Prediction-based fault localization methods aim to perform fault localization by reducing the cost of manually verifying test execution results. Previous studies have shown that this type of method demonstrates certain effectiveness in single-bug programs. However, in multi-bug programs-especially when fault interference exists-the applicability of such methods still requires further investigation. In this paper, we take an existing prediction-based fault localization method that predicts test case execution results as the research subject, analyze its localization performance in multi-bug programs, and particularly examine the impact of fault interference on its effectiveness. We apply this method to six multi-bug program cases, including scenarios both with and without fault interference, and evaluate the localization performance. The experimental results show that, in multi-bug programs without fault interference, prediction-based fault localization methods can generally play a positive role in localization effectiveness. However, In the presence of fault interference, the localization performance of this method exhibits a certain degree of instability.

Large language models often generate plausible but unverifiable reasoning steps, limiting their applicability in domains that demand traceability and adherence to explicit constraints. This paper proposes a unified symbolic-neural framework that integrates entity-attribute-relation knowledge modeling with a transition algebra to support interpretable and verifiable inference. Heterogeneous inputs are normalized into typed entity-attribute-relation structures, which are then lifted into algebraic expressions equipped with operators for sequencing, guarded evaluation, and multimodal mapping. A labeled small-step semantics is defined over these expressions, yielding finite derivation traces that can be independently checked without executing external code. We formally establish three semantic properties-closure, constraint preservation, and replayability-and show how they contribute to stable and auditable reasoning. Experiments on a 576-task benchmark covering structural, constraint-aware, multimodal, ambiguous, and knowledge-augmented scenarios demonstrate that algebraic conditioning improves answer accuracy, trace faithfulness, constraint satisfaction, and interpretability over plain prompting, chain-of-thought, and program-aided baselines. The results show that algebraic semantics provide a practical and formally analyzable mechanism for strengthening the reliability of large language model reasoning.

Parallel fault localization is an advanced technique in software testing and debugging that aims to improve the efficiency and effectiveness of fault localization by analyzing multiple potential fault locations concurrently, ultimately contributing to the enhancement of software quality. A key challenge in this process is selecting an appropriate metric to measure the distance between two SBFL rankings which is crucial for generating high-quality fault-focused clusters and minimizing costs associated with parallel fault localization. Previous studies have shown the potential of metrics such as Jaccard Distance (JD), which considers only partial rankings, alongside commonly used ranking-based metrics like Spearman’s Footrule (SF) and the set intersection-based measure (SBM), in achieving comparable or even superior clustering and fault localization performance compared to Kendall tau (KT). Building on these insights, this paper explores the application of partial rankings to KT, SF, and SBM for parallel fault localization. Rather than evaluating all elements within a ranking, we focus solely on the top-ranking elements when computing distances. To this end, KT, SF, and SBM are redefined to accommodate partial rankings. Case studies are conducted to evaluate the performance of these redesigned metrics, comparing their effectiveness in parallel fault localization against their traditional, full-ranking counterparts.

Software fault localization (FL) is a critical step in debugging, aiming to identify the defective elements in source code with cost-effective effort. Recent advances in large language models (LLMs) have introduced new possibilities for automating and augmenting FL through code understanding, reasoning, and integration with dynamic analysis tools. This paper conducts a literature review of 21 studies on LLM-based FL. Specifically, reviewed studies are classified into five categories: (1) test-free/code-only localization, (2) agentic and tool-augmented frameworks, (3) crash/stack-trace-driven methods, (4) LLM-augmented information-retrieval-based FL, and (5) empirical analysis. For each category, we analyze methodological design choices, including the types of input leveraged (code, tests, logs, bug reports), localization granularity (file, method, statement, token), prompting strategies, fine-tuning vs. incontext learning, and ranking mechanisms. We also summarize datasets, evaluation metrics, and reported performance trends, drawing comparisons across LLMs. Finally, we outline open challenges and research directions for LLM-based FL. This survey provides a knowledge base for researchers and practitioners seeking to advance LLMdriven fault localization.

Large language models (LLMs) have been widely adopted across various industries, but issues such as hallucinations, biases, and erroneous outputs frequently arise, compromising their reliability and safety. Objectively assessing how well an LLM interprets user queries is crucial for selecting the right model for practical problem-solving. This paper proposes an evaluation method that leverages user knowledge graphs to measure an LLM's comprehension of user input. Specifically, we construct both text-based and graphical test cases derived from a user knowledge graph, thereby enabling multi-modal assessment of the LLM's understanding. To implement our approach, a knowledge graph is first built from the user's input and then mutated to produce diverse test cases. A search-based testing method is then applied to evaluate the model's comprehension. We provide a case study demonstrating the framework. Our findings indicate that multi-modal test cases outperform purely text-based test cases in revealing the true understanding capability of LLMs. Among the eight models tested, DeepSeek and Doubao exhibit stronger comprehension than the remaining six.

