Unlocking the Power of Diffusion Models in Sequential Recommendation: A Simple and Effective Approach

Jialei Chen; Yiheng Jiang; Yuanbo Xu

REVIEW 3 major objections 3 minor 36 references

Unlocking the Power of Diffusion Models in Sequential Recommendation: A Simple and Effective Approach

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ADRec claims diffusion recommenders fail via embedding collapse and shows how to stop it.

desk verdict A credible token-level diffusion SR model whose headline numbers need DimeRec added and error bars before I'd trust them. read the letter →

arxiv 2505.19544 v1 pith:W3ZVVMDD submitted 2025-05-26 cs.IR cs.LG

Jialei Chen , Yuanbo Xu , Yiheng Jiang This is my paper · ORCID

classification cs.IRcs.LG

keywords DiffusionmodelSequentialrecommendationEmbeddingcollapseToken-levelAuto-regressiveThree-stagetrainingNext-item

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that diffusion models have underperformed in sequential recommendation not because they are ill-suited, but because existing designs let the item embedding space collapse into a narrow, noise-like shape. To fix this, ADRec runs an independent diffusion process on every token of the shifted target sequence while causal attention supplies clean historical context, and it separates training into embedding pre-training, backbone warm-up, and full fine-tuning. The authors report average relative gains of 15.45% in HR@20 and 13.02% in NDCG@20 over ten baselines on six datasets, with training time cut by 70.98% compared with the strongest diffusion baseline. If correct, this would make diffusion models a practical backbone for next-item recommendation rather than a curiosity that loses to plain transformers.

What carries the argument

The load-bearing object is the token-independent diffusion process: each token $i$ gets its own noise level $t_i$, so the noised sequence is $x^i_{t_i} = \sqrt{\bar{\alpha}^i_{t_i}} x^i_0 + \sqrt{1 - \bar{\alpha}^i_{t_i}} \epsilon$ rather than one shared schedule for the whole sequence. This is what lets the model combine per-token teacher forcing with per-token distribution learning, and it is also what makes the inference shortcut of denoising only the last token legitimate. The second load-bearing piece is the three-stage training schedule, which prevents denoising gradients from crushing randomly initialized embeddings into a uniform Gaussian ball before they have learned any structure.

What would settle it

Train the exact ADRec recipe but replace the random token-level training noise schedule with the fixed inference schedule $t=[0,\ldots,0,T]$, where only the last position is ever noised. If recommendation accuracy stays within noise of the reported numbers, the token-level ELBO is not what carries the result; if it collapses, the inference-time schedule falls outside the training distribution and the claimed alignment fails.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that embedding collapse, not diffusion itself, is why diffusion-based sequential recommenders disappoint. ADRec treats each item token as its own diffusion unit, applying an independent noise level per position in the one-position-offset target sequence, and couples this with autoregressive teacher forcing through a causal attention module that also serves as the denoiser's conditioning. A three-stage schedule first pre-trains the embeddings, then warms up the backbone with embeddings frozen, and only then fine-tunes everything jointly with cross-entropy and mean-squared-error losses. At inference, noise is applied only to the last position while all historical positions stay clean. The paper reports consistent wins across Baby, Beauty, ML-100K, Sports, Toys, and Yelp, plus diagnostic evidence that ADRec's embedding space is more anisotropic, higher-variance, and more linearly separable than those of prior diffusion recommenders.

Load-bearing premise

The load-bearing premise is that applying a different amount of noise to each token remains a correct diffusion training objective when the denoiser is a causal Transformer conditioned on clean embeddings, and that denoising only the last token at inference is the same kind of operation the model was trained for; the paper relies on a prior result for the objective and does not prove the train-inference match itself.

