Conference Program

TUESDAY, OCTOBER 20th

SESSION 1. Information Extraction from Text I: Topic Modeling and Syntactic Analysis.

Alexander Sychev. Функциональная разметка терминов тематических моделей на основе векторных эмбеддингов. (PDF)

Abstract

В работе рассматривается проблема интерпретируемости результатов тематического моделирования текстовых корпусов. Предлагается модельно-независимая схема функциональной разметки терминов, основанная на алгоритме Top2vec, которая классифицирует термины на пять функциональных классов (CORE, PERIPH, MULTI, UNASSIGNED, NOISE) на основе двух ортогональных метрик: центральности и специфичности. Схема использует только матрицу косинусных сходств «термин–тема», что делает ее вычислительно эффективной и применимой к любым эмбеддинговым моделям тематического моделирования. Для объективной оценки качества текстовых описаний тем, генерируемых большими языковыми моделями, предложена метрика WSDD (взвешенная семантическая плотность и различимость), учитывающая взвешенную плотность, абсолютную различимость и покрытие ядра. Эксперимент на корпусе из 32 тысяч новостных сообщений показал, что функциональная разметка может повышать качество LLM-генерации описаний тем по сравнению с плоскими списками топ-50 терминов. Выявлен диагностический потенциал разметки для выявления артефактов моделирования и скрытых структурных слоев корпуса. Предложенный метод не требует переобучения моделей, вычислительно лёгок и может быть интегрирован в существующие конвейеры анализа текстов.

Artem Rostunov and Maksim Stankevich. Identifying Active Event Discussions and Topics through Social Media Analysis. (PDF)

Abstract

This paper investigates the analysis of social media data for the identification of active discussions, periods of increased user activity, and the thematic classification of posts in online news communities of a small city. Using data obtained from the Russian-language social network VKontakte, a dataset consisting of posts and their associated comments was collected for the period from 2019 to 2024. Based on the collected data, a set of discussion activity indicators was proposed and used to construct a discussion activity criterion through principal component analysis. To identify the thematic categories of events, a predefined list of event categories was provided to several large language models together with the texts of the posts. The experimental results demonstrated a high degree of inter-model agreement in category annotation, as well as a relatively high level of agreement with expert judgments. These findings highlight the considerable potential of large language models for monitoring user reactions to various local and global events in online communities.

Драгош Карамела-Аждер and Елена Шамаева. Методы сравнения синтаксических деревьев зависимостей. (PDF)

Abstract

Данная работа посвящена разработке и применению методики выявления структурных ошибок синтаксических анализаторов на основе графовой метрики близости. Традиционные метрики оценки качества LAS и UAS оценивают долю правильно определённых зависимостей, но не учитывают глобальные топологические искажения дерева. В работе исследуются возможности ядра Вайсфейлера–Лемана (WL) для ранжирования предложений по степени структурного отклонения от эталонного разбора. Проведена оценка качества ранжирования с использованием метрики Average Precision, показавшая существенное превосходство графового метода над классическими метриками. На основе анализа причин расхождений между эталонными и предсказанными деревьями формализованы критерии структурных ошибок. Разработана программная утилита, реализующая фильтрацию предложений по графовым метрикам. С помощью разработанной утилиты создан категориальный бенчмарк, разделённый на шесть категорий длины и пять типов структурных ошибок. Проведён сравнительный анализ пяти популярных синтаксических анализаторов на русскоязычных корпусах. Установлено, что на категориальном бенчмарке все рассмотренные анализаторы демонстрируют значения UAS, LAS и WL существенно ниже средних по полному набору данных, при этом наихудшие значения достигнуты на предложениях длиной 1–5 слов.

SESSION 2. Machine Learning in Natural Sciences.

Alexander Marukhin and Sergey Safonov. SINDy-Based Method for Inverse Reconstruction of Time-Dependent Coefficients in Nonlinear ODE Systems. (PDF)

Abstract

Dynamical systems governed by ordinary or partial differential equations with time-dependent coefficients arise in the numerous applications in physics, engineering, control theory and computational biomedicine, where accurate reconstruction of nonstationary dynamics is essential for prediction and interpretation of complex dynamic processes. In this work, we propose a data-driven methodology for solving inverse problems associated with such systems by extending the Sparse Identification of Nonlinear Dynamics (SINDy) method to the case of time-varying coefficients. The proposed approach combines sparse regression based on Sequentially thresholded least squares (STLSQ) for constant coefficients with locally weighted LASSO estimation, enabling reconstruction of temporally evolving coefficients while preserving the sparsity in constant coefficients and interpretable structure of the governing equations. The application of this modified SINDy method is successfully demonstrated on several examples of nonlinear ODE systems. The obtained results provide the potential of the proposed framework for accurate identification of nonstationary dynamical systems, but also the need for further theoretical development in the areas of model selection, numerical stability, convergence analysis, and robustness under strong nonlinear multicollinearity.

Ekaterina Smolina, Dmitry Rostov, Alexander Romanov, Andrey Eliasin, Vasily Kostin, Roman Ponomarev, Leonid Zhukov and Lev Smirnov. Machine Learning for Predicting Focal Characteristics of Lensed Optical Fibers. (PDF)

Abstract

Lensed optical fibers are critical components for efficient coupling between single-mode optical fibers and photonic integrated circuit elements. Accurate prediction of their focal parameters, focal distance � and focal spot diameter � , from a given microlens profile is essential for photonic device design and optimization. Classical numerical methods (FEM and FDTD) provide high accuracy but are computationally prohibitive for large-scale design exploration. In this work, we develop an accelerated 2D simulation framework exploiting the axial symmetry, generate a dataset of synthetic microlens profiles, and train machine learning models, including multiple linear regression and a U-Net convolutional neural network, to predict focal characteristics directly from the lens geometry. The U-Net models trained separately on FEM and FDTD data each achieve an error of about 1% (the larger of the mean absolute percentage errors for � and �), while reducing the per-configuration computation time from several minutes (for the 2D solver) to approximately 3 ms. This framework provides a basis for rapid design-space exploration and tolerance analysis of fiber-to-chip interfaces after validation on experimentally characterized fabrication variations.

[SHORT] Надежда Киселёва, Виктор Дударев, Олег Сенько, Александр Докукин, Андрей Столяренко and Юлиана Кузнецова. Особенности применения методов машинного обучения в неорганической химии. (PDF)

Abstract

Рассмотрены основные особенности применения машинного обучения в контексте неорганической химии и материаловедения. Особое внимание уделено проблемам мультиколлинеарности признакового описания, пропускам в значениях признаков, малому числу экспериментально исследованных примеров для обучения, ошибкам в целевых параметрах объектов, а также способам повышения точности прогнозирования. В статье представлены также разработанные авторами информационные системы конструирования новых неорганических соединений, учитывающие вышеприведённые особенности.

SESSION 3. Information Extraction from Text II: Datasets.

Pavel Mamaev, Natalia Loukachevitch and Grigory Kovalev. Deep Re-annotation and Controlled Candidate Pools for Re-evaluating Ranking Models on MS MARCO. (PDF)

Abstract

This paper studies how additional manual relevance annotation changes the evaluation of retrieval models on MS MARCO, where the original annotation usually provides only one explicitly relevant passage per query. We analyze a manually annotated subset of 1,179 queries and 33,667 query-document pairs, comparing the original MS MARCO labels, extended human labels, and labels produced by a compact cross-encoder judge. Here, deep re-annotation means judging multiple candidate answers within a defined pool, rather than assuming that the original positive passage exhausts the relevant answers. The judge is then applied to a larger set of queries to study pseudo-relevance labels under controlled pooling depths. Manual re-annotation increases the average number of relevant documents per query from 1.07 to 17.47 and yields statistically significant gains in NDCG@10 for all considered ranking sources, for example from 0.597 to 0.877 for BM25 and from 0.701 to 0.967 for E5. On the full set, high pseudo-relevance density appears only after the judge is applied to a large pooled set built from the top BM25, E5, and BGE results. This value therefore cannot be interpreted without specifying the candidate pooling protocol.

Elena Bolshakova and Julia Mescheriakova. Extending Derivational Dataset with Multi-Root Words. (PDF)

Abstract

Lexical derivational resources provide word-formation information useful both for linguistic studies and for NLP tasks, including deriving meaning of new and rare words. The paper describes a rule-based method to automatically enrich the recently constructed dataset containing derivational trees with single-root Russian words. The proposed method extends the derivational trees by multi-root words taken from the dataset of Russian words segmented into morphs (roots and affixes: prefixes, suffixes, endings, postfixes). To recognize derivational relations between words, the method accounts for affixation as the core of Russian word formation and involves detecting intrinsic node in the derivational tree, which is to be parent word for multi-root to be inserted into the tree. Preliminary statistics on the extended derivational trees and scores of their correctness are reported.

[SHORT] Elena Sidorova, Irina Akhmadeeva, Irina Kononenko, Alexey Sery and Yury Zagorulko. Constructing Multi-Level Datasets for Information Reliability Assessment in Popular Science Discourse. (PDF)

Abstract

This paper discusses issues related to assessing the reliability of information in popular science texts. Popular science texts are characterized by a casual style of presentation, loose references to primary sources and expert opinions, and the inclusion of controversial statements and facts. Clarifying these elements in the author’s arguments can significantly enhance the credibility of the text among critical readers. Alternatively, it can help refute the arguments put forward by the author. This study proposes a system for evaluating the reliability of information and the sources used by authors to support their claims. The main components of this system include modules for analyzing arguments, generating critical questions about the presented evidence, searching for facts from scientific sources on the Internet, evaluating sources, and verifying the accuracy of the information. Special attention is paid to contradictory information.The evaluation is conducted using a probabilistic trust model that takes into account various factors: the reputation of the source, the timing and type of information provided, and the degree of ambiguity in the information extracted using language models. The issue of preparing datasets for experiments in this area is also discussed. A three-tiered approach to data collection and annotation is proposed, involving both manual annotation by experts and automated procedures for data collection and synthesis.

SESSION 4. Organization of Experiments and Analytical Workflows.

Dmitry Kovalev. Планирование виртуальных экспериментов с повторным использованием вычисленных фрагментов. (PDF)

Abstract

В работе решается задача экономии вычислений при повторных запусках виртуального эксперимента, то есть проверки конкурирующих научных гипотез с помощью вычислительных конвейеров. Исследователь проверяет не одну гипотезу, а десятки альтернатив, и каждую прогоняет на множестве конфигураций. Полное повторение эксперимента после каждой правки обходится слишком дорого. Часть результатов можно взять из кэша. Однако при изменении одной гипотезы устаревают и все, которые на неё опираются: пересчёты распространяются каскадом. Предложены модели и алгоритмы, которые по графу зависимостей между гипотезами заранее определяют, какие гипотезы обязаны быть пересчитаны, а какие можно безопасно взять из кэша. Перед планированием корректность эксперимента автоматически проверяется по онтологии предметной области. Планировщик работает за время, линейное по размеру графа зависимостей. Главное отличие от существующих средств (Nextflow, Pegasus) в том, что единицей пересчёта служит гипотеза как научное утверждение, а не файл или отдельная задача; доказано, что полученный план минимален: меньше пересчитать нельзя. На 171 реальном потоке работ план ускоряет пересчёт в 1,9 раза относительно полного перезапуска при заполненности кэша 70 % и в 3,9 раза при 90 % (кэш моделируется равномерной случайной выборкой); накладные расходы планирования пренебрежимы. Проверка на вручную построенных графах гипотез nf-core подтверждает перенос оценок на уровень научных утверждений.

Ilya Derezovskiy and Alexey Mitsyuk. Visualization of Decision-Making Processes under Uncertainty. (PDF)

Abstract

This paper proposes an approach for visualizing decision-making processes under uncertainty in dynamic environments, integrating an interactive spatial simulation, a decision tree, and a visualization of decision-making processes that combines events, actions, and object states within a single visualization while preserving causal relationships between them. To account for uncertainty, our approach incorporates simulation modeling with interval-based parameters and multiple execution runs, allowing the analysis of variability in outcomes under different environmental conditions. The proposed approach is implemented as a web-based application that allows emergency planners and simulation researchersto construct simulation scenarios, define response strategies, and analyze their effectiveness based on both qualitative patterns and quantitative metrics. The applicability of the proposed approach is demonstrated through a warehouse fire scenario, where alternative response strategies are compared under uncertainty.