With the growing adoption of knowledge graphs (KGs) across artificial intelligence, knowledge management, and software engineering, the demand for a unified and semantically coherent modeling framework has intensified. Existing paradigms often suffer from representational fragmentation and limited logical expressiveness, hindering cross-domain integration and reuse. This paper proposes a formal set-theoretic modeling approach that unifies entities, attributes, and relations within a structured tripartite framework—Entity–Attribute–Relation (EAR). We define a semantically typed relation system comprising core and extensible relation types, enabling logical inference, attribute aggregation, and recursive modeling. A progressive construction pipeline facilitates systematic KG development from heterogeneous inputs. Using the classic "chicken–rabbit in a cage" puzzle, we illustrate the model’s expressiveness and reasoning capabilities, and provide a comparative analysis with RDF and property graph approaches. Theoretical analysis and illustrative examples suggest that EAR offers improved semantic clarity, structural consistency, and extensibility, indicating strong potential for application in domains such as education, healthcare, and geospatial systems.

As large language models (LLMs) increasingly incorporate role-playing capabilities for personalized interaction and entertainment; a new class of security threats has emerged-role-playing prompt attacks. This paper presents a preliminary investigation into how adversarial users exploit such prompts to bypass safety mechanisms and elicit harmful, unethical, or prohibited responses from state-of-the-art LLMs. Using a curated dataset of 86 attack scenarios spanning diverse personas (e.g., fictional villains, film characters, expert criminals), we analyze prompts, both in Chinese and English, that simulate morally questionable or illegal tasks-including bank robbery, drug manufacturing, cyberterrorism, and targeted violence-by embedding them within cinematic narratives or fictional scripts. Our case study results reveal that when models are instructed to “act as” a character from a movie or TV show, or to “help write a script,” LLMs frequently produce detailed and contextually appropriate responses, offering step-by-step procedures, tools, and planning strategies. These findings raise pressing concerns about the blurred boundary between fictional narrative generation and real-world misuse, highlighting the limitations of current safety filters in recognizing and mitigating contextually embedded adversarial intent.

The software expression model serves as a formalized specification, accurately depicting software behavior and generating test sequences through algebraic operations derived from the model. Typically, automated algebraic manipulation involves constructing an abstract syntax tree (AST) for the expression, followed by traversing it to identify subexpressions. However, this approach introduces a significant amount of redundant algebraic operations, diminishing the efficiency of expression parsing. To address this challenge, this paper introduces HT-EP, an innovative hierarchical tree-based expression parsing algorithm. HT-EP transforms expressions into hierarchical trees, utilizing algebraic operations to process nodes efficiently and generate streamlined test sequences. Compared to ASTs, hierarchical trees exhibit a simplified structure with fewer nodes, enabling faster traversal. Our experiment involved 124 expressions from scholarly papers over the past six decades and core functional expressions from 15 open-source software projects. The goal was to assess the parsing and fault detection capabilities of HT-EP against four other expression parsing algorithms. Additionally, we compared the complexities of hierarchical trees and ASTs, exploring factors influencing hierarchical tree complexity. Experimental results reveal that the HT-EP algorithm excels in parsing and software fault detection capabilities compared to the other four algorithms. Furthermore, for expressions derived from real-world cases, HT-EP achieves an approximate 40% reduction in redundant algebraic operation steps and an average 63% reduction in runtime compared to AST-EP.

Large language models (LLMs) have demonstrated significant prowess in code analysis and natural language processing, making them highly valuable for software testing. This paper conducts a comprehensive evaluation of LLMs applied to software testing, with a particular emphasis on test case generation, error tracing, and bug localization across twelve open-source projects. The advantages and limitations, as well as recommendations associated with utilizing LLMs for these tasks, are delineated. Furthermore, we delve into the phenomenon of hallucination in LLMs, examining its impact on software testing processes and presenting solutions to mitigate its effects. The findings of this work contribute to a deeper understanding of integrating LLMs into software testing, providing insights that pave the way for enhanced effectiveness in the field.

This paper mainly studies the security problems of false data injection attack (FDIA) and external environment interference of unmanned surface vessels. FDIA and environmental disturbance are considered at the same time, and the interference of these two parts is analyzed separately. First, the attack scenario is analyzed. A model of FDIA with great concealment is established. Secondly, we introduce a detection mechanism that identifies system interference by analyzing the residual error of the authentication signal. To mitigate FDIA’s impact on the unmanned vessels control system, we derive and establish a formula for its influence on the ship’s dynamics. Additionally, an observer is designed to estimate the extent of FDIA. Finally,a simulation is provided to confirm the method’s effectiveness.

