Pith. sign in

REVIEW 4 major objections 4 minor 19 references

This paper claims that interleaving sequence-modeling and feature-interaction layers within repeated blocks—not adding depth or parameters—is what drives Bumblebee's predictive gains.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 08:25 UTC pith:6IEUIPKV

load-bearing objection A plausible, clearly written architecture paper whose central interleaving claim rests on a control that the text itself describes inconsistently. the 4 major comments →

arxiv 2607.24804 v1 pith:6IEUIPKV submitted 2026-07-06 cs.IR cs.LG

Bumblebee: Interleaved Mixed-Layer Building Blocks for Large-Scale Recommendation Systems

David Bauer , Cancan Zhang , Wenshun Liu , Xiaoyi Zhang , Weijia Liu , Wanli Ma , Yue Weng , Wei Li
show 6 more authors
Rui Li Jing Qian Huayu Li Xiaoyi Liu Linhong Zhu Jerry Fu
This is my paper
classification cs.IR cs.LG
keywords recommendation systemsinterleaved architecturesequence modelingfeature interactionresidual connectionsablation studynormalized entropyblock design
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Bumblebee is a recommendation architecture built from repeated 'mixed-layer' blocks, each containing sequence personalization, attention-based encoding, and feature crossing. The paper's central claim is that the interleaved arrangement of these heterogeneous components—not the components themselves, and not added parameters—is the primary driver of its gains over baselines. In a controlled comparison, the same functional components arranged in alternating blocks beat a sequential stack by 0.20% mean normalized entropy (a lower-is-better prediction metric), and a residual-connection effect alone accounts for 0.25% of the gain at zero added parameters. The paper also reports consistent improvements over established baselines across numerous classification and regression tasks on a large industrial dataset. If the arrangement claim holds, it offers a reusable design principle: fuse the sequence and feature tracks early and repeatedly instead of composing deep homogeneous stacks.

Core claim

The paper's central discovery is that interleaving heterogeneous functional units—sequence reweighting, self-attention, filtered target attention, cross-attention, and feature crossing—into repeated self-contained blocks yields better recommendation quality than arranging the same units in sequential deep stacks. Each block outputs a joint representation of sequence and non-sequence features for the next block, and the alternating layout turns ordinary residual connections into cross-module information pathways that carry both modalities at zero added parameters. The composition ablation reports a 0.20% mean normalized entropy improvement over a sequential arrangement of the same components,

What carries the argument

The carrying object is the Bumblebee block: a self-contained micro-pipeline that first reweights the user interaction sequence using user-level context, then encodes it with self-attention and candidate-filtered target attention, and finally fuses sequence and non-sequence representations through pooled cross-attention and feature crossing. Blocks stack vertically, each consuming the previous block's joint representation. The interleaved layout is what changes the residual connections: because attention and feature-crossing alternate within each block, attention residuals and feature-crossing residuals each carry information about both modalities, becoming cross-module pathways without new p

Load-bearing premise

The load-bearing premise is that the sequential comparison model contains exactly the same ingredients as Bumblebee—including the step that re-weights a user's history—and costs the same compute; if it does not, the 0.20% gain attributed to interleaving could instead come from missing ingredients.

What would settle it

Check the configuration listing of the Sequential baseline: Section 4.1.1 names one self-attention layer, four target-attention layers, four cross-attention layers, and four feature-interaction layers but does not mention sequence reweighting, while Section 4.4.2 says the baseline includes it. Rerunning the composition ablation with sequence reweighting explicitly toggled in both arms, with equal parameters and compute, would settle whether the 0.20% NE gap disappears or reverses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If arrangement is the driver, future recommender systems can gain accuracy without more parameters by rearranging existing layers into alternating blocks.
  • The zero-parameter residual gain makes layer ordering a first-class design axis alongside module choice and depth.
  • Configurable block specializations (full, lite, long-sequence) provide a controlled quality-versus-throughput trade-off suitable for production serving.
  • The reported results show consistent NE and RMSE improvements over DLRM-style, SIM, and HSTU baselines across consumption and engagement tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The interleaving principle may transfer to other two-modality architectures: whenever two processing tracks alternate, residual connections can become cross-module pathways for free, so the same block recipe could apply beyond recommender systems.
  • Because the feature-interaction module is pluggable, the arrangement claim predicts the 0.20% interleaving advantage should persist if DHEN is swapped for DCN or DeepFM—a test the paper does not run.
  • The paper's evidence comes from one industrial dataset; rerunning the composition ablation on public data with sequence reweighting explicitly present in both arms would clarify whether the gap is due to layout or to component coverage.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces Bumblebee, a stackable recommendation architecture whose blocks interleave sequence personalization, self/target/cross attention, and feature crossing. It is evaluated on an 81-billion-sample industrial dataset against DLRM, SIM, HSTU, and a parameter-matched 'Sequential' baseline. The central claim is that the interleaved arrangement itself, rather than the component set, drives predictive gains: Table 6 reports a 0.20% mean-NE improvement over the Sequential arrangement, and Table 5 attributes 0.25% NE to residual connections acting as cross-module pathways at zero added parameters.

