acceptodds
Under review as a conference paper at ICLR 2027

SNR as a Supervisory Coordinate: Coarse-to-Fine Learning for Diffusion-based Recommendation

Abstract

Diffusion models have emerged as a powerful paradigm for sequential recommendation (SR) due to their ability to model user preferences as complex, multi-modal distributions. However, existing frameworks typically rely on static, point-wise supervision throughout the denoising trajectory. Such constant-resolution objectives are misaligned with the time-varying information that the noisy state retains about the target. Specifically, enforcing a precise point target in high-noise regimes leads to supervisory over-specification, forcing the model to resolve distinctions that exceed the recoverable information. To address this, we propose CloudDiff, a diffusion framework for SR that adaptively modulates its supervisory granularity according to the signal-to-noise ratio (SNR). CloudDiff introduces two SNR-coupled mechanisms: (1) a semantic cloud loss that aligns the denoiser's intermediate representations with set-valued targets of semantic neighbors, with their influence gated by the noise level; and (2) a log-SNR-coupled temperature schedule that dynamically adjusts the sharpness of the contrastive distribution. We provide an information-theoretic foundation showing that both axes of CloudDiff are governed by a single coordinate, log-SNR, jointly motivating the proposed coarse-to-fine scheme. Extensive experiments on multiple benchmarks demonstrate that CloudDiff consistently outperforms state-of-the-art baselines, with detailed analyses revealing how SNR-adaptive supervision stabilizes the refinement trajectory.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sequential Recommendation, Generative Recommendation, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_U7rhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a coarse-to-fine learning for diffusion recommenders, consisting of a semantic cloud construction module and a log-SNR-coupled temperature schedule. The motivation is straightforward and makes sense. My major concerns lie in the experimental evaluation (please refer to the weaknesses and my question). It seems that the baselines had not been tuned well, the results are much lower than those in their original papers.

Critical strengths
  1. The motivation of the proposed method is easy to understand and makes sense.

  2. Extensive theoretical analysis is provided to support the model design.

Critical weaknesses
  1. The results of the baselines present in this paper is not convicing to me. The results of the strongest baseline PreferDiff are significantly lower than those in the original paper, even under the same datasets and settings. This makes me concerned about whether the baselines are tuned properly.

  2. The baseline methods are not SOTA. The strongest baseline is from 2025 while most others are before 2023, which I think cannot sufficiently evaluate the effectiveness of the proposed method in such rapid-developed topic.

Questions for authors
  1. What does the performance of the proposed model change if you use a fixed cloud (maybe in the initial step or some middle steps) for every denoising steps?

  2. Could you show the results of you method, TIGER and DiffuRec under the leave-one-out evaluation setting? It seems that in this paper, these two methods' performances are much lower than in their original papers.

  3. Since you use a very large negtive pool (typically 128 and 256) in your method, I am curious about the number of negatives in PreferDiff in your implementation. I would also suggest to conduct hyperparameter analysis in terms of the number of negative to figure out whether the peroformance gain largely from sampling more negative samples.

  4. You use LLMs to construct soft positive, this makes me concerned about the training cost. Did you try any simple methods to sample soft positives?

  5. You use LLM-derived embeddings in your method. Did you also use these embeddings for other methods or simply use ID embeddings for them?

6 comments
AuthorsOpenReview ↗

Response to Reviewer U7rh (1/2)

We sincerely thank the reviewer for their thorough evaluation and constructive suggestions, particularly for identifying important areas to strengthen our experiments.

Response to W1: PreferDiff Reproduction.

In our original experiments, we searched embedding dimensions up to 1024 for all methods (see App. E.4.1), reporting results with 512 for CloudDiff and 1024 for PreferDiff. However, as noted in the original PreferDiff paper, PreferDiff is sensitive to embedding dimensions and considers up to 3072, which explains the lower performance reported in our paper. To address this, we additionally evaluated both methods at dimensions of 2048 and 3072 under the same settings, as shown below.

DimMethodMetricBeautyToysSportsYelp
2048CloudDiffN@100.03870.03640.01590.0115
R@100.05660.05140.02510.0225
PreferDiffN@100.03670.03570.01510.0087
R@100.05190.04960.02160.0136
3072CloudDiffN@100.03870.03630.01640.0128
R@100.05680.05080.02620.0219
PreferDiffN@100.03710.03690.01610.0107
R@100.05430.05180.02400.0181

As reported in the original PreferDiff paper, larger embedding dimensions substantially improve its performance. Nevertheless, CloudDiff outperforms PreferDiff on all four datasets at 2048 dimensions and three at 3072, with particularly strong gains on sparse Yelp. More importantly, CloudDiff achieves strong performance with only 512 hidden dimensions, without relying on larger representations and their associated computational costs.


