Large Language Models for Software Engineering PDF

1 Large Language Models for Software Engineering: A...

1 Large Language Models for Software Engineering: A Systematic Literature Review XINYI HOU∗ , Huazhong University of Science and Technology, China YANJIE ZHAO∗ , Huazhong University of Science and Technology, China arXiv:2308.10620v6 [cs.SE] 10 Apr 2024 YUE LIU, Monash University, Australia ZHOU YANG, Singapore Management University, Singapore KAILONG WANG, Huazhong University of Science and Technology, China LI LI, Beihang University, China XIAPU LUO, The Hong Kong Polytechnic University, China DAVID LO, Singapore Management University, Singapore JOHN GRUNDY, Monash University, Australia HAOYU WANG† , Huazhong University of Science and Technology, China Large Language Models (LLMs) have significantly impacted numerous domains, including Software Engi- neering (SE). Many recent publications have explored LLMs applied to various SE tasks. Nevertheless, a comprehensive understanding of the application, effects, and possible limitations of LLMs on SE is still in its early stages. To bridge this gap, we conducted a systematic literature review (SLR) on LLM4SE, with a particu- lar focus on understanding how LLMs can be exploited to optimize processes and outcomes. We select and analyze 395 research papers from January 2017 to January 2024 to answer four key research questions (RQs). In RQ1, we categorize different LLMs that have been employed in SE tasks, characterizing their distinctive features and uses. In RQ2, we analyze the methods used in data collection, preprocessing, and application, highlighting the role of well-curated datasets for successful LLM for SE implementation. RQ3 investigates the strategies employed to optimize and evaluate the performance of LLMs in SE. Finally, RQ4 examines the specific SE tasks where LLMs have shown success to date, illustrating their practical contributions to the field. From the answers to these RQs, we discuss the current state-of-the-art and trends, identifying gaps in existing research, and flagging promising areas for future study. Our artifacts are publicly available at https://github.com/xinyi-hou/LLM4SE_SLR. CCS Concepts: General and reference → Surveys and overviews; Software and its engineering → Software development techniques; Computing methodologies → Artificial intelligence. Additional Key Words and Phrases: Software Engineering, Large Language Model, Survey ∗ Co-first authors who contributed equally to this work. † Haoyu Wang is the corresponding author ([email protected]). Authors’ addresses: Xinyi Hou, [email protected], Huazhong University of Science and Technology, Wuhan, China; Yanjie Zhao, [email protected], Huazhong University of Science and Technology, Wuhan, China; Yue Liu, yue.liu1@ monash.edu, Monash University, Melbourne, Australia; Zhou Yang, [email protected], Singapore Management University, Singapore; Kailong Wang, [email protected], Huazhong University of Science and Technology, Wuhan, China; Li Li, [email protected], Beihang University, Beijing, China; Xiapu Luo, [email protected], The Hong Kong Polytechnic University, Hong Kong, China; David Lo, [email protected], Singapore Management University, Singapore; John Grundy, [email protected], Monash University, Melbourne, Australia; Haoyu Wang, [email protected], Huazhong University of Science and Technology, Wuhan, China. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. © 2024 Association for Computing Machinery. 1049-331X/2024/12-ART1 $15.00 https://doi.org/XXXXXXX.XXXXXXX ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:2 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang ACM Reference Format: Xinyi Hou, Yanjie Zhao, Yue Liu, Zhou Yang, Kailong Wang, Li Li, Xiapu Luo, David Lo, John Grundy, and Haoyu Wang. 2024. Large Language Models for Software Engineering: A Systematic Literature Review. ACM Trans. Softw. Eng. Methodol. X, Y, Article 1 (December 2024), 79 pages. https://doi.org/XXXXXXX.XXXXXXX 1 INTRODUCTION In the field of language processing, traditional Language Models (LMs) have been foundational elements, establishing a basis for text generation and understanding. Increased computational power, advanced machine learning techniques, and access to very large-scale data have led to a significant transition into the emergence of Large Language Models (LLMs) [526, 558]. Equipped with expansive and diverse training data, these models have demonstrated an impressive ability to simulate human linguistic capabilities, leading to a sea of changes across multiple domains. With their capacity to learn from massive corpora and generate plausible text, LLMs are blurring the line between human and machine-produced language. They have provided researchers and engineers alike with a powerful tool to explore the complexity and richness of human communication, consequently sparking a transformational period in the field of language processing and beyond. Software Engineering (SE) – a discipline focused on the development, implementation, and maintenance of software systems – is one of those areas reaping the benefits of the LLM rev- olution. The utilization of LLMs in SE primarily emerges from an innovative perspec- tive where numerous SE challenges can be effectively reframed into data, code, or text analy- sis tasks. Using LLMs to address these SE tasks has shown a wealth of potential break- throughs [33, 37, 210, 399, 427, 488, 489, 536]. The applicability of LLMs is particularly pronounced in tasks such as code summarization , which involves yielding an abstract natural language depiction of a code’s functionality, as well as the generation of well-structured code and code artifacts like annotations. Codex, an LLM with 12 billion parameters, has demonstrated the ability to solve 72.31% of complex Python programming challenges posed by humans. GPT-4 , an LLM from OpenAI, has been used with a strong performance in several SE tasks, encompassing code writing, understanding, execution, and reasoning. It not only handles real-world applications and diverse coding challenges but also shows the ability to explain results in natural language and generate code from pseudocode. Simultaneously, researchers have embarked on a series of research activities regarding LLM- related works, where a number of literature reviews or survey papers have been produced [36, 87, 506]. Table 1 summarises some of these. However, these related studies have limitations. They either focus narrowly on a single SE scope, such as the application of LLMs in software testing and natural-language-to-code (NL2Code) tasks , or they are primarily centered on Machine Learn- ing (ML) or Deep Learning (DL) models [452, 466, 509], overlooking more advanced and recently emerged LLM applications, such as ChatGPT , which are increasingly finding applications within the SE field [269, 400, 427, 475]. Alternatively, they merely offer a preliminary exploration of the performance of LLMs in various SE tasks through empirical experiments [74, 275, 400, 493, 521], or analyze existing partially relevant studies to reveal the challenges in this field without conducting a systematic literature survey. Furthermore, some works have investigated the applica- tion of Code LLMs in SE [543, 564], yet have not fully considered general LLMs like ChatGPT and LLaMA , which are also widely applied to various SE tasks [144, 325, 382, 497]. The integration of LLMs within SE is undoubtedly a complex endeavor, requiring key considerations including the choice of the right model, comprehension of the unique features of different LLMs, devising pre-training and fine-tuning strategies, handling of data, evaluation of outcomes, and surmounting implementation challenges. Despite the burgeoning interest and ongoing explorations in the field, a detailed and systematic review of LLMs’ application in SE has been notably absent ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:3 Table 1. State-of-the-art surveys related to LLMs for SE. Reference Year Scope of models1 Scope of SE tasks SLR2 Time frame # Collected Papers Zhang et al. 2023 Code LLM Automated program repair ✓ 2017-2023 185 Zheng et al. 2023 Code LLM General SE scope ✓ 2021-2023 149 Fan et al. 2023 LLM General SE scope × - Not specified Zan et al. 2023 LLM (12M+) NL2Code × 2020-2023 Not specified Wang et al. 2023 LLM (117M+) Software testing ✓ 2019-2023 52 Wang et al. 2022 ML, DL3 General SE scope ✓ 2009-2020 1,209 (ML) + 358 (DL) Yang et al. 2022 DL General SE scope ✓ 2015-2020 250 Watson et al. 2022 DL General SE scope ✓ 2009-2019 128 Our work 2024 LLM General SE scope ✓ 2017-2024 395 1 “M” means million and “B” means billion. The numbers in parentheses indicate the parameter sizes of LLMs. 2 SLR stands for Systematic Literature Review. This column denotes whether the paper follows an SLR process. 