Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Climber: Toward Efficient Scaling Laws for Large Recommendation Models

T0 review · 4 major / 5 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 Climber claims that a Transformer variant with multi-scale sequence extraction, adaptive temperature attention, and bit-wise gating scales more efficiently than DLRM and vanilla Transformer, with balanced sequence-depth expansion…

desk verdict Worth refereeing: real engineering and a plausible balanced-scaling heuristic, but the compute-equivalence axis is not established. read the letter →

arxiv 2502.09888 v2 pith:PRWN724D submitted 2025-02-14 cs.IR

Songpei Xu , Shijia Wang , Da Guo , Xianwen Guo , Qiang Xiao , Bin Huang , Guanlin Wu , Chuanjiang Luo This is my paper · ORCID
classification cs.IR
keywords recommendersystemsscalinglawsTransformermulti-scalesequenceextractionadaptivetemperatureattentionbit-wisegatingfusionKVcachegenerativerecommendation
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

Climber asks how a recommendation model can be scaled up under the tight latency and compute budgets of production serving. The paper claims that a Transformer variant built around multi-scale sequence extraction, adaptive temperature attention, and bit-wise gating fusion scales more efficiently than both DLRM and a vanilla Transformer: it starts from a higher performance point and has a steeper scaling slope, so the same FLOPs buy more AUC. It further claims that under equal FLOPs the way compute is split between sequence length and layer depth matters, and that alternating expansion of the two dimensions, rather than scaling either alone, gives the best offline and online results. The reported payoff is a 12.19% online metric improvement, up to 14.38x faster inference than DLRM, and the ability to deploy models with 100x more complexity at the same serving cost.

What carries the argument

Three architectural mechanisms carry the argument. Multi-scale Sequence Extraction (MSE) splits the user lifecycle sequence into multiple equal-length subsequences and assigns each to its own Transformer block, cutting the per-block attention complexity from $O(n_s^2 d)$ to $O(n_k^2 d)$ and, under parallel execution, to $O(\max(n_k)^2 d)$. Adaptive Transformer Layer (ATL) replaces the fixed $\sqrt{d_k}$ temperature in softmax attention with a learned per-layer temperature that depends on the extraction strategy and scenario. Bit-wise Gating Fusion (BGF) concatenates block outputs through another ATL and applies a squeeze-and-excitation bit-wise gate to fuse multi-interest representations. On the serving side, the paper reorganizes data from 'single user, single item' into 'single user, multiple items' with full-visible masks over history and diagonal masks between items, and combines encoder-level KV cache, FlashAttention-style tiling, and operator fusion; these are the mechanisms behind the reported 5.15x training throughput and lower inference latency.

What would settle it

Recompute actual or profiler-measured FLOPs for the equal-product pairs in Table 3, including the quadratic attention term and the overheads of multi-scale extraction, and re-rank the configurations on measured compute instead of the linear proxy. If, for example, (1600, 1) is materially more expensive than (200, 8) at the same sequence-length-layer product, the best configuration at each budget would shift and the paper's central allocation rule would fail that test.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that Climber exhibits a more ideal scaling curve than DLRM and vanilla Transformer, and that balanced scaling, alternating sequence-length and depth expansions, yields both offline and online metric gains. The empirical basis is Table 3: at equal FLOPs, the configuration (800 sequence, 8 layers) outperforms both (1600, 4) and (400, 16), and the online A/B test shows a monotone progression from +3.65% at (400, 4) to +12.19% at (800, 12). The paper positions this as the first recommendation model to display both offline and online scaling curves while maintaining resource balance, with acceleration techniques making 100x model-complexity deployment feasible at the serving cost of the DLRM baseline.

Load-bearing premise

The scaling plots and the equivalent-FLOPs comparisons assume that attention's quadratic computational cost is minor, so total compute can be treated as linear in sequence length times layer count; if that cost share is not negligible for the shallow, long-sequence configurations, the balanced-scaling recommendation may not survive a full FLOPs accounting.

