Pith. sign in

REVIEW 4 major objections 5 minor 81 references

Tackling Device Data Distribution Real-time Shift via Prototype-based Parameter Editing

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Persona claims on-device models can track real-time distribution shift with cloud-generated parameter edits and no on-device backpropagation, reporting gains over fine-tuning, test-time adaptation, DCCL, and DUET on recommendation and visio

desk verdict Multi-prototype parameter editing is a credible practical extension of DUET, but the routing rule is unvalidated and the theory is circular as written. read the letter →

arxiv 2509.06552 v1 pith:T32D4L2T submitted 2025-09-08 cs.LG cs.CVcs.DCcs.IR

classification cs.LGcs.CVcs.DCcs.IR
keywords on-devicemodelgeneralizationreal-timedistributionshiftparametereditingprototypemodelscloud-devicecollaborationbackpropagation-freeadaptationsequentialrecommendationtest-time
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 sets out to prove that on-device models can keep pace with data that drifts in real time without being retrained on the device. Its proposed system, Persona, sends real-time device data to a cloud-side Parameter Editor that outputs a small matrix of weight changes, which is added to the deployed model's adaptive layers. A multi-prototype variant clusters devices into groups, each with its own editor and prototype model, and routes each incoming sample to the group whose generated edit is smallest. On sequential-recommendation datasets and the CK+ vision benchmark, Persona reports better accuracy than fine-tuning, test-time adaptation, DCCL, and DUET, while a single update costs around 10.78 ms, orders of magnitude faster than retraining. The core bet is that a forward pass through a parameter generator is sufficient adaptation; no gradient computation is ever needed on the device.

What carries the argument

The Parameter Editing Matrix is the load-bearing object: a matrix of weight deltas produced by the cloud Parameter Editor from an embedding of real-time data, clipped to a bounded range and added to the frozen weights of an adaptive layer. Prototype models and per-group editors are obtained by clustering these matrices on historical data; the dynamic assignment rule j' = arg min_j {G^(j)} routes a device to the group whose generated edit is smallest; and a cross-layer sequence module constrains all layer embeddings to the same space so different layers agree on which prototype owns a sample.

What would settle it

On a held-out stream of device batches, compare two routing policies: Persona's arg-min-edit rule and an oracle that evaluates each candidate prototype's loss on the batch and picks the best. If the oracle's accuracy advantage is large, the paper's assignment mechanism is not the source of the reported gains. A sharper test is to construct a batch on which the nearest prototype by edit size is known to be the wrong class, and show the misroute degrades accuracy below the single-prototype variant.

Watch

Extended reading notes

Core claim

Real-time distribution shift on a device, the paper claims, can be fixed by a cloud neural Parameter Editor that writes a small clipped weight-delta into the deployed model's adaptive layers, replacing fine-tuning. Shared layers stay frozen; only adaptive weights change, and clipping keeps the edited model near the pretrained one. Persona clusters historical edits into prototype models and routes live samples to the group whose edit is smallest; a sequential layer module keeps assignments consistent. Tests on six recommendation datasets and CK+ put multi-prototype Persona ahead of fine-tuning, TTA, DCCL, and DUET at about 10 ms per update.

Load-bearing premise

The routing rule chooses the prototype whose generated edit is smallest, assuming that a small edit means the prototype already fits the incoming data best; the paper does not define the norm it compares or show that edit size tracks actual group-model performance.

Editorial extensions

If this is right

  • A deployed recommender or vision model can be personalized continuously with only a cloud forward pass per update; no on-device gradients, optimizer state, or labeled data are required.
  • Because the edit is clipped, the adapted model stays within a bounded region of the pretrained weights, which the paper argues yields a tighter generalization bound than unrestricted fine-tuning.
  • Multi-prototype routing improves over a single global editor in the reported experiments, implying that a device population is better served by several local prototype models than by one universal adapter.
  • The same mechanism transfers across modalities: the paper reports gains on both tabular sequential recommendation and image classification with the same design.
  • Per-update latency stays close to 10 ms, comparable to the previous fastest tuning-free baseline and orders of magnitude below fine-tuning, making real-time response feasible.

