<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Archiving and Interchange DTD with OASIS Tables with MathML3 v1.4 20241031//EN" "https://jats.nlm.nih.gov/archiving/1.4/JATS-archive-oasis-article1-4-mathml3.dtd">
<article xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:ali="http://www.niso.org/schemas/ali/1.0/" dtd-version="1.4" article-type="research-article" xml:lang="en"><front><journal-meta><journal-title-group><journal-title xml:lang="ru">Успехи кибернетики</journal-title></journal-title-group><issn publication-format="electronic">2712-9942</issn></journal-meta><article-meta><article-categories><subj-group><subject>Other</subject></subj-group></article-categories><title-group><article-title xml:lang="ru">Информационное поле объектов в больших языковых моделях: от архитектуры к формированию представлений</article-title><trans-title-group xml:lang="en"><trans-title>Representation of Objects in Large Language Models: From Architecture to Emergence</trans-title></trans-title-group></title-group><contrib-group><contrib contrib-type="author"><name-alternatives><name xml:lang="ru"><surname>Бондарева</surname><given-names>Н. А.</given-names></name><name xml:lang="en"><surname>Bondareva</surname><given-names>N. A.</given-names></name></name-alternatives><xref ref-type="aff" rid="aff1"/><xref ref-type="aff" rid="aff2"/><email>nicibond9991@gmail.com</email><contrib-id contrib-id-type="orcid">https://orcid.org/0000-0002-7586-903X</contrib-id></contrib><contrib contrib-type="author"><name-alternatives><name xml:lang="ru"><surname>Бондарев</surname><given-names>А. Е.</given-names></name><name xml:lang="en"><surname>Bondarev</surname><given-names>A. E.</given-names></name></name-alternatives><xref ref-type="aff" rid="aff1"/><xref ref-type="aff" rid="aff2"/><email>bond@keldysh.ru</email><contrib-id contrib-id-type="orcid">https://orcid.org/0000-0003-3681-5212</contrib-id></contrib><aff-alternatives id="aff1"><aff><institution xml:lang="en">Keldysh Institute of Applied Mathematics, Russian Academy of Sciences</institution></aff></aff-alternatives><aff-alternatives id="aff2"><aff><institution xml:lang="ru">Институт прикладной математики им. М.В. Келдыша РАН</institution></aff></aff-alternatives></contrib-group><pub-date pub-type="epub" iso-8601-date="2026-03-31"><day>31</day><month>03</month><year>2026</year></pub-date><volume>7</volume><issue>1</issue><fpage>12</fpage><lpage>23</lpage><history><date date-type="received" iso-8601-date="2025-12-15"><day>15</day><month>12</month><year>2025</year></date><date date-type="accepted" iso-8601-date="2025-12-22"><day>22</day><month>12</month><year>2025</year></date></history><self-uri xlink:href="https://ru.jcyb.ru/nisii_tech/article/view/474" xlink:title="https://ru.jcyb.ru/nisii_tech/article/view/474">https://ru.jcyb.ru/nisii_tech/article/view/474</self-uri><self-uri content-type="pdf" xlink:href="publication-4635ec60-5001-4ca3-83e3-3ea620d95aa1.pdf" xlink:title="PDF"/><abstract xml:lang="ru"><p>настоящая работа представляет обзор архитектурных принципов и алгоритмов современных генеративных нейронных сетей и больших языковых моделей. Рассматриваются фундаментальные механизмы обработки текста: векторные представления, токенизация, многослойная контекстуализация, авторегрессивная генерация и управление контекстной памятью. Детально анализируется архитектура Transformer с акцентом на механизм самовнимания как ключевую инновацию, обеспечивающую параллельную обработку последовательностей и эффективное улавливание дальних зависимостей. Описываются варианты архитектуры (encoder-only, decoder-only, encoder-decoder), позиционное кодирование и вычислительная сложность компонентов.