Editorial extensions

If this is right

If ADRec's results are right, diffusion models can be a practical backbone for sequential recommendation rather than a method that consistently loses to transformer baselines.
Per-token denoising with causal attention means every position in the history contributes to distribution learning, which should be more sample-efficient than diffusing only the final item or splitting sequences into many subsequences.
The three-stage recipe (pre-train embeddings, freeze them while warming the denoiser, then unfreeze) is a concrete anti-collapse protocol that other generative recommenders could adopt directly.
Because training complexity stays at transformer level and diffusion adds no per-epoch overhead, the reported 70.98% training-time reduction over the strongest diffusion baseline should transfer to datasets of similar size and sparsity.
The embedding diagnostics used in the paper, such as singular-value spread, isotropy, and linear probing, give a cheap way to detect representation collapse without waiting for full ranking evaluation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A directly testable extension is to apply ADRec's inference schedule, noise only the final position while keeping history clean, to existing sequence-level diffusion recommenders; if their gains approach ADRec's, the clean-context trick matters more than the token-level training objective.
The paper's reading of noise as a soft mask suggests an interpolation spectrum between pure autoregression and full-sequence diffusion, where per-position noise budgets could be tuned as a controllable uncertainty signal for recommendation.
The collapse metrics could serve as a screening test for generative recommenders more broadly: if trained embeddings are statistically indistinguishable from random Gaussian draws, ranking metrics alone may overstate the model's health.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's the paper: ADRec applies token-level diffusion to sequential recommendation, with a three-stage training trick and an inference scheme that only noises the last token. It's a sensible adaptation of Diffusion Forcing [1] to a recommendation setting, and the ablations give a credible story for why each piece matters: remove the warm-up, the pre-training, or the MSE loss and you lose performance. The embedding-collapse evidence (t-SNE, singular value stats, linear probe) is more thorough than most papers in this subfield, and the code is linked. The DiffuRec preprocessing adjustment in Appendix B is honest and the right thing to do.

The soft spots are real. DimeRec is a diffusion-based SR baseline that they discuss in the related work and in Table 1, but it never makes it into Table 2. So the claim to beat 'ten baselines' is against a set that omits the strongest diffusion competitor that they know about. I'd want DimeRec in the comparison before believing the 'best diffusion baseline' improvement numbers. Second, they say five independent runs, but Table 2 has no standard deviations or significance tests. Some relative improvements are small (Yelp HR@20: 7.6%; Sports NDCG@20: 8.8%); those could easily vanish with run-to-run variance. Third, there's a conceptual gap: inference uses t=[0,...,0,T], which is a very specific noise pattern. Training samples each token's noise independently, so that exact pattern is unlikely. The paper cites Diffusion Forcing for the ELBO but does not show that the inference distribution aligns with training. This matters for the claim that the last-token-only denoising is valid. It's not a knockout blow—the ablations suggest the model still works—but it is unproven.

For who: people working on diffusion-based recommenders will want to read it; it's not a conceptual breakthrough, but it's a solid engineering contribution. I'd send it to peer review, but with a request for DimeRec and error bars as the first revision. The current conditional verdict is about right.

Referee Report

3 major / 3 minor

Summary. The paper proposes ADRec, an auto-regressive diffusion model for sequential recommendation. Unlike prior diffusion-based recommenders that apply noise only to the last target item with a sequence-level noise process, ADRec applies an independent per-token diffusion process across the entire one-step-shifted target sequence, uses a causal-attention module as a denoiser, and adopts a three-stage training strategy (embedding pre-training, backbone warm-up with frozen embeddings, and full fine-tuning). At inference, noise is applied only to the last position while the historical sequence remains clean. The authors report consistent improvements over ten baselines on six datasets, with average relative gains of 15.45% in HR@20 and 13.02% in NDCG@20, along with a 70.98% average training-time reduction compared with DiffuRec. The paper also provides qualitative and quantitative evidence for improved embedding structure.

