SkillLens framework lets LLM agents reuse skills with 6.31% accuracy boost

Existing skill libraries for LLM agents treat skills as flat, single-resolution prompt blocks, creating a tradeoff between relevance and cost. Coarse skills introduce irrelevant context; rewriting entire skills is expensive. Researchers from multiple institutions present SkillLens, a hierarchical framework that organizes skills into a four-layer directed graph: policies, strategies, procedures, and primitives. Given a task, SkillLens retrieves semantically relevant skill seeds, expands them via degree-corrected random walk over the graph, and uses a verifier to decide whether each visited node should be accepted, decomposed, rewritten, or skipped. This enables the agent to reuse compatible subskills directly while adapting only locally mismatched components.

SkillLens also improves over time by refining its multi-granularity skills and verifier to optimize routing decisions. Theoretical analysis shows mixed-granularity adaptation incurs sublinear cost under sparse mismatch assumptions. Empirically, on MuLocbench, SkillLens achieves up to a 6.31 percentage-point Acc@1 gain for bug localization, and on ALFWorld it raises agent success rate from 45.00% to 51.31%. The framework outperforms strong skill-based baselines without requiring full skill rewrites, making it a practical step toward cost-efficient, adaptive AI agents.