WikiSkill: Compiling Agent Experience into Persistent KB for Skill Evolution (academy.dair.ai)

The recent introduction of WikiSkill presents a groundbreaking framework designed to enhance the skill evolution of AI agents by consolidating their experiences into a persistent knowledge base. This innovative approach not only allows agents to develop specialized skills from their accumulated knowledge but also ensures that these insights are systematically reused across different iterations. By decoupling raw execution experiences from executable skills and continuously integrating these experiences into the Wiki, the framework significantly improves upon existing state-of-the-art methods in skill evolution. For the AI/ML community, WikiSkill represents a significant leap in the way agent skills can be developed and utilized. The framework demonstrates that larger models benefit more from evolved skills, and intriguingly, smaller models can outperform considerably larger unskilled models when equipped with developed skills. Furthermore, the ability for skills to transfer effectively across different models and families could reshape how agents are trained and applied in diverse contexts. With ablation studies confirming that persistent knowledge accumulation is critical for effective skill evolution, WikiSkill opens up new pathways for the future of adaptive and reusable AI agents.