Skill Transfer

Skill transfer

Skill transfer is the practice of taking a procedural skill (a structured instruction document or reusable routine that an AI agent has learned or evolved for some task domain) and applying it through a different agent or underlying model than the one that produced it. The transfer can be across model scales (a skill evolved on a 27B model used by a 9B model), across model families (a Qwen skill used by Gemma, or vice versa), or across agents with different tool access and prompt conventions. The phenomenon matters because it implies that the artifact of skill evolution is partly independent of the model that produced it, which has implications for how skill libraries are built, shared, and priced.

Empirical evidence for cross-model skill transfer emerged in 2026, most prominently in the WikiSkill evaluations from Google Research. There, skills evolved by Qwen-3.6-27B were applied through Qwen-3.5-9B on ALFWorld, raising the smaller model's score to 70.2%, well above the 63.4% it achieved using its own self-evolved skills. The same pattern held across families: skills evolved by one model family often improved a different family, and in several cases the transferred skill outperformed the recipient's own evolved skill. The interpretation is that skill discovery and skill execution are separate capabilities. Discovering what works requires the search and reasoning competence of a stronger model; executing a well-written skill requires only the procedural competence the skill itself describes.

This separation has practical consequences for how agent systems should be designed. If a skill evolved by a large proprietary model can be reused by a smaller open model, then skill libraries become a kind of distillation mechanism, a way to package distilled expertise as procedural text rather than parameter updates. Distributing a Markdown skill file is dramatically cheaper than distilling weights, and the file is inspectable, version-controlled, and easy to audit. The trade-off is that transferred skills may carry assumptions about tool conventions, prompt structure, or error modes that the recipient model handles differently, so the gain is real but not guaranteed across all recipient-task combinations.

Skill transfer also intersects with skill libraries and prompt-engineering practice more broadly. A curated library of high-quality skills, evolved under a strong model, validated on held-out tasks, and released as reusable artifacts, is a form of model-independent knowledge distribution that sits between fine-tuning (which bakes knowledge into weights) and retrieval-augmented generation (which surfaces relevant documents at inference). The skill library's value comes from being portable: the same skill file can be loaded by many different agents and improved once but benefit many. This is the use case that motivates treating agent skills as first-class artifacts to evolve rather than as one-off prompts to author.

The limits of skill transfer are still being mapped. The WikiSkill results show that transfer works well when the recipient model has enough capability to execute the multi-step strategy the skill describes. Smaller models like Qwen-3.5-4B sometimes fell back to default behavior on long-context tasks despite being given an evolved skill, suggesting a floor on the recipient's competence below which even well-written skills do not help. The taxonomy of when transfer helps, when it hurts, and which kinds of skills transfer across which kinds of recipient models is an open research question for the field.