Large language models (LLMs) are increasingly being used for code generation. Yet, both the generated code and the underlying LLM-based systems demand rigorous security evaluation. A common approach to enhancing code generation quality involves multi-agent systems composed of multiple models. This paper evaluates the performance of GPT-OSS 20B, GPT-OSS 120B, and Qwen3-Coder 480B models in single-agent and multi-agent configurations, using two code security benchmarks, SecurityEval and CyberSecEval. The key contribution is SafeAICoder, an extensible and scalable framework for testing LLMs enabling distributed server-side generation of multi-module programs and tests, independent of client-side code.

This study is dedicated to experiments in Word Sense Disambiguation on Russian-language material using generative (decoder-only) models of small and medium size. While the direct application of generative models is not an optimal approach for this task, such models demonstrate potential as semantic annotators of large volumes of raw data. The automation of semantic text annotation using generative models may help overcome the bottleneck of insufficient labeled data for training encoder-based models.

Previous studies show that state-of-the-art English and multilingual models can achieve accuracy above 90% on this task, while smaller models typically reach 80%+. The present study aims to determine whether a comparable task can be successfully addressed for small- and medium-scale models adapted to Russian, that do not require substantial computational resources.

The experiments reported in this paper are conducted both in a basic setting (one/few-shot prompting) with separate tests for narrow and broad context windows of the ambiguous lexeme, as well as under several modifications, such as context enrichment with paradigmatic information (e.g., hypernyms, hyponyms, domain labels of the target word) and ensemble approaches in which one model validates and refines the predictions of another. The study is based on the Russian annotated resource RuSemCor, annotated in terms of the RuWordNet semantic net.

The experiments ultimately pursue the development of an efficient and accessible pipeline for automatic semantic annotation of Russian texts, useful both for direct application and for preparing annotated data for machinelearning tasks.

Task-specific models have long dominated natural language processing tasks. However, general-purpose large language models (LLMs) have recently begun to successfully compete with highly specialized solutions across various NLP domains. In this paper, we investigate the applicability of LLMs to the task of estimating the semantic proximity of a word's meanings in a pair of usages, known as Word-in-Context (WiC). Drawing on the multilingual CoMeDi benchmark, we propose novel approaches to building automated WiC-systems based on LLMs. We conduct a systematic comparison of five different configurations in terms of quality and computational costs. In particular, we propose a configuration where LLM predictions are adjusted using a training set, without the need to fine-tune the LLM itself. Results on test sets across seven languages show that our proposed approaches enable LLMs to outperform all existing specialized systems, establishing a new state-of-the-art (SOTA) on the CoMeDi benchmark. Nevertheless, the achieved high quality comes with a significant increase in computational costs: LLM-based systems require several orders of magnitude more computations compared to compact specialized models (such as XL-DURel). This work represents a step towards understanding the trade-off between accuracy and resource efficiency when using modern LLMs in lexical semantics tasks.

This article addresses the problem of quantifying the confidence of large language models (LLMs) in the automatic semantic classification of arguments with emotive predicates. Using Russian-language social media posts, we analyze verbs of fear (to scare, to be afraid, etc.) and emotional attitude (to like, to love) with the semantic roles of experiencer, causator, and object. The study aims to compare self-assessed confidence of LLM Claude Sonnet 4.5 with expert assessments of the model's reasoning texts in argument classification in the healthcare domain. The experiment utilized a stratified set of 300 examples using chain-of-thoughts reasoning in Russian and four-step confidence scale. The results showed a moderate Spearman correlation between the expert and model assessments. A statistically significant relationship was found only between self-assessment of the model and the actual classification accuracy, while expert assessment of the linguistic characteristics of reasoning was unrelated to accuracy. It was concluded that explicit LLM reasoning is not directly related to its self-assessment of confidence and is separate from the decision-making process; reasoning may be an important functional part of the user interface, but not of the research.

One of the key automation challenges for government agencies is the preparation of official response letters. This article presents an empirical comparison of two approaches to the automatic generation of official response letters: one based on templates defining the letter structure and one based on relevant letter examples selected using RAG. The original GovLetter dataset, which includes real-life business correspondence from government agencies in the Khanty-Mansi Autonomous Okrug – Yugra, was used as a basis. Generation was performed using a locally deployed open-source language model. The quality of the results was assessed according to 12 criteria using the Schema-Guided Reasoning (SGR) framework and the LLM-as-a-Judge methodology. The experimental results demonstrate that the example-based approach outperforms the template-based approach in most metrics, particularly in the accuracy of arguments, formal tone, and correct formatting. The obtained results confirm the potential of solutions based on document history for the effective automation of official response letter preparation.

