All issues

Interpretability in recent deep learning models has become an epicenter of research particularly in sensitive domains such as healthcare, and finance. Concept bottleneck models have emerged as a promising approach for achieving transparency and interpretability by leveraging a set of humanunderstandable concepts as an intermediate representation before the prediction layer. However, manual concept annotation is discouraged due to the time and effort involved. Our work explores the potential of large language models (LLMs) for generating high-quality concept banks and proposes a multimodal evaluation metric to assess the quality of generated concepts. We investigate three key research questions: the ability of LLMs to generate concept banks comparable to existing knowledge bases like ConceptNet, the sufficiency of unimodal text-based semantic similarity for evaluating concept-class label associations, and the effectiveness of multimodal information in quantifying concept generation quality compared to unimodal concept-label semantic similarity. Our findings reveal that multimodal models outperform unimodal approaches in capturing concept-class label similarity. Furthermore, our generated concepts for the CIFAR-10 and CIFAR-100 datasets surpass those obtained from ConceptNet and the baseline comparison, demonstrating the standalone capability of LLMs in generating highquality concepts. Being able to automatically generate and evaluate high-quality concepts will enable researchers to quickly adapt and iterate to a newer dataset with little to no effort before they can feed that into concept bottleneck models.

In this paper, we consider the problem of restoring the correspondence matrix based on the observations of real correspondences in Moscow. Following the conventional approach [Gasnikov et al., 2013], the transport network is considered as a directed graph whose edges correspond to road sections and the graph vertices correspond to areas that the traffic participants leave or enter. The number of city residents is considered constant. The problem of restoring the correspondence matrix is to calculate all the correspondence from the $i$ area to the $j$ area.

To restore the matrix, we propose to use one of the most popular methods of calculating the correspondence matrix in urban studies — the entropy model. In our work, which is based on the work [Wilson, 1978], we describe the evolutionary justification of the entropy model and the main idea of the transition to solving the problem of entropy-linear programming (ELP) in calculating the correspondence matrix. To solve the ELP problem, it is proposed to pass to the dual problem. In this paper, we describe several numerical optimization methods for solving this problem: the Sinkhorn method and the Accelerated Sinkhorn method. We provide numerical experiments for the following variants of cost functions: a linear cost function and a superposition of the power and logarithmic cost functions. In these functions, the cost is a combination of average time and distance between areas, which depends on the parameters. The correspondence matrix is calculated for multiple sets of parameters and then we calculate the quality of the restored matrix relative to the known correspondence matrix.

We assume that the noise in the restored correspondence matrix is Gaussian, as a result, we use the standard deviation as a quality metric. The article provides an overview of gradient-free optimization methods for solving non-convex problems. Since the number of parameters of the cost function is small, we use the grid search method to find the optimal parameters of the cost function. Thus, the correspondence matrix calculated for each set of parameters and then the quality of the restored matrix is evaluated relative to the known correspondence matrix. Further, according to the minimum residual value for each cost function, we determine for which cost function and at what parameter values the restored matrix best describes real correspondence.

Traditional methods for solving the traveling salesman problem are not effective for high-dimensional problems due to their high computational complexity. One of the most effective ways to solve this problem is the decomposition approach, which includes three main stages: clustering vertices, solving subproblems within each cluster and then merging the obtained solutions into a final solution. This article focuses on the third stage — merging cycles of solving subproblems — since this stage is not always given sufficient attention, which leads to less accurate final solutions of the problem. The paper proposes a new modified Sigal algorithm for merging cycles. To evaluate its effectiveness, it is compared with two algorithms for merging cycles — the method of connecting midpoints of edges and an algorithm based on closeness of cluster centroids. The dependence of quality of solving subproblems on algorithms used for merging cycles is investigated. Sigal’s modified algorithm performs pairwise clustering and minimizes total distance. The centroid method focuses on connecting clusters based on closeness of centroids, and an algorithm using mid-points estimates the distance between mid-points of edges. Two types of clustering — k-means and affinity propagation — were also considered. Numerical experiments were performed using the TSPLIB dataset with different numbers of cities and topologies to test effectiveness of proposed algorithm. The study analyzes errors caused by the order in which clusters were merged, the quality of solving subtasks and number of clusters. Experiments show that the modified Sigal algorithm has the smallest median final distance and the most stable results compared to other methods. Results indicate that the quality of the final solution obtained using the modified Sigal algorithm is more stable depending on the sequence of merging clusters. Improving the quality of solving subproblems usually results in linear improvement of the final solution, but the pooling algorithm rarely affects the degree of this improvement.

