MUSE-Autoskill: Self-Evolving LLM Agents via Skill Lifecycle
Researchers propose MUSE-Autoskill Agent, a framework enabling large language model (LLM) agents to continuously improve task-solving by creating, reusing, and refining skills under a unified lifecycle. The system includes skill creation on demand, storage and reuse across tasks, efficient organization and selection, and evaluation through unit tests and runtime feedback. It introduces skill-level memory to accumulate experience per skill across tasks. Initial experiments on SkillsBench show promise. The paper is available on arXiv (2605.27366).
Key facts
- MUSE-Autoskill Agent is a skill-centric framework for LLM agents.
- It unifies skill creation, memory, management, evaluation, and refinement.
- Skills are created on demand and reused across tasks.
- Skill-level memory accumulates experience for each skill.
- Evaluation uses unit tests and runtime feedback.
- Experiments were conducted on SkillsBench.
- The paper is on arXiv with ID 2605.27366.
- The approach aims to improve reusability, reliability, and long-term improvement.
Entities
Institutions
- arXiv