[SHORT] Nikita Ilin, Aleksey Artamonov and Mikhail Ulizko. From Static DAGs to Dynamic Computational Graphs: A New Paradigm for Orchestrating Analytical Processes in Data-Intensive Domains. (PDF)

Abstract

This paper examines the evolution of workflow orchestration in intelligent data analysis systems: from the classical model of a static directed acyclic graph toward dynamic computational graphs and large language model based components. It is argued that a static DAG remains an effective instrument for predictable, regulated data pipelines, but becomes limiting wherever structural uncertainty, next step policy, per step cost, and context and state management are decisive. Conditional branching in batch oriented orchestrators is formally reducible to models with discrete states and transitions. Nevertheless, this does not erase the distinction between design time branches and scenarios with new step types, learnable or model based routing, and a tight human in the loop or agentic feedback loop. Drawing on workflow management, formal models (Petri nets, statecharts, finite state machines), agentic workflow research, and engineering practice, static, dynamic, and hybrid approaches are compared. A hybrid architecture is motivated: a DAG backbone for stable stages and dynamic subgraphs for adaptive branches, quality assurance cycles, and AI agent orchestration. The argument is grounded in an illustrative IDAS case for scientific and technical literature, contrasting deterministic catalog collection with structurally uncertain metadata enrichment. Selection criteria, deployment risks, and mandatory stop policies are outlined.

SESSION 5. Speech and Signal Processing.

Dmitrii Popov and Archil Maysuradze. Uncertainty-Aware Ensemble Pseudo-Labeling for Speaker Diarization. (PDF)

Abstract

Speaker diarization — the task of partitioning an audio recording into speaker-homogeneous segments without prior knowledge of the speakers — remains a difficult problem in speech processing. One of the practical limitations of modern end-to-end diarization systems is their dependence on accurately labeled training data, which is expensive to obtain at scale. In this work, we study a pseudo-labeling approach based on weighted aggregation of several diarization models and recording-level uncertainty estimation. The ensemble produces multi-label pseudo-labels that explicitly represent overlapping speech, and recordings with high ensemble entropy are removed before training the target model. We evaluate the method with a TS-VAD-based student model on VoxConverse, DIHARD, and AMI. The results indicate that training on filtered pseudolabels improves over a supervised baseline, especially when additional unlabeled data are available. An ablation study suggests that a small ensemble already provides most of the benefit, while uncertainty-based filtering gives a stable reduction in DER across the evaluated datasets.

Mikhail Vdovichenko and Rustam Azimov. Multilingual Text-to-Speech Model for Low-Resource Languages. (PDF)

Abstract

This paper studies multilingual text-to-speech synthesis for Turkic languages under data scarcity and severe corpus imbalance. The experiments cover Turkish, Uzbek, Kazakh, Chuvash, Bashkir, and Tatar. The proposed approach combines joint training of a discreteaudio-token generation model with a weighted loss that compensates for differences in language corpus sizes. Quality is evaluated with both automatic metrics and the subjective PSER metric. We first select an effective training ratio for the two largest corpora, Turkish and Uzbek, and then study how multilingual data affect both neural audio codec reconstruction and final speech synthesis quality. The analysis also compares language-specific and joint codec configurations as component-level ablations. The final multilingual model significantly improves PSER over the production Uzbek system, reaches statistical parity with the production Turkish, Chuvash, and Tatar systems, and underperforms the production Kazakh and Bashkir systems while retaining acceptable synthesis quality. Error analysis shows that the most difficult inputs contain numbers, dates, and special symbols, which indicates that text normalization remains a key direction for further improvement.

[SHORT] Victor Nozdrin, Denis Karchkov, Tatiana Levanova and Lev Smirnov. Generation of synthetic ECG of 5 main nosologies using WGAN. (PDF)

Abstract

The availability of sufficiently large and diverse annotated datasets remains one of the main limitations in the development of deeplearning methods for electrocardiogram analysis. In this work, we propose a class-specific framework based on a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) for generating 10-second, 12-lead ECG records corresponding to five rhythm classes: sinus rhythm, sinus bradycardia, sinus tachycardia, sinus arrhythmia, and atrial flutter. The generator is trained using feedback from frozen segmentation networks that describe the P–QRS–T structure and, for atrial flutter, F-wave activity. The generated signals were evaluated using embedding-based Fréchet distance (FD-ECG), Wasserstein distance between RR-interval distributions, spectral metrics, and a segmentation-derived Similarity Score. Across the five complete models, FD-ECG ranged from 10.06 to 48.03, RR Wasserstein distance from 0.021 to 0.138 s, and the Similarity Score from 0.894 to 0.960. For sinus rhythm, the complete model achieved lower spectral error and a substantially higher Similarity Score than clean WGAN-GP. In contrast, the clean baseline obtained slightly lower FD-ECG and RR Wasserstein distance. The downstream utility of the generated ECGs was assessed using an XResNet1D101 classifier. Adding 600 synthetic records to each class increased test accuracy from 0.8280 to 0.8375 and macro-F1 from 0.8272 to 0.8311. However, the overall difference in correctness was not statistically significant according to the exact McNemar test (p = 0.2018), and the effect was class-dependent. These results indicate that synthetic ECG augmentation can provide a moderate practical benefit, but its influence depends on the diagnostic class and the evaluated signal property.

SESSION 6. Recommender Systems and Scientific Knowledge Services.

Alice Novikova and Dmitry Ignatov. Ranking-Aligned Contrastive Consistency Distillation for One-Step Diffusion-Based Sequential Recommendation. (PDF)

Abstract

Diffusion models set a new quality standard for sequential recommendation, but pay for it with iterative inference that is poorly compatible with the latency budgets of online deployment. We propose Ranking-Aligned Contrastive Consistency Distillation (RCCD), a procedure that transfers the quality of a multi-step latent-diffusion teacher into a single-step student. RCCD combines discrete-time consistency distillation along the teacher’s reverse trajectory, a zero-initialised residual refinement head over the teacher’s frozen backbone, the teacher’s own catalogue-level cross-entropy, and an in-batch InfoNCE term aligned with the directional geometry of inner-product retrieval. On Amazon Toys, Amazon Beauty and MovieLens-1M the single-step student reaches 100.9–110.7% of full-trajectory HR/NDCG, whereas truncating the same teacher to one step retains only 48.7–52.1%. The contrastive term is isolated under a fixed architecture and improves the model on all 18 seedmatched contrasts, with every 95% interval excluding zero. Against fulltrajectory diffusion recommenders the one-step student leads on Toys, tracks the strongest baseline within 0.11 HR@10 points on Beauty, and trails DimeRec by 1.49 HR@10 points on MovieLens-1M. End-to-end perrequest latency is 32.1–156.8× lower than the teacher’s; we decompose this factor and show that on the short-history datasets it is attributable entirely to the 32× reduction in function evaluations rather than to a cheaper per-step network.

Vadim Chunikhin, Andrey Khalov and Olga Ataeva. A Scientific Knowledge Recommender System for the SciLib Intelligent Information System. (PDF)

Abstract

This paper proposes an approach to building a scientific knowledge recommender system within the SciLib intelligent information system. The relevance of this work is determined by the rapid growth in the volume of scientific publications and the increasing complexity of relationships among documents, authors, citations, methods, concepts, and mathematical objects. The proposed solution is intended for automated content analysis of scientific document collections, extraction of text and structured metadata, identification of concepts and fine-grained units of scientific knowledge, their normalization using ontologies and graphbased representations, and generation of explainable recommendations for reading, writing, reviewing, and editorial and publishing workflows for scientific materials. Particular attention is paid to multilevel recommendation, in which the objects of analysis include not only publications and their metadata but also formulas, statements, lemmas, theorems, methods, key concepts, and the contexts in which they are used. Several types of embeddings are employed to represent documents in the search and recommendation pipeline, reflecting the document’s core content, title, keywords, and metadata. The Mathematical Information Retrieval Benchmark (MIRB) is used to evaluate mathematical retrieval as one component of the recommender system. A recommendation is formalized as an explainable chain of transitions over a typed graph of objects in the scientific ecosystem, while an experimental evaluation of the recommender’s textual branch on the MIRB NumberTheoryPremiseRetrieval dataset establishes a reproducible baseline and demonstrates that textual similarity alone is insufficient for reliable premise retrieval.

Evgenii Malyh, Andrey Khalov and Olga Ataeva. An Ontological Approach to the Design and Validation of Microservice Interactions. (PDF)

Abstract

This paper addresses the problem of designing a microservice architecture under a growing number of services, increasingly complex inter-service links, and the emergence of intelligent components. The study is carried out as part of the development of the SciLibAi system, which is characterized by heterogeneous data formats, dynamic interactions, and AI agents that use the Model Context Protocol (MCP). Requirements are formulated for describing services, messages, events, APIs, contracts, roles, and dependencies, and the tasks and evaluation criteria of the architectural approach are defined. The limitations of existing approaches to describing distributed systems are considered using the example of the IEEE Std 1516 High Level Architecture standard, which is oriented primarily toward modeling and simulation but does not address the semantic description of a microservice architecture, the validation of message contracts, or the runtime control of interaction admissibility. As a solution, an ontological approach is proposed that formalizes the semantics of a microservice system, links message contracts to the architectural model, describes the roles of components, and specifies rules for admissible data exchange. A solution architecture based on a GraphQL gateway, an application example in the SciLibAi system, and a visualization of the elements of the ontological model are presented. Prior to generating the GraphQL schema and validation rules, SHACL validation of the ontological model is performed, which makes it possible to detect inconsistencies and incomplete descriptions before integration; the article-processing scenario demonstrates a reduction in the number of manual contract-coordination points as services evolve.

WEDNESDAY,, OCTOBER 21th

SESSION 7. Applications of Large Language Models.

Bulat Latypov, Andrei Pavlov, Andrei Petrovski and Igor Kotenko. Automated Financial Intelligence Pipeline: LLM-Powered Extraction and Analysis of Russian Corporate Disclosures. (PDF)

Abstract

This paper presents an automated pipeline for the collection, preprocessing, and LLM-based analysis of corporate financial disclosures from Russian public companies. The system integrates three Scrapy-based web spiders—targeting the Yandex Investor Relations portal, Gazprom IR, and the mandatory e-disclosure.ru regulatory filing platform—with a Go backend for data ingestion and a DeepSeek-powered analytical module for structured information extraction. The pipeline ingested 16,453 documents and performed full LLM analysis on 508 filings, generating sentiment scores, category classifications, and professional-grade abstractive summaries. Average processing latency was 4.85 seconds per document. In a manual field-level evaluation, the key parameters produced by the LLM were compared with reference values obtained from the corresponding source documents. We describe the system architecture, engineering trade-offs, and empirical findings, including a persistent neutrality bias in sentiment scoring attributable to the formal register of regulatory filings. The pipeline demonstrates the feasibility of scalable, automated financial monitoring using large language models.

Dmitry Morozov and Aleksandr Vetrov. Beyond Root Matching: LLM-Based Identification of Russian Same-Root Word Pairs. (PDF)

Abstract

The identification of same-root words is central to Russian morphemics, derivational lexicography, and orthography, yet it remains difficult to formalize because of root alternations, homonymous roots, loanwords, and competing synchronic and diachronic criteria. In this paper, we study whether modern Large Language Models can support the semi-automatic expansion of Russian same-root dictionaries. We formulate the task as word-pair classification and construct a balanced benchmark from the morphemic dictionary used in the Russian National Corpus. The dataset distinguishes true same-root pairs, pairs with homonymous roots, and unrelated pairs, making the task more challenging than simple root-string matching. We compare prompting modern LLMs — Gemini 3.1 Flash Lite, GPT 5.4 Mini, GLM 5.2, and DeepSeek V4 Flash — with two baselines: a rule-based morphonological filter and a hybrid system combining this filter with a fine-tuned ruBert-base classifier. The rule-based filter achieves high recall but low precision, while the hybrid baseline reduces false positives at the cost of many additional false negatives. Gemini 3.1 Flash Lite achieves the highest accuracy and F1-score among the tested methods, reaching 91.6% accuracy and 87.4% F1-score. Error analysis shows that most Gemini’s errors occur at the boundary between lexicographic root identity and broader etymological relatedness. Our results suggest that LLM prompting is a promising tool for candidate generation in same-root dictionary expansion, although expert verification remains necessary.

