Full talk · edited transcript

A Hybrid Generative-Ranking Model Based on Semantic IDs in Yandex Music Recommendations

Daria Tikhonovich23 May 2026

Daria Tikhonovich, head of the R&D team for recommender technologies, Yandex.

The team develops semantic, generative and foundation technologies for recommender systems. The goal is to get from recommender models the same gain that LLMs get from the Next Token Prediction (NTP) task. The difference is the modality: the models are trained on behavior. Their input is a sequence of user actions in the service — items and the feedback users leave on them — and their output should be the next "best" item, one that makes the user happy or brings benefit to the service. The talk describes Gryphon, a hybrid model developed for Yandex Music: what breaks when the generative approach is moved to a real catalog, and how it is fixed architecturally.

Continue reading — it's free

Google or a one-time email code. No password.