acceptodds
Under review as a conference paper at ICLR 2027

When Diffusion Recommenders Stop Sampling: Preserving Generative Diversity

Abstract

A user history can admit multiple plausible next interactions, yet conventional sequential recommenders produce a single deterministic ranking. While diffu- sion models offer a natural way to sample alternative continuations, we show that existing target-item diffusion recommenders often suffer from severe sampling collapse, producing nearly identical outputs across independent draws. We iden- tify two ubiquitous design choices that promote this collapse: joint learning of the latent space and denoiser, and in-trajectory supervision via discrete losses. To address these failure modes, we introduce DARTS (Diverse and Accurate Recommendation Through Sampling), a three-stage latent diffusion framework that explicitly decouples representation learning, generative modeling, and dis- crete ranking by learning and freezing the latent space before diffusion training and postponing ranking supervision to the final decoding stage. We further char- acterize sampling diversity using complementary measures of variation before and after decoding, and use posterior-predictive aggregation to exploit this diversity at inference. Across three Amazon Reviews 2023 categories, DARTS preserves substantially greater sampling diversity while outperforming strong sequential, generative-retrieval, and diffusion baselines by 7–12% on standard ranking met- rics. Aggregating multiple samples further improves ranking accuracy at a fixed slate size.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
sequential recommendation, generative models, sampling diversity

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_eVi8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines whether target-item diffusion recommenders retain stochastic variation across repeated samples for the same user history. DARTS separates sequence-autoencoder pretraining, latent diffusion in a frozen representation space, and decoder fine-tuning. The paper evaluates diversity before and after decoding and reports improvements in Recall and NDCG on three Amazon Reviews 2023 categories. Averaging predictions from multiple samples further improves ranking accuracy without increasing the recommendation list size.

Critical strengths
  1. The focus on repeated-sample behavior is useful. Ranking accuracy alone cannot tell whether a diffusion recommender behaves as a stochastic predictor. The diagnostics in Section 4.3 examine variation in both latent space and the resulting recommendation lists.
  2. DARTS improves ranking accuracy across all three datasets and outperforms baselines using the same hybrid item representations. The evaluation also distinguishes gains from larger recommendation lists from those obtained at a fixed list size.
Critical weaknesses
  1. The contribution of the generated latent to ranking accuracy is not directly tested. As shown in Appendix A.3, the decoder processes the right-shifted interaction history and attends to the encoded history states cc. However, I could not find a history-only variant that retains these inputs but removes the generated target latent. The existing ablations retain the generation branch, making it difficult to separate its contribution from that of the pretrained encoder–decoder and subsequent ranking fine-tuning.
  2. The evidence that different samples provide relevant alternative recommendations remains limited. The diagnostics establish variation, and the aggregation results show some predictive value, but neither tells us whether different samples recover distinct relevant future items. The authors acknowledge that leave-one-out evaluation provides only one observed future item per history. It would be useful to evaluate whether different samples recover different held-out future interactions for the same history.
Questions for authors
  1. How does DARTS compare with a history-only decoder that is fine-tuned under the same ranking objective?
  2. Do different samples recover different relevant future items for the same user history?
  3. On the Scientific dataset, Table 1 reports Recall@5 and Recall@10 of 0.0329 and 0.0524, while Table 10 reports 0.0344 and 0.0545 at M=1M=1. Since union recall reduces to ordinary recall in this setting, were different checkpoints, sampling runs, or evaluation settings used?
1 comment

W1&Q1:

To isolate the contribution of diffusion, we conducted an additional ablation replacing Stage-2 diffusion training with a deterministic latent predictor, while keeping the Stage-1 representations, decoder, and Stage-3 objective unchanged. Importantly, the deterministic baseline has exactly the same number of parameters as DARTS, ensuring a fair comparison. Its deterministic variant performs substantially worse than DART (43–44% lower across all four ranking metrics). This demonstrates that the gains cannot be attributed solely to MLM pretraining and BPR fine-tuning, and highlights the contribution of diffusion to our framework. This ablation is present in the revised version of the paper in Appendix J

ModelR @ 10R @ 5Ndcg @ 10Ndcg @ 5
Deterministic latent predictor0.03650.02280.01890.0145
Darts0.05240.03290.02700.0207

W2 & Q2:

We thank the reviewer for raising this important point. Directly evaluating multimodality is challenging in sequential recommendation, as datasets typically provide only one observed next item rather than multiple valid recommendations for a given context.

