Pith. sign in

REVIEW 3 major objections 6 minor 63 references

IMFuse: Instance-Aware Multi-Layer Fusion for LLM-Enhanced Sequential Recommendation

T0 review · 3 major / 6 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Final-layer LLM item embeddings are not enough: adaptively fusing all layers, item by item, lifts sequential recommendation by about 6.7%.

desk verdict Solid incremental systems paper: multi-layer LLM fusion with item-routed experts gives consistent mid-single-digit gains and low overhead; worth a referee, not a paradigm shift. read the letter →

arxiv 2607.27002 v1 pith:HBRWWQCC submitted 2026-07-29 cs.IR

Yuheng Zheng (1) , Yu Cui (1) , Bin Wu (2) , Jian Zhang (3) , Ye Feng (3) , Can Wang (1) , Jiawei Chen (1) ((1) Zhejiang University , (2) Zhengzhou University
show 2 more authors
(3) University of Science Technology of China)
This is my paper · ORCID
classification cs.IR
keywords sequentialrecommendationlargelanguagemodelsmulti-layerrepresentationsinstance-awarefusionsemanticembeddingenhancementdimensionalcollapse
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

Sequential recommenders that enrich item IDs with LLM text embeddings almost always take only the last Transformer layer. This paper shows that practice is incomplete: deeper layers grow more collapsed in their spectrum, earlier layers keep complementary coarse-to-fine meaning, and different kinds of items evolve differently across depth, so one fixed layer is a poor fit for everyone. IMFuse therefore learns a shared, dimension-wise preference over all layers and then gently reweights that preference per item with a small set of expert templates routed from the item’s final-layer summary. The fused embedding drops into existing ID–semantic pipelines without changing the recommender backbone. On four Amazon catalogs, two backbones, and four enhancement methods, the same module yields consistent gains (about 6.72% relative on average) at negligible extra parameters and runtime.

What carries the argument

IMFuse: a global learnable score matrix over layers (softmaxed per embedding dimension), modulated per item by a soft mixture of shared layer-preference expert templates whose routing comes from the final-layer embedding, then used to weight-sum transformed multi-layer embeddings into one semantic vector.

What would settle it

On the same four datasets and backbones, replace the final-layer router with a fixed or randomly routed mixture (or drop instance modulation entirely) and check whether the reported gains over strong final-layer enhancers and over simple multi-layer averages disappear; or show that shallower layers never improve next-item ranking once dimensional collapse is controlled.

Watch

Extended reading notes

Core claim

Relying solely on the final LLM layer for item semantics is suboptimal for sequential recommendation because that layer suffers stronger dimensional collapse, intermediate layers carry non-redundant signals, and items differ in which depths help them. An instance-aware fusion that combines global dimension-wise layer weights with light expert modulation of those weights produces better item representations and improves standard enhancement pipelines by roughly 6.7% on average.

Load-bearing premise

That an item’s final-layer embedding is a good enough summary to decide how much each earlier layer should matter for that item.

Editorial extensions

If this is right

  • LLM-enhanced sequential recommenders should store or cache intermediate item hidden states, not only the last layer.
  • Existing final-layer adapters (alignment, SVD init, spectral transforms) can be upgraded by swapping in multi-layer fused semantics without redesigning the recommender.
  • Layer usefulness is item-dependent, so uniform depth selection or plain averaging leaves accuracy on the table.
  • The same global-plus-instance pattern can transfer across LLM encoders (e.g., LLaMA-3 and Qwen3) with modest retuning.

