Snapchat has launched SnapLGR, a new generative retrieval system for its short-video recommendation service. This system utilizes large language models (LLMs) to improve content discovery by creating semantic identifiers (SIDs) from multimodal item embeddings and enhancing them with contrastive learning. The system also incorporates continued pretraining and supervised fine-tuning on user interaction data, along with an efficient serving architecture using TensorRT-LLM. In live A/B testing, SnapLGR demonstrated significant improvements in user engagement metrics, including a 0.37% increase in View Time and a 0.18% rise in Deep Sessions. AI
IMPACT Enhances content recommendation systems with LLM capabilities, potentially improving user engagement and discovery on social platforms.
RANK_REASON Deployment of an LLM-based system for content recommendation by a major social media platform.
Read on arXiv cs.IR (Information Retrieval) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