Response to W2: Add Recent SOTA.

MethodMetricBeautyToysSportsYelp
DiffGRM [1] (WWW'26)N@100.02940.03270.01120.0102
R@100.05360.05020.02020.0181
PreferGrow [2] (NeurIPS'25)N@100.02940.01650.01020.0087
R@100.04510.02570.01600.0145
CloudDiff (Ours)N@100.03840.03470.01580.0118
R@100.05600.04840.02490.0209

We additionally compare CloudDiff against two recent diffusion-based recommendation methods, DiffGRM [1] (WWW'26) and PreferGrow [2] (NeurIPS'25). CloudDiff achieves the highest NDCG@10 across all four datasets and the highest Recall@10 on three datasets, demonstrating its competitiveness against recent SOTA baselines.


Response to Q1: Fixed Cloud.

Fig. 4 with p = 0 is exactly a cloud whose weight is held constant over all denoising steps, and the "cloud + static τ\tau" variant of Fig. 3 additionally fixes the temperature; both are below the full model, i.e. adapting the cloud with the noise level is what helps. We will make this correspondence explicit in the text.

We additionally evaluate CloudDiff with the cloud fixed at the midpoint timestep across all denoising steps.

MethodMetricBeautyToysSportsYelp
CloudDiff (midpoint tt Cloud)N@100.03630.03180.01210.0090
R@100.05350.04580.01920.0158
CloudDiff (Full)N@100.03840.03470.01580.0118
R@100.05600.04840.02490.0209

Fixing cloud supervision at an intermediate timestep yields lower performance than full CloudDiff across all datasets. This suggests that while semantic-cloud supervision is beneficial, adapting its strength to the diffusion noise level is important.

AuthorsOpenReview ↗

Response to Reviewer U7rh (2/2)

Response to Q2: Leave-One-Out Evaluation.

The performance gap mainly stems from different evaluation protocols. While the original DiffuRec and TIGER papers adopt leave-one-out (LOO) evaluation, our experiments follow a user-split protocol. Under LOO, earlier interactions from test users are available during training, whereas user-split evaluation holds out entire users, requiring models to generalize to unseen users. These differences make our setting more challenging and help explain the lower baseline performance compared with the original papers.

MethodMetricBeautyToysSportsYelp
DiffuRecN@100.04540.04070.02400.0184
R@100.07580.06510.03930.0289
TIGERN@100.03470.02970.01770.0120
R@100.05990.05310.03280.0222
PreferDiffN@100.04230.04640.02110.0174
R@100.07480.08520.03780.0264
CloudDiffN@100.04470.04880.02320.0189
R@100.07920.09000.04060.0313

Under the standard LOO setting, CloudDiff achieves the best R@10 across all four datasets and the best N@10 on Toys and Yelp, while remaining competitive on Beauty and Sports. These results demonstrate that CloudDiff maintains strong performance under both evaluation protocols.


Response to Q3: Negatives.

PreferDiff uses 256 in-batch negatives in our implementation. Reducing CloudDiff's negatives (main: 128 / 256) to 32 or 64:

# NegativesMetricBeautyToysSportsYelp
32N@100.03720.03290.01390.0109
R@100.05470.04680.02310.0189
64N@100.03770.03340.01370.0111
R@100.05540.04850.02280.0180
Main (128 / 256)N@100.03840.03470.01580.0118
R@100.05600.04840.02490.0209

CloudDiff maintains competitive performance even with substantially fewer negatives. While larger negative pools generally improve performance, the gains are modest, suggesting that CloudDiff’s improvements are not primarily driven by negative pool size.


Response to Q4: Other Soft Positives.

Appendix G.3 demonstrates that the cloud can be constructed without LLMs, using SPPMI co-occurrence or SASRec item-embedding neighbors.

Neighbor SpaceBeautyToysSportsYelp
LLM0.03840.03470.01580.0118
SPPMI-SVD0.03610.03150.01160.0097
SASRec0.03590.03420.01300.0107

The results demonstrate robustness across different soft-positive constructions, with LLM-derived semantics being one option rather than a requirement. Our core contribution lies in noise-adaptive supervision, rather than the specific neighborhood construction. Also, LLM generation incurs only a one-time offline cost.


Response to Q5: LLM Embeddings For Other Methods.