Reading between the lines

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

  • The dynamic assignment rule treats the magnitude of the generated edit as a proxy for fit; a direct test would compare its routing choices against an oracle that picks the prototype with the best held-out accuracy on the current batch. If small edits do not track fit, the multi-prototype gain could come from group fine-tuning rather than from routing.
  • Because only adaptive-layer weights change, Persona could in principle compose with quantization or pruning on the shared layers, but the paper does not test that interaction; whether the 10 ms latency holds for very large adaptive layers or long sequences is a natural extension.
  • The method assumes the cloud can receive raw real-time device data; in privacy-sensitive settings the upload itself becomes the bottleneck, and a testable variant would generate edits from locally embedded or anonymized features instead.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes Persona, a device-cloud framework for real-time distribution shift on on-device models. A cloud-side parameter editor generates a parameter editing matrix from real-time device data; this matrix is clipped and added to a fixed shared backbone to update only an adaptive classifier layer, avoiding backpropagation on the device. A single-prototype variant uses one global editor; a multi-prototype variant partitions historical devices into groups, trains group-specific editors and prototype models, and at inference routes each device to the group whose generated edit is smallest. Cross-layer knowledge transfer couples the per-layer edit generators. Experiments on four recommendation datasets (Beauty, Electronic, Music, MovieLens) and one vision dataset (CK+) compare against fine-tuning, TTA, DCCL, DUET and report that Persona(m) improves over baselines, with update latency of about 10-19 ms. The paper also claims a generalization-error bound (Lemma 1) based on clipping the edit magnitude.

Significance. If the reported results are reproducible, the framework is a practical and interesting step toward tuning-free on-device personalization: the cloud-side generation is efficient, the multi-prototype idea is natural, and the evaluation spans two modalities. The paper's main empirical claims are plausible, but the theoretical support is not self-contained and the routing rule is under-specified. Because the multi-prototype improvements are the primary source of the 'surpassing SOTA' claim, the routing ambiguity and absence of statistical detail are load-bearing. The paper does not currently ship code or the promised appendix.

major comments (4)
  1. [3.2.2, Eq. (8), Lemma 1] Lemma 1 is asserted with an argument that is circular as written: after defining K and K_ΔW, the text states 'adding parameter constraints reduces model complexity, therefore K > K_ΔW', which is exactly the conclusion needed to show ΔR > 0. The proof is deferred to an appendix that is not present in this arXiv version, and the constants K, K_ΔW, n, m_G are undefined. The displayed formula is also ambiguous: as typeset, ΔR = 2·(K − K_ΔW/(√n√m_G)) does not follow from K > K_ΔW without assumptions on n and m_G. Please supply the full proof, define all constants, and either repair the inequality or remove the theoretical claim from the main text.
  2. [3.3, Eq. (12)] The dynamic assignment rule j' = arg min_j {G^(j)} is not well-defined: each G^(j) is a set of per-layer parameter editing matrices, and the manuscript never specifies the norm, aggregation, or scalarization used to compare them. The training objective (Eq. (7)) and cluster initialization (Eq. (9)) do not calibrate edit magnitude to group-model fit, so a group editor with a small output scale or a prototype near the global initialization can win the arg-min regardless of whether it actually fits the incoming data. Table 5 compares global and group prototype models under the same rule, but it does not validate the routing choice against an oracle (e.g., best group by validation loss) or an alternative routing heuristic. Since the multi-prototype gains in Table 1 are the central empirical claim, please define the comparison norm and add a routing ablation.
  3. [3.2.2, Eq. (7)] The training objective is self-referential: it writes Θ_c := G(x_H; Θ_G^s) + Θ_c, with Θ_c on both sides. It is unclear whether the Θ_c on the right is the fixed pretrained classifier, a previous iterate, or the current value being updated. This circularity makes the described optimization procedure ambiguous and should be corrected. In addition, Eq. (10) uses informal arrows ('Sample → Partition → {M_G^j}') without specifying the fine-tuning loss or whether group editors are trained jointly with the clustering; this is needed to reproduce the method.
  4. [4.2, Tables 1-5] All main tables report point estimates with no standard deviations, confidence intervals, or significance tests. Several improvements over DUET are small (e.g., Table 1, GRU4Rec on Electronic: AUC 0.7895 vs. 0.7886; NDCG@5 0.2976 vs. 0.2976), and Figure 4 is the only place error bars appear, without stating the number of runs. Given the claim of 'significant improvements' and 'surpassing SOTA', please report repeated-run statistics and significance tests for the main comparisons.
minor comments (5)
  1. [4.1] The dataset list includes Amazon CDs and Douban Book, but Tables 1-5 report results only for Beauty, Electronic, Music, and MovieLens. Either provide the corresponding results or explain the omission.
  2. [Figure 1(d)] The figure states 'Time Consumption: 1s (Persona)≪10s (Fine-tuning)', which contradicts Table 1's reported 10.78 ms for Persona and 62.63 s for fine-tuning. Please align the figure with the tabulated numbers.
  3. [Eq. (9)] There is a parenthesis mismatch in the displayed equation, and the clustering algorithm (e.g., k-means, choice of k) is never specified. Please clarify the cluster initialization procedure.
  4. [3.1] The notation involving M_F, M_F^(0), Θ_F, Θ_b, Θ_c is very dense and used inconsistently (for example, Eq. (7) reuses Θ_c on both sides). A notation table or a concrete example would substantially improve readability.
  5. [General] The manuscript repeatedly states that 'Part of the methodology and theoretical proof can be found in the Appendix' and that experimental setup details are in the Appendix, but the arXiv version contains no appendix. The reviewer cannot verify the missing material; please include it in the revision.