Significance. If the central claim survives scrutiny, the paper establishes a reusable architectural principle: interleaving heterogeneous functional units can outperform deep homogeneous stacks at fixed parameter count. The evaluation is unusually large in scale (81B samples), and the equal-parameter composition control and component ablation are the right experimental tools for the claim. However, the paper releases neither code nor data, reports no significance information for the headline ablation deltas, and contains a control-description mismatch that directly affects the main inference. These issues must be resolved before the result can be accepted as stated.

major comments (4)
  1. [§4.1.1 vs §4.4.2, Table 6] The composition ablation's control group is not described consistently. §4.1.1 defines Sequential as 'one self attention layer, four layers of target attention, and four layers of cross attention followed by four layers of feature interaction,' with no sequence reweighting; §4.4.2 says Sequential contains 'the same functional components—self attention, target attention, sequence reweighting, and feature crossing.' Table 2 shows dense parameters differ (1.914B vs 1.952B), consistent with a missing module in Sequential. Since Table 5 assigns +0.15% NE to sequence reweighting, the 0.20% interleaving gain could largely reflect component coverage rather than arrangement. Please give the exact Sequential configuration, including residual structure, reconcile the dense-parameter difference, and, if reweighting was absent, rerun the comparison.
  2. [§4.4, Tables 5 and 6] The headline deltas of 0.20% (interleaving) and 0.25% (residual flow) are reported as single point estimates with no variance, error bars, or significance tests. Table 4 reports σ² for c-NE and e-NE, but Tables 5 and 6 provide no such information. Given the paper's central claim is that a 0.20% gap is attributable to arrangement, the authors should report repeated-seed/run variability or a significance test for at least the composition ablation and the residual-flow ablation.
  3. [§4.2 vs §4.4, Tables 4–6] The NE scales are not reconciled. Table 4 reports Bumblebee c-NE = 0.7895 and e-NE = 0.6258, while Tables 5 and 6 report 'Mean NE' of 0.5958 for the same full model. The latter is below both sub-averages, suggesting different task sets, different averaging, or different evaluation windows. Because the percentage deltas in the ablations are relative to these means, the paper must define precisely how 'Mean NE' is computed and how it relates to Table 4.
  4. [§3.7, §4.4.2, §5.2] The relationship between the residual-flow contribution and the interleaving gain needs clarification. §3.7 and Table 5 credit 0.25% NE to 'cross-module residual flow' when residuals are removed from Bumblebee; §4.4.2 attributes the 0.20% NE gap to interleaving. Since residual flow is one of the mechanisms through which interleaving acts, these two numbers are not independent additive contributions. The paper should state whether the 0.25% is already part of the 0.20% arrangement gain or an additional component-level effect, and avoid presenting them in a way that invites double-counting in the Conclusion.
minor comments (4)
  1. [§3.8.3 / Figure 6] The text says the long-sequence block 'drops out self attention mechanisms while keeping filtered and cross attention'; Figure 6(c) caption says it also drops the personalization module. Clarify, since §4.1.1 uses one full + three long-sequence blocks and the presence or absence of reweighting matters for the Table 6 control.
  2. [Table 4(b)] The RMSE table header uses ∆N E; it should be ∆RMSE.
  3. [§4.1.2 / Table 3] The dataset table would benefit from clarifying whether user/item counts are per timestep or cumulative; the current wording 'reflect the amount of training data' is vague.
  4. [Overall] No code or detailed configuration artifacts are released. While industrial constraints make this understandable, a precise configuration table (hidden sizes, number of heads, bucket boundaries, search-feature choices, training hyperparameters) would substantially improve reproducibility.