3 ML and DL refer to Machine Learning and Deep Learning, respectively. in the current literature. This gap signifies a need for understanding the relationship between LLMs and SE. In response, our research aims to bridge this gap, providing valuable insights to the community. In this paper, we conduct an SLR on the utilization of LLMs in SE (LLM4SE). By mapping the current state-of-the-art, pinpointing the key strengths, weaknesses, and gaps in the existing LLM4SE literature, and proposing potential avenues for future research, our review aims to provide researchers and practitioners with a thorough guide to the convergence of LLMs and SE. We anticipate that our findings will be instrumental in guiding future inquiries and advancements in this rapidly evolving field. This work makes the following key contributions: We are the first to present a comprehensive SLR on 395 papers published between January 2017 and January 2024 that focus on the use of LLM-based solutions to address SE challenges. We conducted a detailed analysis of the selected papers based on publication trends, distribution of publication venues, etc. We have classified the LLMs utilized for the reported SE tasks and have provided a summary of the usage and trends of different LLM categories within the SE domain. We describe the reported data processing stages, encompassing data collection, categorization, preprocessing, and representation. We discuss optimizers used for LLM4SE tasks, including tuning techniques, prevalent prompt engineering techniques, and commonly employed evaluation metrics. We describe the key applications of LLM4SE encompassing a diverse range of 85 specific SE tasks, grouped into six core SE activities – requirements engineering, software design, software development, software quality assurance, software maintenance, and software management. We have summarised key challenges that using LLMs encounters within the SE field and have suggested several potential research directions for LLM4SE. Section 2 presents our research questions (RQs) and elaborates on our SLR methodology. The succeeding Sections 3 to 6 are devoted to answering each of these RQs individually. Section 7 discloses the potential threats to the validity of our study. Section 8 discusses the challenges yet to be overcome when employing LLMs to solve SE tasks and highlights promising opportunities and directions for future research. Section 9 concludes the whole paper. 2 APPROACH This SLR follows the methodology proposed by Kitchenham et al. [197, 198], used in most other SE-related SLRs [229, 261, 352, 452]. Following the guidelines provided by Kitchenham et al., our ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:4 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang methodology included three main steps: planning the review (i.e., Section 2.1, 2.2), conducting the review (i.e., Section 2.3, 2.4), and analyzing the basic review results (i.e, Section 2.5). 2.1 Research Questions To provide a comprehensive overview of the LLM4SE field, it is important to fully comprehend how these models are currently being applied in SE, the challenges they face, and their potential future research directions in SE. Thus, we aim to provide an SLR of the application of LLMs to software engineering. This study thus aims to answer the following research questions: RQ1: What LLMs have been employed to date to solve SE tasks? RQ1 is designed to map out the landscape of LLMs applied in the field of SE. It seeks to identify and categorize the various LLM architectures—such as decoder-only, encoder-decoder, and encoder-only models—that have been leveraged to address diverse SE challenges. This RQ aims to provide a comprehensive overview of how these models are being utilized and the implications of their usage in this field. RQ2: How are SE-related datasets collected, preprocessed, and used in LLMs? RQ2 delves into the methodologies behind the assembly, refinement, and application of datasets in the realm of LLMs for SE tasks. It aims to uncover the strategies for dataset collection, the criteria for dataset selection, and the preprocessing steps essential for making the data conducive for LLM training and application. Additionally, this question seeks to explore the types of data that are most prevalent in SE-related LLM research and how these data types influence the modeling and outcomes. RQ3: What techniques are used to optimize and evaluate LLM4SE? RQ3 aims to explore the use of different optimization and evaluation techniques specific to LLMs in the context of SE. This includes an investigation into Parameter Efficient Fine-Tuning (PEFT) methods and various prompting techniques that are tailored to enhance LLM performance on SE tasks. Furthermore, this RQ aims to assess the range of evaluation metrics and methodologies employed to gauge the effectiveness and impact of LLMs in SE, providing insights into how these models are fine-tuned and assessed for their utility and efficiency. RQ4: What SE tasks have been effectively addressed to date using LLM4SE? This RQ aims to identify the SE tasks that have been successfully tackled using LLMs, offering a detailed view of the application spectrum of LLMs in SE. It seeks to identify the specific tasks within SE, such as code generation and program repair, where LLMs have shown significant utility, and to explore the nature and scope of these applications. 2.2 Search Strategy As shown in Fig.1, we employed the “Quasi-Gold Standard” (QGS) approach for paper search. We conducted a manual search to identify a set of relevant studies and extracted a search string from them. This search string was then used to perform an automated search, and subsequently, a snowballing search was employed to further supplement the search results. This approach ensures both search efficiency and maximum coverage, minimizing the risk of omission. Subsequently, we employed a series of relatively strict filtering steps to obtain the most relevant studies. Specifically, we followed five steps to determine the relevance of the studies: (1) Select publication venues for manual search and select digital databases for automated search to ensure coverage of all the selected venues. (2) Establish QGS: Screen all papers for manual search and filter by inclusion/exclusion criteria (defined in Table 3). (3) Subjectively define the search string based on domain knowledge. (4) Conduct an automated search using the search string defined in Step (3). ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:5 Study Identification Automated Search ACM Digital Science Web of IEEE Xplore Springer arXiv DBLP Library Direct Science Large Language Derive search strings Model (LLM) 1,192 papers 10,445 papers 65,290 papers 42,166 papers 85,671 papers 9,966 papers 4,035 papers Complement Evaluate Refine Snowballing Search search Export strings Research Question 1-4 13,565 papers forward backward Identify relevant venues Manual Search Export Export Study selection 3,964 papers 9,601 papers 6 selected SE venues Software 218,765 papers Add 13 papers Total 395 papers Engineering 51 papers Study Selection Filter out studies Check the title, Scan full-text to Remove duplicate Conduct quality with less than 8 abstract, and Identify venue select primary studies assessment pages keywords studies 80,611 papers 4,341 5,078 papers papers 1,172 papers 810 papers 594 papers 382 papers Fig. 1. Study identification and selection process. (5) Conduct snowballing search after performing study selection on the results of manual search and automated search. 2.2.1 Search Items. During the manual search, we selected six of the top SE conferences and journals (i.e., ICSE, ESEC/FSE, ASE, ISSTA, TOSEM, and TSE, as shown in Table 2) and searched for papers that applied LLM4SE. We systematically crawled a list comprising 4,618 published papers from the top venues. Following automated scanning via scripts, we manually verified and identified 51 papers that were relevant to our research objectives. These 51 relevant papers formed the basis for constructing the Quasi-Gold Standard (QGS). Our search string should combine two sets of keywords: one pertaining to SE tasks, and the other related to LLMs. Only if the paper contains both types of keywords, there is a higher probability that it is the paper we need. The complete set of search keywords is as follows: Keywords related to SE tasks: Software Engineering, Software Development, Program*1 , Software Testing, Software Mainten*, SE, Software Lifecycle, Software Design*, Code representation, Code generation, Code comment generation, Code search, Code localization, Code completion, Code summarization, Method name generation, Bug detection, Bug localization, Vulnerability detection, Testing techniques, Test case generation, Program analysis, Bug classification, Defect prediction, Program repair, Code clone detection, Bug report, Software quality evaluation, SATD detection, Code smell detection, Compiled-related, Code review, Software classification, Code classification, Code change, Incident detection, Requirement extraction, Requirement traceability, Requirement validation, Effort cost prediction, Mining GitHub/Github mining, Mining SO (Stack Overflow)/SO mining, Mining app/App mining, Mining tag/Tag mining, Developer-based mining 1 The* symbol serves as a wildcard, representing any characters or character sequence. For example, “Program*” can match “Program”, “Programming”, “Programmer”, and so on. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:6 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang Table 2. Publication venues for manual search. Acronym Venues ASE International Conference on Automated Software Engineering ESEC/FSE Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering ICSE International Conference on Software Engineering ISSTA International Symposium on Software Testing and Analysis TOSEM Transactions on Software Engineering and Methodology TSE Transactions on Software Engineering Table 3. Inclusion criteria and Exclusion criteria. Inclusion criteria 1) The paper claims that an LLM is used. 2) The paper claims that the study involves an SE task. 3) The paper with accessible full text. Exclusion criteria 1) Short papers whose number of pages is less than 8. 2) Duplicate papers or similar studies with different versions from the same authors. 3) Studies belonging to books, thesis, monographs, keynotes, panels, or venues not executing a full peer-review process. 4) Tool demos and editorials. 5) The paper is published in a workshop or a doctoral symposium. 6) The paper is a grey publication, e.g., a technical report or thesis. 7) Non-English written literature. 8) The paper mentions the use of LLMs without describing the employed techniques. 9) The paper leverages SE methods to enhance LLMs, rather than focusing on using LLMs for SE tasks. Keywords related to LLMs: LLM, Large Language Model*, Language Model*, LM, PLM, Pre- trained, Pre-training, Natural Language Processing, NLP, Machine Learning, ML, Deep Learning, DL, Artificial Intelligence, AI, Transformer, BERT, Codex, GPT, T5, Sequence Model*, Attention Model*, Transfer Learning, Neural Network*, ChatGPT, GPT-* It is important to note that the list of keywords related to LLMs that we set up includes Machine Learning, Deep Learning, and other such terms that do not seem to be necessarily related to LLMs. The reason for this is that we want to avoid omitting papers related to our research as much as possible, so the process of performing automated searches expands our search scope. 