Yandex presents new AI research at ICML 2026 on faster model training and more efficient compute

Yandex researchers and engineers are presenting new papers at the International Conference on Machine Learning, one of the world's leading conferences in AI and machine learning. ICML 2026 is taking place this week in Seoul.

The research focuses on practical challenges that matter across the AI industry: making better use of GPU resources, accelerating the training of large models, improving machine learning on complex data, reducing the need for costly manual labeling, and making search and recommendation systems more efficient.

ICML brings together researchers from leading universities, research centers, and technology companies around the world. This year, the conference received 23,918 paper submissions, with 6,352 accepted to the main program — an acceptance rate of 26.6%.

Making AI compute more efficient

As AI models grow larger, one of the industry's central challenges is how to train and run them more efficiently without simply adding more GPUs. Several of Yandex's ICML papers address this problem from different angles.

One paper focuses on graph neural networks (GNNs), a class of models that analyze both objects and the connections between them. These models are used in tasks involving linked data, such as relationships between users and products, documents, or road network segments.

Yandex researchers developed new software modules that help graph neural networks run more efficiently on GPUs. In experiments, they improved overall performance by up to 8.5x, reduced peak memory consumption by up to 76x, and sped up individual operations by 3.9–10x. Yandex has released the code as open source

The paper was selected as a Spotlight paper at ICML, a recognition awarded to papers with some of the highest scores from the program committee. This year, 536 papers received Spotlight recognition, representing 2.2% of all submissions.

Another paper looks at how to accelerate the training of large language models. When LLMs are trained using pipeline parallelism, different GPUs work on different parts of the model. This can improve scalability, but some GPUs may still spend time waiting for others to finish their work.

Asynchronous training can reduce this idle time, but it has often been considered unstable for LLMs because it relies on delayed gradients. Yandex researchers showed that the problem is not the delay itself, but how the training process handles it. In particular, they found that AdamW, a widely used optimizer, is sensitive to stale gradients, while newer methods such as Muon are much more robust.

The team also introduced a lightweight correction method, inspired by Error Feedback, to reduce errors during asynchronous updates. In experiments with 10-billion-parameter Mixture of Experts models trained on 200 billion tokens, the asynchronous approach matched the quality of synchronous training.

A separate paper introduces two new optimization algorithms, SoftSignum and SoftMuon. Optimizers determine how a model updates its parameters during training, which has a major impact on both training efficiency and final model quality. In experiments, the new methods consistently outperformed several widely used approaches, including AdamW.

Working with complex data

Yandex also presented research on graph and tabular data, two important types of data used in real-world machine learning systems.

For graph-based tasks, Yandex researchers developed GraphPFN, a model pretrained on more than 1.6 million synthetic graphs. It shows strong performance even without additional fine-tuning and, when adjusted for a specific task, outperforms all approaches considered in the paper on most real-world datasets. This can make it faster to develop models for new graph-related tasks while reducing the amount of task-specific training data required.

Another paper examines why modern neural networks can perform well on tabular data, which remains one of the most widely used formats in business and applied machine learning. The researchers showed that these models benefit from accounting for uncertainty in the data and proposed a more efficient way to represent numerical features.

Training models with limited labeled data

In many applied machine learning tasks, large amounts of data are available, but high-quality labels are limited. This is common in fields such as medicine, manufacturing, and other expert-driven domains, where labeling data can be expensive and time-consuming.

To address this challenge, Yandex and co-authors proposed a method that combines a small amount of labeled data with a much larger volume of unlabeled data. The approach can help train models in cases where collecting new labels may be prohibitively slow or expensive.

Improving search and recommendations

Another Yandex paper focuses on search and recommendation systems. In many such systems, the most accurate model is too computationally expensive to apply to every possible candidate.

Yandex's method helps identify the most promising candidates before they are passed to a more precise model for final ranking. This can reduce computational costs while preserving quality in large-scale search, recommendation, and retrieval systems.