Yaroslav Zharov

Abhiram Bellur, Mohammed Raihan Ullah, Fraol Batole et al.

The primary value of AI agents in software development lies in their ability to extend the developer's capacity for reasoning and action, not to supplant human involvement. To showcase how to use agents working in tandem with developers, we designed a novel approach for carrying out coordinated renaming. Coordinated renaming, where a single rename refactoring triggers refactorings in multiple, related identifiers, is a frequent yet challenging task. Developers must manually propagate these rename refactorings across numerous files and contexts, a process that is both tedious and highly error-prone. State-of-the-art heuristic-based approaches produce an overwhelming number of false positives, while vanilla Large Language Models (LLMs) provide incomplete suggestions due to their limited context and inability to interact with refactoring tools. This leaves developers with incomplete refactorings or burdens them with filtering too many false positives. Coordinated renaming is exactly the kind of repetitive task that agents can significantly reduce the developers' burden while keeping them in the driver's seat. We designed, implemented, and evaluated the first multi-agent framework that automates coordinated renaming. It operates on a key insight: a developer's initial refactoring is a clue to infer the scope of related refactorings. Our Scope Inference Agent first transforms this clue into an explicit, natural-language Declared Scope. The Planned Execution Agent then uses this as a strict plan to identify program elements that should undergo refactoring and safely executes the changes by invoking the IDE's own trusted refactoring APIs. Finally, the Replication Agent uses it to guide the project-wide search. We first conducted a formative study on the practice of coordinated renaming in 609K commits in 100 open-source projects and surveyed 205 developers ...

Yaroslav Zharov, Evelina Ametova, Rebecca Spiecker et al.

Noise is an important issue for radiographic and tomographic imaging techniques. It becomes particularly critical in applications where additional constraints force a strong reduction of the Signal-to-Noise Ratio (SNR) per image. These constraints may result from limitations on the maximum available flux or permissible dose and the associated restriction on exposure time. Often, a high SNR per image is traded for the ability to distribute a given total exposure capacity per pixel over multiple channels, thus obtaining additional information about the object by the same total exposure time. These can be energy channels in the case of spectroscopic imaging or time channels in the case of time-resolved imaging. In this paper, we report on a method for improving the quality of noisy multi-channel (time or energy-resolved) imaging datasets. The method relies on the recent Noise2Noise (N2N) self-supervised denoising approach that learns to predict a noise-free signal without access to noise-free data. N2N in turn requires drawing pairs of samples from a data distribution sharing identical signals while being exposed to different samples of random noise. The method is applicable if adjacent channels share enough information to provide images with similar enough information but independent noise. We demonstrate several representative case studies, namely spectroscopic (k-edge) X-ray tomography, in vivo X-ray cine-radiography, and energy-dispersive (Bragg edge) neutron tomography. In all cases, the N2N method shows dramatic improvement and outperforms conventional denoising methods. For such imaging techniques, the method can therefore significantly improve image quality, or maintain image quality with further reduced exposure time per image.

Jwalin Bhatt, Yaroslav Zharov, Sungho Suh et al.

Morphological atlases are an important tool in organismal studies, and modern high-throughput Computed Tomography (CT) facilities can produce hundreds of full-body high-resolution volumetric images of organisms. However, creating an atlas from these volumes requires accurate organ segmentation. In the last decade, machine learning approaches have achieved incredible results in image segmentation tasks, but they require large amounts of annotated data for training. In this paper, we propose a self-training framework for multi-organ segmentation in tomographic images of Medaka fish. We utilize the pseudo-labeled data from a pretrained Teacher model and adopt a Quality Classifier to refine the pseudo-labeled data. Then, we introduce a pixel-wise knowledge distillation method to prevent overfitting to the pseudo-labeled data and improve the segmentation performance. The experimental results demonstrate that our method improves mean Intersection over Union (IoU) by 5.9% on the full dataset and enables keeping the quality while using three times less markup.

Yaroslav Zharov, Tilo Baumbach, Vincent Heuveline

In Computed Tomography, machine learning is often used for automated data processing. However, increasing model complexity is accompanied by increasingly large volume datasets, which in turn increases the cost of model training. Unlike most work that mitigates this by advancing model architectures and training algorithms, we consider the annotation procedure and its effect on the model performance. We assume three main virtues of a good dataset collected for a model training to be label quality, diversity, and completeness. We compare the effects of those virtues on the model performance using open medical CT datasets and conclude, that quality is more important than diversity early during labeling; the diversity, in turn, is more important than completeness. Based on this conclusion and additional experiments, we propose a labeling procedure for the segmentation of tomographic images to minimize efforts spent on labeling while maximizing the model performance.

