Large Language Models for Code (LLMs4Code) have been found to exhibit outstanding performance in the software engineering domain, especially the remarkable performance in coding tasks. However, even the most advanced LLMs4Code can inevitably contain incorrect or outdated code knowledge. Due to the high cost of training LLMs4Code, it is impractical to re-train the models for fixing these problematic code knowledge. Model editing is a new technical field for effectively and efficiently correcting erroneous knowledge in LLMs, where various model editing techniques and benchmarks have been proposed recently. Despite that, a comprehensive study that thoroughly compares and analyzes the performance of the state-of-the-art model editing techniques for adapting the knowledge within LLMs4Code across various code-related tasks is notably absent. To bridge this gap, we perform the first systematic study on applying state-of-the-art model editing approaches to repair the inaccuracy of LLMs4Code. To that end, we introduce a benchmark named CLMEEval, which consists of two datasets, i.e., CoNaLa-Edit (CNLE) with 21K+ code generation samples and CodeSearchNet-Edit (CSNE) with 16K+ code summarization samples. With the help of CLMEEval, we evaluate six advanced model editing techniques on three LLMs4Code: CodeLlama (7B), CodeQwen1.5 (7B), and Stable-Code (3B). Our findings include that the external memorization-based GRACE approach achieves the best knowledge editing effectiveness and specificity (the editing does not influence untargeted knowledge), while generalization (whether the editing can generalize to other semantically-identical inputs) is a universal challenge for existing techniques. Furthermore, building on in-depth case analysis, we introduce an enhanced version of GRACE called A-GRACE, which incorporates contrastive learning to better capture the semantics of the inputs.
The increasing demand for spatiotemporal data and modeling tasks in geosciences has made geospatial code generation technology a critical factor in enhancing productivity. Although large language models (LLMs) have demonstrated potential in code generation tasks, they often encounter issues such as refusal to code or hallucination in geospatial code generation due to a lack of domain-specific knowledge and code corpora. To address these challenges, this paper presents and open-sources the GeoCode-PT and GeoCode-SFT corpora, along with the GeoCode-Eval evaluation dataset. Additionally, by leveraging QLoRA and LoRA for pretraining and fine-tuning, we introduce GeoCode-GPT-7B, the first LLM focused on geospatial code generation, fine-tuned from Code Llama-7B. Furthermore, we establish a comprehensive geospatial code evaluation framework, incorporating option matching, expert validation, and prompt engineering scoring for LLMs, and systematically evaluate GeoCode-GPT-7B using the GeoCode-Eval dataset. Experimental results show that GeoCode-GPT outperforms other models in multiple-choice accuracy by 9.1% to 32.1%, in code summarization ability by 1.7% to 25.4%, and in code generation capability by 1.2% to 25.1%. This paper provides a solution and empirical validation for enhancing LLMs' performance in geospatial code generation, extends the boundaries of domain-specific model applications, and offers valuable insights into unlocking their potential in geospatial code generation.
Recent advancements in large language models (LLMs) have significantly enhanced their ability to understand both natural language and code, driving their use in tasks like natural language-to-code (NL2Code) and code summarization. However, LLMs are prone to hallucination-outputs that stray from intended meanings. Detecting hallucinations in code summarization is especially difficult due to the complex interplay between programming and natural languages. We introduce a first-of-its-kind dataset with $\sim$10K samples, curated specifically for hallucination detection in code summarization. We further propose a novel Entity Tracing Framework (ETF) that a) utilizes static program analysis to identify code entities from the program and b) uses LLMs to map and verify these entities and their intents within generated code summaries. Our experimental analysis demonstrates the effectiveness of the framework, leading to a 0.73 F1 score. This approach provides an interpretable method for detecting hallucinations by grounding entities, allowing us to evaluate summary accuracy.