Editorial extensions

If this is right

  • At a fixed compute budget, the best configuration is not the largest single dimension but a balanced pair: in Table 3, (800 sequence, 8 layers) outperforms both (1600, 4) and (400, 16) at the same FLOPs.
  • Online deployment can follow the same alternating-expansion policy: the paper reports online metric gains climbing from +3.65% at (400, 4) to +12.19% at (800, 12), consistent with the offline allocation rule.
  • The acceleration techniques make 100x more complex models serveable under the latency budget of the DLRM baseline, so scaling laws can be exploited in production rather than only in offline simulation.
  • Climber's offline gains on three public datasets beyond the music scenario suggest the architecture, not the specific platform, drives the improvement.

Reading between the lines

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

  • Editorial extension: if the balanced-scaling pattern reflects the structure of sparse, multi-source user behavior sequences rather than Climber's specific components, the same alternating expansion rule should transfer to other sequence-based recommenders; running Table 3's equal-FLOPs grid on HSTU and MARM would test this.
  • Editorial extension: the reported +12.19% online gain bundles the new architecture, the batched sample format, and the scaling policy; holding the sample format fixed while scaling only sequence length and depth would separate the scaling-curve effect from data-efficiency gains.
  • Editorial extension: a full FLOPs audit that includes attention's quadratic term and KV-cache memory cost could reorder some configurations at a fixed latency budget, so the balanced-scaling rule should be re-checked against measured serving cost rather than the linear proxy.
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 Climber, a Transformer variant for recommendation systems that combines multi-scale sequence extraction (MSE), adaptive temperature attention (ATL), and bit-wise gating fusion (BGF), together with deployment-oriented acceleration techniques such as 'single user, multiple items' batching and encoder-level KV caching. The central claims are that Climber exhibits a more favorable scaling curve than DLRM and Transformer, that balanced scaling by alternately increasing sequence length and layer number yields both offline and online gains, and that the framework achieves a 12.19% online metric improvement at NetEase Cloud Music. These claims rest on a FLOPs proxy C ∝ s·l, equivalent-FLOPs comparisons in Table 3, and online A/B results in Table 4.

Significance. If the scaling-law claims are correct, the paper provides rare industrial evidence on how to allocate compute between sequence length and depth for recommendation models, with a concrete deployment showing large efficiency gains. The strengths include real online A/B results, a deployed system, acceleration techniques with measurable throughput improvements, and evaluations on multiple public datasets. The paper is among the few to report online scaling curves for recommender systems. However, the credibility of the central scaling conclusion depends on the validity of the compute proxy and on controlled comparisons; these are not yet established, so the current evidence is promising but incomplete.

