REGER paper adds explicit item relations to generative recommendation via semantic IDs
_reachsumit · x · 2026-10-29
The arXiv paper REGER addresses a weakness of generative recommendation: randomly initialized semantic ID (SID) embeddings and models trained from scratch don't inherit item semantics, and collaborative relations are only learned implicitly. REGER explicitly injects inter-item relations at three stages: generative semantic retrieval pretraining so the model learns to generate semantic neighbors' SIDs; combining global attention with local attention aligned to SID boundaries to capture history context and directed adjacent-item transitions; and a theoretical framework for hierarchical preference decomposition over the SID prefix tree, proving an exact factorization of collaborative preference distributions. Also distills from a discriminative teacher.
More from Research
- K-Dense open-sources 177 science skills that turn coding agents into research scientists — alvelda · 2026-10-29
- Stanford team releases SOTA prompt optimization framework with unsupervised prompt performance measure — BrihiJ · 2026-10-29
- Debate set: Can we make AIs sentient — and should we? Solms vs. Humphrey — cccalum · 2026-10-29
- Samaya's 30B post-trained model beats frontier on financial research at 11x lower cost — maithra_raghu · 2026-10-29
- Scaling cross-situational word learning: 2,441 hours of child-perspective video from Stanford team — RichmanRonald · 2026-10-29
- Curiosity algorithms need world models plus locomotor primitives to replicate animal behavior — aran_nayebi · 2026-10-29