Static analysis tools like ESLint are widely used for code quality assurance, yet optimizing their integration with LLM-based Automated Program Repair remains underexplored. While existing approaches typically use complete static analysis reports, the impact of individual information components and their interactions has received limited attention. This paper presents an empirical analysis of 12 ESLint information components’ impact on GPT-4 Turbo’s repair performance. Through experiments on the complete TFix test dataset (10,504 cases), we find: (1) Concise components (e.g., linter messages) show high performance per token, while verbose components show diminishing returns in combination. (2) Components exhibit interaction effects; some components with low individual performance show high importance in combinations. (3) Strategic component selection based on interaction analysis can achieve comparable performance to using all components while reducing token usage. Our findings suggest that prompt optimization for LLM-based APR can benefit from analyzing both individual component contributions and their interactions.

Deep learning-based automatic program repair (DL-APR) returns a patch code when given a defect code. Recent studies on DL-APR techniques have focused on the training phase to generate more accurate patches; however, a trained model cannot always generate an accurate patch for every new defect code, as the training dataset does not completely represent the new defects to be input in the future. DL-APR researchers should study a method to elicit the best performance on new inputs from the trained and deployed model. A new defect instance (i.e., defect codes and their context codes) is one of the crucial input data that determine the accuracy of the DL-APR, which can be changed and improved. We improve the quality of new input defect instances by focusing on the presence of noise tokens which compromise the defect instances’ quality, thus impairing the accuracy of generated patches. This paper shows that 1) there are noise tokens which prevent correct patch generation (inference) in a new defect instance, and 2) it is necessary to mask these noise tokens to avoid their usage in inferencing patch codes. In order to validate these two assertions, we use a state-of-the-art DL-APR technique and a genetic algorithm to generate near-optimal defect instances which maximize the patch generation accuracy (i.e., the BLEU score) of 4,573 defect instances. Based on optimization results, we found that 1) noise tokens impair patch generation accuracy in approximately 49% of instances, and 2) if these tokens are precluded from inference by masking them, we can improve patch generation accuracy by 88%. The results suggest that future work is required to automatically remove noise tokens from new defect instances so that the trained patch generator generates better patches.

Deep learning-based automated program repair (DL-APR) can automatically fix software bugs and has received significant attention in the industry because of its potential to significantly reduce software development and maintenance costs. The Samsung mobile experience (MX) team is currently switching from Java to Kotlin projects. This study reviews the application of DL-APR, which automatically fixes defects that arise during this switching process; however, the shortage of Kotlin defect-fixing datasets in Samsung MX team precludes us from fully utilizing the power of deep learning. Therefore, strategies are needed to effectively reuse the pretrained DL-APR model. This demand can be met using the Kotlin defect-fixing datasets constructed from industrial and open-source repositories, and transfer learning. This study aims to validate the performance of the pretrained DL-APR model in fixing defects in the Samsung Kotlin projects, then improve its performance by applying transfer learning. We show that transfer learning with open source and industrial Kotlin defect-fixing datasets can improve the defect-fixing performance of the existing DL-APR by 307%. Furthermore, we confirmed that the performance was improved by 532% compared with the baseline DL-APR model as a result of transferring the knowledge of an industrial (non-defect) bug-fixing dataset. We also discovered that the embedded vectors and overlapping code tokens of the code-change pairs are valuable features for selecting useful knowledge transfer instances by improving the performance of APR models by up to 696%. Our study demonstrates the possibility of applying transfer learning to practitioners who review the application of DL-APR to industrial software.

Automated Feedback Generation (AFG) was proposed to automatically generate personalized feedback on students' programming assignments. Existing AFG techniques have been developed mainly for novice programmers, so feedback on complex programs cannot be generated. Therefore, we propose MUNCK, which automatically generates feedback for multiple function programs, one of the complex programs. Our experiment shows that MUNCK can generate feedback for 90% of multiple function programs.