Currently, the mortality rate of Myocardial Infarction (MI) is over 30%, and portable intelligent Electrocardiographic (ECG) monitoring systems are one of the most effective methods to reduce the mortality rate of MI. Embedding deep learning networks into monitoring systems can greatly improve the accuracy of MI detection, but the system will also face challenges such as the limited computational resources of edge devices and how to build a highly accurate ECG-assisted diagnostic model on the cloud devices. This paper proposes a portable intelligent ECG monitoring system based on an end-edge-cloud architecture. On edge devices, we employ a lightweight ECG anomaly detection method based on Morphological Features and Core Density Ratio (MCDR) to alleviate computational pressure. On cloud devices, we utilize a Convolutional Neural Network - Long Short-Term Memory (CNN-LSTM) fusion model to receive ECG signals preliminarily identified as abnormal by edge devices, assisting in Myocardial Infarction (MI) detection and type determination at the edge. This study was tested on the PTB diagnostic ECG database, achieving an accuracy of 88.6% for MI detection on edge devices. With assistance from cloud devices, the accuracy of MI detection and type discrimination reached 99.2%. The system significantly reduces the computational power required by edge devices, providing MI patients with a more portable and accurate intelligent ECG monitoring platform. It can also serve as an auxiliary tool for clinical doctors.

With the rapid development of internet technology, open-source software mirror sites have become indispensable tools for developers and tech enthusiasts, serving as crucial platforms for resource acquisition. We selected and tested two open-source applications by downloading them from the Tsinghua University Open Source Software Mirror Site and the Nanjing University Open Source Software Mirror Site. By implementing black-box testing strategies, we identified several bugs and design flaws in the software. Furthermore, we utilized large language models such as ChatGPT-3.5, Gemini, Kimi, Llama, ERNIE Bot, and Tongyi Qianwen to analyze the potential causes of these issues, exploring new approaches to AI-assisted software quality assurance. Through comparative analysis of feedback from multiple interactions with these large language models, we systematically evaluated their effectiveness in the field of software testing. This study provides empirical evidence for optimizing model applications and enhancing testing efficiency.

Since the advent of Large Language Models (LLMs) such as ChatGPT, natural language processing has seen significant progress, providing timely, context-aware, and intelligent solutions across various domains. While users benefit greatly from these advancements, increasing safety concerns have surfaced, highlighting vulnerabilities and their impact on daily tasks. It is crucial to understand the implications of LLMs and not simply trust their outputs blindly. This paper reviews literature on LLM safety issues including data privacy, security, adversarial attacks, and compliance with laws and ethical guidelines.. Specifically, limitations, risks, testing methods, and evaluation criteria from a safety perspective are discussed. Based on these insights, it proposes suggestions to mitigate current safety challenges. The goal is to aid researchers and developers in designing safer, more reliable LLMs and to enhance user interactions with these technologies.

It has become a popular trend to build software’s regular expression or extended regular expression models in order to generate test sequences from these models. Such test sequences tend to have promising test coverage and fault detection capability. During this process, one critical step is expression parsing based on algebraic operations. However, the parsing can be very challenging as different algebraic systems have different algebraic operators and algebraic operations. Besides, the parsing difficulty continues to grow as software complexity increases. To address the above challenges, this paper proposes a general expression parsing framework for test sequence generation. The proposed framework consists of three stages, expression decomposition, algebraic operations, and subexpression combination. To implement the framework, an expression parsing algorithm based on abstract syntax tree is developed. Case studies based on 117 expressions collected from the literature over the past 30 years as well as 13 software systems are conducted to evaluate the effectiveness of the proposed algorithm. The results indicate that our algorithm is superior to three existing and commonly used algorithms with respect to expression parsing and software fault detection.

Since its public launch at the end of 2022, ChatGPT has garnered global attention, showcasing the diverse capabilities of AI in tackling human tasks. Its rapid growth and widespread adoption have permeated every corner of our daily routine. This paper provides a quick peek at the impact of ChatGPT from the perspective of software engineering education. Specifically, to make our case study creative and interesting, we compare the impact answered by ChatGPT with the real feedback from existing literature. In this way, we explore the potential of ChatGPT as a teaching and learning tool in software engineering.

Heart disease is one of the leading causes of death globally, and early diagnosis and prevention of heart disease are of great significance. The paper provides a correlation analysis of the quantitative and qualitative variables on a heart disease dataset. Based on the Pearson correlation coefficient matrix, the paper determines the base classifiers and uses logistic regression as the meta-classifier to construct a stacking ensemble learning model for heart disease classification prediction. To address the issue of data imbalance, cost-sensitive learning is further introduced, and a stacking heart disease classification prediction model based on threshold optimization is established. The experimental result shows that the stacking heart disease classification prediction model achieves a classification accuracy of 90.16% and demonstrates good generalization ability. Our work provides guidance for further research on heart disease classification prediction.

In automated and autonomous driving, we need to verify the behavior considering critical interactions with the environment. To find such interactions, we need an appropriate test suite generation methodology. This paper summarizes the findings we published last year in a journal comparing combinatorial, search-based, and random testing utilizing the same input model. We focused on detecting faults and combinatorial coverage, considering two automated emergency braking function implementations in our experimental evaluation.