acceptodds
Under review as a conference paper at ICLR 2027

Item Representations Matter: Bridging Hard Semantic IDs and Soft Prototype Mixtures for Generative Retrieval

Abstract

Modern recommender systems rely on generative retrieval techniques to predict the next item either as discrete codes or as a continuous vector. On the one hand, Semantic IDs (SIDs) give discrete models shared parameters and pretrained semantic structure, but discretizing items into hard codes discards some information from the original content embeddings. On the other hand, continuous models avoid quantization, yet typically use flat item lookup tables with no explicit parameter tying across items. We ask whether a continuous recommender can introduce shared structure without making it a discrete bottleneck. We propose **CAP-Mix**, a flow recommender that represents each item as a soft mixture of shared prototypes plus an item-specific vector. The assignments, prototypes, and item vectors are trained jointly, and the soft representation is retained at inference. Pretrained information is optional: SID responsibilities can serve as a prior over assignments, while a full content embedding can initialize the item-specific vector without changing the initial flat table. In matched comparisons on three datasets, CAP-Mix improves over a flat table on the two sparse Amazon datasets and yields a small gain on Steam when trained without pretrained information. Given only an offline SID, it outperforms a hard table built from the same tokenizer by 4.45%–14.90% in HR@5 ↑. Given the full content embedding, it matches or improves a strong flat table while hard assignments regress on the sparse Amazon datasets. Overall, averaged across 6 benchmarks, CAP-Mix improves HR@5 ↑ and NDCG@5 ↑ by +6.9% and +13.6% relative to the best evaluated baseline for each dataset and metric, while delivering 1.4–161× the inference throughput ↑ of the discrete baselines in our experiments.

Primary area:
generative models
Keywords:
Flow-Matching, Generative Recommendation, Semantic ID

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_3ocmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAP-Mix, a prototype-mixture item representation framework for flow-based generative recommendation. The method addresses the limitations of hard Semantic IDs, which introduce quantization errors, and flat continuous item embeddings, which lack explicit cross-item parameter sharing. CAP-Mix represents each item as a soft mixture of shared prototypes combined with an item-specific vector. The assignments and prototypes are jointly optimized with the recommendation model, supported by code-matching and prototype-utilization objectives. The framework can also incorporate pretrained SID information or continuous content embeddings. Experiments on six public datasets demonstrate improvements in recommendation accuracy over several generative retrieval baselines, along with higher inference throughput compared with the evaluated discrete generative methods.

Critical strengths
  1. The paper addresses an important trade-off between hard semantic quantization and independent continuous item representations. Combining shared prototypes with item-specific vectors provides a reasonable way to introduce cross-item parameter sharing without imposing a discrete representation bottleneck.

  2. The experiments cover six public datasets and three pretrained-information regimes. The comparisons with flat embeddings, hard SIDs, and generative retrieval baselines, together with the diagnostic analyses in the appendix, provide useful empirical evidence for the proposed framework.

Critical weaknesses
  1. The central problem of balancing information preservation and cross-item parameter sharing has been extensively explored in representation learning and generative recommendation. Recent studies, such as UniGRec [1] and Differentiable Semantic ID [2], have already investigated adaptive semantic assignments jointly optimized with recommendation objectives. Although extending these ideas to continuous flow-based recommendation is reasonable, the paper does not clearly identify the new challenges introduced by this setting or explain how its research problem fundamentally differs from existing work. As a result, the novelty of the problem formulation is limited.
  2. The main representation in Eq. (5), which combines a soft mixture of shared prototypes with an item-specific vector, is conceptually similar to existing prototype-based representation learning and embedding decomposition approaches. The soft assignments, attention-based assignment module, and prototype-utilization regularization also rely largely on established techniques. More importantly, the item-specific vector remains fully trainable and unconstrained, meaning that it can theoretically represent the entire item embedding without relying on the shared prototypes. Although auxiliary losses are introduced to encourage prototype learning, the paper provides insufficient evidence that this design learns more effective cross-item sharing than simpler alternatives, such as linear softmax assignments or parameter-matched embedding decompositions. Therefore, the methodological contribution and the necessity of the proposed components are not sufficiently established.
  3. Although CAP-Mix achieves improvements on several benchmarks, its advantages over the flat baseline are relatively small or inconsistent when full content embeddings are available. As shown in Table 1, the HR@5 improvements are only 1.49% on Beauty and 2.09% on Steam, while performance decreases by 0.52% on Industrial. More importantly, the paper lacks systematic ablations of the attention-based assignment module, code-matching loss, and prototype-utilization regularization, making it difficult to determine their individual contributions. Direct comparisons with closely related soft semantic ID methods are also missing. Furthermore, the reported inference speedups are primarily measured against discrete generative architectures with different backbones and decoding procedures, making it difficult to separate the benefit of the proposed item representation from that of continuous retrieval itself. These limitations weaken the experimental support for the paper's main conclusions.

[1] Li J, Zhang Y, Bai Y, et al. UniGRec: Unified Generative Recommendation with Soft Identifiers for End-to-End Optimization[J]. arXiv preprint arXiv:2601.17438, 2026.

