WikiSkill proposes a persistent knowledge layer for evolving AI-agent skills

arXiv records “WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution” by Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan and Tu Vu. Version 1 was submitted on 27 August 2026. [1]

The authors describe WikiSkill as a framework for co-evolving agent skills and a persistent knowledge base. In their account, the wiki is intended to preserve insights that might otherwise remain scattered through prior optimization histories, so later skill updates can reuse accumulated knowledge. The supplied record supports this high-level design description but does not establish implementation details. [1]

The authors report that WikiSkill consistently outperformed state-of-the-art skill-evolution methods across diverse benchmarks and models, and improved on no-skill baselines in most tested model-benchmark combinations. They also report that skill evolution complements model scaling: larger models generally benefited more from evolved skills, while smaller models equipped with skills could outperform substantially larger models without them. These are the authors’ findings, not independently verified results. [1]

The abstract does not specify which benchmarks, models or baselines were used, nor does it give effect sizes, metrics or statistical analyses. As a result, the supplied evidence does not allow readers to assess the magnitude, reliability or practical scope of the reported scaling and performance effects. [1]

For the question of reuse across systems, the consequential claim is that evolved skills transfer effectively across models and model families. The authors further say that skills evolved by other models can outperform skills a model evolves for itself. If supported by detailed evaluation, that would suggest the persistent knowledge layer may help preserve skills that remain useful beyond the model that first developed them. This remains an author-reported claim. [1]

The authors also report that their ablation studies found persistent knowledge accumulation in the wiki to be critical for effective skill evolution. The supplied abstract does not state the transfer conditions, the models involved, the size of any advantage, or the ablation setup. It therefore does not establish when cross-model reuse works or why it succeeds. [1]

The preprint addresses the authors’ stated problem that insights guiding skill development can be dispersed across optimization histories and difficult to reuse systematically. Its proposed solution is a durable knowledge layer that accumulates and refines those insights before later skill updates draw on them. [1]

Important unanswered questions include how the wiki is kept accurate as it grows, what resources it requires, how it behaves when prior experience is misleading, and whether its reported gains and transfer effects can be reproduced independently. The supplied record does not answer these questions. [1]

This brief is based on the arXiv record for the version-1 preprint, which provides the title, author list, submission information and abstract. [1]