acceptodds
Under review as a conference paper at ICLR 2027

Generative Recommendations via Latent Diffusion Models

Abstract

Generative recommendation models commonly represent items using sequences of semantic tokens. However, these representations are typically learned independently of the recommendation task, potentially discarding information relevant to next-item predictions. We introduce a latent diffusion framework that jointly learns item representations from both content and interaction history and a masked denoising model for next-item recommendation. Our variational formulation supports discrete, continuous, and hybrid latent variables while encouraging discriminative representations and reducing collisions. Using a straight-through estimator, we allow the denoiser to update the item encoder, while our catalog-retrieval objective reduces collisions. We additionally propose beam-free retrieval based on denoising scores. In our experiments on three Amazon Reviews datasets, our method outperforms comparable two-stage methods. Further, our ablations demonstrate that joint training improves recommendation performance and reduces collisions.

Primary area:
generative models
Keywords:
Generative models, discrete diffusion, recommendations

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 4 reviews, above 27% of the others.
Reviewer_38GTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies task-adapted semantic IDs for sequential recommendation. Starting from fixed sentence-T5 metadata embeddings, it jointly trains an item tokenizer and a masked diffusion recommender, using a straight-through estimator to propagate recommendation gradients to the tokenizer. Its ELBO-motivated training objective combines denoising and catalog-retrieval losses with auxiliary reconstruction and vector-quantization losses.

The proposed beam-free inference first scores valid catalog items from a fully masked target, then rescores 100 candidates using all nonempty target-mask patterns. A validation-fitted ranker combines these neural scores with training-derived transition and popularity features. The final discrete pipeline also includes collision repair followed by denoiser finetuning with the tokenizer frozen. Continuous and hybrid latents are evaluated as representation ablations rather than components of the selected system. Experiments on three Amazon Reviews 2023 categories report higher NDCG than the compared baselines, and an ablation on Musical Instruments supports the benefits of joint tokenizer training.

Critical strengths
  1. The paper addresses a meaningful mismatch between content-based tokenization and the downstream recommendation objective. Allowing recommendation gradients to update the tokenizer is well motivated, and Table 3 provides direct supporting evidence: on Musical Instruments, joint training improves NDCG@10 from 0.0290 to 0.0300 and reduces the fraction of catalog items with colliding IDs from 1.12% to 0.45%.

  2. The catalog-supported proposal and denoising-based rescoring provide a useful alternative to sequential beam expansion. The proposal ensures that candidates correspond to catalog items, while the denoising score evaluates compatibility under multiple masking patterns. Table 1 reports the highest NDCG among the compared methods on all three datasets.

Critical weaknesses
  1. The evidence isolating the contribution of joint representation learning remains incomplete. I appreciate that Table 3 includes a joint-versus-frozen comparison and Table 6 decomposes the ranking components. However, the controlled joint-training comparison is reported only on Musical Instruments, with a modest NDCG@10 gain and no uncertainty estimates. The final ranker also has a substantial effect: Table 6 increases NDCG@10 from 0.0327 with denoising scores alone to 0.0366 with the complete ranker. Since the external baselines in Table 1 are taken from Shi et al. (2025), the available evidence does not fully isolate tokenizer adaptation from differences in ranking and training stages. Moreover, closely related approaches to task-adapted tokenization, such as ETEGRec and DIGER, are discussed in Section 5 but absent from the empirical comparison. These gaps limit the evidence for the specific contribution beyond the complete system; the three Amazon categories also provide limited evidence across domains and catalog scales.

  2. The assumptions behind Eq. (13) need clarification. The Bayes decoder in Eq. (10) gives p(y∣z)=p(y)qϕ(z∣y)/rϕ(z)p(y|z)=p(y)q_\phi(z|y)/r_\phi(z), where rϕ(z)=∑y′p(y′)qϕ(z∣y′)r_\phi(z)=\sum_{y'}p(y')q_\phi(z|y'). This denominator generally remains inside the marginalization over zz. The paper already uses a uniform item prior and hard IDs, and discusses uniqueness in Section 3.1. If the IDs are unique, rϕ(z)r_\phi(z) is constant on valid catalog IDs and the ranking simplification is justified. However, residual collisions or soft assignments can make it depend on zz and change the ranking if omitted. Please clarify whether the final repaired catalog has unique IDs and how the proposal and denoising scores in Eqs. (14)-(15) relate to this derivation.

