pith. machine review for the scientific record. sign in

arxiv: 2604.18146 · v2 · submitted 2026-04-20 · 💻 cs.IR · cs.AI· cs.CL

Recognition: unknown

Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations

Yunjia Xi , Menghui Zhu , Jianghao Lin , Bo Chen , Ruiming Tang , Yong Yu , Weinan Zhang
Authors on Pith no claims yet

Pith reviewed 2026-05-10 04:06 UTC · model grok-4.3

classification 💻 cs.IR cs.AIcs.CL
keywords modular representation compressionlarge language modelsrecommendation systemsmid-layer representation advantagetask adaptationinformation constraintsrepresentation compressionindustrial deployment
0
0 comments X

The pith

Explicit modularity control in LLMs overcomes final-layer degradation to enable efficient recommendation representations.

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

The paper identifies a Mid-layer Representation Advantage where middle layers of LLMs produce better features for recommendation tasks than the final layer. This stems from spontaneous internal functional modularity that causes the final layer to specialize in the original proxy task. The authors introduce MARC to address this by adding compression and adaptation modules while decoupling tasks through information constraints and distinct network structures. If correct, the approach allows compressed LLM outputs to serve industrial recommendation systems with lower storage and compute costs. Experiments confirm improved efficiency and a 2.82% eCPM lift in a live large-scale advertising test.

Core claim

We identify Mid-layer Representation Advantage (MRA) where mid-layer LLM representations outperform final layers on recommendation tasks. Interpreting this as arising from spontaneous functional modularity that forces final-layer specialization in the proxy training task, we propose MARC. It uses Modular Adjustment to insert explicit compression and task-adaptation modules so the LLM functions strictly as a representation learner, and Modular Task Decoupling to separate tasks via information constraints and different network structures, yielding efficient representations validated in offline tests and an online A/B experiment.

What carries the argument

MARC, the two-part method of Modular Adjustment to insert compression and adaptation modules and Modular Task Decoupling via information constraints plus varied network structures to ground each module to its task.

If this is right

  • LLM representations can be compressed to practical sizes for industrial recommendation systems without sacrificing effectiveness.
  • Recommendation models gain access to higher-quality features by drawing from controlled mid-layer outputs rather than degraded final layers.
  • The same modular approach supports real-time deployment in high-volume commercial scenarios such as search advertising.
  • Existing final-layer compression techniques become more effective once the underlying modularity issue is handled explicitly.

Where Pith is reading between the lines

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

  • The same modular adjustment pattern could be tested on LLMs for other transfer tasks where final layers underperform, such as ranking or retrieval.
  • Training regimes that enforce modularity from the start might reduce the need for post-hoc fixes when adapting LLMs to multiple downstream applications.
  • The decoupling mechanism suggests a route to parameter-efficient adaptation that preserves the model's original capabilities while adding new task-specific paths.

Load-bearing premise

LLMs develop spontaneous internal functional modularity that forces the final layer to specialize in the proxy training task and thereby degrade its representations for recommendation tasks.

What would settle it

A controlled experiment in which final-layer LLM representations after standard compression match or exceed mid-layer performance on recommendation metrics, or an online deployment of MARC that shows no measurable eCPM improvement.

Figures

Figures reproduced from arXiv: 2604.18146 by Bo Chen, Jianghao Lin, Menghui Zhu, Ruiming Tang, Weinan Zhang, Yong Yu, Yunjia Xi.

Figure 1
Figure 1. Figure 1: The impact of various LLM representations on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The impact of different sizes of LLMs on down [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mid-layer representation advantage across different datasets and LLMs. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Layer-wise performance on the proxy target task. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison between existing projection-based com [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study on ML-1M dataset. The results indicate that removing any component leads to a decline in performance. The absence of HSIC loss (w/o HSIC) leads to the largest performance drop for both original and compressed representations, emphasizing its importance for information con￾straints. Removing explicit interactions (w/o EI) or the matching network (w/o MN) has a limited effect on the original r… view at source ↗
Figure 7
Figure 7. Figure 7: Neither switching loss functions nor removing the in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance under different dimensions [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
read the original abstract