Yaroslav Zharov, Alexey Ershov, Tilo Baumbach et al.

The technical advances in Computed Tomography (CT) allow to obtain immense amounts of 3D data. For such datasets it is very costly and time-consuming to obtain the accurate 3D segmentation markup to train neural networks. The annotation is typically done for a limited number of 2D slices, followed by an interpolation. In this work, we propose a pre-training method SortingLoss. It performs pre-training on slices instead of volumes, so that a model could be fine-tuned on a sparse set of slices, without the interpolation step. Unlike general methods (e.g. SimCLR or Barlow Twins), the task specific methods (e.g. Transferable Visual Words) trade broad applicability for quality benefits by imposing stronger assumptions on the input data. We propose a relatively mild assumption -- if we take several slices along some axis of a volume, structure of the sample presented on those slices, should give a strong clue to reconstruct the correct order of those slices along the axis. Many biomedical datasets fulfill this requirement due to the specific anatomy of a sample and pre-defined alignment of the imaging setup. We examine the proposed method on two datasets: medical CT of lungs affected by COVID-19 disease, and high-resolution synchrotron-based full-body CT of model organisms (Medaka fish). We show that the proposed method performs on par with SimCLR, while working 2x faster and requiring 1.5x less memory. In addition, we present the benefits in terms of practical scenarios, especially the applicability to the pre-training of large models and the ability to localize samples within volumes in an unsupervised setup.

Kirill Gelvan, Igor Slinko, Felix Steinbauer et al.

LLM-based Software Engineering agents face a critical bottleneck: context length limitations cause failures on complex, long-horizon tasks. One promising solution is to encode context as continuous embeddings rather than discrete tokens, enabling denser information storage. We apply the recently proposed In-Context Autoencoder for this purpose. While the method performs well on single-shot common-knowledge and code-understanding tasks, our experiments demonstrate that it fails on multi-step agentic coding tasks. In this paper, we explore this phenomenon and discuss possible factors contributing to this failure.

Igor Slinko, Ilia Zavidnyi, Egor Bogomolov et al.

Rejection Fine-Tuning (RFT) is a standard method for training LLM agents, where unsuccessful trajectories are discarded from the training set. In the context of SWE-bench tasks, this corresponds to filtering out runs where the submitted patch does not pass the tests. However, this approach discards unresolved trajectories, even though they form a large portion of all trajectories for hard tasks and even then may be partially correct. In this work, we propose Step Rejection Fine-Tuning (SRFT) - a practical way to leverage these unresolved trajectories. For this, we employ a critic LLM to assess the correctness of each step in a trajectory. Consequently, during training, we mask the loss for erroneous steps while retaining them in the context window. This way we ensure the model learns to recover from errors without reproducing them. Evaluation on SWE-bench Verified shows that while RFT improves the resolution rate by 2.4% by excluding unresolved trajectories, SRFT improves it by 3.7% by filtering them instead of discarding completely, reaching the total resolution rate of 32.2%.

Tobias Lindenbauer, Egor Bogomolov, Yaroslav Zharov

Benchmarks for Software Engineering (SE) AI agents, most notably SWE-bench, have catalyzed progress in programming capabilities of AI agents. However, they overlook critical developer workflows such as Version Control System (VCS) operations. To address this issue, we present GitGoodBench, a novel benchmark for evaluating AI agent performance on VCS tasks. GitGoodBench covers three core Git scenarios extracted from permissive open-source Python, Java, and Kotlin repositories. Our benchmark provides three datasets: a comprehensive evaluation suite (900 samples), a rapid prototyping version (120 samples), and a training corpus (17,469 samples). We establish baseline performance on the prototyping version of our benchmark using GPT-4o equipped with custom tools, achieving a 21.11% solve rate overall. We expect GitGoodBench to serve as a crucial stepping stone toward truly comprehensive SE agents that go beyond mere programming.

Abhiram Bellur, Fraol Batole, Mohammed Raihan Ullah et al.

