SkillOpt: A New Method for Self-Evolving Agent Skills
A new paper on arXiv proposes SkillOpt, a systematic method for optimizing agent skills through text-space edits. Unlike current hand-crafted or loosely self-revised skills, SkillOpt treats skills as external state of a frozen agent, using a separate optimizer model to apply bounded edits based on scored rollouts. Edits are accepted only if they improve a validation score. The method includes a textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta updates for stability, adding zero inference-time model calls at deployment. Tested across six benchmarks, seven target models, and three execut.
Key facts
- SkillOpt is introduced as the first systematic controllable text-space optimizer for agent skills.
- It uses a separate optimizer model to make add/delete/replace edits on a single skill document.
- Edits are accepted only when they strictly improve a held-out validation score.
- The method includes a textual learning-rate budget and rejected-edit buffer.
- It adds zero inference-time model calls at deployment.
- Tested across six benchmarks, seven target models, and three execut.
Entities
Institutions
- arXiv