[2] Fu J, Ge X, Karatzoglou A, et al. Differentiable Semantic ID for Generative Recommendation[C]//Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2026: 369-379.

Questions for authors
  1. What are the key methodological differences between CAP-Mix and existing approaches such as UniGRec and Differentiable Semantic ID? Could the authors further clarify the necessity of the proposed prototype-mixture design and provide comparisons with simpler alternatives, such as linear softmax assignments or parameter-matched embedding decompositions?
  2. Could the authors provide ablation studies to quantify the individual contributions of the attention-based assignment module, code-matching loss, and prototype-utilization regularization? Additionally, how much of the reported inference efficiency improvement can be attributed to the proposed item representation rather than differences in backbone architectures and decoding procedures?
4 comments

Dear Reviewer 3ocm,

Thanks very much for your appreciation of our paper’s motivation for cross-item sharing, prototype-mixture design, and breadth of the experimental evaluation. Regarding all of your raised questions in terms of motivation, methodological distinctions from prior work, component ablations, and experimental evidence, for better illustration, we made the following organization with added empirical experiments and extended analysis. We are more than willing to provide any further explanation, please let us know.

Q1

[...] Recent studies, such as UniGRec [1] and Differentiable Semantic ID [2], have already investigated adaptive semantic assignments jointly optimized with recommendation objectives. Although extending these ideas to continuous flow-based recommendation is reasonable, the paper does not clearly identify the new challenges introduced by this setting or explain how its research problem fundamentally differs from existing work. [...]

A: Our challenge is to learn useful cross-item sharing while retaining unrestricted, unquantized item representations, while being able to incorporate SID and/or embedding information. Because per-item vectors can bypass shared prototypes, useful sharing is not automatic, and how to incorporate SID information into continuous recommenders while allowing for adaptation during training is not straightforward either. CAP-Mix jointly learns adaptive prototype mixtures and item-specific vectors to pool recommendation gradients; semantic information enters separately through content initialization or optional tokenizer priors. The contribution is this continuous-table learning design, rather than a new solution to tokenizer non-differentiability.

Q2

[...] the item-specific vector remains fully trainable and unconstrained, meaning that it can theoretically represent the entire item embedding without relying on the shared prototypes.

A: The unrestricted item-specific vector is intentional. In continuous space, we preserve per-item flexibility, including the capacity to represent item-specific patterns. However, the ability to represent the same item table does not imply identical learning behavior: shared parameters introduce gradient coupling that independent item rows lack.

Please note the flat baselines of Table 1 and Table 7 also have this unrestricted item embedding. For example, the scratch scenario initializes all item embeddings randomly and CAP-Mix outperforms the flat baseline. Since both have the same representation capacity, this performance improvement comes from the learning dynamics provided by CAP-Mix.

Furthermore, we separately tested removing the per-item embeddings, as well as replacing them with a compressed item table. The full model has higher mean HR@5 in every tested setting:

RegimeVariantT1 Beauty HR@5T1 Industrial HR@5T7 Beauty HR@5T7 Industrial HR@5
ScratchCAP-Mix5.89 (ref.)2.88 (ref.)5.67 (ref.)2.81 (ref.)
ScratchNo item-specific vectors3.50 (−40.58%)2.80 (−2.78%)1.25 (−77.95%)0.99 (−64.77%)
ScratchHASH10 (1/10 residual rows)3.70 (−37.18%)2.59 (−10.07%)1.97 (−65.26%)1.50 (−46.62%)
Offline SIDCAP-Mix6.32 (ref.)3.47 (ref.)5.89 (ref.)3.21 (ref.)
Offline SIDNo item-specific vectors4.94 (−21.84%)2.92 (−15.85%)3.00 (−49.07%)0.56 (−82.55%)
Offline SIDHASH10 (1/10 residual rows)5.26 (−16.77%)2.73 (−21.33%)4.58 (−22.24%)2.01 (−37.38%)

HASH10 maps items randomly to ceil(V/10) trainable residual rows, so approximately ten items share each row. No-residual removes r_i while retaining learned features, assignments and prototypes. These controls show the utility of the residual in the tested recipes.

Q3

Although auxiliary losses are introduced to encourage prototype learning, the paper provides insufficient evidence that this design learns more effective cross-item sharing than simpler alternatives, such as linear softmax assignments or parameter-matched embedding decompositions. Therefore, the methodological contribution and the necessity of the proposed components are not sufficiently established.

Could the authors further clarify the necessity of the proposed prototype-mixture design and provide comparisons with simpler alternatives, such as linear softmax assignments or parameter-matched embedding decompositions?

A: We added both simpler alternatives: independent linear-logit softmax heads and a parameter-matched residual factorization. CAP-Mix has higher mean HR@5 in all four tested Beauty/Industrial scratch/PCA-text settings:

RegimeVariantBeauty HR@5Industrial HR@5
ScratchCAP-Mix (attention)5.89 (ref.)2.88 (ref.)
ScratchLinear-logit softmax5.81 (−1.36%)2.82 (−2.08%)
ScratchResidual factorization5.76 (−2.21%)2.73 (−5.21%)
PCA-textCAP-Mix (attention)6.79 (ref.)3.79 (ref.)
PCA-textLinear-logit softmax6.70 (−1.33%)3.67 (−3.17%)
PCA-textResidual factorization6.51 (−4.12%)3.68 (−2.90%)

Our design predicts prototype assignments from a compact learned item feature (D_f=32 in these experiments, versus retrieval dimension D=128). The attention head enables assignment representations to interact across codebook tokens, providing an architectural route to coordinating potentially complementary prototype contributions rather than predicting each group solely from the item feature. This is an inductive bias.