Circularity Check

0 steps flagged

No circular derivation: interleaving and residual-flow claims are ablation-based measurements; the flagged sequential-baseline mismatch is a control-validity risk, not a circularity.

full rationale

The paper's load-bearing assertions—that interleaving beats a sequential arrangement by 0.20% NE (Table 6, §4.4.2) and that cross-module residual flow contributes 0.25% NE (Table 5, §3.7)—are empirical ablation measurements on a fixed architecture, not quantities derived from fitted parameters or from a cited theorem. The composition ablation explicitly defines the Sequential comparator as containing "the same functional components" (§4.4.2), and the residual-flow number comes from removing residual connections ("w/o Residual Flow", +0.25% NE in Table 5), so neither claim is equivalent to its input by construction. There are self-citations (HSTU/Zhai et al. 2024 is adopted for self-attention; Hou et al. 2026 is cited for sequence personalization; Ding et al. 2026 appears in related work), but none is invoked as a uniqueness theorem or as the reason the interleaving result holds; they supply components or context and are not load-bearing for the headline. The paper itself discloses in §6.2 that the one-full-plus-three-long block composition "was derived empirically and validated with appropriate ablation studies (Section 4.4)," which is model selection, not circularity. The genuine weakness is a control-validity concern: §4.1.1 describes the Sequential baseline as "one self attention layer, four layers of target attention, and four layers of cross attention followed by four layers of feature interaction" without listing sequence reweighting, while §4.4.2 says the baseline includes "sequence reweighting"; if the baseline lacked that module, part of the 0.20% gap could be component coverage rather than arrangement. That is a confound to be clarified, not a reduction of the prediction to its input. Accordingly, no significant circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical entity is introduced. The central claim rests on architecture hyperparameters chosen by hand and on the proprietary dataset; model weights are learned parameters, not ad hoc constants. The most consequential hand-chosen option is the block composition used in the main evaluation.

free parameters (4)
  • Block composition (1 full + 3 long-sequence blocks) = 1 full + 3 long-sequence
    Selected empirically (Section 6.2); not derived. Main NE comparisons use this composition.
  • Shape-bucketing bucket boundaries = unspecified (e.g., 32, 64, 128, ..., 4096)
    Chosen by hand to limit compiled kernels; affects padding and compute.
  • Search features and max-match cap for filtered target attention = unspecified
    Determine which history entries each candidate attends to; no values or sensitivity analysis are reported.
  • PMA seed count K and seed/LCE dimensions = unspecified
    Control the cross-attention bottleneck; not reported or ablated.
axioms (5)
  • standard math Prior module implementations (HSTU attention, DHEN crossing, PMA, residual/SE blocks) are correct and behave as cited.
    The design reuses these published components and assumes combining them does not introduce pathological interactions.
  • domain assumption Evaluation on 72 timesteps of proprietary industrial data with a held-out window is representative of real-world recommendation quality.
    All conclusions rest on the internal dataset in Table 3; no public benchmark or cross-dataset validation is provided.
  • domain assumption Equal parameter counts and matched FLOPs make the Sequential baseline an isolation control for arrangement.
    Section 4.1.1 claims broadly matched parameters and FLOPs, but component counts and sequence processing may still differ; no formal isolation proof is given.
  • ad hoc to paper The fixed configuration of one full block plus three long-sequence blocks adequately represents Bumblebee's design space.
    Chosen empirically (Section 6.2) rather than derived; all headline results use this configuration.
  • domain assumption Aggregating NE across 26 tasks into a single mean is a meaningful way to rank models.
    Tables 5-6 use mean NE; no task-level variance or weighting is reported for the aggregate.

pith-pipeline@v1.3.0-alltime-deepseek · 17486 in / 16588 out tokens · 162262 ms · 2026-08-02T08:25:36.984622+00:00 · methodology

0 comments
read the original abstract

Recommendation systems have undergone significant transformations in the past years. The transition from traditional feature interaction modules to generative next-action prediction has pushed the boundaries of personalized content. Developments have largely evolved along two separate tracks. Sequence modeling approaches on the one hand and feature interaction methods on the other. In this paper, we introduce Bumblebee, a recommendation architecture that addresses the lack of interaction between the two directions through an interleaved, stackable block design. Each block implements a micro-pipeline of layers combining sequence personalization, attention-based encoding, and feature crossing into a self-contained unit. Every block produces a joint representation of both feature modalities which is consumed by the next block in the sequence. This mechanism encourages early and repeated mixture of modalities and enriches downstream features with additional contextual information. Residual connections between blocks create cross-modal information pathways and yield additional predictive performance without adding additional parameters. Blocks can be specialized by selectively dropping components, enabling flexible trade-offs between quality and throughput. We evaluate our approach on large-scale industrial data and show consistent improvements over comparable baseline models across several classification and regression tasks. Furthermore, we conduct ablation studies to confirm that the interleaved composition itself is the primary driver of these improvements. Our results suggest that interleaving heterogeneous functional units, rather than composing deep stacks, is a promising paradigm for future-generation recommendation architectures.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 2 canonical work pages

  1. [5]

    ISBN 9781713871088

    Curran Associates Inc. ISBN 9781713871088. Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment, 2025.https://arxiv. org/abs/2502.18965. Qin Ding, Kevin Course, Linjian Ma, Jianhui Sun, Ruochen Liu, Zhao Zhu, Chun...

