MUSE-Autoskill: Lifecycle-Managed Self-Evolving Skills
Source: HuggingFace daily papers (2026-05-27, 6 upvotes) · arxiv 2605.27366 arxiv: 2605.27366 Date: 2026-05-27 Raw: raw/huggingface/2026-05-27-muse-autoskill-self-evolving-agents-via-skill-creation-memor.md Tier: 2 (agentic systems, skill evolution)
TL;DR
Existing skill-creation methods treat skills as isolated, static artifacts, which caps reusability and long-term improvement. MUSE-Autoskill (Memory-Utilizing Skill Evolution) manages skills under a unified lifecycle: creation on demand, memory, management/selection, evaluation via unit tests and runtime feedback, and refinement. It adds skill-level memory that accumulates per-skill experience across tasks. On SkillsBench it shows initial evidence that lifecycle-managed skills improve task success, efficiency, reuse, and cross-agent transfer.
Key points
- Skills as long-lived, testable assets, not one-off artifacts: each skill carries accumulated experience and is validated by unit tests plus runtime feedback before reuse.
- Full lifecycle (create, store, select, evaluate, refine) rather than just creation.
- Cross-agent transfer is an explicit target and an evaluated dimension.
Relation to prior wiki state
MUSE-Autoskill is the constructive counter-move in the skill-evolution debate the wiki has been tracking. SkillEvolBench (05-26) delivered the harsh result: distilled skills under-transfer versus raw-trajectory reuse, and larger skill libraries hurt. MUSE-Autoskill's answer is that the failure is one of lifecycle management, not of the skill abstraction itself: if skills are unit-tested, refined from runtime feedback, and carry per-skill memory, the clutter and drift that SkillEvolBench blamed for the negative result are managed away. Whether lifecycle management actually beats raw-trajectory reuse on SkillEvolBench's frozen-deployment-under-context-shift protocol is the test it has not yet run (it evaluates on SkillsBench instead). It also pairs with SkillOpt (05-25, treat the skill document as a trainable parameter): SkillOpt optimizes the skill's text, MUSE-Autoskill manages the skill's lifecycle, and the open question is whether either survives SkillEvolBench's frozen deployment.
Gaps
Evaluated on SkillsBench, not the adversarial SkillEvolBench protocol, so it does not yet refute the negative result it implicitly responds to. "Initial evidence" framing signals early-stage results.