MOVEMETHOD is a hallmark refactoring. Despite a plethora of research tools that recommend which methods to move and where, these recommendations do not align with how expert developers perform MOVEMETHOD. Given the extensive training of Large Language Models and their reliance upon naturalness of code, they should expertly recommend which methods are misplaced in a given class and which classes are better hosts. Our formative study of 2016 LLM recommendations revealed that LLMs give expert suggestions, yet they are unreliable: up to 80% of the suggestions are hallucinations. We introduce the first LLM fully powered assistant for MOVEMETHOD refactoring that automates its whole end-to-end lifecycle, from recommendation to execution. We designed novel solutions that automatically filter LLM hallucinations using static analysis from IDEs and a novel workflow that requires LLMs to be self-consistent, critique, and rank refactoring suggestions. As MOVEMETHOD refactoring requires global, projectlevel reasoning, we solved the limited context size of LLMs by employing refactoring-aware retrieval augment generation (RAG). Our approach, MM-assist, synergistically combines the strengths of the LLM, IDE, static analysis, and semantic relevance. In our thorough, multi-methodology empirical evaluation, we compare MM-assist with the previous state-of-the-art approaches. MM-assist significantly outperforms them: (i) on a benchmark widely used by other researchers, our Recall@1 and Recall@3 show a 1.7x improvement; (ii) on a corpus of 210 recent refactorings from Open-source software, our Recall rates improve by at least 2.4x. Lastly, we conducted a user study with 30 experienced participants who used MM-assist to refactor their own code for one week. They rated 82.8% of MM-assist recommendations positively. This shows that MM-assist is both effective and useful.

Alexander Kovrigin, Aleksandra Eliseeva, Konstantin Grotov et al.

Environment setup-the process of configuring the system to work with a specific software project-represents a persistent challenge in Software Engineering (SE). Automated environment setup methods could assist developers by providing fully configured environments for arbitrary repositories without manual effort. This also helps SE researchers to scale execution-based benchmarks. However, recent studies reveal that even state-of-the-art Large Language Models (LLMs) achieve limited success in automating this task. To address this limitation, we tune a specialized model for environment setup. We combine supervised fine-tuning for generating correct Bash scripts and Reinforcement Learning with Verifiable Rewards (RLVR) to adapt it to the task of environment setup. On EnvBench-Python, our method enables Qwen3-8B (a model runnable on consumer hardware) to perform on par with larger models-Qwen3-32B and GPT-4o. The training code and model checkpoints are available online: https://github.com/JetBrains-Research/PIPer.

Alexander Kovrigin, Aleksandra Eliseeva, Yaroslav Zharov et al.

Recent advancements in code-fluent Large Language Models (LLMs) enabled the research on repository-level code editing. In such tasks, the model navigates and modifies the entire codebase of a project according to request. Hence, such tasks require efficient context retrieval, i.e., navigating vast codebases to gather relevant context. Despite the recognized importance of context retrieval, existing studies tend to approach repository-level coding tasks in an end-to-end manner, rendering the impact of individual components within these complicated systems unclear. In this work, we decouple the task of context retrieval from the other components of the repository-level code editing pipelines. We lay the groundwork to define the strengths and weaknesses of this component and the role that reasoning plays in it by conducting experiments that focus solely on context retrieval. We conclude that while the reasoning helps to improve the precision of the gathered context, it still lacks the ability to identify its sufficiency. We also outline the ultimate role of the specialized tools in the process of context gathering. The code supplementing this paper is available at https://github.com/JetBrains-Research/ai-agents-code-editing.

Egor Bogomolov, Yaroslav Zharov

Software engineering, whether performed by humans or by AI agents, requires reasoning about how software behaves. We call the internal model that supports such reasoning the software world model, and view current code-execution benchmarks as covering one well-studied slice of it -- control flow. In this paper, we take a step toward a broader evaluation by shifting the observable axis to execution resources: alongside test outcome and exception class, we predict peak memory, wall-clock time, and ranked profiler outputs at method and line granularity. We use SWE-bench Verified as the source of data to hold the test close to real-world software engineering tasks. All tested models, frontier ones included, show modest performance and brittle behaviour, suggesting a notable lack of understanding of how software is executed, as opposed to how its source code is written.

Konstantin Grotov, Artem Borzilov, Maksim Krivobok et al.

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. With the rise of code-fluent Large Language Models empowered with agentic techniques, smart bug-fixing tools with a high level of autonomy have emerged. However, those tools are tuned for classical script programming and still struggle with non-linear computational notebooks. In this paper, we present an AI agent designed specifically for error resolution in a computational notebook. We have developed an agentic system capable of exploring a notebook environment by interacting with it -- similar to how a user would -- and integrated the system into the JetBrains service for collaborative data science called Datalore. We evaluate our approach against the pre-existing single-action solution by comparing costs and conducting a user study. Users rate the error resolution capabilities of the agentic system higher but experience difficulties with UI. We share the results of the study and consider them valuable for further improving user-agent collaboration.