To investigate recommendation diversity, as a proxy we conducted an additional experiment using sequences of 15 items, holding out 5 items, considering that they are all possible valid recommendations. We have evaluated how our model recovers these items across multiple independent draws. Coverage@10 increases from 4.49% with one draw to 6.41% with seven draws, indicating that different draws can recover different relevant items.

Number of Draws1234567
Coverage@100.04490.04860.04980.05130.05500.05770.0641

Note that these coverage values should not be directly compared with our standard Recall@10 results. This experiment uses shorter training sequences and a different evaluation objective: the model is still trained for next-item prediction, but evaluated on its ability to recover the five held-out items across multiple draws. Training directly to recover these five items would change the learning objective and evaluate a different model from DART.

This ablation is present in the revised version of the paper in appendix M.

While we acknowledge that this experiment does not establish the existence of multiple distinct modes in the learned conditional distribution, it provides strong empirical evidence that DART generates diverse, relevant recommendations across independent draws, supporting its ability to go beyond deterministic next-item prediction. This motivates a complementary question: does the diversity across draws also translate into diversity in the final recommendations, and does it extend beyond differences in item identity to differences in item semantics?

To investigate this, we performed four independent inference draws for 10,000 test histories and analyzed the top-1 recommendations after Stage 3. At the item level, the top-1 recommendation differs across draws for 61.8% of histories. To assess semantic diversity, we used Amazon's level-2 product categories, corresponding to the second level of its product-category hierarchy and providing a more fine-grained grouping of items. We found that 35.22% of histories receive recommendations spanning multiple level-2 categories, with an average of 1.434 distinct categories across the four draws.

These results show that latent diversity translates into diversity in the final recommendations, both at the item and semantic-category levels.

We have added this ablation to the paper in Appendix L.

Q3:

We thank the reviewer for identifying this inconsistency. The discrepancy resulted from inadvertently mixing results from experiments conducted under different settings. We have corrected this error in the revised manuscript to ensure consistency between Tables 1 and 10.

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

The paper shows that target-item diffusion recommenders suffer from "sampling collapse": independent draws give nearly identical rankings. Using 2x2 ablations, it attributes this to two design choices: learning item representations jointly with the denoiser, and applying discrete catalogue supervision inside the diffusion trajectory. The proposed DARTS trains in three stages: a frozen MLM-pretrained sequence autoencoder, then a latent-regression denoiser, then a BPR-fine-tuned decoder. The paper also adds three sampling diagnostics (J, R, PR). On three Amazon 2023 categories, DARTS beats sequential, semantic-ID and diffusion baselines on Recall/NDCG.

Critical strengths
  1. The diagnosis is convincing. The three diagnostics cover different aspects of sampling: how large the variation is, how it spreads across directions, and how much the decoded sets overlap. The 2x2 ablations change only the two suspected design choices.
  2. The negative controls are well chosen. Gumbel-perturbed SASRec and the union of slates rule out added randomness or longer lists as the source of the gains.
Critical weaknesses
  1. My main concern is that it is unclear what the diffusion component actually adds. The closest discriminative baseline, SASRec+T5, has neither the MLM pretraining nor the Stage-3 BPR fine-tuning, and no deterministic control uses the same three-stage recipe. Since diffusion only helps once draws are aggregated, the 6.8-12.0% margins at M=1 could come entirely from pretraining and fine-tuning. If so, the extra cost is hard to justify, especially as the components are standard and the staging follows LDM/ELF. Training needs three separately tuned stages, and serving is 14.5x slower than a single-pass discriminative model.
  2. The statistical evidence is thin. The three Amazon categories are small and very sparse, and absolute scores are low (NDCG@10 of 2.7-5.3%). All results are single runs under one leave-one-out protocol, so a gap like 0.0207 vs. 0.0190 NDCG@5 could be seed noise.
Questions for authors
  1. Clarification: since diffusion only helps at M>1, how much of the M=1 gain in Table 1 comes from the generative path, and how much from MLM pretraining and BPR fine-tuning?
  2. Stage 3 runs the full 200-step reverse chain on every batch. How does the total training time of the three stages compare with the discriminative baselines?
1 comment