Recently, large language models (LLMs) have advanced recommendation systems (RSs), and recent works have begun to explore how to integrate LLMs into industrial RSs. While most approaches deploy LLMs offline to generate and pre-cache augmented representations for RSs, high-dimensional representations from LLMs introduce substantial storage and computational costs. Thus, it is crucial to compress LLM representations effectively. However, we identify a counterintuitive phenomenon during representation compression: Mid-layer Representation Advantage (MRA), where representations from middle layers of LLMs outperform those from final layers in recommendation tasks. This degraded final layer renders existing compression methods, which typically compress on the final layer, suboptimal. We interpret this based on modularity theory that LLMs develop spontaneous internal functional modularity and force the final layer to specialize in the proxy training task. Thus, we propose \underline{M}odul\underline{a}r \underline{R}epresentation \underline{C}ompression (MARC) to explicitly control the modularity of LLMs. First, Modular Adjustment explicitly introduces compression and task adaptation modules, enabling the LLM to operate strictly as a representation-learning module. Next, to ground each module to its specific task, Modular Task Decoupling uses information constraints and different network structures to decouple tasks. Extensive experiments validate that MARC addresses MRA and produces efficient representations. Notably, MARC achieved a 2.82% eCPM lift in an online A/B test within a large-scale commercial search advertising scenario.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper identifies a Mid-layer Representation Advantage (MRA) in LLMs for recommendation tasks, where middle-layer representations outperform final-layer ones. It attributes MRA to spontaneous internal functional modularity that causes the final layer to specialize in the proxy pre-training task. The authors propose Modular Representation Compression (MARC) with two components: Modular Adjustment (inserting explicit compression and task-adaptation modules so the LLM acts as a representation learner) and Modular Task Decoupling (using information constraints and distinct network structures to ground modules to their tasks). Extensive experiments are claimed to validate that MARC addresses MRA and yields efficient representations, with a reported 2.82% eCPM lift in an online A/B test on a large-scale commercial search advertising platform.

Significance. If the MRA phenomenon and its modularity-based explanation hold, and if MARC's gains are causally tied to explicit modularity control rather than added capacity, the work could enable more efficient deployment of LLMs in industrial recommendation systems by reducing high-dimensional representation storage and compute costs while maintaining or improving effectiveness. The online A/B test result provides a concrete indicator of practical impact in a production setting.

major comments (3)
  1. [Abstract / Introduction] The central motivation for MARC rests on the claim (Abstract and Introduction) that MRA arises specifically from spontaneous LLM internal functional modularity forcing final-layer specialization in the proxy task. No layer-probing, mutual-information, or representational similarity analyses are presented to test this mechanism against alternatives such as generic depth effects or optimization dynamics; without such evidence the specific design of Modular Adjustment and Modular Task Decoupling is not guaranteed to be the appropriate remedy.
  2. [Experiments / Online Evaluation] §5 (Experiments) and the online A/B test description report a 2.82% eCPM lift but supply no details on baseline systems, statistical testing, traffic allocation, or controls for confounding factors in the commercial search advertising scenario. This information is load-bearing for evaluating whether MARC's components, rather than extra parameters or training tricks, drive the observed improvement.
  3. [Experiments / Ablation Studies] The ablation studies claimed to validate that MARC addresses MRA lack quantitative isolation of the Modular Adjustment and Decoupling contributions (e.g., performance when only compression is added versus full MARC). If the gains are not shown to exceed those from generic capacity increases, the modularity-control hypothesis remains unconfirmed.
minor comments (2)
  1. [Method] Notation for the Modular Adjustment module and information constraints in the method section could be clarified with explicit equations or pseudocode to facilitate reproduction.
  2. [Related Work] The manuscript should include references to prior work on LLM layer-wise representation analysis and modularity in neural networks to situate the MRA interpretation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below with clarifications on the existing evidence and commitments to strengthen the manuscript through targeted revisions.

read point-by-point responses

  1. Referee: [Abstract / Introduction] The central motivation for MARC rests on the claim (Abstract and Introduction) that MRA arises specifically from spontaneous LLM internal functional modularity forcing final-layer specialization in the proxy task. No layer-probing, mutual-information, or representational similarity analyses are presented to test this mechanism against alternatives such as generic depth effects or optimization dynamics; without such evidence the specific design of Modular Adjustment and Modular Task Decoupling is not guaranteed to be the appropriate remedy.