This paper presents an experimental study of the application of automatic annotation of text data in the question – answer format (QA pairs) under conditions of limited computing resources and data protection requirements. Unlike traditional approaches based on rigid rules or the use of external APIs, we propose using small language models with a small number of parameters that can function locally without a GPU on standard CPU systems. Two models were selected for testing — Gemma-3-4b and Qwen-2.5-3b (quantized 4-bit versions) — and a corpus of documents with a clear structure and a formally rigorous style of presentation was used as source material. An automatic annotation system was developed that implements the full cycle of QA dataset generation: automatic division of the source document into logically connected fragments, formation of “question – answer” pairs using the Gemma-3-4b model, preliminary verification of their correctness using Qwen-2.5-3b based on evidence span from the context and expert quality assessment. The results are exported in JSONL format. Performance evaluation covers the entire QA pair generation system, including fragment processing by the local language model, text preprocessing and postprocessing modules. Performance is measured by the time it takes to generate a single QA pair, the total throughput of the system, RAM usage, and CPU load, which allows for an objective assessment of the computational efficiency of the proposed approach when running on a CPU. An experiment on an extended sample of 12 documents showed that automatic annotation demonstrates stable performance when processing different types of documents, while manual annotation is characterized by significantly higher time costs and high variability. Depending on the type of document, the acceleration of annotation compared to the manual process ranges from 8 to 14 times. Quality analysis showed that most of the generated QA pairs have high semantic consistency with the original context, with only a limited proportion of data requiring expert correction or exception. Although full manual validation of the corpus (the “gold standard”) was not performed as part of this work, the combination of automatic evaluation and selective expert review allows us to consider the resulting quality level acceptable for preliminary automated annotation tasks. Overall, the results confirm the practical applicability of small language models for building autonomous and reproducible automatic text annotation systems under limited computational resources and provide a basis for further research in the field of effective training corpus preparation for natural language processing tasks.

This paper investigates the effectiveness of Retrieval-Augmented Generation (RAG) combined with various Large Language Models (LLMs) for document retrieval and information access in corporate information systems. We survey typical use-cases of LLMs in enterprise environments, outline the RAG architecture, and discuss the major challenges that arise when integrating LLMs into a RAG pipeline. A system architecture is proposed that couples a text-vector encoder with an LLM. The encoder builds a vector database that indexes a library of corporate documents. For every user query, relevant contextual fragments are retrieved from this library via the FAISS engine and appended to the prompt given to the LLM. The LLM then generates an answer grounded in the supplied context. The overall structure and workflow of the proposed RAG solution are described in detail. To justify the choice of the generative component, we benchmark a set of widely used LLMs — ChatGPT, GigaChat, YandexGPT, Llama, Mistral, Qwen, and others — when employed as the answer-generation module. Using an expert-annotated test set of queries, we evaluate the accuracy, completeness, linguistic quality, and conciseness of the responses. Model-specific characteristics and average response latencies are analysed; the study highlights the significant influence of available GPU memory on the throughput of local LLM deployments. An overall ranking of the models is derived from an aggregated quality metric. The results confirm that the proposed RAG architecture provides efficient document retrieval and information delivery in corporate environments. Future research directions include richer context augmentation techniques and a transition toward agent-based LLM architectures. The paper concludes with practical recommendations on selecting an optimal RAG–LLM configuration to ensure fast and precise access to enterprise knowledge assets.

This article solves the problem of developing a technology for collecting initial data for building models for assessing the functional state of a person. This condition is assessed by the pupil response of a person to a change in illumination based on the pupillometry method. This method involves the collection and analysis of initial data (pupillograms), presented in the form of time series characterizing the dynamics of changes in the human pupils to a light impulse effect. The drawbacks of the traditional approach to the collection of initial data using the methods of computer vision and smoothing of time series are analyzed. Attention is focused on the importance of the quality of the initial data for the construction of adequate mathematical models. The need for manual marking of the iris and pupil circles is updated to improve the accuracy and quality of the initial data. The stages of the proposed technology for collecting initial data are described. An example of the obtained pupillogram is given, which has a smooth shape and does not contain outliers, noise, anomalies and missing values. Based on the presented technology, a software and hardware complex has been developed, which is a collection of special software with two main modules, and hardware implemented on the basis of a Raspberry Pi 4 Model B microcomputer, with peripheral equipment that implements the specified functionality. To evaluate the effectiveness of the developed technology, models of a single-layer perspetron and a collective of neural networks are used, for the construction of which the initial data on the functional state of intoxication of a person were used. The studies have shown that the use of manual marking of the initial data (in comparison with automatic methods of computer vision) leads to a decrease in the number of errors of the 1st and 2nd years of the kind and, accordingly, to an increase in the accuracy of assessing the functional state of a person. Thus, the presented technology for collecting initial data can be effectively used to build adequate models for assessing the functional state of a person by pupillary response to changes in illumination. The use of such models is relevant in solving individual problems of ensuring transport security, in particular, monitoring the functional state of drivers.