Reading between the lines

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

  • If final-layer routing is only a convenient proxy, user- or session-conditioned routing could further personalize which semantic depths matter.
  • The spectral-collapse diagnosis suggests multi-layer fusion may matter most for long-tail or fine-grained catalog items whose distinctions live in subordinate singular directions.
  • Caching all layers is the practical bottleneck; learned sparse layer subsets per catalog could keep the gains while cutting memory.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper argues that LLM-enhanced sequential recommenders underuse intermediate-layer item embeddings by defaulting to the final layer, which exhibits stronger dimensional (spectral) collapse and is not uniformly optimal across items. Motivated by layer-wise singular-value concentration, inter-layer cosine similarity, and cluster-wise trajectory differences (Figs. 1–2; §3.1), it proposes IMFuse: a global dimension-wise layer-preference matrix W (Eqs. 8–11) combined with instance-aware expert modulation that routes from the final-layer embedding to modulate those preferences per item. The fused semantic embedding plugs into existing ID–semantic pipelines without changing the SR backbone. Experiments on four Amazon datasets, SASRec and HSTU, four enhancement methods (RLMRec, LLM-ESR, LLMInit, SpecTran), multi-layer fusion baselines, ablations, efficiency, and a Qwen3-8B transfer study report consistent gains (claimed average relative improvement 6.72%) with small parameter/runtime overhead.

Significance. If the empirical gains hold under standard reproducibility checks, this is a useful, model-agnostic systems contribution for LLM-enhanced sequential recommendation: it reframes “which LLM layer to use” as a recommendation-supervised, dimension- and item-adaptive fusion problem rather than a fixed final-layer choice. Strengths include clear motivation analyses tied to the method, broad coverage (4 datasets × 2 backbones × multiple enhancers), ablations that move in the expected direction (Table 4: last/mean/w/o IM/rand/perm), efficiency numbers (Table 5), cross-LLM transfer (Table 6), and qualitative evidence of non-monotonic global weights and heterogeneous expert routing (Fig. 4). The overhead is modest, which matters for practical adoption. The work is incremental relative to multi-layer embedding fusion outside recsys, but the recommendation-specific design (global dim-wise preferences + item modulation under next-item loss) is a legitimate and timely delta.

major comments (3)
  1. [§4.2, Tables 2–3; Abstract] Tables 2–3 and the Abstract’s 6.72% average relative improvement are the central empirical claim, yet the manuscript does not report multi-seed means/stds, confidence intervals, or significance tests. Relative lifts on sparse Amazon splits (especially Office) can be sensitive to initialization and early stopping. Please add at least 3–5 random seeds (or equivalent paired tests) for the main SASRec/HSTU + SpecTran/RLMRec settings and state whether the average improvement remains stable; without this, the headline percentage is harder to trust at journal standard.
  2. [§3.3, Eq. (9); Appendix A.1, Table 7] Eq. (9) routes instance-aware modulation exclusively from the final-layer embedding E^L_i, while §3.1 and Fig. 1 argue that deeper layers are more collapsed and that useful depth varies by item. Appendix Table 7 also shows standalone NDCG strongly favoring the final layer. Table 4’s w/o IM ablation supports modulation in aggregate, but does not test whether the router input is adequate (e.g., shallow/mid summary, concatenated layer stats, or ID embedding as router features). A small ablation on router input would either validate the weakest design assumption or show the gains are mostly from global W rather than true instance-depth matching.
  3. [§3.1; §4.6, Figure 4; Table 4] Motivation treats spectral collapse and inter-layer dissimilarity as evidence of complementary recommendation-useful signal (§3.1, Eqs. 4–7). Mean fusion is sometimes competitive and sometimes weaker (Table 4; Appendix Table 8), which is consistent with non-uniform layer value but does not directly show that subordinate singular components improve ranking rather than adding noise that the supervised fusion happens to reweight. Briefly linking learned A_i / G mass to layers with less collapse (or to cluster trajectories in Fig. 2) would tighten the causal story between the analyses and the claimed mechanism.
minor comments (6)
  1. [Title page; References] Venue/metadata placeholders remain throughout (Conference acronym ’XX, Woodstock, NY, 2018 copyright, https://doi.org/XXXXXXX.XXXXXXX). Several cited venues are dated 2025–2026 (e.g., SpecTran SIGIR’26, CASE-MLP EACL’26, VA-HS ICML’26); ensure bibliographic accuracy before camera-ready.
  2. [§2.2–§3.3] Notation switches among E_l, eE_l, E_{f,i}, E_sem and uses both L+1 layers (including 0) and “32” as final in figures; define whether layer 0 is embeddings or first block once in §2.2.1.
  3. [Figure 2] Figure 1–2 captions and axis labels are readable but the dual-panel similarity heatmaps in Fig. 2(b) are not explained in the main text beyond a short clause; add one sentence on what structural difference the two groups illustrate.
  4. [§4.2, Table 3] Table 3 reports multi-layer baselines only on Clothing and Toy with two enhancers; a sentence on why Beauty/Office were omitted (space vs. cost) would help.
  5. [§4.1.4; §4.3] Hyperparameters η ∈ {5,10}, β ∈ {0.1,0.2,0.3}, M=3 are stated (§4.1.4) but sensitivity is only partly covered via w/ Rand. and w/ Perm.; a one-row sensitivity on M would strengthen the efficiency narrative.
  6. [§1; §3.1] Minor prose issues: “inevitably leaves” (§1), “dimensional collapse” vs “spectral collapse” used interchangeably—pick one primary term after first definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical multi-layer fusion trained and evaluated on held-out splits; gains are not forced by definition or self-citation.