Questions for authors
  1. Under what assumptions is the normalization term rϕ(z)r_\phi(z) removable in Eq. (13), and does the final system satisfy those assumptions after collision repair? Please clarify how Eqs. (14)-(15) approximate the resulting item-ranking objective, including the treatment of any remaining collisions.

  2. Can the joint-versus-frozen comparison be extended across the three datasets with matched training budgets, collision handling, decoding, and ranking, together with variability across runs? The most informative additional controls would be applying the same ranking procedure and validation budget to the frozen/LLaDA-Rec counterpart, and comparing against a directly relevant task-adapted tokenizer baseline such as ETEGRec or DIGER. Please also confirm that the imported baseline numbers use compatible splits, candidate catalogs, and evaluation filtering.

  3. At which stage are the Table 3 results measured? Please report how recommendation quality and collision rate change after joint training, after repair, and after the 50-epoch frozen continuation, along with their training costs. Does the frozen baseline receive a comparable continuation budget? The final frozen stage does not contradict joint learning, but its contribution should be distinguished from that of tokenizer adaptation.

  4. What are the measured latency, throughput, and peak memory of beam-free retrieval versus native beam decoding under matched hardware, batching, and recommendation quality?

3 comments

We thank the reviewer for their feedback. In the following, we address their points.

The evidence isolating the contribution of joint representation learning remains incomplete. I appreciate that Table 3 includes a joint-versus-frozen comparison and Table 6 decomposes the ranking components. However, the controlled joint-training comparison is reported only on Musical Instruments, with a modest NDCG@10 gain and no uncertainty estimates. The final ranker also has a substantial effect: Table 6 increases NDCG@10 from 0.0327 with denoising scores alone to 0.0366 with the complete ranker. Since the external baselines in Table 1 are taken from Shi et al. (2025), the available evidence does not fully isolate tokenizer adaptation from differences in ranking and training stages. Moreover, closely related approaches to task-adapted tokenization, such as ETEGRec and DIGER, are discussed in Section 5 but absent from the empirical comparison. These gaps limit the evidence for the specific contribution beyond the complete system; the three Amazon categories also provide limited evidence across domains and catalog scales.

We agree that the frozen-tokenizer ablation alone does not establish an advantage over existing task-adapted tokenizers. Therefore, we evaluated DIGER on all three datasets using the same splits, training catalog, and T5 features.

In the following table, we report NDCG@10 of the test split:

DatasetDIGEROurs, complete
Industrial & Scientific0.00640.0280
Musical Instruments0.02470.0368
Video Games0.03410.0596

As we observe, our method achieves higher NDCG@10 in these comparisons.

We acknowledge that joint tokenization and recommendation have been explored previously. However, our contribution focuses on combining semantic-ID learning with masked diffusion within a variational framework and our beam-free retrieval.

The assumptions behind Eq. (13) need clarification. The Bayes decoder in Eq. (10) gives , where . This denominator generally remains inside the marginalization over . The paper already uses a uniform item prior and hard IDs, and discusses uniqueness in Section 3.1. If the IDs are unique, is constant on valid catalog IDs and the ranking simplification is justified. However, residual collisions or soft assignments can make it depend on and change the ranking if omitted. Please clarify whether the final repaired catalog has unique IDs and how the proposal and denoising scores in Eqs. (14)-(15) relate to this derivation. Q1: Under what assumptions is the normalization term removable in Eq. (13), and does the final system satisfy those assumptions after collision repair? Please clarify how Eqs. (14)-(15) approximate the resulting item-ranking objective, including the treatment of any remaining collisions.

In general, the denominator rϕ(z)r_\phi(z) depends on zz and cannot be omitted. With a uniform item prior and deterministic IDs, the correct ranking is:

p(y∣x)∝pθ(z(y)∣x)m(z(y))p(y\mid x)\propto \frac{p_\theta(z(y)\mid x)}{m(z(y))},

where m(z)m(z) counts items sharing the ID zz. However, the approximation is correct when there are no collisions, which is the case for us as our final collision rate is close to 0.

