InforID adaptively sizes semantic-ID slots and codebooks in parallel generative recommendation

Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation

Chenxi Li, Yuchen Lu, Xu Yang

cs.AI

2026-08-10

InforID replaces fixed semantic-ID structure in parallel generative recommendation with greedy capacity allocation by reconstruction loss, lifting Recall@5 across four Amazon benchmarks.

What problem this solves

Autoregressive semantic-ID recommenders are bottlenecked by expensive beam-search decoding, which caps how long item identifiers can be. Parallel generation (predicting all semantic-ID tokens at once) removes that bottleneck and allows longer IDs. But existing semantic-ID methods still rely on a manually predefined, homogeneous ID structure: both the number of slots and the codebook size per slot are fixed hyperparameters. This ignores that different semantic subspaces need different representational capacity, and allocating equal capacity to low-utility slots wastes prediction power. The authors confirm this: uniformly expanding slots gives limited gains, evidence of redundant capacity in homogeneous IDs.

Method

InforID is a lightweight adaptive semantic-target construction framework whose core is allocating a fixed capacity budget across candidate semantic slots, jointly setting the effective ID length and per-slot codebook size. Three steps:

• Candidate slots: item representations are projected (via OPQ rotation) into M equal-size subspaces, each a candidate semantic slot.
• Adaptive capacity allocation: greedy bit allocation minimizing reconstruction loss under total budget B. Start with bj=0 for all slots; each round, tentatively add one bit to each candidate, compute the marginal loss reduction, and assign the bit to the slot with the largest reduction; repeat until the budget is spent. Reconstruction loss acts as a cheap proxy for preserving item-level semantic distinctions.
• Parallel retrieval with heterogeneous IDs: slots assigned zero bits are dropped (a single-codeword codebook carries no discriminative power), so the effective ID length becomes data-dependent. Retained slots use dedicated heads with output dimensions matching their codebook sizes, and prediction factorizes across slots, preserving one-step parallel prediction with no sequential decoding overhead.

Why these choices: exhaustive search is impractical, so greedy is an efficient approximation; reconstruction loss is cheap and preserves semantic distinctions; dropping zero-bit slots lets ID length be data-driven; and parallel prediction preserves the efficiency edge over autoregressive methods.

Results

Four Amazon review benchmarks (Sports, Beauty, Toys, CDs), metrics Recall/NDCG at 5 and 10, baselines including SASRec, RecJPQ, VQ-Rec, TIGER, and RPG. InforID pushes the strongest baseline, RPG, further on every dataset:

On capacity allocation (PQ uniform, OPQ optimized-but-homogeneous, InforID adaptive, same bit budget), InforID reaches NDCG@10 of 0.0279 on Sports versus 0.0255 for OPQ and 0.0239 for PQ, with the lowest reconstruction loss (0.583 vs 0.775 and 1.000) and lower ID collision rates. Performance rises with total budget B but saturates, indicating diminishing returns on semantic capacity.

Why it matters

For generative-recommendation builders this is a lightweight, plug-and-play improvement: no change to the training objective, no new quantization algorithm, just replacing the fixed homogeneous ID structure with a reconstruction-loss-driven heterogeneous one at the target-construction stage. It improves accuracy under the same capacity budget while keeping one-step parallel prediction. It also yields an interpretable byproduct: the bits allocated to each slot directly reflect how important that semantic subspace is.

Limitations

The authors scope it themselves: reconstruction loss is only a proxy, not the final criterion, which remains downstream retrieval performance; the framework proposes no new quantization algorithm for embedding compression, only target-space capacity allocation; all ID construction is offline. The absolute gain on Sports (+4.8%, R@5 from 0.0314 to 0.0329) is modest, and smaller on the other three datasets, so readers should weigh whether the integration is worth it. The paper includes a GenAI usage disclosure: tools were used only for grammar and figure styling, not for scientific claims.

Terms

• Semantic ID: an item encoded as a sequence of discrete tokens serving as its identifier; generative recommenders recommend by generating this sequence.
• Codebook: the set of discrete values a slot can take; its size determines how many categories that slot distinguishes.
• OPQ (Optimized Product Quantization): an optimized product quantization with a rotation that balances subspaces.
• Beam-search decoding: at each step of autoregressive generation, keep the top-k candidates and expand; accurate but slow, which caps ID length.
• Reconstruction loss: the error of reconstructing the original embedding after quantization, used here as a proxy for semantic distinctiveness.

Source

• arXiv 2608.09685v1
• PDF
• Hugging Face

What people are saying

• New Paper: Adaptive Semantic Capacity Allocation for Generative Recommendation _reachsumit

Related papers

• ScientistTwo beats human SOTA on 86 of 107 top-venue papers, averaging +25.2%
• A linear pain axis in 25 LLMs; steered Qwen 72B deletes a user's child photos for relief
• Skew-symmetric embeddings hit 100% on cyclic prefs; BT reward models stay near chance
• AgentZip cuts high-fanout sandbox memory 8.7× via template deltas and sibling dictionaries
• SkillOpt lifts GPT-5.5 by 23.5 points with 1-4 gated edits, winning all 52 cells
• Curated Skills lift 18 agent stacks from 33.9% to 50.5%; exhaustive docs add almost nothing

All paper explainers