核心概念
T5APR introduces a novel multilingual neural repair approach, leveraging transformer models, to efficiently fix bugs across various programming languages.
摘要
T5APR is a novel neural program repair approach that leverages CodeT5 and checkpoint ensemble strategy to provide bug fixes across multiple programming languages. It outperforms existing methods in fixing bugs and demonstrates competitiveness in various benchmarks. The approach fine-tunes the model on a multilingual dataset and uses multiple checkpoints for patch generation and validation.
统计
T5APR correctly fixes 1,985 bugs, including 1,442 bugs identical to developer patches.
T5APR achieves state-of-the-art performance in terms of both repair effectiveness and efficiency.
T5APR generates correct patches for various types of bugs in different languages.
T5APR ranks candidate patches using project test suites to select the most suitable one.