Continuous-Query Generation Removes SID Constraints in Recommendation
October 12, 2026
CQG-Single and CQG-AR methods replace discrete Semantic ID (SID) decoding with continuous-query paradigms in generative recommendation. This approach removes the reachability constraints of finite-beam search and eliminates the need to align discrete SID vocabularies with pretrained vector spaces.
HOW THIS AFFECTS YOU
●
builderYou can improve recommendation recall by allowing models to score every item in a catalog directly via dense retrieval.
●
researcherYou can bypass the discrete vocabulary alignment problem in generative retrieval models.