Circularity Check

1 steps flagged · score 4.0 of 10

Theoretical generalization-bound argument is circular: Lemma 1 assumes K>K_ΔW (i.e., clipping reduces complexity) to conclude ΔR>0; central experimental claims remain grounded in external datasets.

  1. other [Section 3.2.2, Lemma 1 and Eq. (8)]
    "Adding parameter constraints reduces model complexity, therefore K>KΔW. Since K>KΔW>0, n>0, and mG>0, it follows that ΔR>0. Therefore, our Persona has a tighter generalization error bound."

    The lemma defines ΔR=2(K−KΔW)/(√n √mG) and then asserts K>KΔW because 'adding parameter constraints reduces model complexity.' K and KΔW are introduced only as 'constant terms of model complexity under different conditions' with no independent definition or computation in the main text. The claimed bound reduction ΔR>0 is exactly equivalent to the asserted inequality K>KΔW, so the conclusion is assumed rather than derived from the Lipschitz/clip assumption. As presented, this is a self-supporting theoretical step. The empirical evaluation does not depend on this bound and is validated against external datasets and baselines, so the circularity is partial.

full rationale

The paper's main empirical claim—Persona improves device-model generalization under real-time shift—is tested on public datasets (Beauty, Electronic, Music, MovieLens, CK+) against external baselines, and the time-consumption numbers are direct measurements. That part is self-contained and not circular. The one clear circular step is Lemma 1/Eq. (8): the generalization-error-bound reduction is asserted by assuming K>KΔW, which is precisely the conclusion needed (ΔR>0); no proof of the complexity inequality is supplied in the main text, so the bound reduces to its premise. Self-citations (DUET and other prior works by the same authors) are used as baselines or related work, not as load-bearing justification for the central result, so they do not add circularity. The dynamic-assignment rule Eq. (12) is an unvalidated routing heuristic and a robustness/correctness concern, but not a circularity: no equation shows that the routing decision is forced by the training objective or by definition. Overall score 4 reflects one circular theoretical derivation with independent empirical content.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central empirical method rests on a small number of tunable hyperparameters (T, N_M) and on three domain assumptions: Lipschitz smoothness, K>K_ΔW, and representativeness of historical clusters. No invented physical entities are introduced. The theoretical appendix would be needed to verify whether Lemma 1 is a proof or a restatement.

