WikiSkill: Compiling Agent Experience into Knowledge for Skill Evolution (arxiv.org)

Researchers have introduced WikiSkill, a groundbreaking framework designed to enhance the evolution of AI agent skills by systematically compiling experience into a persistent knowledge repository. This innovation addresses a critical limitation in existing methods, which often scatter valuable insights across various optimization attempts, making it challenging to reuse knowledge consistently. By separating raw execution experiences, accumulated knowledge, and executable skills, WikiSkill allows for continuous consolidation of agent experiences into a centralized wiki that improves skill development over time. The significance of WikiSkill lies in its demonstrated capability to outperform leading skill-evolution methods while providing substantial benefits in diverse AI models. Notably, larger models gain more from evolved skills, and smaller models can even surpass larger counterparts without such skills, highlighting the effectiveness and transferability of these evolved capabilities across different models and families. The study underscores that persistent knowledge accumulation is crucial for effective skill evolution, marking a pivotal step toward creating more adaptable and efficient AI agents that can leverage shared experiences for improved performance.