The parameter-matched factorization is X = R + UB, with R ∈ ℝ<sup>V×D</sup>, U ∈ ℝ<sup>V×q</sup> and B ∈ ℝ<sup>q×D</sup>. R remains unrestricted; U contains unconstrained item-specific weights over q shared vectors in B. We choose q to approximately match CAP-Mix’s reported active-training parameter count.

Q4

Although CAP-Mix achieves improvements on several benchmarks, its advantages over the flat baseline are relatively small or inconsistent when full content embeddings are available. [...]

A: We agree that the gains over PCA-text flat tables are small relative to the other settings. However, the HR improvements are consistent, and increasingly so at larger cutoffs in Table 1 (with prior) and Table 7 (without prior). Among flat, hard and CAP-Mix, CAP-Mix ranks first in 4/6 dataset–prior settings at HR@5, 5/6 at HR@10 and 6/6 at HR@20. Industrial in Table 1 moves from second to first at HR@20, and Steam in Table 7 does so at HR@10.

DatasetVariantT1 HR@5T1 HR@10T1 HR@20T7 HR@5T7 HR@10T7 HR@20
BeautyFlat6.699.6613.476.279.5213.52
BeautyHard6.439.3913.116.279.2312.90
BeautyCAP-Mix6.799.7513.536.509.7013.71
IndustrialFlat3.815.798.333.625.548.03
IndustrialHard3.525.417.893.515.427.93
IndustrialCAP-Mix3.795.778.363.685.678.25
SteamFlat8.1313.6520.558.1013.4520.70
SteamHard8.1913.9520.958.2113.4520.38
SteamCAP-Mix8.3014.2821.508.1413.7720.75

A content-initialized flat table is already a strong baseline, which may leave less scope for additional gains from statistics pooling.

Q5

More importantly, the paper lacks systematic ablations of the attention-based assignment module, code-matching loss, and prototype-utilization regularization [...].

Could the authors provide ablation studies to quantify the individual contributions of the attention-based assignment module, code-matching loss, and prototype-utilization regularization?

A: We added individual and joint loss removals across all three Beauty regimes, and provided the attention-to-linear-logit head swap in Q3.

We removed each loss and partially retuned the surviving auxiliary coefficients. As shown, removing either loss lowers mean HR@5.

VariantScratch HR@5Offline SID HR@5PCA-text HR@5
CAP-Mix5.89 (ref.)6.32 (ref.)6.79 (ref.)
No code matching5.78 (−1.87%)5.76 (−8.86%)6.68 (−1.62%)
No utilization loss5.75 (−2.38%)6.08 (−3.80%)6.77 (−0.29%)
Neither loss5.63 (−4.41%)5.76 (−8.86%)6.60 (−2.80%)

Q6

Direct comparisons with closely related soft semantic ID methods are also missing.

A: We added DIGER and UniGRec comparisons. For UniGRec, we use the authors’ newer official v2 implementation without post-hoc reranking.

MethodBeauty HR@5Beauty NDCG@5Industrial HR@5Industrial NDCG@5Steam HR@5Steam NDCG@5
DIGER4.533.023.202.097.804.98
UniGRec5.633.793.522.327.945.10
CAP-Mix6.794.733.792.348.695.57

CAP-Mix has higher mean HR@5 and NDCG@5 than the listed baselines on all three datasets in these comparisons.

Please see the response to reviewer Suzb's Q14 question for more details.

Q7

the reported inference speedups are primarily measured against discrete generative architectures with different backbones and decoding procedures, making it difficult to separate the benefit of the proposed item representation from that of continuous retrieval itself.

Additionally, how much of the reported inference efficiency improvement can be attributed to the proposed item representation rather than differences in backbone architectures and decoding procedures?

A: Thank you for this question. We do not attribute the speedup over discrete generators to CAP-Mix’s assignment head. The assignment heads are used during training, and the resulting item embeddings are materialized as a table for inference. The heads are not evaluated per query and do not accelerate the flow backbone. Thus, CAP-Mix provides training time item representation learning benefits that do not affect inference performance.

Differences from TIGER/MaskGR reflect complete systems with different backbones and decoding procedures. The prior adds per-query cost and remains an accuracy–cost trade-off; it can be omitted, as in Table 7.

