arXiv subjects

Quanjun Zhang

Publications and source records attributed to Quanjun Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2025

Repair-R1: Better Test Before Repair

APR (Automated Program Repair) aims to automatically locate program defects, generate patches and validate the repairs. Existing techniques for APR are often combined with LLMs (Large Language Models), which leverages the code-related knowledge of LLMs to improve repair effectiveness. Current LLM-based APR methods typically utilize test cases only during the inference stage, adopting an iterative approach that performs repair first and validates it through test execution afterward. This conventional paradigm neglects two important aspects: the potential contribution of test cases in the training phase, and the possibility of leveraging testing prior to repair. To address this, we propose Repair-R1, which introduces test cases into the model's training phase and shifts test generation to precede repair. The model is required to first generate discriminative test cases that can distinguish defective behaviors, and then perform repair based on these tests. This enables the model to better locate defects and understand the underlying causes of defects, thereby improving repair effectiveness. We implement Repair-R1 with three different backbone models, using RL (reinforcement learning) to co-optimize test generation and bug repair. Experimental results on four widely adopted benchmarks demonstrate the superiority of Repair-R1. Specially, compared to vanilla models, Repair-R1 improves repair success rate by 2.68\% to 48.29\%, test generation success rate by 16.38\% to 53.28\%, and test coverage by 0.78\% to 53.96\%. We publish the code and weights at https://github.com/Tomsawyerhu/APR-RL and https://huggingface.co/tomhu/Qwen3-4B-RL-5000-step.

cs.SE↗

arXiv2025

LRP4RAG: Detecting Hallucinations in Retrieval-Augmented Generation via Layer-wise Relevance Propagation

Retrieval-Augmented Generation (RAG) has become a primary technique for mitigating hallucinations in large language models (LLMs). However, incomplete knowledge extraction and insufficient understanding can still mislead LLMs to produce irrelevant or even contradictory responses, which means hallucinations persist in RAG. In this paper, we propose LRP4RAG, a method based on the Layer-wise Relevance Propagation (LRP) algorithm for detecting hallucinations in RAG. Specifically, we first utilize LRP to compute the relevance between the input and output of the RAG generator. We then apply further extraction and resampling to the relevance matrix. The processed relevance data are input into multiple classifiers to determine whether the output contains hallucinations. To the best of our knowledge, this is the first time that LRP has been used for detecting RAG hallucinations, and extensive experiments demonstrate that LRP4RAG outperforms existing baselines.

cs.CL↗

arXiv2025

Large Language Models for Unit Testing: A Systematic Literature Review

Unit testing is a fundamental practice in modern software engineering, with the aim of ensuring the correctness, maintainability, and reliability of individual software components. Very recently, with the advances in Large Language Models (LLMs), a rapidly growing body of research has leveraged LLMs to automate various unit testing tasks, demonstrating remarkable performance and significantly reducing manual effort. However, due to ongoing explorations in the LLM-based unit testing field, it is challenging for researchers to understand existing achievements, open challenges, and future opportunities. This paper presents the first systematic literature review on the application of LLMs in unit testing until March 2025. We analyze \numpaper{} relevant papers from the perspectives of both unit testing and LLMs. We first categorize existing unit testing tasks that benefit from LLMs, e.g., test generation and oracle generation. We then discuss several critical aspects of integrating LLMs into unit testing research, including model usage, adaptation strategies, and hybrid approaches. We further summarize key challenges that remain unresolved and outline promising directions to guide future research in this area. Overall, our paper provides a systematic overview of the research landscape to the unit testing community, helping researchers gain a comprehensive understanding of achievements and promote future research. Our artifacts are publicly available at the GitHub repository: https://github.com/iSEngLab/AwesomeLLM4UT.

cs.SE↗

arXiv2025

Security of Language Models for Code: A Systematic Literature Review

Language models for code (CodeLMs) have emerged as powerful tools for code-related tasks, outperforming traditional methods and standard machine learning approaches. However, these models are susceptible to security vulnerabilities, drawing increasing research attention from domains such as software engineering, artificial intelligence, and cybersecurity. Despite the growing body of research focused on the security of CodeLMs, a comprehensive survey in this area remains absent. To address this gap, we systematically review 67 relevant papers, organizing them based on attack and defense strategies. Furthermore, we provide an overview of commonly used language models, datasets, and evaluation metrics, and highlight open-source tools and promising directions for future research in securing CodeLMs.

