Gryphon: A Unified Architecture for Semantic-ID Generation and Item-Level Scoring in Industrial Recommendations
Researchers have developed Gryphon, a new recommendation system architecture that unifies semantic ID generation and item-level scoring. Unlike previous methods that optimized token sequences, Gryphon directly re-scores generated IDs to improve item relevance. In an A/B test on a music service, Gryphon achieved higher recall and simplified the candidate generation system by replacing multiple components, without significantly impacting user listening time. AI
IMPACT Simplifies recommendation systems and improves item relevance by directly scoring items rather than token sequences.