CoSID: Concept-Conditioned Semantic-ID Decoding for Efficient Generative Recommendation

Abstract

Generative recommenders retrieve items by autoregressively decoding semantic IDs (SIDs). The standard autoregressive SID interface (SID-AR) represents each history item with K code tokens, expanding a T-item history to TK tokens, while trie-constrained beam search repeatedly re-enters the full backbone during generation. Composing each item’s K code embeddings into a single input token restores item-level history length, but every decoding step still runs through the full backbone. We introduce CoSID, a concept-conditioned SID decoder that encodes the history once into a compact next-item concept and delegates the entire beam search to a lightweight KV-cached decoder. Across four datasets under a global temporal split, CoSID matches or surpasses the baselines in accuracy, maintains comparable or broader catalog coverage, and delivers up to 6.2 × higher throughput at beam width 100 and 7.3 × at beam width 1000. Because autoregressive SID decoding no longer re-enters the backbone, throughput remains nearly independent of backbone depth.