major comments (4)
  1. [Section 4.3, Eq. C∝s·l] The FLOPs proxy C ∝ s·l is load-bearing for the equivalent-FLOPs comparisons in Table 3 and the online scaling interpretation in Table 4, but it is not justified for this architecture. The paper asserts that attention's quadratic cost is only a minor proportion of overall FLOPs, citing LLM scaling references, yet Section 3.1.2 motivates MSE precisely by reducing O(n_s^2 d) attention cost. Under a more accurate C = α·s·l + β·l·s^2, the configurations in Table 3 are not compute-equivalent: at the 2.55×10^9 row, (1600,4) incurs 10.24M quadratic attention units while (800,8) incurs 5.12M units, a factor of two. The reported preference for (800,8) over (1600,4) could therefore be an artifact of the miscalibrated compute axis rather than a genuine property of balanced scaling. Please report profiler-measured FLOPs for the specific configurations, including an attention vs. non-attention breakdown, and re-run the equivalent-FLOPs analysis with the correct cost model.
  2. [Section 3.1.2 and Figure 3(a)] The scaling comparison between Climber and the baselines confounds architecture with data preprocessing. Climber uses behavior-filtered subsequences with n ≪ n_s, while DLRM and Transformer process raw fixed-length sequences of 2000. The claimed 'higher starting point' of Climber's scaling curve could result from the filtered data providing a higher-quality training signal, not from the model architecture. To support the architectural scaling claim, the authors should include a control where the baselines are trained on the same behavior-filtered sequences, or where Climber is trained on unfiltered sequences.
  3. [Figure 1(b) and Section 1] The 'performance interval' shown in Figure 1(b) is described as simulated, but no simulation procedure, model family, hyperparameters, or data-generating process is specified. This simulation motivates the central claim that factor combinations matter under equivalent FLOPs. Please provide a complete description of the simulation so the result is reproducible, or replace it with real measurements from the industrial dataset.
  4. [Tables 2-4] No error bars, number of independent runs, or significance tests are reported anywhere. The AUC differences in Table 3 that support the balanced-scaling recommendation are small (e.g., 0.8365 vs. 0.8382 and 0.8382 vs. 0.8367), and without variance estimates it is unclear whether the ranking is stable. The online metric in Table 4 is also undefined; please specify the exact business metric and how it was computed. Adding these details is necessary to judge the robustness of the scaling-law conclusions.
minor comments (5)
  1. [Section 4.2.2] The heading 'Abalation Study' should be 'Ablation Study'.
  2. [Section 3.1.3] The phrase 'in mutli-scenario' contains a typo; it should be 'in multi-scenario'.
  3. [Abstract, Section 4.1.2, Section 5] The abstract reports '5.15× throughput gain', Section 4.1.2 says '5.15× training acceleration', and Section 5 states 'sustains 5.15× training acceleration'. Please clarify whether the 5.15× figure applies to training, inference, or both, and align the wording.
  4. [Section 4.3] The sentence 'the FLOPs can be calculated and validated by specific tools of Tensorflow' is vague. Please name the tool, report computed FLOPs for the configurations in Table 3, and state whether the plotted FLOPs are measured or estimated.
  5. [Figure 3(b) and 3(c)] The captions for Figure 3(b) and 3(c) do not specify the fixed sequence length and layer number used when varying the other factor. Please add these details so the scaling curves are interpretable.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the scaling claims are empirical observations on the authors' own architecture, with only minor non-load-bearing self-citations.

full rationale

The paper's load-bearing derivation, if any, is the scaling analysis in Sections 4.3-4.5. Section 4.3 explicitly defines FLOPs as C ∝ s·l and then compares measured AUC across (sequence length, layer number) pairs that share the same product. The conclusion that (800,8) is preferable to (1600,4) at fixed C is an empirical finding about the authors' model family; nothing in that comparison is fitted and then renamed a prediction. The adaptive temperature coefficient and bit-wise gating parameters are learned from data and evaluated through ablations, so they are neither defined in terms of the target metric nor used to manufacture the scaling-law result. The only cited prior work by the same group ([33,34,37]) appears in a literature-review sentence noting that scaling-factor interplay is underexplored in traditional recommender systems; it supplies no equation, no uniqueness argument, and no fitted value, and is therefore not load-bearing. The paper is also benchmarked against external SOTA models on public datasets, so its empirical claims are externally grounded. The main methodological question is the accuracy of the FLOPs proxy: Section 4.3 asserts that quadratic attention cost is a minor proportion and cites external LLM scaling references, while Section 3.1.2 motivates MSE by reducing O(n_s^2 d). If attention cost is actually non-negligible, the 'equivalent FLOPs' rows in Table 3 would be miscalibrated and the balanced-scaling recommendation could be an artifact. That is a measurement/correctness concern, not a circularity: the scaling conclusion is not equivalent to its input by construction, and the contested assumption is imported from external LLM literature rather than from the authors' own prior work. Accordingly, no circular step is identified; the minor self-citations justify a low non-zero score rather than a claim of circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The scaling-law claims rest on learned model components (temperature, bias, gating) and several domain assumptions rather than on derivations from first principles. The paper does not fit power-law exponents, so there are no fitted scaling-law constants; the main free parameters are the learned attention modulators and hand-chosen extraction strategies.

