SkillLens: Adaptive Multi-Granularity Skill Reuse for Cost-Efficient LLM Agents

Short summary

SkillLens proposes a hierarchical four-layer skill framework (policies, strategies, procedures, primitives) that retrieves skills at mixed granularity and uses a verifier to determine whether each unit should be accepted, decomposed, rewritten, or skipped. This approach balances relevance against computational cost. Across benchmarks, it achieves up to 6.31 percentage-point gains on bug localization and improves agent success rates from 45% to 51.31%.

• •Four-layer hierarchical skill framework enables mixed-granularity reuse for LLM agents
• •Verifier decides whether to accept, adapt, or skip each skill component based on relevance and cost
• •Up to 6.31% improvement in bug localization accuracy, 51.31% agent success rate on benchmarks

Generated with AI, which can make mistakes.