Russian Journal of Cybernetics

Commemorating Alexander Bondarev

V. A. Galkin — 2026-03-31

Mathematical Ideas of P. L. Chebyshev and Their Applications in the Natural Sciences and Artificial Intelligence Technologies: International Conference Commemorating Chebyshev’s 205th Anniversary. Call for Papers

Organizing Committee — 2026-03-31

Representation of Objects in Large Language Models: From Architecture to Emergence

N. A. Bondareva — 2026-03-31

we reviewed the architectural principles and algorithms of modern generative neural networks and large language models. We examined the core mechanisms of text processing, including tokenization, embeddings, deep contextualization, autoregressive generation, and context management. We analyzed the transformer architecture in detail and emphasized self-attention as the key innovation that enables parallel sequence processing and captures long-range dependencies efficiently. We described common architecture variants (encoder-only, decoder-only, and encoder–decoder), positional encoding methods, and the computational complexity of major components.
We studied pretraining on large text corpora and subsequent adaptation through fine-tuning and instruction tuning. We examined alignment with human preferences using reinforcement learning from human feedback (RLHF) and reviewed efficient adaptation methods, including low-rank adaptation (LoRA) and quantization. We introduced the concept of an object information context, which we define as the overall amount and distribution of information about a given object in digital data, including all references to that object in the training corpus. We showed that the structure of this context in the training data influences the formation of object representations in the model’s embedding space. We also examined the quality of the data environment and showed that low-quality sources degrade model accuracy, and we discussed approaches to mitigate these effects.

A Control-Volume Scheme for the Magnetic Induction Equation in Spherical Coordinates with Constrained Transport

I. V. Bychin — 2026-03-31

this paper presents a conservative finite-difference scheme for the magnetic induction equation in spherical coordinates. We based the approach on the integral form of Faraday’s law of induction applied to the faces of control volumes. We performed the discretization with the finite volume method using a fully implicit scheme and constrained transport (CT). We derived the grid metric terms and the coefficients of the discrete form of the magnetic induction equation in spherical coordinates. We analyzed a special case of approximating the radial component of the electric field at control-volume edges located on the polar axis (θ = 0,π). We implemented the resulting numerical scheme in the CVMHD code developed by the authors for simulations of magnetohydrodynamic flows and hydromagnetic dynamo in spherical shells.

A Neural Network for the Russian Sign Language Recognition

S. G. Elovoy — 2026-03-31

we implemented a neural network to recognize Russian finger-spelling gestures and described the training process. We developed image-processing procedures for gesture recognition. We used video frames as input vectors for the neural network. The network recognizes Russian finger-spelling. The system provides a scoring method to assess the skills of students learning sign language and motivates them to continue training. We described the mathematical models used to evaluate results and the mechanisms for user interaction. The study aims to improve the accessibility and effectiveness of Russian sign language instruction. The system improves the learning process for deaf and hard-of-hearing users by providing automatic gesture recognition, game-based motivation, and a clear assessment of results.

Key Challenges in Creating an Integrated ERP Environment

R. D. Gimranov — 2026-03-31

we studied how large enterprises develop integrated enterprise resource planning (ERP) systems under import substitution, where they must combine software from multiple vendors into a single solution. We analyzed discussions from 2022–2025 and identified key tensions between integration and adaptability, as well as the lack of a formal approach to business architecture design. We applied the emergent stratification method to structure requirements and proposed a four-layer model: the conceptual layer, the business model layer, the application layer, and the infrastructure layer. At the conceptual layer, we selected ontological models for business objects, applied the Business Capabilities Framework to describe business processes, and established codification principles. We classified AI agents as explorers, constructors, and executors to support adaptability and automation in integrated architectures. We considered an ontologybased platform as a promising foundation for system architecture. Our results provide a basis for further research on methods, modeling, and software implementation of integrated ERP systems.

The Impact of Machine Vision on the Resilience of Critical Infrastructure in Highly Automated Environments

E. M. Ismailov — 2026-03-31

we analyzed common heating systems used in process pipelines and assessed the consequences of their absence or failure for oil-producing companies. We demonstrated the importance of applying machine vision to improve safety in the fuel and energy sector. We examined machine vision algorithms in combination with existing industrial electric heating systems. Based on this analysis, we identified the key conditions for improving the performance and safety of electric heating systems using artificial intelligence technologies.

