A least squares based computational scheme for the approximate solution of an integral equation with the Grünwald–Letnikov fractional integral has been developed. A distinctive feature of this scheme is the use of a neural network to compute the coefficients for the least squares method. The relevance of the study is determined by the fact that artificial intelligence is increasingly being applied to solve many practical problems related to various physical processes. An estimate of the convergence of approximate solutions to the exact solution has been obtained. Possible directions for the further application of artificial intelligence in solving physical problems are also considered.
The paper considers the problem of several traveling salesmen. The task is to find a set of a predetermined number of disjoint cycles on a graph with weighted arcs, in which the weight (the sum of the weights of the arcs) of the largest cycle is minimal. An accurate algorithm for solving the problem based on the branch-and-bound method has been developed. Similar to the well-known Balas–Christofides algorithm for solving the traveling salesman problem, the constructed algorithm uses the Hungarian algorithm for solving the assignment problem. Numerical experiments with large-dimensional random graphs have been carried out.
An analysis of the virtual exhibition “Science of the Soviet Union during the Great Patriotic War” is presented. It was created within the framework of the digital library The Scientific Heritage of Russia (DL SHR), as an element of the developing the Common Digital Space of Scientific Knowledge. The creation and organization of this exhibition are given as a model for integrating heterogeneous digital resources for the systematic presentation of historical and scientific content. Based on a comprehensive approach, the principles of forming an interdisciplinary collection are examined. A quantitative and structural analysis of the data set is conducted, highlighting key parameters, such as the distribution of materials by language, publication year, type, and scientific discipline, using the State Rubricator of Scientific and Technical Information (SRSTI). The architecture of the exhibition’s user interface is presented. It is demonstrated how the technological solutions implemented on the DL SHR platform contribute both conserving and promoting the accessibility of scientific heritage, ensuring its accessibility to researchers, educators, and the general public.
An Erratum to this paper has been published: https://doi.org/10.3103/S0005105526080017
This article reflects the results of novel research on the formation of the Common Digital Space of Scientific Knowledge (CDSSK). This work has been carried out since 2019 in a number of academic organizations, including the Interdepartmental Supercomputer Center of the Russian Academy of Sciences (now the Department of Supercomputer Systems and Parallel Computing at the National Research Center “Kurchatov Institute”). As part of these studies, the structure of the CDSSK ontology, a language for its description, and a number of unified software tools have been developed to ensure the formation of the ontology of individual subspaces and the input of various types and kinds of object attributes and named relationships into the CDSSK. Currently, the formation of CDSSK content is being modeled using the example of a universal and a number of thematic subspaces. The results of this modeling are presented below. The attributes and relationships of the Administrative Units class objects belonging to the Geography subspace, the Organizations and Their Subdivisions class, and the Classification Systems class belonging to the universal subspace are presented. The ability to navigate through the loaded real resources is demonstrated.
This paper is a mosaic of vivid fragments describing various industrial aspects of artificial intelligence (AI). These are sketches of a bigger picture, which may never be completed, as each day brings more information about new achievements, ideas, and threats. Discussions center around issues of civilian AI and jobs, the development of algorithms for intelligent games, the threats and dangers posed by AI, AI ethics, and standards and international norms for artificial intelligence. Each fragment is a review of the latest (mid-January 2026) Russian and international sources, with quotations, translations, screenshots, and links to original documents. This text stands as a large “fragment” on the benefits of AI applications, which was drafted at a great speed. This may well be the beginning of a separate, never-ending study.
