Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents
Subtask-level, text-written skills transferred better than whole-task or code memories in the authors’ controlled agent study.
The paper finds that task-level skills often pushed agents below a no-memory baseline, while subtask-level skills improved performance on average. Text skills transferred more reliably than code skills. The authors also propose a “skill utility” score based on specificity and abstractness, computed from skills and task descriptions before running new tasks. ArXiv · AI/CL/LG's note
The paper finds that task-level skills often pushed agents below a no-memory baseline, while subtask-level skills improved performance on average. Text skills transferred more reliably than code skills. The authors also propose a “skill utility” score based on specificity and abstractness, computed from skills and task descriptions before running new tasks. ArXiv · AI/CL/LG's note
score 4