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.

Long-Term Optimization for Large-Scale Generative Retrieval with Off-Policy REINFORCE

Generative retrieval has become a popular paradigm for large-scale recommendation. However, it is typically trained with supervised next-item prediction objectives that do not directly optimize long-term user satisfaction. In this work, we formulate recommendation as a session-level sequential decision-making problem and introduce an autoregressive approach for training generative retrievers 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 recommendations with the highest predicted long-term returns. Experiments on the public large-scale Yambda-5B dataset show that our RL agent improves offline estimates of cumulative session reward over next-item and next-positive prediction baselines, while largely preserving retrieval quality. Moreover, allocating more inference-time compute to simulating future responses improves model-based long-term return estimates without updating the policy.

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.