Yandex Introduces Sona: A Single Generative Recommender That Replaces Entire Recommendation Cascade

Yandex’s Sona report is a shot across the bow for the old recommendation cascade. Instead of one model finding candidates, another filtering them, and a third making the final call, Sona folds the whole job into a single generative system. Yandex says it tested that switch in a seven-day live experiment on its smart speakers, and the new setup replaced more than 15 candidate generators plus the pre-ranker and ranker.

The cleaner part of the idea is also the bluntest: no hand-engineered features. Sona works from logged event fields such as track ID, artist ID, duration, likes, played time, and surface flags, along with learned Semantic IDs. For Yandex Music, that matters because playback on smart speakers can start without anyone first picking an artist, genre, or mood. The team calls that a pure-recommendation setting, which is a polite way of saying the model has to do the hard part on its own.

Under the hood, Sona leans on a shared user representation. The encoder reads listening history once, the decoder generates candidates, and a separate Ranking Module scores them against the same encoder state. The history handling is split: the most recent 2,048 events get heavier attention, while older events go through a lighter path. Yandex says that keeps most of the quality of full attention while cutting inference cost roughly in half.

The item side is equally engineered, just not by humans hand-crafting features. Each track is turned into three discrete codes. A frozen multimodal LLM reads the first 90 seconds of audio plus title, artists, and tags, then a four-layer refinement transformer lines those signals up with listening behavior. Residual K-means then compresses the result into three codebooks of 32,000 entries each. Yandex says that setup beat a CLMR audio baseline, with Recall@1000 rising from 0.8111 to 0.8524.

What makes this more than a neat architecture diagram is the live result. In the seven-day A/B test, on 15% of randomly selected users per split, Sona lifted active users by 4.53%, total listening time by 6.30%, likes by 11.42%, repeat commands by 17.99%, and deeply engaged users by 7.37%. Those gains sit on top of earlier improvements, and on active users the uplift was 2.35 times the +1.93% improvement Argus had previously delivered on the same surface.