2.2.2 Search Datasets. After determining the search string, we conducted an automated search across seven widely used databases, which are capable of covering all published or latest papers. Given that the first paper about the Transformer architecture , which forms the basis for LLMs, was published in 2017, we focused our search on papers published from that year onward2. Two authors independently performed the automated search, and the search results from each database were merged and deduplicated. Specifically, we obtained 1,192 papers from IEEE Xplore, 10,445 papers from the ACM Digital Library, 62,290 papers from ScienceDirect, 42,166 papers from Web of Science, 85,671 papers from Springer, 9,966 papers from arXiv, and 4,035 papers from DBLP. 2.3 Study Selection 2.3.1 Study Inclusion and Exclusion Criteria. Based on our search strategy, we initially obtained 218,765 papers that potentially relate to our research. Next, we needed to further evaluate the relevance of these papers based on inclusion and exclusion criteria (To ensure that our inclusion and exclusion criteria were sufficiently objective and rational, we designed these criteria following several state-of-the-art SLR papers [302, 452, 466, 509].), as shown in Table 3, so that the selected papers can directly address our research questions. The paper selection process, as illustrated in 2 The cut-off date for the paper collection process of this version is January 31, 2024. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:7 Table 4. Checklist of Quality Assessment Criteria (QAC) for LLM studies in SE. ID Quality Assessment Criteria QAC1 Is the study relevant to SE tasks? QAC2 Does the study utilize LLMs? QAC3 Is the research not a secondary study, such as an SLR, review, or survey? QAC4 Was the research published in a high-repute venue? QAC5 Is there a clear motivation for the research? QAC6 Does the study provide a clear description of the techniques used? QAC7 Are the experimental setups, including experimental environments and dataset information, described in detail? QAC8 Does the study clearly confirm the experimental findings? QAC9 Are the key contributions and limitations of the study discussed? QAC10 Does the study make a contribution to the academic or industrial community? Fig. 1, consists of six phases. In the first phase, we conducted automated filtering to exclude papers with less than 8 pages [23, 452] (Exclusion criteria 1), reducing the number of papers to 80,611. In the second phase, we examined the titles, abstracts, and keywords of the papers to identify those that include relevant LLM-related keywords. We then expanded the search scope to avoid missing relevant papers, including ML, DL, and other related keywords that may not directly correspond to LLM. The purpose of this phase is to narrow down the scope and filter out papers directly related to LLM (Inclusion criteria 1). Papers that are filtered out in this phase are then manually reviewed in the fifth phase. Additionally, we excluded 448 non-English written literature (Exclusion criteria 7). After the second phase, the number of papers was reduced to 5,078. The third phase involves identifying the venues of the papers (Exclusion criteria 3). We extracted publication information such as “journal”, “URL”, “DOI”, and “series” to determine the publication sources. For papers from arXiv in 2023 and 2024, we chose to retain them, considering that this field is emerging and many works are in the process of submission. Although these papers did not undergo peer review, we have a quality assessment process to eliminate papers with low quality. This step resulted in 1,172 papers. In the fourth phase, we merged and deduplicated the remaining papers from the seven databases and the manually searched paper list (Exclusion criteria 2), resulting in 810 papers. We then reviewed the full texts of the papers and excluded 190 papers that were grey publications or were published in workshops or doctoral symposiums (Exclusion criteria 4, 5, 6). By further assessing the quality of the papers, we identified 382 papers directly relevant to our research. This phase primarily involved excluding papers that mentioned LLMs but did not directly apply them, such as papers that only discussed LLMs in future work or focused on evaluating the performance of LLM-enabled tools (Exclusion criteria 8). For systematic views, survey, and review papers, we have retained them and will assess their content during the quality assessment phase to determine their relevance to our research. 2.3.2 Study Quality Assessment. A well-crafted quality assessment can help to prevent biases introduced by low-quality studies and can indicate to readers where caution about conclusions should be drawn. We formulated ten Quality Assessment Criteria (QAC), as shown in Table 4. These aim to assess the relevance, clarity, validity, and significance of included papers. We used a scoring system of -1, 0, 1 (irrelevant/unmet, partially relevant/met, relevant/fully met). The first three questions were designed for the remaining 382 papers in the fifth stage. If QAC1, QAC2, or QAC3 received a score of -1, there is no need to proceed with QAC4-QAC10, and the paper can be excluded directly. QAC4-QAC10 involved assessing the content of the papers using a scoring system of 0, 1, 2, 3 (poor, fair, good, excellent). Finally, we calculated the total score of QAC4-QAC10 for each paper. For published papers, the maximum score for QAC4-QAC10 should be 21 (3 × 7). ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:8 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang TSE, 14 300 ESEC/FSE, 12 273 ICSE, 41 TOSEM, 11 250 ASE, 10 Number of papers 200 SANER, 10 arXiv, 241 ICSME, 9 150 EMNLP, 7 100 ISSTA, 7 56 46 ICML, 6 50 7 13 ICPC, 5 0 Others, 17 NeurIPS, 5 2020 2021 2022 2023 2024 (a) Distribution of papers across venues. (b) Distribution of papers over years. Fig. 2. Overview of the selected 395 papers’ distribution. Fig. 3. Topics discussed in the collected papers. We retained papers with a score of 16.8 (21 × 0.8) or above. For unpublished papers on arXiv, the score for QAC4 is always 0, and the maximum score for QAC5-QAC10 should be 18 (3 × 6). We retained papers with a score of 14.4 (18 × 0.8) or above. After this quality assessment, we obtained a final set of 382 papers. 2.4 Snowballing Search To identify any additional possibly relevant primary studies, we conducted a snowballing search. Snowballing refers to using the reference list of a paper or the citations to the paper to identify additional papers. Snowballing could benefit from not only looking at the reference lists and citations but also complementing them with a systematic way of looking at where papers are actually referenced and where papers are cited. Using the references and the citations respectively is referred to as backward and forward snowballing. Before conducting snowballing, a set of initial papers needs to be prepared. In this study, the initial paper list consists of the remaining 382 papers after the quality assessment. We performed forward and backward snowballing, which resulted in the collection of 3,964 and 9,610 papers, respectively. After initial deduplication, we were left with 5,152 papers. We then conducted the full study selection process on these 5,152 papers, including deduplicating them with the 382 papers from performing snowballing on the initial list. As a result, we obtained an additional 13 papers. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:9 Table 5. Extracted data items and related research questions (RQs). RQ Data Item 1,2,3,4 The category of SE task 1,2,3,4 The category of LLM 1,4 Characteristics and applicability of LLMs 2 The adopted data handling techniques 3 The adopted weight training algorithms and optimizer 3 The selected evaluation metrics 4 The SE activity to which the SE task belongs 4 The developed strategies and solutions 2.5 Data Extraction and Analysis We finally obtained 395 relevant research papers after searching and snowballing. Fig. 2 presents an overview of the distribution of the included papers. As shown in Fig. 2 (a), 154 papers are published in peer-reviewed venues. ICSE is the most common of these venues, with a contribution of 41 papers. Other venues with noteworthy contributions include TSE, ESEC/FSE, and TOSEM, contributing 14, 12, and 11 papers respectively. Meanwhile, the remaining 241 papers are published on arXiv, an open-access platform that serves as a repository for scholarly articles. This finding is not surprising since much new LLM4SE research is rapidly emerging and thus many works are just completed and are likely in the peer review process. Despite the non-peer-reviewed nature of these papers, we have performed a rigorous quality assessment process on all collected papers, to ensure the quality of validity of our findings. This approach allows us to include all high-quality and relevant publications while maintaining high research standards. Fig. 2 (b) shows the temporal distribution of the included papers. The number of publications has seen a rapidly growing trend since 2020. In 2020 and 2021, there are only 7 and 13 relevant papers, respectively. However, by 2022, the number of papers has increased dramatically to 56. What’s surprising is that, in 2023 alone, the number of published papers has already reached 273. And within just one month in 2024, 46 relevant papers are published. This rapid growth trend demonstrates that there is a growing research interest in the domain of LLM4SE. In order to visualize the main content of our collection of papers, we generated a word cloud based on the abstracts of 395 papers as shown in Fig. 3. The most frequently occurring words include “code”, “LLM”, “language”, “model”, “large”, “task”, “software”,“generation”, “performance”, and “program”, clearly indicating the main themes explored in these papers. The terms “code” and “software” emphasize the core elements of software engineering, while “LLM”, “large”, “language” and “model” denote the use of large language models in a variety of tasks. The terms “generation”, “task”, and “program” emphasize the use of the LLM for automatic code generation and other SE tasks. In addition, “performance” reflects the evaluation and assessment of the effectiveness of LLM in SE applications. The word cloud provides further visual evidence that the literature we have collected is closely related to our research topic. We then conducted data extraction during the full-text review. This extraction phase collected all relevant data that would facilitate a comprehensive and insightful response to the RQs outlined in Section 2.1. As depicted in Table 5, we extracted data including the classification of SE tasks, their corresponding activities, as well as the category, characteristics, and applicability of the LLMs. With this collected data, we systematically analyzed the relevant aspects of LLM4SE. