Large Language Models (LLMs) have gained significant traction in software engineering for automating tasks such as unit test generation. Most existing studies prioritize code coverage as the primary metric for enhancing test suite effectiveness. However, prior research has shown that although code coverage can reach approximately 80%, the mutation score, which generally exhibits a stronger correlation with defect detection effectiveness, attains only about 35%. This gap highlights the need to enhance test suite effectiveness guided by mutation score rather than code coverage. Recent studies, including MuTAP and Mu tGe n, explored the use of survived mutants to enhance test suite effectiveness. However, their evaluations were limited to simple standalone methods that rely on built-in functions and standard libraries. Non-standalone methods, which depend on other classes and involve complex user-defined types, are more intricate and commonly found in real-world projects. The limited contextual information and basic repair mechanisms in their prompt designs make it unclear whether their performance can generalize to non-standalone methods. Moreover, the two studies rely on existing language-specific, rule-based mutation techniques, which require specific configurations and incur additional costs when adapting to other programming languages. To bridge this gap, we propose a novel, fully automatic LLM-based approach to enhance test suite effec-tiveness, guided by survived mutants. The approach augments initial test suites by integrating mutation testing with test case generation. It takes focal method information as input and generates test cases targeting survived mutants identified from applying the initial test suites. Our approach incorporates multiple prompt techniques, rich contextual information, and an advanced repair mechanism to effectively generate test cases for non-standalone methods. The evaluation covers 1,035 focal methods, categorized as standalone or non-standalone. On average, the mutation score increases by 16.04% for standalone methods and 8.11% for non-standalone methods. We validate the practical impact of augmented test suites in LLM-based code generation. After test suite augmentation, pass@1 decreased by 0.3152 and 0.1772 on average for standalone and non-standalone methods, respectively, indicating the effectiveness of our approach in reducing false positives caused by insufficient test cases in code generation evaluation.

Background . Code coverage and mutation score serve as pivotal test effectiveness metrics used to assess a test suite’s ability to uncover actual defects. However, prior research has produced inconsistent or even conflicting findings regarding their correlation with defect detection capability, particularly concerning the impact of test suite size. Problem. The extent of the potentially confounding effect of test suite size in test effectiveness evaluation context is not clear, nor is the method to remove the potentially confounding effect, or the influence of this removal on the performance of test suite optimization. Objective . Our goal is to deeply understand how test suite size affects the true relationship between test effectiveness metrics and a test suite’s ability to detect actual defects. Method. We first employ statistical methods to examine the extent of the potentially confounding effect of test suite size in the context of test effectiveness evaluation. After that, we propose a linear regression-based method to remove the potentially confounding effect of test suite size. Finally, we empirically explore the impact of this removal method on test suite optimization. Result. Our experimental results, based on the Defects4J defect dataset, uncovers that: (1) the confounding effect of test suite size on the associations between test effectiveness metrics and defect detection capability in general exists; (2) the proposed linear regression-based method can effectively remove the confounding effect; and (3) after removing the confounding effect, mutation score demonstrates superior effectiveness in predicting test suite effectiveness, while statement coverage is the least effective metric. Furthermore, both coverage-based and mutation-based test suite reduction exhibit enhanced cost-effectiveness in defect detection, and there is a marginal improvement in the speed of defect detection for coverage-based test case prioritization. Conclusion . When using test effectiveness metrics to assess test suite effectiveness, it is crucial to eliminate the influence of test suite size.

The rapid evolution of machine learning model training has outpaced the development of corresponding measurement and testing tools, leading to two significant challenges. Firstly, developers of deep learning frameworks struggle to identify operators that fail to meet precision criteria, as these issues may only manifest in a few data points. Secondly, model trainers lack effective methods to estimate precision loss caused by operators during training. To address these issues, we introduce a Pythonic framework inspired by common network layer definitions in deep learning. Our framework includes two new layers: the Fuzz Layer and the Check Layer, designed to aid in measurement and testing. The Fuzz Layer introduces minor perturbations to tensor inputs for any deterministic layer under testing (LUT). The Check Layer then measures precision by analyzing the differences before and after the perturbation. This approach estimates a lower bound of the maximal relative error and alerts developers or trainers of potential bugs if the difference exceeds a pre-defined tolerance. Additionally, Check Layers can be used independently to conduct precision tests for specific layers, ensuring the precision of operators during runtime. Despite the additional memory and time requirements, this runtime testing ensures proper training of the original model. We demonstrate the utility of our framework, FuMi, through two experiments. First, we tested 21 torch operators across nine popular machine learning models using PyTorch for various tasks, finding that the conv2d and linear operators often fail to meet precision requirements. Second, to showcase the generalizability of our framework, we tested the ATTENTION operator. By comparing different implementations of state-of-the-art ATTENTION operators, we found that the maximum relative error of the ATTENTION operator is not less than 1%, which is 13 times larger than that measured by Predoo (a unit test tool). This framework provides a robust solution for identifying precision issues in deep learning models, ensuring reliable and accurate model training.

