- Статьи
- Internet and technology
- Acceleration maneuver: the Russian Federation accelerated the training of AI models by 40%
Acceleration maneuver: the Russian Federation accelerated the training of AI models by 40%
Russian scientists have developed an effective way to quickly train AI models - it saves computing costs by 40% without loss of quality. To do this, we use a set of algorithms that automate the process without manually testing hundreds of parameters: the system independently selects the most effective configurations based on metrics such as accuracy, generation speed, and compliance with the response format. The areas of technology use are wide: government and corporate tasks, as well as industrial AI services - that is, wherever artificial intelligence needs accelerated learning.
How the new training system works
Scientists at the Scientific and Educational Center of the Federal Tax Service of Russia, together with Bauman Moscow State Technical University, have developed an automated scenario for configuring language models that allows AI to be adapted to specific measurable tasks without manually sorting through parameters. The solution is aimed at improving the accuracy, speed of operation and compliance with the specified formats, depending on the purpose of using the model.
The developers explained that the high efficiency of the model is achieved by optimizing key indicators. To speed up the work, surrogate models are used that evaluate different AI configurations in advance. This makes it possible to reduce the number of comprehensive checks of the system for malicious objects by about 1.6 times, reduce the amount of resource-intensive calculations and speed up setup by about 40% of the time without loss of quality. Experiments have shown that the method consistently outperforms traditional approaches, including in terms of code coverage.
"The system automatically selects optimal configurations for different tasks, from maximum speed to the highest accuracy, saving developers from having to manually test hundreds of options," said Igor Masich, PhD, Leading researcher at the REC Federal Tax Service of Russia and Bauman Moscow State Technical University.
According to him, the development is focused on practical applications and can be used in industrial AI services, as well as in the creation of specialized solutions for government and corporate tasks.
— The model offers the developer not one universal setting, but a whole set of optimized options for different tasks. If speed is important to the project, you can choose a fast configuration with good quality. If accuracy is critical, an option with maximum accuracy is available, which works a little longer. Developers set up AI tools themselves based on their goals," the author explained.
Areas of application of the methodology in the field of AI
The development belongs to the class of Automated Machine Learning systems. Simply put, it is an AI that is able to do part of the work of a data scientist: automatically configure and select models for specific tasks based on data. The very idea of AutoML is not new, but the emergence of large language models (Large language model — artificial intelligence programs that can recognize and generate text) requires new approaches and tools, so the technology can be considered relevant and useful, said Alexander Bukhanovsky, PhD, head of the ITMO National Center for Cognitive Research.
— The technology of Bauman Moscow State Technical University is focused on individual AI models, but in the future it may be especially in demand when creating multi-agent systems based on large fundamental models. In such systems, the same AI model can be used by different agents, but with different settings - in terms of accuracy, speed and other parameters," the expert specified.
The new development allows us to understand in practice exactly which methods — for example, a variety of responses, the use of external data or simplified models — really improve the result, says Igor Terekhin, head of the GenAI DAR Competence Center (CORUS Consulting Group). In his opinion, the solution has serious prospects for implementation.
— AI is no longer an experiment "for the future" and is becoming a tool for increasing business sustainability. Companies will launch AI projects with a short payback period of 9-12 months and immediately realize the effect of reducing manual labor, primarily due to AI agents. This will lead to more ambitious and pragmatic implementations of AI embedded in IT architecture and business models," he noted.
Domestic developments in the field of artificial intelligence will be able to find application in a wide variety of fields, says the technical director of Audit Management Company (SL Soft FabricaONE.AI, shareholder — Softline Group) Yuri Tyurin. According to him, such platforms open the way to the mass use of multi-agent AI systems that work as "digital teams", where several models, performing different roles, solve complex tasks together without constant human involvement.
"With the acceleration of model training, we are approaching the emergence of systems capable of independently setting and performing complex chains of tasks," said Artem Zhadeev, Sales Director at Stakhanovets.
The expert is confident that the effect of such solutions is much broader than it might seem at first glance, and such technologies will become one of the notable trends of 2026.
Переведено сервисом «Яндекс Переводчик»
