EXG: Experience Graph Framework for Self-Evolving LLM Agents
A new research paper presents EXG, an innovative experience graph framework aimed at self-evolving large language model (LLM)-based agents. Released on arXiv, this study highlights the issue that many operational agents exhibit static behavior, as knowledge gained during their operation seldom leads to systematic enhancements. Current self-evolving methods either depend on limited single-task reflection or utilize unstructured memory that gathers fragmented experiences, which can hinder immediate application. In contrast, EXG systematically organizes both successes and failures into a relational structure, facilitating real-time graph expansion during deployment. This framework marks the first experience graph tailored for self-evolving agents, allowing for ongoing learning and development. The paper can be accessed at https://arxiv.org/abs/2605.17721.
Key facts
- EXG is an experience graph framework for self-evolving LLM agents.
- It organizes accumulated successes and failures into a structured, relational representation.
- EXG supports online, real-time graph growth during deployment.
- It is the first experience graph designed for self-evolving agents.
- The paper is published on arXiv with identifier 2605.17721.
- Existing approaches rely on ad hoc reflection or unstructured memory.
- Most deployed agents remain behaviorally static.
- The framework enables continuous learning and improvement over time.
Entities
Institutions
- arXiv