At ICML 2026 in Seoul, Yandex unveiled a suite of breakthrough AI research tackling core bottlenecks in training large models, GPU memory usage, graph neural networks, and limited labeled data. All of Yandex’s papers landed in the conference’s main program, with one earning the prestigious Spotlight designation-an honor granted to just 2.2% of submissions this year (536 out of 23,918).
The standout achievement involves graph neural networks, where Yandex researchers developed software modules that sped up computations by up to 8.5 times while slashing peak GPU memory usage by a staggering 76-fold. This isn’t just academic-graph models power recommendation engines, search, logistics, and traffic network analysis. Dramatically reducing memory use means a single GPU server can handle more tasks simultaneously, boosting efficiency in real-world applications.
Yandex accelerates AI training with graph neural network innovations
Another key study addresses training large language models with pipeline parallelism, a setup where some accelerators remain idle while others finish processing. Yandex tackled the notorious instability of asynchronous training by fine-tuning optimizers and applying update corrections. Their approach matched the accuracy of fully synchronous training on Mixture-of-Experts (MoE) models with 10 billion parameters-architectures favored by major players from Mistral to Google for scaling model size without a linear surge in per-token compute.
Advancements in large language model training and optimizer algorithms
Yandex also introduced two new algorithms, SoftSignum and SoftMuon, which consistently outperformed standard optimizers like AdamW, the long-standing go-to for transformer training. Additionally, their GraphPFN model, pretrained on over 1.6 million synthetic graphs, achieved high accuracy even without fine-tuning and then surpassed other methods on multiple real-world datasets after adaptation.
Training methods addressing limited labeled data and computational costs
The research goes beyond speed, addressing cost efficiency too. Together with partners, Yandex proposed a training method designed for scenarios with scarce labeled data but abundant unlabeled data-a common challenge in medical and industrial fields where expert labeling is expensive. Another technique targets search and recommendation systems by pre-selecting promising candidates for more precise evaluation, significantly cutting computational overhead.
ICML 2026 highlights engineering breakthroughs in AI efficiency
ICML is among the top three global machine learning conferences-alongside NeurIPS and ICLR-where acceptance increasingly reflects not just theory but solid engineering contributions. With the persistent shortage of high-end accelerators and rising AI training costs, innovations that reduce memory consumption, simplify GPU deployment, and minimize manual data labeling are moving rapidly from research papers to practical tools ready for industry use.
The future of cost-effective AI training and deployment
Going forward, Yandex’s results highlight a growing trend: accelerating AI training isn’t just about raw compute power, but smarter algorithms and efficient resource use. The challenge now is how these breakthroughs will scale across diverse AI workloads and hardware setups outside research labs-and whether they can catalyze a new wave of cost-effective AI deployment globally.