WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

Google Research (with a Virginia Tech co-affiliation for senior author Tu Vu) proposes WikiSkill, a skill-evolution framework in which agent skills co-evolve with a persistent, compounding knowledge base (wiki) rather than being re-derived from scattered optimization histories. The agent workspace is split into three layers: an immutable raw/ layer of execution traces, a wiki/ layer of markdown pattern files (failure modes, successful strategies, workarounds) plus an evolution log and a skill-impact tracker, and a skills/ layer holding the active skill set. An orchestrated loop has a Wiki Maintainer consolidate traces into patterns, a wiki-informed Skill Proposer draft skill updates, and validation gating with rollback accept or reject them; the wiki is never reset between iterations, so rejected interventions aren't re-proposed and recurring errors are visible.

Across five benchmarks (LiveMathematicianBench, SealQA, SpreadsheetBench, OfficeQA, ALFWorld) and five inference models (Qwen-3.5-4B/9B, Qwen-3.6-27B, Gemma-4-31B, Gemini 3.5 Flash), WikiSkill beats SkillOpt, EvoSkill and Trace2Skill on every model's average — by 3.3 to 12.0 points over the strongest competitor — e.g. Gemini 3.5 Flash goes 33.0 → 72.6% on LiveMath and 50.5 → 76.6% on SpreadsheetBench; Qwen-3.6-27B 52.8 → 77.6% on ALFWorld. Gains grow with model scale (+12.3 / +17.5 / +23.9 average points across the Qwen sizes), yet skills let Qwen-3.5-9B (47.4%) outrun skill-less Qwen-3.6-27B (39.4%). Evolved skills transfer across models and families and often beat self-evolved ones (Qwen-3.6-27B skills lift Gemma-4-31B to 73.7% on LiveMath vs. 56.7% self-evolved), though model-specific workarounds can transfer negatively (Qwen-3.5-4B spreadsheet skills drop Gemini 3.5 Flash to 18.1%). Ablation: giving the Skill Proposer wiki access raises the Gemini 3.5 Flash average from 48.7% to 63.7%; giving the inference agent wiki access during rollouts hurts (63.7 → 60.9%). Skills are injected directly into the prompt, so retrieval/triggering at scale is left for future work; no code release announced.