full rationale

IMFuse is a systems/empirical paper. Its load-bearing claim is that recommendation-supervised fusion of multi-layer LLM item states (global dimension-wise preferences W plus instance-aware expert modulation via a router on E^L_i) improves next-item ranking over final-layer baselines and generic multi-layer aggregators. Layer weights and experts are optimized under InfoNCE on training sequences and measured by HR/NDCG on chronological leave-one-out test splits across four datasets, two backbones, four enhancement pipelines, and a second LLM encoder. That is ordinary supervised learning, not a derivation that reduces the reported lift to a fitted identity. Motivating analyses (singular-value collapse, inter-layer cosine similarity, item-group trajectories) are observational diagnostics, not uniqueness theorems. Depth-biased initialization of W and final-layer routing are design choices; ablations (w/ Last, w/ Mean, w/o IM, w/ Rand., w/ Perm.) and learned non-monotonic weights / heterogeneous routing (Figure 4) test them rather than smuggle the result in by construction. Overlapping-author citations (e.g., SpecTran) supply baselines and related methods, not a load-bearing uniqueness or ansatz that forces the 6.72% claim. No self-definitional loop, fitted-input-as-prediction, or renaming of a known identity appears in the equation chain (Eqs. 8–11).

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

Load-bearing content is empirical method design plus standard recsys evaluation assumptions. Free knobs (η, β, expert count, init scheme) and domain choices (frozen LLM, final-layer routing, InfoNCE, Amazon leave-one-out) carry the claim more than any invented physical entity. No formal derivation from first principles.

free parameters (5)
  • depth-bias strength η = chosen in {5, 10}
    Initializes global layer scores as W_l = η(l/L)1_d; selected from {5,10} and shapes optimization trajectory (Sec. 3.2, 4.1.4).
  • modulation strength β = chosen in {0.1, 0.2, 0.3}
    Scales instance-specific template adjustment in A_i = softmax_l(W ⊙ (1+β b_i)); grid {0.1,0.2,0.3}.
  • number of experts M = 3
    Size of layer-preference template bank B and router output; fixed to 3 without extensive justification.
  • recommendation embedding dim d and sequence/model hyperparameters = d=128; history=10; batch=256
    d=128, max history 10, batch 256, SASRec 2 blocks/1 head, weight decay/dropout grids—standard but claim-sensitive knobs.
  • global layer-score matrix W and template bank B = trained end-to-end (not closed form)
    Learned parameters that directly define fusion weights; central fitted objects under NDCG/HR supervision.
assumptions (5)
  • domain assumption Frozen multi-layer LLM encodings of item text are valid, fixed semantic features for collaborative sequential ranking.
    Sec. 2.2–3: LLM stays frozen; only Φ, fusion, and IMFuse train.
  • domain assumption Next-item prediction under chronological leave-one-out on Amazon reviews with InfoNCE (64 negatives) is an adequate test of recommendation quality.
    Sec. 4.1 evaluation protocol shared with prior LLM-rec papers.
  • ad hoc to paper Final-layer item embedding is a sufficient router input for choosing layer mixtures for that item.
    Eq. 9 and Sec. 3.3 explicitly choose E^L_i for routing.
  • ad hoc to paper Dimensional collapse / inter-layer dissimilarity imply complementary recommendation-useful signal rather than noise or task-irrelevant features.
    Sec. 1 and 3.1 motivate fusion from spectral and cosine analyses.
  • standard math Standard linear algebra SVD/cosine and K-means++ on embeddings are appropriate probes of representation geometry.
    Eqs. 4–7 and Appendix clustering procedure.
invented entities (2)
  • IMFuse global dimension-wise layer preference matrix G/W
    purpose: Shared per-dimension softmax weights over LLM layers before item modulation.
    Core trainable object introduced in Sec. 3.2; no existence claim beyond the module.
  • Instance-aware expert modulation (template bank B + MLP router)
    purpose: Produce item-specific adjustments b_i to global layer scores without per-item full matrices.
    Sec. 3.3 MoE-style design specific to this paper’s fusion recipe.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IMFuse: Instance-Aware Multi-Layer Fusion for LLM-Enhanced Sequential Recommendation." pith.science (2026). https://pith.science/paper/HBRWWQCC