Baselines were implemented following their original designs. Methods using LLM-derived embeddings (e.g., LLM-ESR) were provided with the same embeddings as CloudDiff, while ID-based methods retained their trainable ID embeddings. CloudDiff also uses trainable ID embeddings for recommendation, with LLM-derived embeddings providing auxiliary supervision through semantic neighborhood and denoiser representation alignment.


References

[1] Liu, Zhao, et al. Diffgrm: Diffusion-based generative recommendation model. In: Proceedings of the ACM Web Conference 2026. 2026. p. 5853-5864.

[2] Hu, Guoqing, et al. Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation. Advances in Neural Information Processing Systems, 2025, 38: 135766-135804.

Thanks for your detailed responses.

My concern on the training cost of LLM-derived soft positive (or other methods) is still not addressed.

Another question is whether the soft positives are selected from all items?

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

This paper studies whether a fixed, item-level training target is appropriate throughout the denoising trajectory of a diffusion-based sequential recommender. The authors argue that, when the signal-to-noise ratio (SNR) is low, the noisy target embedding xtx_t contains insufficient information to support fine-grained discrimination of the ground-truth item ii. They propose CloudDiff, which augments x0x_0-prediction with a semantic-cloud contrastive objective applied to the denoiser's intermediate representation ztz_t. Neighbors Ni(K)\mathcal N_i(K) are constructed from frozen, LLM-derived item embeddings. Their contribution is scaled by Ω(t)=(1−αˉt)p\Omega(t)=(1-\bar\alpha_t)^p, while the contrastive temperature follows a log-SNR-coupled schedule τ(t)=τloκu(t)\tau(t)=\tau_{\mathrm{lo}}\kappa^{u(t)}. The intended effect is broad semantic supervision at high noise and sharper target discrimination at low noise. The paper reports the best HR and NDCG results among its compared methods on four full-ranking benchmarks, with ablations and supplementary analyses supporting the usefulness of both mechanisms.

Critical strengths
  1. Strong and consistent ranking results under a demanding evaluation protocol. CloudDiff achieves the best reported HR and NDCG results at cutoffs 5, 10, and 20 on Beauty, Toys, Sports, and Yelp using full-item ranking rather than sampled candidates. Its NDCG@10 scores are 0.0384, 0.0347, 0.0158, and 0.0118, respectively. The breadth and consistency of these gains are a substantial reason to take the proposed training strategy seriously.

  2. A clear two-axis realization of coarse-to-fine supervision. The semantic-cloud loss LSC(t)=−log⁡Mi(t)+Ω(t)Pi(t)Mi(t)+Ω(t)Pi(t)+Ni(t)\mathcal L_{\mathrm{SC}}(t)=-\log\frac{M_i(t)+\Omega(t)P_i(t)}{M_i(t)+\Omega(t)P_i(t)+N_i(t)} uses Ω(t)=(1−αˉt)p\Omega(t)=(1-\bar\alpha_t)^p to control the contribution of semantic neighbors, while τ(t)=τloκu(t)\tau(t)=\tau_{\mathrm{lo}}\kappa^{u(t)} controls contrastive sharpness. Applying this objective to the intermediate representation ztz_t gives the semantic prior a concrete role in training without requiring LLM inference at recommendation time.

  3. Useful component-level ablations. Figure 3 compares the full model against cloud supervision with a static temperature, point supervision with a log-SNR temperature, and point supervision with a static temperature. Figure 4 varies the fade exponent pp. These experiments provide evidence that the observed gains are not merely due to adding an arbitrary auxiliary contrastive loss, and they help identify the contribution of adapting both supervisory support and sharpness.

Critical weaknesses
  1. The proposed mechanism is not directly isolated by the objective or experiments. The paper argues that exact item-level targets over-specify supervision at high noise, yet CloudDiff retains Ldiff=∥x0−x^0∥2\mathcal L_{\mathrm{diff}}=\|x_0-\hat x_0\|^2 at every timestep and adds the semantic-cloud loss to ztz_t. A point-valued observation does not by itself imply that the population-optimal squared-error predictor must identify that observation; it predicts E[x0∣xt,hu]\mathbb E[x_0\mid x_t,h_u]. Moreover, the denoiser observes the user-history representation huh_u, which may remain informative even when xtx_t has low SNR. The results establish that the auxiliary supervision helps ranking, but do not yet show that its benefit specifically comes from correcting the stated high-noise failure mode.

