Pith. sign in

REVIEW 3 major objections 6 minor 38 references

Device-Cloud Collaborative Correction for On-Device Recommendation

T0 review · 3 major / 6 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 A device-cloud framework called CoCorrRec lets a small recurrent recommender update its own parameters at test time, matching or beating larger Transformer-based recommenders on accuracy while using fewer parameters and fewer FLOPs.

desk verdict A plausible device-cloud recommendation framework with an overstated efficiency claim and a latency analysis that ignores fixed network costs; worth reviewing, but it needs honest revision. read the letter →

arxiv 2506.12687 v1 pith:Y2II6SLD submitted 2025-06-15 cs.IR

Tianyu Zhan , Shengyu Zhang , Zheqi Lv , Jieming Zhu , Jiwei Li , Fan Wu , Fei Wu This is my paper · ORCID
classification cs.IR
keywords device-cloudrecommendationtest-timetrainingsequentialRNNself-correctionnetworkglobalcorrectionon-deviceinferenceefficiencyTTTblock
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 the usual trade-off in on-device recommendation—recurrent networks are cheap but forget, Transformers are accurate but expensive—can be broken by letting a small recurrent model correct its own weights during inference, and by having a cloud network contribute a global correction to stop that self-correction from overfitting to one user's biased history. If this is right, device recommenders can keep the linear cost and small memory of RNN-based models while matching or beating quadratic-cost Transformer baselines on accuracy. The proposed CoCorrRec framework splits the work: a Self-Correction Network (SCN) on the device updates model parameters per input token using self-supervised gradient steps, and a Global Correction Network (GCN) on the cloud ingests the uploaded hidden state and produces corrections that expand the parameter search space and inject collaborative-filtering knowledge. Experiments on three datasets report improvements of 0.06% to 4.56% over the best baselines across UAUC, HitRate@K, and NDCG@K, with the device model staying around 1.95M–2.08M parameters and the cloud correction arriving within the device's tolerance window under idealized bandwidth assumptions.

What carries the argument

The central mechanism is the TTT (test-time training) Block, a sequence-modeling layer whose hidden state is a linear model $W$ updated by a self-supervised gradient-descent step per input token, with mini-batch parallelization and a dual-form computation that lets the whole block run on matrix operations. The update uses three low-rank projections $\theta_Q$, $\theta_K$, and $\theta_V$: $\theta_K$ corrupts the input, $\theta_V$ provides the reconstruction target, and $\theta_Q$ keeps the input aligned with the hidden model's dimension; the output token is produced from the updated $W$. The device-side SCN wraps this block with embedding, RMSNorm, rotary positional encoding (RoPE), layer norm, and a softmax prediction layer, retaining linear-time $O(N \cdot d)$ complexity. The cloud-side GCN is a three-layer MLP that takes the uploaded hidden state $H$ and outputs corrections $K_{attn}$ and $K_b$ to the attention and bias terms of the TTT update, normalized to the statistics of those terms and fused via learnable scalars $P$ and $Q$. Because $H$ is computed before the TTT block, the GCN correction can run concurrently with device inference.

What would settle it

Run CoCorrRec's device-cloud loop over a real 4G or 5G network with a loaded cloud endpoint and measure whether the correction arrives within the reported tolerance window; if the round trip exceeds tolerance on typical conditions, or if removing the GCN drops accuracy below the best baseline on the same data splits, the central balance claim fails.

Watch

Extended reading notes

Core claim

The central claim is that RNN-based sequential recommenders, which are cheap and grow linearly in sequence length, can be made competitive with Transformer-based recommenders on devices if their parameters are updated at test time in a self-supervised way. Concretely, CoCorrRec uses a TTT Block, whose hidden state is a linear model $W$ updated by gradient descent on a reconstruction loss, so the device model adapts to the current user's click sequence instead of relying on fixed pretrained weights. To counter two failure modes of this adaptation—biased local data distributions and reduced search space from mini-batch acceleration—a cloud-side GCN maps the uploaded hidden state $H$ to corrections of the attention and bias terms of the TTT update, normalized and adaptively fused. The paper reports that the resulting model outperforms DIN, GRU4Rec, SASRec, BERT4Rec, and Mamba4Rec on the Beauty, Electronic, and Yelp datasets across UAUC, HitRate@K, and NDCG@K, while staying small and growing FLOPs and memory at a rate comparable to RNN baselines. The implication is that test-time self-correction plus a small cloud-side global correction can replace heavy Transformer architectures for device recommendation.