Significance. The paper targets a real and previously identified problem for diffusion-based sequential recommendation: embedding collapse and the mismatch between sequence-level training noise and inference-time needs. The proposed token-level diffusion is a reasonable adaptation of Diffusion Forcing to the recommendation setting, and the three-stage training strategy is simple and appears effective in the reported ablation. The public code, the fair-preprocessing modification of DiffuRec in Appendix B, and the explicit reporting of training-time reductions are strengths. If the empirical claims are confirmed with a complete baseline set and proper statistical reporting, the work would be a useful step toward making diffusion models practical for recommender systems. However, the main evidence is empirical, and the current evaluation has two load-bearing gaps: a closely related diffusion baseline (DimeRec) is omitted, and no standard deviations or significance tests are reported for any of the five-run averages.

major comments (3)

[§4.1.2, Table 2]
[§4.1.4, Table 2]
[§3.2.3, Eq. (11) and §3.5, Eq. (14)]

minor comments (3)

[Appendix B]
[§4.4, Table 4]
[Table 5]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ADRec's performance claims are benchmarked against external baselines and no predicted quantity reduces to its own inputs by construction.

full rationale

ADRec's central claims are empirical and benchmarked against external baselines: the next-item prediction task, HR@20/NDCG@20 on six public datasets, and the training-time comparison against DiffuRec are all measured rather than constructed. The token-level diffusion objective (Eq. 11) is imported from Diffusion Forcing [1], an external prior work whose authors do not overlap with the present paper, so citing its ELBO reweighting is independent support, not a self-citation chain. The hyperparameters lambda = 1e-3 and the 32-step truncated linear schedule are taken from DiffuRec [16] and are not fitted to ADRec's test outcomes; the three-stage training and inference noise pattern (Eq. 14) are design choices validated by ablations (Table 4), not consequences of the loss definition. The train-inference alignment gap noted by the reviewer is an open validity concern, not a circular reduction. The omitted DimeRec baseline and absence of error bars are legitimate evidence-quality concerns but do not make any derivation circular. I find no step in which a predicted quantity is equal by construction to an input, a fitted parameter is renamed a prediction, or a load-bearing premise reduces to a self-citation.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; 'embedding collapse' is a diagnostic label for existing representation behavior, not a postulated entity. The main free parameters are standard hyperparameters or values imported from DifuRec and Diffusion Forcing, and the key mathematical premise is cited rather than re-derived.

free parameters (5)

lambda (noise guidance weight) = 1e-3
Coefficient in feature aggregation z = c + lambda*(x_t + t_emb); value taken from DiffuRec (Section 3.2.2), and Figure 5 shows performance around this region; not fitted to the ADRec objective.
Maximum diffusion steps T = 32
Noise schedule uses T=32 following DiffuRec; Appendix Figure 9 shows insensitivity, so not a critical fit.
Embedding and hidden dimension = 128
Set for all datasets in Section 4.1.4; standard model capacity choice.
Maximum sequence length = 50
Set for all datasets in Section 4.1.4; affects how many tokens are denoised per sequence.
Stage-2 warm-up epochs = 5
Fixed number of epochs for backbone warm-up; no sensitivity analysis is provided.

assumptions (3)

standard math The token-independent diffusion process in Eq. (11) optimizes a reweighted ELBO on ln p_theta((x^i_{t_i})), as established in Chen et al. [1].
ADRec's diffusion training is justified solely by citation to [1], not derived in this paper.
domain assumption Item embeddings form a Euclidean space where additive Gaussian noise and MSE reconstruction are meaningful objectives for item representation learning.
Assumed throughout Section 3; no analysis of the embedding geometry is provided beyond visualization and SVD metrics.
domain assumption Causal attention without positional encoding can recover token positions, so removing positional encoding does not hurt sequence modeling.
Invoked in Section 3.2.1 and Appendix D, supported by reference [8] and internal ablations, not by a general proof.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Unlocking the Power of Diffusion Models in Sequential Recommendation: A Simple and Effective Approach." pith.science (2026). https://pith.science/paper/W3ZVVMDD