Pretrained language models have shown strong effectiveness in code-related tasks, such as code retrieval, code generation, code summarization, and code completion tasks. In this paper, we propose COde assistaNt viA retrieval-augmeNted language model (CONAN), which aims to build a code assistant by mimicking the knowledge-seeking behaviors of humans during coding. Specifically, it consists of a code structure aware retriever (CONAN-R) and a dual-view code representation-based retrieval-augmented generation model (CONAN-G). CONAN-R pretrains CodeT5 using Code-Documentation Alignment and Masked Entity Prediction tasks to make language models code structure-aware and learn effective representations for code snippets and documentation. Then CONAN-G designs a dual-view code representation mechanism for implementing a retrieval-augmented code generation model. CONAN-G regards the code documentation descriptions as prompts, which help language models better understand the code semantics. Our experiments show that CONAN achieves convincing performance on different code generation tasks and significantly outperforms previous retrieval augmented code generation models. Our further analyses show that CONAN learns tailored representations for both code snippets and documentation by aligning code-documentation data pairs and capturing structural semantics by masking and predicting entities in the code data. Additionally, the retrieved code snippets and documentation provide necessary information from both program language and natural language to assist the code generation process. CONAN can also be used as an assistant for Large Language Models (LLMs), providing LLMs with external knowledge in shorter code document lengths to improve their effectiveness on various code tasks. It shows the ability of CONAN to extract necessary information and help filter out the noise from retrieved code documents.
Recent advancements in developing Pre-trained Language Models for Code (Code-PLMs) have urged many areas of Software Engineering (SE) and brought breakthrough results for many SE tasks. Though these models have achieved the state-of-the-art performance for SE tasks for many popular programming languages, such as Java and Python, the Scientific Software and its related languages like R programming language have rarely benefited or even been evaluated with the Code-PLMs. Research has shown that R has many differences with other programming languages and requires specific techniques. In this study, we provide the first insights for code intelligence for R. For this purpose, we collect and open source an R dataset, and evaluate Code-PLMs for the two tasks of code summarization and method name prediction using several settings and strategies, including the differences in two R styles, Tidy-verse and Base R. Our results demonstrate that the studied models have experienced varying degrees of performance degradation when processing R programming language code, which is supported by human evaluation. Additionally, not all models show performance improvement in R-specific tasks even after multi-language fine-tuning. The dual syntax paradigms in R significantly impact the models' performance, particularly in code summarization tasks. Furthermore, the project-specific context inherent in R codebases significantly impacts the performance when attempting cross-project training.
The increasing prevalence of large language models (LLMs) such as GPT-4 in various applications has led to a surge in the size of prompts required for optimal performance, leading to challenges in computational efficiency. Prompt compression aims to reduce the inference cost by minimizing input tokens without compromising on the task performance. However, existing prompt compression techniques either rely on sub-optimal metrics such as information entropy or model it as a task-agnostic token classification problem that fails to capture task-specific information. To address these issues, we propose a novel and efficient reinforcement learning (RL) based task-aware prompt compression method. To ensure low latency requirements, we leverage existing Transformer encoder-based token classification model while guiding the learning process with task-specific reward signals using lightweight REINFORCE algorithm. We evaluate the performance of our method on three diverse and challenging tasks including text summarization, question answering and code summarization. We demonstrate that our RL-guided compression method improves the task performance by 8% - 189% across these three scenarios over state-of-the-art compression techniques while satisfying the same compression rate and latency requirements.
This paper presents our findings on the automatic summarization of Java methods within Ericsson, a global telecommunications company. We evaluate the performance of an approach called Automatic Semantic Augmentation of Prompts (ASAP), which uses a Large Language Model (LLM) to generate leading summary comments for Java methods. ASAP enhances the $LLM's$ prompt context by integrating static program analysis and information retrieval techniques to identify similar exemplar methods along with their developer-written Javadocs, and serves as the baseline in our study. In contrast, we explore and compare the performance of four simpler approaches that do not require static program analysis, information retrieval, or the presence of exemplars as in the ASAP method. Our methods rely solely on the Java method body as input, making them lightweight and more suitable for rapid deployment in commercial software development environments. We conducted experiments on an Ericsson software project and replicated the study using two widely-used open-source Java projects, Guava and Elasticsearch, to ensure the reliability of our results. Performance was measured across eight metrics that capture various aspects of similarity. Notably, one of our simpler approaches performed as well as or better than the ASAP method on both the Ericsson project and the open-source projects. Additionally, we performed an ablation study to examine the impact of method names on Javadoc summary generation across our four proposed approaches and the ASAP method. By masking the method names and observing the generated summaries, we found that our approaches were statistically significantly less influenced by the absence of method names compared to the baseline. This suggests that our methods are more robust to variations in method names and may derive summaries more comprehensively from the method body than the ASAP approach.