Load-bearing premise

The real-time feasibility claim assumes that uploading the hidden state, running the cloud correction, and downloading the corrections completes within the device's tolerance window under idealized 4G/5G bandwidth and a fixed 1.18 ms cloud compute time, with no live measurement of network jitter, queueing, or variable cloud load.

Editorial extensions

If this is right

  • If CoCorrRec's results hold, device recommenders can drop Transformer-based architectures for sequential reranking without sacrificing accuracy, because test-time parameter correction closes the gap that previously motivated attention models.
  • The RNN-plus-self-correction design keeps per-token FLOPs and memory growth roughly linear, so longer user sequences cost less than attention-based models at the same accuracy level.
  • The cloud-side GCN provides collaborative-filtering signal to the device model without enlarging the deployed device model's parameter count, since it lives in the cloud and corrects only two terms of the TTT update.
  • The device-cloud split can run concurrently: uploading the hidden state, running the cloud correction, and downloading the corrections completes within tolerance in the paper's bandwidth analysis, so the cloud add-on need not add latency in the idealized setting.

Reading between the lines

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

  • The paper's latency budget rests on idealized bandwidth assumptions; a live-network test with jitter, queueing, and variable cloud load is the natural next step to confirm the real-time feasibility claim beyond the paper's arithmetic.
  • The uploaded hidden state $H$ is a low-dimensional summary but still carries information about user behavior, so a privacy-preserving variant of the GCN channel—such as adding noise or obfuscation before upload—is a plausible extension that the paper does not explore.
  • The same pattern of a self-correcting recurrent core plus a cloud-side global corrector could transfer to other low-resource sequence tasks such as on-device keyboard prediction or small language models, where per-user bias is similarly a concern.
  • Since the GCN corrections are fused by learnable scalars, one could train the fusion weights per user or per domain to make the cloud contribution adaptive; the paper does not investigate this direction.
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

3 major / 6 minor

Summary. The paper proposes CoCorrRec, a device-cloud collaborative framework for on-device sequential recommendation. On the device, a self-correction network (SCN) built around a TTT block updates parameters at test time using a mini-batch parallelization strategy. On the cloud, a global correction network (GCN), implemented as a three-layer MLP, takes the device's hidden state and produces corrections for attention and bias terms, which are adaptively fused into the SCN. The authors report experiments on Beauty, Electronic, and Yelp, claiming that CoCorrRec outperforms RNN-, Transformer-, and SSM-based baselines while having fewer parameters and lower FLOPs. They also provide a resource-consumption analysis (Table 4) and a device-cloud communication delay analysis (Table 5) intended to support real-time feasibility.

Significance. If the performance and efficiency claims were fully supported, CoCorrRec would be a useful contribution to on-device recommendation: it demonstrates that a lightweight RNN-style model with test-time adaptation can close much of the gap with heavier Transformer baselines, and the device-cloud correction idea is timely. The paper includes experiments on three public datasets and releases code. However, the headline efficiency claim is contradicted by the paper's own Table 4 for the full CoCorrRec model, and the real-time latency argument in Section 5.3 omits fixed network overhead, so the central 'balance between real-time performance and high efficiency' claim is not currently established. The empirical performance gains are plausible but are reported without variance or significance testing, and some margins are very small.