The article examines current challenges in scientometrics arising from the surge in publication activity and the widespread adoption of generative artificial intelligence. The existing scientometric toolkit for analyzing research activity is reviewed and categorized into quantitative metrics and science mapping methods (citation network analysis, academic genealogy, semantic analysis, etc.). An attempt is made to overcome the limitations of traditional citation analysis, such as “semantic blindness” and vulnerability to manipulation. As a potential solution, a conceptual model is proposed where the unit of analysis shifts from the publication as a whole to an individual “key statement.” This approach involves recording not only the statement’s content but also its type, area of relevance, and its logical relationship with other statements (confirmation, refutation, clarification, generalization, etc.). Within this framework, principles for calculating modified scientometric metrics are introduced. The proposed model was tested on a corpus of 728 articles from the Russian journal Informatika i Obrazovanie (Informatics and Education) (2016–2025). An analysis conducted using large language models revealed that retrospective extraction of statements faces significant hurdles due to established cultures of scientific communication. Consequently, the study highlights the advantages of having authors formulate key statements themselves as a distinct type of metadata. In conclusion, the paper outlines development paths for the concept of an “embedding space of knowledge”, which could eventually complement existing approaches to analyzing the evolution of scientific ideas and theories.
Mathematical models and a variational method are presented for calculating 3D rod structures strengthened under load. These models and the calculation method offer broader capabilities compared to existing ones. Their application makes it possible to carry out calculations for rod systems strengthened by section enlargement of rods, modifying the design model, and altering the deformed state. Complex rod structures with various cross-sectional shapes are examined. The calculations are based on the hypotheses of Timoshenko’s beam theory. For thin-walled rods, the provisions of thin-walled rod theory are additionally used, taking shear deformations into account. It is assumed that the material of a rod element follows a bilinear hardening material model. The stress-strain state (SSS) of the strengthened structure is calculated in stages. At the initial stage, displacements and stresses in the structural elements under the initial loads are determined. At the second stage, the magnitudes of erection forces and stresses arising when strengthening elements are attached to the main structural elements are determined. At the final stage, the strengthened structure is calculated for the effects of additional loads applied after strengthening. Examples are presented of calculating existing structures strengthened by various methods using the proposed models and calculation method.
Contemporary approaches to subword text tokenization are considered in application to the low-resource Tajik language, which is characterised by a complex morphological structure and high variability of word forms. In the course of the study, a large-scale heterogeneous corpus was compiled and preprocessed, comprising 99 books and 134 497 textual articles of various genres and topics, with a total volume exceeding 33 million tokens. The corpus was cleaned of noise, normalised, and used as a basis for training and subsequent testing of subword models. On the basis of this corpus, five tokenization models were trained and analyzed, implementing the BPE, WordPiece, and Unigram algorithms using the Hugging Face Tokenizers and SentencePiece libraries. A comparative assessment was carried out according to a number of key indicators, including the out-of-vocabulary (OOV) rate, text compression ratio, tokenization speed, as well as n-gram distribution characteristics that make it possible to evaluate the ability of the models to reflect the morphological and structural organization of the language. The experimental results revealed the strengths and weaknesses of various approaches to subword segmentation and identified the most effective tokenization strategies under the conditions of the morphological complexity of the Tajik language. The findings may be used in the development of language models and applied NLP tools for Tajik and other low-resource languages, thus contributing to the expansion of their presence in the digital environment.
This article examines key challenges in teaching functional programming to students already familiar with the imperative paradigm. It describes the student models and underlying complexities that arise when teaching functional programming in this context (mutable variables, loops, and sequential computations). An extended example of the transition from the imperative to the functional paradigm is provided. The return of a functional value is examined in detail using examples of numerical differentiation and interpolation. An implementation of deferred evaluation based on anonymous functions is considered. The multiparadigm Lisp language is shown to be a convenient introduction to the functional paradigm.
In the context of digital transformation of organizations and the growing volume of data, there is a demand for more transparent and explainable approaches to employee evaluation. The purpose of the study is to design and validate an ontological model (OWL 2/SHACL) that integrates employees’ cognitive indicators and sociological characteristics into a unified knowledge space to support HR processes. The scientific novelty of the work lies in the development of a unified semantic model linking data from cognitive tests, questionnaires, work context, and performance indicators; in the formulation of competency questions (CQ) that trigger reasoning mechanisms within the knowledge graph; and in the creation of patterns for predicting competency gaps, identifying the risk of overload/burnout, while ensuring ethics and nondiscrimination control. The proposed approach is based on the ontological engineering methodologies METHONTOLOGY and NeOn, semantic web concepts, and psychometric methods.