To make the timing comparison more consistent, we retained native generation while replacing the original testing-loop timers with a common prediction-only benchmark using three actual history items, including complete SID tuples:

ModelBeauty HR@5Beauty queries/sIndustrial HR@5Industrial queries/s
CAP-Mix, with prior6.7929,759.903.7928,221.02
CAP-Mix, no prior (Table 7)6.5040,305.383.6840,619.65
FlowRec, no prior6.2841,396.553.6241,207.25
TIGER3.6119,664.222.136,065.25
MaskGR5.08168.373.46222.28

Accordingly, our efficiency claim is that the prototype-assignment head adds no per-query evaluation after materialization, retaining the inference-time benefits of FlowRec over discrete baselines in our experiments. We will modify our statements to make this clear.

Q8

What are the key methodological differences between CAP-Mix and existing approaches such as UniGRec and Differentiable Semantic ID?

A: UniGRec and DIGER jointly adapt semantic tokenization to recommendation while retaining discrete autoregressive generation. One of their main contributions is how to overcome the non-differentiability of discrete SIDs tokenizers. CAP-Mix instead combines adaptive prototype mixtures with unrestricted per-item vectors in a continuous item table, and thus does not have deal with differentiability complexities. The flow predicts a continuous vector and retrieves against the materialized table without semantic-ID decoding. Semantic information enters separately through PCA initialization or optional tokenizer priors, while parameter sharing and statistics pooling are provided by the CAP-Mix's prototype mixtures, without requiring hard quantization.

Thank you for the detailed responses and additional experiments. The rebuttal addresses several of my concerns, but some key issues remain.

Methodological novelty: The differences from existing discrete Semantic ID methods are clearer, but the novelty over soft prototype mixture and embedding decomposition approaches remains insufficiently established. The additional experiments mainly demonstrate the importance of item-specific vectors rather than the contribution of shared prototypes.

Performance improvements: The new comparisons are helpful, but the gains over Linear Softmax and the Flat Baseline remain modest, especially with full content embeddings. Statistical significance analyses are still missing.

Experimental evidence: The additional ablations, comparisons with UniGRec and DIGER, and efficiency experiments substantially improve the experimental completeness. However, they mainly support empirical effectiveness rather than methodological novelty, while some concerns about comparison fairness remain.

Overall, I appreciate the authors' efforts, but my main concern regarding methodological novelty remains. Therefore, I will maintain my original rating.

Reviewer_DeamOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SID-based discrete recommendation models enable cross-item parameter sharing but introduce information loss and code collisions. Embedding-based continuous recommendation models operate directly on embeddings but lack explicit cross-item parameter sharing. To combine the advantages of both approaches, the authors propose CAP-Mix, a novel flow-based recommender that represents each item using an item-specific vector and a soft mixture of prototype embeddings shared across items. CAP-Mix can optionally incorporate pretrained information from hard SIDs or content embeddings for initialization. The experiments demonstrate the effectiveness and efficiency of CAP-Mix relative to the selected baselines.

Critical strengths
  • Flexibility. CAP-Mix accommodates different levels of pretrained information: it can learn from scratch using interaction data alone, incorporate offline SIDs as priors over soft prototype assignments, or use content embeddings to initialize item-specific vectors. This flexibility allows the framework to exploit available semantic information while keeping its item representations adaptable to the recommendation task.
Critical weaknesses
  • Limited improvements. As reported in Table 1, CAP-Mix offers limited improvements over flat item representations and SIDs in the PCA-Text setting. The authors are encouraged to provide additional empirical evidence to better substantiate its effectiveness in this setting.
  • Lack of ablation studies. The experimental evaluation does not include ablation studies of CAP-Mix's components. The authors should conduct ablations to assess the contributions of the proposed components, such as the item-specific vectors, code-matching loss, and balance loss.
Questions for authors

Please refer to weaknesses.

2 comments

Dear Reviewer Deam,

Thanks very much for your appreciation of our paper’s flexibility to learn from scratch, incorporation of offline SID priors, and use of content embeddings for initialization. Regarding all of your raised questions in terms of PCA-text improvements, and lack of ablation studies, for better illustration, we made the following organization with added empirical experiments and extended analysis. We are more than willing to provide any further explanation, please let us know.

Q1

Limited improvements. As reported in Table 1, CAP-Mix offers limited improvements over flat item representations and SIDs in the PCA-Text setting. The authors are encouraged to provide additional empirical evidence to better substantiate its effectiveness in this setting.

A: In Table 1 (prior enabled) and Table 7 (prior disabled), CAP-Mix ranks first in 4/6 dataset–prior settings at HR@5, 5/6 at HR@10 and 6/6 at HR@20. Industrial in Table 1 moves from second to first at HR@20, and Steam in Table 7 moves from second to first at HR@10. We aggree the improvement in the PCA-text setting is smaller compared to other settings. However, note this improvement is consistent across runs, more as we increase the cutoff depth.

DatasetVariantT1 HR@5T1 HR@10T1 HR@20T7 HR@5T7 HR@10T7 HR@20
BeautyFlat6.699.6613.476.279.5213.52
BeautyHard6.439.3913.116.279.2312.90
BeautyCAP-Mix6.799.7513.536.509.7013.71
IndustrialFlat3.815.798.333.625.548.03
IndustrialHard3.525.417.893.515.427.93
IndustrialCAP-Mix3.795.778.363.685.678.25
SteamFlat8.1313.6520.558.1013.4520.70
SteamHard8.1913.9520.958.2113.4520.38
SteamCAP-Mix8.3014.2821.508.1413.7720.75

Q2

Lack of ablation studies. The experimental evaluation does not include ablation studies of CAP-Mix's components. The authors should conduct ablations to assess the contributions of the proposed components, such as the item-specific vectors, code-matching loss, and balance loss.

