WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
WikiSkill is a framework that co-evolves agent skills with a persistent knowledge base (wiki). It separates raw execution experience, accumulated knowledge, and executable skills, continuously consolidating experience into the wiki for subsequent skill updates. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. 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. Evolved skills transfer effectively across models and model families.
Researchers introduced WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). It separates raw execution experience, accumulated knowledge, and executable skills, continuously consolidating experience into the wiki for subsequent skill updates. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. The study finds 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. Evolved skills transfer effectively across models and model families.
WikiSkill introduces a three-layer separation of raw execution experience, accumulated knowledge, and executable skills, with continuous consolidation of experience into a persistent wiki. This design enables systematic reuse of insights across iterations, unlike prior methods where insights remain scattered in optimization histories. The framework demonstrates consistent improvements over state-of-the-art skill-evolution methods and no-skill baselines, and shows that skill evolution complements model scaling, with smaller models with skills outperforming larger models without them.
The finding that smaller models with evolved skills can outperform substantially larger models without them has significant implications for cost-efficient AI deployment. It suggests that investing in skill evolution may reduce the need for ever-larger models, potentially lowering inference costs and enabling broader adoption. The transferability of evolved skills across models and model families indicates that skill libraries could become valuable reusable assets, similar to pre-trained model weights.
WikiSkill's approach could reduce the cost of building capable AI agents by enabling smaller models to achieve performance comparable to larger ones through evolved skills. This may lower infrastructure costs for AI service providers and make advanced agent capabilities accessible to organizations with limited compute budgets. The transferability of skills across models also suggests potential for a market in pre-built, domain-specific skill packages.
Observable next signals include: (1) open-source releases of WikiSkill code or skill libraries; (2) follow-up research applying WikiSkill to specific domains such as coding, robotics, or scientific discovery; (3) commercial adoption of skill-evolution frameworks in agent platforms; (4) benchmarks comparing skill-evolved small models against large models in production settings; (5) integration of persistent knowledge bases with other skill-learning methods.