Google's WikiSkill Framework Boosts AI Agent Performance Across Five Benchmarks

Google Research has made significant strides in improving AI agent performance with its WikiSkill framework. This innovative approach allows agents to learn from experience and retain knowledge across iterations, a major improvement over previous methods that discarded valuable insights after each cycle.

The framework's three-layer architecture consists of the Raw Layer, which captures execution traces; the Wiki Layer, where raw data is consolidated into accumulated knowledge; and the Skill Layer, hosting executable procedures. Four key components keep the system running: an inference agent, a wiki maintainer, a skill proposer, and a validation gating mechanism.

WikiSkill was validated across five benchmarks designed to test math reasoning, web search, spreadsheet management, long-context question answering, and embodied interaction. The standout results came from Gemini-3.5-Flash, with notable improvements on LiveMathematicianBench (33.0% to 72.6%), SpreadSheetBench (50.5% to 76.6%), and an average gain of 12.0 points across all benchmarks.