    Pith/arXiv arXiv 2025

  2. [9]

    doi: 10.1145/3678004.https://doi.org/10.1145/3678004

    ISSN 1046-8188. doi: 10.1145/3678004.https://doi.org/10.1145/3678004. Dheevatsa Mudigere, Yuchen Hao, Jianyu Huang, Zhihao Jia, Andrew Tulloch, Srinivas Sridharan, Xing Liu, Mustafa Ozdal, Jade Nie, Jongsoo Park, Liang Luo, Jie (Amy) Yang, Leon Gao, Dmytro Ivchenko, Aarti Basant, Yuxi Hu, Jiyan Yang, Ehsan K. Ardestani, Xiaodong Wang, Rakesh Komuravelli, ...

    work page doi:10.1145/3678004.https://doi.org/10.1145/3678004

  3. [10]

    ISBN 9781450386104

    Association for Computing Machinery. ISBN 9781450386104. doi: 10.1145/3470496.3533727.https://doi.org/10.1145/3470496.3533727. Li-Wei Pan, Wei-Ke Pan, Mei-Yan Wei, Hong-Zhi Yin, and Zhong Ming. A survey on sequential recommendation. Frontiers of Computer Science, 20(3):2003606, Oct

    arXiv

  4. [11]

    doi: 10.1007/s11704-025-41329-w

    ISSN 2095-2236. doi: 10.1007/s11704-025-41329-w. https://doi.org/10.1007/s11704-025-41329-w. Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. InProceedings of the 29th ACM International Conference on In...

    doi:10.1007/s11704-025-41329-w 2095

  5. [13]

    ISBN 9781450384469

    Association for Computing Machinery. ISBN 9781450384469. doi: 10.1145/3459637.3481941.https://doi.org/10.1145/3459637.3481941. Kyuyong Shin, Hanock Kwak, Su Young Kim, Max Nihlén Ramström, Jisu Jeong, Jung-Woo Ha, and Kyung-Min Kim. Scaling law for recommendation models: towards general-purpose user representations. InProceedings of the Thirty-Seventh AAA...

    arXiv

  6. [14]

    doi: 10.1609/aaai.v37i4.25582.https: //doi.org/10.1609/aaai.v37i4.25582

    ISBN 978-1-57735-880-0. doi: 10.1609/aaai.v37i4.25582.https: //doi.org/10.1609/aaai.v37i4.25582. Zihua Si, Lin Guan, Zhongxiang Sun, Xiaoxue Zang, Jing Lu, Yiqun Hui, Xingchao Cao, Zeyu Yang, Yichen Zheng, Dewei Leng, Kai Zheng, Chenbin Zhang, Yanan Niu, Yang Song, and Kun Gai. Twin v2: Scaling ultra-long user behavior sequence modeling for enhanced ctr p...

    work page doi:10.1609/aaai.v37i4.25582.https:

  7. [15]

    ISBN 9798400704369

    Association for Computing Machinery. ISBN 9798400704369. doi: 10.1145/3627673.3680030. https://doi.org/10.1145/3627673.3680030. Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. Autoint: Automatic feature interaction learning via self-attentive neural networks. InProceedings of the 28th ACM International Conference...

    arXiv

  8. [17]

    ISBN 9781450351942

    Association for Computing Machinery. ISBN 9781450351942. doi: 10.1145/3124749.3124754.https://doi.org/10.1145/3124749.3124754. Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi. Actions speak louder than words: Trillion-parameter sequential transducers for generative reco...

    arXiv

  9. [18]

    https://proceedings.mlr.press/v235/zhai24a.html. Buyun Zhang, Liang Luo, Xi Liu, Jay Li, Zeliang Chen, Weilin Zhang, Xiaohan Wei, Yuchen Hao, Michael Tsang, Wenjun Wang, Yang Liu, Huayu Li, Yasmine Badr, Jongsoo Park, Jiyan Yang, Dheevatsa Mudigere, and Ellie Wen. Dhen: A deep and hierarchical ensemble network for large-scale click-through rate prediction...

