Skill-to-LoRA: From Using Skills to Learning Behaviors for Token-Efficient LLM Agents

The article introduces Skill-to-LoRA (S2L), a novel approach that transforms traditional skill representations into behavior-centric skill-specific LoRA adapters, enabling more efficient use of procedural knowledge in LLM agents. Evaluated using the Qwen3.6-27B model on a subset of the SWE-Skills-Bench, S2L demonstrates a 2.9 to 5.2 percentage point improvement in pass rates and a 6.6% reduction in token costs compared to full skill text prompting. This method allows practitioners to convert procedural skills into dynamic, trainable modules, enhancing the efficiency and effectiveness of LLM applications.