We will clarify this assumption in the methodology section.

Can the joint-versus-frozen comparison be extended across the three datasets with matched training budgets, collision handling, decoding, and ranking, together with variability across runs? The most informative additional controls would be applying the same ranking procedure and validation budget to the frozen/LLaDA-Rec counterpart, and comparing against a directly relevant task-adapted tokenizer baseline such as ETEGRec or DIGER. Please also confirm that the imported baseline numbers use compatible splits, candidate catalogs, and evaluation filtering.

We follow the reviewer's suggestion and extend the joint-versus-frozen comparison to all three datasets, using both neural-only beam decoding and complete inference. Both training procedures include collision repair.

The following table reports NDCG@10:

DatasetFrozen, neural-onlyJoint, neural-onlyFrozen, completeJoint, complete
Industrial & Scientific0.01950.02160.02750.0280
Musical Instruments0.02900.03000.03700.0368
Video Games0.04660.04710.06000.0596

As we observe, joint training improves neural-only performance across all three datasets. However, when including the ranker, the performance becomes numerically close.

We have confirmed that the splits are equal. While the Llada-Rec implementation used all items for their catalog, we use only the training catalog and count targets outside that catalog as misses.

At which stage are the Table 3 results measured? Please report how recommendation quality and collision rate change after joint training, after repair, and after the 50-epoch frozen continuation, along with their training costs. Does the frozen baseline receive a comparable continuation budget? The final frozen stage does not contradict joint learning, but its contribution should be distinguished from that of tokenizer adaptation.

In Table 3, we report results from the final checkpoints, after the 50-epoch frozen continuation. Both models receive the same training compute, i.e., 150 epochs of denoiser training followed by 50 additional epochs with the tokenizer frozen.

The only difference is that for the joint model, the tokenizer and denoiser are jointly optimized during the 150-epoch stage, while for the frozen baseline, the tokenizer is pretrained using the reconstruction and quantization objectives. The tokenizer remains frozen throughout both denoiser-training stages.

Both rows use neural-only beam decoding without the reranker.

What are the measured latency, throughput, and peak memory of beam-free retrieval versus native beam decoding under matched hardware, batching, and recommendation quality?

We measured both methods on an NVIDIA A100 (40 GB). The times include generation, denoising-based scoring, and reranking.

DatasetMethodLatency (ms, batch 1)Throughput (queries/s, batch 64)Peak GPU memory (GiB, batch 64)
Industrial & ScientificNative-beam140.47.6310.74
Industrial & ScientificBeam-free108.09.781.90
Musical InstrumentsNative-beam153.26.8710.73
Musical InstrumentsBeam-free117.88.811.89
Video GamesNative-beam222.24.7510.75
Video GamesBeam-free172.65.991.91

We again thank the reviewer for their feedback.

Thank you for the additional experiments. Jointly training the tokenizer improves NDCG@10 on all three datasets with neural-only beam decoding. In the full pipeline, the differences relative to a frozen tokenizer are small and mixed +0.0005, −0.0002, and −0.0004. Since joint training is a central contribution and its benefit in the full system remains unclear, I am keeping my rating at 2.

Reviewer_MZh4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers generative recommendations (GR) with diffusion models. The authors start with raising a concern that GR learned representations usually discard the information relevant to next-item predictions. Inspired by the success in diffusion models, they introduce the combination of GR and diffusion models to improve the current arch as well as algorithms for training a good GR model. Experiments on some Amazon review datasets show the improvements over baselines and ablation studies are given.

Critical strengths

The presentation is clear and generally understandable -- the authors introduce each component of their proposed model arch, and they report the performance gains over the existing methods clearly.

Critical weaknesses
  1. There is no clear evidence that this method works well at scale yet -- current experiments are only on some public datasets.

  2. The presentation/motivation might be a bit concerning -- the authors first raise one hypothesis/concern in the beginning via "However, these representations are typically learned independently of the recommendation task, potentially discarding information relevant to next-item predictions.", and directly start to introduce each building blocks including model arch, losses, and math behind the diffusion models. There seems to be a gap here, as one may wonder if there is evidence to support the hypothesis on the downsides of existing methods, and what exactly is the information in " information relevant to next-item predictions".

