories, providing a global planning perspective to ensure coherent and robust tool execution. Theoretically, we establish formal bounds on utility convergence and prove that our continuous formulation fundamentally guarantees robust generalization and error attenuation. Empirical evaluations show that FlowAgent achieves superior robustness and adaptability in long-horizon reasoning tasks.

大语言模型（LLM）在编排工具以完成推理任务方面表现出卓越的能力。然而，现有方法依赖于缺乏全局视角的逐步式范式，这会导致长视距下的错误累积，并限制了对未见工具的泛化能力。为了克服这些局限性，我们提出了“作为连续流的工具：演进式智能体推理”（FlowAgent），它将工具链式调用重新概念化为语义空间中的连续轨迹生成。为了系统地评估这一范式，我们引入了首个致力于动态真实世界环境中计划级智能体推理的计划级闭环基准测试。具体而言，所提出的 FlowAgent 利用条件流匹配来生成连续的潜在轨迹，提供全局规划视角以确保连贯且鲁棒的工具执行。在理论上，我们确立了效用收敛的形式化边界，并证明了我们的连续公式从根本上保证了鲁棒的泛化和误差衰减。实证评估表明，FlowAgent 在长视距推理任务中实现了卓越的鲁棒性和适应性。

核心贡献 (Key Contributions)

  1. Continuous Flow Paradigm: Reconceptualizes traditional step-by-step tool chaining into continuous trajectory generation within a semantic latent space.
  2. Conditional Flow Matching: Utilizes flow-matching techniques to maintain a global planning perspective, significantly reducing error accumulation during long-horizon tasks.
  3. Plan-Level Closed-Loop Benchmark: Introduces the first specialized benchmark designed for evaluating plan-level agentic reasoning within dynamic real-world settings.
  4. Theoretical Guarantees: Formally establishes bounds on utility convergence and proves robust generalization and error attenuation.
  1. 连续流范式： 将传统的逐步工具链式调用重新构想为语义潜在空间内的连续轨迹生成。
  2. 条件流匹配： 利用流匹配技术来保持全局规划视角，显著减少长视距任务过程中的错误累积。
  3. 计划级闭环基准测试： 推出了首个专门用于评估动态现实世界环境中计划级智能体推理的基准测试。
  4. 理论保证： 正式建立效用收敛的边界，并证明了鲁棒的泛化能力与误差衰减。

访问与资源 (Access & Resources)