Speed, Precision, and Self-Correction: HSE Faculty of Computer Science Researchers at ICML-2026

Mishan Aliev, Oleg Desheulin, Denis Rakitin, Anna Karpova

© HSE University

Researchers from the HSE Faculty of Computer Science (FCS) presented their work at the International Conference on Machine Learning (ICML 2026) in Seoul, South Korea, one of the leading scientific events in the field. Several projects by the faculty’s researchers received the prestigious Spotlight distinction.

Alexander Molozhavenko

Among the recognised papers was ‘Tensorion: A Tensor-Aware Generalization of the Muon Optimizer’ by researchers from the Laboratory for Matrix and Tensor Methods in Machine Learning at the HSE AI and Digital Science Institute. The team enhanced the popular Muon optimisation algorithm by enabling it to work with more complex, multidimensional data structures known as tensors. The paper was one of just five to receive Spotlight status at a workshop featuring 75 presentations. ‘This was my first experience presenting at an international conference, and it was an extremely positive one,’ said Alexander Molozhavenko, Research Assistant at the laboratory. ‘The conference brought together a huge number of researchers, including many working in areas closely related to ours. It was also wonderful to see such a large Russian-speaking scientific community and to make new professional connections.’

Another paper awarded Spotlight status was ‘On Efficient Scaling of GNNs via IO-Aware Layers Implementations,’ written by researchers from the Yandex Laboratory at the HSE Faculty of Computer Science. The study addresses the challenge of accelerating graph neural networks on modern graphics processing units (GPUs).

The authors note that existing implementations of graph neural networks are often poorly optimised. To address this, they analysed several of the most widely used architectures and eliminated performance bottlenecks specific to each group of algorithms. As a result, they achieved several-fold improvements in computational performance together with substantial reductions in peak memory usage. For example, the GATv2 model delivered almost a sevenfold increase in computation speed while reducing memory consumption by a factor of one hundred.

The researchers have already made their code publicly available and hope that their approach will stimulate further advances in the efficiency of graph neural networks, much as transformer architectures have undergone rapid optimisation and development in recent years.

Fedor Velikonivtsev

© HSE University

‘The conference was particularly valuable because it brought together so many researchers with a shared purpose: presenting their work, speaking with fellow authors in person, exchanging ideas, and building new professional relationships. In my view, networking is the most important aspect of any conference. Communication by email or online simply cannot match face-to-face interaction,’ said Fedor Velikonivtsev, Junior Research Fellow at the Yandex Laboratory.

He added that conferences tend to foster a more open atmosphere, with fewer preconceptions, making researchers more willing to discuss their work than during a conventional presentation. ‘For example, at one of the meet-ups I had the opportunity to speak with Tri Dao, a leading researcher in efficient GPU computing and Chief Scientist at Together AI, which is exactly the area I am currently working on. I also met many outstanding researchers from a wide range of fields and had the chance to reconnect with colleagues from abroad,’ said Fedor Velikonivtsev.

© HSE University

The delegation from the HSE AI and Digital Science Institute at the Faculty of Computer Science presented 16 research papers spanning a broad range of topics, from fundamental questions in optimisation and optimal transport theory to practical applications in text and image generation and molecular modelling.

In the paper ‘Improved Stochastic Optimization of LogSumExp,’ the researchers addressed a challenge that has limited the application of stochastic methods to a wide class of problems, including optimal transport and robust learning. At the heart of these problems lies the LogSumExp function. However, computing it exactly becomes prohibitively expensive for large datasets, while simplified approximations introduce significant errors. The team developed a new approximation that preserves the function’s key mathematical properties—convexity and smoothness—while making it possible to use fast stochastic gradient methods with virtually no loss of accuracy.

Egor Gladin

‘Previously, optimising the LogSumExp function in high-dimensional problems required extremely time-consuming computations. It was possible to speed up the process only by sacrificing either accuracy or numerical stability. We have eliminated this trade-off: our approximation delivers high accuracy and stability while converging rapidly, successfully handling problems involving hundreds of thousands of terms,’ said Egor Gladin, Senior Research Fellow at the Laboratory for Theoretical Foundations of AI Models and one of the paper’s authors.

