SELM improves AI-agent skill reuse in four benchmarks, authors say

Mengxi Hu and eight co-authors published SELM, an AI-agent framework, as an arXiv cs.AI preprint on Oct. 20 after tests on four agent benchmarks.

The paper says SELM links three library functions—retrieval, updating and lifecycle control—to address a weakness the authors identify in earlier skill-evolution methods: weak reuse of learned skills across domains.

“SELM jointly manages skill retrieval, updating, and lifecycle by maintaining skill confidence and utility signals,” the authors wrote in the abstract.

Large-language-model agents are systems that use LLMs to plan and execute tasks. In this paper, “skills” refer to reusable task-specific knowledge that agents can draw on, rather than relearn from scratch for each new problem.

The authors say SELM uses utility-aware retrieval to choose useful skills, a Map-Reduce process to create updates from task experience, and lifecycle rules to keep or discard skills based on effectiveness. They say the framework outperformed in most settings across the four benchmarks, which covered different tasks and execution environments.

The evidence is preliminary. The arXiv paper is a preprint and has not been peer reviewed.

The preprint was submitted on Oct. 17 as arXiv:2610.24630v1 in artificial intelligence.

Source: arXiv cs.AI preprint, Oct. 20, 2026