Konstantin Grotov, Sergey Titov, Yaroslav Zharov et al.

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.

Tobias Lindenbauer, Igor Slinko, Ludwig Felder et al.

Large Language Model (LLM)-based agents solve complex tasks through iterative reasoning, exploration, and tool-use, a process that can result in long, expensive context histories. While state-of-the-art Software Engineering (SE) agents like OpenHands or Cursor use LLM-based summarization to tackle this issue, it is unclear whether the increased complexity offers tangible performance benefits compared to simply omitting older observations. We present a systematic comparison of these approaches within SWE-agent on SWE-bench Verified across five diverse model configurations. Moreover, we show initial evidence of our findings generalizing to the OpenHands agent scaffold. We find that a simple environment observation masking strategy halves cost relative to the raw agent while matching, and sometimes slightly exceeding, the solve rate of LLM summarization. Additionally, we introduce a novel hybrid approach that further reduces costs by 7% and 11% compared to just observation masking or LLM summarization, respectively. Our findings raise concerns regarding the trend towards pure LLM summarization and provide initial evidence of untapped cost reductions by pushing the efficiency-effectiveness frontier. We release code and data for reproducibility.

Petr Tsvetkov, Aleksandra Eliseeva, Danny Dig et al.

When a Commit Message Generation (CMG) system is integrated into the IDEs and other products at JetBrains, we perform online evaluation based on user acceptance of the generated messages. However, performing online experiments with every change to a CMG system is troublesome, as each iteration affects users and requires time to collect enough statistics. On the other hand, offline evaluation, a prevalent approach in the research literature, facilitates fast experiments but employs automatic metrics that are not guaranteed to represent the preferences of real users. In this work, we describe a novel way we employed to deal with this problem at JetBrains, by leveraging an online metric - the number of edits users introduce before committing the generated messages to the VCS - to select metrics for offline experiments. To support this new type of evaluation, we develop a novel markup collection tool mimicking the real workflow with a CMG system, collect a dataset with 57 pairs consisting of commit messages generated by GPT-4 and their counterparts edited by human experts, and design and verify a way to synthetically extend such a dataset. Then, we use the final dataset of 656 pairs to study how the widely used similarity metrics correlate with the online metric reflecting the real users' experience. Our results indicate that edit distance exhibits the highest correlation with the online metric, whereas commonly used similarity metrics such as BLEU and METEOR demonstrate low correlation. This contradicts the previous studies on similarity metrics for CMG, suggesting that user interactions with a CMG system in real-world settings differ significantly from the responses by human labelers within controlled environments. We release all the code and the dataset to support future research in the field: https://jb.gg/cmg-evaluation.

Yaroslav Zharov, Alexey Ershov, Tilo Baumbach et al.

In synchrotron-based Computed Tomography (CT) there is a trade-off between spatial resolution, field of view and speed of positioning and alignment of samples. The problem is even more prominent for high-throughput tomography--an automated setup, capable of scanning large batches of samples without human interaction. As a result, in many applications, only 20-30% of the reconstructed volume contains the actual sample. Such data redundancy clutters the storage and increases processing time. Hence, an automated sample localization becomes an important practical problem. In this work, we describe two self-supervised losses designed for biological CT. We further demonstrate how to employ the uncertainty estimation for sample localization. This approach shows the ability to localize a sample with less than 1.5\% relative error and reduce the used storage by a factor of four. We also show that one of the proposed losses works reasonably well as a pre-training task for the semantic segmentation.

Yaroslav Zharov, Denis Korzhenkov, Pavel Shvechikov et al.

We introduce a novel approach to feed-forward neural network interpretation based on partitioning the space of sequences of neuron activations. In line with this approach, we propose a model-specific interpretation method, called YASENN. Our method inherits many advantages of model-agnostic distillation, such as an ability to focus on the particular input region and to express an explanation in terms of features different from those observed by a neural network. Moreover, examination of distillation error makes the method applicable to the problems with low tolerance to interpretation mistakes. Technically, YASENN distills the network with an ensemble of layer-wise gradient boosting decision trees and encodes the sequences of neuron activations with leaf indices. The finite number of unique codes induces a partitioning of the input space. Each partition may be described in a variety of ways, including examination of an interpretable model (e.g. a logistic regression or a decision tree) trained to discriminate between objects of those partitions. Our experiments provide an intuition behind the method and demonstrate revealed artifacts in neural network decision making.