free parameters (3)
  • Adaptive temperature function f_tc(a_k, r) = learned from data per block, layer, and scenario
    Controls the softmax temperature in each adaptive Transformer layer; the central performance claim depends on this learned modulation being better than the fixed 1/sqrt(d_k).
  • Relative attention bias f_p,t_b(a_k, r) = learned
    Adds positional and temporal bias to attention scores per extraction strategy and scenario.
  • Number and type of extraction strategies (N_b and behavior set) = not reported; e.g., click/like/share
    Hand-chosen; determines subsequence lengths and the complexity reduction; not tuned systematically in the paper.
assumptions (4)
  • domain assumption Attention quadratic cost is minor relative to linear cost in large recommendation models (Section 4.3).
    Used to define equivalent FLOPs as C ∝ s*l, which underpins Table 3 and the balanced-scaling conclusion.
  • domain assumption Retaining only positive behaviors in extracted subsequences preserves the signal needed for CTR prediction (Section 3.1.2).
    Justifies the large complexity reduction (n << n_s) and makes Climber's effective sequence length 200 vs 2000 for baselines.
  • domain assumption A single fixed temperature 1/sqrt(d_k) is insufficient and a learned per-scenario/behavior temperature is beneficial (Section 3.1.3).
    Assumption behind ATL; no theoretical derivation is provided.
  • ad hoc to paper The simulated 'performance interval' in Figure 1(b) reflects real model behavior.
    The 'performance interval' concept is introduced via a toy simulation without a generation procedure or real-data validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Climber: Toward Efficient Scaling Laws for Large Recommendation Models." pith.science (2026). https://pith.science/paper/PRWN724D

@misc{pith2026250209888,
  author       = {Pith},
  title        = {Pith review of: Climber: Toward Efficient Scaling Laws for Large Recommendation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PRWN724D}},
  note         = {Machine review of arXiv:2502.09888}
}
abstract

Transformer-based generative models have achieved remarkable success across domains with various scaling law manifestations. However, our extensive experiments reveal persistent challenges when applying Transformer to recommendation systems: (1) Transformer scaling is not ideal with increased computational resources, due to structural incompatibilities with recommendation-specific features such as multi-source data heterogeneity; (2) critical online inference latency constraints (tens of milliseconds) that intensify with longer user behavior sequences and growing computational demands. We propose Climber, an efficient recommendation framework comprising two synergistic components: the model architecture for efficient scaling and the co-designed acceleration techniques. Our proposed model adopts two core innovations: (1) multi-scale sequence extraction that achieves a time complexity reduction by a constant factor, enabling more efficient scaling with sequence length; (2) dynamic temperature modulation adapting attention distributions to the multi-scenario and multi-behavior patterns. Complemented by acceleration techniques, Climber achieves a 5.15$\times$ throughput gain without performance degradation by adopting a "single user, multiple item" batched processing and memory-efficient Key-Value caching. Comprehensive offline experiments on multiple datasets validate that Climber exhibits a more ideal scaling curve. To our knowledge, this is the first publicly documented framework where controlled model scaling drives continuous online metric growth (12.19\% overall lift) without prohibitive resource costs. Climber has been successfully deployed on Netease Cloud Music, one of China's largest music streaming platforms, serving tens of millions of users daily.

Figures

Figures reproduced from arXiv: 2502.09888 by the authors.

Figure 1
Figure 1. (a) Scalability: DLRM vs. Transformer. The left part [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Climber Model Architecture. 3.1 Model Architecture 3.1.1 Overall. To address the computational complexity and scal￾ing challenges in recommendation systems, we propose the model from a recommendation perspective. This model integrates rec￾ommendation characteristics into Transformer architecture with resource-aware scalability. It achieves scaling up from three per￾spectives: multi-scale sequence, multi-scenario, an… view at source ↗
Figure 3
Figure 3. Scaling Curve. (a) Scalability: DLRM vs Transformer vs Climber in Industrial Dataset. (b) Model performance of scaling [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LARES: Latent Reasoning for Sequential Recommendation

    cs.IR 2025-05 conditional novelty 6.0 of 10

    LARES applies depth-recurrent latent reasoning to sequential recommendation, refining all item tokens at each step, and reports consistent gains across four Amazon benchmarks.

  2. Melo: A Production LLM-Powered Music Recommendation Agent

    cs.IR 2026-07 conditional novelty 5.5 of 10

    Production music agent Melo cuts entity misID 7.8 pp and recovers 59% of sparse long-tail sessions via named grounding and reflective retry, with >2 pp retention and >1 min engagement lifts online.