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:10 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang Encoder-only ALBERT (6) BERTOverflow (3) seBERT (2) Features BERT (50) CuBERT (3) GraphCodeBERT (25) CodeRetriever (1) Encoder Input Text RoBERTA (24) CostSens BERT (1) CodeBERT (51) Sentence-BERT (2) Trace BERT (3) PRCBERT (1) Encoder-decoder CoTexT (4) AlphaCode (6) CoditT5 (1) Output Text BART PLBART (15) Codetrans (2) NatGen (2) CodeT5+ (7) Decoder Features Encoder T5 (20) CodeT5 (46) UniXcoder (16) SPT-Code (2) CodeReviewer (2) Input Text PolyCoder (8) InstructGPT (5) CodeParrot (6) CodeGeeX (8) CodeGeeX2 (3) StableLM (1) CodeLlama (19) Decoder-only GPT-2 (17) GPT-Neo (13) CodeGPT (26) CodeGen (44) LaMDA (2) ChatGPT (72) WizardCoder (12) CodeLlama2 (1) Output Text GPT-1 (4) GPT-3 (12) GPT-J (13) GPT-3.5 (54) CodeGen2 (7) GPT-4 (53) LLaMA (14) Llama2 (10) Llama2-Chat (2) Decoder GPT-NeoX (5) PaLM2 (1) Bard (2) Vicuna (11) SantaCoder (5) CodeFuse (1) Input Text XLNet (4) Codex (62) InCoder (29) PaLM (4) PaLM-Coder (3) Claude (3) BLOOM (5) DeepSeek Coder (1) Copilot (7) PyCodeGPT (5) PanGu-Coder (1) OPT (5) StarCoder (25) Claude2 (2) DeepSeek (3) 2018 2019 2020 2021 2022 2023 2024 Fig. 4. Distribution of the LLMs (as well as LLM-based applications) discussed in the collected papers. The numbers in parentheses indicate the count of papers in which each LLM has been utilized. 3 RQ1: WHAT LLMS HAVE BEEN EMPLOYED TO DATE TO SOLVE SE TASKS? 3.1 Large Language Models (LLMs) Pre-trained language models (PLMs) have demonstrated impressive capabilities in solving var- ious NLP tasks [202, 381, 468, 558]. Researchers have observed that scaling up the model sizes significantly enhances their capacity, leading to remarkable performance improvements when the parameter scale surpasses a certain threshold [137, 381, 422]. The term “Large Language Model” (LLM) was introduced to distinguish language models based on their parameter size, specifically referring to large-sized PLMs. However, we note that the literature lacks a formal consen- sus on the minimum parameter scale for LLMs, as the model’s capacity is intertwined with both data size and total compute. In this paper, we adopt the LLM scope division and taxonomy introduced by Pan et al. and categorize the mainstream LLMs investigated in this study into three groups according to their architectures: encoder-only, encoder-decoder, and decoder- only LLMs. This taxonomy and relevant models are shown in Fig. 4. We have included the LLMs used by each work and their parameter sizes (if declared in the paper) in our public repository: https://github.com/xinyi-hou/LLM4SE_SLR. Additionally, Table 6 summarizes the LLMs with dif- ferent architectures suitable for different types of SE tasks. Encoder-only LLMs. Encoder-only LLMs are a type of neural network architecture that utilizes only the encoder component of the model. The encoder’s function is to process and encode the input sentence into a hidden representation, capturing the relationships between words and the overall context of the sentence. Notable instances of encoder-only LLMs include BERT and its variants [92, 118, 211, 260]. As an example, BERT’s structure, based on the Transformer’s ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:11 Table 6. Summary of LLMs with different architectures used in SE tasks. Model Type Example of SE tasks Encoder-only Understanding Code Understanding Bug localization Vulnerability detection Encoder-Decoder Understanding and Generation Code summarization Code translation Program repair Decoder-only Generation Code generation Code completion Test case generation encoder architecture, has been referenced in 50 our selected primary studies. Its distinctive bidi- rectional attention mechanism simultaneously considers the left and right context of each word during training. In the SE domain, other prominent models like CodeBERT , GraphCode- BERT , RoBERTa , and ALBERT have been widely employed. Specialized models such as BERTOverflow and CodeRetriever have been specifically developed for SE applications. These models differ from BERT by leveraging program structure, introducing new pre-training tasks, or engaging new modalities, thereby improving the architecture’s application to code-related tasks. For example, CodeBERT integrates a token prediction scheme to comprehend code by predicting subsequent tokens, enhancing its understanding of programming languages for tasks like code completion and bug detection. GraphCodeBERT introduces edge-type predic- tion, recognizing relationships between code elements as a graph. This enables GraphCoderBERT to leverage code structure, improving its effectiveness in tasks like code summarization and program analysis. Encoder-only LLMs have shown efficacy in tasks requiring a nuanced understanding of the entire sentence or code snippet. Examples include code review, bug report understanding, and named entity recognition pertaining to code entities [19, 231, 297, 344, 380, 502]. Encoder-decoder LLMs. Encoder-decoder LLMs incorporate both encoder and decoder mod- ules. The encoder ingests the input sentence and encodes it into a hidden space, effectively capturing the underlying structure and semantics. This hidden representation serves as an interme- diary language, bridging the gap between diverse input and output formats. Conversely, the decoder utilizes this hidden space to generate the target output text, translating the abstract representation into concrete and contextually relevant expressions. Models such as PLBART , T5 , and CodeT5 embodies this architecture. Further advancements are evident in CodeT5+ , while AlphaCode and CoTexT showcase the architecture’s adaptability to various SE tasks. The encoder-decoder design offers flexible training strategies and is proficient in handling multifaceted tasks such as summarization, translation, and question-answering. Within the field of SE, this ability has been successfully applied to tasks like code summarization [9, 115, 287]. The encoder module’s capacity to understand and represent both the structure and semantics of code is pivotal, allowing the decoder to translate this comprehension into concise, human-readable summaries. Decoder-only LLMs. Decoder-only LLMs exclusively utilize the decoder module to generate the target output text, following a distinct training paradigm that emphasizes sequential prediction. Unlike the encoder-decoder architecture, where the encoder processes input text, the decoder- only architecture begins with an initial state and predicts subsequent tokens, gradually building the output text. This approach relies heavily on the model’s ability to understand and anticipate ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:12 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang language structure, syntax, and context. GPT-series models, such as GPT-1 , GPT-2 , GPT- 3 , GPT-3.5 , GPT-4 , as well as their notable derivative, ChatGPT 3 , represent their major implementations. More specialized versions like CodeGPT , InstructGPT , Codex , Copilot 4 , and others have been fine-tuned for specific tasks in SE. Open-source models like GPT-J , GPT-Neo , GPT-NeoX , LLaMA , and Vicuna also follow this architecture. Decoder-only LLMs are usually more suitable for various generation tasks, such as code generation and code completion. These models can generally perform downstream tasks from a few examples or simple instructions without adding prediction heads or fine-tuning, making them valuable tools in SE research. 2022 marked a surge in the development of decoder-only LLMs, a trend that gained further momentum in 2023, notably with the launch of commercial products by leading Internet companies. For example, Google launched Gemini , Meta introduced LLaMA and Llama 2 , and Anthropic unveiled Claude , etc. Contrary to LLMs such as GPT-4 and its derivative application, ChatGPT, released by OpenAI, which were promptly integrated into SE tasks, these new additions have not yet found widespread application within the SE field. Their potential remains largely unexplored, with opportunities for further assessment and utilization in specific tasks and challenges. The continued advancement of these models emphasizes the active exploration and innovation within decoder-only architectures. 