跳转至

文章背景与核心概要

本文桥接了动态图系统与自进化智能体这两个此前相互隔离的研究方向，提出了一种全新的核心视角：将智能体的进化过程构架为动态图变换（dynamic graph transformation）。随着大语言模型（LLM）驱动的智能体逐渐演变为能够管理记忆、利用工具、获取技能、优化工作流以及进行多智能体协作的持久化、自进化系统，传统的静态支撑结构已难以满足需求。

论文摒弃了仅将图视为静态支撑结构或单纯关注智能体级别机制的做法，创新性地将整个智能体状态建模为动态图。其中，记忆、工具、技能、工作流以及智能体间的关系被表示为带类型的节点、边和子图，并通过受模式约束的重写（schema-constrained rewrites）进行更新。该研究不仅系统梳理了智能体进化的四维分类法，还将动态图学习的多个子领域映射至智能体核心能力，为未来具身及认知智能体架构提供了坚实的可复用基础设施和治理框架。

Self-Evolving Agents as Dynamic Graph Transformation: A Survey and New Perspective

arXiv: arXiv:2608.18104 [cs.AI]
Submitted: June 10, 2026
Authors: Yuanyuan Xu, Wenjie Zhang, Yin Chen, Xuemin Lin, Ying Zhang
Project Repository: GitHub - Awesome-Agent-Dynamic-Graphs

arXiv: arXiv:2608.18104 [cs.AI]
Submitted: June 10, 2026
Authors: Yuanyuan Xu, Wenjie Zhang, Yin Chen, Xuemin Lin, Ying Zhang
Project Repository: GitHub - Awesome-Agent-Dynamic-Graphs

📌 Summary

大语言模型（LLM）驱动的智能体正日益转变为持久化的、自进化的系统，能够管理记忆、利用工具、获取技能、优化工作流并在多个智能体之间进行协同。

本文通过引入一个新的核心视角——将智能体进化构架为动态图变换，连接了两个此前相互隔离的研究领域：动态图系统和自进化智能体。

作者没有将图仅仅视为静态的支撑结构，也没有严格局限于智能体层面的机制，而是将整个智能体状态建模为一个动态图。记忆、工具、技能、工作流以及智能体间的关系被表示为类型化的节点、边和子图，它们通过模式约束的重写进行更新。

Large language model (LLM)-based agents are increasingly transforming into persistent, self-evolving systems capable of managing memories, utilizing tools, acquiring skills, refining workflows, and coordinating across multiple agents.

This paper bridges two previously isolated research tracks—dynamic graph systems and self-evolving agents—by introducing a novel core perspective: framing agent evolution as dynamic graph transformation.

Instead of treating graphs merely as static support structures or focusing strictly on agent-level mechanisms, the authors model the entire agent state as a dynamic graph. Memories, tools, skills, workflows, and inter-agent relationships are represented as typed nodes, edges, and subgraphs that update through schema-constrained rewrites.

🔑 Key Contributions & Framework

  1. 智能体进化的新范式：
    将复杂、不断变化的智能体状态（实体、关系、属性、依赖项和执行结构）建模为动态图，这些图根据环境反馈和新证据经历模式约束的重写。

  2. 智能体进化的四重分类法：
    系统地将现有的基于动态图的方法组织为四个关键维度：

  3. 节点/特征进化
  4. 边/拓扑进化
  5. 子图激活

  6. 跨组件协同进化

  7. 动态图学习作为可复用基础设施：
    将动态图学习的九个子领域映射到核心智能体进化能力上，强调了必要的适应性调整以及潜在的失效模式。

  8. 图感知评估与治理：
    从动态图的角度提出了五种评估和治理协议，以补充传统的端到端任务评估。

  1. A New Paradigm for Agent Evolution:
    Models complex, shifting agent states (entities, relations, attributes, dependencies, and execution structures) as dynamic graphs undergoing schema-constrained rewrites based on environmental feedback and new evidence.

  2. A Four-Fold Taxonomy of Agent Evolution:
    Existing dynamic-graph-based methods are systematically organized into four key dimensions:

  3. Node/Feature Evolution
  4. Edge/Topology Evolution
  5. Subgraph Activation

  6. Cross-Component Co-Evolution

  7. Dynamic Graph Learning as Reusable Infrastructure:
    Maps nine subfields of dynamic graph learning to core agent-evolution capabilities, highlighting essential adaptations and potential failure modes.

  8. Graph-Aware Evaluation and Governance:
    Proposes five types of evaluation and governance protocols from a dynamic-graph perspective to complement traditional end-task evaluations.

License & Media Elements

license icon view license

License & Media Elements

license icon view license