What's this about?
LLM agents can induce skills from completed tasks and reuse them to grow more capable with experience. In practice, this paper finds, induced skills transfer unreliably across tasks — the direct follow-up to yesterday's finding that skills work through procedural anchoring.
What exactly makes a skill fail to transfer?
A skill learned from one task encodes both what generalizes and what was incidental to that task. "Check the config file before running the migration" might be a real procedure, or it might be true only of that one project. The agent that induced it can't tell the difference, because it only saw the case where it worked. Transfer fails when the incidental part is load-bearing in the skill's phrasing — and the more specific and useful a skill is for its original task, the more likely it carries specifics that don't hold elsewhere.
Why is this worth noticing?
- It's the reliability question underneath every self-improving agent result this month — accumulating skills only compounds if they transfer, and this says they often don't.
- It matches yesterday's retrieval finding. As skill pools grow, precision collapsed from 29.6% to 3.3%; if the retrieved skills also transfer unreliably, both halves of the pipeline degrade with scale.
- The decomposition angle in the title is the proposed fix: break induced skills into transferable and task-specific parts rather than storing them whole.