Problem
Yandex Music has more than 50 million tracks. Its recommendations come from a stack of many candidate generators plus a pre-ranking module, which is costly to maintain. A generative model on Semantic IDs, trained like an LLM on next-item prediction, could replace the stack. On a real catalog it breaks in two places: similar tracks get identical IDs (collisions), and the usual fix, an index layer that makes IDs unique, overflows as the catalog grows. Without the layer, quality drops.
Example from the article: with the collision-resolution layer the model gets +14% Recall@100 over the production two-tower generator. The same model without the layer gets −44%.
Solution in brief
Gryphon is one encoder-decoder model for candidate generation and ranking. A process reward model (PRM) scores partial Semantic ID prefixes during beam search, so errors do not accumulate. An output reward model (ORM) then scores the concrete items under the generated IDs, which resolves collisions without an extra layer.
Questions this case answers
- How to replace a stack of candidate generators and a pre-ranker with one generative model over discrete item codes for a catalog of tens of millions of items
- How to resolve identical discrete codes shared by different items without an extra index layer that grows with the catalog
- How to replace a multi-model candidate generation and pre-ranking stack with one generative recommender
- How to make a generative recommender on Semantic IDs work on a catalog of tens of millions of items
Tools and methods
- Semantic IDs (RQ-KMeans, 3 × 32,000 codebooks) — hierarchical item tokens
- TIGER-style encoder-decoder — history in, next Semantic ID out
- PRM (process reward model) — scores prefixes inside beam search
- ORM (output reward model) — scores items after generation
- InfoNCE with in-batch negatives and LogQ correction — loss for PRM and ORM
- Qwen 2.5 Omni — item embeddings
Google or a one-time email code. No password.