major comments (3)
  1. [Abstract and Table 4] The abstract claims that CoCorrRec achieves its results 'with fewer parameters and lower FLOPs' compared to existing Transformer-based and RNN-based device recommendation models. This statement is not supported for the full CoCorrRec model: Table 4 reports 2.08M parameters and 82.7M FLOPs/token for CoCorrRec, versus 1.97M parameters and 30.2M FLOPs/token for GRU4Rec and 2.06M parameters and 68.4M FLOPs/token for Mamba4Rec. Thus the full model has more parameters than GRU4Rec and more than twice its per-token FLOPs. The claim holds only for the CoCorrRec (w/o GCN) variant (1.95M parameters, 17.0M FLOPs/token). Because the paper's central selling point is the device-cloud system, this comparison must be re-scoped and stated precisely, e.g., by separating on-device cost from total cloud-inclusive cost.
  2. [Section 5.3, Table 5] The real-time feasibility conclusion does not follow from the reported calculation. Table 5 computes upload and download delays only as payload size divided by nominal bandwidth (for Beauty: 2.56 KB upload and 4.16 KB download) plus a fixed 1.18 ms cloud compute time, yielding 1.25–2.52 ms, and compares this to a 4.21 ms tolerance. This model omits network round-trip time, connection setup, protocol overhead, jitter, and cloud queueing. For sub-10 KB payloads, fixed RTT dominates: typical 4G/5G RTT is tens of milliseconds, which already exceeds the stated tolerance. The sentence 'the time needed for on-device model to receive corrections consistently remains within the tolerance limit' therefore requires either a live end-to-end measurement or a latency model that includes fixed network delays and tail latency, not just bandwidth-limited transfer time.
  3. [Section 5.4, Table 1] The claim that CoCorrRec 'consistently outperforms' the baselines is not supported with statistical evidence. No error bars, repeated runs, or significance tests are reported, and several margins are very small: for example, Yelp UAUC is 0.9533 for CoCorrRec versus 0.9524 for BERT4Rec, and CoCorrRec (w/o GCN) has NDCG@5 of 0.5372 on Yelp, below BERT4Rec's 0.5388. Since the paper already reports gains as low as 0.06%, the reader cannot determine whether the differences are meaningful. Reporting mean and standard deviation over multiple seeds, or at least a significance test, is necessary to substantiate the 'consistently outperforms' language used in the abstract and Section 5.4.
minor comments (6)
  1. [Throughout] Multiple table cross-references appear as 'Table??' (e.g., in Sections 5.1, 5.2, 5.3, and 5.4), and Table 3 is malformed in the text. These need to be fixed before publication.
  2. [Section 5.2] The text states 'for a Transformer-base (d=64, h=12)' when the experimental setting in Table 3 reports embedding dimension 64 and 4 heads for transformer-based models. The numerical explanation should be made consistent with the actual configuration.
  3. [Section 4.2 and Eq. (14)] The normalization of K_attn and K_b using mean and standard deviation of the corresponding values is described, but it is unclear whether these statistics are computed over the current batch, the training set, or a running estimate. This should be specified for reproducibility.
  4. [Algorithm 1] The algorithm uses the notation 'W0', 'Wb', and 'b0', which is inconsistent with the subscripted notation in Section 3.3 and Eq. (6). Please unify the notation.
  5. [Section 5.3] The statement that 'uploading embeddings for real-time samples enhances user privacy' is not substantiated and seems to conflict with the fact that user hidden states are uploaded to the cloud. Clarify what privacy guarantee is intended.
  6. [Figure 1] Figure 1's 'parameter search space' illustration is informal and not backed by any measurement in the paper. Either add evidence for the reduced search space claim or soften the corresponding discussion in Section 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the TTT component is externally grounded, the GCN is a jointly trained function rather than a renamed fit, and the latency concern is an empirical gap rather than a circular derivation.

full rationale

The paper's derivation chain is self-contained. SCN applies the external TTT update rule (Eqs. 1-5, Sun et al. 2024) with a specified mini-batch dual form, and GCN is defined as a three-layer MLP mapping the uploaded hidden state H to corrections K_attn and K_b (Eq. 13), normalized and adaptively fused (Eqs. 14-15). Both modules are trained jointly on global historical data by cross-entropy (Eq. 16) and then evaluated on real-time samples (Eq. 17); no output quantity is fitted to the same target it claims to predict. Self-citations (e.g., Lv et al. 2023/2024/2025, Fu et al. 2024/2025, Zhan et al. 2025, Li et al. 2024) appear in related work and motivation, but the load-bearing equations and Algorithm 1 are specified in the paper itself and are externally falsifiable on public datasets. The Section 5.3 / Table 5 real-time claim omits fixed network RTT, jitter, and cloud queueing, so the 'within tolerance' conclusion has missing empirical support; that is a correctness risk, not a circularity. No equation reduces to its own input by construction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a hand-set TTT learning rate, the choice of a short sequence length, and several domain assumptions about the effectiveness of self-supervised test-time updates and the safety of uploading hidden states. The GCN introduces learnable parameters (fusion weights P and Q) that are trained rather than ad hoc. No invented entities are introduced.