cs.SE↗

arXiv2025

TestART: Improving LLM-based Unit Testing via Co-evolution of Automated Generation and Repair Iteration

Unit testing is crucial for detecting bugs in individual program units but consumes time and effort. Recently, large language models (LLMs) have demonstrated remarkable capabilities in generating unit test cases. However, several problems limit their ability to generate high-quality unit test cases: (1) compilation and runtime errors caused by the hallucination of LLMs; (2) lack of testing and coverage feedback information restricting the increase of code coverage;(3) the repetitive suppression problem causing invalid LLM-based repair and generation attempts. To address these limitations, we propose TestART, a novel unit test generation method. TestART improves LLM-based unit testing via co-evolution of automated generation and repair iteration, representing a significant advancement in automated unit test generation. TestART leverages the template-based repair strategy to effectively fix bugs in LLM-generated test cases for the first time. Meanwhile, TestART extracts coverage information from successful test cases and uses it as coverage-guided testing feedback. It also incorporates positive prompt injection to prevent repetition suppression, thereby enhancing the sufficiency of the final test case. This synergy between generation and repair elevates the correctness and sufficiency of the produced test cases significantly beyond previous methods. In comparative experiments, TestART demonstrates an 18% improvement in pass rate and a 20% enhancement in coverage across three types of datasets compared to baseline models. Additionally, it achieves better coverage rates than EvoSuite with only half the number of test cases. These results demonstrate TestART's superior ability to produce high-quality unit test cases by harnessing the power of LLMs while overcoming their inherent flaws.

cs.SE↗

arXiv2025

Improving Retrieval-Augmented Deep Assertion Generation via Joint Training

Unit testing attempts to validate the correctness of basic units of the software system under test and has a crucial role in software development and testing. Very recent work proposes a retrieve-and-edit approach to generate unit test oracles, i.e., assertions. Despite being promising, it is still far from perfect due to some limitations, such as splitting assertion retrieval and generation into two separate components without benefiting each other. In this paper, we propose AG-RAG, a retrieval-augmented automated assertion generation approach that leverages external codebases and joint training to address various technical limitations of prior work. Inspired by the plastic surgery hypothesis, AG-RAG attempts to combine relevant unit tests and advanced pre-trained language models (PLMs) with retrieval-augmented fine-tuning. AG-RAG builds a dense retriever to search for relevant test-assert pairs (TAPs) with semantic matching and a retrieval-augmented generator to synthesize accurate assertions with the focal-test and retrieved TAPs as input. Besides, AG-RAG leverages a code-aware language model CodeT5 as the cornerstone to facilitate both assertion retrieval and generation tasks. Furthermore, the retriever is optimized in conjunction with the generator as a whole pipeline with a joint training strategy. This unified design fully adapts both components specifically for retrieving more useful TAPs, thereby generating accurate assertions. We extensively evaluate AG-RAG against six state-of-the-art AG approaches on two benchmarks and three metrics. Experimental results show that AG-RAG significantly outperforms previous AG approaches on all benchmarks and metrics, e.g., improving the most recent baseline EditAS by 20.82% and 26.98% in terms of accuracy. AG-RAG also correctly generates 1739 and 2866 unique assertions that all baselines fail to generate, 3.45X and 9.20X more than EditAS.

cs.SE↗

arXiv2025

Improving Deep Assertion Generation via Fine-Tuning Retrieval-Augmented Pre-trained Language Models

Unit testing validates the correctness of the units of the software system under test and serves as the cornerstone in improving software quality and reliability. To reduce manual efforts in writing unit tests, some techniques have been proposed to automatically generate test assertions, with recent integration-based approaches considered state-of-the-art. Despite being promising, such integration-based approaches face several limitations, including reliance on lexical matching for assertion retrieval and a limited training corpus for assertion generation. This paper proposes a novel retrieval-augmented deep assertion generation approach, namely RetriGen, based on a hybrid retriever and a pre-trained language model (PLM)-based generator. Given a focal-test, RetriGen first builds a hybrid assertion retriever to search for the most relevant Test-Assert Pair from external codebases. The retrieval process considers lexical similarity and semantical similarity via a token-based and an embedding-based retriever, respectively. RetriGen then treats assertion generation as a sequence-to-sequence task and designs a PLM-based assertion generator to predict a correct assertion. We conduct extensive experiments to evaluate RetriGen against six state-of-the-art approaches across two large-scale datasets and two metrics. The results demonstrate that RetriGen achieves 57.66% accuracy and 73.24% CodeBLEU, outperforming all baselines with average improvements of 50.66% and 14.14%, respectively.