Mainframe operating systems, despite their inception in the 1940s, continue to support critical sectors like finance and government. However, these systems are often viewed as outdated, requiring extensive maintenance and modernization. Addressing this challenge necessitates innovative tools that can understand and interact with legacy codebases. To this end, we introduce XMainframe, a state-of-the-art large language model (LLM) specifically designed with knowledge of mainframe legacy systems and COBOL codebases. Our solution involves the creation of an extensive data collection pipeline to produce high-quality training datasets, enhancing XMainframe's performance in this specialized domain. Additionally, we present MainframeBench, a comprehensive benchmark for assessing mainframe knowledge, including multiple-choice questions, question answering, and COBOL code summarization. Our empirical evaluations demonstrate that XMainframe consistently outperforms existing state-of-the-art LLMs across these tasks. Specifically, XMainframe achieves 30% higher accuracy than DeepSeek-Coder on multiple-choice questions, doubles the BLEU score of Mixtral-Instruct 8x7B on question answering, and scores six times higher than GPT-3.5 on COBOL summarization. Our work highlights the potential of XMainframe to drive significant advancements in managing and modernizing legacy systems, thereby enhancing productivity and saving time for software developers.
Machine learning models trained on code and related artifacts offer valuable support for software maintenance but suffer from interpretability issues due to their complex internal variables. These concerns are particularly significant in safety-critical applications where the models' decision-making processes must be reliable. The specific features and representations learned by these models remain unclear, adding to the hesitancy in adopting them widely. To address these challenges, we introduce DeepCodeProbe, a probing approach that examines the syntax and representation learning abilities of ML models designed for software maintenance tasks. Our study applies DeepCodeProbe to state-of-the-art models for code clone detection, code summarization, and comment generation. Findings reveal that while small models capture abstract syntactic representations, their ability to fully grasp programming language syntax is limited. Increasing model capacity improves syntax learning but introduces trade-offs such as increased training time and overfitting. DeepCodeProbe also identifies specific code patterns the models learn from their training data. Additionally, we provide best practices for training models on code to enhance performance and interpretability, supported by an open-source replication package for broader application of DeepCodeProbe in interpreting other code-related models.
To support software developers in understanding and maintaining programs, various automatic (source) code summarization techniques have been proposed to generate a concise natural language summary (i.e., comment) for a given code snippet. Recently, the emergence of large language models (LLMs) has led to a great boost in the performance of code-related tasks. In this paper, we undertake a systematic and comprehensive study on code summarization in the era of LLMs, which covers multiple aspects involved in the workflow of LLM-based code summarization. Specifically, we begin by examining prevalent automated evaluation methods for assessing the quality of summaries generated by LLMs and find that the results of the GPT-4 evaluation method are most closely aligned with human evaluation. Then, we explore the effectiveness of five prompting techniques (zero-shot, few-shot, chain-of-thought, critique, and expert) in adapting LLMs to code summarization tasks. Contrary to expectations, advanced prompting techniques may not outperform simple zero-shot prompting. Next, we investigate the impact of LLMs' model settings (including top\_p and temperature parameters) on the quality of generated summaries. We find the impact of the two parameters on summary quality varies by the base LLM and programming language, but their impacts are similar. Moreover, we canvass LLMs' abilities to summarize code snippets in distinct types of programming languages. The results reveal that LLMs perform suboptimally when summarizing code written in logic programming languages compared to other language types. Finally, we unexpectedly find that CodeLlama-Instruct with 7B parameters can outperform advanced GPT-4 in generating summaries describing code implementation details and asserting code properties. We hope that our findings can provide a comprehensive understanding of code summarization in the era of LLMs.
Recent advancements in open-source code large language models (LLMs) have been driven by fine-tuning on the data generated from powerful closed-source LLMs, which are expensive to obtain. This paper explores whether it is possible to use a fine-tuned open-source model to generate additional data to augment its instruction-tuning dataset. We make two observations: (1) A code snippet can serve as the response to different instructions. (2) Instruction-tuned code LLMs perform better at translating code into instructions than the reverse. Based on these observations, we propose Inverse-Instruct, a data augmentation technique that uses a fine-tuned LLM to generate additional instructions of code responses from its own training dataset. The additional instruction-response pairs are added to the original dataset, and a stronger code LLM can be obtained by fine-tuning on the augmented dataset. We empirically validate Inverse-Instruct on a range of open-source code models (e.g. CodeLlama-Python and DeepSeek-Coder) and benchmarks (e.g., HumanEval(+), MBPP(+), DS-1000 and MultiPL-E), showing it consistently improves the base models.
