Yandex published a technical report detailing Sona, a unified generative recommender model designed to replace traditional multi-stage recommendation cascades. Tested in live production on Yandex smart speakers, Sona consolidates candidate generation, pre-ranking, and ranking into a single served transformer architecture without hand-engineered features. The system utilizes learned Semantic IDs generated from multimodal audio features and user interaction logs to drive pure-recommendation feeds.
Why it matters
Offers ML engineers a blueprint to replace complex multi-model candidate pipelines with single end-to-end generative transformers.
Demonstrates significant inference cost savings by employing asymmetric attention depth across long historical user interaction logs.
Source: marktechpost.com