cs.SE↗

arXiv2025

Eliminating Backdoors in Neural Code Models for Secure Code Understanding

Neural code models (NCMs) have been widely used to address various code understanding tasks, such as defect detection. However, numerous recent studies reveal that such models are vulnerable to backdoor attacks. Backdoored NCMs function normally on normal/clean code snippets, but exhibit adversary-expected behavior on poisoned code snippets injected with the adversary-crafted trigger. It poses a significant security threat. Therefore, there is an urgent need for effective techniques to detect and eliminate backdoors stealthily implanted in NCMs. To address this issue, in this paper, we innovatively propose a backdoor elimination technique for secure code understanding, called EliBadCode. EliBadCode eliminates backdoors in NCMs by inverting/reverse-engineering and unlearning backdoor triggers. Specifically, EliBadCode first filters the model vocabulary for trigger tokens based on the naming conventions of specific programming languages to reduce the trigger search space and cost. Then, EliBadCode introduces a sample-specific trigger position identification method, which can reduce the interference of non-backdoor (adversarial) perturbations for subsequent trigger inversion, thereby producing effective inverted backdoor triggers efficiently. Backdoor triggers can be viewed as backdoor (adversarial) perturbations. Subsequently, EliBadCode employs a Greedy Coordinate Gradient algorithm to optimize the inverted trigger and designs a trigger anchoring method to purify the inverted trigger. Finally, EliBadCode eliminates backdoors through model unlearning. We evaluate the effectiveness of EliBadCode in eliminating backdoors implanted in multiple NCMs used for three safety-critical code understanding tasks. The results demonstrate that EliBadCode can effectively eliminate backdoors while having minimal adverse effects on the normal functionality of the model.

cs.CR↗

arXiv2025

ACTesting: Automated Cross-modal Testing Method of Text-to-Image Software

Recently, creative generative artificial intelligence software has emerged as a pivotal assistant, enabling users to generate content and seek inspiration rapidly. Text-to-Image (T2I) software, one of the most widely used, synthesizes images with text input by engaging in a cross-modal process. However, despite substantial advancements in the T2I engine, T2I software still encounters errors when generating complex or non-realistic scenes, including omitting focal entities, low image realism, and mismatched text-image information. The cross-modal nature of T2I software complicates error detection for traditional testing methods, and the absence of test oracles further exacerbates the complexity of the testing process. To fill this gap, we propose ACTesting, an Automated Cross-modal Testing Method of Text-to-Image Software, the first testing method explicitly designed for T2I software. ACTesting utilizes the metamorphic testing principle to address the oracle problem and identifies cross-modal semantic consistency as its fundamental Metamorphic relation (MR) by employing the Entity-relationship (ER) triples. We design three kinds of mutation operators under the guidance of MR and the adaptability density constraint to construct the new input text. After generating the images based on the text, ACTesting verifies whether MR is satisfied by detecting the ER triples across two modalities to detect the errors of T2I software. In our experiments across five popular T2I software, ACTesting effectively generates error-revealing tests, resulting in a decrease in text-image consistency by up to 20% when compared to the baseline. Additionally, an ablation study demonstrates the efficacy of the proposed mutation operators. The experimental results validate that ACTesting can reliably identify errors within T2I software.

cs.SE↗

arXiv2024

A Large-scale Empirical Study on Fine-tuning Large Language Models for Unit Testing

