ellibsЭлектронные библиотекиRussian Digital Libraries Journal1562-5419Казанский (Приволжский) федеральный университет10.26907/1562-5419-2025-28-5-1103-1119ellibs-611Research ArticleСтатьиОценка неопределенности в трансформерных цепях на основе принципа согласованности эффективной информацииMeasuring Uncertainty in Transformer Circuits with Effective Information ConsistencyКрасновскийАнатолий АнатольевичKrasnovskyAnatoly Anatolievicha.a.krasnovsky@gmail.comУниверситет ИннополисInnopolis University20251912202528511031119Copyright © Красновский А.А., 20252025Красновский А.А.Krasnovsky A.A.Данная работа распространяется под лицензией Creative Commons Attribution 4.0.This work is licensed under a Creative Commons Attribution 4.0 License.https://ellibs.elpub.ru/jour/article/view/611

Механистическая интерпретируемость позволяет выявлять функциональные подграфы в больших языковых моделях (LLM), известные как трансформерные цепи (Transformer Circuits, TC), которые реализуют конкретные алгоритмы. Однако отсутствует формальный способ, позволяющий за один проход количественно оценить, когда активная цепь ведет себя согласованно и, следовательно, ее состояние может быть признано корректным. Опираясь на ранее предложенную автором пучково‑теоретическую формализацию причинной эмерджентности (Krasnovsky, 2025), мы специализируем ее для трансформерных цепей и вводим безразмерную однопроходную оценку согласованности эффективной информации (Effective Information Consistency Score, EICS). EICS сочетает нормализованную несогласованность пучка, вычисляемую из локальных якобианов и активаций, с гауссовским прокси EI для причинной эмерджентности на уровне цепи, полученным из того же состояния прямого прохода. Такая конструкция является прозрачной (white‑box), однопроходной и делает единицы измерения явными, так что оценка безразмерна. Представлены практические рекомендации по интерпретации оценки, учету вычислительных затрат (с быстрыми и точными режимами) и анализ простейшего примера для проверки на адекватность.

Mechanistic interpretability has identified functional subgraphs within large language models (LLMs), known as Transformer Circuits (TCs), that appear to implement specific algorithms. Yet we lack a formal, single-pass way to quantify when an active circuit is behaving coherently and thus likely trustworthy. Building on the author’s prior sheaf‑theoretic formulation of causal emergence (Krasnovsky, 2025), we specialize it to transformer circuits and introduce the single‑pass, dimensionless Effective‑Information Consistency Score (EICS). EICS combines (i) a normalized sheaf inconsistency computed from local Jacobians and activations, with (ii) a Gaussian EI proxy for circuit-level causal emergence derived from the same forward state. The construction is white-box, single-pass, and makes units explicit so that the score is dimensionless. We further provide practical guidance on score interpretation, computational overhead (with fast and exact modes), and a toy sanity-check analysis.

