arXiv subjects

Timofey Bryksin

Publications and source records attributed to Timofey Bryksin.

At least 37 records · Page 2Linked to original sources

arXiv2024

Together We Go Further: LLMs and IDE Static Analysis for Extract Method Refactoring

Long methods that encapsulate multiple responsibilities within a single method are challenging to maintain. Choosing which statements to extract into new methods has been the target of many research tools. Despite steady improvements, these tools often fail to generate refactorings that align with developers' preferences and acceptance criteria. Given that Large Language Models (LLMs) have been trained on large code corpora, if we harness their familiarity with the way developers form functions, we could suggest refactorings that developers are likely to accept. In this paper, we advance the science and practice of refactoring by synergistically combining the insights of LLMs with the power of IDEs to perform Extract Method (EM). Our formative study on 1752 EM scenarios revealed that LLMs are very effective for giving expert suggestions, yet they are unreliable: up to 76.3% of the suggestions are hallucinations. We designed a novel approach that removes hallucinations from the candidates suggested by LLMs, then further enhances and ranks suggestions based on static analysis techniques from program slicing, and finally leverages the IDE to execute refactorings correctly. We implemented this approach in an IntelliJ IDEA plugin called EM-Assist. We empirically evaluated EM-Assist on a diverse corpus that replicates 1752 actual refactorings from open-source projects. We found that EM-Assist outperforms previous state of the art tools: EM-Assist suggests the developerperformed refactoring in 53.4% of cases, improving over the recall rate of 39.4% for previous best-in-class tools. Furthermore, we conducted firehouse surveys with 16 industrial developers and suggested refactorings on their recent commits. 81.3% of them agreed with the recommendations provided by EM-Assist.

cs.SE↗

arXiv2024

Untangling Knots: Leveraging LLM for Error Resolution in Computational Notebooks

Computational notebooks became indispensable tools for research-related development, offering unprecedented interactivity and flexibility in the development process. However, these benefits come at the cost of reproducibility and an increased potential for bugs. There are many tools for bug fixing; however, they are generally targeted at the classical linear code. With the rise of code-fluent Large Language Models, a new stream of smart bug-fixing tools has emerged. However, the applicability of those tools is still problematic for non-linear computational notebooks. In this paper, we propose a potential solution for resolving errors in computational notebooks via an iterative LLM-based agent. We discuss the questions raised by this approach and share a novel dataset of computational notebooks containing bugs to facilitate the research of the proposed approach.

cs.SE↗

arXiv2024

Dynamic Retrieval-Augmented Generation

Current state-of-the-art large language models are effective in generating high-quality text and encapsulating a broad spectrum of world knowledge. These models, however, often hallucinate and lack locally relevant factual data. Retrieval-augmented approaches were introduced to overcome these problems and provide more accurate responses. Typically, the retrieved information is simply appended to the main request, restricting the context window size of the model. We propose a novel approach for the Dynamic Retrieval-Augmented Generation (DRAG), based on the entity-augmented generation, which injects compressed embeddings of the retrieved entities into the generative model. The proposed pipeline was developed for code-generation tasks, yet can be transferred to some domains of natural language processing. To train the model, we collect and publish a new project-level code generation dataset. We use it for the evaluation along with publicly available datasets. Our approach achieves several targets: (1) lifting the length limitations of the context window, saving on the prompt size; (2) allowing huge expansion of the number of retrieval entities available for the context; (3) alleviating the problem of misspelling or failing to find relevant entity names. This allows the model to beat all baselines (except GPT-3.5) with a strong margin.

cs.SE↗

arXiv2024

Reassessing Java Code Readability Models with a Human-Centered Approach

To ensure that Large Language Models (LLMs) effectively support user productivity, they need to be adjusted. Existing Code Readability (CR) models can guide this alignment. However, there are concerns about their relevance in modern software engineering since they often miss the developers' notion of readability and rely on outdated code. This research assesses existing Java CR models for LLM adjustments, measuring the correlation between their and developers' evaluations of AI-generated Java code. Using the Repertory Grid Technique with 15 developers, we identified 12 key code aspects influencing CR that were consequently assessed by 390 programmers when labeling 120 AI-generated snippets. Our findings indicate that when AI generates concise and executable code, it is often considered readable by CR models and developers. However, a limited correlation between these evaluations underscores the importance of future research on learning objectives for adjusting LLMs and on the aspects influencing CR evaluations included in predictive models.

cs.SE↗

arXiv2023

The Effect of Perceptual Load on Performance within IDE in People with ADHD Symptoms

