What they did

The authors built a skill-evolution framework in which an agent’s execution experience is continuously consolidated into a persistent wiki. Later skill updates draw on this accumulated knowledge, keeping the execution history, distilled knowledge, and executable skill artifacts distinct.

They evaluated WikiSkill across multiple models and benchmarks, comparing it with state-of-the-art skill-evolution methods and agents operating without evolved skills. Ablation experiments tested whether retaining the persistent wiki was necessary for the reported gains.

Key findings

• WikiSkill consistently outperformed the evaluated prior skill-evolution methods and improved over no-skill baselines in most model–benchmark combinations.
• Larger models generally gained more from evolved skills, suggesting that skill evolution complements rather than substitutes for model scaling.
• Smaller models equipped with evolved skills could substantially outperform larger models without skills.
• Skills transferred across models and model families; in some cases, skills evolved by another model worked better than a model’s self-evolved skills. Ablations identified persistent wiki accumulation as a critical component.

Why it matters

The results suggest that agents can improve more effectively when experience is compiled into durable, reusable knowledge rather than repeatedly summarized from an unstructured history. Cross-model transfer also raises the possibility of building shared skill libraries whose value is not tied to the model that created them.

Caveats

The supplied abstract does not identify the benchmarks, models, effect sizes, computational costs, or statistical uncertainty, so the breadth and practical magnitude of the gains cannot be assessed here. It also remains unclear how well the wiki scales over long deployments, handles conflicting or outdated knowledge, or resists accumulating errors.