Predicting Tasks in Social Groups Using a Gated Recurrent Unit

T. N. Konurbaeva — 2026-03-31

we studied task prediction for informal groups (families, friends, interest groups, etc.) in collaborative planning systems. We developed a method based on a gated recurrent unit (GRU), chosen as a more efficient and robust neural network for short sequences than standard recurrent neural networks or long short-term memory networks. We implemented a prediction algorithm that considers both individual behavioral patterns and the collective dynamics of the group, enabling the identification of personalized and collaborative tasks. We evaluated the model using the F1-score, which is appropriate for this problem because it reduces to a binary classification (task / no task) with a strong class imbalance: most time intervals contain no new tasks, and positive examples are rare. The F1-score, as the harmonic mean of precision and recall, provides a balanced assessment by accounting equally for false positives and false negatives, since missing a real task is as undesirable as generating a false prediction.

Solution of the Inverse Problem with a Known Piecewise Linear Moving Boundary for Discontinuous Heat Flow at the Medium Interface

M. E. Korzhova — 2026-03-31

we studied the inverse problem on a half-line for a composite medium consisting of a protective layer with low thermal conductivity and the underlying material. A temperature sensor is placed at the interface between the two media, allowing the determination of temperature on the outer surface of the protective layer. The boundary moves as the protective layer decays according to a piecewise-linear law, which can be determined experimentally and is considered known. The temperature at the moving boundary is fixed and equal to the destruction temperature of the protective layer.
When the layer finishes decaying, the problem reduces to a case with a homogeneous initial condition on the material, for which the solution is already known. To account for heterogeneity in the initial condition, it is necessary to construct it accurately by solving the problem with the moving boundary. Therefore, we constructed an exact solution for the moving-boundary problem under piecewise-linear motion, after which the problem reduces to the already solved case. The article also briefly summarizes the solution to the inverse problem we previously proposed.

Coordinate and Momentum Representations of Hamilton–Jacobi and Hamilton Equations

V. P. Koshcheev — 2026-03-31

we showed that the Hamilton–Jacobi equation can be formulated in both coordinate and momentum representations using differential one-forms that satisfy Poincaré’s lemma. We also showed that one of Hamilton’s equations follows from the Hamilton–Jacobi equation in the coordinate representation, whereas the other Hamilton equation follows from the Hamilton–Jacobi equation in the momentum representation.

Complex Differential 1-Form and the Hénon–Heiles Potential

V. P. Koshcheev — 2026-03-31

we constructed the family of Hénon–Heiles potentials under the condition that the real or imaginary parts of the complex differentiable 1-form do not satisfy the Poincaré lemma. We showed that the scalar Hénon–Heiles potential serves as the source of the vortex field circulation of the force applied to the test particle. We also showed that the circulation density of the vortex field of the force remains constant even within the region of chaotic or quasi-chaotic motion.

Digital Assessment of the Bone Fracture Riskin Chronic Kidney Disease Patients

N. R. Urmantseva — 2026-03-31

phosphorus and calcium metabolism disorders are one of the leading and late-diagnosed complications of chronic kidney disease. This paper presents the analysis of patient data and a computerassisted procedure for fracture risk assessment in patients with chronic kidney disease.

Solving Fuzzy Problems Using Virtual Lattices

A. V. Myshev — 2026-03-31

in this paper, we present a new method for constructing solutions to fuzzy problems on virtual lattices under conditions of multifactor uncertainty. The method is based on and extends the theory of the virtual perspective approach. We built solutions as cellular complexes on finite node topologies of the virtual lattices within the active memory of the computational environment.

Selection of Criteria to Evaluate the Effectiveness of Community Councils

N. B. Nazina — 2026-03-31

we studied how to evaluate the performance of community councils when no standard effectiveness criteria exist. We note that the lack of clear assessment tools makes it difficult to compare municipalities and hinders efficient resource planning. We developed a set of scientifically grounded evaluation criteria using an expert-based approach. We surveyed nine specialists, including university lecturers, community council chairs, and city administration representatives. We asked the experts to rate seven proposed criteria on a five-point importance scale. We applied a statistical method to measure agreement among the experts and found a high level of consensus. We identified four key priority areas: organizational effectiveness, community engagement, territorial improvement, and economic efficiency. We found that the least significant criteria were related to public order and business involvement. We show that these findings can be used to create an integrated community council effectiveness index, improving transparency in local governance and increasing public trust in democratic institutions.

The Potential of Current Multimodal Transformers for Image Analysis

P. A. Alexandrov — 2026-03-31

we studied the image analysis capabilities of two widely used neural network services: ChatGPT-5 mini and DeepSeek-3.1 Thinking. We measured the quality of feature generation and analogy matching using a new methodology and a unique experimental framework that employed all four training examples for each of two classes. In experiments with 93 proposed sounds and automatically generated Modified Bongard Tests, ChatGPT-5 mini completed 15 (16.1%) tests, and DeepSeek-3.1 Thinking completed 17 (18.3%). These results demonstrate that, despite clear progress in few-shot learning, current multimodal neural network transformers still face fundamental limitations in contextual learning.