Unit testing plays a pivotal role in software development, improving software quality and reliability. However, generating effective test cases manually is time-consuming, prompting interest in unit testing research. Recently, Large Language Models (LLMs) have shown potential in various unit testing tasks, including test generation, assertion generation, and test evolution, but existing studies are limited in scope and lack a systematic evaluation of the effectiveness of LLMs. To bridge this gap, we present a large-scale empirical study on fine-tuning LLMs for unit testing. Our study involves three unit testing tasks, five benchmarks, eight evaluation metrics, and 37 popular LLMs across various architectures and sizes, consuming over 3,000 NVIDIA A100 GPU hours. We focus on three key research questions: (1) the performance of LLMs compared to state-of-the-art methods, (2) the impact of different factors on LLM performance, and (3) the effectiveness of fine-tuning versus prompt engineering. Our findings reveal that LLMs outperform existing state-of-the-art approaches on all three unit testing tasks across nearly all metrics, highlighting the potential of fine-tuning LLMs in unit testing tasks. Furthermore, large-scale, decoder-only models achieve the best results across tasks, while encoder-decoder models perform better under the same parameter scale. Additionally, the comparison of the performance between fine-tuning and prompt engineering approaches reveals the considerable potential capability of the prompt engineering approach in unit testing tasks. We then discuss the concerned issues on the test generation task, including data leakage issues, bug detection capabilities, and metrics comparisons. Finally, we further pinpoint carious practical guidelines for LLM-based approaches to unit testing tasks in the near future.

cs.SE↗

arXiv2024

Can GPT-O1 Kill All Bugs? An Evaluation of GPT-Family LLMs on QuixBugs

LLMs have long demonstrated remarkable effectiveness in automatic program repair (APR), with OpenAI's ChatGPT being one of the most widely used models in this domain. Through continuous iterations and upgrades of GPT-family models, their performance in fixing bugs has already reached state-of-the-art levels. However, there are few works comparing the effectiveness and variations of different versions of GPT-family models on APR. In this work, inspired by the recent public release of the GPT-o1 models, we conduct the first study to compare the effectiveness of different versions of the GPT-family models in APR. We evaluate the performance of the latest version of the GPT-family models (i.e., O1-preview and O1-mini), GPT-4o, and the historical version of ChatGPT on APR. We conduct an empirical study of the four GPT-family models against other LLMs and APR techniques on the QuixBugs benchmark from multiple evaluation perspectives, including repair success rate, repair cost, response length, and behavior patterns. The results demonstrate that O1's repair capability exceeds that of prior GPT-family models, successfully fixing all 40 bugs in the benchmark. Our work can serve as a foundation for further in-depth exploration of the applications of GPT-family models in APR.

cs.SE↗

arXiv2024

TestBench: Evaluating Class-Level Test Case Generation Capability of Large Language Models

Software testing is a crucial phase in the software life cycle, helping identify potential risks and reduce maintenance costs. With the advancement of Large Language Models (LLMs), researchers have proposed an increasing number of LLM-based software testing techniques, particularly in the area of test case generation. Despite the growing interest, limited efforts have been made to thoroughly evaluate the actual capabilities of LLMs in this task. In this paper, we introduce TestBench, a benchmark for class-level LLM-based test case generation. We construct a dataset of 108 Java programs from 9 real-world, large-scale projects on GitHub, each representing a different thematic domain. We then design three distinct types of prompts based on context descriptions, including self-contained context, full context, and simple context. Besides, we propose a fine-grained evaluation framework that considers five aspects of test cases: syntactic correctness, compilation correctness, test correctness, code coverage rate, and defect detection rate. Furthermore, we propose a heuristic algorithm to repair erroneous test cases generated by LLMs. We evaluate CodeLlama-13b, GPT-3.5, and GPT-4 on the TestBench, and our experimental results indicate that larger models demonstrate a greater ability to effectively utilize contextual information, thus generating higher-quality test cases. Smaller models may struggle with the noise introduced by the extensive information contained within the full context. However, when using the simplified version, namely the simple context, which is derived from the full context via abstract syntax tree analysis, the performance of these models improves significantly. Our analysis highlights the current progress and pinpoints future directions to further enhance the effectiveness of models by handling contextual information for test case generation.

