Yandex's Gryphon-v2 replaces 15-stage recommender cascade with one model

Industrial recommender systems typically rely on multi-stage cascades: separate candidate generators, pre-rankers, and final rankers, each processing user history repeatedly and requiring complex feature pipelines and multiple serving stages. Gryphon-v2, from Yandex researchers, collapses this architecture into one model. It encodes a user's history a single time, autoregressively generates Semantic-ID candidates, resolves them to actual catalogue items, and ranks them with an item-level Ranking Module that reuses the shared encoder states. To transfer fine-grained production ranking preferences without adding a second serving model, the team uses rollout distillation: a high-capacity, training-only Teacher Ranker provides the sole ranking supervision, with scores collected over two complementary candidate distributions—decoder rollouts that mirror serving-time generation, and logged impressions showing items users actually saw.

In an online A/B experiment on a large-scale recommendation surface at Yandex Music, a single Gryphon-v2 model successfully replaced a production cascade containing more than 15 candidate generators, pre-ranking, and final ranking. The deployment increased active users by 1.41% while keeping serving latency comparable to the original cascade. This demonstrates that a generative retriever with a distilled Ranking Module is a practical, end-to-end alternative to production cascades, potentially simplifying infrastructure and enabling more holistic optimization of recommendation quality. The paper, authored by Anna Lipkina and eight colleagues, is available on arXiv.