Рассматриваются процессы предобучения на больших текстовых корпусах, адаптации через fine-tuning и instruction tuning, выравнивания с человеческими предпочтениями посредством RLHF, а также методы эффективной адаптации (LoRA, квантизация). Вводится концепция информационного поля объектов как интегральной характеристики качества и плотности информации в цифровом пространстве, а также совокупности всех информационных единиц, содержащих упоминания исследуемого объекта в цифровом пространстве. Демонстрируется связь между структурой информационного поля в обучающих данных и формированием представлений в пространстве эмбеддингов модели. Обсуждаются проблемы качества информационного окружения, включая деградацию точности при включении низкокачественных источников, и подходы к их решению.</p></abstract><abstract xml:lang="en" abstract-type="summary"><p>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.</p></abstract><kwd-group xml:lang="ru"><kwd>большие языковые модели</kwd><kwd>Transformer</kwd><kwd>механизм самовнимания</kwd><kwd>информационное поле</kwd><kwd>обучение с подкреплением от человека</kwd><kwd>векторные представления</kwd><kwd>генерация текста</kwd></kwd-group><kwd-group xml:lang="en"><kwd>large language models</kwd><kwd>transformer</kwd><kwd>self-attention</kwd><kwd>representation of objects</kwd><kwd>reinforcement learning with human feedback</kwd><kwd>vector representations</kwd><kwd>text generation</kwd></kwd-group></article-meta></front><back><ref-list><ref id="ref1"><mixed-citation publication-type="other" xml:lang="ru">Vaswani A., Shazeer N., Parmar N. et al. Attention is All You Need. 2017. arXiv:1706.03762. DOI: 10.48550/arXiv.1706.03762.</mixed-citation></ref><ref id="ref2"><mixed-citation publication-type="other" xml:lang="ru">Bahdanau D., Cho K., Bengio Y. Neural Machine Translation by Jointly Learning to Align and Translate. 2016. arXiv:1409.0473.</mixed-citation></ref><ref id="ref3"><mixed-citation publication-type="other" xml:lang="ru">Radford A., Narasimhan K., Salimans T., Sutskever I. Improving Language Understanding by Generative Pre-Training. 2018. Режим доступа: https://cdn.openai.com/research-covers/languageunsupervised/language_understanding_paper.pdf.</mixed-citation></ref><ref id="ref4"><mixed-citation publication-type="other" xml:lang="ru">Radford A., Wu J., Child R. et al. Language Models are Unsupervised Multitask Learners. 2019. Режим доступа: https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf.</mixed-citation></ref><ref id="ref5"><mixed-citation publication-type="other" xml:lang="ru">Brown T. B., Mann B., Ryder N. et al. Language Models are Few-Shot Learners. 2020. arXiv:2005.14165.</mixed-citation></ref><ref id="ref6"><mixed-citation publication-type="other" xml:lang="ru">GPT-4 Technical Report. arXiv:2303.08774. DOI: 10.48550/arXiv.2303.08774.</mixed-citation></ref><ref id="ref7"><mixed-citation publication-type="other" xml:lang="ru">Devlin J., Chang M.-W., Lee K., Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. 2018. arXiv:1810.04805. DOI: 10.18653/v1/N19-1423.</mixed-citation></ref><ref id="ref8"><mixed-citation publication-type="other" xml:lang="ru">Raffel C., Shazeer N., Roberts A. et al. Exploring the Limits of Transfer Learning with a Unified Textto-Text Transformer. Journal of Machine Learning Research. 2020;21:1–67. arXiv:1910.10683. DOI: 10.48550/arXiv.1910.10683.</mixed-citation></ref><ref id="ref9"><mixed-citation publication-type="other" xml:lang="ru">Touvron H., Lavril T., Izacard G. et al. LLaMA: Open and Efficient Foundation Language Models. 2023. arXiv:2302.13971. DOI: 10.48550/arXiv.2302.13971.</mixed-citation></ref><ref id="ref10"><mixed-citation publication-type="other" xml:lang="ru">Jiang A. Q., Sablayrolles A., Mensch A. et al. Mistral 7B. 2023. arXiv:2310.06825. DOI: 10.48550/arXiv.2310.06825.</mixed-citation></ref><ref id="ref11"><mixed-citation publication-type="other" xml:lang="ru">Kaplan J., McCandlish S., Henighan T. et al. Scaling Laws for Neural Language Models. 2020. arXiv:2001.08361. DOI: 10.48550/arXiv.2001.08361.</mixed-citation></ref><ref id="ref12"><mixed-citation publication-type="other" xml:lang="ru">Hoffmann J., Borgeaud S., Mensch A. et al. Training Compute-Optimal Large Language Models. Advances in Neural Information Processing Systems (NeurIPS). 