[SHORT] Sofia Maltseva and Konstantin Turalchuk. Research into Approaches to Using LLM and RAG to Improve Intelligent Assistants in the Field of Medicine. (PDF)

Abstract

Использование LLM в диалоговых системах, в которых предоставляется ответ на медицинские вопросы пользователей, сопряжено с рисками возникновения галлюцинаций в ответах. Ошибки и неточности могут нанести вред здоровью пациентов. Для снижения вероятности возникновения ошибок используются различные подходы верификации ответов LLM. Глобально их можно разделить на три категории: дополнительное обучение LLM, промпт-инжиниринг и RAG. На основе обзора технологий выбраны шесть LLM, два подхода RAG и метод промпт-инжиниринга CoT и проведено экспериментальное тестирование. Результаты тестирования показали, что на наборах медицинских вопросах лучше всего справляется LLM GigaChat3.1-10B-A1.8B в сочетании с CoT и методом RAG, предполагающим сортировку документа перед формированием контекста в промпте. Эти достижения легли в основу прототипа информационной системы в виде вебприложения, в котором пользователи в формате чата получают от интеллектуального ассистента ответы на медицинские вопросы.

[SHORT] Алёна Антипова and Елена Шамаева. Применение больших языковых моделей в задаче упрощения русскоязычных текстов. (PDF)

Abstract

Автоматическое упрощение текстов представляет собой важную задачу обработки естественного языка, имеющую практическое значение для повышения доступности образовательных, научно-популярных и юридических материалов. Несмотря на активное развитие методов на основе больших языковых моделей, область упрощения русскоязычных текстов остаётся изученной недостаточно. В данной работе проводится систематическое исследование применимости больших языковых моделей семейства Qwen3 к задаче упрощения русскоязычных текстов различных жанров. Анализируется влияние трёх факторов: адаптации модели к русскому языку, количества её параметров и стратегии промптирования. Оценка качества упрощения выполняется комплексно и включает классические метрики (индексы удобочитаемости, SARI, BARTScore), а также, впервые для данной задачи, статистические метрики на основе распределений синтаксических триплетов (дивергенция ЙенсенаШеннона и энтропия Шеннона). Эксперименты проведены на корпусах литературных произведений и энциклопедических текстов датасета RuAdapt. Результаты показывают, что выбор оптимальной модели существенно зависит от стратегии промптирования. Установлено, что в среднем наиболее высокие результаты демонстрирует Qwen3-4B. Синтаксические триплеты, как частный случай синтаксических n-грамм, подтверждают свою применимость к оценке качества упрощения. Разработанная в данной статье система автоматизации исследований по оценке упрощения текста большими языковыми моделями опубликована по адресу https://github.com/WhiteWhiteFoxX/ru-text-simplification.

[SHORT] Grigory Drozhashchikh and Valentin Malykh. LLM-Based Stance and Argument Classification for Russian. (PDF)

Abstract

Large language models are widely used on English corpora for downstream argument mining tasks, but research dedicated to the Russian language remains limited, and systematic studies applying large language models to this field are only beginning to emerge. We use two large language models (Gemma 4 31B IT and Gemini 3.1 Flash Lite) to assess zero-shot and two-shot classification with and without Chain-of-Thought (CoT), on a corpus annotated for stance detection and argument classification simultaneously: the professionally translated ArgMicro corpus and the natively Russian RuArg-2022 comments corpus. On RuArg-2022 our best configuration reaches 0.7379 macro-F1 on stance detection, above the 0.6968 obtained in the benchmark paper, while on argument classification it reaches 0.5874 against 0.7404. Chainof-Thought prompting changes the metric by +0.0204 on average for argument classification and by +0.0108 for stance detection on RuArg2022, where the effect is consistent across all configurations, whereas on ArgMicro it is mixed and not statistically significant. The sign of the effect therefore follows the task rather than the model.

SESSION 8. Stochastic Models and Optimization.

Valeria Kuznetsova and Elena Kuvykina. Investigation of the Conflict Flow Control System Using an Algorithm with an Informative Pedestrian Flow. (PDF)

Abstract

The paper is devoted to the study of a control system for three conflicting flows at an intersection. Two of the flows are traffic flows, and one is a pedestrian flow. The pedestrian flow is not merged with any of the traffic flows. The traffic signal controller acts as the server. The switching of the server modes occurs in accordance with the pedestrian arrival and service processes. Thus, this system is a feedback system. The numerical solution of the optimization problem for the control algorithm parameters is based on simulation modeling. The minimum average delays of an arbitrary request in the system are chosen as the optimality criterion. The dependence of the optimal control parameters on the arrival rates and the probabilistic structure of the input flows has been studied.

Varvara Kovaleva and Alexander Ponomarenko. A Benchmark Generator of Realistic Logistic Transport Networks with Nonlinear Edge Costs and Transshipment Capacities. (PDF)

Abstract

Routing many commodities through a logistic network at minimum joint cost of vehicle movements and transshipment at intermediate storages becomes substantially harder—and substantially more realistic—once two features are present together: edge costs that are nonlinear in the transported volume, because freight travels in fixedcapacity vehicles, and hard upper bounds on the volume reloaded at each storage. No public benchmark captures this combination, and the locations of commercial distribution centres are proprietary. We close this gap with a parameterised generator of synthetic logistic networks, calibrated on real road graphs of nine European countries, five U.S. states and the European part of Russia extracted from OpenStreetMap. The generator reproduces the structural signatures of upper-level distribution networks—near-planar topology, low vertex degree, short edges and large diameters—by combining Zipf-distributed city sizes, a spatial-network budget model trading edge-building cost against routing convenience, centrality-driven storage capacities, and a doubly constrained gravity demand model. We release 35 fully synthetic instances (10–150 storages) and 15 real-geometry instances in an open, self-describing, fully reproducible file format. An empirical characterisation computed directly from the released files shows that synthetic and real-geometry instances are structurally consistent (two-sample Kolmogorov–Smirnov distances of 0.10 on degree and edge-length distributions, with small and quantified mean deviations), and a baseline-solver study demonstrates that the corpus discriminates instance difficulty across orders of magnitude. The benchmark is a faithful, reusable testbed for the development and comparison of exact and heuristic algorithms.

Alexander Ponomarenko and Iliya Antonov. Tuning Dispatch Thresholds for Fixed Last-Mile Routes: A Simulation-Based Pareto Analysis of a Production Policy. (PDF)

Abstract

Many parcel networks dispatch vehicles on fixed routes using a simple load-accumulation rule: a truck leaves the depot for a fixed route as soon as the volume (or item count) waiting for that route crosses a threshold. The threshold is usually parameterised as an affine function of route length, τr = β + γ dr, and the pair (β, γ) is chosen once and frozen into production. This paper studies how good that frozen choice actually is, treating the question as a data-intensive, data-driven decision-making problem over a full month of real operational flow. Using a discrete-event simulator that replays the recorded arrival stream and reconstructs every trip, we sweep the (β, γ) design space, evaluate the two competing objectives—company operating cost and average parcel lead time—and recover the Pareto frontier of efficient policies for two deployed variants (volume-triggered and item-count-triggered). The two policies turn out to be in strikingly different states of tune. The volume-threshold configuration lies on its own Pareto frontier: the simulator finds no (β, γ) pair that strictly dominates it, so the deployed policy is already Paretoefficient—an unusual positive audit result. The item-count configuration is the opposite: it is dominated by a concrete simulated configuration that is both faster and cheaper, and the available cost saving at equal lead time is about 5.0 %. We trace the item-count policy’s inefficiency to a base that is too large and a length coefficient that is too small for the deployed truck capacity, and show that a steeper threshold—lower base, higher slope—is preferable. Because the remedy is a two-scalar reconfiguration, the analysis converts directly into an actionable, zero-capital recurring saving.

[SHORT] Andrei V. Zorine. On estimating the shape of periodic intensity in a Poisson process. (PDF)

Abstract

The short paper examines a non-homogeneous Poisson process with a periodic instantaneous intensity function and presents work in progress. It considers the positions of jump times within a period, i.e., the remainders of the jump times modulo the period. It is shown that the resulting sequence of random variables forms a Harris positive recurrent general Markov chain on an interval. It is established that the stationary distribution density is proportional to the intensity function. This makes it possible to establish the asymptotic normality of the sequence of kernel density estimators of the stationary density, to prove the convergence of the finite-dimensional distributions of the empirical process to a Brownian bridge, and to justify the application of Pearson’s chi-square goodness-of-fit test. The main limit theorems are proved using the central limit theorem for martingales.

SESSION 9. Large Language Models for Software Engineering.

Wang Zihan and Rustam Azimov. Graph-Orchestrated LLM Agents for Automated Configuration of Reproducible Python Execution Environments. (PDF)

Abstract

Configuring executable environments for unfamiliar opensource repositories remains difficult and error-prone: metadata are often incomplete, dependency constraints inconsistent, and failures appear only once the environment is executed. This paper presents a graphorchestrated LLM agent framework for repository analysis, executionenvironment repair, and Dockerfile synthesis in Python projects, combining LangGraph orchestration, a ReAct-style reasoning loop, structured test feedback, Docker isolation, and container snapshot rollback. Rather than relying on static declarations or unstructured error interpretation, it represents execution as a controlled state-transition process in which planner, tool, and observer nodes refine the environment until test collection succeeds or a terminal failure is reached. On 98 public GitHub Python repositories with heterogeneous project types and dependency structures the system configured 91.8% successfully, with every successful case re-verified by rebuilding the generated Dockerfile from scratch. This exceeds our rerun of Repo2Run on the same benchmark, although the confidence intervals overlap, and is far above the static and directgeneration baselines reported in prior work. The results suggest that explicit state modeling and execution-feedback-driven routing improve the reliability of LLM agents for reproducible environment construction.

Artur Ziazetdinov, Vladimir Zunin and Pavel Pashentsev. Reproducible Pipeline for Evaluating RAG-Based Verilog RTL Generation. (PDF)

Abstract

This study examines Verilog RTL generation from naturallanguage specifications with large language models (LLMs) and retrievalaugmented generation (RAG). The main analysis covers 76 tasks, six models, two RAG conditions (RAG-off and RAG-on), and ten attempts for every task, model, and condition combination. Empty responses and request failures are included among the attempts. The resulting comparison therefore characterizes the generation and verifiation process rather than the quality of successfully produced RTL code alone. Results are evaluated by the proportion of tasks solved in at least one of ten attempts. Functional pass@10 requires an explicit passing testbench verdict, whereas pipeline pass@10 additionally requires syntax checking, linting, and synthesis of the same candidate. The maximum pipeline pass@10 reached 46.1% for GPT-5.4-nano without RAG. For Qwen3- Coder-30B, the value increased from 17.1% to 21.1% after adding fragments from a mixed knowledge base (mixed-RAG); several other models showed zero or negative changes. The results concern the selected tasks and knowledge-base configuration and do not establish a universal improvement from RAG.

[SHORT] Vadim Voevodkin. Call-Graph Augmented Three-Stage Pipeline for Repository-Level Code Search. (PDF)

Abstract

Repository-level code search—locating the exact function or method that implements a given developer intent—is a core capability required by AI coding assistants, automated debugging tools, and code review systems. Existing benchmarks evaluate code retrieval at the dataset level (e.g., CodeSearchNet, COIR) and systematically ignore the rich structural information encoded in a program’s call graph. We present a three-stage retrieval pipeline that combines (1) a bi-encoder for dense candidate retrieval, (2) call-graph neighbor expansion via AST analysis, and (3) cross-encoder reranking. To support evaluation we contribute a new benchmark of 601 intent-style questions with ground-truth line annotations across nine real-world open-source Python repositories (42 to 27,849 functions). On hard questions—where the relevant snippet is rarely in the dense top-N—AST expansion raises final Hit@5 from 71.7% to 77.7% (+6.0 pp). For large repositories the gain is larger: infiniflow/ragflow improves by +16.7 pp and vllm-project/vllm by +8.8 pp on hard questions. We analyse when the reranker discards AST-rescued snippets and derive practical guidance for call-graph hop depth and corpus size.

SESSION 10. Time Series Analysis and Forecasting.

