Google's WikiSkill: Giving an Agent a Wiki Lifted Gemini 3.5 Flash from 49.5% to 68.1%

WikiSkill is a skill-evolution framework from Google Research, released on August 27, 2026 as arXiv 2608.27454, that compiles agent execution experience into a persistent wiki and rewrites skills from what that wiki records. Averaged across five benchmarks, Gemini-3.5-Flash rose from 49.5% with no skills to 68.1%, and Qwen-3.6-27B rose from 39.4% to 63.3%. The framework beat the existing skill-evolution methods Trace2Skill, EvoSkill, and SkillOpt at every model size tested. The active ingredient is not the skill file but the knowledge store kept beside it.

What a Skill File Cannot Hold

Agent skills package the procedures and domain knowledge for a task into reusable documents, and through 2026 they became the standard extension surface on major agent platforms. Recent work moved from hand-authored skills toward mining them automatically out of an agent's own execution history.

The paper's complaint is structural. The reasoning behind each skill edit, which approach failed and under what conditions another one held, ends up scattered across an optimization history with no path back into the next iteration. Each round then relearns the same lessons.

WikiSkill's answer is to split the workspace into three layers. A raw layer holds immutable execution traces from every iteration. A wiki layer holds a patterns directory, evolution logs, and a skill-impact tracker recording every proposal and its outcome. A skills layer holds only the executable procedures that get injected into the agent prompt. Skills stay short, and the wiki carries the reasoning behind them.

What Happens Inside One Iteration

A WikiSkill iteration is a fixed sequence of four roles that runs the agent, updates the wiki, proposes exactly one skill edit, and gates that edit on validation. An inference agent runs training tasks carrying only the current skills and leaves traces behind, with no wiki access during execution. A Wiki Maintainer samples both successful and failing traces and updates the patterns and logs. A Skill Proposer, reading the wiki and operating in ReAct mode, produces an atomic proposal that targets exactly one skill. Gating then accepts the proposal only if held-out validation performance improves.

What happens on rejection is the point of the design. The skills roll back, but the wiki persists. The fact that a direction lost on validation becomes input to the next round rather than being discarded. The paper's runs typically span seven to eight iterations.

Withholding the wiki from the inference agent is deliberate too. Handing an agent more reference material at execution time makes it impossible to tell whether a gain came from skill evolution or simply from extra context. Keeping the wiki on the offline optimization side removes that confound.

The Per-Benchmark Spread Behind the Average

The aggregate jump is large, but the five benchmarks behave differently. For Gemini-3.5-Flash, LiveMath more than doubled from 33.0% to 72.6% and SpreadSheet went from 50.5% to 76.6%. SealQA managed only 29.4% to 44.7%, and ALFWorld stayed exactly where it was at 85.9%.

There is a readable pattern in that spread. The benchmarks that moved most are the procedural ones. Mathematical manipulation and spreadsheet operations have a repeatable shape, so a procedure written down once transfers to the next instance. A task like SealQA, where each question demands different facts verified in different places, gives a procedure document much less to work with. Missing knowledge is not fixed by better process.

ALFWorld stalls for a different reason. Gemini-3.5-Flash was already at 85.9% without any skills, and the remaining 14.1 points were not the kind of error a procedure document reaches. On the same benchmark Qwen-3.5-4B went from 24.4% to 53.7% and Qwen-3.6-27B from 52.8% to 77.6%. Gains scale with available headroom rather than with task difficulty, which means the question worth asking before adopting this is not how much it lifts a score but whether the current failures are procedural at all.

The Wiki Itself Is Worth 15.0 Points

The ablation is the firmest result in the paper. With the wiki withheld from every component, Gemini-3.5-Flash averages 48.7%. With the wiki open to the Skill Proposer alone, which is the default configuration, it averages 63.7%. The 15.0-point gap comes from persistent knowledge accumulation on an otherwise identical pipeline.

The third configuration is the interesting one. Opening the wiki to both components drops the average to 60.9%. The paper uses this to justify its default, but the wider implication matters more: exposing a knowledge store at more points does not monotonically help. Accumulated context is a basis for judgment in one position and noise in another.

That maps onto a common mistake when teams wire internal documents into an agent. Pushing every execution log and decision record into the runtime prompt feels obvious, and WikiSkill does the opposite. The accumulated material shapes procedures offline, and only the short refined procedure travels with the agent at execution time.

When Someone Else's Evolved Skills Beat Your Own

The transfer experiments are the most practically useful result in the paper, and their sharpest number is Gemma-4-31B reaching 73.1% on LiveMath using skills authored by the far smaller Qwen-3.5-4B. On SpreadSheet, Qwen-3.5-9B scored 24.3% with no skills and 33.6% with skills it evolved itself, but reached 50.5% using skills evolved by Qwen-3.6-27B. On LiveMath, Gemma-4-31B scored 56.7% with self-evolved skills and 73.1% with skills produced by the far smaller Qwen-3.5-4B.

That second case inverts the expected direction. Skills from a larger model are not automatically better; a smaller model that fails more often while writing its procedures can produce something more useful to a larger one. More failure forces more explicit instructions, and explicit instructions travel across model families.

The paper's own framing is that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while a smaller model with skills can outperform a substantially larger one without them. Qwen-3.5-9B at 47.4% with WikiSkill against Qwen-3.6-27B at 39.4% without is the example. Reading that as "9B beats 27B" is a mistake, because the two sides are not comparable in cost. Each evolution round spends rollouts plus two LLM agent calls, and the paper does not put a token budget on that overhead.

The Limits the Authors Drew Themselves

The four limitations the WikiSkill authors state in the 2026 paper are more useful than the results table for anyone evaluating adoption. First, the experiments inject every evolved skill, so the retrieval step a real deployment needs was never evaluated. Second, strict validation gating can discard proposals that are neutral now and enabling later. Third, the wiki accumulates indefinitely with no automated pruning. Fourth, the benchmarks have relatively short horizons and do not cover tasks spanning hundreds of environment actions.

The third and fourth belong together. A store that only grows enlarges what the Skill Proposer must read each round. At seven or eight iterations that costs little, but an agent running for months in production under the same design could make the wiki the bottleneck. The 60.9% from opening the wiki to both components reads as the first signal in that direction.

Two things are established. Separating a layer that compiles execution history into persistent knowledge is worth 15.0 points on an otherwise identical pipeline, and the procedures it produces transfer across models and model families. Pruning policy and long-horizon behavior remain open.

Source: arXiv 2608.27454, "WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution" (Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan and Tu Vu, Google Research, submitted August 27, 2026; three-layer raw, wiki and skills workspace with an inference agent, Wiki Maintainer, Skill Proposer and gating with rollback, typically seven to eight iterations, benchmarks LiveMathematicianBench, SealQA, SpreadsheetBench, OfficeQA and ALFWorld, average scores of 26.2% to 38.5% for Qwen-3.5-4B, 29.9% to 47.4% for Qwen-3.5-9B, 39.4% to 63.3% for Qwen-3.6-27B, 41.3% to 54.9% for Gemma-4-31B and 49.5% to 68.1% for Gemini-3.5-Flash, gains over Trace2Skill, EvoSkill and SkillOpt, wiki ablation results of 48.7%, 63.7% and 60.9%, cross-model transfer results and the four limitations stated by the authors), compiled by ASAP.