free parameters (2)
  • TTT learning rate eta = 1
    Set to 1 in Algorithm 1 and used in the dual-form computation (Eq 6-7). No ablation or tuning is reported, so the result depends on this hand-picked value.
  • Sequence length used in experiments = 10
    All models are trained with sequence length 10 (Table 3), far below Yelp's average interaction length of 75.3. Accuracy and efficiency figures are tied to this short length.
assumptions (3)
  • domain assumption The self-supervised reconstruction loss used in TTT (Eq 2) is a useful surrogate for next-item prediction in sequential recommendation.
    The paper adopts TTT-Linear without an analysis linking the reconstruction objective to recommendation quality; the benefit is only shown empirically.
  • domain assumption Uploading hidden state H to the cloud does not violate user privacy.
    Section 5.3 states that uploading embeddings enhances privacy, but the framework uploads H, which encodes user behavior, and no privacy analysis is given.
  • ad hoc to paper A three-layer MLP (GCN) is expressive enough to produce corrections that improve the device model across users.
    No capacity or generalization analysis is provided; the GCN architecture is chosen heuristically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Device-Cloud Collaborative Correction for On-Device Recommendation." pith.science (2026). https://pith.science/paper/Y2II6SLD

@misc{pith2026250612687,
  author       = {Pith},
  title        = {Pith review of: Device-Cloud Collaborative Correction for On-Device Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y2II6SLD}},
  note         = {Machine review of arXiv:2506.12687}
}
read the original abstract

With the rapid development of recommendation models and device computing power, device-based recommendation has become an important research area due to its better real-time performance and privacy protection. Previously, Transformer-based sequential recommendation models have been widely applied in this field because they outperform Recurrent Neural Network (RNN)-based recommendation models in terms of performance. However, as the length of interaction sequences increases, Transformer-based models introduce significantly more space and computational overhead compared to RNN-based models, posing challenges for device-based recommendation. To balance real-time performance and high performance on devices, we propose Device-Cloud \underline{Co}llaborative \underline{Corr}ection Framework for On-Device \underline{Rec}ommendation (CoCorrRec). CoCorrRec uses a self-correction network (SCN) to correct parameters with extremely low time cost. By updating model parameters during testing based on the input token, it achieves performance comparable to current optimal but more complex Transformer-based models. Furthermore, to prevent SCN from overfitting, we design a global correction network (GCN) that processes hidden states uploaded from devices and provides a global correction solution. Extensive experiments on multiple datasets show that CoCorrRec outperforms existing Transformer-based and RNN-based device recommendation models in terms of performance, with fewer parameters and lower FLOPs, thereby achieving a balance between real-time performance and high efficiency.

Figures

Figures reproduced from arXiv: 2506.12687 by the authors.

Figure 1
Figure 1. Parameter space variation in traditional gradient descent [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CoCorrRec. (a) describes the pipeline of CoCorrRec inference, the on-device SCN uploads [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The overview of on-device SCN. projection is performed using θQ, θK, and θV : Xi = θiH i ∈ {Q, K, V } (9) Rotary positional encodings (RoPE) [Su et al., 2024] are then applied to XQ and XK as follows, in which rotary embedding is calculated by XV : Xi = RoP E(Xi , XV ) i ∈ {Q, K} (10) TTT Block employs self-supervised learning to process the input X. When using a mini-batch and dual form, the calcu￾lation of Wb and … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The FLOPs and Max-memory usage of models as the se [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Model performance and inference time with different [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 34 canonical work pages

  1. [1]

    Diet: Customized slimming for incompatible networks in sequential recommendation

    [Fuet al., 2024 ] Kairui Fu, Shengyu Zhang, Zheqi Lv, Jingyuan Chen, and Jiwei Li. Diet: Customized slimming for incompatible networks in sequential recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 816–826,

    work page 2024

  2. [9]

    User-oriented fair- ness in recommendation

    [Liet al., 2021 ] Yunqi Li, Hanxiong Chen, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. User-oriented fair- ness in recommendation. InWWW,

    work page 2021

  3. [11]