@misc{pith2026250519544,
  author       = {Pith},
  title        = {Pith review of: Unlocking the Power of Diffusion Models in Sequential Recommendation: A Simple and Effective Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3ZVVMDD}},
  note         = {Machine review of arXiv:2505.19544}
}

read the original abstract

In this paper, we focus on the often-overlooked issue of embedding collapse in existing diffusion-based sequential recommendation models and propose ADRec, an innovative framework designed to mitigate this problem. Diverging from previous diffusion-based methods, ADRec applies an independent noise process to each token and performs diffusion across the entire target sequence during training. ADRec captures token interdependency through auto-regression while modeling per-token distributions through token-level diffusion. This dual approach enables the model to effectively capture both sequence dynamics and item representations, overcoming the limitations of existing methods. To further mitigate embedding collapse, we propose a three-stage training strategy: (1) pre-training the embedding weights, (2) aligning these weights with the ADRec backbone, and (3) fine-tuning the model. During inference, ADRec applies the denoising process only to the last token, ensuring that the meaningful patterns in historical interactions are preserved. Our comprehensive empirical evaluation across six datasets underscores the effectiveness of ADRec in enhancing both the accuracy and efficiency of diffusion-based sequential recommendation systems.

Figures

Figures reproduced from arXiv: 2505.19544 by the authors.

**Figure 2.** Method overview. The left diagram illustrates that time series and diffusion processes are two orthogonal directions of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The three-stage training strategy and the inference strategy of ADRec. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Comparison of training time between ADRec and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Comparison of ADRec with linear integration using [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 7.** Figure 7: Performance impact of introducing the PCGrad [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 9.** Figure 9: Recall@20 and NDCG@20 at different diffusion [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

**Figure 8.** Figure 8: Diagrams of different auto-regressive strategies and [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 10.** Figure 10: t-SNE results on six datasets [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 15 canonical work pages

[1]

Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. 2024. Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. doi:10.48550/arXiv.2407.01392 arXiv:2407.01392

arXiv doi:10.48550/arxiv.2407.01392 2024
[2]

Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. 2024. Deconstruct- ing denoising diffusion models for self-supervised learning.arXiv preprint arXiv:2401.14404(2024)

arXiv 2024
[3]

Hanwen Du, Huanhuan Yuan, Zhen Huang, Pengpeng Zhao, and Xiao- fang Zhou. 2023. Sequential Recommendation with Diffusion Models. arXiv:2304.04541 [cs.IR] https://arxiv.org/abs/2304.04541

arXiv 2023
[4]

Xinyu Du, Huanhuan Yuan, Pengpeng Zhao, Jianfeng Qu, Fuzhen Zhuang, Guanfeng Liu, and Victor S. Sheng. 2023. Frequency Enhanced Hybrid At- tention Network for Sequential Recommendation. arXiv:2304.09184 [cs.IR] https://arxiv.org/abs/2304.09184

work page Pith review arXiv 2023
[5]

Xinyan Fan, Zheng Liu, Jianxun Lian, Wayne Xin Zhao, Xing Xie, and Ji-Rong Wen. 2021. Lighter and better: low-rank decomposed self-attention networks for next-item recommendation. InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 1733–1737

2021
[6]

Fischer, Vincent Tao Hu, and Bjorn Ommer

Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S. Fischer, Vincent Tao Hu, and Bjorn Ommer. 2024. Diffusion Models and Representation Learning: A Survey. arXiv:2407.00783 [cs.CV] https://arxiv.org/abs/2407.00783

arXiv 2024
[7]

Zhujin Gao, Junliang Guo, Xu Tan, Yongxin Zhu, Fang Zhang, Jiang Bian, and Linli Xu. 2024. Empowering Diffusion Models on the Embedding Space for Text Generation. arXiv:2212.09412 [cs.CL] https://arxiv.org/abs/2212.09412

arXiv 2024
[8]

Adi Haviv, Ori Ram, Ofir Press, Peter Izsak, and Omer Levy. 2022. Transformer language models without positional encodings still learn positional information. arXiv preprint arXiv:2203.16634(2022)

arXiv 2022

Show all 36 references

[9]

Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk
[10]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Proba- bilistic Models. InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

2020
[11]