cs.SE↗

arXiv2024

A Survey on Large Language Models for Software Engineering

Software Engineering (SE) is the systematic design, development, maintenance, and management of software applications underpinning the digital infrastructure of our modern world. Very recently, the SE community has seen a rapidly increasing number of techniques employing Large Language Models (LLMs) to automate a broad range of SE tasks. Nevertheless, existing information of the applications, effects, and possible limitations of LLMs within SE is still not well-studied. In this paper, we provide a systematic survey to summarize the current state-of-the-art research in the LLM-based SE community. We summarize 62 representative LLMs of Code across three model architectures, 15 pre-training objectives across four categories, and 16 downstream tasks across five categories. We then present a detailed summarization of the recent SE studies for which LLMs are commonly utilized, including 947 studies for 112 specific code-related tasks across five crucial phases within the SE workflow. We also discuss several critical aspects during the integration of LLMs into SE, such as empirical evaluation, benchmarking, security and reliability, domain tuning, compressing and distillation. Finally, we highlight several challenges and potential opportunities on applying LLMs for future SE studies, such as exploring domain LLMs and constructing clean evaluation datasets. Overall, our work can help researchers gain a comprehensive understanding about the achievements of the existing LLM-based SE studies and promote the practical application of these techniques. Our artifacts are publicly available and will be continuously updated at the living repository: https://github.com/iSEngLab/AwesomeLLM4SE.

cs.SE↗

arXiv2024

No Man is an Island: Towards Fully Automatic Programming by Code Search, Code Generation and Program Repair

Automatic programming attempts to minimize human intervention in the generation of executable code, and has been a long-standing challenge in the software engineering community. To advance automatic programming, researchers are focusing on three primary directions: (1) code search that reuses existing code snippets from external databases; (2) code generation that produces new code snippets from natural language; and (3) program repair that refines existing code snippets by fixing detected bugs. Despite significant advancements, the effectiveness of state-of-the-art techniques is still limited, such as the usability of searched code and the correctness of generated code. Motivated by the real-world programming process, where developers usually use various external tools to aid their coding processes, such as code search engines and code testing tools, in this work, we propose \toolname{}, an automatic programming framework that leverages recent large language models (LLMs) to integrate the three research areas to address their inherent limitations. In particular, our framework first leverages different code search strategies to retrieve similar code snippets, which are then used to further guide the code generation process of LLMs. Our framework further validates the quality of generated code by compilers and test cases, and constructs repair prompts to query LLMs for generating correct patches. We conduct preliminary experiments to demonstrate the potential of our framework, \eg helping CodeLlama solve 267 programming problems with an improvement of 62.53\%. As a generic framework, \toolname{} can integrate various code search, generation, and repair tools, combining these three research areas together for the first time. More importantly, it demonstrates the potential of using traditional SE tools to enhance the usability of LLMs in automatic programming.

cs.SE↗

arXiv2024

ESALE: Enhancing Code-Summary Alignment Learning for Source Code Summarization

(Source) code summarization aims to automatically generate succinct natural language summaries for given code snippets. Such summaries play a significant role in promoting developers to understand and maintain code. Inspired by neural machine translation, deep learning-based code summarization techniques widely adopt an encoder-decoder framework, where the encoder transforms given code snippets into context vectors, and the decoder decodes context vectors into summaries. Recently, large-scale pre-trained models for source code are equipped with encoders capable of producing general context vectors and have achieved substantial improvements on code summarization. However, although they are usually trained mainly on code-focused tasks and can capture general code features, they still fall short in capturing specific features that need to be summarized. This paper proposes a novel approach to improve code summarization based on summary-focused tasks. Specifically, we exploit a multi-task learning paradigm to train the encoder on three summary-focused tasks to enhance its ability to learn code-summary alignment, including unidirectional language modeling (ULM), masked language modeling (MLM), and action word prediction (AWP). Unlike pre-trained models that mainly predict masked tokens in code snippets, we design ULM and MLM to predict masked words in summaries. Intuitively, predicting words based on given code snippets would help learn the code-summary alignment. Additionally, we introduce the domain-specific task AWP to enhance the ability of the encoder to learn the alignment between action words and code snippets. The extensive experiments on four datasets demonstrate that our approach, called ESALE significantly outperforms baselines in all three widely used metrics, including BLEU, METEOR, and ROUGE-L.

cs.SE↗

arXiv2024

A Critical Review of Large Language Model on Software Engineering: An Example from ChatGPT and Automated Program Repair