    Authors: We acknowledge that the current manuscript relies on empirical demonstration of MRA via layer-wise performance curves (Section 4) and the theoretical grounding from modularity literature rather than direct mechanistic probes such as mutual-information estimation or CKA-based representational similarity. The design of MARC is motivated by the consistent observation that final-layer representations underperform middle layers across multiple LLMs and datasets, which existing compression methods fail to address. In revision we will add a dedicated analysis subsection that includes layer-probing results, mutual-information estimates between layer representations and recommendation labels, and similarity metrics comparing middle vs. final layers to distinguish modularity effects from generic depth or optimization artifacts. This will provide stronger support for why the modular adjustment and decoupling components are the appropriate remedy. revision: yes

  2. Referee: [Experiments / Online Evaluation] §5 (Experiments) and the online A/B test description report a 2.82% eCPM lift but supply no details on baseline systems, statistical testing, traffic allocation, or controls for confounding factors in the commercial search advertising scenario. This information is load-bearing for evaluating whether MARC's components, rather than extra parameters or training tricks, drive the observed improvement.

    Authors: We agree that the online evaluation section requires substantially more detail to allow readers to assess causality. In the revised manuscript we will expand Section 5.4 with: explicit descriptions of all baseline systems (production RS, uncompressed LLM representations, and non-modular compression baselines); statistical testing methodology including p-values and confidence intervals for the reported 2.82% eCPM lift; traffic allocation (randomized 50/50 split with user-level randomization); and controls for confounding factors such as time-of-day, campaign effects, and user cohort balancing. These additions will clarify that the observed gains are attributable to MARC rather than incidental capacity or training differences. revision: yes

  3. Referee: [Experiments / Ablation Studies] The ablation studies claimed to validate that MARC addresses MRA lack quantitative isolation of the Modular Adjustment and Decoupling contributions (e.g., performance when only compression is added versus full MARC). If the gains are not shown to exceed those from generic capacity increases, the modularity-control hypothesis remains unconfirmed.

    Authors: The existing ablation studies (Section 5.2 and associated tables) compare full MARC against several variants and show consistent gains, but we recognize they do not fully isolate the individual contributions or rule out generic capacity effects. In revision we will augment the ablation suite with: (1) a compression-only variant (Modular Adjustment without task adaptation), (2) a capacity-matched baseline that adds equivalent parameters without the modular structure or information constraints, and (3) quantitative metrics tracking how each component reduces the MRA gap. These new results will be presented alongside the current tables to provide clearer isolation of the modularity-control mechanism. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical observation drives method, validated externally

full rationale

The paper first reports an empirical observation (MRA: middle layers outperforming final layers on recommendation tasks), offers an interpretation drawn from modularity theory without reducing it to a fitted parameter or self-citation chain, then introduces MARC components (Modular Adjustment and Modular Task Decoupling) as an explicit intervention. Validation rests on separate experiments and a live A/B test reporting a 2.82% eCPM lift, none of which are shown to be tautological with the input data or definitions. No equations, predictions, or uniqueness claims collapse back to the inputs by construction, satisfying the requirement for independent content against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The central claim rests on the modularity theory as an explanatory axiom and introduces new conceptual entities (MRA and MARC modules) whose primary support comes from the paper's own experiments rather than independent external evidence.

axioms (1)
  • domain assumption LLMs develop spontaneous internal functional modularity
    Used to interpret the cause of the mid-layer representation advantage and the suboptimality of final-layer representations.
invented entities (3)
  • Mid-layer Representation Advantage (MRA) no independent evidence
    purpose: To name and explain the observed superiority of mid-layer LLM representations for recommendation tasks
    Introduced as a counterintuitive phenomenon identified during compression experiments.
  • Modular Adjustment module no independent evidence
    purpose: To explicitly introduce compression and task adaptation capabilities into the LLM
    Core component of the proposed MARC method.
  • Modular Task Decoupling no independent evidence
    purpose: To separate tasks via information constraints and differing network structures
    Second core component of MARC to ground each module to its task.