free parameters (2)
  • clip threshold T = 1.0 (swept over {0.1, 0.5, 1.0, 5.0})
    Controls the range of parameter edits in Eq. (6); Figure 5 shows performance is sensitive to T and T=1.0 is best. No independent selection rule is given in the main text.
  • number of prototypes N_M = chosen from {2, 3, 5, 10} per dataset
    Tables 3-5 show performance varies with N_M and the optimal value changes with dataset; the main experiments do not state how N_M was selected for the headline results.
assumptions (4)
  • domain assumption Assumption 1: the forward function f of the adaptive model is Lipschitz continuous in parameters, satisfying ||f(x; Clip(Δ)+Θ)-f(x;Θ)|| <= p ||Clip(Δ)||.
    Stated in Sec. 3.2.2 and used only to justify Lemma 1; it is not verified for MobileNet, SASRec, or GRU4Rec models.
  • ad hoc to paper The complexity constants K and K_ΔW satisfy K > K_ΔW after clipping.
    Text after Eq. (8) asserts this is because clipping reduces complexity; the inequality is exactly the monotonicity needed to conclude ΔR>0, so the lemma's conclusion is assumed rather than derived in the main text.
  • domain assumption Historical data D_H and its cluster structure remain representative of the real-time distribution shift encountered at inference.
    Multi-prototype training partitions D_H (Eqs. 9-10) and dynamic assignment uses prototypes during inference; if real-time shifts leave the historical cluster support, the assignments and edits have no training coverage.
  • domain assumption CK+ captures on-device data distribution shift for vision.
    Sec. 4.1 justifies dropping CIFAR-10/ImageNet because CK+ 'can reflect changes in user distribution'; no shift statistics or controlled shift analysis are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tackling Device Data Distribution Real-time Shift via Prototype-based Parameter Editing." pith.science (2026). https://pith.science/paper/T32D4L2T

@misc{pith2026250906552,
  author       = {Pith},
  title        = {Pith review of: Tackling Device Data Distribution Real-time Shift via Prototype-based Parameter Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T32D4L2T}},
  note         = {Machine review of arXiv:2509.06552}
}
read the original abstract

The on-device real-time data distribution shift on devices challenges the generalization of lightweight on-device models. This critical issue is often overlooked in current research, which predominantly relies on data-intensive and computationally expensive fine-tuning approaches. To tackle this, we introduce Persona, a novel personalized method using a prototype-based, backpropagation-free parameter editing framework to enhance model generalization without post-deployment retraining. Persona employs a neural adapter in the cloud to generate a parameter editing matrix based on real-time device data. This matrix adeptly adapts on-device models to the prevailing data distributions, efficiently clustering them into prototype models. The prototypes are dynamically refined via the parameter editing matrix, facilitating efficient evolution. Furthermore, the integration of cross-layer knowledge transfer ensures consistent and context-aware multi-layer parameter changes and prototype assignment. Extensive experiments on vision task and recommendation task on multiple datasets confirm Persona's effectiveness and generality.

Figures

Figures reproduced from arXiv: 2509.06552 by the authors.

Figure 1
Figure 1. (a) describes the on-device static model. (b) describes traditional paradigm which can be used to solve the on-device real [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Persona Overview. (a) Prototype Model includes shared layers and adaptive layers. The parameters of the adaptive [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. t-SNE visualization of the data distribution and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance w.r.t. Personalized methods. this table, we have the following findings: (1) Almost all Personal￾ized methods can improve the baseline’s ("-") performance, which reveals the significance of model generalization on the device. (2) In all cases, the effect of…
Figure 5
Figure 5. Figure 5: The impact of the threshold of Parameter Editing Matrix. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 61 canonical work pages

  1. [1]

    Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. 2022. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 18511–18521

  2. [2]

    Xu Chen, Hanxiong Chen, Hongteng Xu, Yongfeng Zhang, Yixin Cao, Zheng Qin, and Hongyuan Zha. 2019. Personalized fashion recommendation with visual explanations based on multimodal attention network: Towards visually explainable recommendation. InProceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval...