During software maintenance and evolution, developers spend more than half of their time on code comprehension activities. In order to understand an unfamiliar code base, they would naturally ask different types of questions related to code snippets and try to find the answers. In this paper, we conduct an initial work to explore the possibility of automatic question generation for program comprehension. We construct a large-scale data set containing pairs of source code and questions that are automatically transformed from inline comments based on dependency analysis and semantic role labeling. We also build a comprehensive taxonomy of question types so as to generate questions concerning different aspects of code snippets, such as purpose, implementation details and so on. Then, we propose a deep learning-based prototype CodeQG to automatically generates multiple types of questions for code snippets. We evaluate CodeQG by using both typical performance metrics and manual evaluation. The results show that (1) we can achieve a value of 42.02 on BLEU4 and 60.81 on ROUGE-L for the generated questions; (2) overall, the questions are very correct in grammatical, semantic and format; (3) the questions are related to the corresponding code snippet and are helpful for developers in source code comprehension activities. Our work gives insights into automatically generating multiple types of questions for code comprehension. We expect this exploration will improve the applicability and generality of machine code comprehension.

Mutation testing aims to simulate real-world defects, but existing tools often struggle to replicate method invocation defects accurately. To address this, we propose MIN (Method INvocation mutator), which uses a mapping strategy to pair method names with corresponding values, ensuring that methods share argument and return types. This method enhances the feasibility and realism of mutants by considering factors such as library methods, access control, inheritance, and static methods. Experimental results show that integrating MIN into Major (a popular mutation tool) improves semantic similarity to real defects by 11%, increases mutant set diversity to 97.5%, and reduces undetected faults by 38.5%. Furthermore, MIN’s performance rivals that of state-of-the-art machine learning-based mutators like CodeBERT, with a 10x speed advantage over CodeBERT and 4x over DeepMutation in generating compilable mutants. These findings demonstrate that MIN can significantly enhance defect simulation and improve the efficiency of mutation testing.

Background. Software testing is a critical activity for ensuring the quality and reliability of software systems. To evaluate the effectiveness of different test suites, researchers have developed a variety of metrics. Problem. However, comparing these metrics is challenging due to the lack of a standardized evaluation framework including comprehensive factors. As a result, researchers often focus on single factors (e.g., size), which finally leads to different or even contradictory conclusions. After comparing dozens of pieces of work in detail, we have found two main problems most troubling to our community: (1) researchers tend to oversimplify the description of the ground truth they use, and (2) data involving real defects is not suitable for analysis using traditional statistical indicators. Objective. We aim at scrutinizing the whole process of comparing test suites for our community. Method. To hit this aim, we propose a framework ASSENT (ev A luating te S t S uite E ffective N ess me T rics) to guide the follow-up research for evaluating a test suite effectiveness metric. ASSENT consists of three fundamental components: ground truth, benchmark test suites, and agreement indicator. Its functioning is as follows: first, users clarify the ground truth for determining the real order in effectiveness among test suites. Second, users generate a set of benchmark test suites and derive their ground truth order in effectiveness. Third, users use the metric to derive the order in effectiveness for the same test suites. Finally, users calculate the agreement indicator between the two orders derived by two metrics. Result. With ASSENT, we are able to compare the accuracy of different test suite effectiveness metrics. We apply ASSENT to evaluate representative test suite effectiveness metrics, including mutation score and code coverage metrics. Our results show that, based on the real faults, mutation score, and subsuming mutation score are the best metrics to quantify test suite effectiveness. Meanwhile, by using mutants instead of real faults, test effectiveness will be overestimated by more than 20% in values. Conclusion. We recommend that the standardized evaluation framework ASSENT should be used for evaluating and comparing test effectiveness metrics in the future work.

In [1, p. 4967], a figure citation is incorrect and “Fig. 3(c)” should be “Fig. 1(c)” in the left column, the fourth line from the bottom. It is corrected below.