Despite the rapid development in the field of automatic speech recognition (ASR) systems, the recognition quality remains poor for audio recordings with acoustic degradation (music, crowd shouts, sounds of machinery, etc.). This becomes especially important when implementing models in critical areas that are characterized by increased attention to control and reliability of the results (aviation, medicine, autopilots, etc.). Error in such areas is very expensive and the use of models becomes impossible, even if the error occurs rarely. To reduce risks, it is advisable to evaluate the expected recognition quality in advance.

This article proposes an approach to predicting the Word Error Rate (WER) based on the acoustic characteristics of the signal and calculating the perplexity of language models. The proposed method involves the creation of diverse sets of audio data by applying various types of acoustic observations to pure speech samples at various levels of quality and intelligibility. Unlike previous studies, a complete set of speech features is extracted and analyzed: prediction of the value of the signal-to-noise ratio (SNR), neural network sound quality metrics (NISQA, etc.) as well as the perplexity of the text of the ASR hypothesis using the language model as an additional feature for training a unified model.

Experiments are being conducted using modern speech recognition architectures to demonstrate the effectiveness of the proposed method in predicting WER in various acoustic conditions. It is shown that the inclusion of perplexity significantly improves the quality of the WER prediction, in particular for data where acoustic features are weakly correlated with recognition errors. The results are applicable for automatic evaluation of the expected quality of speech recognition and filtering of audio inputs.

The paper addresses the task of taxonomy expansion – hierarchical structures for organizing concepts. We propose an architecture based on the ReAct (Reasoning + Acting) approach that enables taxonomy expansion in a zero-shot setting without fine-tuning large language models. The system is implemented in two scenarios: autonomous navigation from root nodes and verification of hypotheses generated by other models. Experiments on the diachronic RuWordNet dataset show that direct navigation from the root faces limitations due to graph complexity (MAP@3 = 24.6%). However, using the system as a verifier improves the performance of baseline models: MAP@3 gains of 9.5 pp for FastText and 1.1 pp for TaxoYandexGPT-5-Lite. The key advantages of the approach are its universality, the absence of fine-tuning requirements, and interpretability through explicit reasoning chains.

The issue of generating synthetic data for psychological research remains relevant and complex. The problems of confidentiality, reliability, validity, and accuracy of conclusions are unevenly represented in different areas of psychology and are actually related to the use of synthetic data in related sciences such as medicine, sociology, history, political science, and economics. The study of various psychological phenomena in large social groups is associated with the analysis of complex and difficult-to-formalize constructs. Synthetic data refers to artificially generated data based on algorithms and modeling.

1. Rosenzweig's classification of types of reactions to frustration was chosen as the basis for this study. When analyzing online discourse, there is a problem of the low number of certain types. This is especially true for the class of impunitive reactions. The paper analyzes the possibility of creating text corpora using synthetic data of reactions to frustration generated by large language models.

During the experiments, the experts created prompts and generated examples of impulsive reactions using four large language models, with 10 examples of each type of reaction. They then evaluated the contextual validity and quality of the generated responses.

The results obtained allowed them to identify the weaknesses in generating texts with complex psychological phenomena for training neural network models.

In the field of syntactic parsing, fine-tuning large language models to generate a sentence's syntactic structure in bracket form is a promising approach. Existing fine-tuned models demonstrate high values of syntax parsing metrics; however, in some cases they generate an incorrect bracket sequence (for example, with an unbalanced number of opening and closing parentheses). Therefore, this study aims to develop a method to reduce the number of incorrectly generated sequences. A system of constraints has been introduced on the set of tokens generated at each stage of the large language model. The implemented method was tested on four fine-tuned models (adapted and non-adapted with 4 and 8 billion parameters) on the SynTagRus dependency tree dataset. For all models, the number of incorrectly generated sequences decreased. However, some sentences were not parsed, because the LLMs generated an excessively long token sequence and were interrupted. The code for this study is publicly available.

