EGR: Single shared LLM boosts ad conversion by 2.91% in production

Current generative retrieval methods for large-scale recommendation and advertising are complex: Semantic-ID approaches require quantization, mutable identifier vocabularies, and token-to-item grounding, while embedding-based pipelines train the item encoder separately from the query generator, limiting alignment. EGR (Embedding-Native Generative Retrieval) proposed by Xiaodong Liu and 15 others at Snap simplifies this by using a single shared large language model (LLM) to learn both item representations (from item metadata) and user representations (from interaction histories) in one unified embedding space. Items are indexed as dense vectors, user histories become dense retrieval queries, and joint contrastive training groups related items while aligning queries with target items.

EGR was evaluated on public benchmarks (Amazon Reviews), industrial data (Snap DPA), and in live production. It outperforms published baselines on Amazon Reviews, scales with data on Snap DPA, handles cold-start items effectively, and benefits from multimodal input (e.g., images). In a production advertising system at Snap, EGR delivered a +2.91% conversion-rate lift while dramatically simplifying the overall retrieval architecture. The paper has been accepted to RecSys 2026, signaling strong peer recognition. This work promises to reduce system complexity and improve retrieval quality simultaneously, making it highly relevant for any large-scale recommender or ad platform.