Yupeng Hou, Binbin Hu, Zhiqiang Zhang, and Wayne Xin Zhao. 2022. CORE: Simple and Effective Session-based Recommendation within Consistent Repre- sentation Space. InSIGIR

2022
[12]

Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

2018
[13]

Anton Klenitskiy and Alexey Vasilev. 2023. Turning Dross Into Gold Loss: is BERT4Rec really better than SASRec?. InProceedings of the 17th ACM Conference on Recommender Systems. 1120–1125

2023
[14]

Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. 2024. Autoregressive Image Generation without Vector Quantization.NeurIPS 2024 (2024)

2024
[15]

Wuchao Li, Rui Huang, Haijun Zhao, Chi Liu, Kai Zheng, Qi Liu, Na Mou, Guorui Zhou, Defu Lian, Yang Song, Wentian Bao, Enyun Yu, and Wenwu Ou. 2024. DimeRec: A Unified Framework for Enhanced Sequential Recommendation via Generative Diffusion Models. doi:10.48550/arXiv.2408.121...

2024 doi arXiv
[16]

Zihao Li, Aixin Sun, and Chenliang Li. 2023. DiffuRec: A Diffusion Model for Sequential Recommendation. doi:10.48550/arXiv.2304.00686 arXiv:2304.00686

2023 doi arXiv
[17]

Jianghao Lin, Jiaqi Liu, Jiachen Zhu, Yunjia Xi, Chengkai Liu, Yangtian Zhang, Yong Yu, and Weinan Zhang. 2024. A Survey on Diffusion Models for Recom- mender Systems. doi:10.48550/arXiv.2409.05033 arXiv:2409.05033 [cs]

2024 doi arXiv
[18]

Yong Niu, Xing Xing, Zhichun Jia, Ruidi Liu, Mindong Xin, and Jianfu Cui. 2024. Diffusion Recommendation with Implicit Sequence Influence. InCompanion Proceedings of the ACM Web Conference 2024(Singapore, Singapore)(WWW ’24). Association for Computing Machinery, New York, NY, ...

2024
[19]

Liwei Pan, Weike Pan, Meiyan Wei, Hongzhi Yin, and Zhong Ming. 2024. A Survey on Sequential Recommendation. arXiv:2412.12770 [cs.IR] https://arxiv. org/abs/2412.12770

2024 arXiv
[20]

Noveen Sachdeva, Giuseppe Manco, Ettore Ritacco, and Vikram Pudi. 2019. Se- quential variational autoencoders for collaborative filtering. InProceedings of the twelfth ACM international conference on web search and data mining. 600–608

2019
[21]

Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2020. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456(2020)

2020 arXiv
[22]

Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang
[23]

Zhen Tian, Wayne Xin Zhao, Changwang Zhang, Xin Zhao, Zhongrui Ma, and Ji- Rong Wen. 2024. EulerFormer: Sequential User Behavior Modeling with Complex Vector Attention. arXiv:2403.17729 [cs.IR] https://arxiv.org/abs/2403.17729

2024 arXiv
[24]

Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, 11 (2008)

2008
[25]

A Vaswani. 2017. Attention is all you need.Advances in Neural Information Processing Systems(2017)

2017
[26]

Shoujin Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z Sheng, and Mehmet Orgun. 2019. Sequential recommender systems: challenges, progress and prospects. In28th International Joint Conference on Artificial Intelligence, IJCAI

2019
[27]

Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua. 2023. Diffusion Recommender Model. doi:10.48550/arXiv.2304.04971 arXiv:2304.04971

2023 doi arXiv
[28]

Yu Wang, Zhiwei Liu, Liangwei Yang, and Philip S. Yu. 2023. Conditional De- noising Diffusion for Sequential Recommendation. arXiv:2304.11433 [cs.LG] https://arxiv.org/abs/2304.11433

2023 arXiv
[29]

International Joint Conferences on Artificial Intelligence, 6332–6338
[30]

Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiangnan He. 2023. Generate What You Prefer: Reshaping Sequential Recom- mendation via Guided Diffusion. doi:10.48550/arXiv.2310.20453 arXiv:2310.20453

2023 doi arXiv
[31]

Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient surgery for multi-task learning.Advances in Neural Information Processing Systems33 (2020), 5824–5836. A Related Works A.1 Traditional Sequential Recommendation Recent sequ...

