SkillLens: Adaptive Multi-Granularity Skill Reuse for Cost-Efficient LLM Agents
Researchers have introduced SkillLens, a novel hierarchical skill-evolution framework designed to enhance the cost-efficiency and performance of Large Language Model (LLM) agents. Addressing the limitations of existing systems that treat skills as flat, single-resolution blocks, SkillLens organizes skills into a four-layer graph comprising policies, strategies, procedures, and primitives. The system retrieves semantically relevant skill seeds and expands them via degree-corrected random walks, using a verifier to determine whether to accept, decompose, rewrite, or skip specific units. This approach allows agents to reuse compatible subskills while adapting only mismatched components, theoretically ensuring sublinear costs under sparse mismatch assumptions. Empirical evaluations on MuLocbench and ALFWorld benchmarks demonstrate significant improvements over strong baselines. SkillLens achieved a 6.31 percentage-point gain in Acc@1 for bug localization and increased agent success rates from 45.00% to 51.31%. This development represents a significant advancement in optimizing procedural experience reuse for AI agents, balancing relevance and computational cost effectively.
Wire timeline
SkillLens: Adaptive Multi-Granularity Skill Reuse for Cost-Efficient LLM Agents
Researchers have introduced SkillLens, a novel hierarchical skill-evolution framework designed to enhance the cost-efficiency and performance of Large Language Model (LLM) agents. Addressing the limitations of existing systems that treat skills as flat, single-resolution blocks, SkillLens organizes skills into a four-layer graph comprising policies, strategies, procedures, and primitives. The system retrieves semantically relevant skill seeds and expands them via degree-corrected random walks, using a verifier to determine whether to accept, decompose, rewrite, or skip specific units. This approach allows agents to reuse compatible subskills while adapting only mismatched components, theoretically ensuring sublinear costs under sparse mismatch assumptions. Empirical evaluations on MuLocbench and ALFWorld benchmarks demonstrate significant improvements over strong baselines. SkillLens achieved a 6.31 percentage-point gain in Acc@1 for bug localization and increased agent success rates from 45.00% to 51.31%. This development represents a significant advancement in optimizing procedural experience reuse for AI agents, balancing relevance and computational cost effectively.
cs.AI updates on arXiv.org