Agentic Self-Evolution for Large Language Models: Taxonomy, Techniques, and Applications

Abstract

The rapid advancement of Large Language Models (LLMs) has empowered autonomous agents with advanced reasoning, planning, and tool-use capabilities. However, traditional agent systems rely heavily on human-guided training. Despite the reduced data volume in post-training stages, the increasing demand for high-quality supervision creates a critical scalability bottleneck. This dependency not only restricts scalability but also confines agents to the upper bound of human expertise, severely limiting their evolutionary potential. To overcome these limitations, research on autonomous agents is shifting towards Agentic Self-Evolution, a paradigm in which agents autonomously coordinate their own improvement loops with minimal human supervision. To unify this rapidly expanding field, we propose a taxonomy that categorizes methods into three core components of an autonomous evolutionary cycle: (i) Model-Centric Self-Evolution, where agents enhance internal capabilities through inference scaling or parameter bootstrapping; (ii) Environment-Centric Self-Evolution, where agents achieve continual self-evolution by interacting with the environment to obtain external knowledge and experience-based feedback; and (iii) Model-Environment Co-Evolution, where agents and their environments jointly evolve through sustained interaction. We highlight Model-Environment Co-Evolution as a key emerging direction for Agentic Self-Evolution, where environment definitions and their co-evolution with agents will become central challenges in future research. We provide a systematic technical foundation of Agentic Self-Evolution, current implementations, and identify key technical challenges and promising research directions. All related resources, including research papers, open-source data, and projects, are collected for the community in https://github.com/XMUDeepLIT/ Awesome-Agentic-Self-Evolution.