WikiSkill: Google's Framework for Reusable AI Skills

Google Research and Virginia Tech have developed a new framework called WikiSkill that helps AI agents learn from their past experiences without losing valuable knowledge. The system creates a structured wiki from the agent's past attempts, allowing it to reuse and build upon existing skills.

The researchers tested WikiSkill on five benchmarks, covering various tasks such as math reasoning, web search, and interactive household tasks. They found that WikiSkill outperformed existing skill-evolution methods by 3.3 to 12 percentage points.

One of the key features of WikiSkill is its ability to preserve diagnoses and failed interventions, allowing later iterations to build upon previous knowledge. This is in contrast to other systems that discard useful information after proposing a patch.