Google's WikiSkill framework boosts AI agent performance across five benchmarks
In brief
- Google Research introduced WikiSkill, a framework giving AI agents persistent memory across iterations
- WikiSkill uses a three-layer architecture: Raw Layer, Wiki Layer, and Skill Layer
- Performance jumped 39.6 points on math reasoning and 26.1 points on spreadsheet tasks
- Average gain across five benchmarks was 12.0 points; ablation studies confirmed wiki criticality
- Skills learned by one model transferred effectively to different model families
How WikiSkill Works
The framework is built on a three-layer architecture, each serving a distinct purpose. The Raw Layer captures immutable execution traces, the Wiki Layer stores refined skill insights, and the Skill Layer applies those insights to guide future agent behavior. This structure lets agents accumulate knowledge rather than starting from scratch each iteration.
Previous skill-evolution methods for AI agents would generate useful insights during execution, then discard them after each cycle rather than carrying them forward. WikiSkill fixes that problem by maintaining a persistent repository of learned skills that agents can reference and refine over time.
Benchmark Results Across Five Domains
WikiSkill was validated across five benchmarks designed to test fundamentally different capabilities: math reasoning, web search, spreadsheet management, long-context question answering, and embodied interaction.
The gains were substantial. On LiveMathematicianBench, Gemini-3.5-Flash performance jumped from 33.0% to 72.6%. On SpreadSheetBench, scores climbed from 50.5% to 76.6%. Across all benchmarks, the average performance gain was 12.0 points.
Ablation studies confirmed that the persistent wiki is the critical component driving performance gains. Remove it, and the improvements largely disappear.
Cross-Model Skill Transfer
One finding stands out for its implications. Skills learned by one model outperformed self-evolved skills when applied to a different model entirely, including models from different families. This suggests that knowledge acquired by one AI system can transfer meaningfully to others, opening possibilities for shared skill repositories across the AI ecosystem.
The research was authored by Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, and Da-Cheng Juan from Google Research, alongside Tu Vu from Google Research and Virginia Tech.