A: We added the requested ablations and controls. Reviewer Suzb’s Q8 reports individual and joint removal of code matching and balance regularization; Q6–Q7 examines how SID information enters the baselines and CAP-Mix; Q9 compares adaptive and frozen assignments; Q11 reports the parameter-matched residual factorization; and Q10 provides per-item assignment-entropy diagnostics.

For the item-specific residuals, we report removal and a compressed hashed-residual control in both the Table 1 and Table 7 settings:

RegimeVariantT1 Beauty HR@5T1 Industrial HR@5T7 Beauty HR@5T7 Industrial HR@5
ScratchCAP-Mix5.89 (ref.)2.88 (ref.)5.67 (ref.)2.81 (ref.)
ScratchNo item-specific vectors3.50 (−40.58%)2.80 (−2.78%)1.25 (−77.95%)0.99 (−64.77%)
ScratchHASH10 (1/10 residual rows)3.70 (−37.18%)2.59 (−10.07%)1.97 (−65.26%)1.50 (−46.62%)
Offline SIDCAP-Mix6.32 (ref.)3.47 (ref.)5.89 (ref.)3.21 (ref.)
Offline SIDNo item-specific vectors4.94 (−21.84%)2.92 (−15.85%)3.00 (−49.07%)0.56 (−82.55%)
Offline SIDHASH10 (1/10 residual rows)5.26 (−16.77%)2.73 (−21.33%)4.58 (−22.24%)2.01 (−37.38%)

HASH10 uses a fixed random map to a trainable residual bank with approximately one-tenth as many rows, so roughly ten items share each residual. Both removing the item-specific vectors and randomly hashing them lower mean HR@5 in the reported comparisons. We will incorporate these ablations and controls into the revised paper.

I thank the authors for their detailed responses, which have resolved my concerns about the lack of ablation studies. However, the performance gains of CAP-Mix remain modest. I therefore retain my original scores.

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

This paper proposes CAP-Mix, a continuous flow-based recommender whose item table combines learned soft mixtures of shared prototypes with unrestricted item-specific vectors. Assignments, prototypes, and item vectors are jointly trained using the recommendation objective, code-matching supervision, and a prototype-utilization penalty. Optional pretrained information enters through tokenizer posterior responsibilities or content-based initialization, while a co-visitation prior contributes to retrieval scores. Experimental comparisons across six datasets show modest improvements over independent item embeddings on most evaluated settings and larger gains over fixed hard semantic-ID assignments.

Critical strengths

The paper examines whether continuous item embeddings can retain item-specific flexibility while benefiting from shared structure commonly associated with semantic IDs. This is a meaningful design question beyond changing the generative backbone.

The three information regimes help distinguish pre-trained content from learned sharing, and zero-initializing prototypes makes the PCA-text CAP-Mix and flat models begin from identical item embeddings. The random hard-assignment control and the no co-visitation prior experiment provide additional diagnostic evidence.

Critical weaknesses

The relationship between the controlled flat baseline and the main FlowRec baseline is unclear. This affects the interpretation of the gains. Table 1 reports PCA-text flat HR@5 scores of 6.69 on Beauty and 3.81 on Industrial, compared with CAP-Mix scores of 6.79 and 3.79. These correspond to changes of +1.49% and −0.52%. However, Table 2 reports FlowRec scores of 6.27 and 3.62, producing substantially larger system-level gains. Those FlowRec scores coincide with the no co-visitation prior flat results in Table 7. The scratch Beauty comparison exhibits the same pattern: Table 1 reports 5.77 for flat, whereas Table 7 and the full FlowRec results report 5.61. These differences need reconciliation with Appendix B’s statement that FlowRec uses the same encoder, velocity field, and decoder as CAP-Mix. They could reflect different prior settings or configuration selection, but the paper does not make the distinction sufficiently clear. If the main FlowRec comparison also changes the retrieval prior, it remains a potentially useful complete-system comparison, but cannot establish the effect of replacing the item table alone. A matched flat baseline should appear alongside the main results, with the prior settings and tuning procedures specified explicitly.

The offline-SID comparison changes the available information as well as whether assignments are hard or adaptive. The hard baseline receives stored discrete codes, whereas CAP-Mix receives full posterior responsibilities over codewords. These responsibilities preserve relative preferences among alternative prototypes that a single code does not retain. Appendix C further explains that the responsibilities follow a greedy residual path rather than the balanced assignment path used to construct the stored codes; their modes agree with the stored codes for only 92%–94% of Beauty items at each level. Thus, using the same tokenizer does not make this an information-matched comparison. The results support the combined use of posterior information and adaptive soft representations, but do not isolate the benefit of softness or co-adaptation.

