Reproduce “Sona Technical Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sona Team, Alexandr Udeneev, Aleksei Krasilnikov, Alexey Nadtochiy, Andrey Semenov, Andrey Tsyrkunov, Anna Krivonos, Anna Lipkina and 27 more
Yandex Music's Sona Team introduced SONA, a single-model generative recommender that consolidates candidate generation and ranking within a transformer-based architecture, operating without hand-engineered features. Online A/B tests on the "My Vibe" smart speaker surface demonstrated a +4.53% increase in Active Users and an +11.42% increase in Likes, outperforming previous production models.
Key insights
- A unified generative model can effectively replace complex, multi-stage production recommender cascades, achieving superior performance in real-world applications.
- Large-scale sequence models, trained exclusively on logged event fields and learned Semantic IDs, can obviate the need for hand-engineered features, simplifying system development and maintenance.
- Integrating History Compression with a shared user representation allows for efficient processing of long user histories, making single-model architectures viable for production inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.