Deep Neural Networks (DNNs) are increasingly deployed in safety-critical systems, yet their numerical reliability remains a pressing concern due to defects like NaN/INF occurrences during training/inference. Existing static analyzers detect operator-level risks but lack the ability to establish contextual mappings between operator-level anomalies and their originating code implementations, thereby limiting their practical applicability in real-world debugging scenarios. To address this limitation, this paper introduces a novel hybrid method that synergizes risk sensitive paths derived from static analysis with Large Language Model(LLM)-guided semantic reasoning. By anchoring LLMs to identified risk-sensitive computation paths, the proposed method precisely correlates operator-level numerical defects with code-level vulnerabilities, while mitigating LLM hallucinations. Evaluated on a widely-used real-world benchmark, the proposed method achieves $\mathbf{1 0 0} \boldsymbol{\%}$ defect detection accuracy, matching state-of-the-art static analyzers, but significantly outperforms them in code-level root-cause localization. This work pioneers a new paradigm for DNN reliability assurance by combining the static numerical analysis with the semantic reasoning power of LLMs, establishing a practical numerical defect analysis pipeline.

As one of the state-of-the-art parameter-efficient fine-tuning~(PEFT) methods, Low-Rank Adaptation (LoRA) enables model optimization with reduced computational cost through trainable low-rank matrix. However, the low-rank nature makes it prone to produce a decrease in the representation ability, leading to suboptimal performance. In order to break this limitation, we propose RidgeLoRA, a lightweight architecture like LoRA that incorporates novel architecture and matrix ridge enhanced full-rank approximation, to match the performance of full-rank training, while eliminating the need for high memory and a large number of parameters to restore the rank of matrices. We provide a rigorous mathematical derivation to prove that RidgeLoRA has a better upper bound on the representations than vanilla LoRA. Furthermore, extensive experiments across multiple domains demonstrate that RidgeLoRA achieves better performance than other LoRA variants, and can even match or surpass full-rank training.

Deep learning frameworks play a key rule to bridge the gap between deep learning theory and practice. With the growing of safety- and security-critical applications built upon deep learning frameworks, their reliability is becoming increasingly important. To ensure the reliability of these frameworks, several efforts have been taken to study the causes and symptoms of bugs in deep learning frameworks, however, relatively little progress has been made in investigating the fault triggering conditions of those bugs. This paper presents the first comprehensive empirical study on fault triggering conditions in three widely-used deep learning frameworks (i.e., TensorFlow, MXNET and PaddlePaddle). We have collected 3,555 bug reports from GitHub repositories of these frameworks. A bug classification is performed based on fault triggering conditions, followed by the analysis of frequency distribution of different bug types and the evolution features. The correlations between bug types and fixing time are investigated. Moreover, we have also studied the root causes of Bohrbugs and Mandelbugs and investigated the important consequences of each bug type. Finally, the analysis of regression bugs in deep learning frameworks is conducted. We have revealed 12 important findings based on our empirical results and have provided 10 implications for developers and users.

Abstract The goal of software development is to deliver software products with high quality and free from defects, but resource and time constraints often cause the developers to submit incomplete or temporary patches of codes and further bear the additional burden. Therefore, the investigations on identifying self‐admitted technical debt (SATD) to improve code quality have been conducted in recent years. However, missing syntactic structure information and the imbalance distribution bias shorten the SATD identification performance. Addressing to this issue, we present a graph neural network based SATD identification model (GNNSI) to improve the performance. Specifically, we obtain the structure information of the missing SATD in a compositional way to obtain different feature maps for different comments, and use focal loss to handle the imbalance between SATD and non‐SATD classes in the comments. Then extensive experiments on 10 open source projects are conducted, and the results show that GNNSI outperforms the baselines and can help developers to better predict SATDs.

Recognizing and repairing defects to enhance quality in software life circle has become a critical research topic. Unfortunately, it is difficult to guarantee the validity of the defect prediction method based on manually designed features proposed in previous studies. Numerous scholars have endeavored to use a single model to obtain prediction results for different types of fault, but this is difficult to perform. This article improves the defect representation and prediction model in software defect prediction, proposing Augmented-Code Property Graph (CPG) based defect prediction method (ACGDP). Augmented-CPG is a novel encoding graph format introduced in this article. Based on Augmented-CPG, we suggested defect region candidate extraction approach linked to the defect category. Graph neural networks are used for obtaining defect characteristics. Experiments on three distinct types of defects indicate that ACGDP can predict certain classed of defects effectively.

