This survey presents a comprehensive overview of self-evolution approaches in large language models (LLMs). The paper introduces a conceptual framework for self-evolution, describing it as an iterative cycle of experience acquisition, refinement, updating, and evaluation. The framework outlines how LLMs can autonomously learn from their own experiences, refine them, and improve their performance through continuous learning. The study categorizes the evolution objectives of LLMs and LLM-based agents, summarizing existing methods and providing a taxonomy for each module. It also identifies existing challenges and proposes future directions to enhance self-evolution frameworks, offering critical insights for researchers aiming to accelerate the development of self-evolving LLMs. The paper discusses the evolution objectives of LLMs, which include evolving abilities such as instruction following, reasoning, math, coding, and role-play, as well as evolution directions like improving performance, adapting to feedback, expanding knowledge bases, and reducing bias. It then delves into the process of experience acquisition, which involves generating new tasks, solving them, and obtaining feedback. The paper further explores solution evolution, where models generate solutions that are relevant and informative, and feedback mechanisms that help refine the learning process. Experience refinement is then discussed, focusing on filtering and correcting experiences to ensure quality and reliability. The paper outlines various methods for filtering, including metric-based and metric-free approaches, and for correcting, including critique-based and critique-free methods. Finally, the paper covers the updating phase, which involves in-weight and in-context learning methods to improve model performance. In-weight learning includes techniques like replay-based, regularization-based, and architecture-based methods, while in-context learning leverages the model's ability to learn from experiences without extensive training. The survey highlights the potential of self-evolving LLMs to adapt, learn, and improve autonomously, similar to human evolution in response to changing environments. It emphasizes the importance of self-evolution in overcoming the limitations of static, data-bound models and moving towards more dynamic, robust, and intelligent systems. The paper concludes by outlining future research directions and the need for systematic organization and analysis of self-evolution methods to advance the field.This survey presents a comprehensive overview of self-evolution approaches in large language models (LLMs). The paper introduces a conceptual framework for self-evolution, describing it as an iterative cycle of experience acquisition, refinement, updating, and evaluation. The framework outlines how LLMs can autonomously learn from their own experiences, refine them, and improve their performance through continuous learning. The study categorizes the evolution objectives of LLMs and LLM-based agents, summarizing existing methods and providing a taxonomy for each module. It also identifies existing challenges and proposes future directions to enhance self-evolution frameworks, offering critical insights for researchers aiming to accelerate the development of self-evolving LLMs. The paper discusses the evolution objectives of LLMs, which include evolving abilities such as instruction following, reasoning, math, coding, and role-play, as well as evolution directions like improving performance, adapting to feedback, expanding knowledge bases, and reducing bias. It then delves into the process of experience acquisition, which involves generating new tasks, solving them, and obtaining feedback. The paper further explores solution evolution, where models generate solutions that are relevant and informative, and feedback mechanisms that help refine the learning process. Experience refinement is then discussed, focusing on filtering and correcting experiences to ensure quality and reliability. The paper outlines various methods for filtering, including metric-based and metric-free approaches, and for correcting, including critique-based and critique-free methods. Finally, the paper covers the updating phase, which involves in-weight and in-context learning methods to improve model performance. In-weight learning includes techniques like replay-based, regularization-based, and architecture-based methods, while in-context learning leverages the model's ability to learn from experiences without extensive training. The survey highlights the potential of self-evolving LLMs to adapt, learn, and improve autonomously, similar to human evolution in response to changing environments. It emphasizes the importance of self-evolution in overcoming the limitations of static, data-bound models and moving towards more dynamic, robust, and intelligent systems. The paper concludes by outlining future research directions and the need for systematic organization and analysis of self-evolution methods to advance the field.