In this paper, we describe the research on how perceptual load can affect programming performance in people with symptoms of Attention Deficit / Hyperactivity Disorder (ADHD). We asked developers to complete the Barkley Deficits in Executive Functioning Scale, which indicates the presence and severity levels of ADHD symptoms. After that, participants solved mentally active programming tasks (coding) and monotonous ones (debugging) in the integrated development environment in high perceptual load modes (visually noisy) and low perceptual load modes (visually clear). The development environment was augmented with the plugin we wrote to track efficiency metrics, i.e. time, speed, and activity. We found that the perceptual load does affect programmers' efficiency. For mentally active tasks, the time of inserting the first character was shorter and the overall speed was higher in the low perceptual load mode. For monotonous tasks, the total time for the solution was less for the low perceptual load mode. Also, we found that the effect of perceptual load on programmers' efficiency differs between those with and without ADHD symptoms. This effect has a specificity: depending on efficiency measures and ADHD symptoms, one or another level of perceptual load might be beneficial. Our findings support the idea of behavioral assessment of users for providing appropriate accommodation for the workforce with special needs.

cs.SE↗

arXiv2023

From Commit Message Generation to History-Aware Commit Message Completion

Commit messages are crucial to software development, allowing developers to track changes and collaborate effectively. Despite their utility, most commit messages lack important information since writing high-quality commit messages is tedious and time-consuming. The active research on commit message generation (CMG) has not yet led to wide adoption in practice. We argue that if we could shift the focus from commit message generation to commit message completion and use previous commit history as additional context, we could significantly improve the quality and the personal nature of the resulting commit messages. In this paper, we propose and evaluate both of these novel ideas. Since the existing datasets lack historical data, we collect and share a novel dataset called CommitChronicle, containing 10.7M commits across 20 programming languages. We use this dataset to evaluate the completion setting and the usefulness of the historical context for state-of-the-art CMG models and GPT-3.5-turbo. Our results show that in some contexts, commit message completion shows better results than generation, and that while in general GPT-3.5-turbo performs worse, it shows potential for long and detailed messages. As for the history, the results show that historical information improves the performance of CMG models in the generation task, and the performance of GPT-3.5-turbo in both generation and completion.

cs.SE↗

arXiv2023

Overcoming the Mental Set Effect in Programming Problem Solving

This paper adopts a cognitive psychology perspective to investigate the recurring mistakes in code resulting from the mental set (Einstellung) effect. The Einstellung effect is the tendency to approach problem-solving with a preconceived mindset, often overlooking better solutions that may be available. This effect can significantly impact creative thinking, as the development of patterns of thought can hinder the emergence of novel and creative ideas. Our study aims to test the Einstellung effect and the two mechanisms of its overcoming in the field of programming. The first intervention was the change of the color scheme of the code editor to the less habitual one. The second intervention was a combination of instruction to "forget the previous solutions and tasks" and the change in the color scheme. During the experiment, participants were given two sets of four programming tasks. Each task had two possible solutions: one using suboptimal code dictated by the mental set, and the other using a less familiar but more efficient and recommended methodology. Between the sets, participants either received no treatment or one of two interventions aimed at helping them overcome the mental set. The results of our experiment suggest that the tested techniques were insufficient to support overcoming the mental set, which we attribute to the specificity of the programming domain. The study contributes to the existing literature by providing insights into creativity support during problem-solving in software development and offering a framework for experimental research in this field.

cs.SE↗

arXiv2023

Out of the BLEU: how should we assess quality of the Code Generation models?

In recent years, researchers have created and introduced a significant number of various code generation models. As human evaluation of every new model version is unfeasible, the community adopted automatic evaluation metrics such as BLEU to approximate the results of human judgement. These metrics originate from the machine translation domain and it is unclear whether they are applicable for the code generation tasks and how well they agree with the human evaluation on this task. There are also other metrics, CodeBLEU and RUBY, developed to estimate the similarity of code, that take into account the properties of source code. However, for these metrics there are hardly any studies on their agreement with the human evaluation. Despite all that, minimal differences in the metric scores have been used in recent papers to claim superiority of some code generation models over the others. In this paper, we present a study on the applicability of six metrics -- BLEU, ROUGE-L, METEOR, ChrF, CodeBLEU, and RUBY -- for evaluation of code generation models. We conduct a study on two different code generation datasets and use human annotators to assess the quality of all models run on these datasets. The results indicate that for the CoNaLa dataset of Python one-liners, none of the metrics can correctly emulate human judgement on which model is better with >95% certainty if the difference in model scores is less than 5 points. For the HearthStone dataset, which consists of classes of a particular structure, a difference in model scores of at least 2 points is enough to claim the superiority of one model over the other. Our findings suggest that the ChrF metric is a better fit for the evaluation of code generation models than the commonly used BLEU and CodeBLEU. Yet, finding a metric for code generation that closely agrees with humans requires additional work.

cs.SE↗

arXiv2023

Detecting Code Quality Issues in Pre-written Templates of Programming Tasks in Online Courses

