FasterPy: An LLM-based Code Execution Efficiency Optimization Framework

A research team has introduced FasterPy, a framework that applies large language models to automatically improve the execution speed of Python programs, according to a paper posted on arXiv. The system combines retrieval-augmented generation with parameter-efficient fine-tuning to suggest performance-enhancing code edits [1]. The paper, authored by Peng Liang and colleagues, was first submitted in December 2025 and revised in June 2026 [1]. The authors argue that conventional rule-based optimization tools require manual upkeep for each class of performance bug, such as redundant loops or repeated computations, making them labor-intensive and narrow in scope [1]. Machine-learning approaches have offered alternatives, but they typically depend on specialized program representations and carefully curated training datasets, which raises development costs and limits scalability [1]. FasterPy addresses these constraints by pairing a retrieval-augmented generation module with Low-Rank Adaptation, a technique that fine-tunes a large language model by updating only a small set of parameters [1]. The retrieval component draws on a knowledge base built from existing pairs of slow and fast Python code, along with their corresponding performance measurements, to ground the model’s suggestions in prior optimization examples [1]. The researchers evaluated FasterPy on the Performance Improving Code Edits benchmark, a standardized test used to compare models on code-optimization tasks [1]. Language model benchmarks of this kind typically supply a dataset and a set of evaluation metrics, which can include not only accuracy but also throughput and energy efficiency [5]. The paper states that FasterPy outperformed existing models on multiple metrics, though the authors did not disclose specific numerical results in the abstract [1]. The framework’s design emphasizes low cost and efficiency, a combination the authors say makes it practical for broader adoption [1]. The code and experimental data have been released on GitHub, allowing other researchers to reproduce the findings or adapt the system for related tasks [1]. The work lands as interest grows in using large language models for software engineering beyond code generation, including maintenance and performance tuning, though independent validation of the reported gains remains pending.