In the paper ‘Guided Star-Shaped Masked Diffusion,’ the researchers reimagined the way diffusion models generate text or code by selecting tokens sequentially. A key limitation of existing approaches is their irreversibility: once a model makes an incorrect choice at an early stage, it cannot revise it later. The new G-Star algorithm, based on a star-shaped generation paradigm, enables the model to revisit and revise earlier decisions whenever subsequent context reveals an error. This capability is particularly valuable in real-world applications where output quality, speed, and computational cost are equally important. Moreover, the method can be integrated into existing models without requiring expensive retraining.

Viacheslav Meshchaninov

‘In conventional discrete diffusion models, every selected token is effectively ‘frozen.’ We instead proposed a star-shaped trajectory in which every token remains available for revision at every stage, allowing the model to revisit any earlier decision if later context suggests it was incorrect. This flexibility is crucial for generating long, coherent texts and software code, where a single mistake in the first token can undermine the entire logical structure. At the same time, our method requires only a lightweight fine-tuning of a single layer, making it practical for deployment in existing models,’ explained Viacheslav Meshchaninov, Junior Research Fellow at the Centre of Deep Learning and Bayesian Methods.

In the paper ‘Your GFlowNet Secretly Learns an Optimal Transport Plan,’ the researchers uncovered an unexpected connection between Generative Flow Networks (GFlowNets)—which construct discrete objects step by step for tasks such as molecular design and combinatorial optimisation—and optimal transport, the mathematical framework used to compare probability distributions while taking the geometry of the underlying space into account.

Ian Maximov

‘We demonstrated that Generative Flow Networks (GFlowNets), originally developed for generating objects, are in fact solving an optimal transport problem on a graph. This creates an unexpected bridge between two research areas and opens up new possibilities for scalable solutions to transport problems on discrete structures, where traditional methods cannot be applied,’ explained Ian Maximov, Research Assistant at the Centre of Deep Learning and Bayesian Methods.

At the SPIGM workshop held as part of the conference, Kirill Korolev, a Research Assistant at the International Laboratory of Stochastic Algorithms and High-Dimensional Inference, presented the paper ‘Stop the Sampler! Classifier-Based Adaptive Stopping for Sampling Kernels’ together with his colleagues. The researchers proposed a new approach to data generation in which a specially trained classifier determines the optimal moment to stop the computational process. This enables the algorithm to avoid systematic errors while significantly increasing its speed. In addition, the team improved the mechanisms underlying diffusion models, allowing the algorithm to allocate computational resources more flexibly by devoting additional processing steps only to the most complex regions of the data.

Kirill Korolev

‘At the conference, I focused on sessions devoted to reinforcement learning, sampling, and probabilistic methods, where I came across several particularly interesting papers and new ideas. Alongside broadening my scientific horizons, I believe one of the main purposes of attending conferences like this is to meet fellow researchers (especially those working in my own field), exchange ideas, and present the projects our laboratory is developing,’ said Kirill Korolev.

The researcher was particularly impressed by the sampling session, which featured many of the leading experts in the field, several of whom he had the opportunity to meet in person. ‘During the poster session, our colleagues also shared a number of important open research problems and suggested several promising directions for future work,’ Kirill Korolev added.

© HSE University

Another team from the same laboratory presented the poster ‘Optimality of FSQ Tokens for Continuous Diffusion for Categorical Data with Application to Text-to-Speech,’ which explored how diffusion models can process complex data more effectively at a deeper level. The researchers focused on the neural network’s latent space—the abstract representation in which an AI model encodes information. Their study showed that the Finite Scalar Quantisation (FSQ) encoding method organises this latent space in the most effective way for such architectures. To demonstrate the practical value of the approach, the team successfully applied it to the task of text-to-speech synthesis.

Tasnima Sadekova

According to Tasnima Sadekova, a Research Assistant at the International Laboratory of Stochastic Algorithms and High-Dimensional Inference, many of the conference presentations focused on diffusion language modelling, interpretability, multimodal analysis, agent-based systems, and advances in training methodologies for large language models. ‘The breadth of topics encouraged interdisciplinary exchange, allowing researchers not only to discover new ideas within their own fields but also to look beyond their areas of specialisation,’ said Tasnima Sadekova. ‘Alongside the main programme, the conference featured a series of talks devoted to the current challenges facing the field. I was particularly interested in a presentation on the future of research in an era of rapid technological progress and increasingly capable AI agent systems.’ The speaker, a professor from Princeton University, expressed optimism about the near future, arguing that AI technologies will remain primarily supporting tools, reducing the routine aspects of programming while allowing researchers to focus on generating new ideas. In his view, the key challenge ahead is developing robust methods for assessing emerging competencies in artificial intelligence.

