简体中文 · English
Sona: Replacing a music recommendation cascade with one model¶
Fidelity: core-mechanism reproduction. Trains a compressed-history encoder, three-level autoregressive Semantic ID head, and item ranker. The negative local transfer result is retained.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2608.11015 |
| Institution | Yandex |
| First public release | 2026-08-11 (arXiv v1) |
| Author code | No author-released code found (checked 2026-08-13). |
| Local adapter or method key | sona |
| Local reproduction code | src/auto_research/reproductions/sona/ |
The core training implementation is in latest_20260813_common.py.
Original paper summary¶
Background and main changes¶
Yandex Music's previous stack had more than 15 candidate generators and separate pre-ranking and ranking stages. Sona represents users with a chronological event encoder, compresses history to limit long-sequence cost, autoregressively generates Semantic IDs, and ranks items using a Ranking Module conditioned on the same state. A high-capacity teacher supplies distillation supervision only during training.
flowchart LR
H["chronological events"] --> C["history compression"]
C --> E["shared encoder"]
E --> D["autoregressive SID decoder"]
E --> R["item ranking module"]
T["training-only teacher"] --> R
D --> O["single serving model"]
R --> OKey figure from the paper¶
Figure 2: The core architecture, its main modules, and their connections. From the paper; copyright belongs to the authors. Click for the source.
Core equations¶
Offline and online results¶
In a live My Vibe smart-speaker A/B test, Active Users rose +4.53%, Total Listening Time +6.30%, and Likes +11.42%. The paper reports all three as statistically significant. One model replaced the prior candidate generation and ranking cascade.
Local reproduction¶
Local comparison: A same-budget shared-encoder ranker is compared with history compression + SID decoder + item ranker. NDCG@10 changes by −92.05%.
The MovieLens-1M run uses 260 users, 420 items, 50 steps, and seed 42. The shared-encoder baseline scores 0.02083 NDCG@10, versus 0.00166 (−92.05%) for compressed SID + ranker, which runs 7,200 SID-token targets with 4× history compression. Random ID decomposition and short training cannot recover the neighborhoods of the private semantic tokenizer. This is a negative result.
Stable metrics: metrics/movielens1m-seed42.json.
Scope¶
The reproduction does not include the private music semantic tokenizer, full teacher, 15+ production candidate sources, or smart-speaker serving. It verifies only the core training and inference data flow.