Search papers, labs, and topics across Lattice.
This paper introduces SkillTrace, a novel approach for composing reusable skills in large language model agents by utilizing a three-level query-skill graph. By organizing user queries into a semantic hierarchy and propagating skill dependencies, SkillTrace effectively identifies complete and executable skill compositions. Experimental results show that SkillTrace achieves state-of-the-art performance, with success rates of 53.17% on SkillsBench and 91.43% on ALFWorld, highlighting its robustness across various language model backbones.
SkillTrace redefines skill composition for LLM agents, achieving unprecedented success rates by leveraging a structured query-skill graph.
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.
Tsinghua AI