The WikiSkill framework addresses a limitation in current AI agent development where insights from skill optimization often remain scattered, hindering their systematic reuse. The new approach integrates a persistent knowledge base, referred to as a "wiki," that co-evolves with the agent's skills. This wiki consolidates experience, enabling subsequent skill updates to build upon accumulated knowledge.
The framework organizes an agent's operational space into three distinct layers. The "Raw Layer" serves as an immutable repository, storing every interaction an agent undertakes during training. This includes reasoning steps, tool calls, and outputs, ensuring that all raw data remains intact for future analysis. Above this, the "Wiki Layer" distills the raw data into structured insights, documenting both successful strategies and failure patterns. This knowledge layer is designed to be permanent and cumulative, growing with each iteration without being reset. The third layer, the "Skill Layer," contains the active procedural instructions that guide the agent's task execution. Unlike the wiki, skills in this layer can be modified or rolled back if updates negatively impact performance.
The process within WikiSkill involves an inference agent executing tasks using its current skills, generating execution traces. A "Wiki Maintainer" component then analyzes these traces to identify patterns of success and failure, subsequently updating the wiki with these findings. A "Skill Proposer" then uses the updated wiki and execution traces to generate or modify candidate skills.
Empirical evaluations of WikiSkill were conducted across five benchmarks, encompassing tasks such as mathematical reasoning (LiveMath), web search (SealQA), and embodied tasks (ALFWorld). The results indicate that WikiSkill consistently improves agent performance, often surpassing existing skill-evolution methods like Trace2Skill and SkillOpt. For instance, the framework boosted the performance of Gemini-3.5-Flash from 49.5 percent to 68.1 percent and Qwen-3.6-27B from 39.4 percent to 63.3 percent on average. On specific benchmarks, the improvements were more pronounced, with Gemini-3.5-Flash showing a jump from 33.0 percent to 72.6 percent on LiveMath.
A key finding from the research is the complementary relationship between model scale and skill evolution. Larger models generally exhibited greater benefits from evolved skills. However, smaller models equipped with these evolved skills were observed to outperform substantially larger models that lacked such skills. This suggests that skill evolution can serve as a method to enhance the capabilities of smaller models, potentially closing the performance gap with larger counterparts.
Furthermore, the study found that evolved skills can transfer effectively across different models and even model families. In some cases, skills evolved by one model demonstrated superior performance when used by another model, even outperforming skills that the latter model evolved itself. This finding suggests the potential for shared skill registries, where skills developed by one agent could benefit a broader range of AI systems. Ablation studies confirmed that the persistent accumulation of knowledge within the wiki is a critical factor for effective skill evolution.
This work draws inspiration from the concept of an "LLM Wiki," which advocates for compiling experience into cumulative, persistent knowledge. WikiSkill applies this principle to the automated development of AI agent skills, providing a structured mechanism for agents to learn from their operational history and continuously refine their abilities. The researchers involved in this project include Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, and Da-Cheng Juan from Google Research, and Tu Vu from Google Research and Virginia Tech.