pith-pipeline@v0.9.0 · 5586 in / 1433 out tokens · 41541 ms · 2026-05-10T04:06:13.603347+00:00 · methodology

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

84 extracted references · 35 canonical work pages · 5 internal anchors

  1. [1]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He

  2. [2]

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation.arXiv preprint arXiv:2305.00447(2023)

    work page arXiv 2023

  3. [3]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders.arXiv preprint arXiv:2404.05961(2024)

    work page arXiv 2024

  4. [4]

    Jin Chen, Zheng Liu, Xu Huang, Chenwang Wu, Qi Liu, Gangwei Jiang, Yuanhao Pu, Yuxuan Lei, Xiaolong Chen, Xingmei Wang, et al. 2023. When large language models meet personalization: Perspectives of challenges and opportunities.arXiv preprint arXiv:2307.16376(2023)

    work page arXiv 2023

  5. [5]

    Qian Dong, Yiding Liu, Qingyao Ai, Zhijing Wu, Haitao Li, Yiqun Liu, Shuaiqiang Wang, Dawei Yin, and Shaoping Ma. 2024. Unsupervised large language model Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia alignment for information retrieval via contrastive fe...

    2024

  6. [6]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    Wenqi Fan, Zihuai Zhao, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Jiliang Tang, and Qing Li. 2023. Recommender systems in the era of large language models (llms).arXiv preprint arXiv:2307.02046(2023)

    work page arXiv 2023

  8. [8]

    Binzong Geng, Zhaoxin Huan, Xiaolu Zhang, Yong He, Liang Zhang, Fajie Yuan, Jun Zhou, and Linjian Mo. 2024. Breaking the length barrier: Llm-enhanced CTR prediction in long textual user behaviors. InSIGIR’24. 2311–2315

    2024

  9. [9]

    Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. 2005. Measuring statistical dependence with Hilbert-Schmidt norms. InInternational conference on algorithmic learning theory. Springer, 63–77

    2005

  10. [10]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction.arXiv preprint arXiv:1703.04247(2017)

    work page arXiv 2017

  11. [11]

    R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. 2018. Learning deep represen- tations by mutual information estimation and maximization.arXiv preprint arXiv:1808.06670(2018)

    work page Pith review arXiv 2018

  12. [12]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685(2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  13. [13]

    Jun Hu, Wenwen Xia, Xiaolu Zhang, Chilin Fu, Weichang Wu, Zhaoxin Huan, Ang Li, Zuoli Tang, and Jun Zhou. 2024. Enhancing sequential recommendation via llm-based semantic embedding learning. InCompanion Proceedings of the ACM on Web Conference 2024. 103–111

    2024

  14. [14]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. InProceedings of the 22nd ACM international conference on Information & Knowledge Management. 2333–2338

    2013

  15. [15]

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

    2019

  16. [16]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. 2024. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186(2024)

    work page internal anchor Pith review arXiv 2024

  17. [17]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated Gain-Based Evaluation of IR Techniques.ACM Trans. Inf. Syst.(2002), 422–446

    2002

  18. [18]

    Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al . 2023. Phi-2: The surprising power of small language models.Microsoft Research Blog1, 3 (2023), 3

    2023

  19. [19]

    Jian Jia, Yipei Wang, Yan Li, Honggang Chen, Xuehan Bai, Zhaocheng Liu, Jian Liang, Quan Chen, Han Li, Peng Jiang, et al. 2024. Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application. arXiv preprint arXiv:2405.03988(2024)

    work page arXiv 2024

  20. [20]

    Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang

  21. [21]

    Scaling sentence embeddings with large language models.arXiv preprint arXiv:2307.16645(2023)

    work page arXiv 2023

  22. [22]

    Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. 2019. Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351(2019)

    work page arXiv 2019

  23. [23]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, Vol. 1. Minneapolis, Minnesota

