HSE FCS Researchers Showcase AI and Bioinformatics Breakthroughs at ICLR 2026

© HSE University

Researchers from the AI and Digital Science Institute at the HSE Faculty of Computer Science, along with students from the AI360: Artificial Intelligence Engineering track of the Applied Mathematics and Information Science bachelor’s programme, took part in ICLR, one of the world’s most prestigious international conferences on machine learning and representation learning. This year’s event was held in Rio de Janeiro, Brazil.

Main Track: Breakthroughs in Generative AI and Bioinformatics

ICLR 2026 featured a wide range of research by HSE Faculty of Computer Science representatives. Researchers presented papers in the main track, delivered talks in the newly introduced Blog Track, and participated in thematic workshops. Their scientific contributions covered a broad spectrum of topics, from accelerating image generation and weather forecasting to DNA design and solving quantum equations.

Alexander Oganov, Research Assistant at the Centre of Deep Learning and Bayesian Methods, presented the paper GAS: Improving Discretization of Diffusion ODEs via Generalized Adversarial Solver, which introduces a new method for speeding up image generation using diffusion models. The researchers optimised mathematical computations for specific tasks through adversarial machine learning techniques. Thanks to this innovation, high-quality images can now be generated in just 4–10 steps, as opposed to standard approaches that typically require 50–100 steps.

Alexander Oganov, Eva Neudachina, and Alexander Tolmachev (Moscow Institute of Physics and Technology)

© HSE University

‘Image generation is often formulated as a process in which each new step gradually improves the result. Modern models use neural networks with over one hundred million parameters and involve highly complex multi-step processes. In our paper, we show that by correctly adjusting around forty numbers, it is possible to simplify the generation process and produce images of cats four times faster,’ Alexander Oganov explained.

Main Track Publications

 Tight Bounds for Schrödinger Potential Estimation in Unpaired Data Translation Problems

 GeomMotif: A Benchmark for Arbitrary Geometric Preservation in Protein Generation

 LoRA meets Riemannion: Muon Optimizer for Parametrization-independent Low-Rank Adapters

 GENLINK: Ancestry Inference with GNNs on IBD Graphs for Genetically Similar Populations

Block Track: Computing Singular Value Decomposition Much Faster

ICLR introduced the experimental Blog Track this year, designed to make complex research more accessible through simplified explanations suitable for social media.

Askar Tsyganov

© HSE University

Askar Tsyganov, Research Assistant at the HSE International Laboratory of Stochastic Algorithms and High-Dimensional Inference, presented findings from the study Faster SVD via Accelerated Newton–Schulz Iteration. The work focuses on accelerating singular value decomposition (SVD), a fundamental operation widely used in applications ranging from data compression to recommendation systems. The authors proposed a method based on simple matrix multiplication operations rather than traditional algorithms, which are often poorly adapted to modern GPUs. This approach could significantly speed up neural network training, signal processing, and other computationally intensive tasks.

‘ICLR brings together leading AI companies and researchers from around the world. By attending the conference’s poster sessions, you can see future trends that are likely to shape the field and transform the world. Posters on such topics always attract a large audience. This year, research on world models stood out in particular—a concept in which neural networks learn to predict the dynamics of their environment,’ Askar Tsyganov noted.

Workshops: From Weather Physics to Neural Network Optimisation

In the study Physics-Constrained Neural Networks for Improved Short-Term Weather Forecasting: A Case Study over the South Pacific, researchers proposed a solution to one of the key challenges facing neural network-based weather forecasting models—so-called hallucinations, where the model produces physically impossible atmospheric states.

Neural network models do not always obey fundamental physical laws. As a result, forecasts may appear plausible from a data perspective while contradicting the equations of hydrodynamics. To overcome this limitation, the researchers developed a hybrid architecture combining numerical solutions of partial differential equations with modern neural network components. This approach preserves the physical validity of the model while retaining the advantages of machine learning.

‘The developed model predicts atmospheric conditions step by step and was trained on the open WeatherBench dataset using supercomputer resources. This method maintains the physical consistency of forecasts by imposing a rigid framework based on fundamental equations, thereby reducing the risk of non-physical results. It opens up promising opportunities for the development of hybrid physics-AI models in climatology,’ said Egor Bugaev, Research Assistant at the HSE Laboratory of Methods for Big Data Analysis (LAMBDA).

Egor Bugaev, Fedor Buzaev

© HSE University

In the paper Flow-Matching Sampling in Physics-Informed Neural Networks for PDEs with Sharp Source Terms, researchers tackled the challenge of training neural networks that model physical equations containing ‘sharp’ sources—abrupt local peaks in the right-hand side terms. Physics-informed neural networks (PINNs) often lose numerical stability when dealing with such equations and require vast numbers of sample points in regions with steep gradients. Uniform or random point distributions typically fail to adequately cover the areas where the greatest errors occur. To address this issue, the researchers proposed a new adaptive sampling strategy based on diffusion models: using a flow-matching technique with optimal transport, the neural network learns to generate additional sample points precisely where equation residuals are highest.

‘The developed method, FMS PINN, operates iteratively: at each stage, the network is retrained on an expanded set of points generated by a vector field that transforms a simple Gaussian distribution into one that mirrors the shape of residuals. The approach was tested on the Poisson equation with singular sources, linear elasticity equations in materials with complex geometry, as well as on the Klein–Gordon equation. In problems involving sharp peaks, FMS PINN achieved accuracy up to ten times greater than normalising flows and other competing methods. This opens the door to the reliable application of PINNs in engineering tasks with localised features—from stress analysis in composite materials to electrostatics with point sources,’ commented Fedor Buzaev, Junior Research Fellow at the Laboratory of Methods for Big Data Analysis (LAMBDA).

Publications Presented at Thematic Workshops

 Log-density Hessian estimation without the curse of dimensionality via denoising score matching

 Schrödinger bridge problem via empirical risk minimization

 Guided Star-Shaped Masked Diffusion

 Evolutionary Two-Stage Hyperparameter Optimization Strategies for Physics-Informed Neural Networks

 Discrete Flow Matching for Regulatory DNA Sequence Design

 Unlocking the Potential of Weighting Methods in Federated Learning Through Communication Compression

Thanks to the support of HSE Faculty of Computer Science partners—Yandex and Sber—second-year students from the AI360: Artificial Intelligence Engineering track of the Applied Mathematics and Information Science bachelor’s programme were also able to attend the conference as participants. Many students noted that the ICLR format and programme served as an important source of inspiration and fresh ideas for their future research.

‘I received a great deal of positive feedback at the conference regarding an idea I had been considering recently: the use of the frequency domain in model architecture. I was especially fortunate that this year’s conference included a dedicated workshop on time series,’ shared Oleg Kudashin.

‘ICLR was my first experience attending an A* level conference. I was genuinely impressed by its scale and the sheer number of participants. There were researchers from all over the world, as well as many colleagues from Russia and HSE,’ added Maxim Lomakin.