  3. [3]

    Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. 2020. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. 11030–11039

  4. [4]

    Yucheng Ding, Chaoyue Niu, Fan Wu, Shaojie Tang, Chengfei Lyu, and Guihai Chen. 2023. DC-CCL: Device-Cloud Collaborative Controlled Learning for Large Vision Models.arXiv preprint arXiv:2303.10361(2023)

  5. [5]

    Tan M Dinh, Anh Tuan Tran, Rang Nguyen, and Binh-Son Hua. 2022. Hyper- inverter: Improving stylegan inversion via hypernetwork. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 11389–11398

  6. [6]

    Kairui Fu, Zheqi Lv, Shengyu Zhang, Fan Wu, and Kun Kuang. 2025. Forward Once for All: Structural Parameterized Adaptation for Efficient Cloud-coordinated On-device Recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 318–329

  7. [7]

    Kairui Fu, Shengyu Zhang, Zheqi Lv, Jingyuan Chen, and Jiwei Li. 2024. DIET: Customized Slimming for Incompatible Networks in Sequential Recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

  8. [9]

    Yulu Gan, Mingjie Pan, Rongyu Zhang, Zijian Ling, Lingran Zhao, Jiaming Liu, and Shanghang Zhang. 2023. Cloud-device collaborative adaptation to contin- ual changing environments in the real-world. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12157–12166

Show all 81 references
  1. [10]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction. International Joint Conference on Artificial Intelligence(2017)

  2. [11]

    Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. 2020. Ghostnet: More features from cheap operations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1580–1589

  3. [12]

    Junjun He, Zhongying Deng, and Yu Qiao. 2019. Dynamic multi-scale filters for semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 3562–3572

  4. [14]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition. 770–778

  5. [15]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639–648

  6. [16]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  7. [17]

    Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingx- ing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. 2019. Searching for mobilenetv3. InProceedings of the IEEE/CVF International Conference on Computer Vision. 1314–1324

  8. [18]

    Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications.CoRR abs/1704.04861 (2017). arXiv:1704.04861 http://arxiv.org/a...

  9. [19]

    Jeremy Howard and Sebastian Ruder. 2018. Universal language model fine-tuning for text classification. InACL 2018-56th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference (Long Papers), Vol. 1. Association for Computational Linguistic...

  10. [20]

    Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. 2016. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size.arXiv preprint arXiv:1602.07360(2016)

  11. [21]

    Wei Ji, Li Li, Zheqi Lv, Wenqiao Zhang, Mengze Li, Zhen Wan, Wenqiang Lei, and Roger Zimmermann. 2025. Backpropagation-free multi-modal on-device model adaptation via cloud-device collaboration.ACM Transactions on Multimedia Computing, Communications and Applications21, 2 (2025), 1–17

  12. [22]

    Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. 2016. Dynamic filter networks.Advances in neural information processing systems29 (2016)

  13. [23]

    Penghao Jiang, Ke Xin, Chunxi Li, and Yinsi Zhou. 2023. High-efficiency Device- Cloud Collaborative Transformer Model. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition. 2203–2209

  14. [24]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE International Conference on Data Mining (ICDM). IEEE, 197–206

  15. [25]

    2023.{RECL}: Responsive{Resource-Efficient} continuous learning for video analytics

    Mehrdad Khani, Ganesh Ananthanarayanan, Kevin Hsieh, Junchen Jiang, Ravi Netravali, Yuanchao Shu, Mohammad Alizadeh, and Victor Bahl. 2023.{RECL}: Responsive{Resource-Efficient} continuous learning for video analytics. In20th USENIX Symposium on Networked Systems Design and Im...