W1&Q1

To isolate diffusion's contribution, we replaced Stage-2 diffusion training with a deterministic latent predictor, keeping Stage-1 representations, the decoder, and the Stage-3 objective unchanged. Despite having the same number of parameters, the deterministic variant performs 43–44% worse across all four ranking metrics. This highlights diffusion's contribution beyond MLM pretraining and BPR fine-tuning. We report this ablation in Appendix J.

ModelR @ 10R @ 5Ndcg @ 10Ndcg @ 5
Deterministic latent predictor0.03650.02280.01890.0145
Darts0.05240.03290.02700.0207

Regarding computational cost, we acknowledge that DART is more expensive at inference than single-pass discriminative models. However, a central objective of our work is to investigate why existing diffusion-based recommendation methods often fail to improve recommendation diversity and to address this limitation. From this perspective, comparing DART with other diffusion-based methods is particularly relevant. DART achieves approximately 3.5× faster inference than competing diffusion baselines (Appendix G), while providing improved recommendation performance and diversity. We believe this represents a favorable efficiency–performance trade-off within diffusion-based recommendation.

W2

Our results are averaged across multiple random seeds. We also conducted additional multi-seed experiments for DART and the baselines, obtaining consistent results. The standard deviation is approximately 10−510^{-5} across metrics, suggesting low sensitivity to random initialization. Details are provided in Appendix A.6.

Second, we chose the three Amazon categories because they are established benchmarks in generative recommendation and have been adopted as reference benchmarks in recent work, including Never Lose the Thread, ChainRec, Decision4Rec, and PCTX. Thus, although these datasets are sparse, they allow comparison with a substantial body of prior work.

To further assess robustness across dataset characteristics, we evaluated MovieLens-1M (ML-1M), which has a different sparsity profile. As semantic information is less informative here, we omitted T5 embeddings. DART consistently outperforms the baselines, improving by 4.8–15.2% over the strongest baseline across the four metrics. Results are reported in Appendix K.

ModelHit @ 5Hit @ 10Ndcg @ 5Ndcg @ 10
Sasrec11.1%19.02%6.45%9.06%
Dream rec16.05%24.63%10.61%13.36%
DiffuRec16.02%24.28%10.41%13.05%
ours18.5%25.8%12.2%14.7%
(+/-)+ 15.2%+4.8%15.0%10.0%

Together, the multi-seed results and the additional ML-1M experiments provide further evidence that the performance gains are robust and are not specific to a single random seed or dataset.

[R1] Never Lose the Thread (2026) Never Lose the Thread: Collaborative Latent Reasoning for Generative Recommendation. OpenReview submission. OpenReview

[R2] ChainRec (2026) Li, F., Li, Q., Gao, X., Pan, B., Wu, Y., Zhang, J., Yu, H., Jiang, J., Xiao, J., & Shi, H. (2026). ChainRec: An Agentic Recommender Learning to Route Tool Chains for Diverse and Evolving Interests. arXiv:2602.10490

[R3] Decision4Rec (2026) Polignano, M., Mehta, H. D., Lops, P., De Luca, E. W., & Semeraro, G. (2026). Decision4Rec: A Jev-Inspired Framework of Decision Models on Knowledge and Active Graphs for Multi-Agent Recommender Systems. arXiv:2610.26035

[R4] PCTX (2025) Zhong, Q., Su, J., Ma, Y., McAuley, J., & Hou, Y. (2025). PCTX: Tokenizing Personalized Context for Generative Recommendation. arXiv:2510.21276

Q2:

We acknowledge that DART incurs substantially higher training costs than single-pass discriminative baselines, particularly because Stage 3 requires running the reverse diffusion process. We have not conducted a systematic wall-clock training-time comparison and therefore cannot provide a reliable quantitative estimate of this overhead.