  3. Ablation studies to some extent address the concern in 2, but they are mainly focusing on how each part could benefit the final performance gains instead of a deeper understanding of the source of gains. Given that all techniques already exit in literature, it might be difficult to show the real novelty here in this paper.

Questions for authors

See Weaknesses above. Also are the improvements in all tables statistically significant?

1 comment

We thank the reviewer for their valuable feedback. In the following, we address their points.

There is no clear evidence that this method works well at scale yet -- current experiments are only on some public datasets.

Our evaluation covers three Amazon categories with approximately 25 thousand catalog items and more than 300 thousand training interactions per dataset (Table 7). We use these, as they are standard benchmarks, allowing us to compare with other methods. Unfortunately, we are not able to evaluate on private datasets. We stated this limitation explicitly in the updated manuscript.

The presentation/motivation might be a bit concerning -- the authors first raise one hypothesis/concern in the beginning via "However, these representations are typically learned independently of the recommendation task, potentially discarding information relevant to next-item predictions.", and directly start to introduce each building blocks including model arch, losses, and math behind the diffusion models. There seems to be a gap here, as one may wonder if there is evidence to support the hypothesis on the downsides of existing methods, and what exactly is the information in " information relevant to next-item predictions".

We thank the reviewer for this comment and would like to clarify certain points. By “information relevant to next-item prediction”, we mean information that helps predict the next interaction given the history.

A tokenizer trained to reconstruct metadata embeddings is optimized to preserve information for the corresponding reconstruction. Thus, it does not explicitly prioritize preserving information that improves next-item prediction. As the discrete representations are limited, different objectives may favor different information. Our motivation is therefore a potential mismatch between the tokenization and recommendation objectives. In Table 3, we observe that joint learning improves recommendation performance relative to the frozen-tokenizer procedure.

We will clarify this distinction and make the motivation more precise.

Ablation studies to some extent address the concern in 2, but they are mainly focusing on how each part could benefit the final performance gains instead of a deeper understanding of the source of gains. Given that all techniques already exit in literature, it might be difficult to show the real novelty here in this paper.

Our main contribution is joint semantic-ID learning, combined with masked diffusion, unified within a variational framework. Further, we propose beam-free retrieval.

In Tables 3–4, we show the effects of the different components on recommendation performance and collisions. Further, in Table 6, we examine the contribution of reranking.

Are the results statistically significant?

We have added standard deviations over three training seeds to the main results table. For convenience, we report them below.

DatasetRecall@1Recall@5Recall@10NDCG@5NDCG@10
Industrial & Scientific0.0123 ±\pm 0.00010.0337 ±\pm 0.00010.0492 ±\pm 0.00030.0232 ±\pm 0.00010.0281 ±\pm 0.0001
Musical Instruments0.0154 ±\pm 0.00030.0437 ±\pm 0.00040.0651 ±\pm 0.00040.0298 ±\pm 0.00020.0367 ±\pm 0.0001
Video Games0.0254 ±\pm 0.00020.0715 ±\pm 0.00060.1043 ±\pm 0.00060.0488 ±\pm 0.00040.0594 ±\pm 0.0004

We thank the reviewer again for their feedback.

Reviewer_3VzbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a latent diffusion framework that jointly optimizes the item tokenizer and a conditional denoising model, with an additional retrieval objective to reduce semantic ID collisions. Experiments on three Amazon Reviews datasets show improved recommendation performance over the compared two-stage methods, while ablations support the benefits of joint training for both recommendation accuracy and collision reduction.

