Snapchat's SnapLGR uses LLMs to boost view time by 0.37%

Snapchat researchers have released details on SnapLGR, a production LLM-based generative retrieval system for short-video recommendation. The core idea is to treat a pretrained large language model as a retrieval engine that directly generates video IDs, rather than relying on traditional two-stage candidate generation. To make this feasible, SnapLGR constructs semantic identifiers (SIDs) from multimodal item embeddings, then enriches them with Personalized PageRank-based co-engagement contrastive learning — this improves codebook usage, reduces identifier collisions, and injects collaborative filtering signals into the semantic space.

Deployment concerns are addressed with continued pretraining to ground the newly introduced SID tokens before supervised fine-tuning on user interactions, and serving is optimized via TensorRT-LLM CUDA-backed beam search alongside a decentralized worker-loop architecture. In a live A/B test against the existing TIGER-style generative retrieval baseline, SnapLGR delivered a 0.37% increase in View Time, 0.09% in Time Spent, 0.18% in Deep Sessions, and 0.11% in Deep Sessions Unique User. The offline decomposition reveals that joint design of representation learning, vocabulary grounding, and efficient training/serving is critical for production viability.