In this work, we developed an algorithm for detecting code quality issues in the templates of online programming tasks, validated it, and conducted an empirical study on the dataset of student solutions. The algorithm consists of analyzing recurring unfixed issues in solutions of different students, matching them with the code of the template, and then filtering the results. Our manual validation on a subset of tasks demonstrated a precision of 80.8% and a recall of 73.3%. We used the algorithm on 415 Java tasks from the JetBrains Academy platform and discovered that as much as 14.7% of tasks have at least one issue in their template, thus making it harder for students to learn good code quality practices. We describe our results in detail, provide several motivating examples and specific cases, and share the feedback of the developers of the platform, who fixed 51 issues based on the output of our approach.

cs.SE↗

arXiv2023

What Writing Assistants Can Learn from Programming IDEs

With the development of artificial intelligence, writing assistants (WAs) are changing the way people interact with text, creating lengthy outputs that can be overwhelming for users. The programming field has long addressed this issue, and Integrated Development Environments (IDEs) have been created for efficient software development, helping programmers reduce the cognitive load. This experience could be employed in the development of WAs. IDEs can also be used to test assumptions about interventions that help people interact with WAs efficiently. Previous works have successfully used self-written IDE plugins to test hypotheses in the field of human-computer interaction. The lessons learned can be applied to the building of WAs.

cs.SE↗

arXiv2023

Optimizing Duplicate Size Thresholds in IDEs

In this paper, we present an approach for transferring an optimal lower size threshold for clone detection from one language to another by analyzing their clone distributions. We showcase this method by transferring the threshold from regular Python scripts to Jupyter notebooks for using in two JetBrains IDEs, Datalore and DataSpell.

cs.SE↗

arXiv2023

Judging Adam: Studying the Performance of Optimization Methods on ML4SE Tasks

Solving a problem with a deep learning model requires researchers to optimize the loss function with a certain optimization method. The research community has developed more than a hundred different optimizers, yet there is scarce data on optimizer performance in various tasks. In particular, none of the benchmarks test the performance of optimizers on source code-related problems. However, existing benchmark data indicates that certain optimizers may be more efficient for particular domains. In this work, we test the performance of various optimizers on deep learning models for source code and find that the choice of an optimizer can have a significant impact on the model quality, with up to two-fold score differences between some of the relatively well-performing optimizers. We also find that RAdam optimizer (and its modification with the Lookahead envelope) is the best optimizer that almost always performs well on the tasks we consider. Our findings show a need for a more extensive study of the optimizers in code-related tasks, and indicate that the ML4SE community should consider using RAdam instead of Adam as the default optimizer for code-related deep learning tasks.

cs.SE↗

arXiv2023

Just-in-Time Code Duplicates Extraction

Refactoring is a critical task in software maintenance, and is usually performed to enforce better design and coding practices, while coping with design defects. The Extract Method refactoring is widely used for merging duplicate code fragments into a single new method. Several studies attempted to recommend Extract Method refactoring opportunities using different techniques, including program slicing, program dependency graph analysis, change history analysis, structural similarity, and feature extraction. However, irrespective of the method, most of the existing approaches interfere with the developer's workflow: they require the developer to stop coding and analyze the suggested opportunities, and also consider all refactoring suggestions in the entire project without focusing on the development context. To increase the adoption of the Extract Method refactoring, in this paper, we aim to investigate the effectiveness of machine learning and deep learning algorithms for its recommendation while maintaining the workflow of the developer. The proposed approach relies on mining prior applied Extract Method refactorings and extracting their features to train a deep learning classifier that detects them in the user's code. We implemented our approach as a plugin for IntelliJ IDEA called AntiCopyPaster. To develop our approach, we trained and evaluated various popular models on a dataset of 18,942 code fragments from 13 Open Source Apache projects. The results show that the best model is the Convolutional Neural Network (CNN), which recommends appropriate Extract Method refactorings with an F-measure of 0.82. We also conducted a qualitative study with 72 developers to evaluate the usefulness of the developed plugin. The results show that developers tend to appreciate the idea of the approach and are satisfied with various aspects of the plugin's operation.

cs.SE↗

arXiv2023

Analyzing the Quality of Submissions in Online Programming Courses

Programming education should aim to provide students with a broad range of skills that they will later use while developing software. An important aspect in this is their ability to write code that is not only correct but also of high quality. Unfortunately, this is difficult to control in the setting of a massive open online course. In this paper, we carry out an analysis of the code quality of submissions from JetBrains Academy - a platform for studying programming in an industry-like project-based setting with an embedded code quality assessment tool called Hyperstyle. We analyzed more than a million Java submissions and more than 1.3 million Python submissions, studied the most prevalent types of code quality issues and the dynamics of how students fix them. We provide several case studies of different issues, as well as an analysis of why certain issues remain unfixed even after several attempts. Also, we studied abnormally long sequences of submissions, in which students attempted to fix code quality issues after passing the task. Our results point the way towards the improvement of online courses, such as making sure that the task itself does not incentivize students to write code poorly.