Large Language Models for Code (code LLMs) have demonstrated remarkable performance across various software engineering (SE) tasks, increasing the application of code LLMs in software development. Despite the success of code LLMs, there remain significant concerns about the actual capabilities and reliability of these models, "whether these models really learn the semantics of code from the training data and leverage the learned knowledge to perform the SE tasks". In this paper, we introduce EMPICA, a comprehensive framework designed to systematically and empirically evaluate the capabilities of code LLMs in understanding code semantics. Specifically, EMPICA systematically introduces controlled modifications/transformations into the input code and examines the models' responses. Generally, code LLMs must be robust to semantically equivalent code inputs and be sensitive to non-equivalent ones for all SE tasks. Specifically, for every SE task, given an input code snippet c and its semantic equivalent variants, code LLMs must robustly produce consistent/equivalent outputs while they are expected to generate different outputs for c and its semantic non-equivalent variants. Our experimental results on three representative code understanding tasks, including code summarization, method name prediction, and output prediction, reveal that the robustness and sensitivity of the state-of-the-art code LLMs to code transformations vary significantly across tasks and transformation operators. In addition, the code LLMs exhibit better robustness to the semantic preserving transformations than their sensitivity to the semantic non-preserving transformations. These results highlight a need to enhance the model's capabilities of understanding code semantics, especially the sensitivity property.
(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.
Binary malware summarization aims to automatically generate human-readable descriptions of malware behaviors from executable files, facilitating tasks like malware cracking and detection. Previous methods based on Large Language Models (LLMs) have shown great promise. However, they still face significant issues, including poor usability, inaccurate explanations,and incomplete summaries, primarily due to the obscure pseudocode structure and the lack of malware training summaries. Further, calling relationships between functions, which involve the rich interactions within a binary malware, remain largely underexplored. To this end, we propose MALSIGHT, a novel code summarization framework that can iteratively generate descriptions of binary malware by exploring malicious source code and benign pseudocode. Specifically, we construct the first malware summary dataset, MalS and MalP, using an LLM and manually refine this dataset with human effort. At the training stage, we tune our proposed MalT5, a novel LLM-based code model, on the MalS and benign pseudocode datasets. Then, at the test stage, we iteratively feed the pseudocode functions into MalT5 to obtain the summary. Such a procedure facilitates the understanding of pseudocode structure and captures the intricate interactions between functions, thereby benefiting summaries' usability, accuracy, and completeness. Additionally, we propose a novel evaluation benchmark, BLEURT-sum, to measure the quality of summaries. Experiments on three datasets show the effectiveness of the proposed MALSIGHT. Notably, our proposed MalT5, with only 0.77B parameters, delivers comparable performance to much larger Code-Llama.
Recently, there has been increasing activity in using deep learning for software engineering, including tasks like code generation and summarization. In particular, the most recent coding Large Language Models seem to perform well on these problems. In this technical report, we aim to review how these models perform in code explanation/summarization, while also investigating their code generation capabilities (based on natural language descriptions).
We introduce AutoCoder, the first Large Language Model to surpass GPT-4 Turbo (April 2024) and GPT-4o in pass@1 on the Human Eval benchmark test ($\mathbf{90.9\%}$ vs. $\mathbf{90.2\%}$). In addition, AutoCoder offers a more versatile code interpreter compared to GPT-4 Turbo and GPT-4o. It's code interpreter can install external packages instead of limiting to built-in packages. AutoCoder's training data is a multi-turn dialogue dataset created by a system combining agent interaction and external code execution verification, a method we term \textbf{\textsc{AIEV-Instruct}} (Instruction Tuning with Agent-Interaction and Execution-Verified). Compared to previous large-scale code dataset generation methods, \textsc{AIEV-Instruct} reduces dependence on proprietary large models and provides execution-validated code dataset. The code and the demo video is available in \url{https://github.com/bin123apple/AutoCoder}.