19 94 2024 Encoder-only 52 2023 8 2022 9 24 2021 85 2020 Encoder-decoder 17 2 0 77 432 Decoder-only 73 9 2 0 40 80 120 160 200 240 280 320 360 400 440 480 Number of instances utilizing an LLM in the collected papers Fig. 5. Trends in the application of LLMs with different architectures in SE tasks over time. 3.2 Trend Analysis As shown in Fig. 5, in the span from 2020 to 2024, the architecture of LLMs has witnessed notable shifts in preference and application within SE tasks. The specific choices between decoder-only, encoder-decoder, and encoder-only structures have shaped the direction of research and solutions in the SE domain. This analysis explores trends in the adoption of these architectures over the years, reflecting the evolving dynamics of LLM for SE tasks. Evolution of LLM architectures in 2021. The year 2020 saw research papers predominantly concentrating on encoder-only LLMs for SE tasks, evidenced by a total of eight papers. Decoder- only LLMs or encoder-decoder LLMs were scarcely featured in that year’s research. A marked change occurred in 2021. Out of 19 papers in 2021, nine were dedicated to decoder-only LLMs, 3 ChatGPT is a conversational agent built upon the GPT architecture, with GPT-3.5 and GPT-4 being specific versions of the architecture, each representing successive advancements. 4 Copilot is an application built upon LLMs tailored for coding tasks. For convenience, all subsequent references in this paper to LLMs and their applications, such as ChatGPT and Copilot, will collectively be referred to as LLMs. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:13 constituting 47.37% of the research. Additionally, two papers, or 10.53%, focused on encoder-decoder LLMs. Encoder-only LLMs witnessed a slight decline, representing 42.1% of the field with eight papers. This rapid transition can be linked to the generative capability of decoder-only LLMs. Researchers [212, 369, 400] found that these models, e.g., GPT series, requiring minimal fine-tuning, could produce not only syntactically correct but also functionally relevant code snippets. Their proficiency in grasping the context of code quickly made them a preferred choice. Diversity of LLM architectures in 2022. 2022 experienced a significant increase in diversity, with more varied LLM architectures finding representation. Out of a total of 142 papers, 73 were centered around decoder-only LLMs, comprising 51.41% of the studies. Encoder-decoder LLMs made their presence known in 17 papers, accounting for 11.97%. Meanwhile, encoder-only LLMs led the field slightly with 52 papers, capturing 36.62% of the research interest. This diverse distri- bution suggests an exploration phase where researchers were actively assessing and leveraging different architectures to suit varied needs and challenges. The near-equal interest across different architectures underscores the field’s richness, indicating that no single approach had become the definitive choice. Dominance of the decoder-only architecture in 2023. 2023 signaled a strong shift towards decoder-only LLMs. An impressive 432 instances of utilizing decoder-only LLMs were recorded across 195 unique papers, reflecting that a single paper might employ multiple such models. These papers focusing on decoder-only LLMs constituted a significant 70.7% of the total research this year. In comparison, encoder-decoder LLMs were the subject of 85 papers, contributing 13.91%, while encoder-only LLMs appeared to stabilize, with 94 papers, representing 15.39% of the 2023 research landscape. This trend signifies a shift in focus and resources toward exploring and harnessing the decoder-only architecture as the primary approach in many current and future LLM4SE research and applications. Exploration of the LLM architecture in 2024. The initial trends in January 2024 showcase the ongoing evolution of LLM architectures. Among the 120 papers examined, decoder-only LLMs con- tinued to maintain a prominent position, with 77 papers dedicated to this architecture, constituting 64.17% of the research. Encoder-decoder LLMs appeared in 24 papers, representing 20% of the total, while encoder-only LLMs were featured in 19 papers, making up 15.83%. Although there is a slight decrease in the dominance of decoder-only architectures compared to the previous year, they still hold a central role. The persistent exploration of encoder-decoder and encoder-only architectures suggests an enduring interest in diverse configurations within the SE research community. Criteria for LLM selection in SE tasks. The selection of an LLM for SE tasks should involve careful consideration rather than arbitrary choice. Key factors guiding this selection encompass the model’s proficiency in understanding the context of code, its ability to generate relevant content, responsiveness to fine-tuning, and demonstrated performance on SE-specific benchmarks [224, 238, 491]. Given the stringent syntactical rules and functional requirements inherent to SE tasks, models capable of seamlessly integrating these complex aspects were typically favored. Task-specific fine-tuning. A notable trend is the customization of LLMs for precise SE tasks [160, 231, 535]. By fine-tuning models with datasets tailored to specific functions such as bug detection or code review, researchers were able to achieve marked performance improvements [55, 204]. In conclusion, the evolution of LLMs for SE, transitioning from encoder-only to decoder-only architectures, highlights the field’s vibrancy and adaptability. This shift has fundamentally altered the approach to SE tasks, reflecting the ongoing innovation within the discipline. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:14 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang RQ1 - Summary (1) There are more than 70 different LLMs used for SE tasks in our selected primary studies. Based on the underlying architecture or principles of different LLMs, we classified the summarized LLMs into three categories, i.e., decoder-only, encoder-decoder, and encoder-only LLMs. (2) We observed that each LLM architecture serves a specific purpose in SE tasks, with encoder- only LLMs focusing on comprehensive understanding, encoder-decoder LLMs used for tasks requiring understanding of input information followed by content generation, and decoder-only LLMs being more suitable for generation tasks. (3) We analyzed the trend of LLM usage for SE tasks. The most widely used LLMs are with decoder-only architectures. There are over 45 LLMs in the decoder-only category and 195 papers have researched the application of decoder-only LLMs to SE tasks. 4 RQ2: HOW ARE SE-RELATED DATASETS COLLECTED, PREPROCESSED, AND USED IN LLMS? Data plays a crucial role in the model training phase. First, data is collected to obtain diversity and richness to ensure that the model can cope with different scenarios and situations. Second, data is classified to clarify the training objectives of the model and avoid confusion and misinformation. The preprocessing of data is indispensable to clean and transform the data to improve its quality. Finally, data is formatted into a structure suitable for model processing, allowing the LLM to learn the data’s features and patterns effectively. We analyze the reported processes of data collection, data classification, data preprocessing, and data representation in our selected primary studies on LLM4SE. 4.1 How are the datasets for training LLMs sourced? Data is an indispensable and critical factor in training LLMs, which determines the generalization ability, effectiveness, and performance of the models. Adequate, high-quality, and diverse data is critical to allow models to fully learn features and patterns, optimize parameters, and ensure reliability in validation and testing. We first investigate the methods used to obtain the dataset. By analyzing the methods of data collection, we divided the data sources into four categories: open-source datasets, collected datasets, constructed datasets, and industrial datasets. Open-source datasets [38, 189, 449, 528] refer to publicly accessible collections of data that are often disseminated through open-source platforms or repositories. For example, datasets like HumanEval , which consists of 164 manually crafted Python problems, each accompanied by its respective unit tests. The open-source nature of these datasets ensures their credibility and allows for community-driven updates, making them a reliable resource for academic research. Collected datasets [149, 285, 380, 427] are those that researchers compile directly from a multitude of sources, including but not limited to, major websites, forums, blogs, and social media platforms. For instance, researchers [35, 373, 473, 502] often scrape data from Stack Overflow threads or GitHub issues comments to create a dataset tailored to their specific research questions. Constructed datasets [83, 185, 201, 532] are specialized datasets that researchers create by modifying or augmenting collected datasets to better align with their specific research objectives. These modifications can be carried out through manual or semi-automatic methods and may include the generation of domain-specific test sets, annotated datasets, or synthetic data. For example, researchers often take a collected dataset of code snippets and manually annotate them with bug types to create a constructed dataset for studying automated program repair techniques [88, 173, 483]. Industrial datasets [11, 290, 462] are those obtained from commercial or industrial entities and often contain proprietary business ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:15 data, user behavior logs, and other sensitive information. These datasets are particularly valuable for research that aims to address real-world business scenarios. However, the acquisition of such datasets is often complicated by issues related to business confidentiality and data privacy. For example, in a collaborative effort with China Merchants Bank (CMB), Wang et al. were able to access 21 projects from CMB’s repositories. Access to such data would likely require non-disclosure agreements and other legal safeguards to protect business interests. Each of these dataset types offers unique advantages and challenges, and the choice between them should be guided by the specific