Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents

TL;DR - This controlled study finds that LLM agents transfer induced skills more reliably when skills represent subtasks rather than whole tasks and are stored as text rather than code. It also proposes an execution-free utility score for diagnosing whether a skill memory is likely to help on new tasks.

• Task-level skills generally lowered performance below the no-memory baseline, while subtask-level skills improved it on average.
• Text-form skills transferred more effectively than code-form skills.
• Skill success depended on the combined effects of specificity and abstractness; neither property was predictive alone.
• The proposed skill utility score consistently correlated with transfer success using only skills and task descriptions.