Vyacheslav Manevich. Метод тройной поправки (TCM) в задачах многошагового прогнозирования финансовых временных рядов. (PDF)

Abstract

Многошаговое прогнозирование финансовых временных рядов остаётся фундаментально сложной задачей из-за экспоненциального накопления ошибок и низкого отношения сигнал/шум, характерного для рыночных данных. В настоящей работе проводится систематическое сравнение авторского метода тройной поправки (TCM) с ведущими эталонными моделями — StatsForecast, Nixtla (TimeGPT), XGBoost, сверточными нейронными сетями и Prophet — на двух принципиально разных типах данных: средние дневные цены закрытия (ACP) и реализованная волатильность (RV). Исследование охватывает 89 активов из 13 рыночных категорий за период 2018–2023 гг. и опирается на строгий протокол скользящего окна (400 дней обучения, 30 шагов прогноза) с применением двух взаимодополняющих стратегий ранжирования — подсчета побед в окнах и метода Коупленда. Результаты показывают, что TCM является высокоэффективным “спринтером”: на коротких горизонтах (1–7 шагов) он уверенно опережает XGBoost и приближается к лидерам — StatsForecast и Nixtla, а при прогнозировании волатильности удерживает преимущество над бустинговыми алгоритмами на всей 30-шаговой дистанции. Однако на среднесрочных горизонтах (13+ шагов) накопление ошибки приводит к его отставанию от StatsForecast и Nixtla. Установлено, что рекурсивный подход для TCM оказывается устойчивее индивидуальных моделей, поскольку последовательный перенос адаптивных коэффициентов лучше сохраняет внутреннюю структуру ряда. С вычислительной точки зрения TCM, XGBoost и StatsForecast укладываются в 5 секунд на актив, что делает их пригодными для практических приложений реального времени. Полученные результаты определяют границы применимости TCM как точного и быстрого инструмента краткосрочного прогнозирования

Anton Voskresenskii, Gulnara Siraeva and Dmitry Kovalev. CL-Occam: Clustered-Local Forecasting with Per-Cluster Model Selection for Heterogeneous Time-Series Panels. (PDF)

Abstract

Forecasting a panel of related time series poses a dilemma. A single global model, trained on all series at once, smooths away local behaviour; a separate model per series keeps local detail but loses cross-series information and overfits. CL-Occam takes the middle road. It groups the series by Dynamic Time Warping shape similarity and fits one model per group. Within each group, cross-validation picks between CatBoost and Ridge, with a bias towards the simpler model on nearties. We evaluate this reproducible pipeline on four open daily panels, at horizons of one week to one month, against seven baselines including per-series models and entity-ID-augmented globals. On the three heterogeneous panels (electricity load, ATM cash, retail sales), CL-Occam reduces the median per-series error of every CatBoost-based global model, and every gain passes Holm-corrected significance testing. The gain grows with the shape heterogeneity of the panel. Against a linear global model the picture is mixed. On a homogeneous weather panel the sign reverses, as predicted. Adding exogenous regressors on the retail panel does not remove the gain: clustering still helps on top of strong features.

[SHORT] Irina Kudryashova. Forecasting Life Expectancy in Russia Using Socio-Economic and Health Indicators: Comparing Linear Regression, Principal Components, and Machine Learning Models. (PDF)

Abstract

Life expectancy is a compact indicator that reflects longterm changes in public health, economic conditions, and social well-being. Its dynamics are driven by heterogeneous and strongly correlated factors, which makes direct regression on raw indicators unstable, especially for small country-level time series. This paper studies annual data for Russia and compares three groups of predictive approaches: direct linear regression on selected raw features, principal-component-based regression, and machine learning models. The dataset combines demographic, medical, and socio-economic indicators derived from WHO, World Bank, UNDP, and Rosstat sources and covers the period from 2000 to 2022. We use a temporal split with 2000-2015 for training and 2016-2022 for testing. The baseline model is built after preliminary screening by descriptive statistics, correlation analysis, and variance inflation factors. The component-based stage uses principal component analysis to mitigate multicollinearity and builds a regression model on the selected components. The machine learning stage includes Ridge, Lasso, Elastic Net, PLS regression, and Random Forest. The best predictive quality on the test period is achieved by Elastic Net (RMSE = 0.453, MAE = 0.334, MAPE = 0.468%), followed closely by Lasso. A regression on two principal components also provides a strong and interpretable result (RMSE = 0.628), while direct regression on raw predictors performs notably worse on the test interval. The results show that for a short, highly collinear annual series, regularized linear models offer the best balance between predictive accuracy, robustness, and interpretability.

[SHORT] Marwan Bouabidi, Dmitry I. Ignatov and Robiul Islam. Brain Activity Time Series Analysis with Neural Networks and Interpretable Machine Learning. (PDF)

Abstract

This study investigates different deep learning architectures for classifying brain states based on Electroencephalography (EEG) signals recorded during visual perception tasks involving varying brightness levels of two visual stimuli, the Necker cube and the Mona Lisa. The raw EEG data was separated into five standard frequency bands (delta δ, theta θ, alpha α, beta β and gamma γ) with a 4th order Butterworth filter, each band representing a separate dataset for a given visual stimuli. We evaluated a standard Multilayer Perceptron (MLP), a novel hybrid Recurrent Neural Network (RNN) incorporating a convolutional front-end and attention mechanisms, and a CNN-Transformer model based on EEGNet and the Transformer model. The continuous EEG data was processed in short temporal sequences, containing 400 ms worth of data. We compared the models’ performances using standard quantitative metrics (accuracy, F1, precision, recall and training time). For interpretability, we employed the SHAP technique to identify the most influential EEG channels. Results showed clear differences in model performance across the five frequency bands. The hybrid RNN and CNN-Transformer performed significantly better, achieving higher accuracies almost across all bands, and improved generalisation by limiting overfitting. The SHAP analysis revealed that different models focused on distinct brain regions and activity patterns.

SESSION 11. Reliability of Neural Networks.

Ivan Mineev, Andrey Lebedev and Sergey Stasenko. Investigating Separation Methods in High-Dimensional Spaces for Image Classification Error Correction. (PDF)

Abstract

This study investigates the use of separation methods in high-dimensional feature spaces for correcting image classification errors without retraining the base neural network. The work focuses on detecting a previously unseen class in the latent space of a ResNet-18 model trained on the CIFAR-10 dataset with one class excluded. Feature representations were extracted from an intermediate layer of the neural network and reduced using Principal Component Analysis. A simplified AI-corrector architecture was then implemented, in which a single global corrector distinguishes between known classes and a target unknown class. Several linear and nonlinear separation methods were compared, including Linear Discriminant Analysis, LinearSVC, logistic regression, Random Forest, XGBoost, and k-nearest neighbors. The experimental results show that the performance of most methods is relatively close, with logistic regression and other linear methods often achieving results comparable to more complex nonlinear algorithms. This suggests that the structure of the latent feature space formed by the neural network plays a more significant role than the specific choice of separator. However, confusion matrix analysis revealed that correcting unknown-class predictions can also degrade performance on previously recognized classes. The study concludes that single global correctors are simple and practically applicable, but further improvements may require multicorrector architectures and error-clustering strategies.

Sergei Vorobyov and Eugene Ilyushin. Scaling Neural Network Verification with Tensor Parallelism and Fully Sharded Data Parallelism. (PDF)

Abstract

GPU memory limits the size of networks that formal verification can handle, and the usual assumption is that the weight matrices are what has to be split across GPUs. We tested this assumption in auto_-LiRPA / α, β-CROWN. We added two sharding methods from large-model training and measured where the memory actually goes. Tensor Parallelism (TP) splits the weights together with the CROWN A-matrices. On 2 GPUs it lowered peak memory by 1.97× and made propagation 1.90× faster; on 4 GPUs the factors were 3.83× and 3.49×. Its bounds are exact in the first sharded zone and get looser in later zones, which use IBP. Fully Sharded Data Parallelism (FSDP) splits only the weights. Its bounds match the single-GPU bounds bit for bit, but peak memory grew by 26–34% and propagation took up to 3.2× longer. Copying the weights from host memory one layer at a time needed less memory and less time than FSDP. A per-node memory trace explains this. At h=4096 the A-matrices take 5.6 times more memory than the weights. In β-CROWN with Branch-and-Bound the peak grows linearly with the number of subdomains B, as 21.3+ 0.372 B MB, and weights are 0.13% of it. We therefore expect more from splitting the subdomains across GPUs, and estimate 45.0% less memory per GPU at P=2 and 67.5% at P=4.

[SHORT] Andrey Lebedev, Sergey Stasenko, Alina Lebedev and Sofia Polevaya. Post-hoc correction of errors in the RR interval classifier when transferring to data with extrasystolic distortions. (PDF)

Abstract

This paper evaluates an external post-hoc layer for correcting the decisions of a frozen RR-interval classifier of emotional maladjustment on RR data from patients with extrasystoles. The experiment uses a reproducible RR-window gradient-boosting proxy inspired by previous work [1]; it is not presented as a reproduction of the unavailable serialized dynamic spectrum model. After source cohort training, the proxy is frozen and only the external corrector is fitted on target development folds. The target cohort comprised 53 patients (48 class 1 and 5 class 0). Five-fold patient-level cross-fitting produced one out-of-fold prediction per patient. The frozen proxy predicted every target patient as positive: balanced accuracy 0.500, sensitivity 1.000, specificity 0.000, and macroF1 0.475. The linear probabilistic corrector had the highest exploratory balanced accuracy, 0.723, with paired ∆BA 0.223 (95% bootstrap CI [−0.019, 0.372]). Fully nested automatic selection gave balanced accuracy 0.523 and paired ∆BA 0.023 (95% CI [−0.207, 0.313]). For every method with a positive ∆BA estimate, the paired 95% CI included zero; the results support technical feasibility but not a statistically reliable clinical improvement.

[SHORT] Andrey Kozachok, Anatoliy Bakaev, Aleksandr Kozachok and Artem Noev. Experimental Analysis of AutoML Platform Vulnerabilities in Environmental Monitoring Tasks and Risk Mitigation Strategies. (PDF)

Abstract

This paper presents an exploratory comparison of the AutoML platforms Auto-sklearn, FLAML, and AutoGluon in regression tasks on the environmental datasets Beijing, UCI, and AURN. A reproducible pilot stress-test protocol is used to examine how data quality violations (noise and missing values) and computational constraints affect the models obtained in specified runs. Adding 5% noise in the FLAML scenario on the Beijing dataset increased RMSE by 165%, reducing its search time from 10 to 1 minute increased RMSE by 34%, and combining noise with a constrained budget produced outcomes that differed from either isolated condition, including a change of the selected model. Because each reported value comes from a single run with a fixed seed, these results are descriptive and do not establish statistically significant differences or a robustness ranking among platforms. They nevertheless demonstrate the practical value of evaluating AutoML systems under representative data defects in addition to reporting a baseline metric.

[SHORT] Нина Родикова and Борис Добров. Methods for Detecting Hallucinations in Large Language Models Summarization. (PDF)

Abstract

This work is devoted to studying the problem of automatically detecting hallucinations of large language models in English texts for the task of text summarization. As part of the study, several modern methods for classifying texts for the presence of hallucinations were implemented: text comparison using a knowledge graph, a method based on elementary judgments, a method based on embedding context similarity, the LLM-as-a-judge approach, a method based on fine-tuning an NLI model, as well as a newly developed combined method. The best results were shown by the combined method with a CatBoost gradient boosting meta-model (quality across all metrics exceeds 90%) and the method with fine-tuned RoBERTa model (quality across all metrics exceeds 90%). The LLM-as-a-judge approach achieved average quality (approximately 60-70% across various metrics). The remaining methods showed low performance on one of the classes (F1-score in the range of 35-40%).

SESSION 12. Educational Data Analysis.

Sergei Gorshkov, Dmitry Ignatov and Julia Aleksandrova. Beyond Aggregation: Validating a Student Digital Profile with Repeated Split-Half Reliability and Multimodal Structural Analysis. (PDF)

Abstract

Digital profiles built from social-media subscriptions are increasingly used as predictors in educational and behavioral analytics, yet they are rarely treated as measurements in a psychometric sense. We analyze a topic-based digital profile of 4,172 Russian university students (the open-profile subsample of the N = 6,276 students surveyed) from 24 institutions and put it through a validation pipeline aimed at psychological-values criteria. We replace per-topic features with nine content-aggregated indices and evaluate their reliability with a 100-fold pair-balanced split-half design. Aggregation lifts reliability from a noisy per-feature regime into a workable range, and the resulting indices remain predictive of psychological criteria. We then characterize the wider digital-psychological space with three multivariate techniques – CP tensor decomposition with permutation tests, checkerboard and contrastive sparse biclustering, and formal concept analysis – and recover interpretable thematic modules, local niches and lattice-level implications. The combined evidence reframes the digital profile not as a vector of standalone features but as a structured measurement space in which content-aggregated indices act as moderately reliable proxies for latent informational orientations. The pipeline is designed to be transferable to other socio-digital studies.