requirements and constraints of the research project at hand. 300 235 240 Number of papers 180 120 84 60 49 6 0 Open-source Collected Constructed Industrial datasets datasets datasets datasets Fig. 6. The collection strategies of LLM-related datasets. Fig. 6 shows the collection strategies of LLM-related datasets. As can be seen from the data in the figure, 235 studies used open-source datasets for training LLMs. One of the main reasons for using open-source datasets in LLM training is their authenticity and credibility. Open-source datasets usually contain real-world data collected from various sources (such as relevant studies that have been conducted), which makes them highly reliable and representative of real-world scenarios. This helps LLMs learn from real examples to better understand real-world applications and improve their performance. Second, since LLMs are a topic that has just recently emerged, a lack of suitable training sets does exist. Therefore, researchers often collect data from sites such as Stack Overflow and GitHub and build datasets to make the data more composite for SE tasks. Among the 395 papers we studied, we discovered that merely six studies utilized industrial datasets. This suggests a potential misalignment between the properties of datasets used in academic research and those encountered in real-world industrial contexts. This divergence underscores the need for future research to investigate industrial datasets, thereby ensuring that LLMs are applicable and robust across both academic and industrial scenarios. Note that some papers use multiple datasets that span different categories, e.g., Xu et al. evaluated the performance of Codex, GPT-J, GPT-Neo, and other LLMs on SE tasks, and Mastropaolo et al. investigated the use of T5 in several code-related tasks such as fixing bugs and generating code comments. For different LLMs or different SE tasks, researchers may use different training datasets. On the other hand, some papers focus on exploring how existing LLMs (e.g., ChatGPT) are used in SE tasks and do not specify the dataset used for model training, as these LLMs like ChatGPT often do not require users to prepare training data themselves for general usage scenarios. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:16 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang Table 7. Data types of datasets involved in prior studies. Category Data type Total Text-based Programming tasks/problems (42) Prompts (33) 151 datasets SO (i.e. Stack Overflow) posts (12) Bug reports (11) Requirements documentation (9) APIs/API documentation (8) Q&A pairs (6) Vulnerability descriptions (4) Reviews (4) Logs (3) Methods (3) Project issues (3) Code comments (2) Theorems (2) Buggy text (1) Dockerfiles (1) Outage descriptions (1) Semantic merge conflicts (1) Site text (1) Software development tasks (1) User intents (1) Software specifications (1) User reviews (1) Code-based Source code (60) Bugs/Buggy code (16) 103 datasets Vulnerable source code (8) Patches (4) Code changes (3) Test suites/cases (3) Bug-fix pairs (2) Error code (2) Error-fix pairs (1) Flaky test cases (1) Identifiers (1) Labeled clone pairs (1) Packages (1) Graph-based GUI Images (1) 1 datasets Software Code repository (9) Android apps (3) 20 repository Issues and commits (3) Pull-requests (2) -based datasets Industrial projects (1) Open-source projects (1) Web applications (1) Combined Programming tasks and test suites/cases (17) Source code and comments (12) 55 datasets Programming tasks and solutions (8) Source code and description (3) Code-text pairs (2) Souce code and API usage sequences (2) Source code and test suites/cases (2) Bug report and test suites/cases (1) Buggy code and comments (1) Buggy code and solutions (1) Code files and summaries (1) Binary code and related annotations (1) Failing test code and error messages (1) Source code and Q&A pairs (1) Source code, methods, and logs (1) Vulnerable code and description (1) *See Appendix A for the full table including references. 4.2 What types of SE datasets have been used in existing LLM4SE studies? Data types play a pivotal role in shaping the architecture and selection of LLMs, as they directly influence the extraction of implicit features and subsequent model decisions[35, 106, 390, 504]. The choice of data types can significantly impact the overall performance and generalization ability of the LLMs. We examine and classify the types of SE datasets employed in LLM4SE studies. By investigating the relationship between data types, model architectures, and performance, we seek to shed light on the critical role of data types in the success of LLM4SE applications. Data type categorization. We classified the data types of all datasets into five categories: code- based, text-based, graph-based, software repository-based, and combined data types. Table 7 de- scribes the specific data included in the data types corresponding to the datasets we summarized from the 395 studies. We can find that most of the studies used text-based datasets, accounting for a total of 151. The dominance of text-based datasets in training LLMs for SE tasks highlights the models’ exceptional natural language processing capabilities. These LLMs excel in understanding and processing textual data, making them an ideal choice for tasks that involve code comprehension, bug fixing, code generation, and other text-oriented SE challenges. Their ability to process and ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:17 learn from vast amounts of text data enables them to provide powerful insights and solutions for various SE applications. The most prevalent type of data utilized in training LLMs for SE tasks is programming tasks/problems with 42 instances observed among the surveyed papers. This dominance can be attributed to the diverse and challenging nature of programming problems, which provide LLMs with opportunities to generalize knowledge and skills across various SE challenges, fostering a robust understanding of software concepts and enhancing performance across a wide range of tasks, including code generation, code completion, and code summarization, etc. Prompts follow closely behind programming tasks, with 33 instances observed in the surveyed papers, providing task-specific guidance to LLMs, serving as cues or instructions for the models, and helping them understand the context and requirements of SE tasks. This combination helps the models develop a robust understanding of software concepts and perform well in a wide range of tasks. There are also SO (i.e., Stack Overflow) posts (12), bug reports (11), etc., which are among the more numerous data types in text-based datasets. The predominance of source code (60) as the most abundant data type in code-based datasets can be attributed to its fundamental role in SE. Source code serves as the foundation of any software project, containing the logic and instructions that define the program’s behavior. Therefore, having a large volume of source code data is crucial for training LLMs to understand the intricacies of software development, enabling them to effectively generate, analyze, and comprehend code in various SE tasks. There are also common data types, such as bugs/buggy code (16) and patches (4), for program repair tasks. Additionally, vulnerable source code (8) is used for vulnerability detection tasks. Graph-based datasets are used in some research studies for SE tasks, e.g., Kolthoff et al. used a dataset composed of screenshots from Google Play Android applications to construct a graphical user interface (GUI) repository in their study on LLM for the rapid prototyping task. These datasets represent code using graph structures, capturing relationships and dependencies between code components. Software repository-based datasets are compilations of data extracted from version control systems, such as Git repositories, containing code, documentation, and related artifacts. This data includes Code repository (3), issues and commits (3), and so on. The data in software repositories can provide a wealth of information covering all aspects of the software development process, including code evolution history, records of issue fixes and feature improvements, code quality assessments, and so on. These data are valuable for studying behaviors and trends in the software development process, improving software quality and development efficiency, and evaluating the performance of software engineering techniques. Therefore, many studies have used software repository-based datasets for empirical analysis and model training. Some studies employed combined datasets containing multiple datatypes. Among them, the most common type is “programming tasks and test suites/cases”. Other combinations of data types include “source code and comments”, “programming tasks and solutions”, “source code and description ”, “code-text pairs”, etc. 4.3 How do data types influence the selection of data-preprocessing techniques? For the training and application of LLMs, the raw dataset needs to be subjected to data processing to obtain a clean and suitable dataset for model training. The data processing steps [216, 279] involve operations such as data cleaning, noise removal, normalization, etc. To ensure consistency and quality of the data, different data types may require different processing methods to improve the performance and effectiveness of LLMs in SE tasks. In this section, we aim to detail the data preprocessing procedures for the two most used types of datasets, i.e., text-based datasets and code-based datasets. