WikiSkill: Persistent Wiki as a Knowledge Accumulation Layer for Skill Evolution

WikiSkill introduces a persistent wiki as an explicit, revisable knowledge store that is co-evolved with agent skills, separating raw experience, accumulated knowledge, and executable skills. The paper reports consistent gains over existing skill-evolution methods and no-skill baselines in most model–benchmark settings, positive interactions between evolved skills and model scaling, cross-model transfer of skills, and an ablation result attributing the gains to persistent knowledge accumulation.

Hugging Face Daily Papers — WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

August 29, 20261 source60% confidence

65Strategic

Strategic score

Evidence strength40

Technical novelty76

Commercial potential55

WikiSkill, introduced in an arXiv preprint, proposes a skill-evolution framework that co-evolves an AI agent's executable skills with a persistent wiki of accumulated knowledge. The framework separates raw execution experience, accumulated knowledge, and executable skills, and continuously consolidates experience into the wiki. The abstract reports consistent improvements over state-of-the-art skill-evolution methods across diverse benchmarks and models, and better-than-no-skill performance in most model–benchmark settings.¹

The substantive novelty is the persistent wiki itself. Earlier automatic skill discovery typically treats optimization histories as epiphenomenal byproducts; WikiSkill instead makes consolidated knowledge a first-class object that survives across skill-update iterations. By routing experience through an external, editable store, the framework moves the bottleneck of cross-iteration memory out of the model weights and into an artifact that can be directly inspected and revised.¹

The paper's most consequential claims are that evolved skills complement model scaling and that smaller models equipped with skills can exceed substantially larger models without them. If these effects are real, they would change the practical cost model of agent deployment by making skill libraries a substitute for some parameter-count investment. The decisive counter-evidence would be a failure to reproduce these interactions on held-out tasks or a demonstration that the wiki contributes nothing beyond logging; neither is available in the abstract.¹

Context

WikiSkill separates the pipeline into three levels—raw execution experience, accumulated knowledge, and executable skills—and maintains a continuous consolidation path from experience into the wiki. Subsequent skill updates can build on the accumulated wiki rather than starting from scratch. An ablation is reported in which persistent knowledge accumulation is critical for effective skill evolution.¹

Caveats

AI hypothesisGiven that only the abstract is available, we hypothesize that the reported cross-model transfer benefits depend on the diversity and capability level of the source models; a source pool that is too homogeneous might collapse the advantage. This is a conjecture, not a finding, and would need to be tested against the paper's eventual supplementary material.¹

Source text

Hugging Face Daily Papers — WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution · August 27, 2026

Computer Science > Artificial Intelligence

Title:WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

Abstract:Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, which enables 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. We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, while continuously consolidating experience into the wiki, which subsequent skill updates can build on. 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. We find 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. We also find that evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills. Finally, our ablation studies confirm that persistent knowledge accumulation in the wiki is critical for effective skill evolution. These results demonstrate the benefits of systematically accumulating and refining agent experience for developing reusable and transferable skills.

Access Paper:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The word 'complements' in the scaling result covers an interaction that the abstract does not quantify: larger models 'generally' benefit more from evolved skills, yet the per-size effect size and task type are unspecified. Without this decomposition, the practical recommendation from this result is underdetermined—one cannot tell whether skills widen or narrow gaps between model sizes. The same passage should not be read as evidence that skills remove scaling requirements.

The report that skills evolved by other models can outperform self-evolved skills implies an interoperability assumption: for cross-model transfer to be meaningful, skill representations must be serialized in a form that survives model-family changes. The abstract does not specify the source–target pairs, whether the effect is asymmetric, or whether transfer is mediated by the wiki or by the skill container itself. This is a substantive open question about the mechanism.

Every headline comparison in the source is qualified rather than universal: gains are 'consistent' but the no-skill baseline improves in 'most' settings, not all. Since no benchmark identity, baseline names, or confidence intervals are given, the claim should be treated as promising rather than established. The abbreviation 'SOTA' also implies a moving target that is unspecified here.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

August 27, 2026endistinct angle

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

↑ Back to top