SKILLER: Reinforcement Learning Framework for Skill Extraction in Small Language Models

A new arXiv paper (2608.10538v1) introduces SKILLER, a natural-language-driven reinforcement learning framework designed to extract reusable skills in small language models. The paper argues that agent skills—standardized formats for packaging procedural knowledge and domain expertise—are essential for constraining a language model's behavior space to ensure repeatable, high-quality task execution. However, deploying these skills with strong closed-source models like Codex and OpenClaw is prohibitively expensive due to high inference costs. The rapid advancement of open-source models that can run on consumer-grade GPUs offers a compelling opportunity to reduce these costs by leveraging skill-based behavioral constraints. Yet, automatically generating effective skills for such compact models remains a significant challenge. SKILLER addresses this by using natural language to guide reinforcement learning, enabling small models to acquire skills without the need for expensive proprietary systems. The paper is available on arXiv under the identifier 2608.10538.