Parameter Efficient Fine-Tuning (PEFT) methods are proposed as an alternative fine-tuning approach for Large Language Models (LLM) to minimize high training costs. While prior research demonstrates the effectiveness of PEFT methods in knowledge transfer using smaller language models, their application to larger LLMs, particularly in low-resource and unseen programming languages such as R, remains under-explored. In this work, we evaluate PEFT methods, LoRA, Compacter, and IA^3 on LLMs for code summarization and generation, with a particular emphasis on knowledge transfer to R as an unseen under-explored target language. Our experiments reveal that LoRA consistently outperforms Compacter and IA^3 in all settings, while Compacter offers significant resource efficiency with minimal performance trade-offs. Additionally, we find that the number of trainable parameters has a greater influence on the functional accuracy of the generated code than PEFT architecture. Our study can direct future research in developing code intelligent tasks for unseen languages including R, as well as the choice of PEFT methods for knowledge transfer, especially when balancing the computational cost and performance.
A brief, fluent, and relevant summary can be helpful during program comprehension; however, such a summary does require significant human effort to produce. Often, good summaries are unavailable in software projects, which makes maintenance more difficult. There has been a considerable body of research into automated AI-based methods, using Large Language models (LLMs), to generate summaries of code; there also has been quite a bit of work on ways to measure the performance of such summarization methods, with special attention paid to how closely these AI-generated summaries resemble a summary a human might have produced. Measures such as BERTScore and BLEU have been suggested and evaluated with human-subject studies. However, LLM-generated summaries can be inaccurate, incomplete, etc.: generally, too dissimilar to one that a good developer might write. Given an LLM-generated code summary, how can a user rationally judge if a summary is sufficiently good and reliable? Given just some input source code, and an LLM-generated summary, existing approaches can help judge brevity, fluency and relevance of the summary; however, it's difficult to gauge whether an LLM-generated summary sufficiently resembles what a human might produce, without a "golden" human-produced summary to compare against. We study this resemblance question as calibration problem: given just the code & the summary from an LLM, can we compute a confidence measure, that provides a reliable indication of whether the summary sufficiently resembles what a human would have produced in this situation? We examine this question using several LLMs, for several languages, and in several different settings. Our investigation suggests approaches to provide reliable predictions of the likelihood that an LLM-generated summary would sufficiently resemble a summary a human might write for the same code.
In this work, we begin to investigate the possibility of training a deep neural network on the task of binary code understanding. Specifically, the network would take, as input, features derived directly from binaries and output English descriptions of functionality to aid a reverse engineer in investigating the capabilities of a piece of closed-source software, be it malicious or benign. Given recent success in applying large language models (generative AI) to the task of source code summarization, this seems a promising direction. However, in our initial survey of the available datasets, we found nothing of sufficiently high quality and volume to train these complex models. Instead, we build our own dataset derived from a capture of Stack Overflow containing 1.1M entries. A major result of our work is a novel dataset evaluation method using the correlation between two distances on sample pairs: one distance in the embedding space of inputs and the other in the embedding space of outputs. Intuitively, if two samples have inputs close in the input embedding space, their outputs should also be close in the output embedding space. We found this Embedding Distance Correlation (EDC) test to be highly diagnostic, indicating that our collected dataset and several existing open-source datasets are of low quality as the distances are not well correlated. We proceed to explore the general applicability of EDC, applying it to a number of qualitatively known good datasets and a number of synthetically known bad ones and found it to be a reliable indicator of dataset value.
Large language models (LLMs) such as Llama 2 perform very well on tasks that involve both natural language and source code, particularly code summarization and code generation. We show that for the task of code summarization, the performance of these models on individual examples often depends on the amount of (subword) token overlap between the code and the corresponding reference natural language descriptions in the dataset. This token overlap arises because the reference descriptions in standard datasets (corresponding to docstrings in large code bases) are often highly similar to the names of the functions they describe. We also show that this token overlap occurs largely in the function names of the code and compare the relative performance of these models after removing function names versus removing code structure. We also show that using multiple evaluation metrics like BLEU and BERTScore gives us very little additional insight since these metrics are highly correlated with each other.