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:18 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang Duplicated Data Initial data Unqualified Text Data Data instance extraction segmentation data deletion preprocessing tokenization segmentation deletion Fig. 7. The data preprocessing procedure for text-based datasets. Unqualified Duplicated Uncompilable Data Data Code Data data instance data extraction compilation representation segmentation deletion deletion deletion Fig. 8. The data preprocessing procedure for code-based datasets. The data preprocessing procedure for text-based datasets. As displayed in Fig. 7, the steps of text-based dataset preprocessing consist of seven steps in total, yet there are some differences from the code-based dataset preprocessing steps. The process begins with data extraction [54, 55, 83, 504], where relevant text is carefully extracted from SE documentation from a variety of sources, including bug reports , requirements documents , code comments , and API documentation. This step ensures that the dataset captures diverse, task-specific textual information. After data extraction, the text is initially segmented and categorized according to the specific requirements of the research task. For example, the text can be segmented into sentences or further broken down into individual words as needed for analysis [129, 204]. To ensure the quality and relevance of the dataset, substandard data deletion is performed to eliminate any invalid or irrelevant text. For example, the dataset used by Lee et al. was constructed from bug reports, and in the “unqualified data deletion” process the researchers filtered out bug reports with fewer than 15 words because the text was too short to contain contextual information. Next, preprocessing operations are performed on the text to standardize and clean it. Common preprocessing steps include removing certain symbols, stop words, and special characters [351, 462]. This standardized form of text facilitates the efficient processing of LLMs. To avoid introducing bias and redundancy in the dataset, researchers eliminated duplicate instances by removing any duplicate text samples [129, 204, 493]. This step enhances the diversity of the dataset and helps the model to generalize better to new inputs. “Data tokenization” is a key step in preparing the text for LLMs. Text is labeled into smaller units, such as words or subwords, so that LLMs are easier to manage and process efficiently. Finally, the preprocessed dataset is partitioned into different subsets, usually including a training set, a validation set, and a test set. The data preprocessing procedure for code-based datasets. We now summarize the process of preprocessing a code-based dataset, which consists of seven steps. Fig. 8 describes the individual data processing steps in detail and gives examples. The first step is data extraction, which involves retrieving relevant code segments from different sources such as software repositories or version control systems [183, 504]. Depending on the requirements of the research task [287, 522], code segments can be extracted at different levels of granularity, ranging from individual methods and functions to entire source code files or even complete software projects. The next step is to remove any code segments that do not meet predefined criteria or quality standards [223, 343, 390]. This filtering process ensures that the extracted code is relevant to the specific SE task under study, thus eliminating incomplete or irrelevant code snippets. To avoid introducing bias and redundancy during model training, the third step involves removing duplicate instances [56, 493, 560]. Any duplicate code instances are identified and removed from the dataset, thus increasing the diversity and uniqueness of the data. After the data extraction and filtering steps, the fourth step, data compilation, comes into play. The extracted and filtered code segments are merged and compiled ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:19 into a unified code dataset. This compilation process simplifies data storage and access and facilitates subsequent analysis and model training [35, 283]. In the fifth step, the problem of invalid or non-executable code is solved by removing data that cannot be compiled. Any code segments that cannot be compiled or executed are removed from the dataset to ensure that the remaining code instances are valid and usable during model training and evaluation. The sixth step is code representation, which consists of converting the code segments into a suitable representation that can be processed by the LLMs. This conversion can take different forms: token-based representation involves tokenizing the source or binary code into distinct tokens; tree-based representation parses the code into Abstract Syntax Trees (AST); and graph-based representation generates a Program Dependence Graph (PDG), encompassing Control Flow Graphs (CFG) and Call Graphs (CG). Finally, in the “data segmentation” step, the preprocessed dataset is partitioned into different subsets for training, validation, and testing [56, 473]. The training set is used to train the LLM, the validation set helps to tune the hyperparameters and optimize the model performance, and the testing set evaluates the model’s ability on unseen data. By strictly adhering to these seven preprocessing steps, researchers can create structured and standardized code-based datasets, thus facilitating the effective application of LLMs for a variety of SE tasks such as code completion, error detection, and code summarization. It is worth emphasizing that the order of these steps is not fixed and can be adjusted based on the specific research task and its associated requirements. Researchers need to carefully consider the objectives, characteristics of the dataset, and the desired outcomes when determining the optimal sequence for these preprocessing techniques. 4.4 What input formats are the datasets for LLM training converted to? Once suitable datasets have been carefully chosen and clean data has been achieved through the preprocessing steps, the next critical aspect is the transformation of the data into appropriate formats that can effectively serve as inputs for LLMs. Table 8 shows four distinct data input types that emerged during the research: Token-based input, Tree/Graph-based input, Pixel-based input, and Hybrid-based input. We now detail each as follows: Table 8. The various input forms of LLMs proposed in prior studies. See Appendix B for the full table including references. Category Input forms Total Token-based input Text in tokens (150) Code in tokens (118) 347 Code and text in tokens (78) Tree/Graph-based input Code in tree structure (2) Code in graph structure (3) 5 Pixel-based input Pixel (1) 1 Hybrid-based input Hybrid input forms (2) 2 Token-based input. Token-based input [7, 9, 19] involves representing code and text as sequences of tokens, which are smaller units like words or subwords. Text in tokens refers to the tokenization of textual data, such as documentation, bug reports, or requirements, enabling the LLMs to process and analyze natural language descriptions effectively. Code and text in tokens combine both code and its associated textual context, allowing the model to capture the relationships between code elements and their descriptions. Code in tokens refers to the representation of code snippets broken down into meaningful tokens, allowing the LLMs to understand programming language syntax and semantics at a fine-grained level. Tree/Graph-based input. Tree-based input [275, 315, 555] represents code as hierarchical tree structures, capturing the syntactic relationships between code elements. Each node in the tree ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:20 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang represents a code element, and the edges represent the hierarchical nesting of control flow state- ments and other code structures. This form of input allows the LLMs to understand the code’s hierarchical structure and perform tasks like code completion and bug fixing. Graph-based input represents code as a graph structure, where nodes represent code elements and edges represent the relationships between them. Unlike trees, graphs allow more flexible and complex relationships between code elements, enabling the model to capture non-linear dependencies in the code. This form of input is used in tasks like code summarization and vulnerability detection by considering the code’s intricate relationships. Pixel-based input. Pixel-based input visualizes code as images, where each pixel represents a code element or token. This visual representation allows the LLMs to process and understand code through image-based learning. In this input form, LLMs learn from the visual patterns and structures in the code to perform tasks like code translation or generating code visualizations. Hybrid-based input. Hybrid-based input combines multiple modalities to provide LLMs with diverse perspectives for better code comprehension. For example, a hybrid input may combine code in tokens with visual representations of code, allowing the model to learn both from the fine- grained details in the tokenized code and from the overall visual structure of the code. This approach enhances the model’s ability to understand complex code patterns and improve performance in tasks such as code comprehension and code generation. During our investigation of LLM-based models for SE tasks, we observed distinct trends in the usage of different input forms during the training process. Token-based input forms, namely code in tokens and text in tokens were the most prevalent, collectively constituting approximately 97.75% of the studies5. Specifically, code in tokens was widely adopted in 118 studies, accounting for approximately 33.24% of the total studies, demonstrating its popularity as a primary choice for representing code snippets. This approach allowed LLMs to grasp programming language syntax and semantics effectively, making it suitable for a wide range of code-related tasks. Similarly, text in tokens was utilized in 150 studies, comprising around 42.25% of the total studies. This input form allowed LLMs to process natural language descriptions, bug reports, and documentation with greater efficiency and accuracy. The popularity of