SkillPager: Query-Adaptive Intra-Skill Navigation via Semantic Node Retrieval
Researchers have developed SkillPager, a novel two-stage framework designed to improve the efficiency of Large Language Model (LLM) agents when processing long procedural documents. SkillPager parses skill documents into semantic nodes and uses Maximal Marginal Relevance (MMR) for query-conditioned node selection, significantly reducing token usage while maintaining high context sufficiency. This approach demonstrates that the efficiency gains are primarily due to the typed semantic granularity of the nodes rather than the retrieval algorithm itself, outperforming graph-based baselines by over 12%. The findings highlight typed intra-document retrieval as a critical challenge for skill-based agents. AI
IMPACT Enhances LLM agent efficiency in handling complex documents, potentially improving performance in task-oriented applications.