  16. [26]

    Haoxuan Li, Yi Bin, Junrong Liao, Yang Yang, and Heng Tao Shen. 2023. Your negative may not be true negative: Boosting image-text matching with false negative elimination. InProceedings of the 31st ACM international conference on multimedia. 924–934

  17. [27]

    Shawn Li, Huixian Gong, Hao Dong, Tiankai Yang, Zhengzhong Tu, and Yue Zhao

  18. [28]

    Siyuan Liang, Hao Wu, Li Zhen, Qiaozhi Hua, Sahil Garg, Georges Kaddoum, Mohammad Mehedi Hassan, and Keping Yu. 2022. Edge YOLO: Real-time intelli- gent object detection system based on edge-cloud cooperation in autonomous vehicles.IEEE Transactions on Intelligent Transportati...

  19. [29]

    Ji Lin, Ligeng Zhu, Wei-Ming Chen, Wei-Chen Wang, Chuang Gan, and Song Han

  20. [30]

    Kai Liu, Zhihang Fu, Chao Chen, Sheng Jin, Ze Chen, Mingyuan Tao, Rongxin Jiang, and Jieping Ye. 2023. Category-extensible out-of-distribution detection via hierarchical context descriptions.Advances in Neural Information Processing Systems36 (2023), 33241–33261

  21. [31]

    Kai Liu, Zhihang Fu, Sheng Jin, Chao Chen, Ze Chen, Rongxin Jiang, Fan Zhou, Yaowu Chen, and Jieping Ye. 2024. Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

  22. [32]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692 (2019)

  23. [33]

    Chengfei Lv, Chaoyue Niu, Renjie Gu, Xiaotang Jiang, Zhaode Wang, Bin Liu, Ziqi Wu, Qiulin Yao, Congyu Huang, Panos Huang, et al. 2022. Walle: An End- to-End, General-Purpose, and Large-Scale Production System for Device-Cloud Collaborative Machine Learning.arXiv preprint arXi...

  24. [34]

    Zheqi Lv, Shaoxuan He, Tianyu Zhan, Shengyu Zhang, Wenqiao Zhang, Jingyuan Chen, Zhou Zhao, and Fei Wu. 2024. Semantic Codebook Learning for Dynamic Recommendation Models. InProceedings of the 32nd ACM International Conference on Multimedia

  25. [35]

    Zheqi Lv, Tianyu Zhan, Wenjie Wang, Xinyu Lin, Shengyu Zhang, Wenqiao Zhang, Jiwei Li, Kun Kuang, and Fei Wu. 2025. Collaboration of Large Language Models and Small Recommendation Models for Device-Cloud Recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowle...

  26. [36]

    Zheqi Lv, Wenqiao Zhang, Zhengyu Chen, Shengyu Zhang, and Kun Kuang. 2024. Intelligent model update strategy for sequential recommendation. InProceedings of the ACM on Web Conference 2024. 3117–3128

  27. [37]

    Zheqi Lv, Wenqiao Zhang, Shengyu Zhang, Kun Kuang, Feng Wang, Yongwei Wang, Zhengyu Chen, Tao Shen, Hongxia Yang, Beng Chin Ooi, and Fei Wu

  28. [38]

    Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. 2018. Shufflenet v2: Practical guidelines for efficient cnn architecture design. InProceedings of the European conference on computer vision (ECCV). 116–131

  29. [39]

    Ahmed Cherif Mazari, Nesrine Boudoukhani, and Abdelhamid Djeffal. 2024. BERT-based ensemble learning for multi-aspect hate speech detection.Cluster Computing27, 1 (2024), 325–339

  30. [40]

    Christos Profentzas, Magnus Almgren, and Olaf Landsiedel. 2022. MiniLearn: On-Device Learning for Low-Power IoT Devices.. InEWSN. 1–11

  31. [41]

    Xufeng Qian, Yue Xu, Fuyu Lv, Shengyu Zhang, Ziwen Jiang, Qingwen Liu, Xiaoyi Zeng, Tat-Seng Chua, and Fei Wu. 2022. Intelligent Request Strategy Design in Recommender System. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 3772–3782

  32. [42]

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang- Chieh Chen. 2018. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition. 4510–4520. MM ’25, October 27–31, 2025, Dublin, I...