However, computational efficiency is not the primary objective of this work. Our primary objective is instead to address the limitations of diffusion-based recommendation, particularly their difficulty in improving recommendation diversity while maintaining ranking performance. The additional cost should be considered in this context.

We adopted DDPM as a well-established framework, consistent with prior diffusion-based recommendation methods. Reducing the computational overhead, for instance through accelerated sampling or alternative generative formulations such as flow matching, is an interesting direction that we leave for future work.

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

This paper finds that the evaluated target-item diffusion recommenders often produce nearly identical outputs across repeated draws conditioned on the same user history. DARTS separates representation learning, diffusion training in a frozen latent space, and final ranking-oriented decoding, and uses diagnostics in both latent and item space to study collapse. It improves ranking on three Amazon categories, and aggregating predictive distributions from multiple draws provides additional gains at a fixed slate size.

Critical strengths
  • I appreciate the distinction between using diffusion and retaining useful sampling variation. Table 2 and the controlled interventions in Appendices C–D give this observation explanatory value beyond another performance comparison.
  • The empirical study is fairly complete. It includes feature-matched baselines, aggregation at a fixed slate size, an unsuccessful candidate-union alternative, and serving costs. In particular, the roughly 3.2% aggregation gain shows that the sampling variation has a practical use.
Critical weaknesses
  • The ranking gains have not been adequately separated from the broader changes to representation learning and training. A deterministic latent predictor using the same encoder, decoder, and ranking fine-tuning is missing, so the overall advantage cannot be attributed entirely to diffusion diversity.
  • The mechanistic conclusions are broader than the evidence. Most diagnostics and ablations use a particular Scientific configuration. A CE intervention through a frozen decoder and one Gumbel control do not establish the general claim that discrete supervision should be excluded from diffusion training.
  • Figure 1 as main figure could be further refined to improve its visual appeal.
Questions for authors
  • If the Stage-1 representations, decoder, and Stage-3 objective are held fixed, how much of the ranking and aggregation advantage remains when diffusion is replaced by a deterministic latent predictor? This control would directly affect my interpretation of the mechanism.
  • Clarification: Were loss or gradient scales matched in the CE ablations, and was the final decoder re-adapted for each variant? The answer would affect whether I attribute the degradation to discrete supervision itself or to the particular training recipe.
1 comment

W1&Q1: We thank the reviewer for this important suggestion. We conducted the requested ablation by replacing Stage-2 diffusion training with a deterministic latent predictor, while keeping the Stage-1 representations, decoder, and Stage-3 objective unchanged.

The results show a substantial performance gap between the deterministic variant and the full model. DART improves over the deterministic predictor by 43–44% across all four ranking metrics (Hit@10, Hit@5, NDCG@10, and NDCG@5).

ModelR @ 10R @ 5Ndcg @ 10Ndcg @ 5
Deterministic latent predictor0.03650.02280.01890.0145
Darts0.05240.03290.02700.0207

We also observed substantial overfitting when training the deterministic predictor, which proved difficult to control. Nevertheless, the results provide evidence that the performance gains are not simply due to latent prediction and subsequent decoding, but that the diffusion-based training contributes substantially to recommendation performance. We have added this ablation and the corresponding discussion to the revised manuscript in appendix J.

W2:

We thank the reviewer for pointing out that our mechanistic conclusions should be more carefully scoped. We do not claim that discrete supervision should be universally excluded from diffusion training. Rather, our claim is specific to recommendation, where CE supervision can encourage sampling collapse and limit recommendation diversity. This concern is also motivated by prior work reporting sampling collapse in diffusion-based recommendation [1–4].

Recent works such as ELF, Cola DLM, and TextLDM further suggest that similar issues may arise in text generation, although this remains an active research question. We have therefore revised our introduction to clarify the scope of our claims and avoid overgeneralizing beyond the recommendation setting.