Flexible software for modeling polymeric film production by use of blown extrusion has been developed. It consists of library of mathematical models for extrusion and forming blown film, sub-system for changeover to new type of film and sub-system for investigation of extrusion and forming for film quality control under film production. The sub-system for changeover allows to choose the equipment of extrusion line on technical and economic indices, to synthesize 3D model of the line and to generate regulation ranges of regime parameters for given type of film. The sub-system for investigation allows to calculate temperature profiles of heating and cooling material, geometrical and optical characteristics of film depending on regime parameters for stages of extrusion and forming and to evaluate regime parameters ensuring given quality of polymeric film.

Analysis of regional economic indicators plays a crucial role in management and development planning, with Gross Regional Product (GRP) serving as one of the key indicators of economic activity. The application of artificial intelligence, including neural network technologies, enables significant improvements in the accuracy and reliability of forecasts of economic processes. This study compares three neural network algorithm models for predicting the GRP of a typical region of the Russian Federation — the Udmurt Republic — based on time series data from 2000 to 2023. The selected models include a neural network with the Bat Algorithm (BA-LSTM), a neural network model based on backpropagation error optimized with a Genetic Algorithm (GA-BPNN), and a neural network model of Elman optimized using the Particle Swarm Optimization algorithm (PSO-Elman). The research involved stages of neural network modeling such as data preprocessing, training model, and comparative analysis based on accuracy and forecast quality metrics. This approach allows for evaluating the advantages and limitations of each model in the context of GRP forecasting, as well as identifying the most promising directions for further research. The utilization of modern neural network methods opens new opportunities for automating regional economic analysis and improving the quality of forecast assessments, which is especially relevant when data are limited and for rapid decision-making. The study uses factors such as the amount of production capital, the average annual number of labor resources, the share of high-tech and knowledge-intensive industries in GRP, and an inflation indicator as input data for predicting GRP. The high accuracy of the predictions achieved by including these factors in the neural network models confirms the strong correlation between these factors and GRP. The results demonstrate the exceptional accuracy of the BA-LSTM neural network model on validation data: the coefficient of determination was 0.82, and the mean absolute percentage error was 4.19%. The high performance and reliability of this model confirm its capacity to predict effectively the dynamics of the GRP. During the forecast period up to 2030, the Udmurt Republic is expected to experience an annual increase in Gross Regional Product (GRP) of +4.6% in current prices or +2.5% in comparable 2023 prices. By 2030, the GRP is projected to reach 1264.5 billion rubles.

The paper presents a comparative analysis of the enterprises of the Arctic Zone of the Russian Federation (AZ RF) on economic indicators in accordance with the rating of the Polar index. This study includes numerical data of 193 enterprises located in the AZ RF. Machine learning methods are applied, both standard, from open source, and own original methods — the method of Optimally Reliable Partitions (ORP), the method of Statistically Weighted Syndromes (SWS). Held split, indicating the maximum value of the functional quality, this study used the simplest family of different one-dimensional partition with a single boundary point, as well as a collection of different two-dimensional partition with one boundary point on each of the two combining variables. Permutation tests allow not only to evaluate the reliability of the data of the revealed regularities, but also to exclude partitions with excessive complexity from the set of the revealed regularities. Patterns connected the class number and economic indicators are revealed using the SDT method on one-dimensional indicators. The regularities which are revealed within the framework of the simplest one-dimensional model with one boundary point and with significance not worse than p < 0.001 are also presented in the given study. The so-called sliding control method was used for reliable evaluation of such diagnostic ability. As a result of these studies, a set of methods that had sufficient effectiveness was identified. The collective method based on the results of several machine learning methods showed the high importance of economic indicators for the division of enterprises in accordance with the rating of the Polar index. Our study proved and showed that those companies that entered the top Rating of the Polar index are generally recognized by financial indicators among all companies in the Arctic Zone. However it would be useful to supplement the list of indicators with ecological and social criteria.