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

Authors: Yongliang Miao, Ziyang Yu, Liang Zhao et al.

arXiv · PDF

Summary

SkillLens organizes skills into a four-layer hierarchy (policies → strategies → procedures → primitives) and retrieves them at mixed granularity to balance relevance and cost. Given a task, it retrieves seed skills, expands via random walk over the skill graph, then uses a verifier to decide for each node whether to accept, decompose, rewrite, or skip. This lets agents reuse compatible subskills directly while adapting only mismatched parts. The system also refines skills and the verifier over time. Results show up to 6.31 percentage-point accuracy gains on bug localization and improved agent success on ALFWorld.

Main takeaways:

• Hierarchical skill graph (four layers from high-level policies down to primitives) enables mixed-granularity retrieval
• Expand via degree-corrected random walk, then verify each node: accept whole, decompose, rewrite, or skip
• Mixed-granularity adaptation has sublinear cost when mismatches are sparse (theoretical result)
• Evolutionary update rule monotonically improves validation objective to local optimum
• Consistent gains over flat skill baselines: +6.31pp on bug localization, +6.31pp agent success (45% → 51.31%) on ALFWorld

Relevance

Tangential—focuses on hierarchical skill reuse for agentic tasks, not persona installation. Could be loosely relevant if I ever think about whether behavioral patterns (like markers) have hierarchical structure or whether partial persona reuse is possible, but no direct connection to my current marker-implantation and leakage experiments.

Abstract

arXiv:2605.08386v1 Announce Type: new Abstract: Skill libraries have become a practical way for LLM agents to reuse procedural experience across tasks. However, existing systems typically treat skills as flat, single-resolution prompt blocks. This creates a tension between relevance and cost: injecting coarse skills can introduce irrelevant or misleading context, while rewriting entire skills is expensive and often unnecessary. We propose SkillLens, a hierarchical skill-evolution framework that organizes skills into a four-layer graph of policies, strategies, procedures, and primitives, and retrieves them at mixed granularity. Given a task, SkillLens first retrieves semantically relevant skill seeds, expands them through degree-corrected random walk over the skill graph, and then uses a verifier to decide whether each visited unit should be accepted, decomposed, rewritten, or skipped. This enables the agent to reuse compatible subskills directly while adapting only locally mismatched components. To improve the system over time, SkillLens further refines multi-granularity skills and verifier in order to improve its routing decisions. We provide theoretical analysis showing that mixed-granularity adaptation incurs sublinear cost under sparse mismatch assumptions and that the evolutionary update rule monotonically improves the validation objective until a local optimum. Across MuLocbench and ALFWorld, SkillLens consistently improves over strong skill-based baselines, achieving up to a 6.31 percentage-point Acc@1 gain for bug localization and raising agent success rate from 45.00% to 51.31%.