  2. The theoretical results motivate the design but do not establish the claimed necessity of its specific schedules. Proposition 1 is an upper bound on posterior MSE, not a matching lower bound for arbitrary item distributions; it therefore cannot alone establish the severity of over-specification at intermediate timesteps. The posterior-deviation scaling motivates the choice p≥1/2p\geq 1/2, but does not prove that smaller pp must fail for the conditional recommendation task. Proposition 3 gives a local KL expansion for changing log⁡τ\log\tau, but does not uniquely imply the proposed log-linear coupling to log-SNR. In Appendix B.4, the claim that the displayed Fano-type lower bound approaches one as SNR vanishes is also inaccurate for fixed cluster size MM: the bound approaches 1−log⁡2/log⁡M1-\log 2/\log M. I would prefer the paper to distinguish mathematical implications from design hypotheses supported empirically.

  3. A temperature-schedule comparator may run in the opposite direction. Appendix E.5 defines the αˉt\bar\alpha_t-linear baseline as τ(t)=τlo+(τhi−τlo)αˉt\tau(t)=\tau_{\mathrm{lo}}+(\tau_{\mathrm{hi}}-\tau_{\mathrm{lo}})\bar\alpha_t. As written, this assigns low temperature to high-noise states and high temperature to low-noise states, reversing the intended coarse-to-fine sharpening. If this is the implemented schedule, Figure 5 does not fairly isolate the advantage of using log-SNR rather than αˉt\bar\alpha_t as the scheduling coordinate. If it is a typographical error, the actual implementation and a direction-matched comparison should be provided.

  4. The necessity and cost of the Gemma-based semantic pipeline are insufficiently established. The paper uses Gemma-3-27B-it to generate preference-oriented descriptions before encoding them, but does not compare against feeding the original title and description into the same embedding model. Consequently, the gain from this large generative step is not separated from the gain from the semantic-cloud objective. The manuscript also mentions associated reviews in constructing descriptions without making their temporal availability fully clear, and does not report the offline generation cost alongside training and 100-step DDIM inference costs. These omissions limit assessment of both experimental validity and practical value.

Questions for authors
  1. Temperature comparator: Is the αˉt\bar\alpha_t-linear formula in Appendix E.5 a typographical error or the schedule actually used in Figure 5? Please provide the implemented formula and compare against a direction-matched schedule with the same temperature endpoints, for example τlo+(τhi−τlo)(1−αˉt)\tau_{\mathrm{lo}}+(\tau_{\mathrm{hi}}-\tau_{\mathrm{lo}})(1-\bar\alpha_t).

  2. Role and provenance of generated text: How does CloudDiff perform when the original title and description are encoded directly by the same embedding model, with no Gemma-generated intermediate description? Were item reviews used to produce the descriptions, and, if so, were they restricted according to the temporal training split? This would clarify both the contribution of Gemma-3-27B-it and whether future information could enter the semantic representations.

  3. Cost of the improvement: Could the authors report offline description-generation cost, training time, and per-user inference latency alongside the ranking gains? The proposed method avoids running the LLM at inference, but still uses iterative DDIM sampling.

4 comments
AuthorsOpenReview ↗

Response to Reviewer AaaF (1/3)

We thank the reviewer for the careful reading of our paper, attention to detail, and thoughtful questions. In particular, your insightful comments on the theoretical analysis have helped us strengthen the clarity of our paper.

Response to W1: High-Noise Failure Mode and Empirical Evidence.

We thank the reviewer for this important observation. We agree that the optimal x0x_0-MSE predictor estimates the conditional posterior mean, and that user history remains informative even at high noise. Our motivation is therefore not that MSE inherently forces exact-item identification, but that point-wise auxiliary supervision may be overly restrictive under high uncertainty. CloudDiff addresses this by introducing semantic neighborhood-level supervision that adapts to the noise level, while retaining the original diffusion MSE. We have revised the manuscript to clarify and appropriately qualify this motivation.

(a) Isolating the contribution of high-noise supervision. We conduct an additional ablation that restricts cloud supervision to different SNR regimes, while maintaining the diffusion MSE across all timesteps.

Cloud SupervisionBeauty N@10Yelp N@10
None (λ=0\lambda=0)0.03190.0050
Low-noise only (SNR≥1\mathrm{SNR}\geq 1)0.03290.0059
Mid-noise only0.03410.0068
High-noise only (SNR<0.01\mathrm{SNR}<0.01)0.03610.0086
Full (CloudDiff)0.03840.0118

High-noise-only supervision achieves the strongest performance among the restricted settings, recovering most of the gains obtained with full supervision. These results highlight the substantial contribution of semantic-cloud supervision in the high-noise regime, while full-range supervision provides additional benefits.

