Generative recommendation (GR) on social media platforms usually retrieves items by autoregressively decoding discrete Semantic IDs (SIDs) with a finite‑width beam search. This paper examines how such an output interface limits the inclusion of held‑out targets in the finite returned list, distinguishing catalog‑wide scoreability from empirical top‑K support. We introduce two continuous‑query paradigms: CQG‑Single generates a single retrieval query in one forward pass, while CQG‑AR autoregressively produces continuous residual updates, preserving step‑wise decoding yet removing the SID token vocabulary, code lookup, and discrete beam expansion. These designs eliminate the SID‑specific reachability constraint imposed by finite‑beam decoding and avoid the technical challenge of aligning a discrete SID vocabulary with the pretrained backbone’s vector space. Every encoded item becomes directly scoreable via dense or approximate‑nearest‑neighbor retrieval. Under a unified protocol with valid‑prefix constrained decoding, CQG‑Single outperforms the TIGER‑style baseline in all seven runs while using only 55.3% of the trainable parameters. CQG‑AR matches performance on ML‑1M and shows stronger numbers on Beauty and Sports datasets. A controlled depth ablation reveals that additional continuous refinement does not monotonically improve shallow accuracy. These findings suggest decoupling output representation, autoregressive modeling, and retrieval budget when evaluating generative recommendation systems.
Review