Simple Cognitive Process Formation Model

V. G. Redko — 2026-03-31

we constructed and analyzed a model of the formation of simple cognitive processes. The model examined a persistence hunting scenario in which a hunter pursued an antelope in very hot conditions. The hunt was successful when the antelope became exhausted before the hunter or when the hunter caught up with the antelope. The hunt was unsuccessful when the hunter lost track of the antelope or became exhausted first.
Both the hunter and the antelope had their own energy resources. The hunter followed the antelope’s tracks and mentally simulated the antelope’s behavior to estimate its movement and energy expenditure. In doing so, the hunter formulated and used hypotheses about the antelope’s movement.
We studied the model using a computer simulation. We analyzed hunter learning, changes in the energy resources of both agents, and the dynamics of successful and unsuccessful hunts. The results show that constructing and using hypotheses about animal behavior plays a key role in hunting and suggest that human cognition in this context resembles scientific reasoning.

Simulation of Fluid Flow Behind the Plate in a 2D Channel

T. V. Gavrilenko — 2026-03-31

we studied 2D fluid flow around a flat plate using computational fluid dynamics based on lattice methods. We analyzed dimensionless time characteristics of transient processes during the transition from laminar to turbulent flow. We determined the relationships between key physical parameters, flow velocity, and the Reynolds number. The study examines the formation and stabilization of vortex structures when the flow reaches critical hydrodynamic conditions as it passes an obstacle in a channel.

Dynamic Data Analysis Based on Segmentation of Graphical Representations and Structural Features

E. G. Tunyan — 2026-03-31

we studied a method for analyzing dynamic data based on segmentation of graphical representations and analysis of structural features. We first converted the original time series into a plotted graph. The method then automatically divided the curve into segments at points where the shape changed noticeably, such as changes in slope, curvature, or small discontinuities. For each segment, we fitted an approximation model (linear, polynomial, exponential, and others) that best described the local behavior of the data. We selected the model by minimizing the approximation error while applying a penalty for excessive model complexity, which prevented overfitting.
We also applied a set of quantitative characteristics to compare neighboring segments. These characteristics describe both statistical properties of the data and the parameters of the fitted models. In this framework, an anomaly is a segment or individual point whose behavior differs substantially from the behavior predicted by models fitted to the other segments. The method, therefore, detects both isolated outliers and sequences of unusual observations, known as collective anomalies, in time-dependent data.
We demonstrated the approach using a synthetic time series and presented a visualization of the resulting segmentation and detected anomalies. A comparison with classical change-point detection methods shows that the proposed method provides interpretable segmentation and remains flexible in its treatment of different structural features, which distinguishes it from purely statistical tests.

Neural Network Approach to the Simulation of Liquid Oxygen Droplet Evaporation in a Gaseous Hydrogen Atmosphere

A. Z. Zakriya — 2026-03-31

this paper presents a neural network approach for modeling the non-equilibrium evaporation of a liquid oxygen droplet in a gaseous hydrogen atmosphere. We developed a deep neural network to approximate solutions of a system of transcendental equations describing the quasi-stationary state of the gas mixture near the droplet. The network used a fully connected architecture with four hidden layers and a piecewise linear activation function. We trained the model on a dataset of 5,000 points generated by numerically solving the system using the fsolve method. The input parameters include pressure, ambient temperature, oxygen mass fraction, liquid temperature, and the non-equilibrium parameter.
The model achieved high approximation accuracy, with a coefficient of determination exceeding 0.99 for both the Peclet number, which characterizes the mass evaporation rate, and the droplet surface temperature. The mean absolute error for surface temperature was 1.18 K. The neural network provides a computational speedup of more than 150 times compared to a direct numerical solution while maintaining accuracy within 3%. We incorporated the physical constraint on the critical temperature of oxygen (154.58 K).
We demonstrated that accounting for non-equilibrium effects increases droplet lifetime by 2–5 times, which must be considered in the design of liquid rocket engine combustion chambers.

Voice-Enabled Intelligent Decision Support System for Phlebology with 3D Vein Reconstruction

N. R. Urmantseva — 2026-03-31

we present an intelligent decision support system for phlebology that addresses subjectivity and delays in diagnosing chronic venous disease (CVD) of the lower extremities. We implemented the system in Python and integrated three modules. First, we used convolutional neural networks to classify images and determine the CEAP class. Second, we developed a voice interface based on the SpeechRecognition and pyttsx3 libraries to support structured dictation of examination reports and to generate clinical recommendations. Third, we reconstructed the venous system in 3D from axial MRI slices and performed a quantitative assessment of the cross-sectional area of the inferior vena cava (IVC). The system improves diagnostic accuracy, reduces report preparation time from 15 to 7 minutes, and enables documentation without manual entry. We tested all components on clinical data, and the system meets the requirements of modern non-invasive phlebology.