(b) Direct evaluation of high-noise predictions. We further evaluate the denoiser's x^0\hat{x}_0 prediction directly at the maximum noise level (t=Tt=T), without iterative denoising, by ranking all candidate items.

Auxiliary TargetBeauty GT Rank ↓Beauty N@10 ↑Yelp GT Rank ↓Yelp N@10 ↑
None (λ=0\lambda=0)0.4270.02830.4330.0042
Exact-item (point)0.3920.03290.3250.0071
Semantic cloud0.3720.03750.2720.0109

CloudDiff consistently achieves the best high-noise predictions. Relative to no auxiliary supervision, NDCG@10 improves by 32.5% on Beauty and 159.5% on Yelp.

Together with Figs. 3-4, which examine semantic-cloud targets and noise-dependent scheduling, these results provide complementary evidence that CloudDiff's gains are associated with improved high-noise prediction and noise-adaptive supervision, rather than merely the presence of an additional contrastive objective. We will incorporate these analyses and clarify the motivation accordingly.

We have included these experimental results in Appendix F.4, as they directly support the core contributions of our paper.


Response to W3 & Q1: Temperature Schedule.

We thank the reviewer for identifying this typographical error in Appendix E.5. We confirm that the actual implementation uses the correct direction-matched schedule:

τ(t)=τhi−(τhi−τlo)αˉt\tau(t)=\tau_{\mathrm{hi}}-(\tau_{\mathrm{hi}}-\tau_{\mathrm{lo}})\bar{\alpha}_t,

which is mathematically equivalent to your suggested formula:

τ(t)=τlo+(τhi−τlo)(1−αˉt)\tau(t)=\tau_{\mathrm{lo}}+(\tau_{\mathrm{hi}}-\tau_{\mathrm{lo}})(1-\bar{\alpha}_t).

Thus, the results in Figure 5 were obtained using the correct coarse-to-fine temperature schedule, ensuring a fair comparison. We have corrected the formula in Appendix E.5 of the revised pdf.

AuthorsOpenReview ↗

Response to Reviewer AaaF (2/3)

W2. Theoretical Results vs. Design Choices.

Our propositions provide theoretical motivation for SNR-adaptive supervision, with the specific design choices further validated empirically.

(a) Proposition 1: Posterior uncertainty. Proposition 1 characterizes the scale of posterior uncertainty through an MMSE upper bound, rather than establishing the severity of over-specification. Appendix F.1 empirically supports this uncertainty pattern: posterior entropy decreases from 9.304 to 6.506, and the effective candidate count from 10,978 to 679 as noise decreases (Table 4). Figure 6 further shows lower per-sample loss variance than PreferDiff, particularly in the high-noise region. Accounting for user-history conditioning, the MMSE satisfies

E∥x0−E[x0∣xt,hu]∥2≤Ehutr⁡[Σh(I+SNR(t)Σh)−1],\mathbb{E}\|x_0-\mathbb{E}[x_0\mid x_t,h_u]\|^2 \le \mathbb{E}_{h_u}\operatorname{tr}\left[\Sigma_h(I+\mathrm{SNR}(t)\Sigma_h)^{-1}\right],

where Σh=Cov⁡(x0∣hu=h)\Sigma_h=\operatorname{Cov}(x_0\mid h_u=h). At zero SNR, the MMSE equals E∥x0−E[x0∣hu]∥2\mathbb{E}\|x_0-\mathbb{E}[x_0\mid h_u]\|^2, showing that uncertainty can persist despite user-history conditioning.

(b) Proposition 2: The condition p≥1/2p\ge 1/2. Proposition 2 establishes E(t)≤c21−αˉt\sqrt{\mathcal E(t)}\le\sqrt{c_2}\sqrt{1-\bar\alpha_t}, motivating the low-noise scaling criterion Ω(t)=O(1−αˉt)\Omega(t)=O(\sqrt{1-\bar\alpha_t}). For our power-law fade Ω(t)=(1−αˉt)p\Omega(t)=(1-\bar\alpha_t)^p, this criterion holds if and only if p≥1/2p\ge1/2. This provides a principled design guideline rather than a necessary condition for recommendation performance. Figure 4 further supports this choice, showing degraded performance for p<1/2p<1/2 and the best results at p=0.5p=0.5.

