ByteDance's SepLRM splits recommendation models around a shared KV cache
_reachsumit · x · 2026-11-11
ByteDance researchers propose SepLRM, a recommendation foundation model (RFM) paradigm on arXiv addressing diminishing ROI in large recommendation models. One shared encoder-decoder RFM encodes long user histories once and exports its final-layer encoder state as a candidate-independent KV cache; multiple independently optimized vertical models query this cache while preserving the same query-KV interaction as their end-to-end counterparts, avoiding repeated long-sequence encoding costs across recommendation stages and business scenarios.
More from Research
- Codex tool-use agent trajectory dataset trends on Hugging Face — israellaguan · 2026-11-11
- Three new math books from Kyoritsu: flow matching for statistical inference, optimal transport, information geometry — FrnkNlsn · 2026-11-11
- EmoHarness: scaffolding lifts small models' multimodal emotion recognition by up to 13.4 points — UIUC-CS · 2026-11-11
- Reasoning as pathfinding in token space: a new post on weird RL reasoning behaviors — StanfordAILab · 2026-11-11
- UC Berkeley's SubWM watermark detects LLM distillation even after laundering attacks — sumanthd17 · 2026-11-11
- New grokking study: generalization emerges from spontaneous structural collapse and compression — burny_tech · 2026-11-11