    Mergenet: Knowledge migration across heterogeneous models, tasks, and modalities

    [Liet al., 2024 ] Kunxi Li, Tianyu Zhan, Shengyu Zhang, Kun Kuang, Jiwei Li, Zhou Zhao, and Fei Wu. Mergenet: Knowledge migration across heterogeneous models, tasks, and modalities. InAAAI Conference on Artificial Intelli- gence,

    work page 2024

  4. [12]

    K-plet recurrent neural networks for sequen- tial recommendation

    [Linet al., 2018 ] Xiang Lin, Shuzi Niu, Yiqiao Wang, and Yucheng Li. K-plet recurrent neural networks for sequen- tial recommendation. InThe 41st International ACM SI- GIR Conference on Research and Development in Infor- mation Retrieval, SIGIR ’18, pages 1057–1060,

    work page 2018

  5. [13]

    Autoseqrec: Autoencoder for efficient sequential recommendation

    [Liuet al., 2023 ] Sijia Liu, Jiahao Liu, Hansu Gu, Dong- sheng Li, Tun Lu, Peng Zhang, and Ning Gu. Autoseqrec: Autoencoder for efficient sequential recommendation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM ’23, pages 1493–1502,

    work page 2023

  6. [14]

    Mamba4rec: Towards efficient sequential recommendation with selec- tive state space models,

    [Liuet al., 2024 ] Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee. Mamba4rec: Towards efficient sequential recommendation with selec- tive state space models,

    work page 2024

  7. [15]

    Diffusion- based cloud-edge-device collaborative learning for next POI recommendations

    [Longet al., 2024b ] Jing Long, Guanhua Ye, Tong Chen, Yang Wang, Meng Wang, and Hongzhi Yin. Diffusion- based cloud-edge-device collaborative learning for next POI recommendations. InKDD, pages 2026–2036. ACM,

    work page 2026

  8. [16]

    DUET: A tuning-free device-cloud collaborative parame- ters generation framework for efficient device model gen- eralization

    [Lvet al., 2023 ] Zheqi Lv, Wenqiao Zhang, Shengyu Zhang, Kun Kuang, Feng Wang, Yongwei Wang, Zhengyu Chen, Tao Shen, Hongxia Yang, Beng Chin Ooi, and Fei Wu. DUET: A tuning-free device-cloud collaborative parame- ters generation framework for efficient device model gen- eralization. InWWW, pages 3077–3085. ACM,

    work page 2023

Show all 38 references
  1. [17]

    Intelligent model update strategy for sequential recommendation

    [Lvet al., 2024 ] Zheqi Lv, Wenqiao Zhang, Zhengyu Chen, Shengyu Zhang, and Kun Kuang. Intelligent model update strategy for sequential recommendation. InWWW, pages 3117–3128. ACM,

    2024

  2. [18]

    Collaboration of large language models and small recommendation models for device- cloud recommendation

    [Lvet al., 2025 ] Zheqi Lv, Tianyu Zhan, Wenjie Wang, Xinyu Lin, Shengyu Zhang, Wenqiao Zhang, Jiwei Li, Kun Kuang, and Fei Wu. Collaboration of large language models and small recommendation models for device- cloud recommendation. InKDD (1), pages 962–973. ACM,

    2025

  3. [19]

    Fedfast: Going beyond average for faster training of federated rec- ommender systems

    [Muhammadet al., 2020 ] Khalil Muhammad, Qinqin Wang, Diarmuid O’Reilly-Morgan, Elias Tragos, Barry Smyth, Neil Hurley, James Geraci, and Aonghus Lawlor. Fedfast: Going beyond average for faster training of federated rec- ommender systems. InKDD,

    2020

  4. [20]

    Col- laborative sequence prediction for sequential recom- mender

    [Niu and Zhang, 2017] Shuzi Niu and Rongzhi Zhang. Col- laborative sequence prediction for sequential recom- mender. InProceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 2239– 2242,

    2017

  5. [22]

    A systematic review and replicability study of bert4rec for sequential recommendation

    [Petrov and Macdonald, 2022] Aleksandr Petrov and Craig Macdonald. A systematic review and replicability study of bert4rec for sequential recommendation. InProceed- ings of the 16th ACM Conference on Recommender Sys- tems, RecSys ’22, pages 436–447,

    2022

  6. [23]

    Intelligent request strategy design in recommender system

    [Qianet al., 2022 ] Xufeng Qian, Yue Xu, Fuyu Lv, Shengyu Zhang, Ziwen Jiang, Qingwen Liu, Xiaoyi Zeng, Tat-Seng Chua, and Fei Wu. Intelligent request strategy design in recommender system. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,...