(c) Proposition 3: Log-SNR coupling. Proposition 3 motivates log⁡τ\log\tau as a natural coordinate, while scale equivariance (S↦cSS\mapsto cS implies log⁡τ↦log⁡τ+log⁡c\log\tau\mapsto\log\tau+\log c) supports its parameterization. Although the affine coupling with log-SNR is not uniquely implied, it is empirically supported: our schedule performs best on three of four datasets (Fig. 5), and the optimal log⁡τ⋆\log\tau^\star exhibits a strong linear relationship with log-SNR (R2=0.867R^2=0.867–0.9040.904, Appendix F.2), even for checkpoints trained without this schedule.

(d) Fano lemma. We acknowledge the incorrect limit statement. For fixed MM, the Fano lower bound approaches 1−log⁡2/log⁡M1-\log2/\log M, not 1. Under a uniform prior over a fixed candidate cluster, the optimal error at zero SNR is 1−1/M1-1/M. Hence, the conclusion that exact item identification remains difficult at high noise is unchanged. We have corrected the statement and provide the general conditional bound:

Pe≥H(V⋆∣Hu)−I(V⋆;xt∣Hu)−log⁡2log⁡M.P_e\ge\frac{H(V^\star\mid H_u)-I(V^\star;x_t\mid H_u)-\log2}{\log M}.

We have revised Introduction, Sections 3.3 and 4.4 and Appendix B.4 to clarify the theoretical implications and distinguish them from design choices supported by empirical evidence.

AuthorsOpenReview ↗

Response to Reviewer AaaF (3/3)

Response to W4 & Q2: Semantic Description Ablation and Clarification.

(a) w/o Gemma-generated descriptions. We directly encoded the original item metadata (e.g., title, brand, and description) using mxbai-embed-large-v1, without Gemma-based description generation. All other settings remained unchanged.

MethodMetricBeautyToysSportsYelp
Raw Metadata (w/o Gemma)N@50.03420.03240.01280.0081
R@50.04390.04040.01810.0115
N@100.03780.03450.01470.0098
R@100.05540.04800.02410.0170
CloudDiff (w/ Gemma)N@50.03510.03250.01390.0101
R@50.04500.04140.01900.0147
N@100.03840.03470.01580.0118
R@100.05600.04840.02490.0209

On the three Amazon datasets, where item metadata is relatively rich, using raw metadata achieves performance close to the original CloudDiff without Gemma-generated descriptions. This suggests that the improvements mainly come from the semantic-cloud objective rather than description generation. In contrast, Yelp provides only sparse metadata (e.g., business names and categories), where generating preference-oriented descriptions helps improve recommendation performance.

(b) Reviews. We apologize for the incorrect statement in the paper. Gemma uses only item metadata to generate descriptions, without access to item reviews or interaction data. Thus, no temporal leakage occurs during description generation. We have corrected the statement in Appendix E.3 of the revised pdf.


Response to Q3: Cost.

We report training and inference costs measured on a single RTX A6000 (median over 3 epochs).

DatasetModelHidden Dim.Train (s/epoch)Inference (s/test set)
BeautyCloudDiff51274
PreferDiff51254
PreferDiff1024 (our main table)85
SportsCloudDiff512107
PreferDiff51296
PreferDiff1024 (our main table)148

Training: CloudDiff requires less training time than PreferDiff in its original configuration, as it uses a smaller hidden dimension (512 vs. 1024).

Inference: CloudDiff uses the same 100-step DDIM sampler as PreferDiff, with comparable latency ≈1.8–2.0\approx 1.8–2.0 ms/user. Neither LLM embeddings nor cloud supervision are required at inference.

Offline preprocessing: Gemma-3-27B (vLLM) generates descriptions at ≈0.45\approx 0.45 s/item, followed by description encoding at ≈2.6\approx 2.6 ms/item using mxbai. Semantic neighbors are then constructed via a one-time kNN search.

We have added the computational cost analysis to Appendix H.

Reviewer_AaeFOpenReview ↗

Response to authors

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts in addressing my concerns, particularly the new high-noise ablations, the clarification of the theoretical claims, and the correction of the temperature schedule.

However, my primary concern remains. While the additional results support the effectiveness of semantic-cloud supervision, they do not fully establish that the improvements specifically arise from mitigating supervisory over-specification. Moreover, the theoretical analysis still serves primarily as motivation rather than a rigorous justification for the specific SNR-dependent design choices.

Overall, the rebuttal has clarified several issues, but my concerns about the underlying mechanism and theoretical justification have not been sufficiently resolved. Therefore, I will maintain my original rating.