механистическая интерпретируемостьтрансформерные цепитеория пучковпричинная эмерджентностьколичественная оценка неопределенностибольшие языковые модели (LLM)mechanistic interpretabilityransformer circuitssheaf theorycausal emergenceuncertainty quantificationlarge language models (LLMs)ReferencesOlsson C., Elhage N., Nanda N., et al. In-context Learning and Induction Heads. 2022. arXiv: 2209.11895.Olsson C., Elhage N., Nanda N., et al. In-context Learning and Induction Heads. 2022. arXiv: 2209.11895.Anthropic. Circuit Tracing / Attribution Graphs: Methods & Applications: Transformer Circuits Team. 2025. Access mode: https://transformer-circuits.pub/2025/attribution-graphs/ (accessed: 2025-08-20).Anthropic. Circuit Tracing / Attribution Graphs: Methods & Applications: Transformer Circuits Team. 2025. Access mode: https://transformer-circuits.pub/2025/attribution-graphs/ (accessed: 2025-08-20).Yao Y., Zhang N., Xi Z., Wang M., Xu Z., Deng S., and Chen H. Knowledge Circuits in Pretrained Transformers // Advances in Neural Information Processing Systems (NeurIPS). 2024. Vol. 37. P. 118571–118602.Yao Y., Zhang N., Xi Z., Wang M., Xu Z., Deng S., and Chen H. Knowledge Circuits in Pretrained Transformers // Advances in Neural Information Processing Systems (NeurIPS). 2024. Vol. 37. P. 118571–118602.Krasnovsky A.A. Sheaf-Theoretic Causal Emergence for Resilience Analysis in Distributed Systems. 2025. arXiv : 2503.14104.Krasnovsky A.A. Sheaf-Theoretic Causal Emergence for Resilience Analysis in Distributed Systems. 2025. arXiv : 2503.14104.Hansen J., Ghrist R. Toward a Spectral Theory of Cellular Sheaves // Journal of Applied and Computational Topology. 2019. Vol. 3, No. 4. P. 315–358.Hansen J., Ghrist R. Toward a Spectral Theory of Cellular Sheaves // Journal of Applied and Computational Topology. 2019. Vol. 3, No. 4. P. 315–358.Robinson M. Topological Signal Processing. Springer, 2014.Robinson M. Topological Signal Processing. Springer, 2014.Rosas F.E., Mediano P.A.M., Jensen H.J., Seth A.K., Barrett A.B., Carhart-Harris R.L., and Bor D. Reconciling Emergences: An Information-Theoretic Approach to Identify Causal Emergence in Multivariate Data // PLOS Computational Biology. 2020. Vol. 16, No. 12. P. e1008289.Rosas F.E., Mediano P.A.M., Jensen H.J., Seth A.K., Barrett A.B., Carhart-Harris R.L., and Bor D. Reconciling Emergences: An Information-Theoretic Approach to Identify Causal Emergence in Multivariate Data // PLOS Computational Biology. 2020. Vol. 16, No. 12. P. e1008289.Tononi G., Sporns O. Measuring Information Integration // BMC Neuroscience. 2003. Vol. 4. P. 31.Tononi G., Sporns O. Measuring Information Integration // BMC Neuroscience. 2003. Vol. 4. P. 31.Oizumi M., Albantakis L., Tononi G. From the Phenomenology to the Mechanisms of Consciousness: Integrated Information Theory 3.0 // PLOS Computational Biology. 2014. Vol. 10, No. 5. P. e1003588.Oizumi M., Albantakis L., Tononi G. From the Phenomenology to the Mechanisms of Consciousness: Integrated Information Theory 3.0 // PLOS Computational Biology. 2014. Vol. 10, No. 5. P. e1003588.Angelopoulos A.N., Bates S. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. 2021. arXiv : 2107.07511.Angelopoulos A.N., Bates S. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. 2021. arXiv : 2107.07511.Guo C., Pleiss G., Sun Y., and Weinberger K.Q. On Calibration of Modern Neural Networks // Proceedings of the 34th International Conference on Machine Learning (ICML). PMLR. 2017. P. 1321–1330.Guo C., Pleiss G., Sun Y., and Weinberger K.Q. On Calibration of Modern Neural Networks // Proceedings of the 34th International Conference on Machine Learning (ICML). PMLR. 2017. P. 1321–1330.Lakshminarayanan B., Pritzel A., Blundell C. Simple and Scalable Predictive Uncertainty Estimation Using Deep Ensembles // Advances in Neural Information Processing Systems (NeurIPS). 2017. Vol. 30.Lakshminarayanan B., Pritzel A., Blundell C. Simple and Scalable Predictive Uncertainty Estimation Using Deep Ensembles // Advances in Neural Information Processing Systems (NeurIPS). 2017. Vol. 30.Bayesian Low-rank Adaptation for Large Language Models (Laplace-LoRA). 2023. ICLR 2024 version. arXiv : 2308.13111.Bayesian Low-rank Adaptation for Large Language Models (Laplace-LoRA). 2023. ICLR 2024 version. arXiv : 2308.13111.Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models / Hase P., Bansal M., Kim B., and Ghandeharioun A. // Advances in Neural Information Processing Systems (NeurIPS). 2023. Vol. 36. P. 17643–17668.Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models / Hase P., Bansal M., Kim B., and Ghandeharioun A. // Advances in Neural Information Processing Systems (NeurIPS). 2023. Vol. 36. P. 17643–17668.Huang L., Yu W., Ma W., Zhong W., Feng Z., Wang H., Chen Q., Peng W., Feng X., Qin B., et al. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions // ACM Transactions on Information Systems. 2025. Vol. 43, No. 2. P. 1–55.Huang L., Yu W., Ma W., Zhong W., Feng Z., Wang H., Chen Q., Peng W., Feng X., Qin B., et al. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions // ACM Transactions on Information Systems. 2025. Vol. 43, No. 2. P. 1–55.

The authors declare that there are no conflicts of interest present.