cs.SE↗

arXiv2023

Predicting Tags For Programming Tasks by Combining Textual And Source Code Data

Competitive programming remains a very popular activity that combines both software engineering and education. In order to prepare and to practice, contestants use extensive archives of problems from past contents available on various competitive programming platforms. One way to make this process more effective is to provide an automatic tag system for the tasks. Prior works do that by either using the tasks' problem statements or the code of their solutions. In this study, we investigate which information source is more valuable for tag prediction. To answer that question, we compare existing approaches of both types on the same dataset and with the same set of tags. Then, we propose a novel approach, which is an ensemble of the Gated Graph Neural Network model for analyzing solutions and the Bidirectional Encoder Representations from Transformers model for processing statements. Our experiments show that our approach outperforms previously proposed models by 0.175 of the PR-AUC metric.

cs.SE↗

arXiv2022

So Much in So Little: Creating Lightweight Embeddings of Python Libraries

In software engineering, different approaches and machine learning models leverage different types of data: source code, textual information, historical data. An important part of any project is its dependencies. The list of dependencies is relatively small but carries a lot of semantics with it, which can be used to compare projects or make judgements about them. In this paper, we focus on Python projects and their PyPi dependencies in the form of requirements.txt files. We compile a dataset of 7,132 Python projects and their dependencies, as well as use Git to pull their versions from previous years. Using this data, we build 32-dimensional embeddings of libraries by applying Singular Value Decomposition to the co-occurrence matrix of projects and libraries. We then cluster the embeddings and study their semantic relations. To showcase the usefulness of such lightweight library embeddings, we introduce a prototype tool for suggesting relevant libraries to a given project. The tool computes project embeddings and uses dependencies of projects with similar embeddings to form suggestions. To compare different library recommenders, we have created a benchmark based on the evolution of dependency sets in open-source projects. Approaches based on the created embeddings significantly outperform the baseline of showing the most popular libraries in a given year. We have also conducted a user study that showed that the suggestions differ in quality for different project domains and that even relevant suggestions might be not particularly useful. Finally, to facilitate potentially more useful recommendations, we extended the recommender system with an option to suggest rarer libraries.

cs.SE↗

arXiv2022

All You Need Is Logs: Improving Code Completion by Learning from Anonymous IDE Usage Logs

In this work, we propose an approach for collecting completion usage logs from the users in an IDE and using them to train a machine learning based model for ranking completion candidates. We developed a set of features that describe completion candidates and their context, and deployed their anonymized collection in the Early Access Program of IntelliJ-based IDEs. We used the logs to collect a dataset of code completions from users, and employed it to train a ranking CatBoost model. Then, we evaluated it in two settings: on a held-out set of the collected completions and in a separate A/B test on two different groups of users in the IDE. Our evaluation shows that using a simple ranking model trained on the past user behavior logs significantly improved code completion experience. Compared to the default heuristics-based ranking, our model demonstrated a decrease in the number of typing actions necessary to perform the completion in the IDE from 2.073 to 1.832. The approach adheres to privacy requirements and legal constraints, since it does not require collecting personal information, performing all the necessary anonymization on the client's side. Importantly, it can be improved continuously: implementing new features, collecting new data, and evaluating new models - this way, we have been using it in production since the end of 2020.

cs.SE↗

arXiv2022

AntiCopyPaster: Extracting Code Duplicates As Soon As They Are Introduced in the IDE

We developed a plugin for IntelliJ IDEA called AntiCopyPaster, which tracks the pasting of code fragments inside the IDE and suggests the appropriate Extract Method refactoring to combat the propagation of duplicates. Unlike the existing approaches, our tool is integrated with the developer's workflow, and pro-actively recommends refactorings. Since not all code fragments need to be extracted, we develop a classification model to make this decision. When a developer copies and pastes a code fragment, the plugin searches for duplicates in the currently opened file, waits for a short period of time to allow the developer to edit the code, and finally inferences the refactoring decision based on a number of features. Our experimental study on a large dataset of 18,942 code fragments mined from 13 Apache projects shows that AntiCopyPaster correctly recommends Extract Method refactorings with an F-score of 0.82. Furthermore, our survey of 59 developers reflects their satisfaction with the developed plugin's operation. The plugin and its source code are publicly available on GitHub at https://github.com/JetBrains-Research/anti-copy-paster. The demonstration video can be found on YouTube: https://youtu.be/_wwHg-qFjJY.

cs.SE↗