Yandex unveils Sona, a single generative recommender replacing the whole cascade

Yandex detailed Sona, a generative model that merges candidate generation and ranking into one system instead of a multi-stage cascade. In a 7-day A/B test on its smart speakers, Sona replaced 15+ candidate generators, the pre-ranker and the ranker with a single served transformer: active users +4.53%, listening time +6.30%, likes +11.42%.

• Sona replaced 15+ candidate generators, pre-ranking and ranking with one model
• A/B test: active users +4.53%, total listening time +6.30%
• Likes rose 11.42%, "repeat" commands 17.99%
• A frozen 0.6B-parameter teacher trains the ranker and is dropped at serving