Improving Automated Code Reviews: Learning from Experience
Hong Yi Lin, Patanamon Thongtanunam, Christoph Treude, Wachiraphan Charoenwet
TL;DR
The paper tackles the challenge of high workload in code reviews by enhancing automated code review models through experience-aware oversampling, biasing training toward reviews from experienced developers. Using CodeReviewer and a targeted oversampling strategy, the study shows improvements in semantic accuracy, actionable feedback, and explanatory content without introducing new data. Key findings demonstrate that leveraging underutilized high-quality reviews can elevate review quality, especially for logic, validation, and resource-related issues, offering a resource-efficient path to better automated reviews. The work contributes a principled method for exploiting reviewer experience signals and provides a data-driven foundation for future refinements in automated code review systems.
Abstract
Modern code review is a critical quality assurance process that is widely adopted in both industry and open source software environments. This process can help newcomers learn from the feedback of experienced reviewers; however, it often brings a large workload and stress to reviewers. To alleviate this burden, the field of automated code reviews aims to automate the process, teaching large language models to provide reviews on submitted code, just as a human would. A recent approach pre-trained and fine-tuned the code intelligent language model on a large-scale code review corpus. However, such techniques did not fully utilise quality reviews amongst the training data. Indeed, reviewers with a higher level of experience or familiarity with the code will likely provide deeper insights than the others. In this study, we set out to investigate whether higher-quality reviews can be generated from automated code review models that are trained based on an experience-aware oversampling technique. Through our quantitative and qualitative evaluation, we find that experience-aware oversampling can increase the correctness, level of information, and meaningfulness of reviews generated by the current state-of-the-art model without introducing new data. The results suggest that a vast amount of high-quality reviews are underutilised with current training strategies. This work sheds light on resource-efficient ways to boost automated code review models.