Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation
2026-08-10 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors explain that traditional recommenders use a slow method called beam-search to predict item IDs, which limits how long these IDs can be. To make predictions faster, they use parallel methods that predict all parts of the ID at once but assume all parts of the ID are the same size and importance. The authors point out this assumption wastes capacity on less useful parts. They propose InforID, which smartly assigns different sizes to each ID part based on its importance, improving accuracy without needing more prediction steps.
autoregressive recommendationbeam search decodingsemantic IDparallel generationcodebook sizesemantic slotscapacity allocationInforIDrecommendation accuracyitem identifiers
Authors
Chenxi Li, Yuchen Lu, Xu Yang
Abstract
Autoregressive semantic ID recommenders are constrained by expensive beam-search decoding, which limits the practical length of item identifiers. Parallel generation methods alleviate this bottleneck by predicting all semantic ID tokens simultaneously, enabling longer IDs. However, existing semantic ID methods still rely on manually predefined and homogeneous ID structures, where both the number of semantic slots and the codebook size of each slot are treated as fixed hyperparameters. This ignores the heterogeneous capacity demands of different semantic subspaces and may allocate prediction capacity to slots with limited utility. We show that uniformly expanding semantic slots can provide limited gains, indicating redundant capacity in homogeneous semantic IDs. We propose InforID, a lightweight adaptive semantic target construction framework for parallel generative recommendation. InforID allocates a fixed capacity budget across candidate semantic slots, thereby jointly determining the effective ID length and slot-specific codebook sizes. Experiments demonstrate improved recommendation accuracy under comparable capacity budgets while preserving one-step parallel prediction.