Google AI System WikiSkill Outperforms Large Models by Learning from Experience

Google researchers have developed an AI system called WikiSkill that improves performance by learning from experience. Released on August 27, 2026, WikiSkill uses a unique approach to update execution procedures based on successes and failures.

The key innovation is the separation of access rights: the wiki supports diagnosing and fixing procedures more than it supports directly answering tasks. This design allows the system to accumulate learning history without rolling back the wiki.

In experiments across five benchmarks and five models, WikiSkill outperformed existing skill-evolution methods, with a 9B model plus skills achieving a higher score than a 27B model without skills.

The evaluation targets included math reasoning, web search, spreadsheet tasks, long-document QA, and interactive environments. The system's performance was consistent across different models, but the ability to discover procedures and execute them are distinct capabilities.