The problem of effective navigation and search for relevant information within growing volumes of scientific publications necessitates a shift from classical full-text search methods to semantic models. This work proposes an approach to structuring a heterogeneous corpus of scientific texts by constructing a Knowledge Graph (KG). A data processing pipeline is developed that encompasses the extraction of metadata, keywords, and structural elements of articles, followed by their integration into a unified graph. Based on the constructed Knowledge Graph, methods for analyzing explicit connections and extracting implicit connections between publications are implemented. The research results demonstrate the effectiveness of the graph-based representation of scientific information for uncovering hidden patterns within subject domains and supporting intelligent navigation.

This paper presents a new Russian-language corpus with coreference annotation, HaRuCo (Habr Russian Coreference Corpus). The corpus is based on popular science articles related to the subject area of “Computational Linguistics.” The paper proposes a method for coreference annotation in texts from narrow subject areas, which includes four main stages: syntactic analysis of the text, assembly of noun phrases and identification of pronouns for constructing mentions (spans), classification of mentions by subject area classes, and clustering of mentions according to chains of coreferentially linked spans. Coreferential links were annotated using a syntactic parser and a large language model, followed by manual verification and correction. The created corpus includes 3.727 entities, 9.905 mentions, and 2.683 coreferential chains. The created corpus can be used for training and evaluating coreference resolution models for the Russian language.

Large language models (LLMs) are increasingly being used for code generation. However, both the generated code and the underlying LLM-based systems demand rigorous security evaluation. A common approach to enhancing code generation quality involves multi-agent systems composed of multiple models. This paper evaluates the performance of GPT-OSS 20B, GPT-OSS 120B, and Qwen3-Coder 480B models in single-agent and multi-agent configurations, using two code security benchmarks, SecurityEval and CyberSecEval. The key contribution is SafeAICoder, an extensible and scalable framework for testing LLMs enabling distributed server-side generation of multi-module programs and tests, independent of client-side code.

The article proposes a fuzzy regulator for calculating sliding window parameters to prepare training data for Large Language Models. The traditional approach sets the stride and context length parameters as fixed constants, uniform for the entire text, and does not account for the linguistic characteristics of individual fragments, such as dense scientific text and monotonous, repetitive text. The proposed method utilizes two automatically computed features of a fragment: lexical diversity and the average BPE token length. Based on the Mamdani algorithm with a base of 9 fuzzy logic rules and defuzzification using the center of gravity method, the fuzzy regulator adaptively calculates the stride and context length values for each fragment. The proposed approach has a cognitive interpretation, as it mimics the mechanism of adaptive human attention during reading, for example, complex fragments are processed more attentively with a small stride size.

The problem of improving the quality of intelligent search in a corpus of educational institution documents, including curricula, course syllabus, and regulatory documents, was solved. Classic architectures of the Retrieval-Augmented Generation (RAG) approach, based on a single plaintext search module, demonstrate low accuracy when used with documents containing tables, logical relationships between entities, and strict regulatory document wording. An adaptive RAG architecture consisting of four layers is proposed, each taking into account the specifics of data storage in such documents. The results show that considering document structure and adaptive query routing significantly improve the actual accuracy of responses. The proposed architecture can be used in the design of intelligent assistants for administrative and educational services at higher education institutions.

Based on the analysis of future digital technologies and modern solutions for supporting scientific research in geology, the authors of the State Geological Museum named after Vladimir Vernadsky of RAS team have developed and proposed an architecture for a digital intelligent framework of geological knowledge. The main concept of the work is to create an integrated digital ecosystem where geological collections, scientific data, and the expertise of natural science museums become part of a single, accessible, and interactive space. This is not just about digitizing exhibits, but about building a connected environment for science, education, and outreach.

Automating the verification of building information models for compliance with fire safety regulations remains a pressing issue in the architectural and construction industry. Existing automated verification systems rely on rule-based approaches that ignore the building's topological context and are poorly adapted to new projects. This paper examines the development and experimental verification of methods for predicting door fire protection parameters in BIM models using graph neural networks and validating the approach using real-world design data from seven residential buildings provided by a major real estate developer. A methodology for predicting door fire resistance classes based on relational graph convolutional networks is proposed, along with a pipeline for extracting data from a specialized format, constructing a graph, and generating features that take into account geometric, semantic, and topological characteristics. Experiments were conducted to predict the presence and fire resistance class with cross-project validation using the "one building out of sample" principle. The developed approach enables automated verification of fire protection parameters and reduces the time required to analyze building models. The use of graph neural networks ensures that topological context is taken into account and high prediction accuracy is achieved, while the use of real data from a major real estate developer confirms the practical applicability of the method.

