Abstract Due to the high cost of repairing defective programs, many researches focus on automatic program repair (APR). In recent years, the new trend of APR is to apply neural networks to mine the relations between defective programs and corresponding patches automatically, which is known as neural program repair (NPR). The community, however, ignores some important properties that could impact the applicability of NPR systems, such as robustness. For semantic‐identical buggy programs, NPR systems may produce totally different patches. In this paper, we propose an evaluation tool named RobustNPR, the first NPR robustness evaluation tool. RobustNPR employs several mutators to generate semantic‐identical mutants of defective programs. For an original defective program and its mutant, it checks two aspects of NPR: (a) Can NPR fix mutants when it can fix the original defective program? and (b) can NPR generate semantic‐identical patches for the original program and the mutant? Then, we evaluate four SOTA NPR models and analyze the results. From the results, we find that even for the best‐performing model, 20.16% of the repair success is unreliable, which indicates that the robustness of NPR is not perfect. In addition, we find that the robustness of NPR is correlated with model settings and other factors.

Recently, the emerging trend in automatic program repair is to apply deep neural networks to generate fixed code from buggy ones, called NPR (Neural Program Repair). However, the existing NPR systems are trained and evaluated under very different settings (e.g., different training data, inconsistent evaluation data, wide-ranged candidate numbers), which makes it hard to draw fair-enough conclusions when comparing them. Motivated by this, we first build a standard benchmark dataset and an extensive framework tool to mitigate threats for the comparison. The dataset consists of a training set, a validation set and an evaluation set with 144,641, 13,739 and 13,706 bug-fix pairs of Java respectively. The tool supports selecting specific training, validation, and evaluation datasets and automatically conducting the pipeline of training and evaluating NPR models, as well as easily integrating new NPR models by implementing well-defined interfaces. Then, based on the benchmark and tool, we conduct a comprehensive empirical comparison of six SOTA NPR systems w.r.t the repairability, inclination and generalizability. The experimental results reveal deeper characteristics of compared NPR systems and subvert some existing comparative conclusions, which further verify the necessity of unifying the experimental setups in exploring the progresses of NPR systems. Meanwhile, we reveal some common features of NPR systems (e.g., they are good at dealing with code-delete bugs). Finally, we identify some promising research directions derived from our findings.