2022;35:30016-30030. arXiv:2203.15556.</mixed-citation></ref><ref id="ref13"><mixed-citation publication-type="other" xml:lang="ru">Ouyang L., Wu J., Jiang X. et al. Training Language Models to Follow Instructions with Human Feedback. Advances in Neural Information Processing Systems (NeurIPS). 2022;35:27730–27744. arXiv:2203.02155.</mixed-citation></ref><ref id="ref14"><mixed-citation publication-type="other" xml:lang="ru">Wei J., Bosma M., Zhao V. Y. et al. Finetuned Language Models Are Zero-Shot Learners. International Conference on Learning Representations (ICLR). 2022. arXiv:2109.01652.</mixed-citation></ref><ref id="ref15"><mixed-citation publication-type="other" xml:lang="ru">Chung H. W., Hou L., Longpre S. et al. Scaling Instruction-Finetuned Language Models. 2022. arXiv:2210.11416. DOI: 10.48550/arXiv.2210.11416.</mixed-citation></ref><ref id="ref16"><mixed-citation publication-type="other" xml:lang="ru">Bai Y. et al. Constitutional AI: Harmlessness from AI Feedback. 2022. arXiv:2212.08073. DOI: 10.48550/arXiv.2212.08073.</mixed-citation></ref><ref id="ref17"><mixed-citation publication-type="other" xml:lang="ru">Hu E. J., Shen Y., Wallis P. et al. LoRA: Low-Rank Adaptation of Large Language Models. International Conference on Learning Representations (ICLR). 2023. arXiv:2106.09685. DOI: 10.48550/arXiv.2106.09685.</mixed-citation></ref><ref id="ref18"><mixed-citation publication-type="other" xml:lang="ru">Dettmers T., Lewis M., Belkada Y., Zettlemoyer L. LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale. Advances in Neural Information Processing Systems (NeurIPS). 2022;35:30318– 30332. arXiv:2208.07339.</mixed-citation></ref><ref id="ref19"><mixed-citation publication-type="other" xml:lang="ru">Shazeer N., Mirhoseini A., Maziarz K. et al. Outrageously Large Neural Networks: The SparselyGated Mixture-of-Experts Layer. International Conference on Learning Representations (ICLR). 2017. arXiv:1701.06538.</mixed-citation></ref><ref id="ref20"><mixed-citation publication-type="other" xml:lang="ru">Wei J., Wang X., Schuurmans D. et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems (NeurIPS). 2022;35:24824– 24837. arXiv:2201.11903.</mixed-citation></ref><ref id="ref21"><mixed-citation publication-type="other" xml:lang="ru">Yao S., Yu D., Zhao J. et al. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. 2023. arXiv:2305.10601. DOI: 10.48550/arXiv.2305.10601.</mixed-citation></ref><ref id="ref22"><mixed-citation publication-type="other" xml:lang="ru">Бондарева Н. А., Бондарев A. E., Андреев С. В., Рыжова И. Г. Информационная плотность объектов в цифровой среде: теоретические основы. Научная визуализация. 2025;17(4):87–98. DOI: 10.26583/sv.17.4.09.</mixed-citation></ref><ref id="ref23"><mixed-citation publication-type="other" xml:lang="ru">Bommasani R., Hudson D. A., Adeli E. et al. On the Opportunities and Risks of Foundation Models. 2021. arXiv:2108.07258. DOI: 10.48550/arXiv.2108.07258.</mixed-citation></ref><ref id="ref24"><mixed-citation publication-type="other" xml:lang="ru">Goodfellow I., Bengio Y., Courville A. Deep Learning. MIT Press; 2016. ISBN: 9780262035613.</mixed-citation></ref><ref id="ref25"><mixed-citation publication-type="other" xml:lang="ru">McEnery T., Hardie A. Corpus Linguistics: Method, Theory and Practice. Cambridge University Press; 2011. ISBN: 9780521761942.</mixed-citation></ref><ref id="ref26"><mixed-citation publication-type="other" xml:lang="ru">Liu B. Sentiment Analysis: Mining Opinions, Sentiments, and Emotions. Cambridge University Press; 2015. ISBN: 9781107017894.</mixed-citation></ref><ref id="ref27"><mixed-citation publication-type="other" xml:lang="ru">Lewis P., Perez E., Piktus A. et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems (NeurIPS). 2020;33:9459-9474. arXiv:2005.11401.</mixed-citation></ref></ref-list></back></article>