Reference graph

Works this paper leans on

43 extracted references · 10 canonical work pages · cited by 2 Pith papers

  1. [1]

    Newsha Ardalani, Carole-Jean Wu, Zeliang Chen, Bhargav Bhushanam, and Adnan Aziz. 2022. Understanding scaling laws for recommendation models. arXiv preprint arXiv:2208.08489 (2022)

    arXiv 2022

  2. [2]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision- language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)

    arXiv 2023

  3. [3]

    Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. 2024. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954 (2024)

    arXiv 2024

  4. [4]

    Brian Brost, Rishabh Mehrotra, and Tristan Jehan. 2019. The music streaming sessions dataset. In The World Wide Web Conference. 2594–2600

    work page 2019

  5. [5]

    Adam Casson. 2023. Transformer FLOPs. (2023). https://adamcasson.com/posts/ transformer-flops

    work page 2023

  6. [6]

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage interest network for lifelong user behavior modeling in CTR prediction at kuaishou. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3785–3794

    2023

  7. [7]

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev

  8. [8]

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashat- tention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems 35 (2022), 16344–16359

    2022

Show all 43 references
  1. [9]

    Harry Dong, Xinyu Yang, Zhenyu Zhang, Zhangyang Wang, Yuejie Chi, and Beidi Chen. 2024. Get More with LESS: Synthesizing Recurrence with KV Cache Compression for Efficient LLM Inference. arXiv preprint arXiv:2402.09398 (2024)

    2024 arXiv

  2. [10]

    Wei Guo, Hao Wang, Luankang Zhang, Jin Yao Chin, Zhongzhou Liu, Kai Cheng, Qiushi Pan, Yi Quan Lee, Wanqi Xue, Tingjia Shen, et al . 2024. Scal- ing New Frontiers: Insights into Large Recommendation Models. arXiv preprint arXiv:2412.00714 (2024)

    2024 arXiv

  3. [11]

    Yu-Lin He, Xiao-Liang Zhang, Wei Ao, and Joshua Zhexue Huang. 2018. Deter- mining the optimal temperature parameter for Softmax function in reinforcement learning. Applied Soft Computing 70 (2018), 80–85

    2018

  4. [12]

    Geoffrey Hinton. 2015. Distilling the Knowledge in a Neural Network. arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  5. [13]

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. 2022. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022)

    2022 arXiv

  6. [14]

    Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learning vector-quantized item representation for transferable sequential recommenders. In Proceedings of the ACM Web Conference 2023 . 1162–1171

    2023

  7. [15]

    Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards universal sequence representation learning for recommender systems. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Dis- covery and Data Mining . 585–593

    2022

  8. [16]

    Jie Hu, Li Shen, and Gang Sun. 2018. Squeeze-and-excitation networks. InProceed- ings of the IEEE conference on computer vision and pattern recognition . 7132–7141

    2018

  9. [17]

    Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: combining fea- ture importance and bilinear feature interaction for click-through rate prediction. In Proceedings of the 13th ACM conference on recommender systems . 169–177

    2019

  10. [18]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  11. [19]

    Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guoliang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. 2019. Multi-interest network with dynamic routing for recommendation at Tmall. In Proceedings of the 28th ACM international conference on information and kn...

    2019

  12. [20]

    Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian McAuley. 2023. Text is all you need: Learning language representations for sequential recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1258–1267

    2023

  13. [21]

    Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al . 2024. Deepseek- v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434 (2024)

    2024 arXiv

  14. [22]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  15. [23]

    Zichang Liu, Aditya Desai, Fangshuo Liao, Weitao Wang, Victor Xie, Zhaozhuo Xu, Anastasios Kyrillidis, and Anshumali Shrivastava. 2024. Scissorhands: Ex- ploiting the persistence of importance hypothesis for llm kv cache compression at test time. Advances in Neural Information...

