Experts announced a new stage in the development of recommendation systems

Developers of recommendation systems began to replace cascades of specialized algorithms with unified AI models. This approach can speed up the development of recommendations and simplify their use in different services, experts said.

Nikita Zelinsky, Director of machine learning and data research at MTS Web Services, says that recommendation systems can already talk about the beginning of their "LLM moment." This is not about turning recommendations into language models, but about changing their architecture.

"Instead of a cascade of many specialized models, one large Transformer backbone appears, trained on a long history of user behavior and reused for different tasks," he explained.

According to Zelinsky, such models prove the "laws of scaling" — they show an increase in quality as the size, volume of data and computing resources increase. Meta (recognized as extremist and banned in the Russian Federation), USTC and Huawei are working in this direction, among others. MTS Web Services also creates its own unified Transformer model for recommendations.

A practical example of the new approach is the generative Sona model, which was trained from scratch by Yandex. The AI model is able to replace dozens of recommendation algorithms — it completely takes over the preparation of recommendations, from the selection of options to their ranking. This is the world's first model that replaces a cascade of algorithms, does not require manual feature selection and shows good results in practice, the press service comments. The model was tested in smart speakers with "Alice". According to the company, in the test group, the listening time for recommended tracks increased by 6.3%, the number of requests to repeat a song increased by almost 18%, and the audience of active listeners increased by 4.5%.

Switching to a single model can reduce development time. As Zelinsky explained, teams will not have to train a separate model from scratch for each new scenario and maintain a large number of algorithms. "As a result, it's easier for the team to launch new scenarios.: Instead of building a separate recommendation system for each product, you can reuse the overall model and adapt it to a new surface or goal. This shortens the cycle from hypothesis to production and makes experiments cheaper and faster," the expert emphasized.

Alexey Sidoryuk, Advisor to the Director General of Artificial Intelligence at the Fintech Association, drew attention to the practical value of this approach. "For businesses, this means the opportunity to develop recommendation services faster, transfer them to new products more easily, and work more efficiently with huge catalogs of content and products. It is especially gratifying that such a development appeared in Russia, was confirmed by an experiment on real users and published in the form of a scientific article," he said.

Pavel Golosov, director of the Institute of Social Sciences at the Presidential Academy, called SONA an example of "mature sovereign development in AI" and pointed out that the result was obtained by the Russian engineering team on its own infrastructure and data, but at the same time integrated into a global scientific context.

"Recommendation systems are one of the most knowledge—intensive and closed areas of applied AI: the architectures underlying the feeds and recommendations of large platforms have been considered the prerogative of Western companies for years, and the details of their device are rarely publicly published," Golos concluded.