Yandex Music replaced 15+ candidate generators with one transformer, +6.3% listening time

SettingAccording8986 · reddit · 2026-10-05

Yandex Music shares Sona, a paper where a single end-to-end transformer replaced 15+ candidate generators plus pre-ranking and ranking models in an A/B test.

Key techniques

The model reads up to 8,192 events. To cut inference cost of full attention, they propose History Compression: history is split into an older 6,144 block and the most recent 2,048; the blocks exchange information via cross-attention plus one full-history self-attention layer, after which a 7-layer stack runs only on the recent 2,048. This roughly halves inference cost while retaining most of the quality.
Decoder and Ranking Module share the same encoder output, so the encoder runs once per request; candidates emerge from beam search as Semantic IDs and are scored immediately.

A/B results (smart speakers, 7 days, 15% of users per arm): +4.53% Active Users and +6.30% Total Listening Time, both significant at p<0.01. Catalog coverage is lower than production; the cause is under investigation. A long-term A/B test is underway; not yet at full traffic. Paper: arxiv.org/abs/2608.11015.

Original post →

Yandex Music replaced 15+ candidate generators with one transformer, +6.3% listening time

More from Models