arXiv subjects

Artem Matveev

Publications and source records attributed to Artem Matveev.

6 recordsLinked to original sources

arXiv2026

Sona Technical Report

We introduce Sona, a single-model generative recommender for Yandex Music. In an online A/B test, Sona replaced the entire production cascade, comprising more than 15 candidate generators followed by pre-ranking and ranking models that consume hundreds of features, including signals from large transformer models such as Argus and target-attention scorers, while significantly improving key engagement metrics. The architecture of Sona unifies candidate generation and ranking around a shared user representation. Its encoder transforms the user's chronological sequence of logged engagement events into hidden states consumed by both the autoregressive decoder and the Ranking Module. The next-token-prediction and distillation objectives jointly update the encoder, coupling generation and ranking through the same user state. Neither Sona nor its Teacher Ranker uses hand-engineered features; both operate on logged event fields and learned item representations. In the final Sona configuration, the larger teacher supplies ranking targets during training but is absent from serving, leaving the encoder, decoder, and Ranking Module as a single deployed model. We evaluate Sona in an online A/B experiment using live traffic from My Vibe on smart speakers, one of Yandex Music's largest recommendation surfaces. Relative to the production control, Sona produced statistically significant uplifts of 4.53% in Active Users, the primary metric, 6.30% in Total Listening Time, and 11.42% in Likes. These effects were incremental to improvements retained from preceding deployments. The Active Users uplift was 2.35 times the increment previously delivered by Argus, the strongest model deployed on this surface before Sona. These results show that a single jointly trained model can replace a mature multi-stage recommendation cascade while improving recommendation quality on live traffic.

cs.IR↗

arXiv2026

Embedding Items at Scale: Comparing GNN-Based and ID-Based Item Embeddings in the Yandex Ecosystem

Transformer-based sequential recommendation models, which process sequences of user-item interactions, rely heavily on the item embedding strategy. Existing approaches either use pretrained item embeddings or learn them end-to-end with the transformer. To the best of our knowledge, no prior work has compared these options from both cost and quality perspectives in a large-scale industrial setting. This paper is a case study that compares pretrained industrial graph neural network item embeddings with end-to-end trainable item embeddings across two mature production recommendation systems at Yandex: Yandex Market and Yandex Music. We additionally evaluate both approaches on a low-resource dataset sampled from Yandex Lavka production logs, for which both the data and code are publicly available for demonstration purposes. Our results show that a separate pretraining stage helps when training data is limited, but provides no worthwhile benefit for large-scale models trained on extensive datasets.

cs.IR↗

arXiv2026

Session-Level Optimization for Large-Scale Retrieval using REINFORCE with Multi-Step Off-Policy Correction

Two-tower models are a widely used paradigm for large-scale retrieval in recommendation. However, they are typically trained with myopic supervised objectives, such as next-item prediction, that do not directly optimize long-term user satisfaction. In this work, we formulate recommendation as a session-level sequential decision-making problem and train a two-tower retriever autoregressively with off-policy REINFORCE on pre-collected data. Unlike the one-step off-policy correction used in prior work, we propose a multi-step approximation of importance weights enabled by the autoregressive formulation. To support offline evaluation, we train a user feedback model that simulates user responses to generated recommendations. This lets us adapt doubly robust off-policy evaluation for sequential decision-making to recommendation, a setting that has received limited attention. We further introduce a feedback-model-based test-time scaling procedure that simulates future responses and selects the recommendation with the highest predicted long-term return. Experiments on the public large-scale Yambda-5B dataset show that our RL agent achieves higher off-policy estimates of cumulative session reward than next-item and next-positive prediction baselines, while remaining competitive on conventional retrieval metrics. Moreover, allocating more inference-time compute to simulating future responses yields higher model-based long-term returns without updating the policy.

cs.IR↗

arXiv2026

Gated Bidirectional Linear Attention for Generative Retrieval

In recommender systems, generative retrieval typically uses an encoder-decoder setup: an encoder processes a user interaction history, and an autoregressive decoder then generates recommended items. In large-scale streaming services, active users accumulate very long histories over time. As histories grow, the encoder becomes a major latency bottleneck because softmax attention scales quadratically with sequence length. In our experiments, using bidirectional attention in the encoder substantially improves quality. However, most sub-quadratic attention methods focus on causal attention. We propose Gated Bidirectional Linear Attention (GBLA), a linear-time bidirectional attention layer that extends kernelized linear attention with three lightweight components: local causal mixing (Conv1D), sequence-level key gating for soft forgetting, and a gated RMSNorm output. On a large-scale Yandex Music dataset, a hybrid encoder that interleaves self-attention (SA) and GBLA in a 1:2 ratio (one SA block followed by two GBLA blocks) matches bidirectional self-attention quality. On H100 GPUs, GBLA reaches up to an $8.2\times$ single-layer speedup at a history length of 32768, compared to FlashAttention-v3. Finally, we show that the same hybrid design generalizes beyond our proprietary setting, consistently preserving self-attention retrieval quality on public Amazon benchmarks.

cs.IR↗

arXiv2025

Scaling Recommender Transformers to One Billion Parameters

While large transformer models have been successfully used in many real-world applications such as natural language processing, computer vision, and speech processing, scaling transformers for recommender systems remains a challenging problem. Recently, Generative Recommenders framework was proposed to scale beyond typical Deep Learning Recommendation Models (DLRMs). Reformulation of recommendation as sequential transduction task led to improvement of scaling properties in terms of compute. Nevertheless, the largest encoder configuration reported by the HSTU authors amounts only to ~176 million parameters, which is considerably smaller than the hundreds of billions or even trillions of parameters common in modern language models. In this work, we present a recipe for training large transformer recommenders with up to a billion parameters. We show that autoregressive learning on user histories naturally decomposes into two subtasks, feedback prediction and next-item prediction, and demonstrate that such a decomposition scales effectively across a wide range of transformer sizes. Furthermore, we report a successful deployment of our proposed architecture on a large-scale music platform serving millions of users. According to our online A/B tests, this new model increases total listening time by +2.26% and raises the likelihood of user likes by +6.37%, constituting (to our knowledge) the largest improvement in recommendation quality reported for any deep learning-based system in the platform's history.

cs.IR↗

arXiv2025

Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval

Two-tower neural networks are a popular architecture for the retrieval stage in recommender systems. These models are typically trained with a softmax loss over the item catalog. However, in web-scale settings, the item catalog is often prohibitively large, making full softmax infeasible. A common solution is sampled softmax, which approximates the full softmax using a small number of sampled negatives. One practical and widely adopted approach is to use in-batch negatives, where negatives are drawn from items in the current mini-batch. However, this introduces a bias: items that appear more frequently in the batch (i.e., popular items) are penalized more heavily. To mitigate this issue, a popular industry technique known as logQ correction adjusts the logits during training by subtracting the log-probability of an item appearing in the batch. This correction is derived by analyzing the bias in the gradient and applying importance sampling, effectively twice, using the in-batch distribution as a proposal distribution. While this approach improves model quality, it does not fully eliminate the bias. In this work, we revisit the derivation of logQ correction and show that it overlooks a subtle but important detail: the positive item in the denominator is not Monte Carlo-sampled - it is always present with probability 1. We propose a refined correction formula that accounts for this. Notably, our loss introduces an interpretable sample weight that reflects the model's uncertainty - the probability of misclassification under the current parameters. We evaluate our method on both public and proprietary datasets, demonstrating consistent improvements over the standard logQ correction.

cs.IR↗