EffiSkill: Agent Skill Based Automated Code Efficiency Optimization
Published in Under review via ACL Rolling Review., 2026
EffiSkill studies automated code-efficiency optimization with language-model agents. It mines reusable Operator and Meta Skills from paired slow and optimized programs, retrieves relevant skills for new programs, and composes optimization plans without execution feedback.
On EffiBench-X, the system improves optimization success rate over the strongest baseline across multiple model and language settings. The current public version is available on arXiv.
Recommended citation: Zimu Wang, Yuling Shi, Mengfan Li, Zijun Liu, Jie M. Zhang, Chengcheng Wan, Xiaodong Gu. Under review via ACL Rolling Review.