SkillAligner Adapts AI Agent Skills for Better Task Performance.

Language agents often rely on general-purpose skills to perform tasks, but these retrieved skills can sometimes conflict with the specific task, execution environment, or other skills. This issue, termed "skill-execution misfit," hinders effective agent performance. To address this, researchers propose SkillAligner, a novel framework that adapts retrieved skills dynamically at execution time without requiring additional training. Instead of treating skills as rigid instructions, SkillAligner views them as adaptable drafts. It performs a one-time joint adaptation process before execution. During adaptation, SkillAligner specializes relevant skill fragments to meet task requirements, aligns their underlying assumptions with the available execution interface, and resolves any dependencies, conflicts, or redundancies among multiple skills. The refined guidance is then consolidated into a compact execution guide, which the agent reuses throughout its trajectory. Extensive experiments across various agent benchmarks and model backbones demonstrate that SkillAligner significantly boosts task performance, mitigates skill-induced regressions, and lowers overall inference costs.