EffiSkill: An Agent Skill-Based Framework for Automatic Code Efficiency Optimization

EffiSkill models slow-to-fast code transformations as reusable agent skills. Through execution-agnostic diagnosis and skill retrieval, it achieves a 3.69 to 12.52 percentage point improvement over the strongest baseline on EffiBench-X, demonstrating the value of mechanism-level skill reuse in the field of code optimization.

Code efficiency is a core dimension of software quality, but program optimization using Large Language Models (LLMs) remains challenging. Existing methods rely on one-time rewriting, example retrieval, or prompt-based search, lacking explicit reusable optimization knowledge, which limits generalization ability. Human programmers rely on transferable optimization skills (e.g., loop unrolling, replacing linear search with hash tables), and EffiSkill aims to equip LLM systems with this skill reuse capability.

The key insight of EffiSkill is to model recurring slow-to-fast transformations as agent skills, decoupling optimization knowledge from specific code. It adopts a two-stage architecture: 1) Mine operational skills (specific transformation patterns) and meta-skills (high-level strategies) from large-scale slow/fast program pairs to build a skill library; 2) Apply the skill library to unseen programs, completing optimization through execution-agnostic diagnosis, retrieval, plan combination, and candidate generation without runtime feedback.

Skill Mining and Representation: The skill library includes operational skills (e.g., converting list comprehensions to generators, set implementation for O(1) lookups) and meta-skills (e.g., optimizing algorithm complexity first then constant factors). Skills are semantically represented via contrastive learning to support effective retrieval. Optimization Process: Diagnosis phase (static analysis to identify performance bottlenecks) → Skill retrieval and adaptation (matching and adapting skills to current context) → Plan combination (meta-skills guide skill order to generate optimization candidates).

In the EffiBench-X benchmark test, EffiSkill achieves a 3.69 to 12.52 percentage point improvement over the strongest baseline, with consistent improvements across different models and language settings. The execution-agnostic design increases optimization speed by several orders of magnitude compared to methods relying on runtime feedback. The skill library is highly scalable: the optimization success rate continues to improve after adding more program pairs.

1. Skill reuse paradigm: Explicit skill representation improves interpretability and controllability, and can be extended to tasks like code completion and bug fixing; 2) Possibility of execution-agnostic optimization: Static analysis combined with knowledge-driven methods can achieve effective optimization; 3) Human-machine collaboration mode: The system provides candidates for human review, suitable for production environments.

Application Scenarios: CI/CD performance regression protection (detecting and fixing performance degradation), legacy code modernization (replacing outdated implementations), programming education (demonstrating optimization techniques). Future Research: Automatic expansion of the skill library, cross-language skill transfer, integration with runtime analysis, improving skill representation accuracy.