[1] FlowRec (2025) Li, L., Cheng, M., Ye, Y., Liu, Z., & Chen, E. (2025). Preference Trajectory Modeling via Flow Matching for Sequential Recommendation. arXiv:2508.17618

[2] DiffuRec (2023) Li, Z., Sun, A., & Li, C. (2023). DiffuRec: A Diffusion Model for Sequential Recommendation. arXiv:2304.00686

[3] ADRec (2025) Chen, J., Xu, Y., & Jiang, Y. (2025). Unlocking the Power of Diffusion Models in Sequential Recommendation: A Simple and Effective Approach. arXiv:2505.19544

[4] InDiRec (SIGIR 2025)

Qu, Y., & Nobuhara, H. (2025). Intent-aware Diffusion with Contrastive Learning for Sequential Recommendation. DOI: 10.1145/3726302.3730010

[5] ELF (2026) Hu, K., Qiu, L., Lu, Y., Zhao, H., Li, T., Kim, Y., Andreas, J., & He, K. (2026). ELF: Embedded Language Flows. arXiv:2605.10938

[6] Cola DLM (2026)

Guo, H., Zhao, Q., Zhao, Y., Nie, S., Zhu, R., Guo, Q., Wang, F., Yang, T., Zhao, H., Wei, G., & Zeng, Y. (2026). Continuous Latent Diffusion Language Model. arXiv:2605.06548

[7] TextLDM (2026)

Jiang, J., Ren, J., Li, W., Wang, B., Sun, H., Yang, Y., Liu, J., Zhang, Y., Zheng, S., Zhang, Y., Huang, H., Duan, N., & Zuo, W. (2026). TextLDM: Language Modeling with Continuous Latent Diffusion. arXiv:2605.07748

Q2:

We thank the reviewer for this clarification. The requested ablation is reported in Table 7 of the manuscript (Section 4.4 and Appendix C). Starting from the same Stage-1 checkpoint, we retrained Stage 2 using a weighted combination of latent diffusion loss and cross-entropy loss, varying the CE weight (α\alpha) from 0 to 1. All other components, including the initialization, architecture, and noise schedule, were kept unchanged.

Importantly, the decoder was re-adapted independently for each value of \alpha using the same Stage-3 training procedure. The results show that even a small CE contribution (α\alpha=0.1) substantially reduces sampling diversity (R: 0.30 to 0.14; PR: 4.70 to 1.50) and recommendation accuracy (NDCG@5: 2.07% to 1.51%). Increasing the CE weight further degrades both.

Following your suggestion, we have made this experimental protocol explicit in Appendix C, clarifying that the decoder is re-adapted for every value of α\alpha.

Reviewer_8ELuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies why diffusion-based recommenders often generate nearly identical recommendations when sampling multiple times for the same user. The authors propose DARTS, a three-stage framework that separates representation learning, diffusion training, and ranking, and show that it improves both recommendation accuracy and sampling diversity on three Amazon Reviews 2023 datasets.

Critical strengths
  1. The paper identifies an interesting problem in diffusion-based recommendation: repeated sampling often produces nearly identical recommendations. Studying this issue beyond conventional accuracy metrics is valuable.
  2. The three-stage design is straightforward and well motivated, with clear separation between representation learning, diffusion training, and ranking.
  3. The experiments are reasonably comprehensive, including comparisons with strong baselines, ablation studies, and analyses of sampling diversity and inference efficiency.
Critical weaknesses
  1. The novelty over existing diffusion recommendation methods, particularly ADRec, is not sufficiently clear. The authors should better explain what fundamentally distinguishes their sampling collapse analysis from previous work on embedding collapse.
  2. The chosen diffusion noise schedule may not sufficiently remove information from the original latent. With 200 steps and a linear noise schedule, the terminal signal coefficient remains non-negligible αˉT≈0.132\bar{\alpha}_T \approx 0.132, while inference directly initializes from a standard Gaussian. The authors should investigate this potential distribution mismatch.
  3. The final item is always masked during autoencoder pretraining, but the diffusion target is obtained by encoding the unmasked full sequence. It is unclear whether the frozen encoder can reliably produce this target representation, and this design choice deserves further justification.
Questions for authors
  1. In Stage 1, the final item is always masked, whereas the diffusion target in Stage 2 is obtained from the encoder output of the unmasked full sequence. What is the motivation for this choice, and have the authors investigated alternative target representations?
  2. Given that the terminal diffusion signal coefficient remains non-negligible under the chosen noise schedule, how do the authors justify initializing the reverse process directly from a standard Gaussian distribution? Have they evaluated alternative noise schedules?
  3. How do the authors demonstrate that the model learns a conditional distribution with multiple plausible modes rather than simply generating random perturbations around the conditional mean?
  4. Before and after Stage 3, how much of the diversity in the latent space is actually reflected in item semantics and recommendation outputs?
3 comments