    2022

  7. [24]

    Factorizing person- alized markov chains for next-basket recommendation

    [Rendleet al., 2010 ] Steffen Rendle, Christoph Freuden- thaler, and Lars Schmidt-Thieme. Factorizing person- alized markov chains for next-basket recommendation. InProceedings of the 19th International Conference on World Wide Web, TheWebConf ’10, pages 811–820,

    2010

  8. [25]

    Enhancing hierarchy-aware graph networks with deep dual clustering for session-based recommendation

    [Suet al., 2023b ] Jiajie Su, Chaochao Chen, Weiming Liu, Fei Wu, Xiaolin Zheng, and Haoming Lyu. Enhancing hierarchy-aware graph networks with deep dual clustering for session-based recommendation. InProceedings of the ACM web conference 2023, pages 165–176,

    2023

  9. [26]

    Roformer: En- hanced transformer with rotary position embedding.Neu- rocomputing, 568:127063,

    [Suet al., 2024 ] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: En- hanced transformer with rotary position embedding.Neu- rocomputing, 568:127063,

    2024

  10. [27]

    Bert4rec: Se- quential recommendation with bidirectional encoder rep- resentations from transformer

    [Sunet al., 2019 ] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Se- quential recommendation with bidirectional encoder rep- resentations from transformer. InProceedings of the 28th ACM international conference on information and knowl-...

    2019

  11. [29]

    Learning to (learn at test time): Rnns with expressive hidden states,

    [Sunet al., 2024 ] Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xin- lei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, and Carlos Guestrin. Learning to (learn at test time): Rnns with expressive hidden states,

    2024

  12. [30]

    Person- alized top-n sequential recommendation via convolutional sequence embedding

    [Tang and Wang, 2018] Jiaxi Tang and Ke Wang. Person- alized top-n sequential recommendation via convolutional sequence embedding. InProceedings of the 11th ACM In- ternational Conference on Web Search and Data Mining, WSDM ’18, pages 565–573,

    2018

  13. [31]

    Test-time training on video streams.arXiv preprint arXiv:2307.05014,

    [Wanget al., 2023a ] Renhao Wang, Yu Sun, Yossi Gandels- man, Xinlei Chen, Alexei A Efros, and Xiaolong Wang. Test-time training on video streams.arXiv preprint arXiv:2307.05014,

    arXiv

  14. [32]

    On-device learning for model personalization with large-scale cloud-coordinated domain adaption

    [Yanet al., 2022 ] Yikai Yan, Chaoyue Niu, Renjie Gu, Fan Wu, Shaojie Tang, Lifeng Hua, Chengfei Lyu, and Guihai Chen. On-device learning for model personalization with large-scale cloud-coordinated domain adaption. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discov...

    2022

  15. [33]

    Device- cloud collaborative learning for recommendation

    [Yaoet al., 2021 ] Jiangchao Yao, Feng Wang, Kunyang Jia, Bo Han, Jingren Zhou, and Hongxia Yang. Device- cloud collaborative learning for recommendation. InKDD, pages 3865–3874. ACM,

    2021

  16. [34]

    Device-cloud collaborative recommendation via meta con- troller

    [Yaoet al., 2022 ] Jiangchao Yao, Feng Wang, Xichen Ding, Shaohu Chen, Bo Han, Jingren Zhou, and Hongxia Yang. Device-cloud collaborative recommendation via meta con- troller. InKDD, pages 4353–4362. ACM,

    2022

  17. [35]

    Preliminary evaluation of the test-time train- ing layers in recommendation system (student abstract)

    [Zhanet al., 2025 ] Tianyu Zhan, Zheqi Lv, Shengyu Zhang, and Jiwei Li. Preliminary evaluation of the test-time train- ing layers in recommendation system (student abstract). In Proceedings of the AAAI Conference on Artificial Intelli- gence, volume 39, pages 29554–29557,

    2025

  18. [36]

