Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation

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. Code is available at https://anonymous.4open.science/r/inforID-F582.

Improvements to AI systems based on this paper:

1. Adaptive semantic ID construction for generative recommenders – Replace fixed-length, fixed-codebook semantic IDs with a budget-aware allocation that automatically determines the optimal number of tokens and per-token vocabulary size. This reduces wasted prediction capacity on low-utility slots, improving recommendation accuracy without increasing inference cost.

2. Parallel decoding with heterogeneous codebooks – Enable one-step parallel generation where each position can have a different codebook size, allowing the model to allocate more expressive power to semantically rich subspaces and less to redundant ones. This makes the system faster (no beam search) and more accurate than uniform parallel baselines.

3. Automatic capacity budgeting – Introduce a lightweight module that, given a fixed total capacity (e.g., total number of possible ID combinations), distributes it across semantic slots based on their information gain. This removes the need for manual hyperparameter tuning of ID length and codebook sizes.

4. Improved long-tail item recommendation – By avoiding over-allocation to trivial slots, the system can better represent rare or complex items that require longer, more detailed IDs, improving coverage and diversity in recommendations.

What the improved AI system can do:

• Generate item recommendations in a single forward pass (no iterative decoding), with latency comparable to simple classification but with richer, adaptive ID representations.

• Automatically adapt its ID structure to the dataset’s semantic complexity—e.g., using longer IDs for categories with many fine-grained distinctions and shorter IDs for coarse categories.

• Achieve higher hit rate and NDCG under the same computational budget as existing parallel recommenders, while being more robust to datasets with varying item distributions.

• Serve as a drop-in replacement for current semantic ID–based generative recommenders, requiring only the addition of a capacity-allocation layer, without changing the underlying transformer architecture.