W1:

ADRec focuses on item representation diversity, whereas our work addresses recommendation-generation diversity. Specifically, ADRec addresses embedding collapse in the item-representation space. In contrast, we investigate sampling collapse, where different diffusion trajectories conditioned on the same user history produce nearly identical recommendations, limiting the ability to capture diverse user interests.

Diverse item embeddings do not necessarily guarantee diverse generated recommendations. We clarified this distinction in the revised Introduction and Related Work. Thus, the distinction is between diversity in the learned item space and diversity among generated outputs.

W2&Q2:

We thank the reviewer for highlighting the potential mismatch between the forward process's terminal distribution and the standard Gaussian used at inference. We evaluated three noise schedules: linear, cosine, and exponential, with terminal signal coefficients of approximately 0.13, 0, and 0.006, respectively.

The results reveal a trade-off between recommendation quality and sampling diversity. Linear achieves the highest NDCG@5 (2.07%), while cosine gives comparable performance (2.02%) with higher sampling diversity (J = 0.32). Exponential performs worse on both metrics (NDCG@5 = 1.99%, J = 0.25%).

These findings suggest that cosine better mitigates the terminal distribution mismatch while maintaining competitive performance. We added these experiments to Appendix H.

W3:

This design is motivated by two complementary considerations:

  • We use the unmasked full sequence to construct the diffusion target because it preserves information about the actual next item. A final-item-masked sequence would provide little or no signal about the next item.
  • Following your remark, we added an ablation in which the final item is masked with the same probability as other items during Stage 1. This encourages masked-item modeling at the expense of next-item prediction. The ablation confirms a performance drop of approximately 28–29% across all metrics (see table).
ModelHit @ 10Hit @ 5Ndcg @ 10Ndcg @ 5
Dart with Stage 1(40% masking last item)4.07%2.57%2.09%1.60%
Dart5.24%3.29%2.70%2.07%
+/-+28%+28%+29.1%+29.3%

We added this ablation to Appendix I.

  • Finally, our goal is not to obtain the best latent representation in isolation, but to learn a latent space in which diffusion can effectively model the target distribution and the decoder can accurately recover the corresponding item.

Q3:

Directly evaluating diversity in sequential recommendation is challenging, as datasets typically provide only one observed next item rather than multiple valid recommendations for a given context.

As a proxy, we ran an experiment using 15-item sequences, holding out five as alternative relevant recommendations. We measured their recovery across multiple independent draws. Coverage@10 increases from 4.49% with one draw to 6.41% with seven, indicating that different draws can recover different relevant items.

Number of Draws1234567
Coverage@100.04490.04860.04980.05130.05500.05770.0641

Although this experiment does not establish multiple distinct modes in the learned conditional distribution, it provides strong empirical evidence that DART generates diverse, relevant recommendations across independent draws.

These Coverage@10 values should not be directly compared with standard Recall@10. This experiment uses shorter training sequences and a different evaluation objective: the model retains the next-item training objective but is evaluated on recovery of five held-out items across multiple draws. Training directly to recover these five items would change the objective and evaluate a different model. We added this experiment to Appendix M.

Q4:

To assess how latent diversity translates into recommendation diversity, we performed four independent inference draws for 10,000 test histories and analyzed the top-1 recommendations.