    Devlbert: Learning deconfounded visio-linguistic representations

    [Zhanget al., 2020 ] Shengyu Zhang, Tan Jiang, Tan Wang, Kun Kuang, Zhou Zhao, Jianke Zhu, Jin Yu, Hongxia Yang, and Fei Wu. Devlbert: Learning deconfounded visio-linguistic representations. InProceedings of the 28th ACM International Conference on Multimedia, pages 4373–4382,

    2020

  19. [37]

    Deep interest network for click- through rate prediction

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

    2018

  20. [38]

    Filter-enhanced mlp is all you need for se- quential recommendation

    [Zhouet al., 2022 ] Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. Filter-enhanced mlp is all you need for se- quential recommendation. InProceedings of the ACM Web Conference 2022, TheWebConf ’22, pages 2388–2399, 2022

    2022

  21. [2016]

    Self-attentive sequential recommendation

    [Kang and McAuley, 2018] Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation. In 2018 IEEE International Conference on Data Mining (ICDM), pages 197–206. IEEE,

    2018

  22. [2017]

    Modeling sequences as star graphs to address over-smoothing in self-attentive sequential recommendation.ACM Transactions on Knowledge Discovery from Data,

    [Penget al., 2023 ] Bo Peng, Ziqi Chen, Srinivasan Parthasarathy, and Xia Ning. Modeling sequences as star graphs to address over-smoothing in self-attentive sequential recommendation.ACM Transactions on Knowledge Discovery from Data,

    2023

  23. [2018]

    Enhancing transformers without self-supervised learning: A loss landscape perspective in sequential recommendation

    [Laiet al., 2023 ] Vivian Lai, Huiyuan Chen, Chin- Chia Michael Yeh, Minghua Xu, Yiwei Cai, and Hao Yang. Enhancing transformers without self-supervised learning: A loss landscape perspective in sequential recommendation. InProceedings of the 17th ACM Conference on Recommender...

    2023

  24. [2019]

    Test-time training with self-supervision for generalization under dis- tribution shifts

    [Sunet al., 2020 ] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under dis- tribution shifts. InInternational Conference on Machine Learning, pages 9229–9248. PMLR,

    2020

  25. [2020]

    Test-time training on nearest neighbors for large language models

    [Hardt and Sun, 2023] Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large language models. arXiv preprint arXiv:2305.18466,

    2023 arXiv

  26. [2021]

    Strec: Sparse transformer for sequen- tial recommendations

    [Liet al., 2023 ] Chengxi Li, Yejing Wang, Qidong Liu, Xi- angyu Zhao, Wanyu Wang, Yiqi Wang, Lixin Zou, Wenqi Fan, and Qing Li. Strec: Sparse transformer for sequen- tial recommendations. InProceedings of the 17th ACM Conference on Recommender Systems, RecSys ’23, pages 101–111,

    2023

  27. [2022]

    Edgerec: recommender system on edge in mobile taobao

    [Gonget al., 2020 ] Yu Gong, Ziwen Jiang, Yufei Feng, Bin- bin Hu, Kaiqi Zhao, Qingwen Liu, and Wenwu Ou. Edgerec: recommender system on edge in mobile taobao. InProceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 2477– 2484,

    2020

  28. [2023]

    Session- based recommendations with recurrent neural networks

    [Hidasiet al., 2016 ] Bal´azs Hidasi, Alexandros Karat- zoglou, Linas Baltrunas, and Domonkos Tikk. Session- based recommendations with recurrent neural networks. International Conference on Learning Representations 2016,

    2016

  29. [2024]

    Forward once for all: Structural parameterized adaptation for efficient cloud-coordinated on-device recommendation

    [Fuet al., 2025 ] Kairui Fu, Zheqi Lv, Shengyu Zhang, Fan Wu, and Kun Kuang. 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 Mini...

    2025

  30. [2025]

    [Gandelsmanet al., 2022 ] Yossi Gandelsman, Yu Sun, Xin- lei Chen, and Alexei A

    Association for Computing Machinery. [Gandelsmanet al., 2022 ] Yossi Gandelsman, Yu Sun, Xin- lei Chen, and Alexei A. Efros. Test-time training with masked autoencoders.Advances in Neural Information Processing Systems,

    2022

Pith tools

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