Critical strengths
  1. Well-motivated joint learning: Jointly optimizing the SID tokenizer and the denoising recommender allows item representations to adapt to the recommendation task. The item-discrimination objective and candidate denoising scores also provide practical ways to reduce collisions and improve ranking.
  2. Extensive ablation studies: The paper examines latent representations, joint training, auxiliary losses, decoding choices, and ranking components. These experiments help clarify which design choices contribute to the reported performance.
Critical weaknesses
  1. Limited evidence for the benefit of joint training: Table 3 shows modest gains from joint training on Musical Instruments, with relative improvements of about 3.5% in Recall@10 and NDCG@10, but reports no variability across runs. The final pipeline also includes collision repair, additional finetuning, and reranking, making it difficult to attribute the main results to joint training. Controlled comparisons across all three datasets, with repeated runs, are needed to establish the reliability and contribution of this central component.
  2. Limited evaluation scope. The ablation studies are restricted to Musical Instruments, making it unclear whether the findings generalize to other datasets. Evaluating additional datasets, such as the Amazon dataset used in Image-based Recommendations on Styles and Substitutes, and reporting the corresponding main results would strengthen the empirical evidence.
Questions for authors
  1. The paper motivates joint training as a solution to the mismatch between item tokenization and recommendation, but does not evaluate against methods such as DIGER and ETEGRec that address the same problem. The frozen-tokenizer ablation shows that joint training helps, but does not establish the advantage of the proposed approach over existing alternatives.
1 comment

We thank the reviewer for their thoughtful feedback. In the following, we address their concerns.

Limited evidence for the benefit of joint training: Table 3 shows modest gains from joint training on Musical Instruments, with relative improvements of about 3.5% in Recall@10 and NDCG@10, but reports no variability across runs. The final pipeline also includes collision repair, additional finetuning, and reranking, making it difficult to attribute the main results to joint training. Controlled comparisons across all three datasets, with repeated runs, are needed to establish the reliability and contribution of this central component.

To provide additional insights, we provide the validation Recall@10 and NDCG@10 after each stage in the following table across the three datasets.

Industrial & Scientific:

Recall@10NDCG@10Collision rate
After joint training0.04830.02614.04%
After repair0.04810.02610.42%
After frozen continuation0.04850.02670.42%
After reranking0.06060.03490.42%

Musical Instruments:

Recall@10NDCG@10Collision rate
After joint training0.06060.03244.45%
After repair0.06120.03270.45%
After frozen continuation0.06210.03270.45%
After reranking0.07260.04040.45%

Video Games:

Recall@10NDCG@10Collision rate
After joint training0.09940.05317.23%
After repair0.10010.05351.45%
After frozen continuation0.10100.05401.45%
After reranking0.12020.06801.45%

As we observe, each stage has a positive benefit on the results and either improves recommendations or reduces collisions.

Finally, we include standard deviations over three training seeds to the main results table.

DatasetRecall@1Recall@5Recall@10NDCG@5NDCG@10
Industrial & Scientific0.0123 ±\pm 0.00010.0337 ±\pm 0.00010.0492 ±\pm 0.00030.0232 ±\pm 0.00010.0281 ±\pm 0.0001
Musical Instruments0.0154 ±\pm 0.00030.0437 ±\pm 0.00040.0651 ±\pm 0.00040.0298 ±\pm 0.00020.0367 ±\pm 0.0001
Video Games0.0254 ±\pm 0.00020.0715 ±\pm 0.00060.1043 ±\pm 0.00060.0488 ±\pm 0.00040.0594 ±\pm 0.0004

Limited evaluation scope. The ablation studies are restricted to Musical Instruments, making it unclear whether the findings generalize to other datasets. Evaluating additional datasets, such as the Amazon dataset used in Image-based Recommendations on Styles and Substitutes, and reporting the corresponding main results would strengthen the empirical evidence.

We have extended the joint vs. frozen comparison to all three datasets. However, our evaluation remains limited to these three categories, as this allows us to directly compare to Llada-Rec [1], i.e., our closest related work.

DatasetFrozen, neural-onlyJoint, neural-onlyFrozen, completeJoint, complete
Industrial & Scientific0.01950.02160.02750.0280
Musical Instruments0.02900.03000.03700.0368
Video Games0.04660.04710.06000.0596

As we observe, when using the neural prediction, we observe a consistent improvement across all datasets. After reranking, the results are numerically close.

The paper motivates joint training as a solution to the mismatch between item tokenization and recommendation, but does not evaluate against methods such as DIGER and ETEGRec that address the same problem. The frozen-tokenizer ablation shows that joint training helps, but does not establish the advantage of the proposed approach over existing alternatives.