The top-1 item differs across draws for 61.8% of histories. For semantic diversity, we used Amazon's level-2 product categories, the second level of its product-category hierarchy. We found that 35.22% of histories receive recommendations spanning multiple categories, with an average of 1.434 distinct categories across four draws. This category-level measure captures semantic variation beyond changes in item identity alone.

These results show that latent diversity is reflected in final recommendations at both item and semantic-category levels. We added this analysis to Appendix L.

Thank you for the detailed response and additional experiments. The paper identifies an interesting phenomenon, and the experimental results are promising. However, I still have concerns about the distinction between harmful sampling collapse and reasonable concentration of the conditional distribution. After considering all the responses and comments from other reviewers, I have decided to maintain my original score.

We thank the reviewer for raising this important point. We agree that concentration in the conditional distribution is not necessarily undesirable. When a user's interests are strongly concentrated, generating similar recommendations across multiple draws is perfectly reasonable. Indeed, our results show that DARTS can preserve such consistency: 38.2% of histories produce the same top-1 item across all four draws, while 64.78% remain within a single semantic category.

However, sampling collapse becomes problematic when multiple relevant items or interests are plausible for a given history. In such cases, a model that systematically collapses may fail to capture these alternative possibilities. Our experiments demonstrate that DARTS can preserve this ability: different draws recover different relevant future items (Coverage@10 increasing from 4.49% to 6.41% in appendix M), while 61.8% of histories exhibit different top-1 recommendations across draws (Appendix L). Furthermore, 35.22% of histories receive recommendations spanning multiple semantic categories.

Therefore, the objective of DARTS is not to maximize diversity indiscriminately, but to preserve the ability to generate alternative relevant recommendations when appropriate, while maintaining consistency when user preferences are concentrated. The improvements in ranking accuracy further suggest that this diversity is achieved without sacrificing recommendation relevance.

Comments on the paper

2 comments

We thank all reviewers for their comments and suggestions. We have carefully addressed their concerns, conducted several additional experiments, and revised the manuscript accordingly (changes highlighted in blue in the manuscript).

The additional experiments strengthen our main conclusions in four directions:

  1. Contribution of diffusion (Appendix J). To disentangle the contribution of diffusion from pretraining and fine-tuning, we introduced a deterministic latent predictor using the same Stage-1 representations, decoder, and Stage-3 ranking objective. DART outperforms this controlled baseline by 43–44% across all four ranking metrics, providing evidence that diffusion contributes substantially to recommendation performance.

  2. Diversity and relevance of generated recommendations (Appendices L–M). Independent diffusion draws produce different top-1 recommendations for 61.8% of user histories, including recommendations from different semantic categories for 35.2%. Coverage of relevant held-out future items increases from 4.49% to 6.41% as the number of draws increases from one to seven. These results show that sampling diversity translates into different, relevant recommendations, rather than merely variations in latent representations.

  3. Robustness and generalization (Appendices H and K). Additional experiments on MovieLens-1M confirm improvements of 4.8–15.2% over the strongest baseline. Multi-seed evaluations further support the robustness of our results, while noise-schedule experiments clarify the trade-off between ranking accuracy and sampling diversity.

  4. Methodological clarifications (Appendices C and I). Additional analyses support our masking strategy and clarify the impact of discrete supervision on sampling diversity. We also refined the distinction between sampling collapse and embedding collapse and clarified the scope of our claims.

Overall, these additional results reinforce our central contribution: diffusion-based recommendation can generate diverse and relevant alternatives while achieving strong ranking performance. We also acknowledge the additional computational cost and discuss this limitation in the revised manuscript.

As the rebuttal discussion period approaches its conclusion, we would like to kindly follow up on our response posted on 16 November and the additional experiments we provided.

In particular, we conducted further experiments comparing our approach with a deterministic model, evaluating the relevance and diversity of recommendations, and assessing our approach on the MovieLens dataset.

We would be happy to address any remaining concerns or provide further clarifications if needed. If our responses and additional evidence have satisfactorily addressed your concerns, we would sincerely appreciate your reconsideration of the rating.

Thank you again for your time, thoughtful feedback, and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.