Reviewer_yppbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses diffusion-based sequential recommendation, aiming to predict the next item by modeling dynamic user preferences through a generative diffusion process. It targets the mismatch between static point-wise supervision and the time-varying information structure of diffusion. Existing methods force exact target prediction in high-noise regimes where information is scarce, leading to supervisory over-specification, high variance, and unstable optimization. To solve this, the proposed workflow involves a forward noising and denoising process with key modules: (1) a semantic cloud using frozen LLM embeddings to retrieve nearest neighbors as soft positives; (2) a dynamic fade function modulating the cloud's contribution based on noise levels; (3) a log-SNR coupled temperature schedule adjusting contrastive sharpness; and (4) an SNR-gated semantic cloud contrastive loss generalizing standard InfoNCE. Experimental findings on benchmark datasets demonstrate that CloudDiff achieves state-of-the-art performance, outperforming strong baselines. Ablations confirm that both the spatial support and predictive sharpness must be dynamically adjusted. Trajectory analyses validate the intended coarse-to-fine behavior, transitioning from broad semantic supervision at high noise to exact target discrimination at low noise, while maintaining more stable training signals.

Critical strengths

Conventional methods apply a single fixed point-wise target throughout the denoising trajectory, leading to the so-called supervisory over-specification. This paper achieves SNR-adaptive supervisory granularity by adapting the spatial support and sharpness of the supervisory target itself, rather than merely re-weighting the loss contribution of a fixed target.

Critical weaknesses

Elevating SNR from a tuning parameter for loss weighting to a coordinate for the shape of the supervisory target is a perspective shift with methodological value. However, the method assumes by default that all training samples benefit from a coarse-to-fine supervisory transition; when user intent itself is highly concentrated, the necessity of the broad supervision provided by the semantic cloud is left undiscussed. Moreover, prior work in diffusion-based recommendation has introduced curriculum learning strategies into the training process, such as CCDRec. The paper does not discuss the essential differences between the two, nor does it explain why SNR-coupled adjustment of supervisory granularity is superior to curriculum strategies based on training stages.

Questions for authors
  1. Why not compare with the baselines of Masked Diffusion.

  2. Why not compare with the model presented in the paper of Plug- in diffusion model for sequential recommendation?

  3. Is the broad supervision provided by the semantic cloud still necessary when the user intent itself is highly concentrated?

  4. What are the implementation details and hyperparameter tuning strategies of the compared algorithms? Are all models kept consistent in terms of LLM usage?

4 comments
AuthorsOpenReview ↗

Response to Reviewer yppb (1/2)

We thank the reviewer for their insightful and important questions regarding our methodology and its key distinctions from other works.

Response to W & Q3: Semantic Cloud under Concentrated User Intent.

To assess the benefit of semantic cloud supervision under concentrated user intent, we compare CloudDiff with its w/o cloud variant. We divide users into quartiles based on the mean pairwise cosine similarity of their history-item embeddings (Q1: dispersed; Q4: concentrated).

Table. Relative improvement (%) of CloudDiff over w/o Semantic Cloud across intent concentration quartiles.

DatasetMetricQ1 (Dispersed)Q2Q3Q4 (Concentrated)
BeautyR@10+19%+33%+25%+18%
N@10+17%+26%+20%+10%
YelpR@10+100%+45%+148%+120%
N@10+75%+56%+140%+74%

CloudDiff consistently improves performance across all intent concentration levels. Notably, even for highly concentrated users (Q4), it achieves substantial gains on the sparse Yelp dataset (+120% R@10 and +74% N@10), highlighting the value of semantic-cloud supervision even when user intent is relatively focused.


Response to W: Difference from Curriculum Learning (CCDRec)..

CloudDiff fundamentally differs from curriculum learning approaches such as CCDRec [1], as summarized below.

AspectCCDRec [1]CloudDiff (Ours)
Adjustment scopeGlobal: across training epochsLocal: per-sample, within each training iteration
Adjustment criterionTraining progressDiffusion timestep (SNR)
What is adjusted?Negative sampling difficultySupervisory target granularity
StrategyEasy-to-hard negativesHigh noise: broad supervision; low noise: fine-grained supervision

Unlike epoch-level curricula, CloudDiff adapts supervision to each sample's noise level within the same training iteration. Since high- and low-noise samples coexist throughout diffusion training, SNR-coupled supervision directly accounts for timestep-dependent uncertainty, regardless of training progress.

We have clarified the distinction between curriculum learning-based approaches and our method in Appendix A.2.

AuthorsOpenReview ↗

Response to Reviewer yppb (2/2)

Response to Q1: Masked Diffusion Baseline.

