How Russian Scientists Are Making Artificial Intelligence More Efficient

Researchers at the Higher School of Economics have developed a method that makes generative flow networks faster, less energy-intensive, and more adaptable—paving the way for greater technological independence.

A New Breakthrough in Russian AI Research

Scientists at the National Research University Higher School of Economics (HSE) have unveiled a method that significantly boosts the efficiency of training generative flow networks—models that build complex objects step by step using forward and backward passes. The new approach improves the balance between these two components, enabling more flexible, faster, and less resource-intensive training.

Generative flow networks (GFlowNets) are a unique type of machine learning algorithm designed to construct objects sequentially. They are already being applied to protein discovery, drug development, and optimizing transportation systems. Traditional training methods consume vast amounts of computing power and time. The new counter-based technique reduces strain on hardware and accelerates model development. This isn’t just an algorithm tweak—it is a step toward a more sustainable and accessible digital infrastructure, and a major leap forward in generative AI training.

Strategic Value for Russian Tech

The breakthrough carries strategic weight for Russia’s technology sector. First, it reduces dependence on foreign high-performance hardware—a critical advantage under conditions of limited access to advanced computing resources. Second, it strengthens the position of Russian research centers globally by demonstrating their ability to create competitive, innovative AI solutions. Such achievements boost the appeal of domestic AI for both business and government, fueling the growth of national AI ecosystems.

We made the search for an optimal solution more like a negotiation, where both sides are willing to shift their positions. In high-uncertainty tasks, the backward model is just a supporting tool that enhances the forward model’s results. We were searching for a way to make the backward model more flexible—and we finally achieved it

Timofey Gritsaev

co-author of the study and research intern at the HSE Faculty of Computer Science’s Center for Deep Learning and Bayesian Methods

Efficiency translates directly into cost savings: less energy, fewer servers, and shorter training cycles. That means Russian companies and public agencies can adopt advanced AI tools without investing millions in hardware, an especially crucial factor in the era of import substitution and digital sovereignty.

Impact on People and Global Science

For ordinary citizens, the method promises faster rollouts of technologies critical to daily life, from medical research to smarter public transport and environmental analysis. For Russia as a whole, it reinforces technological sovereignty, signaling the country’s ability not only to adapt but also to set trends in global AI research. Internationally, the method could find applications in projects across medicine, logistics, and climate science—fields where efficiency and scalability are paramount.

The Russian AI School: Building Toward Efficiency

This discovery is part of a broader movement in Russia toward resource-efficient AI. Over the past five years, researchers from Sberbank and MIPT have developed algorithms for data compression and sparse synchronization in distributed training. Teams from HSE and the Artificial Intelligence Research Institute (AIRI) proposed the 'Group-and-Shuffle' method, which accelerates fine-tuning without sacrificing accuracy.

All these projects share a common goal: making AI not just smarter, but cheaper to operate. HSE’s latest method continues this trajectory, positioning efficiency as the cornerstone of Russia’s AI research tradition.

From Labs to Real-World Applications

The new method is already a candidate for deployment under Russia’s national Digital Economy program. It could be adopted by startups and major firms in areas like pharmaceuticals, logistics, and education—domains where AI model speed and efficiency make a decisive difference.

One major opportunity lies in combining the approach with other domestic technologies, such as the AltAI neuromorphic processors. Together, they could form a closed ecosystem of Russian AI—from algorithms to hardware. In a broader perspective, Russia could transition from being a supplier of raw resources to an exporter of cutting-edge digital technologies, with AI as a central driver.

Efficiency as a Competitive Edge

HSE’s development is more than an academic achievement—it is a signal that Russia is positioning itself among global leaders in energy-efficient, scalable, and practical AI solutions. In an era of limited computing capacity and surging demand for AI, efficiency may prove to be the most valuable competitive advantage.