Ramil Khabibrakhmanov, Bulat Gizatullin, Olga Fedorova and Dmitry Polyakov. Comparative analysis of student representation methods for alumni-based retrieval and thesis topic suggestion. (PDF)

Abstract

We present a comparative analysis of student representation methods on a dataset of graduates of a Russian higher education institution. The methods cover four families: feature-based and semantic coursevector representations, autoencoder-based latent representations (basic, variational, sparse), triplet-loss metric learning, and Node2Vec over a bipartite student-course graph. Retrieval quality is measured against constructed thesis topic clusters using HitRate@K, MRR, NDCG@K, AP@K, Mean Cosine Similarity (MeanCos) and Semantic Precision; robustness is checked by varying the clustering used to construct the ground truth. We additionally examine how different retrieval methods behave in LLM-based thesis topic generation. Pure Sem Vec cosine gives the best overall retrieval performance, leading NDCG@5, HitRate@5, MRR and AP@5, while the SemVec Profile + Gower hybrid gives the best MeanCos and Sparse AE gives the best Semantic Precision. In the downstream evaluation, retrieved alumni context improves generation relative to generation without retrieval, but absolute scores remain low and the ranking of retrieval conditions depends on the relevance threshold. Student-level retrieval and generation scores are weakly associated, so strong retrieval performance does not automatically translate into proportionally better LLM-based thesis-topic generation under the evaluated setup.

[SHORT] Natalya Pustovalova and Tatiana Avdeenko. Psychometric-Based Learner Profiling for Educational Personalization Using PCA and Cluster Analysis. (PDF)

Abstract

This paper proposes a psychometric-based approach to learner profiling for educational personalization aimed at supporting adaptive educational systems under cold-start conditions. Using a dataset of students from the Novosibirsk State Technical University, multidimensional learner representations were constructed from psychometric characteristics. To improve clustering quality and interpretability, Principal Component Analysis (PCA) was applied prior to clustering to derive latent dimensions representing learner traits. Clustering quality was assessed using the Silhouette coefficient, Calinski–Harabasz index, and Davies–Bouldin index, resulting in the selection of a seven-cluster solution. Statistical validation confirmed significant differences among the identified profiles, with the strongest effects observed for emotional and self-regulation characteristics, including neuroticism, emotional intelligence, conscientiousness, behavioral inhibition, and achievement motivation. Based on the obtained profiles, personalized recommendations for educational content selection were developed. The proposed approach enables interpretable learner differentiation and supports recommendation generation at early stages of interaction with adaptive educational systems. This study focuses on constructing and interpreting learner profiles under cold-start conditions, as well as verifying the designed interventions against results obtained by other researchers. The next stage of the research involves testing a software product based on the proposed approach and empirically evaluating the effectiveness of the developed recommendations.

[SHORT] Alexander Varnavsky. Ontological Model of the Didactic Paradox “Success Without Value” for the Sequence of Educational Tools. (PDF)

Abstract

This paper identifies and formalizes the didactic paradox of “Success without Value” that arises when using digital educational tools sequentially: objective learning performance improves significantly, while the subjective value of the tool and overall user experience decrease. Based on an experiment with two groups of master’s students (102 participants) solving decision theory problems using a step by step trainer and a dialogue chatbot, the study shows that contrast effects and success attribution lead to the devaluation of the easier tool when used second. To overcome the descriptive nature of existing psychological theories, we develop the EduParadox ontology in OWL 2 DL, integrating objective and subjective metrics. The ontology is supplemented with SWRL inference rules that automatically classify specific interactions as paradoxical. The proposed model can be embedded into intelligent pedagogical design support systems for the predictive prevention of demotivating configurations of learning trajectories. It is not strictly tied to a specific subject domain, ensuring its applicability across various educational technologies.

THURSDAY,, OCTOBER 22nd

SESSION 13. Biomedical Signal Analysis.

Andrey Vasenkov, Denis Karchkov, Tatiana Levanova and Lev Smirnov. Exact ECG segmentation using deep learning approach. (PDF)

Abstract

Accurate delineation of P-QRS-T complexes in electrocardiogram (ECG) recordings is fundamental for clinical diagnosis and automated analysis. Current evaluation practices for segmentation methods often rely on a tolerance-based scheme, where a predicted boundary is deemed correct if it falls within an arbitrarily chosen δ-neighborhood of the ground truth. While this approach usefully masks small, clinically tolerable misalignments, it inflates performance metrics and introduces a critical methodological flaw: the reported accuracy becomes researcher-dependent and conflates true model predictive ability with subjective error tolerance, thereby preventing fair cross-study comparisons. To address this issue, we propose a deep learning approach that achieves exact ECG segmentation with zero tolerance (δ = 0). We adapt U-Net-like architectures and systematically evaluate different training strategies, frame lengths, and data partitioning schemes on multichannel ECG recordings. By eliminating the tolerance window entirely, our method provides an unambiguous and unbiased assessment of segmentation fidelity, enabling rigorous comparison across models and studies.

Aleksandr Smyshlaev, Nikolay Gromov, Albina Lebedeva, Artem Sharkov, Anna Grebenyukova, Lev Smirnov, Tatiana Levanova and Alexander Pisarchik. Sleep Spindle Segmentation using Hypergraph Learning. (PDF)

Abstract

Sleep spindles are distinct neurophysiological patterns in electroencephalogram (EEG) signals that serve as crucial biomarkers for cognitive function and memory consolidation. While automated detection methods, including deep learning and graph neural networks (GNNs), have advanced EEG analysis, they often struggle with limited interpretability and fail to model high-order, non-pairwise interactions among brain regions. To address these limitations, this paper proposes an alternative framework for sleep spindle segmentation based on topological feature extraction from hypergraph representations of multi-channel EEG data. Instead of training computationally demanding hypergraph neural networks directly, we transform each EEG segment into a hypergraph whose hyperedges are the maximal cliques of the thresholded connectivity graph, and describe it by interpretable topological coefficients. These features are classified using ensemble machine learning models, thereby maintaining a balance between modeling complex interactions and ensuring model interpretability. The proposed pipeline was optimized for scalability using parallel computing and validated on the custom database of clinical records and the MASS dataset. The method achieved eventbased F1-scores of 0.795 and 0.704, respectively. The results indicate that incorporating topological data analysis offers a viable and interpretable approach to biomedical signal segmentation.

Yulia Veselova, Elizaveta Sibireva, Nikolay Gromov, Artem Sharkov, Anna Grebenyukova, Albina Lebedeva, Tatiana Levanova and Lev Smirnov. Active learning approach for EEG segmentation. (PDF)

Abstract

Segmentation of epileptiform activity patterns in electroencephalography (EEG) recordings typically relies on supervised deep learning models requiring large amounts of manually annotated data. Manual labeling of EEG recordings is time-consuming, expensive, and prone to inter-rater variability, which limits the scalability of fully supervised approaches. To address this bottleneck, we propose an active learning framework for EEG segmentation that selects the most informative unlabeled segments in a single acquisition round. Our approach combines two query strategies — maximum predictive entropy and committee disagreement — with a one-dimensional U-Net segmentation backbone. We evaluate the method on a custom data set of EEG excerpts from seven pediatric patients with epileptiform activity. At a matched annotation budget, entropy-based querying outperformed random querying in all eight paired comparisons (paired Wilcoxon signed-rank test, two-sided p = 0.0078, one-sided p = 0.0039 for both F1 and recall), with the gain concentrated entirely in recall relative to random querying at the same budget (median ∆ = +0.09) and no detectable difference in precision (mean ∆ = 0.000, 95% CI [−0.055, +0.068]), and it reduced the mean absolute error of the clinically used spike-wave index to 0.069, against 0.112 for random querying and 0.099 for the pre-trained model. Committeebased querying reduced the number of missed epileptiform events from 22 to 5 (event-level, hold-out record P10) using 5 min of annotated EEG. Relative to the pre-trained model itself, however, single-round active learning did not yield a statistically significant improvement (p = 0.62), and we characterize the regimes in which it helps and in which it does not. Because the expert oracle is simulated from pre-existing masks, our results quantify annotation budget rather than annotation time; this is a feasibility study and no claim of population-level generalization is made.

SESSION 14. Information Security I: Threat Modeling and Vulnerability Detection.

Igor Kotenko, Igor Saenko and Evgenii Mityakov. Hybrid Automaton Model of a Digital Twin for Simulating Attack Scenarios against Critical Information Infrastructure Objects. (PDF)

Abstract

The paper presents a hybrid automaton model of a digital twin for simulating cyberattack scenarios on critical information infrastructure (CII) objects. The relevance of the work stems from the need for an adequate description of cyber-physical systems that combine discrete events and continuous physical processes, as well as the requirement for safe testing of protective measures without interfering with the operation of real objects. A formal model of a CII element is proposed in the form of a hybrid automaton, merging a finite set of states with continuous dynamics. Composition operations (direct product, superposition, system composition) are defined, enabling the construction of hierarchical models of complex systems. The digital twin is represented as the result of a system composition of hybrid automata of physical assets and their virtual copies, incorporating a two-way synchronization mechanism. For the verification of security properties, Linear Temporal Logic is used, allowing for the formulation of requirements regarding the unreachability of dangerous states. A connection is established between the proposed model and the stages of a threat analysis methodology. Using the example of detecting cyberattacks against the telemetry infrastructure of railway transport, the model’s applicability is demonstrated for generating synthetic scenarios and adaptively updating detection algorithms. The scientific novelty lies in the development of a formal basis for building digital twins oriented towards cyber threat analysis. The limitations of the approach and directions for further research are discussed.

Nikolai Kalinin and Nikolay Skvortsov. An Ontology-Based Knowledge Infrastructure for Cyber Threat Response. (PDF)

Abstract

Cybersecurity defenders rely on many separate standards for threat intelligence, adversary behavior, vulnerabilities, prioritization, detection rules, and response playbooks. However, these standards use different schemas, identifiers, and metadata models, which makes it difficult to connect threat understanding with practical response actions. Existing cybersecurity ontologies and knowledge graphs usually cover only a limited part of this ecosystem. We present CyberResponse-KB, a modular ontology and data-integration pipeline that organize cybersecurity resources into connected semantic layers. The ontology is supported by validation constraints for checking multi-standard data and by graph queries that link threat context with candidate CACAO playbooks and Sigma-oriented detection hints. The pipeline combines selected data from ATT&CK, KEV, EPSS, Sigma, CAPEC, CWE, and curated CACAO playbooks. This is intended as a domain knowledge layer for reusing cybersecurity resources in research infrastructures. The proof-of-concept demonstration focuses on cross-standard retrieval through competency queries and illustrative incident scenarios.

Artem Antonov, Vladislav Burovin and Vadim Voevodkin. GNN-ReGVD+: A Vulnerability Detection Pipeline for C/C++ Code with Execution-Based Verification. (PDF)

Abstract

Automatic vulnerability detection in C/C++ source code using graph neural networks is characterized by a significant level of false positives in the absence of execution-based exploitability confirmation. This paper proposes the GNN-ReGVD+ pipeline, which integrates four modules: (1) neural network detection based on the ReGVD model with parameter-efficient LoRA adaptation and hybrid scoring via FAISS; (2) retrieval of relevant exploit templates from a vector database; (3) cascading template adaptation using large language models; (4) execution-based verification in a Docker-isolated environment with multi-sanitizer support and coverage-guided fuzzing. Experimental evaluation on the Devign (27,318 functions) and MegaVul (339,548 functions) datasets demonstrated that the hybrid architecture does not degrade the baseline model performance (∆ < 0.5 p.p.), LoRA adaptation of 0.12% of parameters yields a ROC-AUC improvement of +0.179 during domain transfer with a 1.8× reduction in VRAM consumption, and the sandbox verification module covers 18 vulnerability types (23 CWEs) with a four-level evidence classification. The system is implemented as 9,372 lines of Python code with 203 automated tests.

SESSION 15. Medical Imaging.

Ljubov Lavrova and Konstantin Turalchuk. Development of a methodology for combining significance maps obtained by different interpretation methods in the context of medical image analysis. (PDF)

Abstract

Предметом исследования являются методы интерпретации результатов классификации. Цель исследования – разработка подхода к объединению интерпретаций, полученных методами интерпретации разных классов, для формирования более информативной карты значимости в контексте МРТ-срезов головного мозга. Предлагается подход с выделением ROI, объединением результатов различных методов интерпретации на уровне общих областей с учетом достоверности, стабильности и согласованности методов. Экспериментальное сравнение с исходными методами и с базовым усреднением методов на наборе данных МРТ-срезов головного мозга с ручной разметкой опухолей показало, что предложенный подход превосходит их по метрикам Mass Accuracy, recall, Dice, IoU. Наилучшие гиперпараметры для всех методов подбирались автоматически через сеточный поиск.