We agree with the reviewer that joint training of tokenization and recommendation has been explored before. We mention ETEGRec and DIGER in our related work section. ETEGRec couples tokenizer and autoregressive recommender learning through alignment objectives and alternating optimization. DIGER enables recommendation gradients to influence semantic-ID learning through differentiable assignments.

Our contribution, however, is the joint representation learning within a latent diffusion framework. The tokenizer receives gradients from a masked denoising objective that learns to recover target tokens under different masking patterns, conditioned on interaction history. Our objectives are derived from a variational formulation and help the model maintain discriminative representations, which further include continuous, discrete, and hybrid states. Further, we provide a beam-free retrieval mechanism.

We will clarify the positioning of our contribution in the introduction.

We again thank the reviewer for their feedback.

[1] Shi, Teng, et al. "Llada-rec: Discrete diffusion for parallel semantic id generation in generative recommendation." Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing. 2026.

Reviewer_R6j3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper develops a diffusion-based sequence modeling framework for recommendation based on joint training of the tokenizer and the generative recommendation backbone. It also introduces a classification component to address semantic ID collisions.

Critical strengths

The core idea of this paper is diffusion for recommendation with semantic representation learning. It connects methods from these two areas and achieves better performance than approaches based on either area alone.

The figures are clear and visually appealing, and the mathematical derivations are sound.

Critical weaknesses

1.The final method incorporates not only diffusion scores but also interaction transition statistics from the training set, item popularity, and a linear reranking model fitted on the validation set. These components can themselves improve recommendation performance. Therefore, the reported results do not convincingly establish that the improvements are attributable to the paper’s methodological contributions.

2.ETEGRec (2025) has already investigated end-to-end item tokenization learning, while DIGER (2026) further explores mechanisms through which recommendation gradients directly influence discrete codes. This paper differs by applying this idea to masked diffusion and combining it with a retrieval discrimination loss. This is still a valuable combination, but its novelty requires stronger experimental support.

3.The paper introduces numerous loss functions and a complex training procedure, creating a substantial burden for optimization and validation. However, it does not provide tables or curves showing the behavior of the five losses, their gradients, or their interactions during training.

4.Furthermore, in the four-stage optimization procedure, the tokenizer’s learning rate during backbone training is only 9 × 10⁻⁸, and the EMA coefficient is 0.9999999. Can this reasonably be considered joint training? Evidence showing how the codebook changes during this stage is needed to support this claim. Without such evidence, the codebook could effectively be regarded as a pretrained, frozen codebook during backbone training.

Questions for authors

The four-stage optimization procedure may substantially influence the final results. However, the paper does not adequately explain why these stages are necessary, how these unusually conservative hyperparameters were determined, or whether they are supported by prior literature. In particular, how was the duration of 50 epochs for the final stage selected?

5 comments

We thank the reviewer for their feedback. In the following, we address their remarks.

The final method incorporates not only diffusion scores but also interaction transition statistics from the training set, item popularity, and a linear reranking model fitted on the validation set. These components can themselves improve recommendation performance. Therefore, the reported results do not convincingly establish that the improvements are attributable to the paper’s methodological contributions.

In Table 3,4, and 6, we ablate different components of our model, including the auxiliary ranking features. Further, in Table 6, we separate the ranking signals: NDCG@10 is 0.0298 for training transitions alone, 0.0327 for denoising scores alone, and 0.0366 for all components. Thus, the denoising scores are helpful and receive a substantial benefit from reranking. We clarified this in the updated manuscript.

ETEGRec (2025) has already investigated end-to-end item tokenization learning, while DIGER (2026) further explores mechanisms through which recommendation gradients directly influence discrete codes. This paper differs by applying this idea to masked diffusion and combining it with a retrieval discrimination loss. This is still a valuable combination, but its novelty requires stronger experimental support.

We acknowledge that joint learning of tokenization and recommendation has been explored before. We see our work as a complementary line of work, focusing on discrete diffusion. Our contribution is a variational formulation that natively includes semantic-ID learning, masked diffusion, and catalog retrieval. Further, we propose a beam-free retrieval. In Tables 3–4, we provide supporting evidence for our framework and the derived objectives.