This paper investigates the problem of constructing a mapping between the spiking activity of a convolutional spiking neural network and the class of an input image. This problem arises after unsupervised training: the convolutional layer forms an event-based representation of the image, but the network itself does not contain a ready-made mechanism for assigning a class label. The aim of the study is to evaluate how informative the spike counters of the convolutional layer are after training with the STDP rule (Spike-Timing-Dependent Plasticity, a biologically plausible learning rule for impulse-based, or spiking, neural networks) and to assess the performance of a linear readout classifier as the number of training presentations increases.

The experiments were conducted on the MNIST dataset, a standard dataset of handwritten digit images. A single-layer architecture was used: one convolutional layer with 32 feature maps, a 5 × 5 kernel, and an output dimensionality of 32 × 24 × 24, corresponding to 18432 LIF neurons, where LIF denotes the leaky integrate-and-fire neuron model. The input images were encoded using a deterministic Poisson encoder; the Poisson encoding multiplier was set to 0.006. The presentation duration, that is, the time during which an object was shown to the network, was 100 time steps. Training was performed from scratch using the STDP rule, without resuming from a previously saved state. For evaluation, a protocol was used in which spike counters were collected on 10000 training images and 10000 test images.

In the main experiment, with 15000 STDP training presentations, an accuracy of 0.8883 was achieved. The baseline label-map approach, close to the method of Diehl and Cook, produced an accuracy of about 0.48 under the same configuration. The coverage of the calibrated label map was 219 out of 18,432 neurons, while the average activity was 4,146.9416 spikes per image. The neighboring scale point with 10,000 training presentations yielded an accuracy of 0.8876. The results show that the chosen method for reading out activity has a substantial effect on the final classification quality: a linear readout classifier based on the full vector of network spike counters extracts distributed information that is lost when labels are assigned directly to individual neurons.

This paper investigates quantized approaches to language model adaptation for the task of automatic step-by-step verification of the correctness of quadratic equation solutions. The study examines the effectiveness of Parameter-Efficient Fine-Tuning (PEFT) methods when adapting the DeepSeek-R1-Distill-Qwen-1.5B and InternLM2-Math-Plus-1.8B language models to build a mathematical verifier (Process-supervised Reward Model, PRM). Experiments were conducted on a synthetic dataset of quadratic equations augmented with negative sampling to simulate learner errors. A comparative evaluation of standard (LoRA, DoRA, rsLoRA) and quantized (QLoRA, QDoRA, LoftQ) fine-tuning algorithms was performed. Zero-shot transfer generalization was additionally assessed on a structurally distinct dataset of linear equations. Results show that quantization resolves numerical stability issues for non-standard architectures (InternLM2) while achieving quality comparable to standard methods. For the DeepSeek-R1 model, QLoRA, LoftQ, and QDoRA reached accuracies of 97.77%, 98%, and 98% respectively, only marginally below the standard LoRA (98.67%). Similarly, for the non-standard InternLM2 architecture, QLoRA achieved 92.67% accuracy (vs. 93% for baseline LoRA). However, full-precision algorithms (LoRA) tend to preserve richer representations of learned patterns, providing better knowledge transfer for Reasoning-class models (DeepSeek-R1 accuracy 66.8% vs. 61.4% for QLoRA on unseen data).

The work is a pioneer in studying of the chemical composition at the micro-impurity level of archaeological copper coins as an information basis for identifying the features of monetary circulation in the medieval state. 

An interdisciplinary study of the numismatic material from the Bulgarian settlement of 98 copper coins, represented by 91 Golden Horde, 6 foreign coins of the same period and one coin of the pre-Mongol period of the Zengid dynasty, was conducted. The article contains attribution with images of the obverse and reverse of coins and data of archaeometric chemical composition study of monetary artifacts using two independent methods – X-ray fluorescence and emission spectral.

This is the first interdisciplinary study in our country at present time, covering a representative collection of copper coins of various rulers and mints from archaeological materials of the Bulgarian settlement. As a result of studying the chemical composition of the coins, the ratios of impurities in copper alloys were determined, which divide the collection into ten main groups, including coins of various minted courts and rulers of a certain date. The determination of the impurity composition allows not only to classify coins into groups, but also to put forward reasonable hypotheses about the geographical location of copper deposits used for minting.

The database, compiled from the research results and combining numismatic attribution, coin images, and chemical composition data, is published as an open dataset in the Zenodo repository and is intended for the application of machine learning and artificial intelligence methods in the tasks of automatic identification and provenance analysis of medieval coins.