We present SpeedGen, a novel framework that uses Large Language Models (LLMs) to automate code performance optimization. SpeedGen is designed to address software performance bottlenecks using a feedback-driven approach that profiles code to identify inefficiencies and iteratively refines the code to improve execution speed.

We conducted a comprehensive evaluation of SpeedGen’s capabilities across diverse codebases, benchmarking its performance against a leading large language model. Our results show that SpeedGen consistently reduces execution time and delivers significant performance improvements in various scenarios. The framework’s ability to adapt to different domains underscores its scalability and robustness, making it a valuable tool for optimizing code in a wide range of applications.

A key strength of SpeedGen is its ability to maintain the functional correctness of the code while achieving significant performance gains. This feature ensures that the optimized code remains reliable even when it undergoes significant transformations. By automating the optimization process, SpeedGen minimizes the need for manual intervention, streamlining the software development lifecycle and reducing time-consuming performance tuning efforts.

The introduction of SpeedGen marks a major step forward in the integration of LLMs into software engineering and paves the way for future research and development in this area. With its ability to improve performance without compromising code integrity it lays the foundation for more advanced automated code optimization techniques, simplifying software development.