To optimize the process of software testing and to improve software quality and reliability, many attempts have been made to develop more effective methods for predicting software defects. Previous work on defect prediction has used machine learning and artificial software metrics. Unfortunately, artificial metrics are unable to represent the features of syntactic, semantic, and context information of defective modules. In this article, therefore, we propose a practical approach for identifying software defect patterns via the combination of semantics and context information using abstract syntax tree representation learning. Graph neural networks are also leveraged to capture the latent defect information of defective subtrees, which are pruned based on a fix-inducing change. To validate the proposed approach for predicting defects, we define mining rules based on the GitHub workflow and collect 6052 defects from 307 projects. The experiments indicate that the proposed approach performs better than the state-of-the-art approach and five traditional machine learning baselines. An ablation study shows that the information about code concepts leads to a significant increase in accuracy.

Machine learning is an effective method for software defect prediction. The performance of learning models can be affected by irrelative and redundant features. Feature selection techniques select a subset of most impactful relevant features that will result in higher accuracy and efficiency of models. This paper proposed a Cluster-based Hybrid Feature Selection method (CHIFS) for software defect prediction. A spectral cluster-based Feature Quality coefficient (FQ) was defined as a comprehensive measurement of feature relevance and redundancy. The final feature subset was iteratively selected from feature sequence ranked by FQ. The proposed CHIFS method was validated in the experiments using 3 classifiers with 15 open datasets from Promise Repository. Experimental results showed that the CHIFS method performed better than traditional methods in terms of accuracy and efficiency on a wide range of datasets.

Since the development of computer science, our lives have become increasingly dependent on software. While we enjoy the benefits and convenience that software programs provide, we cannot ignore issues with software reliability, complexity, and security. Since the introduction of complex networks, people have been using software network to analyze software problems; however, traditional software network models are not currently capable of analyzing software with large scale and complex structures. In this paper, a new software network model is proposed, with which each node in the network can be assigned a set of coordinates that reflect its function-call information and make the disorder of the network graph more orderly. Additionally, characteristics and derivatives of the model are thoroughly examined and analyzed. A case study using the coordinate model combined with bug information is then conducted to analyze five different software programs. The results show that the proposed model can be used to analyze the relationship between nodes or defects distribution and software network parameters, as well as high-risk module excavation through a defect density analysis. Compared to traditional software network models, the model maintains the inner logic relationship of the software systems better, which makes it easier to analyze many aspects of software.

Software testing has become an essential activity to guarantee software quality. To reduce the overall cost of software testing, model-based testing has been widely studied in the past two decades and Finite State Machine (FSM) is used to build the model of software behaviors. However, due to the inadequacy of the modeling ability of FSM, FSM-based testing cannot be taken as a test oracle to solve all issues in software testing. To improve the modeling capability of the model, a few researchers have proposed using Extended Regular Expressions (ERE) to model software behaviors. This paper reviews the method of the ERE-based testing and presents six modeling rules to convert program codes to the ERE model. Then, a case is adopted to illustrate the process of generating executable paths from the ERE model and the method of designing test cases by those paths. Compared with the traditional graphic traversal method of constructing executable paths from the program, ERE not only has robust modeling capability to describe more types of software behaviors than FSM, but also can be used to construct effective executable paths to detect program errors.

Accidents caused by defective software systems have long been a nightmare. Though engineers utilize advanced techniques and rigorous quality control procedures, we still have to admit that the increasing complexity and expanding scale of software systems make it extremely difficult to guarantee high quality deliverables. Since large-scale software systems exhibit the characteristics of complex networks, applying the principles of complex networks to evaluate the quality of software systems has attracted attention from both academia and industry. Unfortunately, most current research studies focus only on one or a limited number of attributes of software structures which makes them ineffective in providing comprehensive and insightful quality evaluation for software structures. To overcome this problem, we propose an approach based on various software structural characteristics to evaluate software structures from modularity, hierarchy, complexity, and fault propagation points of view. A model based on these four aspects is proposed to better understand software structural quality. A prediction model is also proposed to provide insights on the nature of software evolution and its current status. Experiments using two software projects were performed against the thresholds obtained by evaluating more than 5,000 versions of open source projects. Our results suggest that the approach described in this paper can help us analyze real-world software projects for better quality evaluation.

With the increase of scale and complexity of software systems as well as the long existing threats of software accidents such as Therac-25 radiation exposure and the Toyota Prius braking system failure, improving the reliability and quality has become the major pressure in developing safety-or security-critical software systems. Since more and more large-scale software systems exhibit the properties of complex network, how can we employ the concepts and metrics of complex networks to the area of software engineering? What structural patterns can be used as indications for software quality and reliability? Though applying the concept of complex network to the field of software engineering has attracted attentions from both the academia and industry, these questions have not been researched thoroughly with satisfactory unanimity. In this study, we analyzed the bug distribution in 1,047 versions of four open source software projects downloaded from GitHub with complex network theory and focused on the correlation between software bugs and a specific type of software network structure -- network motifs. Our results indicate that the functions containing bugs are more likely to be involved in feedforward loop motifs, one type of network motifs with highest degree of uniqueness. This paper could serve as a guide to further investigate the nature of software failures and a powerful tool for software fault prediction and quality evaluation.