Hugging Face researchers have unveiled SKILLER, a reinforcement learning framework designed to automatically generate tailored skills for small open-source language models, addressing the high inference costs associated with closed-source models. The framework leverages natural language as the sole communication medium between a strong model (acting as actor and critic) and the small-model agent system, which serves as the environment.
Agent skills are standardized formats for packaging procedural knowledge and domain expertise, constraining a model's behavior for repeatable, high-quality task execution. However, popular agent harnesses like Codex and OpenClaw rely on expensive closed-source models, making skill deployment costly. SKILLER aims to change this by enabling small models, deployable on consumer-grade GPUs, to use skills effectively.
In experiments across five benchmarks using Qwen3.5-9B and Qwen3.5-4B, SKILLER outperformed three open-source and one closed-source skill generation or evolution methods. It achieved absolute gains of 4.3 to 20.4 percentage points for the 9B model and 1.8 to 13.3 points for the 4B model, while remarkably matching the performance of strong closed-source models on single-skill tasks in SkillsBench.
The project is open-sourced and available on GitHub, offering a practical solution for reducing inference costs without sacrificing task performance.