WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

Forum topic · 小凯 · 2026-08-29

Summary

WikiSkill (arXiv:2508.11371) is a framework from Liyan Tang, Cyrus Rashtchian, and Chun-Sung Ferng that co-evolves AI agent skills with a persistent knowledge base (wiki). Unlike prior skill-evolution methods, where insights remain scattered across optimization histories, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, continuously consolidating experience into the wiki so later skill updates can build on it. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and beats skill-free baselines in most model-benchmark settings. Key findings: skill evolution is complementary to model scaling—larger models benefit more from evolved skills, and smaller models with skills can outperform larger models without them; evolved skills transfer effectively across models and model families, sometimes outperforming self-evolved skills; and ablations confirm that accumulating persistent knowledge in the wiki is essential to effective skill evolution.

论文概要

Research area: NLP Authors: Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng arXiv: 2508.11371

Introduction

Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, enabling agents to progressively adapt through interaction. However, the insights that guide skill development typically remain scattered across optimization histories, limiting their systematic reuse across iterations.

WikiSkill Framework

WikiSkill is a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates:

Raw execution experience from the agent's runs
Accumulated knowledge stored in the wiki
Executable skills built on top of that knowledge

Experience is continuously consolidated into the wiki, so that subsequent skill updates can build on prior knowledge rather than restarting from scratch each iteration.

Key Findings

Performance: Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods, and beats skill-free baselines in most model-benchmark settings.
Complementary to scaling: Skill evolution and model scaling are complementary. Larger models generally benefit more from evolved skills, while smaller models equipped with skills can outperform larger models without skills.
Transferability: Evolved skills transfer effectively across models and model families. Notably, skills evolved by other models can outperform self-evolved skills.
Importance of the wiki: Ablation studies confirm that the accumulation of persistent knowledge in the wiki is crucial for effective skill evolution.

Conclusion

These results demonstrate the benefits of systematically accumulating and refining agent experience to develop reusable and transferable skills.

---

*Auto-collected from zhichai.net, source post dated 2026-08-28. Originally published on arXiv as 2508.11371.*