    Pith/arXiv arXiv 2022

  10. [19]

    Ruifeng Zhang, Zexi Huang, Zikai Wang, Ke Sun, Bohang Zheng, Zhen Ouyang, Huimin Xie, Phil Shen, Junlin Zhang, Wentao Guo, and Qinglei Wang

    https://proceedings.mlr.press/v235/zhang24ao.html. Ruifeng Zhang, Zexi Huang, Zikai Wang, Ke Sun, Bohang Zheng, Zhen Ouyang, Huimin Xie, Phil Shen, Junlin Zhang, Wentao Guo, and Qinglei Wang. Zenith: Scaling up ranking models for billion-scale livestreaming recommendation, 2026.https://arxiv.org/abs/2601.21285. 20

    arXiv 2026

  11. [2016]

    ISBN 9781450347952

    Association for Computing Machinery. ISBN 9781450347952. doi: 10.1145/2988450.2988454.https://doi.org/10.1145/2988450.2988454. Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: fast and memory-efficient exact attention with io-awareness. InProceedings of the 36th International Conference on Neural Information Processing...

    arXiv

  12. [2017]

    ISBN 9780999241103. Bojian Hou, Xiaolong Liu, Xiaoyi Liu, Jiaqi Xu, Yasmine Badr, Mengyue Hang, Sudhanshu Chanpuriya, Junqing Zhou, Yuhang Yang, Han Xu, Qiuling Suo, Laming Chen, Yuxi Hu, Jiasheng Zhang, Huaqing Xiong, Yuzhen Huang, Chao Chen, Yue Dong, Yi Yang, Shuo Chang, Xiaorui Gan, Wenlin Chen, Santanu Kolay, Darren Liu, Jade Nie, Chunzhi Yang, Ellie...

    Pith/arXiv arXiv 2026

  13. [2019]

    ISBN 9781450369763

    Association for Computing Machinery. ISBN 9781450369763. doi: 10.1145/3357384.3357925.https://doi.org/10. 1145/3357384.3357925. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lampl...

    arXiv 2017

  14. [2020]

    ISBN 9781450368599

    Association for Computing Machinery. ISBN 9781450368599. doi: 10.1145/3340531.3412744. https://doi.org/10.1145/3340531.3412744. Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, and Xiaoqiang Zhu. One model to serve all: Star topology adaptive recommender for multi-domain ctr p...

    arXiv

  15. [2021]

    ISBN 9781450383325

    Association for Computing Machinery. ISBN 9781450383325. doi: 10.1145/3447548.3467133.https://doi.org/10.1145/3447548.3467133. Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant input. InProceedings of the 36th International Conference on Machine Lea...

    arXiv

  16. [2022]

    ISBN 9781450392365

    Association for Computing Machinery. ISBN 9781450392365. doi: 10.1145/3511808.3557082.https://doi.org/10.1145/3511808. 3557082. Zheng Chai, Qin Ren, Xijun Xiao, Huizhi Yang, Bo Han, Sijun Zhang, Di Chen, Hui Lu, Wenlin Zhao, Lele Yu, Xionghang Xie, Shiru Ren, Xiang Sun, Yaocheng Tan, Peng Xu, Yuchao Zheng, and Di Wu. Longer: Scaling up long sequence model...

    arXiv

  17. [2023]

    ISBN 9798400702419

    Association for Computing Machinery. ISBN 9798400702419. doi: 10.1145/3604915.3608857.https://doi.org/10.1145/3604915.3608857. Yue Cao, Xiaojiang Zhou, Jiaqi Feng, Peihao Huang, Yao Xiao, Dayao Chen, and Sheng Chen. Sampling is all you need on modeling long-term user behaviors for ctr prediction. InProceedings of the 31st ACM International Conference on I...

    arXiv

  18. [2024]

    ISBN 978-3-031-56059-0

    Springer-Verlag. ISBN 978-3-031-56059-0. doi: 10.1007/978-3-031-56060-6_24. https://doi.org/10.1007/978-3-031-56060-6_24. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020.https://arxiv.org/abs/ 2001.08361. Lang Lang, ...

    Pith/arXiv arXiv 2020

  19. [2025]

    ISBN 9798400713644

    Association for Computing Machinery. ISBN 9798400713644. doi: 10.1145/3705328.3748065.https://doi.org/10.1145/3705328.3748065. Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, and Kun Gai. Twin: Two-stage interest network for lifelong user behavior modeling in ctr prediction at kuaishou....

    arXiv 2026