The idea of using the available large arrays of ionogram processing results from vertical radiosonde of the ionosphere as training datasets for building predictive models using machine learning methods is put forward. The most common formats for saving the results of ionogram processing are considered, as well as some Internet resources with archives of freely available files of these formats. These datasets are used by us to build predictive models, including time series of critical frequencies of ionospheric layers. It is also possible to use some datasets of ionogram processing results to train models designed for automatic ionogram processing.
The heat flux distribution in the North Atlantic calculated using a stochastic difference equation scheme, namely, a first-order autoregressive scheme with random coefficients, is studied. The ERA5 database, which contains geophysical data for the 40 years from 1979 to 2018, is used. The coefficients for the autoregressive series were previously determined based on these data, and it is shown that the conditions on the coefficients ensure the existence and uniqueness of a solution to this difference equation. The method for calculating distributions is based on successive integration using an autoregressive scheme. Computational experiments are conducted and analyzed. Moreover, it is shown that the theoretically calculated distributions are in good agreement with their empirical counterparts. Further, after the division of the original time series into a distinguished mean (trend) and a residual, the latter is analyzed as a stationary random process. Selected correlation functions were calculated and it is shown that they are well approximated by known analytical expressions. Those approximations allow explicitly filtering and predicting the process under study. Numerical calculations were performed on the Lomonosov-2 supercomputer at Moscow State University.
In recent years, the presentation of full-text scientific articles in HTML has become widespread. This format offers several advantages for online publication compared to the traditional PDF format, owing to its more advanced tools for structuring material, embedding multimedia content, and implementing various interactive and dynamic features. Therefore, the task of converting manuscripts from the traditionally used MS Word and LaTeX formats into a high-quality HTML version that is capable of realizing the advantages of this format has become relevant. This paper presents the results of applying the approach for converting scientific articles from MS Word to HTML, proposed in previous studies, to Keldysh Institute of Applied Mathematics’ preprints. The interactive capabilities of the resulting HTML versions are described.
Several combinatorial exercises have been considered, which artificial intelligence solves with errors. The representatives of artificial intelligence examined are ChatGPT and DeepSeek systems. Questions (prompts) to these systems are provided, and the obtained answers are analyzed. Hypotheses are proposed regarding the reasons for the errors made by artificial intelligence when solving the tasks under consideration. It is suggested that similar errors may occur when using artificial intelligence for software development and other applications. Topics for further research are proposed, which may be of interest for determining the conditions for the continued use of artificial intelligence.
This paper proposes an approach to multicriteria decision support that is based on a cognitively oriented multiagent information-analytical system. Cognitive modeling methods are developed, including a formal ontological representation of knowledge concerning production planning and a coalition–holonic agent architecture that ensures adaptability and transparency of computations. A hybrid evolutionary multicriteria algorithm is introduced, where agents generate alternative plans at the local level using a parallel genetic algorithm that optimizes a combination of several criteria. At a global level, a multistage selection of alternatives is implemented with filtering of resource overloads and similar solutions, followed by final aggregation using the PROMETHEE and ELECTRE multicriteria decision-making methods. An experimental study is carried out that compares manual planning with planning supported by the developed system, as well as analyzing the impact of the dynamic adaptation of the genetic algorithm parameters. The results show that the use of the system makes it possible to reduce plan generation time by a factor of 20–30 while maintaining or improving solution quality. At the same time, resource overloads are completely eliminated, and the early termination of evolutionary computations is ensured without loss of solution quality. The system and proposed algorithms are intended for use in planning project activities at manufacturing enterprises.
A previously developed pipeline for enriching news articles with structured data is summarized, and an updated configuration is presented in which GPT-3–OpenAI’s third-generation natural language processing model is replaced with Qwen-Coder. As previously, the updated enrichment pipeline uses a dataset of 400 pages selected from Google News, a free news aggregator provided by Google, remains compatible with the Google Rich Results Test (Google’s tool for validating eligible structured results), and demonstrates that GPT-3-comparable output quality can be achieved on a low-power desktop PC. We describe how this substitution reduces dependence on paid GPT services and report an evaluation comparing the similarity of outputs produced by Qwen-Coder against the GPT-based baseline. The results also show the better performance of the new algorithm compared with the GPT version. The proposed tools lower the barrier to adopting semantic markup practices and thereby broaden their application in digital journalism. Overall, the findings support Qwen-Coder as a cost-effective alternative to large proprietary models for metadata enrichment tasks.