Aleksei Samarin, Egor Kotenko, Anastasia Mamaeva, Aleksei Toropov, Artem Nazarenko, Alexander Savelev, Elena Mikhailova, Dmitry Nazarenko, Julia Afonina and Valentin Malykh. Structured Pruning for Efficient Ultra-Early Ischemic Stroke Lesion Segmentation in Non-Contrast Brain CT. (PDF)

Abstract

Accurate segmentation of ultra-early ischemic stroke lesions in non-contrast computed tomography (NCCT) remains challenging due to low contrast, small lesion size, and ambiguous boundaries between ischemic core and penumbra. We propose a task-aware pruning framework that formulates model compression as a lesion-preserving problem by combining activation-aware importance, lesion- and boundary-aware scoring, and sensitivity-guided sparsity allocation within an iterative prune–finetune scheme. The method achieves a favorable accuracy–efficiency tradeoff, maintaining strong performance at 50% sparsity (Dice 0.722 core, 0.654 penumbra) and stable accuracy at 60% sparsity (0.701 / 0.633), while reducing FLOPs from 126.7G to 47.8G and latency from 42.9 ms to 21.6 ms. Boundary quality is also better preserved, as indicated by lower HD95 and ASSD values. Overall, the proposed approach enables efficient and clinically relevant NCCT stroke segmentation suitable for time-sensitive and resource-constrained environments.

[SHORT] Yelena Grevtseva, Mikhail Melnikov, Oksana Kabenova and Almira Kustubayeva. Воксельный анализ возраста мозга: 3D-свёрточная нейросеть на картах GM и WM. (PDF)

Abstract

Предсказание биологического возраста мозга по МРТ-данным является актуальной биомедицинской задачей. В работе оценивается эффективность 3D-свёрточной нейросети (3D-CNN) в предсказании возраста испытуемых по их структурным МРТ-изображениям. Использовались данные 176 здоровых добровольцев в возрасте 7-20 лет. Сегментированные карты серого (GM) и белого (WM) вещества приведены к размерности 26×26×26 вокселей. Разработана модель VBMCNN, принимающая на вход два канала (GM, WM) МРТ-изображения и пол испытуемых. Обучение выполнено с 5-кратной кросс-валидацией, функцией потерь Хубера и механизмом ранней остановки. Для оценки вклада пространственной информации построена контрольная модель на основе простой полносвязной сети, использующей только суммарные объёмы GM, WM и спинномозговой жидкости (CSF), а также пол. Основная модель достигла среднего R² = 0.77, в то время как контрольная показала R² = 0.12. Разница в точности моделей подтверждает информативность пространственного распределения морфометрических индексов тканей мозга. С помощью карт значимости (saliency maps) выявлены ключевые для прогноза регионы в лобных, височных и затылочных долях, а также в дополнительной моторной области; желудочки мозга не вносят вклада в результат предсказания. Корреляционный анализ не выявил статистически значимых связей между ошибкой предсказания возраста (BAG) и результатами теста интеллекта. Таким образом, 3D-свёрточные архитектуры на основе воксельных карт GM и WM значимо эффективнее методов, опирающихся только на общие объёмы вещества мозга, и могут служить основой для дальнейшего изучения возрастных изменений мозга.

[SHORT] Руслан Махнев, Екатерина Пройдакова and Ксения Петрова. Методология количественного анализа ОКТ-изображений кожи при атрофических заболеваниях на основе кластеризации К-средних. (PDF)

Abstract

В работе предложена методология автоматизированного анализа ОКТ-изображений кожи, основанная на кластеризации пикселей методом К-средних (𝐾 = 3) в цветовом пространстве CIE Lab. Значение самого яркого центроида используется как числовой диагностический признак. На выборке из 200 ОКТ-изображений двух групп пациентов (здоровая кожа и склеродермия, по 100 снимков) показано, что среднее значение яркого центроида в группе патологии устойчиво выше нормы: 190,7 против 170,3 единиц яркости для метода К-средних и 174,3 против 149,1 для ручной разметки. Разработан программный комплекс с хранением результатов в реляционной базе данных. Датасет опубликован в открытом доступе.

[SHORT] Yaroslav Somov and Ekaterina Proydakova. A Comparative Study of UNet and UNet++ for Skin Layer Segmentation in 2D Optical Coherence Tomography. (PDF)

Abstract

In this paper, we review two convolutional network architectures, UNet and UNet++, in the task of skin segmentation in 2D optical coherence tomography images. Manual layer annotation is laborintensive and subjective, motivating an automatic solution. Using a small dataset of annotated healthy skin scans, both models were trained with a composite objective function combining pixel-wise multiclass crossentropy and a differentiable variant of Jaccard index, alongside heavy data augmentation to prevent model overfitting. A deep supervision approach was employed to train the UNet++ model. Performance was measured by the Jaccard index on the validation healthy skin set and then visually evaluated by dermatologists. Qualitative analysis demonstrates promising results for both UNet and UNet++ models. Both models then were evaluated on inflamed skin samples to test generalization properties, as pathological changes affect layer boundaries. UNet++ shows more promising generalization properties. The results confirm that a deep learning approach can offer a valuable assistive tool for dermatologists, potentially accelerating their research speed and providing a strong baseline for visual assessment.

SESSION 16. Information Security II: Security of AI Systems and Data Protection.

Bulat Nutfullin, Vladimir Evgrafov and Dmitry Namiot. When the Defense Writes the Refusal: Auditing Keyword-Scored Evaluation of Inference-Time Defenses for Multimodal Large Language Models. (PDF)

Abstract

Prompt-based inference-time defenses for multimodal large language models (MLLMs) emit safety language into the very answers they defend, and a lexical keyword scorer whose refusal list contains that language then counts defended answers as refusals. We audit one such evaluator, the legacy keyword rule of our own experimental archive, over three executed defenses-static defensive prompting, single-pass rationale-conditioned prompting, and a five-sample Gaussian perturband-vote image adapter-evaluated across eight InternVL and Qwen-VL models. On the canonical 28,000-output benign grid, 7,532 outputs carry a keyword flag, a separate LLM-judge estimator attributes about 138.1 refusals, and the pooled refusal point estimate is 0.52%: the keyword-flag count was 54.6 times the estimated refusal count, and most keywordpositive answered cases echo the defense’s own safety scaffold. The mis-fires concentrate in the rationale-wrapper arms; this defense-evaluator interaction is therefore arm-dependent, not a constant offset, and invalidates keyword-scored comparison of prompt-based inference-time MLLM defenses. We formalize the audit with three claim-admissibility levels (input-valid, output-bound, construct-validated) under which three of seven planned benchmark branches are excluded. The retained pipeline profiles vary across models and corpora and set priorities for paired semantic re-evaluation.

Hailin Liu, Eugene Ilyushin, Jie Ni and Min Zhu. SafeAgent: A Runtime Protection Architecture for Agentic Systems. (PDF)

Abstract

Large language model (LLM) agents are vulnerable to promptinjection attacks that propagate through tool interactions, observations, and context. This paper presents SafeAgent, a runtime protection architecture that treats agent security as a stateful decision problem over evolving execution trajectories. The architecture separates execution governance from security reasoning through a Runtime Controller and a stateful SafeAgent Core. The Controller mediates lifecycle boundaries and tool calls, while the Core maintains persistent risk state, predicts future security states of the agent context, and selects interventions under configurable safety–utility preferences. Evaluation on Agent Security Bench and InjecAgent shows statistically significant reductions in prompt injection, memory poisoning, direct harm, and staged data stealing relative to an undefended agent and text-level guardrails. The observed benign-task utility difference is not statistically significant. Runtime profiling identifies LLM-backed risk encoding and security-world-model reasoning as the primary sources of computational cost. The findings support stateful runtime governance as a complementary protection layer for tool-using LLM agents.

Andrey Kozachok, Anatoliy Bakaev, Aleksandr Kozachok and Artem Noev. Comparative Analysis of Machine Learning Algorithm Robustness Against Adversarial Attacks in Air Quality Forecasting Tasks. (PDF)

Abstract

This study compares six machine-learning algorithm classes under three adversarial stress tests in air-quality event prediction. Random Forest, XGBoost, LightGBM, SVM-RBF, MLP, and LSTM were evaluated on data from Beijing / Wanliu, UCI Air Quality, and AURN / Aberdeen Erroll Park. The protocol comprised label-flipping poisoning, hidden-trigger injection, and ten-step PGD generated against the benchmark MLP; consequently, PGD evaluation is direct-source for that MLP and transfer-based for the other models. Core metrics were measured in five fixed-data runs with different seeds and summarized by the mean, sample standard deviation, and 95% Student-t interval. On Beijing / Wanliu, XGBoost had the highest observed clean-accuracy mean (95.51%), while RF retained the highest mean accuracy under the specified PGD@0.05 transfer inputs (66.51%) and the lowest mean target-label rate on triggered samples (22.59%). Under the same transfer interface, the highest PGD@0.05 means were 72.39% for LightGBM on UCI and 83.72% for RF on AURN, the latter below the 88.55% majority-class accuracy of the AURN test set. A single-run adversarial-training ablation increased accuracy on examples transferred from the corresponding standard neural model, but did not test an adaptive attack against the trained model. The results define protocol-relative robustness profiles rather than an intrinsic ranking or certification of physical and operational robustness

Maxim Egorov and Dmitry Namiot. A new approach to detecting sensitive data in cloud environments. (PDF)

Abstract

This paper examines automated detection of sensitive data in unstructured text using a hybrid ensemble of small language models. The architecture is language-independent at the system level, while the implementation and experiments in this paper are focused on Russian-language data. The proposed solution combines a Rule Engine, a general SLM-NER model, a document-context classifier, 18 class-specific SLM specialists, and an aggregator with class-specific thresholds. The system is designed for offline inference and is evaluated on 18 classes of sensitive entities, including explicit, obfuscated, and context-dependent cases. Experimental results show that the proposed ensemble provides a favorable accuracy-efficiency trade-off. The implementation and reproducibility materials are openly available.

FRIDAY,, OCTOBER 23nd

SESSION 17. Data Models and Database Management.

Vladislav Sukhomlinov and Oleg Sabinin. Application of Methods for Determining Stratum Boundaries to Improve Cardinality Estimation in Query Execution Plans for Numeric Columns. (PDF)

Abstract

This paper explores methods for determining stratum boundaries to improve the accuracy of cardinality estimation in query execution plans for numeric columns with asymmetric data distributions. We consider attributes with highly skewed value distributions, for which the traditional assumptions of the cost-based optimizer and conventional statistical methods can lead to significant estimation errors. We propose an algorithm based on stratification methods that constructs compact column statistics in linear time and requires limited additional memory. We conducted a preliminary experimental evaluation using a modified build of PostgreSQL 18.3 and a new benchmark that generates a database table of up to 100 GB with various data distributions. Preliminary results show the effectiveness of the proposed approach for a single numeric attribute with a non-uniform data distribution and indicate its potential to improve the quality of cost-based optimization in the future.

Evgenii Dmitriev, Roman Samarev and Sergey Stupnikov. Динамическое совместное размещение фрагментов таблиц в распределённых СУБД с учётом графа соединений схемы базы данных и координированным разделением фрагментов. (PDF)

Abstract

В распределённых реляционных СУБД данные часто размещаются по узлам кластера в виде фрагментов отношений. При выполнении запросов с операцией соединения двух отношений, фрагменты которых находятся на разных узлах, возникает межузловой сетевой обмен, являющийся основным источником задержек. В работе предлагается подход к размещению фрагментов, нацеленный на снижение объема межузлового обмена, адаптирующийся к изменениям профиля запросов посредством скользящего окна наблюдений, использующий взвешенный граф соединений отношений схемы базы данных для принятия решений о размещении фрагментов и координированное разделение фрагментов(при росте данных локальность сохраняется автоматически). Подход реализован в виде надстройки над СУБД TiDB. Проведенные эксперименты показывают улучшение локальности выполнения операций соединения и сокращение максимальной задержки транзакций на вычислительном кластере при сохранении пропускной способности.