  33. [43]

    Aviv Shamsian, Aviv Navon, Ethan Fetaya, and Gal Chechik. 2021. Personalized federated learning using hypernetworks. InInternational Conference on Machine Learning. PMLR, 9489–9502

  34. [44]

    Karen Simonyan and Andrew Zisserman. 2015. Very Deep Convolutional Net- works for Large-Scale Image Recognition. In3rd International Conference on Learning Representations, ICLR 2015, Yoshua Bengio and Yann LeCun (Eds.)

  35. [45]

    Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. 2019. Pixel-adaptive convolutional neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 11166–11175

  36. [46]

    Jiajie Su, Chaochao Chen, Zibin Lin, Xi Li, Weiming Liu, and Xiaolin Zheng

  37. [47]

    Jiajie Su, Chaochao Chen, Weiming Liu, Fei Wu, Xiaolin Zheng, and Haoming Lyu. 2023. Enhancing hierarchy-aware graph networks with deep dual clustering for session-based recommendation. InProceedings of the ACM web conference

  38. [48]

    Jiajie Su, Qiyong Zhong, Yunshan Ma, Weiming Liu, Chaochao Chen, Xiaolin Zheng, Jianwei Yin, and Tat-Seng Chua. 2025. Distilling Transitional Pattern to Large Language Models for Multimodal Session-based Recommendation.arXiv preprint arXiv:2504.10538(2025)

  39. [49]

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

  40. [50]

    Mingxing Tan and Quoc Le. 2019. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning. PMLR, 6105–6114

  41. [51]

    InProceedings of the 31st ACM international conference on multimedia

    Personalized behavior-aware transformer for multi-behavior sequential recommendation. InProceedings of the 31st ACM international conference on multimedia. 6321–6331

  42. [52]

    Yunze Tong, Junkun Yuan, Min Zhang, Didi Zhu, Keli Zhang, Fei Wu, and Kun Kuang. 2023. Quantitatively Measuring and Contrastively Exploring Heterogene- ity for Domain Generalization. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

  43. [53]

    Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F. Grewe. 2020. Continual learning with hypernetworks. In8th International Con- ference on Learning Representations, ICLR 2020

  44. [54]

    Jiaqi Wang, Kai Chen, Rui Xu, Ziwei Liu, Chen Change Loy, and Dahua Lin. 2019. Carafe: Content-aware reassembly of features. InProceedings of the IEEE/CVF International Conference on Computer Vision. 3007–3016

  45. [55]

    Jiawei Wang, Shaofei Lu, Da Cao, Dongyu Wang, Yuquan Le, Zhe Quan, and Tat-Seng Chua. 2025. Neural Causal Graph for Interpretable and Intervenable Classification. InThe Thirteenth International Conference on Learning Representa- tions

  46. [56]

    Jiawei Wang, Zhanchang Ma, Da Cao, Yuquan Le, Junbin Xiao, and Tat-Seng Chua

  47. [57]

    Zihao Tang, Zheqi Lv, Shengyu Zhang, Fei Wu, and Kun Kuang. 2024. ModelGPT: Unleashing LLM’s Capabilities for Tailored Model Generation.arXiv preprint arXiv:2402.12408(2024)

  48. [58]

    Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. 2020. Solov2: Dynamic and fast instance segmentation.Advances in Neural information pro- cessing systems33 (2020), 17721–17732

  49. [59]

    Tao Wu, Jingyuan Chen, Wang Lin, Mengze Li, Yumeng Zhu, Ang Li, Kun Kuang, and Fei Wu. 2025. Embracing Imperfection: Simulating Students with Diverse Cognitive Levels Using LLM-based Agents.arXiv preprint arXiv:2505.19997 (2025)

  50. [60]

    Zhou Xian, Shamit Lal, Hsiao-Yu Tung, Emmanouil Antonios Platanios, and Katerina Fragkiadaki. 2021. HyperDynamics: Meta-Learning Object and Agent Dynamics with Hypernetworks. In9th International Conference on Learning Representations, ICLR 2021

  51. [61]

    Yikai Yan, Chaoyue Niu, Renjie Gu, Fan Wu, Shaojie Tang, Lifeng Hua, Chengfei Lyu, and Guihai Chen. 2022. On-Device Learning for Model Personalization with Large-Scale Cloud-Coordinated Domain Adaption. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data ...