The evidence does not establish which parts of the proposed representation are necessary and its advantage over simpler shared parameterizations. The paper does not report direct ablations of code matching, utilization regularization, or assignment adaptation. These are central to the claim that useful shared structure emerges jointly with recommendation learning. In particular, high entropy of the catalog-average assignment does not establish informative item grouping: identical uniform assignments for every item also satisfy the utilization objective. A parameter-matched residual factorization, e.g., an independent item table plus a learned low-rank shared component, would also be an important control setting. CAP-Mix introduces additional item features and shared parameters during training; folding them away at inference does not resolve that training-time difference. The experimental baselines omit the closest jointly optimized SID approaches, such as UniGRec and DIGER, despite discussing them in the motivation. Keeping mixtures soft during continuous retrieval is a meaningful distinction, but the current comparisons do not establish its incremental benefit over these related approaches or a simpler shared representation.

Questions for authors

Please see critical weaknesses. Additional question are listed below:

  1. What explains the differences between the flat results in Table 1 and FlowRec in Tables 2 and 10–11? Does FlowRec use the co-visitation prior during both training and inference? Please specify which non-table settings and tuning budgets are shared, and whether a fully matched flat baseline preserves the main conclusions.

  2. How much of the offline-SID gain remains with frozen soft posterior assignments, or when adaptive soft assignments receive only smoothed versions of the stored hard codes? These controls would distinguish additional posterior information from the benefit of learning the sharing structure.

  3. Does CAP-Mix outperform a training-parameter-matched residual low-rank factorization? What happens when code matching, utilization regularization, or assignment updates are individually removed? Paired seed differences for the controlled Table 1 comparisons would be particularly helpful because several gains are small.

  4. Table 5 measures inference using three history items for every model, while the main quality comparisons retain the discrete baselines’ 120-token input limit. Can the authors report accuracy and throughput at the same history settings, and clarify whether co-visitation prior construction, lookup, and fusion are included in timing and memory measurements?

4 comments

Dear Reviewer Suzb,

Thanks very much for your appreciation of our paper’s meaningful design question, three information regimes, and diagnostic controls. Regarding all of your raised questions in terms of baseline comparability, pretrained information and representation components, and evaluation fairness, for better illustration, we made the following organization with added empirical experiments and extended theoretical analysis. We are more than willing to provide any further explanation, please let us know.

Q1

What explains the differences between the flat results in Table 1 and FlowRec in Tables 2 and 10–11?

A: Thank you for highlighting this ambiguity. The difference is the co-visitation prior. FlowRec’s Table 2 scores correspond to the no-prior flat results in Table 7: PCA-text on Beauty and Industrial, and scratch on Steam. Table 2 therefore compares complete systems, not the item representation alone. We will correct Appendix B’s “same decoder” wording.

Q2

Does FlowRec use the co-visitation prior during both training and inference?

A: No. The original FlowRec method does not use our co-visitation prior during training or inference; accordingly, it does not use it in Table 2.

Q3

Please specify which non-table settings and tuning budgets are shared, [...]

A: The shared FlowRec settings (backbone, FlowRec’s loss weights, history length, embedding width, flow-step count, learning rate and batch size) are fixed across all tables, unless stated otherwise. For Tables 1 and 7, since the prior changes the scoring and matching objectives, we separately tune item-table-specific hyperparameters with and without it.

Q4

[...] and whether a fully matched flat baseline preserves the main conclusions.

A: Yes. Table 7 reports gains over the flat baseline without the prior, while Table 1 reports the prior-enabled comparisons. Both tables compare item representations with the shared backbone and prior settings held fixed within each comparison.

Q5

The relationship between the controlled flat baseline and the main FlowRec baseline is unclear. This affects the interpretation of the gains. [...] A matched flat baseline should appear alongside the main results, with the prior settings and tuning procedures specified explicitly.

A: As explained in Q1–Q4, Tables 1 and 7 already compare item representations with a common flow backbone and the prior enabled and disabled, respectively. We will incorporate Table 7, the no-prior flat comparison, into the main paper alongside Table 1 and correct Appendix B’s decoder description.

Q6

The offline-SID comparison changes the available information as well as whether assignments are hard or adaptive. The hard baseline receives stored discrete codes, whereas CAP-Mix receives full posterior responsibilities over codewords. [...] Thus, using the same tokenizer does not make this an information-matched comparison. [...]

A: Thank you for highlighting the information difference. CAP-Mix retains higher mean HR@5 than the hard-baseline variant given access to full posterior responsibilities.

We ran the “hard” baseline in Table 1’s “offline SID” scenario with fixed full posterior responsibilities instead of stored discrete codes on Beauty, Industrial and Steam. This substitution gives small mean HR@5 changes: +1.03%, +0.33% and −1.44%, respectively. CAP-Mix retains a larger advantage over this posterior-informed hard baseline.

VariantBeauty HR@5Industrial HR@5Steam HR@5
Hard5.80 (ref.)3.02 (ref.)8.32 (ref.)
Hard + posterior5.86 (+1.03%)3.03 (+0.33%)8.20 (−1.44%)
CAP-Mix6.32 (+8.97%)3.47 (+14.90%)8.69 (+4.45%)

Q7

or when adaptive soft assignments receive only smoothed versions of the stored hard codes? These controls would distinguish additional posterior information from the benefit of learning the sharing structure.