Alexander Solovyev. Конструктор информационных систем как средство создания структур данных, независимых от реализации модели данных. (PDF)

Abstract

Создание информационной системы является трудоемкой задачей, поэтому в условиях цифровизации существенные преимущества получает тот, кто обладает технологиями быстрого создания информационных систем. Быстрое создание необходимо по причине стремительного развития цифровых сервисов и платформ, что определяет скорость неизбежного перехода экономики и общества на новый цифровой уклад. В данном исследовании предложен подход к быстрому созданию основы информационных систем, а именно средству создания структур данных и их адаптации при изменениях, независимых от реализации модели данных для информационных систем. Выполнен краткий обзор средств для быстрой разработки информационных систем, рассмотрены их достоинства и недостатки. Сделана формальная постановка задачи исследования, рассмотрены теоретические пути ее решения, приведена принципиальная схема средства создания и адаптации структур данных, независимых от реализации модели данных. Приведен пример практической реализации средства создания и адаптации структур данных, разработанный сотрудниками ФИЦ ИУ РАН. Показана практическая применимость теоретических положений исследования. Определены пути дальнейших исследований.

[SHORT] Sergej Znamenskij. Simplifying editable markup for semi-structured data. (PDF)

Abstract

Redundant markup of stored data hinders human comprehension. The need for manual editing has led to the emergence of humanreadable markup languages such as CSV, TOML, HJSON, and YAML. This article proposes an alternative approach that does not require the development, maintenance, or user learning of specialized languages: Human Oriented Markup Representation (HOMR) is safe, optimal data reformatting from XML, HTML, and JSON, facilitating human perception of the data structure in a text editor with minimal learning effort. Improved perception is achieved through extremely concise markup that takes user preferences into account. A neat visible markup of the hierarchical structure of the text with different positions of roots in branches also was found. In practice, the transformation can be implemented via pre- and postprocessing and/or text editor macros.

SESSION 18. Data Analysis in Astronomy and Space Research I.

Vladimir Samodurov. Database of many years survey observations of the LPI BSA radio telescope at a frequency of 110 MHz: creation, maintenance, and methods of use. (PDF)

Abstract

В работе сообщается о создании базы данных многолетних ежедневных обзорных наблюдений на 128-лучевой диаграмме радиотелескопа БСА ФИАН на частоте 109.0 – 111.5 МГц. Излагается сквозная методика обработки данных – от сжатия данных для внесения в базу до использования базы данных для различных научных задач. Объем переработанных начальных данных 11.54 ТБ (для т.н. малой моды данных с дискретным шагов во времени 0.1 сек и разделенная на 6 частотных полос), сама база данных после сжатия данных в сетку параметров данных с шагом 10 секунд составила 1.3 ТБ. Описаны способы использования базы данных – от мониторинга небесной сферы и радиоисточников в склонениях от −8◦ до +56◦ до поиска в данных транзиентных явлений (пульсары, FRB, ШАЛ и т.д.). Непрерывные круглосуточные обзорные наблюдения ведутся с 6.07.2012, на конец мая 2026 накоплено 397.36 ТБ данных (в малой моде наблюдений 11.54 ТБ, в большой моде наблюдений 385.82 ТБ), и они хранятся на серверах обсерватории. Но использовать их напрямую для ряда научных задач громоздко и неудобно, именно поэтому для облегчения потоковой статистической обработки данных и устранения в них разрывов (данные во время наблюдений регистрируются на трех разных приемниках, формируя на каждом из них свои файлы данных) и создана новая база данных. Объем базы данных составляет примерно 1.3 ТБ на конец мая 2026, и каждый год будет прибавляться по 128 ГБ в автоматическом режиме. Многолетняя база данных открывает перед нами широкие возможности как по поиску и отождествлению уже известных и новых радиоисточников, так и многолетнему мониторингу их параметров на частоте 110 МГц. Публичная страница базы данных расположена по адресу https://astro.prao.ru/

Eugene Shekotihin, Alexei Pozanenko, Nicolay Pankov, Alexander Moskvitin, Alina Volnova and Pavel Minaev. Sky Inpainting and Subtraction with the Brownian Bridge Diffusion Model. (PDF)

Abstract

The paper proposes the usage of a Brownian bridge diffusion (BBDM) model for astronomical images reference-based reconstruction as a possible solution to the problem of precise transient detection and photometry. Results are examined through training and applying a BBDM model to reference-based inpainting and subtraction of astronomical images from ZTF survey, Zeiss-1000 and BTA SAO RAS, ZTSH CRAO, AZT-33IK SSO, AZT-20 ATO, and other instruments. We demonstrate the applicability of the proposed method to transient photometry and search, and discuss its limitations.

Arkady Mkrtchyan, Alexey Pozanenko and Pavel Minaev. Reconstruction of compton events in segmented scintillation polarimeters at high gamma-ray fluxes. (PDF)

Abstract

Комптоновская гамма-поляриметрия основана на том, что вероятность рассеяния линейно поляризованных фотонов зависит от направления плоскости поляризации. В сегментированном сцинтилляционном детекторе направление рассеяния оценивается по координатам двух сегментов, в которых фотон оставил часть или всю свою энергию в результате комптоновского рассеяния и фотопоглощения. Азимутальное распределение таких направлений используется для определения степени и угла линейной поляризации падающего излучения. При большой скорости падения фотонов на детектор сигналы от нескольких независимых частиц могут быть объединены в одно зарегистрированное событие из-за наложения импульсов, конечного временного разрешения и мертвого времени детекторов. В результате часть пар сегментов ошибочно интерпретируется как результат взаимодействия одного фотона. Такие случайные совпадения искажают азимутальное распределение углов рассеяния и создают систематическую ошибку при восстановлении поляризации. В работе исследуется процедура реконструкции, в которой зарегистрированное событие разбивается на все возможные пары сработавших сегментов, а каждая пара классифицируется как реальное комптоновское рассеяние или случайное совпадение. В качестве основного классификатора используется случайный лес. Отклик матрицы из 256 сцинтилляционных сегментов моделируется в Geant4. В качестве спектрального примера используются параметры гамма-всплеска GRB 221009A. Показано, что наиболее информативными признаками являются расстояние между сегментами, разности их координат и суммарная зарегистрированная энергия в сегментах детектора. Применение классификатора уменьшает вклад случайных совпадений, позволяет использовать «многосегментные» события и улучшает восстановление поляризации при потоках вплоть до величин порядка 10−3 эрг см−2 с−1. Полученная граница относится к исследуемой конфигурации детектора, электроники и не является универсальной характеристикой алгоритма.

Artur Zhemchuzhnikov, Artem Prosvetov, Aleksander Andreev, Daniil Aleksandrov, Vladimir Nazarov, Marina Evdokimova, Mikhail Mogilevsky, Dmitry Chugunin and Alexander Chernyshov. Automatic Detection of Whistlers in NVK-2 Data Onboard the Ionosphere-M Satellites Based on a Convolutional Neural Network. (PDF)

Abstract

The results of developing a prototype hardware-software complex for extracting signals with predefined properties from a large volume of data are presented. A method for automatic detection of whistlers based on a convolutional neural network is presented for processing data of the NVK-2 instrument onboard the Ionosphere-M satellites of the Ionozond mission. The processing pipeline includes suppression of impulsive interference from the onboard LAERT topside sounder using a running median absolute deviation (MAD) filter, computation and normalization of spectrograms in the 0–1 kHz band using 1-second time windows, and binary classification of the spectrograms by a neural network model based on EfficientNetV2. On 977 randomly sampled frames from three periods disjoint in time from the training data (98 of them containing whistlers), the model achieves Accuracy 96%, Precision 95.4% (95% confidence interval: 87.3–98.4) and Recall 63.3% (95% confidence interval: 53.4–72.1), with ROC-AUC 0.926. On a separate stratified control sample the purity of the resulting catalogue is 99.5% (95% confidence interval: 97.1–99.9). The method was applied to data of two spacecraft over the period from 20 January to 1 November 2025 for the first spacecraft and from 20 January to 1 September 2025 for the second spacecraft (about eight months in total): 35,728,462 one-second spectrograms were processed, of which 3,452,697 were classified as containing whistlers. From the resulting database, spatio-temporal distributions of whistlers were obtained that are reproducible on both spacecraft: a stable drop in detection probability near the magnetic equator and a substantial increase in probability in the night sector as a function of the solar zenith angle. The results demonstrate the feasibility of moving from manual case-by-case event selection to a statistically representative catalogue for subsequent monitoring of the plasmasphere and plasmapause.

SESSION 19. Research Data Infrastructures.

Nikolay Skvortsov. Формирование планов управления данными на основе моделей требований. (PDF)

Abstract

Планы управления данными стали обязательным элементом современных исследовательских проектов, однако в своей традиционной форме они остаются документами, заполняемыми вручную и слабо связанными с реальным исследовательским процессом. Существующие подходы к машиноуправляемым планам управления данными обеспечивают структурирование информации, но не решают проблему формального обоснования принимаемых решений о сборе, обработке, хранении и публикации данных. В данной работе предлагается подход к формированию семантически обоснованных планов управления данными на основе моделей требований решаемых задач. Подход базируется на совместном применении онтологий предметной области для выражения функциональных требований к данным и методам, специализированных онтологий для выражения нефункциональных требований, в частности, к качеству и происхождению данных, а также онтологии планов управления данными для описания политик хранения, доступа, публикации и долговременного хранения артефактов. План управления данными формируется посредством логического вывода над онтологиями: требования задачи классифицируются в онтологиях, на основе классифицированных требований выводятся применимые политики, которые затем конкретизируются для найденных релевантных ресурсов. Формальный подход обеспечивает доказуемость соответствия плана заданным политикам управления данными. Предложенный подход встраивается в жизненный цикл решения задач над данными и предусматривает сохранение данных, артефактов интеграции, методов для их повторного использования в будущих исследованиях. Разработанный метод демонстрируется на примере решения задачи построения кривых блеска звёзд на основе данных из трёх фотометрических обзоров неба, где формируется план управления данными, включающий политики выбора источников, фильтрации по качеству, слияния данных, сохранения артефактов интеграции и публикации результирующего каталога.

Nikolai Kalinin and Nikolay Skvortsov. A Reference Architecture for Semantic Layers in FAIR Research Infrastructures. (PDF)

Abstract

Research infrastructures produce metadata in heterogeneous formats, which limits cross-domain interoperability. Although the FAIR principles require machine-actionable metadata, existing approaches often address only separate aspects of FAIR implementation. This paper proposes an integrated reference architecture for the semantic layer of FAIR research infrastructures. The architecture combines reusable vocabularies, application profile validation, ontology-based enrichment, FAIR Data Point publication, and FAIR assessment. It uses a vocabulary kernel based on DCAT v3, Dublin Core Terms, PROV-O, and FOAF, extended with optional links to domain ontologies such as ENVO, EDAM, and AGROVOC. It also introduces RI-AP, a domain-agnostic application profile with tiered SHACL conformance levels, and RICo, an optional lightweight ontology that provides research-process terms alongside DCAT. Metadata from two sources is transformed, semantically enriched, and published through a FAIR Data Point. The architecture is illustrated on 40 records from Zenodo and DataCite and evaluated with Core SHACL validation and 14 selected RDA FAIR Data Maturity Model indicators. The experiment validates the architecture concept by exercising its components together in an end-to-end workflow.

Vladimir Korenkov, Andrey Kondratyev and Aleksey Bondyakov. Multifactor Author Identification in a Digital Repository. (PDF)

Abstract

Ensuring the reliability of authorship attribution in digital repositories represents a critical challenge in scientific data management. Existing automated ingestion systems, including the publication management suite developed by JINR based on DSpace, effectively address tasks related to metadata collection and standardization. However, these systems primarily rely on unique identifiers (such as ORCID and Scopus ID) and name matching, which does not fully eliminate the risk of erroneous attribution. This paper proposes and implements a multifactor author identification module that integrates the TF-IDF algorithm as an additional verification layer based on textual content analysis of publications. The developed system employs a two-stage architecture within the import module: the first stage filters potential authors using deterministic rules (identifier matching and Levenshtein distance exceeding 95%); the second stage resolves ambiguous cases by computing cosine similarity between TF-IDF vectors derived from the uploaded document and existing author profiles. Experimental deployment on a sample of 15,240 publications demonstrated a statistically significant improvement in accuracy: the author matching precision increased from 98.1% to 99.3%, while the false positive rate decreased by 63%. Meanwhile, the average processing time per publication increased only marginally (by 45 ms), confirming the feasibility of applying this method in large-scale data environments without compromising system performance.