@misc{pith2026260727002,
  author       = {Pith},
  title        = {Pith review of: IMFuse: Instance-Aware Multi-Layer Fusion for LLM-Enhanced Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBRWWQCC}},
  note         = {Machine review of arXiv:2607.27002}
}
read the original abstract

Recent advancements in Large Language Models (LLMs) have significantly enhanced sequential recommendation by encoding rich item textual information into semantic representations. However, existing methods typically rely on the final-layer hidden states of LLMs, overlooking potentially useful semantic signals encoded in other layers. Through empirical analysis, we reveal the limitations of this practice: final-layer representations often suffer from dimensional collapse, whereas intermediate layers preserve complementary, coarse-to-fine semantic knowledge. Furthermore, we observe that different items exhibit heterogeneous layer-wise representation evolution, making a uniform layer selection sub-optimal. To bridge this gap, we propose IMFuse, an instance-aware multi-layer fusion strategy designed for LLM-enhanced recommendation. Instead of relying on a single layer, IMFuse adaptively aggregates multi-layer semantic information by learning global dimension-wise layer preferences to capture general semantic contributions. To address item-level heterogeneity, IMFuse introduces an instance-aware expert modulation mechanism that dynamically adjusts these global preferences, generating personalized, item-specific semantic representations. Extensive experiments across four real-world datasets demonstrate the effectiveness of IMFuse. It consistently outperforms state-of-the-art baselines with an average relative improvement of 6.72%, while introducing limited parameter and computational overhead.

Figures

Figures reproduced from arXiv: 2607.27002 by the authors.

Figure 1
Figure 1. The Top-𝐾 singular value proportion and inter￾layer cosine similarity analyses on Amazon Clothing with LLaMA-3-8B [10] as the LLM backbone. Deeper LLM layer embeddings exhibit stronger spectral collapse, and different layers exhibit significant semantic disparities. lose potentially beneficial semantic knowledge for recommenda￾tion, inevitably leaves the capabilities of LLMs underutilized. This raises a natural ques… view at source ↗
Figure 2
Figure 2. Item-group heterogeneity across LLM layers. (a) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overall framework of the proposed IMFuse. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The global layer score evolution and expert-routing [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Selecting the number of item groups 𝐾 using the elbow method on the Clothing dataset [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 5 canonical work pages

  1. [1]

    David Arthur and Sergei Vassilvitskii. 2007. K-means++: The Advantages of Careful Seeding. InProceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms. 1027–1035

    2007

  2. [2]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014. doi:10.1145/3604915.3608857

    arXiv 2023

  3. [4]

    Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. HLLM: Enhancing Sequential Recommendations via Hierarchical Large Language Models for Item and User Modeling. arXiv:2409.12740 doi:10.48550/arXiv.2409.12740

    arXiv doi:10.48550/arxiv.2409.12740 2024

  4. [5]

    Sishuo Chen, Xiaohan Bi, Rundong Gao, and Xu Sun. 2022. Holistic Sentence Em- beddings for Better Out-of-Distribution Detection. InFindings of the Association for Computational Linguistics: EMNLP. 6676–6686. doi:10.18653/v1/2022.findings- emnlp.497

    doi:10.18653/v1/2022.findings- 2022

  5. [6]

