Yandex Introduces Sona: A Single Generative Recommender That Replaces Entire Recommendation Cascade
One transformer replaced Yandex's entire recommendation pipeline—and lifted engagement 11.42% in A/B test.
Why it matters
Yandex Music deployed a single generative model (Sona) to replace multi-stage candidate-generation-and-ranking cascade, eliminating hand-engineered features and achieving measurable lift. This is a concrete capability shift: end-to-end learned ranking at scale, not architectural novelty alone.
The key facts
13 to knowSingle transformer model replaced multi-stage recommendation cascade
Eliminated hand-engineered features entirely
A/B test result: 11.42% lift in likes (user engagement metric)
Deployed in Yandex Music production
Model performs both candidate generation and ranking in one pass
No architectural details or model size disclosed
No comparison to baseline latency, compute cost, or other metrics beyond engagement
Yandex Music A/B test: single transformer for candidate generation and ranking
Engagement lift: 11.42% increase in likes
No hand-engineered features required
Model: Sona (generative recommender)
Replaces traditional multi-stage cascade architecture
Real deployment context: music streaming service
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: One transformer ran candidate generation and ranking in Yandex Music's A/B test without hand-engineered features, lifting likes 11.42%.