Our original baseline selection focused on (1) conventional sequential recommendation models, (2) methods leveraging semantic information, and (3) continuous diffusion-based recommenders, which are methodologically closest to CloudDiff. Unlike continuous diffusion, masked diffusion operates on discrete tokens and employs a different corruption and generation process.

MethodMetricBeautyToysSportsYelp
DiffGRM [2] (WWW'26)N@100.02940.03270.01120.0102
R@100.05360.05020.02020.0181
PreferGrow [3] (NeurIPS'25)N@100.02940.01650.01020.0087
R@100.04510.02570.01600.0145
CloudDiff (Ours)N@100.03840.03470.01580.0118
R@100.05600.04840.02490.0209

To broaden our evaluation, we additionally compare CloudDiff with DiffGRM [2], a masked discrete diffusion-based generative recommender, and PreferGrow [3], a discrete diffusion-based recommender using preference fading and growing.

These comparisons extend our evaluation beyond continuous diffusion and demonstrate CloudDiff’s competitiveness across different diffusion paradigms.


Response to Q2: Comparison with PDRec.

Our original baseline selection focused on methods that directly employ diffusion models for recommendation. In contrast, PDRec [4] uses diffusion as a plug-in to enhance conventional sequential recommenders through behavior reweighting, positive augmentation, and negative sampling. Nevertheless, we additionally evaluate PDRec to broaden our comparison.

MethodMetricBeautyToysSportsYelp
PDRec [4] (AAAI'24)N@100.03140.01850.01290.0127
R@100.05270.03300.02480.0223
CloudDiff (Ours)N@100.03840.03470.01580.0118
R@100.05600.04840.02490.0209

CloudDiff outperforms PDRec in 6 of 8 comparisons, particularly on Beauty and Toys. PDRec performs better on Yelp, potentially benefiting from its behavior reweighting and diffusion-based positive augmentation, which help alleviate data sparsity.

We have revised Appendix A.2 to clarify the distinctions from PDRec and discuss discrete-space diffusion-based recommendation methods, and added performance comparisons in Appendix G.6.


Response to Q4: Implementation details, Hyperparameter Tuning, and LLM Usage.

(a) Implementation and hyperparameter tuning. As detailed in Appendix E.4, all baselines follow their original implementations. We conduct a common grid search over learning rates, embedding dimensions, and in-batch negatives, while also tuning model-specific hyperparameters using the search spaces provided in Appendix E.4.

(b) Consistency of LLM usage. Baselines follow their original designs. Methods using LLM-derived embeddings (e.g., LLM-ESR) receive the same embeddings as CloudDiff, while ID-based methods retain their trainable ID embeddings. CloudDiff likewise uses trainable ID embeddings for recommendation, with LLM-derived embeddings used only for auxiliary supervision through semantic neighborhoods and denoiser representation alignment.


References

[1] Yang, Yimeng, et al. Curriculum conditioned diffusion for multimodal recommendation. In: Proceedings of the AAAI conference on artificial intelligence. 2025. p. 13035-13043.

[2] Liu, Zhao, et al. Diffgrm: Diffusion-based generative recommendation model. In: Proceedings of the ACM Web Conference 2026. 2026. p. 5853-5864.

[3] Hu, Guoqing, et al. Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation. Advances in Neural Information Processing Systems, 2025, 38: 135766-135804.

[4] Ma, Haokai, et al. Plug-in diffusion model for sequential recommendation. In: Proceedings of the AAAI conference on artificial intelligence. 2024. p. 8886-8894.

The provided response does not clarify directly on the theoretical principles of the proposed method. My concerns are not addressed.

We thank the reviewer for the follow-up and would like to further clarify the theoretical motivation underlying our design.

First, concentrated user intent does not necessarily imply low item-level uncertainty. Even when historical preferences are concentrated, high diffusion noise can make fine-grained item discrimination more uncertain. Semantic-cloud supervision provides broader auxiliary supervision under such uncertainty, while noise-dependent fading progressively emphasizes fine-grained discrimination as SNR increases.

Second, SNR directly characterizes the corruption level of each diffusion sample, whereas training progress does not. Since different noise levels coexist within each training iteration, SNR-dependent supervision adapts supervisory granularity to sample-specific uncertainty rather than the global training stage.

These theoretical considerations motivate our design, while our empirical results on concentrated user intent and noise-dependent supervision further support its effectiveness.

We would appreciate it if the reviewer could clarify which specific theoretical aspects require further justification, so that we can address them directly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.