New EGR framework uses shared LLM for improved generative retrieval

Researchers have developed EGR, a new framework for embedding-native generative retrieval designed for large-scale recommendation and advertising systems. EGR utilizes a single Large Language Model (LLM) to learn both item and user representations within a shared embedding space, directly indexing items as dense vectors and encoding user histories as retrieval queries. This joint contrastive training approach aims to improve user-item alignment and has demonstrated superior performance on benchmarks, including a 2.91% conversion-rate lift in production by simplifying system design and enhancing retrieval quality. AI

IMPACT This framework could streamline recommendation and advertising systems by improving user-item alignment and performance through a unified LLM approach.

RANK_REASON The cluster contains a research paper detailing a new technical framework for generative retrieval. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.IR (Information Retrieval) →

• alphaXiv
• Amazon Reviews
• arXiv
• CatalyzeX
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• LLM
• ScienceCast
• Snap DPA

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →