Hugging Face researchers have introduced WikiSkill, a novel framework designed to co-evolve reusable agent skills with a persistent knowledge base, or “wiki.” The approach aims to systematically accumulate experience from agent interactions, enabling continuous improvement across different models and benchmarks.
Traditional skill-evolution methods often leave insights scattered across optimization histories, limiting their reuse. WikiSkill addresses this by separating raw execution experience, accumulated knowledge, and executable skills. Experience is continuously consolidated into the wiki, which subsequent skill updates can build upon, creating a virtuous cycle of learning.
In evaluations across diverse benchmarks and models, WikiSkill consistently outperformed state-of-the-art skill-evolution methods and improved over no-skill baselines in most settings. Notably, the research found that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them.
The study also revealed that evolved skills transfer effectively across models and model families, and that skills evolved by other models can outperform self-evolved skills. Ablation studies confirmed that persistent knowledge accumulation in the wiki is critical for effective skill evolution, highlighting the benefits of systematically refining agent experience for developing reusable and transferable skills.