    Yu Cui, Feng Liu, Pengbo Wang, Bohao Wang, Heng Tang, Yi Wan, Jun Wang, and Jiawei Chen. 2024. Distillation Matters: Empowering Sequential Recommenders to Match the Performance of Large Language Models. InProceedings of the 18th ACM Conference on Recommender Systems. 507–517. doi:10.1145/3640457.3688118

    arXiv 2024

  6. [7]

    Yu Cui, Feng Liu, Zhaoxiang Wang, Changwang Zhang, Jun Wang, Can Wang, and Jiawei Chen. 2026. SpecTran: Spectral-Aware Transformer-Based Adapter for LLM-Enhanced Sequential Recommendation. InProceedings of the 49th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval. 234–245. doi:10.1145/3805712.3809701

    arXiv 2026

  7. [8]

    Kounianhua Du, Jizheng Chen, Jianghao Lin, Yunjia Xi, Hangyu Wang, Xinyi Dai, Bo Chen, Ruiming Tang, and Weinan Zhang. 2024. DisCo: Towards Har- monious Disentanglement and Collaboration between Tabular and Semantic Space for Recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 666–676. doi:10.1145/363752...

    arXiv 2024

  8. [9]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). InProceedings of the 16th ACM Conference on Recommender Systems. 299–315. doi:10.1145/3523227.3546767

    arXiv 2022

Show all 63 references
  1. [10]

    Aaron Grattafiori et al. 2024. The Llama 3 Herd of Models. arXiv:2407.21783 doi:10.48550/arXiv.2407.21783

    2024 doi arXiv

  2. [11]

    Jiho Gwak and Yuchul Jung. 2025. Layer-Aware Embedding Fusion for LLMs in Text Classifications. arXiv:2504.05764 doi:10.48550/arXiv.2504.05764

    2025 doi arXiv

  3. [12]

    Yingzhi He, Xiaohao Liu, An Zhang, Yunshan Ma, and Tat-Seng Chua. 2025. LLM2Rec: Large Language Models Are Powerful Embedding Models for Sequen- tial Recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. 896–907. doi:10.114...

    2025

  4. [13]

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

  5. [14]

    Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards Universal Sequence Representation Learning for Recom- mender Systems. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 585–593. doi:10.1145/353...

    2022

  6. [15]

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large Language Models Are Zero-Shot Rankers for Recommender Systems. InAdvances in Information Retrieval. 364–381. doi:10. 1007/978-3-031-56060-6_24

    2024

  7. [16]

    Guoqing Hu, An Zhang, Shuo Liu, Zhibo Cai, Xun Yang, and Xiang Wang. 2025. AlphaFuse: Learn ID Embeddings for Sequential Recommendation in Null Space of Language Embeddings. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information...

    2025

  8. [17]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated Gain-Based Evaluation of IR Techniques.ACM Transactions on Information Systems20, 4 (2002), 422–446. doi:10.1145/582415.582418

    2002

  9. [18]

    Ganesh Jawahar, Benoît Sagot, and Djamé Seddah. 2019. What Does BERT Learn about the Structure of Language?. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 3651–3657. doi:10.18653/v1/P19-1356

    2019 doi

  10. [19]

    Jian Jia, Yipei Wang, Yan Li, Honggang Chen, Xuehan Bai, Zhaocheng Liu, Jian Liang, Quan Chen, Han Li, Peng Jiang, and Kun Gai. 2025. LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Indus- trial Application. InProceedings of the AAAI Confe...

    2025 doi

  11. [20]

    Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. 2022. Understanding Dimensional Collapse in Contrastive Self-Supervised Learning. InInternational Conference on Learning Representations

    2022

  12. [21]

    Tianjie Ju, Weiwei Sun, Wei Du, Xinwei Yuan, Zhaochun Ren, and Gongshen Liu. 2024. How Large Language Models Encode Context Knowledge? A Layer- Wise Probing Study. InProceedings of LREC-COLING. 8235–8246. doi:10.63317/ 2IB4REXE35CA

    2024

  13. [22]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Zheng et al. 197–206

    2018

  14. [23]

    Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large Language Models Meet Collaborative Filtering: An Efficient All-Round LLM-Based Recommender System. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Da...

