Published in RecSys 2025
Kirill Khrylchenko, Vladimir Baikalov, Sergei Makeev, Artem Matveev, Sergei Liamaev
A refined logQ correction for in-batch sampled softmax that removes the residual bias from the always-present positive item, giving consistent gains for large-scale two-tower retrieval on public and proprietary datasets.
Published in KDD 2026, Jeju Island, Republic of Korea
Kirill Khrylchenko, Artem Matveev, Sergei Makeev, Vladimir Baikalov
A recipe for scaling generative-recommender transformers to one billion parameters by decomposing autoregressive user-history modeling into feedback and next-item prediction; deployed on a large-scale music platform for +2.26% listening time and +6.37% likes in online A/B tests.
Published in SIGIR 2026
Artem Matveev, Vladislav Tytskiy, Sergei Makeev, Sergei Liamaev
A linear-time bidirectional attention layer (GBLA) for generative retrieval that matches full self-attention quality while reaching up to 8.2× single-layer speedup on very long user histories.
Published in KDD 2026, 5th Workshop on End-to-End Customer Journey Optimization
Artem Matveev, Sergei Makeev, Aleksei Krasilnikov, Vladimir Baikalov, Sergei Liamaev, Kirill Khrylchenko
Training generative retrievers with off-policy REINFORCE to optimize long-term user satisfaction instead of next-item accuracy, with doubly robust evaluation and test-time scaling that improves cumulative session reward on Yambda-5B.
Published in arXiv preprint
Sergei Makeev, Artem Matveev, Vladimir Baikalov, Kirill Khrylchenko
An industrial case study comparing pretrained GNN item embeddings with end-to-end trainable ID embeddings for transformer-based sequential recommenders across Yandex Market, Yandex Music, and a public low-resource Yandex Lavka dataset.
Published in arXiv preprint
Sona Team: Alexandr Udeneev, Aleksei Krasilnikov, Alexey Nadtochiy, Andrey Semenov, Andrey Tsyrkunov, Anna Krivonos, Anna Lipkina, Artem Matveev, et al. (34 authors)
A single generative recommender that replaced the entire production cascade at Yandex Music — 15+ candidate generators plus pre-ranking and ranking models — while lifting Active Users by 4.53%, Total Listening Time by 6.30%, and Likes by 11.42% in an online A/B test.
Published:
Talk at Yandex For ML on graph neural networks for item representation in recommender systems.