token-based input forms underscores their significance in leveraging the power of LLMs for software engineering applications. In contrast, tree/graph-based input forms, such as code in tree-structure, were used in only seven studies, making up approximately 1.4% of the total. Although less prevalent, this input type emerged as a promising choice to represent the hierarchical structure and syntactic rela- tionships within code. Its adoption indicated an ongoing exploration of tree-based representations in specialized tasks, such as code completion and bug fixing. Pixel-based input and hybrid-based input were relatively less common, each found in one study, contributing approximately 0.28% of the total studies each. While their adoption rates were lower, these input forms presented intriguing possibilities for specific applications. Pixel-based input offered a unique visual representation of code, potentially advantageous for code translation tasks. Meanwhile, hybrid-based input, combining multiple modalities (e.g., code in tree structure and text in tokens in Niu et al.’s work ), showcased the potential for enhancing code comprehension tasks by offering diverse perspectives for the models to learn from. In summary, the trends in input form usage reveal a strong preference for token-based input, demonstrating its versatility and effectiveness in various SE tasks. However, ongoing exploration of other input forms, such as tree/graph-based, pixel-based, and hybrid-based, suggests a dynamic and evolving landscape in the application of LLMs for SE, with potential for further innovation and 5 This refers to studies that explicitly state input forms of LLMs, i.e., a total of 355 papers as shown in Table 8. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:21 improvement in specialized domains. Each of these input forms caters to specific characteristics of the SE tasks being addressed, enabling LLMs to perform effectively across a wide range of code-related applications with a more comprehensive understanding of the input data. RQ2 - Summary (1) We divided the datasets into four categories based on the source of data: open-source, collected, constructed, and industrial datasets. The use of open-source datasets is the most prevalent, constituting approximately 62.83% of the 374 papers that explicitly state the dataset. (2) We categorized the data types within all datasets into five groups: code-based, text-based, graph-based, software repository-based, and combined. Text-based and code-based types are the most frequently used in applying LLMs to SE tasks. This pattern indicates that LLMs are particularly adept at handling text and code-based data in SE tasks, leveraging their natural language processing capabilities. (3) We summarized the data preprocessing procedures for different data types and found several common preprocessing procedures, i.e., data extraction, unqualified data deletion, duplicated instance deletion, and data segmentation. 5 RQ3: WHAT TECHNIQUES ARE USED TO OPTIMIZE AND EVALUATE LLM4SE? 5.1 What tuning techniques are used to enhance the performance of LLMs in SE tasks? Through surveying research related to LLM4SE, we found that while many general-purpose LLMs (e.g., ChatGPT) can be directly applied to software engineering tasks such as code generation [73, 248, 514], code summarization [388, 408, 507], and program repair [37, 102, 489] without fine- tuning, the hidden potential of LLMs often needs to be realized through tuning to be fully exploited. Specifically, this requires training LLMs with task-specific data to learn knowledge relevant to the task context to perform better. We observed that out of 83 studies, LLMs were fine-tuned using full fine-tuning techniques to adapt to downstream SE tasks, with the majority being BERT series models [56, 83, 90, 165, 204, 216, 246, 271, 373, 438, 463, 469, 532]. The cost of training these LLMs is expensive, requiring a large amount of computational resources and massive amounts of data. It is also costly to train and deploy the fine-tuned models separately for each downstream task, as the traditional fine-tuning approach would need to copy a model and perform full-parameter fine-tuning for each downstream task [34, 62, 83, 160, 165, 216]. To reduce this computational burden, some researchers have previously used In-Context Learn- ing (ICL) [102, 104, 142, 150, 170], which feeds the model with manually designed “prompts” that are overly reliant on human design and do not require updating model parameters at all. However, ICL only operates at the time of inference and does not involve learning task-specific parameters, which experimentally proved to give the model limited improvement in downstream tasks. To address this problem, researchers have begun to apply Parameter Efficient Fine-Tuning (PEFT) techniques to LLMs. PEFT aims to improve the performance of pre-trained models on new tasks by optimizing the subset of parameters fine-tuned, thereby reducing the overall com- putational complexity. This approach maintains the majority of the pre-trained model’s parameters in a fixed state, focusing fine-tuning efforts on a minimal yet impactful set of parameters. Prior code intelligence research has demonstrated the capabilities of PEFT techniques, frequently revealing their superiority over full fine-tuning on a variety of tasks. Four common techniques of PEFT include Low-Rank Adaptation (LoRA) , prompt tuning , prefix tuning , and adapter tuning. We now elaborate on each as follows: Low-Rank Adaptation (LoRA). LoRA injects low-rank trainable matrices into the attention layers of the Transformer architecture to significantly reduce the number of trainable parameters. ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. 1:22 X Hou, Y Zhao, Y Liu, Z Yang, K Wang, L Li, X Luo, D Lo, J Grundy, and H Wang We observed that eight studies [19, 267, 324, 386, 388, 397, 461, 538] utilized LoRA to enhance the performance of LLMs in SE tasks. For instance, Pan et al. trained SteloCoder, specifically designed for translating multiple programming languages into Python code, which is based on the StarCoder LLM. LoRA technology was employed during the modification of the StarCoder model architecture to adjust the parameter count. Additionally, Silva et al. applied LoRA to LLaMA, resulting in a highly effective “program repair adapter” for fixing bugs through fine-tuning. Prompt tuning. Prompt tuning involves appending learnable tokens to the model’s input, guiding it towards better task performance. This method keeps the model’s architecture unchanged, leveraging adaptable prompts to influence outputs without altering internal parameters. In the surveyed papers, three research works [267, 461, 570] utilized prompt tuning. For instance, Zhu et al. proposed a method named AUMENA, which automates method naming tasks through context-aware prompt tuning. Prefix tuning. Prefix tuning adapts pre-trained language models by adding trainable tokens not just to the input but also across internal layers, affecting the model’s intermediate representations. This approach modifies the model’s processing with minimal changes to its original parameters, allowing for task-specific customization. This technique was utilized in the following two studies: Lu et al. fine-tuned LLaMA-Reviewer for automating code review, while Wang et al. fine-tuned CodeT5+ for multiple downstream tasks such as code completion, code generation, and code search. Adapter tuning. Adapter tuning adds small neural network modules to the original model, then fine- tuning them on specific tasks without altering the original model’s parameters. Agarwal et al. fine-tuned LLMs using adapter tuning techniques to make them suitable for code representation tasks. Wang et al. indicated that LLMs refined through adapter tuning perform exceptionally well in code search and code summarization tasks. In addition to the above-mentioned tuning methods, other techniques have been used for tuning LLMs in the LLM4SE domain, such as Reinforcement Learning (RL) [156, 157, 161, 403, 504], Supervised Fine Tuning (SFT) [71, 157, 285, 403, 504], an unsupervised data augmentation method called syntax fine-tuning , knowledge preservation fine-tuning , and task-oriented fine-tuning , etc. 5.2 What prompt engineering techniques are applied to improve the performance of LLMs in SE tasks? Prompt engineering is a method of enhancing model performance by using task-specific instruc- tions, known as prompts, without modifying the core model parameters. This approach enables LLMs to seamlessly integrate into downstream tasks solely based on the given prompts, guiding model behavior without the need to update model parameters. Fig. 9 presents eight prompt engineering techniques currently applied in the LLM4SE domain. Few-shot prompting. Few-shot prompting involves providing a limited number of examples or instructions to the model to perform a specific task. The model learns from these examples and generalizes to similar tasks with minimal training data. In the surveyed LLM4SE research, 88 studies utilized few-shot prompting [7, 91, 95, 104, 185, 469, 494, 550]. For instance, Geng et al. adopted an in-context learning paradigm and providing a specific number of prompts simultaneously significantly outperformed state-of-the-art supervised learning methods in generating comments with multiple intents. Zero-shot prompting. In zero-shot prompting , the model is expected to perform a task without any explicit training on that task. Instead, it relies on the prompt provided during inference to generate the desired output. Following few-shot prompting in terms of usage frequency, 79 studies ACM Trans. Softw. Eng. Methodol., Vol. X, No. Y, Article 1. Publication date: December 2024. Large Language Models for Software Engineering: A Systematic Literature Review 1:23 Few-shot prompting 88 Zero-shot prompting 79 Prompt engineering Chain-of-Thought 18 Automatic Prompt Engineer 2 Chain of Code 2 Automatic Chain-of-Thought 1