    2024

  15. [24]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. InInternational Conference on Learning Representations

    2015

  16. [27]

    Xinyu Lin, Wenjie Wang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. 2024. Bridging Items and Language: A Transition Paradigm for Large Language Model-Based Recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1816–1...

    2024

  17. [28]

    Liu, Matt Gardner, Yonatan Belinkov, Matthew E

    Nelson F. Liu, Matt Gardner, Yonatan Belinkov, Matthew E. Peters, and Noah A. Smith. 2019. Linguistic Knowledge and Transferability of Contextual Represen- tations. InProceedings of NAACL-HLT. 1073–1094. doi:10.18653/v1/N19-1112

    2019 doi

  18. [30]

    Qidong Liu, Xian Wu, Yejing Wang, Zijian Zhang, Feng Tian, Yefeng Zheng, and Xiangyu Zhao. 2024. LLM-ESR: Large Language Models Enhancement for Long-Tailed Sequential Recommendation. InAdvances in Neural Information Processing Systems, Vol. 37. 26701–26727. doi:10.52202/079017-0839

    2024 doi

  19. [31]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

  20. [32]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Distributed Representations of Words and Phrases and their Compositionality. In Proceedings of the 26th International Conference on Neural Information Processing Systems. 3111–3119

    2013

  21. [33]

    Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, ...

    2019

  22. [34]

    Lutz Prechelt. 1998. Early Stopping—But When? InNeural Networks: Tricks of the Trade. Springer, 55–69. doi:10.1007/3-540-49430-8_3

    1998 doi

  23. [35]

    Qwen Team. 2025. Qwen3 Technical Report. arXiv:2505.09388 doi:10.48550/arXiv. 2505.09388

    2025 doi arXiv

  24. [36]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. 3982–3992. doi:10.18653/v1/D19-1410

    2019 doi

  25. [37]

    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 the ACM Web Conference. 3464–

    2024

  26. [38]

    Yankun Ren, Zhongde Chen, Xinxing Yang, Longfei Li, Cong Jiang, Lei Cheng, Bo Zhang, Linjian Mo, and Jun Zhou. 2024. Enhancing Sequential Recommenders with Augmented Knowledge from Aligned Large Language Models. InProceedings of the 47th International ACM SIGIR Conference on R...

    2024

  27. [39]

    Leheng Sheng, An Zhang, Yi Zhang, Yuxin Chen, Xiang Wang, and Tat-Seng Chua

  28. [40]

    Wentao Shi, Xiangnan He, Yang Zhang, Chongming Gao, Xinyue Li, Jizhi Zhang, Qifan Wang, and Fuli Feng. 2024. Large Language Models Are Learnable Planners for Long-Term Recommendation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in I...

    2024

  29. [41]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional En- coder Representations from Transformer. InProceedings of the 28th ACM In- ternational Conference on Information and Knowledge Managemen...

    2019

  30. [42]

    Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, and Jun Xu. 2024. Large Language Models Enhanced Collaborative Filtering. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2178–2188. doi:10.1145/3627673.3679558

    2024

  31. [43]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. IDGenRec: LLM-RecSys Alignment with Textual ID Learning. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364. doi:10.11...

    2024

  32. [44]

    Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining. 565–573. doi:10.1145/ 3159652.3159656

    2018

  33. [45]

    Yixuan Tang and Yi Yang. 2024. Pooling and Attention: What Are Effective Designs for LLM-Based Embedding Models? arXiv:2409.02727 doi:10.48550/ arXiv.2409.02727

    2024 doi arXiv

  34. [46]

    Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT Rediscovers the Classical NLP Pipeline. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 4593–4601. doi:10.18653/v1/P19-1452

    2019 doi

  35. [47]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 doi:10.48550/arXiv.1807. 03748

    2018 doi arXiv

  36. [48]

    Bin Wang and C.-C. Jay Kuo. 2020. SBERT-WK: A Sentence Embedding Method by Dissecting BERT-Based Word Models.IEEE/ACM Transactions on Audio, Speech, and Language Processing28 (2020), 2146–2157. doi:10.1109/TASLP.2020.3008390

    2020

  37. [49]

    Bohao Wang, Feng Liu, Jiawei Chen, Xingyu Lou, Changwang Zhang, Jun Wang, Yuegang Sun, Yan Feng, Chun Chen, and Can Wang. 2025. MSL: Not All Tokens Are What You Need for Tuning LLM as a Recommender. InProceedings of the 48th International ACM SIGIR Conference on Research and D...

