Yandex introduced the generative model Sona, which completely takes over the preparation of recommendations – from selecting options to ranking. This is the first model that replaces an entire cascade of algorithms and does not require manual feature selection. TASS was informed about this by the company.

Sona analyzes what the user chose, what they rejected, and what they returned to, and immediately offers suitable options. Previously, each algorithm was trained separately. The first service for testing was music recommendations in smart speakers with Alice: for a week, the new model replaced the existing system of dozens of algorithms.

The listening time for recommended tracks increased by 6.3%, requests to repeat a track – by almost 18%. Active listeners increased by 4.5% – 2.4 times more than after the previous update.

Abandoning the cascade in favor of a single generative model simplifies development and speeds up updates. Previously, improving recommendations required configuring each algorithm separately. Now the system learns from the user's behavior as a whole.