    2024

  16. [24]

    Jinwei Luo, Mingkai He, Xiaolin Lin, Weike Pan, and Zhong Ming. 2022. Dual- task learning for multi-behavior sequential recommendation. In Proceedings of the 31st ACM international conference on information & knowledge management . 1379–1388

    2022

  17. [25]

    Xiao Lv, Jiangxia Cao, Shijie Guan, Xiaoyou Zhou, Zhiguang Qi, Yaqiang Zang, Ming Li, Ben Wang, Kun Gai, and Guorui Zhou. 2024. MARM: Unlocking the Future of Recommendation Systems through Memory Augmentation and Scalable Complexity. arXiv preprint arXiv:2411.09425 (2024)

    2024 arXiv

  18. [26]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel

  19. [27]

    Dheevatsa Mudigere, Yuchen Hao, Jianyu Huang, Zhihao Jia, Andrew Tulloch, Srinivas Sridharan, Xing Liu, Mustafa Ozdal, Jade Nie, Jongsoo Park, et al. 2022. Software-hardware co-design for fast and scalable training of deep learning rec- ommendation models. In Proceedings of th...

    2022

  20. [28]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21, 140 (2020), 1–67

    2020

  21. [29]

    Zihua Si, Lin Guan, ZhongXiang Sun, Xiaoxue Zang, Jing Lu, Yiqun Hui, Xingchao Cao, Zeyu Yang, Yichen Zheng, Dewei Leng, et al. 2024. Twin v2: Scaling ultra- long user behavior sequence modeling for enhanced ctr prediction at kuaishou. In Proceedings of the 33rd ACM Internatio...

    2024

  22. [30]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  23. [31]

    Roberto Turrin, Massimo Quadrana, Andrea Condorelli, Roberto Pagano, Paolo Cremonesi, et al. 2015. 30Music Listening and Playlists Dataset. RecSys Posters 75 (2015)

    2015

  24. [32]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  25. [33]

    Shijia Wang, Tianpei Ouyang, Yunfan Zhou, Qiang Xiao, Yintao Ren, Yifei Pan, Fangjian Li, and Chuanjiang Luo. 2025. Enhanced Emotion-aware Music Recom- mendation via Large Language Models. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining ...

    2025

  26. [34]

    Shijia Wang, Yi Zheng, Qiang Xiao, Yilong Zhao, Qimeng Yang, and Chuanjiang Luo. 2024. Sparsity-Aware Personalized Pattern Extractor Network for Music Multi-task Learning. InInternational Conference on Database Systems for Advanced Applications. Springer, 352–363

    2024

  27. [35]

    Yunli Wang, Zixuan Yang, Zhen Zhang, Zhiqiang Wang, Jian Yang, Shiyang Wen, Peng Jiang, and Kun Gai. 2024. Scaling Laws for Online Advertisement Retrieval. arXiv preprint arXiv:2411.13322 (2024)

    2024

  28. [36]

    Zhibo Xiao, Luwei Yang, Wen Jiang, Yi Wei, Yi Hu, and Hao Wang. 2020. Deep multi-interest network for click-through rate prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management . 2265–2268

    2020

  29. [37]

    Qimeng Yang, Shijia Wang, Da Guo, Dongjin Yu, Qiang Xiao, Dongjing Wang, and Chuanjiang Luo. 2024. Cascading Multimodal Feature Enhanced Contrast Learning for Music Recommendation. In 2024 IEEE International Conference on Data Mining (ICDM). IEEE, 905–910

    2024

  30. [38]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152 (2024)

    2024 arXiv

  31. [39]

    Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. 2022. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12104–12113

    2022

  32. [40]

    Buyun Zhang, Liang Luo, Yuxin Chen, Jade Nie, Xi Liu, Daifeng Guo, Yanli Zhao, Shen Li, Yuchen Hao, Yantao Yao, et al. 2024. Wukong: Towards a Scaling Law for Large-Scale Recommendation. arXiv preprint arXiv:2403.02545 (2024)

    2024 arXiv

  33. [41]

    Gaowei Zhang, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, and Ji- Rong Wen. 2024. Scaling law of large sequential recommendation models. In Proceedings of the 18th ACM Conference on Recommender Systems . 444–453

    2024

  34. [2015]

    In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

    Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52

  35. [2023]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2818–2829

Pith tools

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