acceptodds
Under review as a conference paper at ICLR 2027

RELAX: Semantic IDs Need Not Identify a Single Item

Abstract

Generative recommendation with Semantic IDs commonly uses generation targets that uniquely identify the target item. We identify an exact item generation bottleneck: even when the correct base ID remains in the beam, exact completion can still lose the target. Conditioned on base ID success, exact completion rates are 32.25–45.08 percentage points lower for multi-item base IDs than for single-item base IDs across four benchmarks. We introduce Relax, which relaxes exact item generation through Shared IDs with an explicit capacity bound. Relax constructs a capacity-constrained Semantic ID hierarchy and assigns each item the shortest prefix shared by at most items as its generation target. At inference, each generated Shared ID expands to its item bucket, and a continuous scorer ranks the union of these buckets. The generation target therefore specifies a bounded item set rather than requiring unique item identification. Relative to the strongest reproduced baseline for each metric and benchmark, Relax improves Recall@10 and NDCG@10 by 11.4% and 19.7% on average, respectively. In controlled comparisons using the same frozen continuous scorer, Relax outperforms the variants that retain exact item generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
sequential recommendation, generative recommendation, item semantic ID, discrete token generation, retrieval-augmented recommendation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WncZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether generative recommendation necessarily requires Semantic IDs (SIDs) to uniquely identify individual items. The authors identify an exact item generation bottleneck, where the correct base SID survives beam search but the target item is subsequently eliminated during the final disambiguation step. To address this issue, they propose RELAX, which constructs a capacity-constrained SID hierarchy and assigns each item the shortest prefix representing at most CC items. During inference, generated Shared IDs are expanded into candidate item buckets, which are subsequently ranked by a frozen continuous recommender. Experiments on four recommendation benchmarks demonstrate improvements over existing sequential, generative, and hybrid recommendation methods. Additional controlled comparisons and ablation studies investigate the effectiveness of the proposed framework.

Critical strengths
  • The paper raises a meaningful question regarding whether generative recommendation must resolve exact item identity during autoregressive decoding. The distinction between base ID retention and exact item completion provides a useful perspective for analyzing candidate elimination. The authors further validate this phenomenon across multiple datasets, beam widths, and an additional TIGER configuration. These analyses offer empirical insights into the limitations of conventional SID-based generation.
  • RELAX integrates capacity-constrained hierarchical clustering, variable-length Shared IDs, and continuous item-level ranking into a unified framework. The explicit capacity bound provides a useful guarantee on candidate set size, while the prefix-free property enables unambiguous constrained decoding. The experimental evaluation covers a relatively broad collection of baselines and includes controlled comparisons, recovery analyses, and sensitivity studies. The consistent performance improvements across four benchmarks are encouraging.
Critical weaknesses
  • The bottleneck analysis does not convincingly establish a limitation specific to exact item generation. The main evidence is that base IDs shared by multiple items have considerably lower conditional exact completion rates than those associated with a single item. However, this gap is partly expected given the nature of the decoding process. When a base ID corresponds to only one item, the final disambiguation step generally has a single valid continuation. In contrast, a shared base ID requires the decoder to distinguish among multiple candidates, making it more likely that the target item will be dropped under a limited beam budget. This can occur even without any particular weakness in the model itself. While Appendix B provides additional analyses of earlier decoding transitions and different beam widths, these results do not clearly separate the claimed bottleneck from the increased branching complexity. It is also unclear whether the observed errors stem from the requirement for exact item identification, inadequate learning of disambiguation tokens, or the decoding procedure. Thus, although the empirical results reveal a clear performance gap, they do not fully support the claim that exact item generation introduces a distinct bottleneck that needs to be relaxed. As a result, the paper's central motivation remains somewhat underdeveloped.

