Yandex has put a new kind of recommendation engine to the test, and the results point toward a major shift in how personalized media systems are built. Its Sona model replaces a long chain of recommendation stages with one generative AI system, then proves its value with real smart-speaker users.
The seven-day live production experiment produced gains across every key engagement measure. Listening time climbed, likes rose, replay requests surged, and more users reached a deeper level of engagement with recommended tracks.
One Model Replaces the Recommendation Pipeline
Traditional recommendation systems divide the job into several stages. Candidate generators first find possible choices, a pre-ranking stage narrows them down, and a ranking stage selects what the user sees or hears. Each part can use its own model and optimization process.
Sona takes a different path. In an online A/B test, the generative AI model replaced more than 15 candidate generators, the pre-ranking stage, and the ranking stage with one served transformer. Candidate generation and ranking now belong to one end-to-end system rather than a multi-stage recommendation cascade.
That change also removes the need for hand-engineered features in the system described by Yandex. Sona learns the recommendation task as one connected process, rather than passing decisions through independently optimized models.
Nikolai Savushkin, Head of the Recommendation Systems team at Yandex, described the test as a direct challenge to the standard approach:
“With Sona, we tested whether a single system that learns the recommendation task end to end can replace a complex pipeline of independently optimized models. The production experiment shows that this approach can work not just in research settings, but with real users and a mature recommendation product.”
Smart Speakers Showed the Model’s Reach
Yandex tested Sona on its smart speakers, where playback can begin without the user first selecting an artist, genre, or mood. That makes the recommendation system central to the listening experience: the model must decide what to play before the user provides a detailed request.
The online A/B test delivered a broad set of improvements:
- Active Users increased by 4.53%.
- Total Listening Time increased by 6.30%.
- Likes increased by 11.42%.
- “Repeat” Commands increased by 17.99%.
- Deeply Engaged Users increased by 7.37%.
Yandex also summarized the results in rounded figures. Listening time for recommended tracks increased by 6.3%, requests to replay a track rose by nearly 18%, active listeners increased by 4.5%, likes rose by 11.4%, and highly engaged listeners increased by 7.4%.
These numbers show more than one isolated improvement. A longer listening session points to stronger track selection, while the rise in replay commands shows that users found more recommendations worth hearing again. The gains in likes and highly engaged listeners add another signal: Sona did not only keep people listening, it helped create stronger responses to the music.
A Production Test With a Larger Lift
The increase in active listeners was 2.4 times higher than the gain delivered by the previous recommendation-system update. That comparison gives the result extra weight because it measures Sona against an earlier improvement in the same recommendation environment.
Yandex called Sona the first publicly documented recommendation system to demonstrate in live production that a single model can replace a full multi-stage recommendation cascade without hand-engineered features while delivering statistically significant gains.
That distinction matters. Recommendation research can show that a new architecture works on a controlled benchmark, but a production system faces real users and a mature product. Sona’s seven-day experiment brought the single-model approach into that setting, with changes measured through active listening, total listening time, likes, replay commands, and deep engagement.
“Recommendation systems are a core focus for Yandex and an area where we have built deep expertise over more than a decade,” Savushkin said.
Sona now stands as a test of whether recommendation systems can become simpler in structure while improving the experience they deliver. Yandex’s results suggest that one generative model can handle work once divided among more than 15 candidate generators, pre-ranking, and ranking, opening a new direction for recommendation systems built around end-to-end learning.
Based on
- Yandex Introduces Sona: A Single Generative Recommender That Replaces Entire Recommendation Cascade — marktechpost.com
- Yandex Sona: World’s First AI Model to Replace Full Recommendation Pipeline Without Hand-Engineered Features | Currency News | Financial and Business News | Markets Insider — markets.businessinsider.com