2020
[32]

Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. 2023. Denoising Diffusion Autoencoders are Unified Self-supervised Learners. InProceedings of the IEEE/CVF International Conference on Computer Vision

2023
[35]

SASRec+ [ 13], an enhancement of SASRec, replaces the original binary cross-entropy loss with cross- entropy loss, significantly improving performance

adopts a multi-layer transformer architecture to efficiently model sequence interactions. SASRec+ [ 13], an enhancement of SASRec, replaces the original binary cross-entropy loss with cross- entropy loss, significantly improving performance. BERT4Rec [22] extends this approach...
[36]

pseudo-patterns,

introduces joint training by applying both denoising loss and Unlocking the Power of Diffusion Models in Sequential Recommendation: A Simple and Effective Approach KDD ’25, August 3–7, 2025, Toronto, ON, Canada 2 4 8 16 32 64 Diffusion steps 7.5 10.0 12.5 15.0 17.5 20.0 22.5HR...

2025
[2016]

arXiv:1511.06939 [cs.LG] https://arxiv.org/abs/1511.06939

Session-based Recommendations with Recurrent Neural Networks. arXiv:1511.06939 [cs.LG] https://arxiv.org/abs/1511.06939

arXiv
[2019]

InProceedings of the 28th ACM international conference on information and knowledge management

BERT4Rec: Sequential recommendation with bidirectional encoder rep- resentations from transformer. InProceedings of the 28th ACM international conference on information and knowledge management. 1441–1450

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. 2024. Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. doi:10.48550/arXiv.2407.01392 arXiv:2407.01392

arXiv doi:10.48550/arxiv.2407.01392 2024

[2] [2]

Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. 2024. Deconstruct- ing denoising diffusion models for self-supervised learning.arXiv preprint arXiv:2401.14404(2024)

arXiv 2024

[3] [3]

Hanwen Du, Huanhuan Yuan, Zhen Huang, Pengpeng Zhao, and Xiao- fang Zhou. 2023. Sequential Recommendation with Diffusion Models. arXiv:2304.04541 [cs.IR] https://arxiv.org/abs/2304.04541

arXiv 2023

[4] [4]

Xinyu Du, Huanhuan Yuan, Pengpeng Zhao, Jianfeng Qu, Fuzhen Zhuang, Guanfeng Liu, and Victor S. Sheng. 2023. Frequency Enhanced Hybrid At- tention Network for Sequential Recommendation. arXiv:2304.09184 [cs.IR] https://arxiv.org/abs/2304.09184

work page Pith review arXiv 2023

[5] [5]

Xinyan Fan, Zheng Liu, Jianxun Lian, Wayne Xin Zhao, Xing Xie, and Ji-Rong Wen. 2021. Lighter and better: low-rank decomposed self-attention networks for next-item recommendation. InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 1733–1737

2021

[6] [6]

Fischer, Vincent Tao Hu, and Bjorn Ommer

Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S. Fischer, Vincent Tao Hu, and Bjorn Ommer. 2024. Diffusion Models and Representation Learning: A Survey. arXiv:2407.00783 [cs.CV] https://arxiv.org/abs/2407.00783

arXiv 2024

[7] [7]

Zhujin Gao, Junliang Guo, Xu Tan, Yongxin Zhu, Fang Zhang, Jiang Bian, and Linli Xu. 2024. Empowering Diffusion Models on the Embedding Space for Text Generation. arXiv:2212.09412 [cs.CL] https://arxiv.org/abs/2212.09412

arXiv 2024

[8] [8]

Adi Haviv, Ori Ram, Ofir Press, Peter Izsak, and Omer Levy. 2022. Transformer language models without positional encodings still learn positional information. arXiv preprint arXiv:2203.16634(2022)

arXiv 2022