  52. [62]

    Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. 2019. Condconv: Conditionally parameterized convolutions for efficient inference.Advances in Neural Information Processing Systems32 (2019)

  53. [63]

    InProceedings of the 31st ACM International Conference on Multimedia

    Deconfounded Multimodal Learning for Spatio-temporal Video Grounding. InProceedings of the 31st ACM International Conference on Multimedia. 7521–7529

  54. [64]

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. 2022. Continual test-time domain adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7201–7211

  55. [65]

    Jiangchao Yao, Feng Wang, Xichen Ding, Shaohu Chen, Bo Han, Jingren Zhou, and Hongxia Yang. 2022. Device-cloud Collaborative Recommendation via Meta Controller. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 1...

  56. [66]

    Jiangchao Yao, Feng Wang, Kunyang Jia, Bo Han, Jingren Zhou, and Hongxia Yang

  57. [67]

    Chris Zhang, Mengye Ren, and Raquel Urtasun. 2019. Graph HyperNetworks for Neural Architecture Search. In7th International Conference on Learning Represen- tations, ICLR 2019

  58. [68]

    Jinghao Zhang, Guofan Liu, Qiang Liu, Shu Wu, and Liang Wang. 2024. Modality- Balanced Learning for Multimedia Recommendation. InProceedings of the 32nd ACM International Conference on Multimedia. 7551–7560

  59. [69]

    Jinghao Zhang, Yuting Liu, Wenjie Wang, Qiang Liu, Shu Wu, Liang Wang, and Tat-Seng Chua. 2025. Personalized Text Generation with Contrastive Activation Steering.arXiv preprint arXiv:2503.05213(2025)

  60. [70]

    Linxiao Yang, Yunze Tong, Xinyue Gu, and Liang Sun. 2024. Explain tempo- ral black-box models via functional decomposition. InProceedings of the 41st International Conference on Machine Learning

  61. [71]

    Yuhao Yang, Chao Huang, Lianghao Xia, Yuxuan Liang, Yanwei Yu, and Chenliang Li. 2022. Multi-Behavior Hypergraph-Enhanced Transformer for Sequential Recommendation. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14...

  62. [72]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ...

  63. [73]

    Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. 2021. Decoupled dynamic filter networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6647–6656

  64. [74]

    Yun Zhu, Haizhou Shi, Xiaotang Wang, Yongchao Liu, Yaoke Wang, Boci Peng, Chuntao Hong, and Siliang Tang. 2025. Graphclip: Enhancing transferability in graph foundation models for text-attributed graphs. InProceedings of the ACM on Web Conference 2025. 2183–2197

  65. [78]

    Jinghao Zhang, Yanqiao Zhu, Qiang Liu, Shu Wu, Shuhui Wang, and Liang Wang

  66. [79]

    InProceedings of the 29th ACM international conference on multimedia

    Mining latent structures for multimedia recommendation. InProceedings of the 29th ACM international conference on multimedia. 3872–3880

  67. [80]

    Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. 2018. Shufflenet: An ex- tremely efficient convolutional neural network for mobile devices. InProceedings of the IEEE conference on computer vision and pattern recognition. 6848–6856

  68. [2016]

    Session-based recommendations with recurrent neural networks.Interna- tional Conference on Learning Representations 2016(2016)

  69. [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

  70. [2021]

    InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining

    Device-cloud collaborative learning for recommendation. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 3865–3874

  71. [2022]

    On-device training under 256kb memory.Advances in Neural Information Processing Systems35 (2022), 22941–22954

  72. [2023]

    InProceedings of the ACM Web Conference 2023

    DUET: A Tuning-Free Device-Cloud Collaborative Parameters Generation Framework for Efficient Device Model Generalization. InProceedings of the ACM Web Conference 2023

  73. [2025]

    InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR)

    DPU: Dynamic Prototype Updating for Multimodal Out-of-Distribution De- tection. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR). 10193–10202

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.