    2025

  38. [50]

    Bohao Wang, Feng Liu, Changwang Zhang, Jiawei Chen, Yudi Wu, Sheng Zhou, Xingyu Lou, Jun Wang, Yan Feng, Chun Chen, and Can Wang. 2025. LLM4DSR: Leveraging Large Language Model for Denoising Sequential Recommendation. ACM Transactions on Information Systems44, 1 (2025), 1–32. ...

    2025 doi

  39. [51]

    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 Web Conference 2024. 3876–3887. doi:10.1145/3589334.3645671

    2024

  40. [52]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. LLMRec: Large Language Models with Graph Augmentation for Recommendation. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 8...

    2024

  41. [53]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models. InProceedings of the 18th ACM Conference on Recommender S...

    2024 doi

  42. [54]

    Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive Learning for Sequential Recommendation. In2022 IEEE 38th International Conference on Data Engineering. 1259–1273. doi:10. 1109/ICDE53745.2022.00099

    2022

  43. [55]

    Wujiang Xu, Qitian Wu, Zujie Liang, Jiaojiao Han, Xuying Ning, Yunxiao Shi, Wenfang Lin, and Yongfeng Zhang. 2025. SLMRec: Distilling Large Language Models into Small for Sequential Recommendation. InInternational Conference on Learning Representations

    2025

  44. [56]

    Zhengyi Yang, Xiangnan He, Jizhi Zhang, Jiancan Wu, Xin Xin, Jiawei Chen, and Xiang Wang. 2023. A Generic Learning Framework for Sequential Recom- mendation with Distribution Shifts. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in In...

    2023

  45. [57]

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to Go Next for Recommender Systems? ID- vs. Modality-Based Recommender Models Revisited. InProceedings of the 46th International ACM SIGIR Conference on Research and Devel...

    2023

  46. [58]

    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. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. InProceedings of the 41st In...

    2024

  47. [59]

    Gaifan Zhang, Yi Zhou, and Danushka Bollegala. 2026. CASE: Condition-Aware Sentence Embeddings for Conditional Semantic Textual Similarity Measurement. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long...

    2026

  48. [61]

    Weizhi Zhang, Liangwei Yang, Wooseong Yang, Henry Peng Zou, Yuqing Liu, Ke Xu, Sourav Medya, and Philip S. Yu. 2025. LLMInit: A Free Lunch from Large Language Models for Selective Initialization of Recommendation. InProceedings of the 2025 Conference on Empirical Methods in Na...

    2025 doi

  49. [62]

    Yeqin Zhang, Yunfei Wang, Jiaxuan Chen, Ke Qin, Yizheng Zhao, and Cam- Tu Nguyen. 2026. LLM-Based Embeddings: Attention Values Encode Sentence Semantics Better Than Hidden States. InProceedings of the 43rd International Conference on Machine Learning. To appear

    2026

  50. [63]

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, and Qing Li. 2024. Recommender Systems in the Era of Large Language Models (LLMs).IEEE Transactions on Knowledge and Data Engineering36, 11 (2024), 6889–690...

    2024

  51. [64]

    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. 1435–1448. doi:10.1109/ICDE60...

    2024

  52. [2015]

    InProceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Image-Based Recommendations on Styles and Substitutes. InProceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval. 43–52. doi:10.1145/2766462.2767755

  53. [2016]

    In International Conference on Learning Representations

    Session-Based Recommendations with Recurrent Neural Networks. In International Conference on Learning Representations

  54. [2025]

    InInternational Conference on Learning Representations

    Language Representations Can Be What Recommenders Need: Findings and Potentials. InInternational Conference on Learning Representations

  55. [3475]

    doi:10.1145/3589334.3645458

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.