  • The methodological novelty is modest, and the benefits of the proposed components are not convincingly demonstrated. RELAX shares much of its overall design with existing generate-and-rank methods, especially Gryphon and Gryphon-v2, which have already explored shared SIDs followed by item-level ranking. Its main differences lie in the capacity-constrained hierarchy construction and the selection of the shortest qualifying prefix. These are sensible design choices, but it is unclear how much they contribute beyond existing approaches. In Table 2, RELAX outperforms the strongest controlled variant (Active + PrefixStop + Scorer) by only 0.0023–0.0030 in absolute Recall@10 across the four datasets. Moreover, the comparison changes both the history representation and the generation targets, so the individual effects of these modifications cannot be determined. Although Appendix A.1 mentions that results are averaged over three random seeds, the paper reports neither standard deviations nor significance tests, making it difficult to assess the reliability of these relatively small gains. The capacity-constrained hierarchy also deserves closer examination. Under the default configuration, the constraints at earlier hierarchy levels seem fairly permissive relative to the catalog sizes considered. The paper does not show how often these constraints actually alter the original clustering assignments, or whether the resulting hierarchy performs better than simpler alternatives. Without such evidence, it is difficult to determine whether RELAX's distinctive design choices offer meaningful and reliable improvements over existing shared-SID generation and ranking methods.

Questions for authors
  • Q1. On the exact item generation bottleneck (W1). To what extent can the observed gap in conditional completion rates be attributed to the number of valid disambiguation-token continuations? Could the authors analyze completion rates across different base-ID collision-group sizes, ideally controlling for factors such as target item frequency or token probabilities? This would help determine whether the performance gap reflects an actual modeling limitation or simply the increased difficulty of decoding among multiple candidates.

  • Q2. On the effectiveness of RELAX's additional components (W2). Could the authors provide standard deviations and statistical significance tests for the improvements over Active + PrefixStop + Scorer? It would also be helpful to evaluate the use of Shared IDs in interaction histories and generation targets separately, rather than changing both at once. Such analyses would make it easier to assess whether these design choices provide independent benefits and whether the reported improvements are statistically reliable.

  • Q3. On the need for capacity-constrained hierarchy construction (W2). How often does the Active assignment change the original nearest-centroid assignments at each level of the hierarchy? Could the authors also compare RELAX with a standard unconstrained residual K-means hierarchy combined with capacity-aware prefix selection, or with a simpler approach to resolving collisions? These comparisons would help clarify whether the proposed hierarchy construction is necessary or whether similar improvements could be achieved with a less complicated design.

Reviewer_RhFDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines whether generative recommendation requires Semantic IDs to uniquely identify individual items. It identifies an exact item generation bottleneck: a target can be discarded during final disambiguation even when its correct base ID remains in the beam. To address this, the authors propose RELAX, which constructs a residual K-means hierarchy with explicit capacity constraints and assigns each item the shortest prefix shared by at most C items. These Shared IDs are used in both interaction histories and generation targets. A frozen continuous user representation provides additional conditioning for generation. At inference, generated IDs are expanded into candidate buckets, and a frozen continuous scorer ranks the union of their associated items. Experiments on four sequential recommendation benchmarks under full-catalog evaluation report average relative improvements of 11.4% in Recall@10 and 19.7% in NDCG@10 over the strongest reproduced baselines. Validation includes controlled comparisons using the same frozen scorer, analyses of recovery from exact completion failures, and ablations examining capacity constraints, prefix selection, beam width, and continuous user guidance.

Critical strengths

RELAX defines generation targets as the shortest prefixes whose associated item sets satisfy a prescribed capacity bound. The hierarchy construction guarantees a valid target for every indexed item, while the resulting prefix-free identifiers permit unambiguous termination and bound the candidate set by bC (b is the beam size and C is a hyperparam). This provides a concrete mechanism for allocating candidate selection to discrete generation and item discrimination to continuous ranking. The connection between identifier design and a predictable candidate budget is a meaningful methodological contribution.

Critical weaknesses

A common beam width does not equalize the number of item candidates: at (b=20), RELAX ranks approximately 38-54 candidates, whereas Unique + Scorer ranks approximately 19 (Table 2). The comparison with Active + PrefixStop and the results in Appendix C provide useful partial controls, but the paper lacks a systematic evaluation across matched candidate budgets and comparable inference latency, particularly against the strongest baselines. Candidate counts alone cannot establish computational fairness either, since shorter Shared IDs may reduce decoding cost. It therefore remains unclear how much of the reported advantage persists when competing methods are allowed to adjust their decoding budgets to comparable operating points. Comparisons of accuracy against candidate count and end-to-end latency would be consequential to my assessment: without them, the results establish improvements at selected configurations more clearly than a consistently better allocation of retrieval and ranking computation.

