SkillReranker Improves LLM Agent Skill Selection.

Modern agent systems rely heavily on skill libraries to tackle complex tasks, but the increasing size of these libraries makes accurate skill selection challenging. Existing methods often struggle with ambiguous semantic matches and fail to account for dynamic task difficulty or skill applicability. To address these limitations, a new framework called SkillReranker has been proposed. This inference-time reranking system improves adaptive skill selection by first performing semantic decomposition of both the task and candidate skills. This process generates detailed subtask and execution-state descriptions, as well as transition-state descriptions for each skill. These descriptions are then used to build a directed acyclic execution graph, mapping intermediate task states to nodes and candidate skills to edges. SkillReranker identifies subtask intervals and employs a cross-encoder to score candidate skills within each interval, ultimately selecting the most suitable set of skills. Experiments show that SkillReranker significantly boosts task performance, reduces interaction steps, and lowers token consumption across various benchmarks.