Large language model agents increasingly rely on skills as a lightweight way to acquire task-specific knowledge. Recent work has enabled these skills to evolve from task trajectories, but according to a new arXiv preprint, such methods often lack systematic management.
The paper, titled "Selm: Self-Evolving Library Management for Multi-Skill Agent Optimization," proposes Selm as a response to that gap. Based on the abstract, Selm focuses on managing a library of skills in a self-evolving manner, rather than simply generating new skills from trajectories.
Because this is a single source with a brief abstract, no experimental results or comparisons are available to report. The contribution, as described, is a management framework for multi-skill agents rather than a new skill-learning algorithm. The preprint is available on arXiv under identifier 2610.24630.