[SHORT] Виктор Дударев and Надежда Киселёва. Практический подход к публикации исследовательских данных в соответствии с принципами FAIR в области материаловедения с использованием MatInf RDMS. (PDF)

Abstract

Рассматриваются требования современных журналов к публикации исследовательских данных, что направлено как на улучшение воспроизводимости научных исследований, так и на более эффективное использование их результатов. В качестве примера рассматривается публикация в MatInf Research Data Management System актуальных данных, связанных объектами из подграфа исследования, совместно с архивным снимком данных на момент публикации, размещаемым в Zenodo, для обеспечения воспроизводимости результатов. Кратко обсуждается дополнение публикации исполняемыми документами на базе Jupyter4NFDI. Приводятся типовые тексты для доступности данных и кода, которые могут быть использованы в качестве шаблона для соответствующих секций в исследовательских статьях. В заключении подводятся краткие итоги и направления дальнейшего развития.

SESSION 20. Data Analysis in Astronomy and Space Research II.

Мария Ефремова, Алексей Позаненко, Николай Панков, Алина Вольнова and Александр Москвитин. Методика поиска оптического компонента GW190425 в данных обзора ZTF. (PDF)

Abstract

GW190425 — второе гравитационно-волновое событие, связанное со слиянием двойной системы нейтронных звёзд, зарегистрированное в ходе наблюдательного цикла О3 гравитационноволновыми детекторами LIGO-Virgo. Несмотря на широкую наблюдательную кампанию и поиски в данных наблюдений, оптический транзиент для этого события не обнаружен до сих пор. Вероятно, это во многом обусловлено большой областью локализации гравитационно-волнового источника. Короткий гамма-всплеск GRB 190425, совпадающий по времени с регистрацией GW190425, был найден в данных гамма-детектора SPI-ACS обсерватории INTEGRAL. Это позволило уточнить область локализации. В работе мы проводим поиск оптического компонента GW190425 в данных обзора ZTF, полученных в течение двух суток после регистрации GW190425, в этой уточнённой области. В отличие от предыдущего поиска коллаборации GROWTH, мы используем более низкий порог отношения сигнала к шуму (SNR) 3σ для идентификации объектов на разностных изображениях, но при этом ограничиваемся только изображениями, центрированными на известных галактиках в области локализации GW190425. Список из 26849 целевых галактик составлен на основе объединения каталогов SDSS-DR16, DESI-DR8, GLADE+, NED-LVS, CatWISE и Gaia-DR3. Из 58289 изображений 311 имеют SNR ≥ 3; среди них выделен список из 4 кандидатов.

Art Prosvetov, Sergey Grebenev, Sergey Belkin and Alexey Pozanenko. Correlations and Classification in Light Curves of Type Ic Supernovae with GRB Association. (PDF)

Abstract

We present a rigorous comparative analysis of the bolometric light-curve parameters for a sample of 61 strippedenvelope transients, comprising 12 gamma-ray burst-associated supernovae (GRB-SNe) and 49 ordinary Type Ic events. Supported by bootstrap resampling, our analysis examines previously anticipated global 1D correlations between peak bolometric luminosity (𝐿max) and temporal parameters, such as the decay timescale (𝐷) or full-width at half-maximum (FWHM). The corresponding 95 percent bootstrap confidence intervals broadly encompass zero, and we therefore find no statistically significant evidence for a correlation between peak energy output and the photon diffusion timescale across core-collapse scenarios. Furthermore, while we find no direct linear correlation between the individual rise (𝑅) and decay (𝐷) timescales, we identify a highly constrained empirical scaling between 𝐷 and FWHM. This structural stability provides a new, practical phenomenological tool to accurately reconstruct unobserved early rising phases of incomplete light curves using solely the well-sampled decay tail. Most crucially, we demonstrate that combining temporal and energetic parameters reveals a partial separation in a multidimensional parameter space. Exploiting this partial separation, we developed a physically-driven machine-learning classification framework restricted strictly to intrinsic astrophysical parameters. The resulting model achieves a robust predictive power (mean ROC AUC of 0.88), separating GRB-SNe from standard SNe Ic based solely on light curve morphology. This multiparameter diagnostic tool will be valuable for systematically mining archival data and identifying GRB-SN-like candidates in large-scale, un-targeted surveys (e.g., LSST) without the need for prompt high-energy triggers.

Ivan Chelovekov. Search for Hard X-ray Bursts from SGR 1935+2154 in the RSDC Archive of the INTEGRAL Orbital Observatory. (PDF)

Abstract

We present a catalog-oriented search for hard X-ray activity from the Galactic magnetar SGR 1935+2154 using data products of the International Gamma-Ray Astrophysics Laboratory (INTEGRAL) orbital observatory stored and processed in the Russian Science Data Center (RSDC). The paper is focused on data-processing methods and on cross-instrument verification rather than on a final physical population study. The analysis has two complementary parts. First, a deduplicated literature-based list of 943 reported burst events from SGR 1935+2154 is compared with light curves from the Anti-Coincidence Shield (ACS) of the SPectrometer on INTEGRAL (SPI), hereafter SPI/ACS. This gives 23 ACS events temporally associated with the reference list and provides a basis for future direction-dependent calibration of the ACS response to magnetar bursts. Second, an all-mission search is performed in pixel-illumination-function (PIF) based light curves of the INTEGRAL Soft Gamma-Ray Imager (ISGRI), the low-energy detector of the Imager on Board the INTEGRAL Satellite (IBIS), at the position of SGR 1935+2154. The search uses 20954 PIF light curves corresponding to science windows where the source is inside the adopted partial-coding field of view of the IBIS telescope. A ±5 s comparison with this reference burst list gives 18 ISGRI/PIF associations with reported bursts; all associated peaks were visually inspected and found to be plausible. Four representative ISGRI events are shown as image-domain examples with excesses at the position of SGR 1935+2154. To prioritize further verification, we define a dimensionless nearest-reference similarity metric that combines scalar trigger parameters with normalized light-curve profile morphology.

[SHORT] Maxim Pupkov, Art Prosvetov, Vasiliy Marchuk, Alexander Govorov, Olga Yushkova, Alexander Andreev and Vladimir Nazarov. 3D Lunar Surface Recovery from Generated Satellite Views via Deep Learning. (PDF)

Abstract

As lunar exploration targets wider territories, generating precise 3D topographies over expansive areas has become a primary engineering challenge. This study scales up elevation modeling by processing large orbital datasets covering regions broad enough to exhibit measurable planetary curvature. To extract complex terrain features across these distances, we deployed recently developed mapping algorithms, specifically Surface-Aligned Gaussian Splatting alongside an upgraded photogrammetric pipeline. Because orbital imagery is heavily affected by extreme shadow variations, we explicitly factored solar incidence angles into the reconstruction step. This specific addition noticeably improved the extraction of underlying geometry by decoupling surface textures from temporary lighting conditions. Moving beyond standard visual comparisons, we evaluated the output using physical metrics designed to quantify surface roughness and local unevenness. Calculating these physical parameters is strictly required to assess landing hazards and determine rover mobility. The results show that combining wide-area curvature adjustments, sun-angle awareness, and modern mesh-extraction techniques produces highly reliable topographic maps. This updated methodology offers a practical approach to analyzing lunar geomorphology on a macro scale while preserving the detailed structural data needed for future surface operations.

SESSION 21. Image Segmentation.

Aleksei Samarin, Anastasia Mamaeva, Egor Kotenko, Aleksei Toropov, Alexander Savelev, Artem Nazarenko, Elena Mikhailova, Dmitry Nazarenko, Julia Afonina and Valentin Malykh. Task-Aware Structured Pruning for Infrared Iris and Pupil Segmentation. (PDF)

Abstract

Attention-augmented encoder–decoder networks have shown strong performance in infrared iris and pupil segmentation, but their computational cost limits practical deployment in real-time eye-tracking systems. In this work, we study pruning as a means of improving the efficiency of such models while preserving segmentation quality. We propose a task-aware pruning strategy that combines activation-based importance estimation, spatial redundancy suppression, sensitivity-guided sparsity allocation, and attention module pruning. The method accounts for the structural properties of encoder–decoder architectures and the domain-specific characteristics of infrared eye images. Experiments on the InP dataset evaluate the trade-off between segmentation accuracy and computational efficiency using mIoU, Dice coefficient, parameter count, FLOPs, and inference latency. At 60% sparsity, the proposed method reduces parameters from 24.376M to 10.562M, FLOPs from 18.942G to 6.982G, and latency from 21.384 ms to 11.504 ms, while preserving mIoU of 0.879 and mDice of 0.932.

Aleksei Samarin, Anastasia Mamaeva, Egor Kotenko, Aleksei Toropov, Alexander Savelev, Artem Nazarenko, Elena Mikhailova, Dmitry Nazarenko, Julia Afonina and Valentin Malykh. Foreground-Aware Model Compression for Endoscopic Polyp Analysis. (PDF)

Abstract

Accurate recognition and segmentation of polyps in endoscopic images are critical for early detection of gastrointestinal diseases. Attention-augmented encoder–decoder architectures achieve strong performance by modelling both local details and global context, but their computational cost limits real-time clinical deployment. We propose a task-aware pruning framework that combines activation-aware importance estimation, foreground-aware scoring, sensitivity-guided sparsity allocation, structured pruning of attention components, latent-space background suppression, and iterative pruning with fine-tuning. Experiments demonstrate that the proposed approach achieves a better accuracy–efficiency trade-off compared to baseline pruning strategies. At 60% sparsity, the method reduces model size from 34.672M to 15.388M parameters, FLOPs from 78.541G to 35.264G, and latency from 24.318 ms to 11.876 ms, while maintaining strong segmentation (IoU 0.809, Dice 0.876) and classification performance (F1 0.880), outperforming all evaluated baselines at this compression rate. These results demonstrate the suitability of the proposed framework for real-time endoscopic applications.

Pavel Arkhipov, Sergey Philippskih and Maxim Tsukanov. Метод оценки нейросетевых моделей для сегментации фона аэрофотоснимков с помощью перцептивных метрик. (PDF)

Abstract

В статье рассматривается задача разработки метода оценки нейросетевых моделей для сегментации фона аэрофотоснимков. Предлагаемый подход основан на использовании перцептивных метрик MultiScale Structural Similarity Index, Learned Perceptual Image Patch Similarity и Fréchet Inception Distance, позволяющих учитывать структурные и семантические особенности результатов сегментации, а также их соответствие визуальному восприятию. В исследовании в качестве нейросетевых моделей были использованы современные архитектуры: PSPNet, DeepLabv3+, Mask2Former и AerialFormer. Оценка и сравнительный анализ моделей проводились на наборе данных LoveDA, ориентированном на задачи сегментации аэрофотоснимков. В ходе экспериментов выполнена количественная оценка качества сегментации с использованием выбранных метрик, что позволило провести сравнительный анализ рассматриваемых архитектур. Полученные результаты обеспечивают обоснование выбора наиболее эффективных моделей с учетом требований к качеству сегментации фона и могут быть использованы при разработке и адаптации нейросетевых методов обработки аэрофотоснимков. Установлено, что используемые перцептивные метрики играют различные роли в оценке качества сегментации нейросетевых моделей, по-разному отражая характеристики получаемых результатов и дополняя друг друга при комплексном анализе.

[SHORT] Valeriy Leonov, Alexander Meshcheryakov and Alexandr Kashnitskii. Neural-Network-Based Land-Use Mapping of Moscow Region from Sentinel-2 and Multi-Source Labels. (PDF)

Abstract

High-quality expert annotation is a major bottleneck for regional land-use mapping. A practical objective is therefore to obtain an image-based model capable of producing detailed maps of comparable quality while reducing dependence on spatially complete expert labels. We investigate whether existing global land-cover products can provide the training supervision required for such a model and whether the trained model can recover spatial information beyond the source labels. Three global 10 m products–Dynamic World, ESA WorldCover, and ESRI/Impact Observatory LULC–are harmonized to six IPCC land-use categories and used to construct Consensus labels and supplementary WorldCover labels. A six-channel U-Net is trained on cloud- and shadowmasked Sentinel-2 composites with confidence-weighted cross-entropy. The resulting model is evaluated for Moscow region using the Consensus labels as a pseudo-reference, a local qualitative comparison, and independent expert arable-field vectors. Agreement with Consensus labels over Moscow region reaches OA = 0.9689, wIoU = 0.9463, and wF1 = 0.9711. Independent cropland validation with VEGA-Science field contours gives F1 = 0.9671 and IoU = 0.9363. These results show that heterogeneous existing maps can provide effective training supervision while the imagebased model produces a spatially complete regional map and recovers details that are absent from the conservative source consensus.