Google Releases WikiSkill Paper

Google Research released a WikiSkill preprint for long-running agents, which maintains execution histories as a knowledge base and compiles them into reusable skill files. The system consists of an original execution trace, a wiki that integrates successful strategies and failure patterns, and a proposer that adopts only improved skills through a validation gate.

In benchmark results, the Qwen-3.5-9B model with skills applied achieved an average score of 47.4%, surpassing the 39.4% score of the Qwen-3.6-27B model without skills. However, the 27B model's performance rose to 63.3% when its own skills were applied, suggesting that procedural memory and model capacity are complementary.

Risks of Cross-Model Skill Transfer

In the ALFWorld benchmark, performance improved when the 9B model used skills evolved by the 27B model. However, applying spreadsheet skills written by a weaker model to Gemini-3.5-Flash caused the score to plummet from 50.5% to 18.1%. This demonstrates that fragile workarounds learned by weaker models can become harmful instructions when executed literally by stronger models.

The study was conducted by directly inserting skills into prompts within limited benchmark environments, and did not include automatic wiki cleanup features.