Code coverage has been widely used as the basis for various software quality assurance techniques. Therefore, it is of great importance to ensure that coverage profilers provide reliable code coverage. However, it is challenging to validate the correctness of the code coverage generated due to the lack of an effective oracle. In this paper, we propose an effective approach based on control flow constraint solving to test coverage profilers and have implemented a coverage bug hunting tool, DOG (finD cOverage buGs). Our core idea is to leverage inherent control flow features to generate control flow constraints that the resulting coverage statistics should respect. If DOG identifies any unsatisfiable constraints, it signifies the presence of incorrect coverage statistics. In such cases, DOG provides detailed diagnostic information about the suspicious coverage statistics for manual inspection. Compared with the state-of-the-art works, DOG has the following prominent advantages: (1) wide applicability: DOG eliminates the need for multiple coverage profilers (as required by differential testing) and program variants (as needed in metamorphic testing), making it highly versatile; (2) unique testing capability: DOG effectively analyzes and utilizes relationships among available coverage statistics, boosting its testing capabilities; and (3) enhanced interpretability: DOG provides clear control flow explanations for incorrect code coverage, enabling the localization of suspicious coverage areas. During our testing period with DOG, we successfully identified and reported 27 bugs in Gcov and llvm-cov, both widely-used coverage profilers. Of these, 17 bugs have been confirmed (11 have been fixed), 3 were deemed expected behaviors by developers, and 7 remain unresolved. Remarkably, 21 out of 24 unexpected bugs had been latent for over two and a half years, and nearly half of the coverage bugs (10 out of 24) were undetectable by state-of-the-art coverage profiler validators. These results demonstrate the effectiveness and importance of using DOG to improve the reliability of code coverage profilers.

Background. Mutation testing is a commonly used defect injection technique for evaluating the effectiveness of a test suite. However, it is usually computationally expensive. Therefore, many mutation reduction strategies, which aim to reduce the number of mutants, have been proposed. Problem. It is important to measure the ability of a mutation reduction strategy to maintain test suite effectiveness evaluation. However, existing evaluation indicators are unable to measure the “order-preserving ability”, i.e., to what extent the mutation score order among test suites is maintained before and after mutation reduction. As a result, misleading conclusions can be achieved when using existing indicators to evaluate the reduction effectiveness. Objective. We aim to propose evaluation indicators to measure the “order-preserving ability” of a mutation reduction strategy, which is important but missing in our community. Method. Given a test suite on a Software Under Test (SUT) with a set of original mutants, we leverage the test suite to generate a group of test suites that have a partial order relationship in defect detecting ability. When evaluating a reduction strategy, we first construct two partial order relationships among the generated test suites in terms of mutation score, one with the original mutants and another with the reduced mutants. Then, we measure the extent to which the partial order under the original mutants remains unchanged in the partial order under the reduced mutants. The more partial order is unchanged, the stronger the Order Preservation ( OP ) of the mutation reduction strategy is, and the more effective the reduction strategy is. Furthermore, we propose Effort-aware Relative Order Preservation ( EROP ) to measure how much gain a mutation reduction strategy can provide compared with a random reduction strategy. Result. The experimental results show that OP and EROP are able to efficiently measure the “order-preserving ability” of a mutation reduction strategy. As a result, they have a better ability to distinguish various mutation reduction strategies compared with the existing evaluation indicators. In addition, we find that Subsuming Mutant Selection (SMS) and Clustering Mutant Selection (CMS) are more effective than the other strategies under OP and EROP. Conclusion. We suggest, for the researchers, that OP and EROP should be used to measure the effectiveness of a mutant reduction strategy, and for the practitioners, that SMS and CMS should be given priority in practice.

Knowing the effectiveness of a test suite is essential for many activities such as assessing the test adequacy of code and guiding the generation of new test cases. Mutation testing is a commonly used defect injection technique for evaluating the effectiveness of a test suite. However, it is usually computationally expensive, as a large number of mutants (buggy versions) are needed to be generated from a production code under test and executed against the test suite. In order to reduce the expensive testing cost, recent studies proposed to use supervised models to predict the effectiveness of a test suite without executing the test suite against the mutants. Nonetheless, the training of such a supervised model requires labeled data, which still depends on the costly mutant execution. Furthermore, existing models are based on traditional supervised learning techniques, which assume that the training and testing data come from the same distribution. But, in practice, software systems are subject to considerable concept drifts, i.e., the same distribution assumption usually does not hold. This can lead to inaccurate predictions of a learned supervised model on the target code as time progresses. To tackle these problems, in this paper, we propose a Coverage-Based Unsupervised Approach (CBUA) for evaluating the effectiveness of a test suite. Given a production code under test, the corresponding mutants, and a test suite, CBUA first collects the coverage information of the mutated statements in the target production code under the execution of the test suite. Then, CBUA employs coverage to estimate the probability of each mutant being alive. As such, a mutation score is computed to evaluate the test suite effectiveness and the predicted labels (i.e., killed or alive) are obtained. The whole process only requires a one-time execution of the test suite against the target production code, without involving any mutant execution and any training data. CBUA can ensure the score monotonicity property (i.e., adding test cases to a test suite does not decrease its mutation score), which may be violated by a supervised approach. The experimental results show that CBUA is very competitive with the state-of-the-art supervised approaches in prediction accuracy. In particular, CBUA is shown to be more effective in finding mutants that are covered but not killed by a test suite, which is helpful in identifying the weaknesses in the current test suite and generating new test cases accordingly. Since CBUA is an easy-to-implement approach with a low cost, we suggest that it should be used as a baseline approach for comparison when any novel prediction approach is proposed in future studies.