A: We ran CAP-Mix in the same “offline SID” scenario with this smoothed one-hot prior. Even without re-tuning, this retains mean HR@5 gains over hard of +3.62%, +14.24% and +0.84%, respectively. Full CAP-Mix has higher mean scores on all three datasets.

VariantBeauty HR@5Industrial HR@5Steam HR@5
Hard5.80 (ref.)3.02 (ref.)8.33 (ref.)
CAP-Mix with one-hot prior6.01 (+3.62%)3.45 (+14.24%)8.40 (+0.84%)
CAP-Mix6.32 (+8.97%)3.47 (+14.90%)8.69 (+4.32%)

Replacing posterior responsibilities with one-hot codes changes the prior’s sharpness, so the smoothing and prior-trust settings may need retuning (e.g., more uniform smoothing or less prior sharpening). As explained in Appendix D, the log-prior needs full support since a zero prior probability rules out that code regardless of the learned evidence. To answer your question, we have not tuned these values, and instead used a fixed, untuned smoothing coefficient ε_H = 0.01:

ρ<sub>i,k</sub> = (1 − ε<sub>H</sub>) 1

k=c<sub>i</sub>k = c<sub>i</sub>
  • ε<sub>H</sub>/K, with ε<sub>H</sub> = 0.01.

Q8

The evidence does not establish which parts of the proposed representation are necessary and its advantage over simpler shared parameterizations. The paper does not report direct ablations of code matching, utilization regularization, [...]

What happens when code matching, utilization regularization, [...] are individually removed?

A: We conducted loss ablations on Beauty across the scratch, offline-SID and PCA-text regimes, removing code matching, utilization regularization, and both losses. Each removal lowers mean HR@5 in the reported comparisons.

VariantScratch HR@5Offline SID HR@5PCA-text HR@5
CAP-Mix5.89 (ref.)6.32 (ref.)6.79 (ref.)
No code matching5.78 (−1.87%)5.76 (−8.86%)6.68 (−1.62%)
No utilization loss5.75 (−2.38%)6.08 (−3.80%)6.77 (−0.29%)
Neither loss5.63 (−4.41%)5.76 (−8.86%)6.60 (−2.80%)

Q9

The paper does not report direct ablations of [...] assignment adaptation. These are central to the claim that useful shared structure emerges jointly with recommendation learning.

What happens when [...] assignment updates are individually removed?

A: Adaptive CAP-Mix has higher mean HR@5 than the frozen-prior alternative on Beauty and Industrial in both the offline SID and PCA-text regimes. Frozen assignments are fixed to the tokenizer posterior prior while prototypes, item-specific residuals and the flow backbone remain trainable.

RegimeVariantBeauty HR@5Industrial HR@5
Offline SIDAdaptive6.32 (ref.)3.47 (ref.)
Offline SIDFrozen6.17 (−2.37%)3.40 (−2.02%)
PCA-textAdaptive6.79 (ref.)3.79 (ref.)
PCA-textFrozen6.46 (−4.86%)3.39 (−10.55%)

Q10

In particular, high entropy of the catalog-average assignment does not establish informative item grouping: identical uniform assignments for every item also satisfy the utilization objective.

A: Thank you for pointing this out. We agree that catalog-average entropy alone does not establish informative grouping. The per-item weights are not uniform: for example, in every examined full-model codebook on Beauty, at least 95% of items have entropy below 0.03% of the uniform benchmark.

We report the 95th percentile of normalized per-item entropy, 100 H(A<sub>i</sub>)/ln K. Uniform weights over K prototypes give 100%, while one-hot weights give 0%:

VariantScratchOffline SIDPCA-text
CAP-Mix0.02%0.00%0.00%
No code matching27.04%19.65%31.22%
No utilization loss13.69%6.48%0.00%
Neither loss19.77%19.65%30.69%

As it can be observed, the settings without no code matching have more uniform per-item assignments.

Q11

A parameter-matched residual factorization, e.g., an independent item table plus a learned low-rank shared component, would also be an important control setting.

Does CAP-Mix outperform a training-parameter-matched residual low-rank factorization?

A: Yes, CAP-Mix has higher mean HR@5 in all tested settings.

RegimeVariantBeauty HR@5Industrial HR@5
ScratchCAP-Mix5.89 (ref.)2.88 (ref.)
ScratchResidual factorization5.76 (−2.21%)2.73 (−5.21%)
PCA-textCAP-Mix6.79 (ref.)3.79 (ref.)
PCA-textResidual factorization6.51 (−4.12%)3.68 (−2.90%)

We tested a parameter-matched residual low-rank factorization X = R + UB, retaining the unrestricted item table R and the same flow backbone. Here U ∈ ℝ<sup>V×q</sup>, B ∈ ℝ<sup>q×D</sup> and R ∈ ℝ<sup>V×D</sup>, where V is the number of item-table rows. Equivalently, B is a codebook of q shared vectors, and each row of U contains learned item-specific mixing weights. The shared width q was selected to match CAP-Mix’s reported active-training parameter count.

