LLM Agents Improve Skill Transfer with Subtask-Level and Text Formats

Large language model (LLM) agents have the potential to learn and reuse skills from completed tasks, becoming more capable over time. However, the reliability of this skill transfer and its potential to sometimes hinder performance remain open questions. This research conducts a detailed study to understand how different methods of inducing skills influence their transferability across various tasks. The study specifically compares two key aspects: whether skills are induced at the task level or the subtask level, and whether they are stored as text or code. The findings indicate that subtask-level skills generally improve agent performance, while task-level skills often degrade it. Furthermore, text-based skills demonstrate better transferability than code-based skills. To explain these results, the researchers examined skill specificity (how closely a skill matches real tasks) and abstractness (how evenly its relevance spreads). They propose a "skill utility score" that combines these properties, which consistently correlates with task success during skill transfer. This score is valuable because it can be computed using only skill and task descriptions, without needing actual task execution, offering a practical diagnostic tool for evaluating skill memory.