Research Papers Presented by the HSE Faculty of Computer Science Delegation

International Laboratory of Stochastic Algorithms and High-Dimensional Inference

Kirill Korolev: ‘Stop the Sampler! Classifier-Based Adaptive Stopping for Sampling Kernels’

Tasnima Sadekova: ‘Optimality of FSQ Tokens for Continuous Diffusion for Categorical Data with Application to Text-to-Speech’

Centre of Deep Learning and Bayesian Methods

Denis Rakitin: ‘One-Step Optimal Transport via Regularized Distribution Matching Distillation’

Nikita Morozov: ‘Learning Shortest Paths with Generative Flow Networks’

Viacheslav Meshchaninov: ‘Guided Star-Shaped Masked Diffusion’

Ian Maximov: ‘Your GFlowNet Secretly Learns an Optimal Transport Plan’

Aleksandr Oganov, Ilya Bykov, Mishan Aliev: ‘ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE’

Laboratory for Theoretical Foundations of AI Models

Egor Gladin: ‘Improved Stochastic Optimization of LogSumExp’

Stepan Pavlenko: ‘ICML Uni-Bond: Learning Chemical Bonds from Atomic Coordinates’

Artsyom Patarusau: ‘Dimension-Free Bounds for Covariance Estimation with Tensor-Train Structure’

Nikita Puchkin: ‘On Approximation and Estimation of Schrödinger Potentials without the Curse of Dimensionality | OpenReview’

Alexander Shabalin, Viacheslav Meshchaninov, Dmitry Vetrov: ‘Smoothie: Smoothing Diffusion on Token Embeddings for Text Generation’

Anna Zykova-Myzina, Timofey Gritsaev, Daniil Tiapkin, Nikita Morozov: ‘Proximal Policy Optimization for Amortized Discrete Sampling’

David Vaneev, Ivan Shchekotov, Denis Rakitin, Dmitry Vetrov: ‘Efficient One-to-Many Domain Translation via Diffusive Entropic Optimal Transport | OpenReview’

Big Data and Information Retrieval School

Varvara Nazarenko, Timur Lidzhiev, Alexander Tarakanov: ‘PIEFS: Physics-Informed Eigenfunction Features with Learnable Scaling’

Laboratory for Matrix and Tensor Methods in Machine Learning

Alexander Molozhavenko: ‘Tensorion: A Tensor-Aware Generalization of the Muon Optimizer’

Vladimir Bogachev: ‘AMIGO: Adapters Meet Information Geometry | OpenReview’

Nikolay Yudin: ‘Two-Sided Low-Rank SOAP for Efficient LLM Training | OpenReview’

Viktoriia Chekalina, Daniil Moskovskiy, Tatiana Matveeva, Andrey Kuznetsov, Evgeny Frolov: ‘Generalized Fisher-Weighted SVD: Scalable Kronecker-Factored Fisher Approximation for Compressing Large Language Models’

School of Data Analysis and Artificial Intelligence

Ilya Makarov: ‘PA-CoT: Profile-Adaptive Chain-of-Thought for Personalized Nutritional Consulting | OpenReview’

Yandex Laboratory

Artem Babenko: ‘Registers Matter for Pixel-Space Diffusion Transformers’

Dmitry Eremeev: ‘GraphPFN: A Prior-Data Fitted Graph Foundation Model’

Dmitrii Feoktistov: ‘Softsign: Smooth Sign in Your Optimizer For Better Parameter Heterogeneity Handling | OpenReview’

Gleb Bazhenov: ‘A Fair Evaluation of Graph Foundation Models for Node Property Prediction’ 

Fedor Velikonivtsev: ‘On Efficient Scaling of GNNs via IO-Aware Layers Implementations | OpenReview’

Ivan Rubachev: ‘Unveiling the Role of Data Uncertainty in Tabular Deep Learning | OpenReview’

Nikolay Kartashev: ‘Unveiling the Role of Data Uncertainty in Tabular Deep Learning | OpenReview’

Akim Kotelnikov: ‘TabPack: Efficient Hyperparameter Ensembles for Tabular Deep Learning’