We evaluate this factorization control in scratch and PCA-text, since in those settings q can be freely selected as it does not depend on a SID tokenizer.

Q12

CAP-Mix introduces additional item features and shared parameters during training; folding them away at inference does not resolve that training-time difference.

A: We agree. Folding the item table is an inference optimization; it does not eliminate CAP-Mix’s additional training parameters. We will make this clear in the revised paper.

Q13

Paired seed differences for the controlled Table 1 comparisons would be particularly helpful because several gains are small.

A: Thank you for this suggestion. We report paired CAP-Mix-minus-baseline differences from Table 1. We report the mean, as well as the min and max values in brackets.

RegimeBaselineBeauty ΔHR@5Industrial ΔHR@5Steam ΔHR@5
ScratchFlat+0.12 [+0.04, +0.18]+0.12 [+0.07, +0.17]+0.06 [−0.23, +0.36]
Offline SIDHard+0.51 [+0.38, +0.70]+0.44 [+0.43, +0.47]+0.36 [+0.00, +0.70]
PCA-textFlat+0.10 [−0.04, +0.30]−0.01 [−0.05, +0.03]+0.08 [−0.03, +0.21]
PCA-textHard+0.35 [+0.27, +0.50]+0.27 [+0.19, +0.36]+0.11 [−0.20, +0.44]

Note Table 1’s PCA-text flat model is a strong baseline: item embeddings already start from rich content information, and the decoder also receives an interaction-derived co-visitation prior. This plausibly leaves less room for additional gains from cross-item gradient/statistical pooling.

Q14

The experimental baselines omit the closest jointly optimized SID approaches, such as UniGRec and DIGER, despite discussing them in the motivation. Keeping mixtures soft during continuous retrieval is a meaningful distinction, but the current comparisons do not establish its incremental benefit over these related approaches or a simpler shared representation.

A: Thank you for suggesting these comparisons. We added DIGER and UniGRec and will include them in the revised paper. For UniGRec, we use the authors’ newer official v2 implementation without post-hoc reranking. CAP-Mix has higher mean HR@5 and NDCG@5 than these two baselines on all three tested datasets.

MethodBeauty HR@5Beauty NDCG@5Industrial HR@5Industrial NDCG@5Steam HR@5Steam NDCG@5
DIGER4.533.023.202.097.804.98
UniGRec5.633.793.522.327.945.10
CAP-Mix6.794.733.792.348.695.57

They use the same setting as the other discrete baselines, i.e., 120 SID-token histories and 20 beam width. They use the same text embedding as used by the other baselines. The CAP-Mix row is from Table 2.

We tuned the following:

  • SID: K in {256, 512}, and L in {3,4} for both methods.
  • DIGER presets: for Industrial and Steam, we selected among the authors’ Beauty/Instruments/Yelp presets, varying tokenizer learning rate, assignment temperature, frequency threshold and training budget.
  • DIGER on Steam: we additionally searched recommendation learning rate over {0.001, 0.0003, 0.0001}, since the released presets did not perform well. 0.0003 won.
  • UniGRec: other training/backbone settings followed the authors’ v2 recipe.

Q15

Table 5 measures inference using three history items for every model, while the main quality comparisons retain the discrete baselines’ 120-token input limit. Can the authors report accuracy and throughput at the same history settings, and clarify whether co-visitation prior construction, lookup, and fusion are included in timing and memory measurements?

A: Yes. We report accuracy and prediction-stage throughput using three actual history items, including complete SID tuples. Following the reviewer’s suggestion, we repeated the benchmark while retaining the selected models and native generation, replacing the timing of the original testing loops with a common prediction-only protocol. We report Beauty and Industrial results here and will update all remaining results in the paper.

Co-visitation prior lookup and fusion are included in the timings; global graph construction is excluded. The prior remains an accuracy–cost trade-off, and CAP-Mix can be used with it (Table 1) or without it (Table 7).

Note CAP-Mix does not accelerate the flow backbone, but provides representation-learning benefits during training without affecting the inference performance. This is achieved thanks to materializing the item table before inference, which only needs to be done once. Thus retaining the inference throughput benefit of continuous recommenders over discrete recommenders in our experiments. We will make this clear in the paper.

ModelBeauty HR@5Beauty queries/sIndustrial HR@5Industrial queries/s
CAP-Mix, with prior6.7929,759.903.7928,221.02
CAP-Mix, no prior (Table 7)6.5040,305.383.6840,619.65
FlowRec, no prior6.2841,396.553.6241,207.25
TIGER3.6119,664.222.136,065.25
MaskGR5.08168.373.46222.28

HR@5 ×100, three-seed means; queries/s are medians over five process repetitions at a representative checkpoint. No-prior CAP-Mix accuracy is from Table 7; other scores are the matched-history re-evaluation.

The shared protocol uses H200, batch 128, top-20, FP32 storage with TF32 matmul, discarded warm-up and explicit CUDA synchronization. Per-query prior formation, sparse lookup/diffusion, mixing, logarithm and fusion are included.

Global graph construction/cache loading and item-table materialization are outside the query timer. Input preparation, transfers and discrete SID-to-item conversion are also excluded, so these are model-stage throughputs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.