一句话:提出融合静态分析结构化知识与学习型模型的混合方法,在多种LLM流水线阶段集成,提升自动代码审查的准确率与覆盖率。
代码审查在软件开发中至关重要,但复杂、主观且耗时。早期自动化依赖知识型系统(KBS),能提供精确的规则驱动反馈,却难以处理复杂且依赖上下文的问题;近期基于学习的系统(LBS),尤其是微调预训练语言模型,覆盖面更广但常常牺牲精度。本提案提出一种混合方法,将结构化的外部知识(静态分析结果)与LBS相结合,以生成准确且全面的代码审查意见。作者在大型语言模型(LLM)流水线的不同阶段探索了多种集成策略,并在真实数据集上与独立的KBS和LBS基线进行对比评估。结果表明,该方法在准确率和覆盖率上均取得了一致提升。
Code review is a crucial but often complex, subjective, and time-consuming task in software development. Early automation efforts relied on knowledge-based systems (KBS), which offer precise rule-driven feedback but are limited in handling complex, context-dependent issues. More recent approaches have focused on learning-based systems (LBS), particularly fine-tuned pre-trained language models, enabling broader issue coverage but frequently sacrificing precision. In this proposal, we introduce a hybrid approach that integrates structured external knowledge with LBS to generate accurate and comprehensive code reviews. We explore multiple integration strategies at different stages of the large language model (LLM) pipeline and evaluate them against standalone KBS and LBS baselines on a real-world dataset. Our results show consistent improvements in both accuracy and coverage.