SkillTrace: Traversing a Query-Skill Graph for Composable LLM Agents
TL;DR
This paper introduces SkillTrace, which organizes the user query into a semantic hierarchy, matches skill queries and candidates, and propagates over the skill dependencies, and delivers consistent improvements across different backbone language models, demonstrating the generality and robustness of graph-based skill retrieval.
Abstract
Large language model agents increasingly solve complex tasks by composing reusable skills from a library. To address this, the key challenge is not merely to retrieve individually relevant skills, but to identify a complete and executable skill composition. In this paper, we argue that this problem can be solved in a graph with three levels: compositional relations among skill queries, similarity between queries and candidates in the skill library, and the dependencies among the selected candidates. We introduce SkillTrace, which organizes the user query into a semantic hierarchy, matches skill queries and candidates, and propagates over the skill dependencies. Experiments on SkillsBench and ALFWorld demonstrate that SkillTrace achieves state-of-the-art performance, reaching a success rate of 53.17% on SkillsBench and 91.43% on ALFWorld. SkillTrace also delivers consistent improvements across different backbone language models, demonstrating the generality and robustness of graph-based skill retrieval.