PatchZero: Zero-Shot Automatic Patch Correctness Assessment

Automated Program Repair (APR) techniques have shown more and more promising results in fixing real-world bugs. Despite the effectiveness, APR techniques still face an overfitting problem: a generated patch can be incorrect although it passes all tests. It is time-consuming to manually evaluate the correctness of generated patches that can pass all tests. To address this problem, many approaches have been proposed to automatically assess the correctness of patches generated by APR techniques. However, existing approaches require a large set of manually labeled patches as the training data. To mitigate the issue, in this study, we propose PatchZero, the patch correctness assessment by adopting large pre-trained models. Specifically, for patches generated by a new or unseen APR tool, PatchZero does not need labeled patches of this new or unseen APR tool for training (i.e., zero-shot) but directly queries the large pre-trained model to get predictions on the correctness labels without training. In this way, PatchZero can reduce the manual labeling effort when building a model to automatically assess the correctness of generated patches of new APR tools. To provide knowledge regarding the automatic patch correctness assessment (APCA) task to the large pre-trained models, we also design an instance-wise demonstration formation strategy by using contrastive learning. Specifically, PatchZero selects semantically similar patches to help the large pre-trained model to give more accurate predictions on the unlabeled patches. Our experimental results showed that PatchZero can achieve an accuracy of 82.7% and an F1-score of 86.0% on average although no labeled patch of the new or unseen APR tool is available. In addition, our proposed technique outperformed the prior state-of-the-art by a large margin.