Researcher

Georgii Novikov

Georgii Novikov is a researcher in machine learning whose work spans tensor-train methods and Mistral AI models. He coauthored a tensor-train density estimator in 2021 and an optimization algorithm in 2022, then published work on neural-network memory and point-cloud search in 2024 and earned a Skoltech PhD in Computational and Data Science and Engineering in December 2024. In 2025 and 2026, he coauthored Magistral, Ministral 3, Voxtral TTS and Shieldstral, research on reasoning, efficient language models, speech generation and multimodal safety.

Timeline

  1. 2021

    Tensor-train density estimation

    With Maxim E. Panov and Ivan V. Oseledets, he coauthors Tensor-train density estimation in 2021. The paper introduces an efficient tensor-train model for estimating probability densities and trains it with Riemannian optimization.

  2. 2022

    Optimizing tensor-train functions

    In 2022, he coauthors Optimization of Functions Given in the Tensor Train Format with Andrei Chertkov, Gleb Ryzhakov and Ivan Oseledets. Their algorithm approximates the minimum and maximum tensor element in the TT format.

  3. 2024

    Reducing activation memory

    In 2024, he and Ivan Oseledets publish Inverted Activations, a method that saves the output of pointwise nonlinearities during the forward pass to reduce neural-network training memory.

  4. October 2024

    Compressing point clouds

    In October 2024, he coauthors a tensor-train method that represents point clouds and enables fast approximate nearest-neighbor searches in vector databases.

  5. December 2024

    A Skoltech PhD

    In December 2024, Skoltech awards him a PhD in Computational and Data Science and Engineering for a thesis on tensor-train decomposition, supervised by Professor Ivan Oseledets and co-supervised by Assistant Professor Maxim Panov.

  6. June 2025

    Mistral's first reasoning model

    In June 2025, he coauthors Magistral, Mistral's first reasoning model, and its scalable reinforcement-learning pipeline, built around Mistral's own models and infrastructure.

  7. January 2026–July 2026

    Ministral, Voxtral and Shieldstral

    In January 2026, he coauthors Ministral 3, a parameter-efficient dense language-model series in 3B, 8B, and 14B sizes. In March 2026, he coauthors Voxtral TTS, an expressive multilingual speech-generation model. In July 2026, he coauthors Shieldstral, a 3B multimodal safety classifier that frames content moderation as binary question-answering.

Sources and credits

Photo: We have no published photo of Georgii Novikov yet. Send us a free-licensed photo.

  1. Tensor-train density estimation proceedings.mlr.press. Checked 19 October 2026.
  2. Optimization of Functions Given in the Tensor Train Format arxiv.org. Checked 19 October 2026.
  3. Inverted Activations: Reducing Memory Footprint in Neural Network Training arxiv.org. Checked 19 October 2026.
  4. Tensor-Train Point Cloud Compression and Efficient Approximate Nearest-Neighbor Search arxiv.org. Checked 19 October 2026.
  5. Georgii Novikov | Skoltech skoltech.ru. Checked 19 October 2026.
  6. Magistral arxiv.org. Checked 19 October 2026.
  7. Ministral 3 arxiv.org. Checked 19 October 2026.
  8. Voxtral TTS arxiv.org. Checked 19 October 2026.
  9. Shieldstral arxiv.org. Checked 19 October 2026.