    2019

  24. [24]

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. 2022. Matryoshka representation learning.Advances in Neural Information Processing Systems35 (2022), 30233–30249

    2022

  25. [25]

    Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, et al. 2024. Gecko: Versatile text embeddings distilled from large language models.arXiv preprint arXiv:2403.20327 (2024)

    work page arXiv 2024

  26. [26]

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2023. Large Language Models for Generative Recommendation: A Survey and Visionary Discussions.arXiv preprint arXiv:2309.01157(2023)

    work page arXiv 2023

  27. [27]

    Xianming Li and Jing Li. 2024. AoE: Angle-optimized embeddings for semantic textual similarity. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1825–1839

    2024

  28. [28]

    Xianming Li and Jing Li. 2024. BeLLM: Backward Dependency Enhanced Large Language Model for Sentence Embeddings. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 792–804

    2024

  29. [29]

    Zekun Li, Zeyu Cui, Shu Wu, Xiaoyu Zhang, and Liang Wang. 2019. Fi-gnn: Modeling feature interactions via graph neural networks for ctr prediction. In CIKM’19. 539–548

    2019

  30. [30]

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1754–1763

    2018

  31. [31]

    Jianghao Lin, Bo Chen, Hangyu Wang, Yunjia Xi, Yanru Qu, Xinyi Dai, Kangning Zhang, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. ClickPrompt: CTR models are strong prompt generators for adapting language models to CTR prediction. InProceedings of the ACM Web Conference 2024. 3319–3330

    2024

  32. [32]

    Jianghao Lin, Xinyi Dai, Rong Shan, Bo Chen, Ruiming Tang, Yong Yu, and Weinan Zhang. 2025. Large language models make sample-efficient recommender systems.Frontiers of Computer Science19, 4 (2025), 194328

    2025

  33. [33]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al . 2025. How can recommender systems benefit from large language models: A survey.ACM Transactions on Information Systems43, 2 (2025), 1–47

    2025

  34. [34]

    Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recom- mendation. InProceedings of the ACM on Web Conference 2024. 3497–3508

    2024

  35. [35]

    Chengkai Liu, Jianghao Lin, Hanzhou Liu, Jianling Wang, and James Caverlee

  36. [36]

    InProceedings of the 33rd ACM international conference on information and knowledge management

    Behavior-dependent linear recurrent units for efficient sequential recom- mendation. InProceedings of the 33rd ACM international conference on information and knowledge management. 1430–1440

  37. [37]

    Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee

  38. [38]

    Mamba4rec: Towards efficient sequential recommendation with selective state space models.arXiv preprint arXiv:2403.03900(2024)

    work page arXiv 2024

  39. [39]

    Peng Liu, Lemei Zhang, and Jon Atle Gulla. 2023. Pre-train, prompt and recom- mendation: A comprehensive survey of language modelling paradigm adaptations in recommender systems.arXiv preprint arXiv:2302.03735(2023)

    work page arXiv 2023

  40. [40]

    Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024. Once: Boosting content-based recommendation with both open-and closed-source large language models. InWSDM’24. 452–461

    2024

  41. [41]

    Weiqing Luo, Chonggang Song, Lingling Yi, and Gong Cheng. 2024. KELLM- Rec: Knowledge-Enhanced Large Language Models for Recommendation.arXiv preprint arXiv:2403.06642(2024)

    work page arXiv 2024

  42. [42]

    Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen, Christopher Leung, Jiajie Tang, and Jiebo Luo. 2023. Llm-rec: Person- alized recommendation via prompting large language models.arXiv preprint arXiv:2307.15780(2023)

    work page arXiv 2023

  43. [43]

    Wan-Duo Kurt Ma, JP Lewis, and W Bastiaan Kleijn. 2020. The HSIC bottleneck: Deep learning without back-propagation. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 5085–5092

    2020

  44. [44]

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2024. Fine- tuning llama for multi-stage text retrieval. InSIGIR’24. 2421–2425

    2024

  45. [45]

    Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal components analysis (PCA).Computers & Geosciences19, 3 (1993), 303–342

    1993

  46. [46]

    MindSpore Team. 2020. MindSpore: A new open source deep learning train- ing/inference framework that could be used for mobile, edge and cloud scenarios. https://www.mindspore.cn/ Accessed: 2026-04

    2020

  47. [47]

    Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Aman- preet Singh, and Douwe Kiela. 2024. Generative representational instruction tuning.arXiv preprint arXiv:2402.09906(2024)

    work page arXiv 2024

  48. [48]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  49. [49]

    Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu Ou, et al. 2019. Personalized re-ranking for recommendation. InRecsys’19. 3–11

    2019

  50. [50]

    Yunchen Pu, Zhe Gan, Ricardo Henao, Xin Yuan, Chunyuan Li, Andrew Stevens, and Lawrence Carin. 2016. Variational autoencoder for deep learning of images, labels and captions.Advances in neural information processing systems29 (2016)

    2016

  51. [51]

    Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Representation learning with large language models for recommendation. InProceedings of WWW 2024. 3464–3475

    2024

  52. [52]

    Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. InProceedings of the 28th ACM international conference on information and knowledge management. 1161–1170

    2019

  53. [53]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, and Li. 2024. Idgenrec: Llm-recsys alignment with textual id learning. InSIGIR’24’. 355–364

    2024

  54. [54]

    Changxin Tian, Binbin Hu, Chunjing Gan, Haoyu Chen, Zhuo Zhang, Li Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, and Jiawei Chen. 2024. ReLand: Integrating Large Language Models’ Insights into Industrial Recommenders via a Controllable Reasoning Pool. InRecsys’24. 63–73

    2024

  55. [55]

    Naftali Tishby and Noga Zaslavsky. 2015. Deep learning and the information bottleneck principle. In2015 IEEE Information Theory Workshop (ITW). IEEE, 1–5. https://doi.org/10.1109/ITW.2015.7133169 SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia Yunjia Xi et al

    work page doi:10.1109/itw.2015.7133169 2015

  56. [56]

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  57. [57]

    Zhizhong Wan, Bin Yin, Junjie Xie, Fei Jiang, Xiang Li, and Wei Lin. 2024. LARR: Large Language Model Aided Real-time Scene Recommendation with Semantic Understanding. InProceedings of Recsys. 23–32

    2024

  58. [58]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17. 1–7

    2017

  59. [59]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. InWWW’21. 1785–1797

    2021

  60. [60]

    Shuai Wang, Shengyao Zhuang, Bevan Koopman, and Guido Zuccon. 2024. 2D Matryoshka Training for Information Retrieval.arXiv preprint arXiv:2411.17299 (2024)

    work page arXiv 2024

  61. [61]

    Yuling Wang, Changxin Tian, Binbin Hu, Yanhua Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, Liang Pang, and Xiao Wang. 2024. Can Small Language Models be Good Reasoners for Sequential Recommendation?. InProceedings of the ACM on Web Conference 2024. 3876–3887

    2024

  62. [62]

    Yasi Wang, Hongxun Yao, and Sicheng Zhao. 2016. Auto-encoder based dimen- sionality reduction.Neurocomputing184 (2016), 232–242

    2016

  63. [63]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2023. A Survey on Large Language Models for Recommendation.arXiv preprint arXiv:2305.19860(2023)

    work page arXiv 2023

  64. [64]

    Yunjia Xi, Jianghao Lin, Menghui Zhu, Yongzhao Xiao, Zhuoying Ou, Jiaqi Liu, Tong Wan, Bo Chen, Weiwen Liu, Yasheng Wang, et al. 2025. InfoDeepSeek: Benchmarking Agentic Information Seeking for Retrieval-Augmented Generation. arXiv preprint arXiv:2505.15872(2025)

    work page arXiv 2025

  65. [65]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, Rui Zhang, et al. 2023. Towards open-world recommendation with knowledge augmentation from large language models. arXiv preprint arXiv:2306.10933(2023)

    work page arXiv 2023

  66. [66]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. MemoCRS: Memory-enhanced Sequential Conversational Recom- mender Systems with Large Language Models.arXiv preprint arXiv:2407.04960 (2024)