The paper introduces numerous loss functions and a complex training procedure, creating a substantial burden for optimization and validation. However, it does not provide tables or curves showing the behavior of the five losses, their gradients, or their interactions during training.

The five objectives serve different purposes and motivations. The variational formulation results in the diffusion and classification losses. The history loss is equal to the diffusion loss, just applied to the whole history, and is used to provide additional training signal to the denoiser model and avoid overfitting. The reconstruction and quantization commitment loss are standard for VQ-VAEs and are used to stabilize training.

In Table 4, we ablate the classification and reconstruction terms in a direct comparison on Musical Instruments and show that removing both substantially decreases NDCG and increases the catalog collision rate. Removing either individually produces a smaller degradation.

Furthermore, in the four-stage optimization procedure, the tokenizer’s learning rate during backbone training is only 9 × 10⁻⁸, and the EMA coefficient is 0.9999999. Can this reasonably be considered joint training? Evidence showing how the codebook changes during this stage is needed to support this claim. Without such evidence, the codebook could effectively be regarded as a pretrained, frozen codebook during backbone training.

Although the learning rate is low and the EMA coefficient is close to 1, the tokens change substantially during joint training. In the following table, we show how many semantic IDs change from before to the end of joint training.

DatasetItems whose semantic ID changed
Industrial & Scientific54.92%
Musical Instruments57.39%
Video Games62.76%

As we observe, more than half of the codes change during joint training.

The four-stage optimization procedure may substantially influence the final results. However, the paper does not adequately explain why these stages are necessary, how these unusually conservative hyperparameters were determined, or whether they are supported by prior literature. In particular, how was the duration of 50 epochs for the final stage selected?

Each stage serves a different purpose. The stabilization results in initial representations, while joint training adapts the semantic IDs to the recommendation task. Finally, the repair reduces catalog collisions, and frozen continuation adapts the denoiser to the repaired, fixed IDs. While some tokenizer updates are conservative, they still lead to substantial semantic ID changes in the joint training stage, as shown in the following table.

DatasetItems whose semantic ID changed
Industrial & Scientific54.92%
Musical Instruments57.39%
Video Games62.76%

We selected 50 epochs across all datasets and did not ablate this parameter.

We again thank the reviewer for their valuable feedback and are happy to address any remaining concerns.

W1: I noticed that reviewers 3Vzb (W1) and 38GT (W1/Q2) also raised similar concerns about the attribution of performance improvements. The newly added comparison between joint training and a frozen tokenizer shows that joint training consistently improves the performance of the neural model alone. However, my concern about the attribution of the final performance gains remains.

W2: The response addresses my concern well.

W3/Q1: I would like the authors to provide training and validation loss curves for all five losses. During the rebuttal stage, the authors could present the loss values at different training stages in a table. I hope the data can be as detailed as possible so that I can directly plot the curves.

W4/Q1: I would like the authors to provide a table showing how the number of items whose Semantic IDs have changed evolves over the course of training. I hope the data can be as detailed as possible so that I can directly plot the curve.

I believe the curves requested in W3 and W4 are particularly important, as they could provide substantial information for observing and analyzing the behavior of the proposed method.

Thank you for your response.

W1: I noticed that reviewers 3Vzb (W1) and 38GT (W1/Q2) also raised similar concerns about the attribution of performance improvements. The newly added comparison between joint training and a frozen tokenizer shows that joint training consistently improves the performance of the neural model alone. However, my concern about the attribution of the final performance gains remains.

We agree with the reviewer that the final ranking diminishes the gains of the joint model over the two-stage model. However, we still want to highlight that, in a controlled comparison, the joint model consistently outperforms the two-stage model.

W3/Q1: I would like the authors to provide training and validation loss curves for all five losses. During the rebuttal stage, the authors could present the loss values at different training stages in a table. I hope the data can be as detailed as possible so that I can directly plot the curves. W4/Q1: I would like the authors to provide a table showing how the number of items whose Semantic IDs have changed evolves over the course of training. I hope the data can be as detailed as possible so that I can directly plot the curve.

We have added these plots to the updated manuscript in Appendix B.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.