Yandex's X-split Argus compresses user history into tokens for cheaper CTR ranking

_reachsumit · x · 2026-11-20

Yandex researchers present X-split Argus, an ad CTR ranking framework that compresses long user histories offline into reusable tokens; an online ranker scores ads via cross-attention over these tokens plus recent events, trained end-to-end with an autoregressive click loss and distillation from a large teacher. Offline NE improves 0.81% over the production baseline; deployed in a large-scale RTB system, online inference cost rises 15% but an A/B test shows +4.65% on a revenue-per-conversion business metric.

Original post →

More from Research

Research channel →