SkillLens: Hierarchical Skill Reuse for Cost-Efficient LLM Agents
SkillLens is a structured framework designed for the evolution of skills in LLM agents, featuring a four-layer graph that includes policies, strategies, procedures, and primitives. It effectively retrieves skills at various levels of detail to optimize both relevance and cost. When given a task, SkillLens identifies semantically related skill seeds and expands them through a degree-corrected random walk across the skill graph. A verifier then determines whether to accept, decompose, rewrite, or bypass each encountered unit. This method allows for the reuse of compatible subskills while only adapting components that are locally mismatched. The framework addresses the challenge of irrelevant context from broad skills and the high cost of rewriting entire skills. The concept is detailed in a paper on arXiv (2605.08386).
Key facts
- SkillLens is a hierarchical skill-evolution framework for LLM agents.
- It organizes skills into a four-layer graph: policies, strategies, procedures, and primitives.
- Skills are retrieved at mixed granularity.
- The framework uses degree-corrected random walk over the skill graph.
- A verifier decides acceptance, decomposition, rewriting, or skipping of each unit.
- It aims to balance relevance and cost in skill reuse.
- The paper is available on arXiv with identifier 2605.08386.
- Existing systems treat skills as flat, single-resolution prompt blocks.
Entities
Institutions
- arXiv