Skill-to-LoRA: From Using Skills to Learning Behaviors for Token-Efficient LLM Agents
Skill-to-LoRA transforms skill text into LoRA adapters, improving pass rate by 5.2 percentage points.
Key Findings
Methodology
Skill-to-LoRA (S2L) converts skill text into LoRA adapters, reducing runtime text injection. Offline, it uses complete SKILL.md to generate skill-guided demonstrations; online, it dynamically loads LoRA adapters to activate learned skill behavior.
Key Results
- S2L improves pass rate by 2.9 and 5.2 percentage points over no-skill and Full Skill Text baselines, respectively, while reducing per-step token cost by 6.6%.
- S2L matches or improves Full Skill Text on 18/21 skills.
- Control experiments show skill-specific adapter alignment is crucial for performance gains.
Significance
S2L reduces runtime context redundancy and token cost by converting skill text into trainable behavioral modules. This method not only improves model efficiency but also mitigates interference from long skill documents during inference.
Technical Contribution
S2L introduces a new skill representation method by converting skill text into trainable parameters representing behavioral bias. Unlike existing methods, S2L automatically generates skill-guided demonstrations without large-scale human annotation.
Novelty
S2L is the first to transform skill text into LoRA adapters, dynamically loading them to activate skill behavior. It models skill-induced behavioral change as a learnable functional bias.
Limitations
- S2L heavily relies on skill-specific adapter alignment; wrong or shared LoRA adapters reduce performance.
- The method may not suit complex tasks requiring multi-skill composition.
Future Work
Future work can explore adapter designs for multi-skill composition and validate S2L's effectiveness on larger skill libraries.
AI Executive Summary
In modern agent systems, skills are often distributed as human-readable SKILL.md files. However, this design requires the same reusable procedure to be repeatedly injected into the runtime context, leading to inefficiencies. Skill-to-LoRA (S2L) addresses this by transforming skill text into behavioral modules.
S2L uses the complete skill document offline to generate skill-guided demonstrations and dynamically loads the corresponding LoRA adapter online to activate the learned skill behavior. Experimental results show that S2L improves pass rates on a subset of SWE-Skills-Bench while reducing token costs.
The significance of this method lies in its ability to improve model efficiency and mitigate interference from long skill documents during inference. Future research can explore adapter designs for multi-skill composition to further extend S2L's applicability.
Deep Analysis
Background
As agent systems evolve, the concept of skill libraries has emerged. Each skill combines natural language guidance, scripts, and resources. However, traditional methods require repeated runtime skill text injection, leading to inefficiencies.
Core Problem
Existing methods repeatedly inject skill text at runtime, increasing token costs and potentially introducing mismatched templates or expanding the action space.
Innovation
S2L reduces runtime text injection by converting skill text into LoRA adapters. This method models skill-guided behavioral change as a learnable functional bias.
Methodology
- �� Offline: Use complete skill document to generate skill-guided demonstrations.
- �� Online: Dynamically load corresponding LoRA adapter to activate learned skill behavior.
- �� Use LoRA training to convert behavioral patterns into lightweight parameters.
Experiments
Evaluated on a 21-skill subset of SWE-Skills-Bench using the Qwen3.6-27B model, comparing no-skill and Full Skill Text baselines.
Results
S2L improves pass rates while reducing token costs, highlighting the importance of skill-specific adapters for performance gains.
Applications
S2L can be applied in agent systems requiring efficient skill invocation, particularly in software engineering tasks.
Limitations & Outlook
S2L heavily relies on skill-specific adapter alignment and may not suit complex tasks requiring multi-skill composition.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. Traditional methods are like having to read the recipe every time, while S2L is like having memorized the recipe, only recalling key steps when needed. This saves time and reduces errors.
ELI14 Explained like you're 14
Imagine playing a game where you have to check the guide every time. Annoying, right? S2L is like having mastered the guide's essence, using it only when crucial. This makes you play faster and smoother!
Glossary
LoRA (Low-Rank Adaptation)
A technique for adapting skills by training small low-rank updates.
Used to convert skill text into behavioral modules.
SKILL.md
Human-readable procedural documents describing workflows and tools.
Foundation of traditional skill representation.
SWE-Skills-Bench
A benchmark for evaluating the impact of skill injection on task execution.
Used to validate S2L's effectiveness.
Qwen3.6-27B
The frozen base model used for S2L evaluation.
Serves as the base model in experiments.
Self-distillation
A method of training using data generated by the model itself.
Used to generate skill-guided behavioral demonstrations.
Open Questions Unanswered questions from this research
- 1 How to effectively apply S2L in complex tasks requiring multi-skill composition?
- 2 Is S2L still effective on larger skill libraries?
Applications
Immediate Applications
Software Engineering Agents
S2L can improve skill invocation efficiency in software engineering tasks, reducing runtime token costs.
Long-term Vision
General Intelligent Agents
S2L's success may drive the development of general intelligent agents, enabling more efficient skill learning and application.
Abstract
Agent skills are commonly distributed as SKILL.md files: human-readable procedural documents that describe workflows, tools, resources, and domain conventions. While convenient for inspection and reuse, this design requires the same reusable procedure to be repeatedly injected into the runtime context. We propose Skill-to-LoRA(S2L), a behavior-centric skill representation that replaces runtime skill text with skill-specific LoRA adapters. Rather than compressing the skill document itself, S2L models the behavioral change induced by the skill text: offline, the complete SKILL.md is used to synthesize skill-guided demonstrations; online, the full document is omitted and the corresponding LoRA adapter is dynamically loaded to activate the learned skill behavior. We evaluate S2L with Qwen3.6-27B on a 21-skill subset of SWE-Skills-Bench. Compared with the no-skill and Full Skill Text baselines, S2L improves pass rate by 2.9 and 5.2 percentage points, respectively, while reducing per-step token cost by 6.6% relative to Full Skill Text prompting. S2L matches or improves Full Skill Text on 18/21 skills and the no-skill baseline on 15/21 skills. Control experiments further show that the gains depend on skill-specific adapter alignment: Wrong-LoRA and Shared-LoRA both reduce performance. These results suggest that many procedural agent skills can be converted from runtime instructions into trainable, dynamically loadable behavioral modules. Code will be released upon acceptance.