On October 5, Asif Razzaq reported in MarkTechPost that Yandex has launched a generative recommendation system named Sona. The core takeaway is straightforward: in A/B testing conducted on Yandex Music, a single Transformer handles both candidate generation and ranking simultaneously, requires no hand-engineered features, and delivered an 11.42% increase in user likes.

A conventional recommendation system is built as a multi-stage pipeline. It first retrieves hundreds to thousands of candidate items from a catalog containing millions or even billions of entries, then goes through coarse ranking, fine ranking, and sometimes re-ranking and business rule adjustments. Each stage relies on independent models and features. While this architecture is stable for engineering, it comes with high maintenance costs, and the objectives of different stages can easily become misaligned. If a relevant item is missed during the initial retrieval stage, no amount of ranking power in later stages can bring it back.

Sona was built specifically to address this problem. It unifies candidate generation and ranking into a single generative model, where the Transformer directly outputs recommendation results and eliminates the need for large volumes of manually engineered features. From an engineering perspective, this means fewer pipeline stages, fewer separate feature processing pipelines, and less error accumulation across stages.

Only one confirmed result is available so far: the 11.42% lift in likes from the online A/B test on Yandex Music. The report has not disclosed the model size, inference cost, testing duration, or control group baseline, so it is too early to conclude that this approach works for all recommendation scenarios.

The 11.42% lift is measured in user likes, which does not automatically translate to increases in session length, user retention, or revenue. Music is inherently a high-frequency, strong-interest consumption scenario, where user feedback is dense and well-suited for models to learn user preferences from behavior sequences. For other domains like news, video, or e-commerce, additional challenges need to be addressed, including content cold start, timeliness, diversity, ad load requirements, and latency constraints.

The key insight from this work is that the deployment of large models in recommendation does not have to start with a chat interface. Instead, it can begin directly by replacing the traditional multi-stage recommendation cascade. If a single model can consistently outperform conventional pipelines in online A/B testing, it will not only eliminate redundant modules but also pay off the massive technical debt accumulated from years of adding handcrafted features.

Source: [Asif Razzaq / MarkTechPost](https://www.marktechpost.com/2026/10/05/yandex-introduces-sona-a-single-generative-recommender-that-replaces-entire-recommendation-cascade/)