Snapchat Launches LLM-Based Generative Retrieval System for Short-Video Recommendation
_reachsumit · x · 2026-08-03
Snapchat introduced SnapLGR, an LLM-based generative retrieval system designed for short-video recommendation.
The system features three main designs:
- Multimodal Semantic IDs: Constructs semantic identifiers (SIDs) from multimodal item embeddings, enhanced with Personalized PageRank (PPR)-based co-engagement contrastive learning to improve codebook utilization and reduce collisions.
- Two-stage Training: Uses continued pretraining (CPT) to ground SID tokens before supervised fine-tuning (SFT) on user interaction sequences.
- Practical Serving: Employs TensorRT-LLM CUDA-backed beam search and a decentralized worker-loop architecture to meet strict latency and cost constraints.
Live A/B tests confirmed significant improvements in recommendation performance.
More from Research
- NeurIPS Peer Review is Broken: Exponential Submissions Meet LLM-Generated Reviews — RexDouglass · 2026-08-03
- LLM Safety Trilemma: Useful Capability, Reliable Safety, and Open Access Cannot Coexist — Pingyu Wu · 2026-08-03
- Replicating the Test: Can LLMs Spell Words Based on Audio? — theshawwn · 2026-08-03
- Analysis: Why AI Math Breakthroughs Lagged Behind Coding and What's Next — bookwormengr · 2026-08-03
- StatsMLlib: A Lean 4 Library Formally Verifying Probability, Statistics, and ML Theorems — ChengleiSi · 2026-08-03
- Don't build custom tools for agents: apply the Bitter Lesson — willccbb · 2026-08-03