Paper Overview

Research area: AI Agent / Skill Evolution Authors: Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu Published: 2026-08-30 arXiv: 2608.27454

---

WikiSkill: When AI Learns to Write Its Own Skill Manual

The Problem: Lost Craftsmanship

Imagine a master carpenter retiring after forty years. His judgment of wood grain, his fine adjustments of joinery angles—all of it lived in his hands and eyes, never systematically documented. His apprentices must start from zero, repeating the master's early mistakes.

This mirrors today's AI Agent dilemma. We train increasingly capable assistants that write code, search the web, and analyze data—but after each conversation ends, hard-won experience ("this API needs a specific parameter format," "this site searches best with particular keywords") evaporates into chat logs. The same AI, in its next session, must learn from scratch again.

WikiSkill tackles exactly this problem of 'craft inheritance.'

What Is a Skill?

Think of learning to cook braised pork. After your first attempt, you note: "add cooking wine when blanching to remove gaminess," "better too light than too dark with caramelized sugar—it turns bitter." These distilled insights are a skill—reusable knowledge that makes the next attempt faster and more reliable.

AI agent skills work the same way. After completing tasks, an agent accumulates experience: which search filters work, how to read papers efficiently, how to handle API rate limits. If captured, stored, and recalled, the agent has acquired a skill.

Prior Methods: Sticky Notes Everywhere

Existing approaches let agents repeatedly perform similar tasks and extract general patterns from successful execution traces. The problem: each skill update is based only on recent traces, like recipe notes scribbled on sticky notes plastered across a kitchen with no unified notebook.

As the paper puts it, "the insights that guide skill development typically remain scattered across optimization histories." Consequences:

• The same mistakes get repeated
• Good experience fails to transfer across tasks
• Skills lack connections and never form a knowledge system

WikiSkill: Give the AI a Living Encyclopedia

WikiSkill's core idea in one sentence: "Write experience into a wiki, and let skills evolve on top of accumulated knowledge."

| Component | Analogy | Function | |---|---|---| | Raw Execution Experience | Employee work logs | Records the raw process of each task | | Accumulated Knowledge (Wiki) | Company knowledge base / wiki | Continuously consolidated experience | | Executable Skills | Standard operating procedures (SOPs) | Executable workflows generated from the wiki |

The elegance lies in separation. Traditional methods jump straight from work logs to SOPs. WikiSkill inserts an intermediate knowledge layer: after each task, experience is consolidated and added to the wiki; skill updates are then generated from the wiki's full accumulated knowledge.

• Traditional: after each cooking session, jot notes on a sticky note; next time, read the most recent notes.
• WikiSkill: after each session, organize insights into a steadily growing cooking encyclopedia; next time, consult the encyclopedia.

How the Wiki Gets Written

From the abstract, several key mechanisms can be inferred:

4.1 Experience Consolidation — After each task, raw records (dialogue history, tool-call sequences, success/failure flags) are processed into structured knowledge entries: what problems arose, how they were solved, what is reusable, what pitfalls to avoid.

4.2 Skill Update — When skills need updating, the system queries the wiki and generates new or improved skill descriptions. As the paper notes, "subsequent skill updates can build on" the wiki—skills evolve on a foundation of knowledge rather than in isolation.

4.3 Co-evolution — Wiki and skills reinforce each other: a richer wiki yields better skill updates; better skills improve agent performance, producing higher-quality experience that further enriches the wiki. A virtuous cycle.

Experimental Results: Can Small Models Beat Big Ones?

The results are striking, with some counterintuitive findings:

5.1 Consistent superiority — Across multiple benchmarks and models, WikiSkill consistently beats current state-of-the-art skill evolution methods, confirming the knowledge-layer design works.

5.2 Skill evolution complements model scaling — The most interesting finding:

> "Larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them."

In other words: small model + skills > large model without skills. Like a seasoned engineer with an ordinary degree beating a top-school rookie. For budget-constrained applications, a mid-tier model plus a WikiSkill-style skill system can match or exceed frontier-model performance.

5.3 Cross-model transferability —

> "evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills."

Like a senior engineer's operations manual that new hires use directly—sometimes more effectively than their own explorations. This transferability is precisely the value of the wiki mechanism.

5.4 Ablations — Persistent knowledge accumulation in the wiki is essential; removing the knowledge layer significantly degrades results.

Deeper Reflection: AI's 'Cultural Inheritance'

WikiSkill evokes humanity's cumulative culture. Humans dominate not because individuals are brilliant, but because we accumulate knowledge—no generation reinvents the wheel. From cuneiform to printing, libraries to the internet, we keep improving how knowledge is preserved.

WikiSkill builds essentially the same mechanism for AI. A skill learned by one agent can be inherited and improved by others—a prototype of an AI civilization. It's still simple: it doesn't proactively ask questions, make cross-domain associations, or achieve Newton-style breakthroughs by "standing on giants' shoulders." But the direction—systematically accumulating, transmitting, and evolving knowledge—is clearly a key path toward more advanced intelligence.

Summary

Core contributions of WikiSkill: 1. A three-layer architecture separating raw experience, knowledge, and skills 2. More systematic skill evolution through continuous wiki accumulation 3. Evidence that skill evolution complements (and can substitute for) model scaling 4. Demonstrated skill transferability, opening new possibilities for AI collaboration

One-line verdict: This is not a paper about making AI smarter, but about making AI better at learning. As AI capabilities race forward, efficiently accumulating, transmitting, and reusing knowledge may matter more than simply stacking parameters.

---

Reference:

• Tang, L., Rashtchian, C., Ferng, C. S., Tomkins, A., Juan, D. C., & Vu, T. (2026). WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution. *arXiv preprint arXiv:2608.27454*.