Questions for authors
  1. Clarification and additional evidence on retrieval budgets. The comparison with Active + PrefixStop and the results in Appendix C provide useful partial controls. Does RELAX retain its advantage over Unique + Scorer and the strongest reproduced baselines when their decoding budgets are adjusted to achieve comparable candidate counts or, separately, comparable inference latency? Could the authors also clarify the latency measurement protocol in Table 10, including batch size and whether continuous user encoding, beam search, candidate expansion, and final scoring are all included?

  2. Targeted validation of dependence on the final scorer. The comparisons using the same frozen scorer provide valuable internal controls. However, the underlying continuous recommender also supplies the representations used for indexing and user guidance. Are results available comparing the candidate generators under a common, independently trained final scorer, while holding each generator’s candidate sets fixed? Such evidence would help determine whether the relative gains depend on the particular alignment between the indexing and scoring representations. This is a request for additional robustness evidence, rather than a claim that scorer controls are absent.

Reviewer_k4XnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes RELAX, a generative recommendation framework in which a Semantic ID can represent a bounded set of items. It constructs a capacity-constrained hierarchy and assigns each item the shortest prefix shared by at most CC items. At inference, generated Shared IDs are expanded into item candidates, which are ranked by a frozen continuous scorer. The paper motivates this design through an analysis of target loss during exact ID completion and reports improvements on four recommendation benchmarks.

Critical strengths
  1. The paper investigates a concrete failure mode by distinguishing base-ID retention from exact item completion, with diagnostic results across multiple datasets and beam widths.

  2. The Shared-ID construction provides an explicit per-identifier capacity guarantee and an overall candidate-count bound of bCbC.

  3. The experiments include controlled variants using the same frozen scorer and analyses of capacity, beam width, and continuous user guidance.

Critical weaknesses

1. Concerns about the contribution of the item-level ranking stage.

I have concerns about the contribution of the additional item-level scorer. In my understanding, this is essentially a retrieval stage followed by coarse ranking, a standard pipeline in industrial cascaded recommendation systems. Based on practical experience with such systems, an improvement over retrieval alone is expected when a ranking stage is added. Gains obtained this way seem somewhat like an easy win and do not, by themselves, demonstrate that the generator has better retrieval capability.

Table 2 does include Unique + Scorer, which is an appropriate control. However, the improvements attributable to the additional ranking stage should be clearly distinguished from those attributable to the proposed Shared-ID generation scheme. Better final Recall/NDCG should not automatically be interpreted as better generative retrieval.

2. The item-level candidate budget is not properly controlled.

A crucial missing experiment is a comparison that explicitly controls the number of retrieved candidate items. Although the beam width is held constant, this does not provide an equal item-level retrieval budget. Unique-ID generation returns at most bb items, whereas RELAX expands all items in the bb generated buckets and can therefore return up to bCbC items. In Table 2, Unique + Scorer produces approximately 19 candidates on average, while RELAX produces approximately 38–54.

I do not consider this a fair comparison at the item level: the final gains could arise from giving the scorer access to more candidate items, rather than from a better generation scheme. Appendix C provides approximately matched average candidate counts on Toys and Sports, but these are not explicitly controlled equal-count comparisons and do not fully resolve this concern. I would like to see RELAX compared with Unique + Scorer under the same item-level candidate budgets, using the same frozen scorer, to establish whether the proposed design provides gains beyond candidate expansion.

Questions for authors
  1. Can you compare RELAX with Unique + Scorer while fixing the number of candidate items per request and using the same frozen scorer?

  2. How much of the reported improvement comes from Shared ID generation, rather than candidate expansion or the additional ranking stage?

  3. Can you increase the beam widths of the baselines that use beam search in the main table to match the number of candidate items retrieved by RELAX?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.