Large Language Models (LLMs) have been gaining increasing attention and demonstrated promising performance across a variety of Software Engineering (SE) tasks, such as Automated Program Repair (APR), code summarization, and code completion. For example, ChatGPT, the latest black-box LLM, has been investigated by numerous recent research studies and has shown impressive performance in various tasks. However, there exists a potential risk of data leakage since these LLMs are usually close-sourced with unknown specific training details, e.g., pre-training datasets. In this paper, we seek to review the bug-fixing capabilities of ChatGPT on a clean APR benchmark with different research objectives. We first introduce {\benchmark}, a new benchmark with buggy and the corresponding fixed programs from competitive programming problems starting from 2023, after the training cutoff point of ChatGPT. The results on {\benchmark} show that ChatGPT is able to fix 109 out of 151 buggy programs using the basic prompt within 35 independent rounds, outperforming state-of-the-art LLMs CodeT5 and PLBART by 27.5\% and 62.4\% prediction accuracy. We also investigate the impact of three types of prompts, i.e., problem description, error feedback, and bug localization, leading to additional 34 fixed bugs. Besides, we provide additional discussion from the interactive nature of ChatGPT to illustrate the capacity of a dialog-based repair workflow with 9 additional fixed bugs. Inspired by the findings, we further pinpoint various challenges and opportunities for advanced SE study equipped with such LLMs (e.g.,~ChatGPT) in the near future. More importantly, our work calls for more research on the reevaluation of the achievements obtained by existing black-box LLMs across various SE tasks, not limited to ChatGPT on APR.

cs.SE↗

arXiv2024

Pre-trained Model-based Actionable Warning Identification: A Feasibility Study

Actionable Warning Identification (AWI) plays a pivotal role in improving the usability of static code analyzers. Currently, Machine Learning (ML)-based AWI approaches, which mainly learn an AWI classifier from labeled warnings, are notably common. However, these approaches still face the problem of restricted performance due to the direct reliance on a limited number of labeled warnings to develop a classifier. Very recently, Pre-Trained Models (PTMs), which have been trained through billions of text/code tokens and demonstrated substantial success applications on various code-related tasks, could potentially circumvent the above problem. Nevertheless, the performance of PTMs on AWI has not been systematically investigated, leaving a gap in understanding their pros and cons. In this paper, we are the first to explore the feasibility of applying various PTMs for AWI. By conducting the extensive evaluation on 10K+ SpotBugs warnings from 10 large-scale and open-source projects, we observe that all studied PTMs are consistently 9.85%~21.12% better than the state-of-the-art ML-based AWI approaches. Besides, we investigate the impact of three primary aspects (i.e., data preprocessing, model training, and model prediction) in the typical PTM-based AWI workflow. Further, we identify the reasons for current PTMs' underperformance on AWI. Based on our findings, we provide several practical guidelines to enhance PTM-based AWI in future work.

cs.SE↗

arXiv2024

APPT: Boosting Automated Patch Correctness Prediction via Fine-tuning Pre-trained Models

Automated program repair (APR) aims to fix software bugs automatically without human debugging efforts and plays a crucial role in software development and maintenance. Despite promising, APR is still challenged by a long-standing overfitting problem (i.e., the generated patch is plausible but overfitting). Various techniques have thus been proposed to address the overfitting problem. Recently, researchers have employed BERT to extract code features, which are then used to train a classifier for patch correctness prediction. However, BERT is restricted to feature extraction for classifier training without benefiting from the training process, potentially generating sub-optimal vector representations for patched code snippets. In this paper, we propose APPT, a pre-trained model-based automated patch correctness assessment technique by both pre-training and fine-tuning. APPT adopts a pre-trained model as the encoder stack, followed by an LSTM stack and a deep learning classifier. More importantly, the pre-trained model is fine-tuned in conjunction with other components as a whole pipeline to fully adapt it specifically for reasoning about patch correctness. We conduct an extensive experiment on 1,183 Defects4J patches and the experimental results show that APPT achieves prediction accuracy of 79.7% and recall of 83.2%, outperforming CACHE by 4.3% and 6.7%. Our additional investigation on 49,694 real-world patches shows that APPT achieves the optimum performance compared with existing representation learning techniques. We further investigate the impact of each component and find that they all positively contribute to APPT, e.g., the fine-tuning process and the LSTM stack increase F1-score by 10.22% and 4.11%, respectively. We also prove that adopting advanced pre-trained models can further provide substantial advancement, highlighting the generalizability of APPT.

cs.SE↗