Sona Technical Report
Published in arXiv preprint
Sona is a single-model generative recommender for Yandex Music that unifies candidate generation and ranking around a shared user representation. Its encoder turns the user’s chronological sequence of logged engagement events into hidden states consumed by both the autoregressive decoder and the Ranking Module, with next-token-prediction and distillation objectives jointly updating the encoder so that generation and ranking are coupled through the same user state. Neither Sona nor its Teacher Ranker uses hand-engineered features; the larger teacher supplies ranking targets during training but is absent from serving. In an online A/B test on My Vibe on smart speakers, Sona replaced the entire production cascade — more than 15 candidate generators followed by pre-ranking and ranking models consuming hundreds of features, including signals from Argus — and delivered statistically significant uplifts of 4.53% in Active Users, 6.30% in Total Listening Time, and 11.42% in Likes, with the Active Users uplift 2.35× the increment previously delivered by Argus on this surface.