In accordance with the requirement of the Higher Attestation Commission (HAC), metadata for issues of journals from the List of Peer-Reviewed Scientific Publications in which the main scientific results of dissertations for the degree of Candidate of Sciences and for the degree of Doctor of Sciences must be published (HAC List) have been regularly deposited in the Russian Science Citation Index (RSCI) in the bibliographic database eLibrary.ru for more than 20 years. In March 2023, the editorial offices of journals from the HAC List, following a recommendation from the HAC, began uploading information on their 2022 issues to the Russian Scientific Journals (RSJ) database, created by the Russian Research Institute of Economics, Politics and Law in Science and Technology (RIEPL). In April 2025, an order from the Ministry of Science and Higher Education of the Russian Federation added a new requirement: journals in the HAC List must now be registered not only in the RSCI eLibrary.ru but also in the Metaphora Information System (IS), developed by the Russian Center for Scientific Information (RCSI). Journals from the HAC List are recommended to regularly transfer metadata of published issues to the Metaphora IS through specially organized interfaces. What role do the RSJ database and the Metaphora IS play in the infrastructure of scientific publications? In addition to developing the Metaphora IS, the RCSI, by order of the Government of the Russian Federation, acts as the operator of the White List of scientific journals. The White List was formed in 2023 by the Interdepartmental Working Group (IWG) of the Ministry of Science and Higher Education of the Russian Federation. The White List is intended to be used for monitoring and evaluating the publication activity of Russian scientists. Initially, the White List included about 29 000 English-language international journals and about 1000 Russian-language journals from the RSCI database. In September 2025, the Russian-language part of the White List expanded significantly with the inclusion of journals from the HAC List. We would like to receive detailed information from the originators of the White List on how the levels of White List journals (L1, L2, L3, L4) will correspond to the categories of journals in the HAC List (K1, K2, K3).
We propose a method for constructing scale-invariant representations of retail revenue time series based on three-bar Drummond geometry (DG) computed over three adjacent periods, extended with a multitimeframe context (day, partial calendar week, and a rolling 7-day window). Self-supervised pretraining on these “patches” is performed using a joint-embedding predictive architecture (JEPA) with spatiotemporal masking, followed by fine-tuning with output heads that quantify predictive uncertainty for next-day and next-week forecasts. The work analyzes the properties of affine invariance of the features and the identifiability of the weekly phase; empirical improvement over strong baseline models on real-world data is demonstrated.
Using Vision Transformer (ViT) models in real medical practice such as, for example, in hospitals or diagnostic centers, is often difficult because doctors’ work computers usually do not have powerful graphics processors (GPUs), and computing resources are limited. This work investigates a complete practical pipeline for model inference, aimed at reducing computational costs without significant loss of predictive performance. The proposed approach combines several optimization techniques. First, knowledge distillation (KD) is used, where a compact student model learns to mimic the behavior of a larger, more accurate teacher model. Second, an exponential moving average (EMA) of the model weights is determined to stabilize training and improve generalization. Third, posttraining INT8 quantization (PTQ) is explored to reduce model size and accelerate inference. Additionally, a simplified quantization-aware training variant (QAT-lite) is considered, where the effects of quantization are partially incorporated during fine-tuning. Experiments are conducted on the ISIC dataset, which contains dermoscopic images of skin lesions. Model performance is evaluated using standard classification metrics, including accuracy, macroaveraged F1 score, and area under the ROC curve (ROC-AUC). CPU performance is also analyzed, including inference latency, throughput, memory consumption, and the final model size. The results show that posttraining INT8 quantization preserves performance close to the FP32 baseline while substantially reducing memory and computational requirements. In contrast, QAT-lite does not consistently provide reproducible improvements over PTQ.
