ResearchPod Summary
Modern AI agents extend their problem-solving capabilities by retrieving and invoking specialized, reusable skills from large repositories. However, accurate skill retrieval creates an asymmetric matching problem: users typically issue short, colloquial, and underspecified queries, whereas developers write skill documents using formal technical descriptions, parameter declarations, constraints, and usage examples. While pseudo-query generation can help bridge this representation gap, existing document-level approaches treat the skill document as flat, unstructured text. This leaves rich internal relations implicit, leading to queries that are topically relevant to the skill as a whole but lack capability grounding and parameter consistency. This paper investigates whether explicitly exploiting the internal structure of skill documents can produce more effective retrieval signals for agent skill retrieval.
To address the skill-structure utilization gap, the authors propose Skill2Query, a framework that parses skill documents into structured representations and generates pseudo-queries through a three-stage progressive pipeline. The approach is evaluated across four benchmarks (TheoremQA, LogicBench, ToolQA, and CHAMP) using nearly 30,000 skills from diverse domains to generate 700,000 category-diverse pseudo-queries.
First, Skill2Query parses each skill document into a Skill Knowledge Graph (SKG) containing four node types—Skill, Capability, Parameter, and Example—connected by explicit typed edges. This graph transforms implicit internal relationships into explicit generation constraints. Second, the framework executes a three-stage generation architecture:
The generated pseudo-queries can be deployed across three distinct retrieval settings: offline index augmentation, online query expansion, and query-skill supervision for retriever training. Empirical evaluations show that Skill2Query consistently improves sparse, dense, and skill-routing retrieval baselines, achieving an average Recall@1 gain of 6.70 percentage points across retrieval settings. Furthermore, Skill2Query-generated training data outperforms evaluated generation baselines in Recall@1 and nDCG@1, and end-to-end evaluations demonstrate that improved retrieval accuracy directly translates into higher task success rates for LLM agents.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.