WikiSkill

WikiSkill is a framework for AI agent self-improvement that introduces a persistent, structured knowledge base (the wiki) between raw execution experience and the agent's reusable skills. It organizes the agent's workspace into three layers: a Raw Layer that stores immutable execution traces from past task rollouts, a Wiki Layer that consolidates those traces into structured knowledge about failure modes, successful strategies, and rejected proposals, and a Skill Layer that holds the evolving procedural instructions the agent follows when solving tasks. The framework's core insight is that skill development becomes more effective when the agent maintains persistent, cumulative context about what it has already learned, rather than starting each iteration from scratch.

The WikiSkill loop runs four components in sequence. An Inference Agent executes tasks using the current skill set and produces execution traces. A Wiki Maintainer analyzes those traces alongside the existing wiki, distills new failure patterns and successful strategies, and updates the wiki accordingly. A Skill Proposer then reads the updated wiki and the latest traces to propose candidate skill updates. A Gating and Rollback mechanism evaluates each candidate against a held-out validation set, accepting modifications that improve performance and reverting those that do not. While individual skills can be rolled back, the wiki itself never resets. Rejected proposals are preserved as recorded history, so future iterations can build on knowledge of what was tried and why it failed.

The framework was introduced by Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, and Tu Vu at Google Research in August 2026, and evaluated across five benchmarks (LiveMath, SealQA, SpreadSheetBench, OfficeQA, and ALFWorld) using Qwen, Gemma, and Gemini model families ranging from 4B to 27B parameters. WikiSkill consistently outperformed prior skill-evolution methods and improved over no-skill baselines in most model-benchmark combinations, with particularly large gains on mathematical reasoning and spreadsheet manipulation tasks. On Gemini-3.5-Flash, average benchmark performance rose from 49.5% to 68.1%. On Qwen-3.6-27B, it rose from 39.4% to 63.3%.

Larger models generally benefit more from evolved skills. Within the Qwen family, WikiSkill improved average performance by 12.3%, 17.5%, and 23.9% on the 4B, 9B, and 27B variants respectively, with gains scaling with capability. Smaller models with WikiSkill can match or exceed the performance of substantially larger models without it: Qwen-3.5-9B with WikiSkill outperformed Qwen-3.6-27B without skills (47.4% vs. 39.4%). The skills themselves also transfer across model families. In some cases, skills evolved by a larger model perform better on a smaller recipient model than the smaller model's own self-evolved skills, suggesting that skill discovery and skill execution are distinct capabilities that benefit from separation.

The persistence of the wiki is what the ablation studies identify as the critical ingredient. When the Skill Proposer loses access to the accumulated wiki (effectively removing the Wiki Maintainer from the loop), average benchmark performance drops from 63.7% to 48.7%, a 15-percentage-point drop that confirms persistent knowledge accumulation is doing the heavy lifting. Giving the Inference Agent direct access to the wiki during training rollouts degrades performance, because the agent begins to retrieve task-solving knowledge directly from the wiki rather than from the skills, which produces less informative trajectories for skill development. The framework's design keeps the Inference Agent skill-only during training and reserves wiki access for the Skill Proposer, which edits the skills on behalf of the agent.