    work page arXiv 2024

  67. [67]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Muyan Weng, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Yong Yu, et al. 2024. Efficient and Deploy- able Knowledge Infusion for Open-World Recommendations via Large Language Models.arXiv preprint arXiv:2408.10520(2024)

    work page arXiv 2024

  68. [68]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Chuhan Wu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Play to Your Strengths: Collaborative Intelli- gence of Conventional Recommender Models and Large Language Models.arXiv preprint arXiv:2403.16378(2024)

    work page arXiv 2024

  69. [69]

    Yunjia Xi, Weiwen Liu, Yang Wang, Ruiming Tang, Weinan Zhang, Yue Zhu, Rui Zhang, and Yong Yu. 2023. On-device integrated re-ranking with heteroge- neous behavior modeling. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5225–5236

    2023

  70. [70]

    Yunjia Xi, Hangyu Wang, Bo Chen, Jianghao Lin, Menghui Zhu, Weiwen Liu, Ruiming Tang, Zhewei Wei, Weinan Zhang, and Yong Yu. 2025. Efficiency un- leashed: Inference acceleration for LLM-based recommender systems with spec- ulative decoding. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1891–1901

    2025

  71. [71]

    Yunjia Xi, Hangyu Wang, Bo Chen, Jianghao Lin, Menghui Zhu, Weiwen Liu, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. A Decoding Acceleration Framework for Industrial Deployable LLM-based Recommender Systems.arXiv preprint arXiv:2408.05676(2024)

    work page arXiv 2024

  72. [72]

    Yunjia Xi, Muyan Weng, Wen Chen, Chao Yi, Dian Chen, Gaoyang Guo, Mao Zhang, Jian Wu, Yuning Jiang, Qingwen Liu, et al. 2025. Bursting filter bubble: Enhancing serendipity recommendations with aligned large language models. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 5059–5070

    2025

  73. [73]

    Chaojun Xiao, Zhengyan Zhang, Chenyang Song, Dazhi Jiang, Feng Yao, Xu Han, Xiaozhi Wang, Shuo Wang, Yufei Huang, Guanyu Lin, et al. 2024. Configurable foundation models: Building llms from a modular perspective.arXiv preprint arXiv:2409.02877(2024)

    work page arXiv 2024

  74. [74]

    Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Yanqiao Zhu, May D Wang, Joyce C Ho, Chao Zhang, and Carl Yang. 2024. Bmretriever: Tuning large language models as better biomedical text retrievers.arXiv preprint arXiv:2404.18443(2024)

    work page arXiv 2024

  75. [75]

    Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Jundong Li, and Zi Huang. 2023. Self-supervised learning for recommender systems: A survey.IEEE Transactions on Knowledge and Data Engineering(2023)

    2023

  76. [76]

    Yisong Yue, Thomas Finley, Filip Radlinski, and Thorsten Joachims. 2007. A support vector method for optimizing average precision. InSIGIR’07. 271–278

    2007

  77. [77]

    Chao Zhang, Shiwei Wu, Haoxin Zhang, Tong Xu, Yan Gao, Yao Hu, and En- hong Chen. 2024. NoteLLM: A Retrievable Large Language Model for Note Recommendation. InCompanion Proceedings of WWW. 170–179

    2024

  78. [78]

    Zhengyan Zhang, Zhiyuan Zeng, Yankai Lin, Chaojun Xiao, Xiaozhi Wang, Xu Han, Zhiyuan Liu, Ruobing Xie, Maosong Sun, and Jie Zhou. 2023. Emergent modularity in pre-trained transformers.arXiv preprint arXiv:2305.18390(2023)

    work page arXiv 2023

  79. [79]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448

    2024

  80. [80]

    Zhi Zheng, Wenshuo Chao, Zhaopeng Qiu, Hengshu Zhu, and Hui Xiong. 2024. Harnessing large language models for text-rich sequential recommendation. In Proceedings of the ACM on Web Conference 2024. 3207–3216

    2024

Showing first 80 references.