LLM-Enhanced Topical Trend Detection at Snapchat

Snapchat has unveiled a production-grade system for automatically detecting emerging topical trends on its short-video platform. Published as arXiv:2604.27131, the system integrates three key components: multimodal topic extraction from video and text, time-series burst detection to spot sudden interest spikes, and LLM-based consolidation and enrichment to produce coherent trend summaries. This marks the first end-to-end published system for short-video trend detection at scale, addressing a major challenge in maintaining a dynamic content ecosystem.

Continuous offline human evaluation over six months confirmed high precision in identifying meaningful trends. The system is already deployed globally and applied to downstream surfaces like content ranking and search, driving measurable improvements in content freshness and user experience. By leveraging LLMs to consolidate fragmented signals, Snapchat ensures that trending topics are accurate, timely, and actionable for both the platform and its users.