Pith. sign in

REVIEW 4 major objections 4 minor 25 references

REPREC: Representation Driven Parameter-Efficient Recommendation System

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read REPREC: a small learned injector lets a frozen LLM and a frozen recommender beat LoRA

desk verdict A straightforward, useful recipe for freezing both backbones and learning a small soft-token injector; the empirical claims need tightening, but the method is real and production-relevant. read the letter →

arxiv 2607.24845 v2 pith:YOSNOOHA submitted 2026-07-24 cs.IR cs.AI

Harshini Kavuru , Dwipam Katariya , Giri Iyengar , Pranab Mohanty , Kalanand Mishra , Raghu Machiraju This is my paper
classification cs.IRcs.AI
keywords sequentialrecommendationlargelanguagemodelsparameter-efficientfine-tuningfrozenbackbonessoftprompttokensuserrepresentationalignmentLoRAcomparisonsparsedata
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

REPREC asks whether a frozen sequential recommender and a frozen language model can be connected by a single small trainable module and still outperform fine-tuning. The paper answers yes: an MLP injector maps a fixed-size user embedding into a handful of soft tokens prepended to the LLM, and only that MLP is trained. Across five Amazon benchmarks, REPREC beats LoRA and LLaRA under matched parameter budgets, with the largest margins for casual and core users. Trained on short prompts and evaluated on long ones, REPREC keeps 85–100% of LoRA's accuracy while cutting per-epoch training time by about 1.5×. If right, it means high-quality LLM recommendation does not require touching either backbone.

What carries the argument

The load-bearing mechanism is the MLP injector gφ: R^d → R^{m×D}, which maps the user embedding from a frozen sequential encoder (SASRec or BERT4Rec) into m soft tokens prepended to the frozen LLM's input. A LayerNorm over token embeddings stabilizes the projection. The loss is cross-entropy on the answer token (Yes/No) of a binary candidate-relevance prompt, so gradients propagate through the LLM into the injector only. The fixed-size user embedding keeps computational cost independent of interaction history length.

What would settle it

A full-corpus ranking evaluation (or an online A/B test) on the same five datasets that reverses the HIT@K ordering between REPREC and LoRA would falsify the central claim; more cheaply, re-running the comparison with negatives sampled only from non-interacted items, or with a much larger negative set (e.g., K=1000), would test whether the margin survives.

Watch

Extended reading notes

Core claim

The central discovery is that a frozen sequential encoder and a frozen LLM, connected only by a lightweight MLP injector, can match or surpass LoRA-finetuned LLMs for next-item recommendation. The injector converts the user embedding into m=6 soft tokens in the LLM embedding space; a binary Yes/No prompt ranks candidate items, and training updates only the injector parameters. REPREC reports consistent gains over LoRA across five sparse Amazon datasets and two backbones, with relative HIT@10 gains up to 10.86% for casual users and 10.26% for core users, and 85–100% of LoRA performance when trained on short histories—while leaving both pretrained backbones unchanged.

Load-bearing premise

The comparisons rely on ranking the ground-truth item against 200 fixed negatives (half popularity-weighted, half random) drawn from the whole item pool without excluding previously interacted items; if this sampled-decision protocol doesn't reflect true next-item ranking or online metrics, the 'beats LoRA' conclusion may not transfer.

Editorial extensions

If this is right

  • LLM-based recommenders can be built without fine-tuning or modifying the LLM, so the same frozen LLM can serve many tasks.
  • The modular design lets practitioners swap the sequential encoder or the LLM backend without retraining the whole stack.
  • Training only the injector cuts trainable parameters to roughly 2.39M (comparable to LoRA r=8) and speeds per-epoch training by 1.43–1.81×.
  • Low-activity users, who are the majority in sparse catalogs, benefit most: HIT@10 improves over LoRA by up to about 10.9% for casual users.
  • Training on short prompt histories and evaluating with longer contexts preserves 85–100% of LoRA accuracy, enabling cheaper training.

Reading between the lines

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

  • If the alignment is truly task-agnostic, the same injector principle could connect frozen sequence encoders to frozen LLMs for other domains (e.g., time-series or graphs), where the encoder produces a fixed-size summary.
  • The paper's protocol fixes 200 negatives (50% popularity-weighted) sampled without excluding interacted items; the claim that REPREC 'outperforms LoRA' would be on stronger ground if verified under full-corpus ranking or with interacted items removed from the negative set.
  • A testable extension: vary the number of soft tokens m along with LoRA rank to map the full Pareto frontier; the paper's own figure suggests diminishing returns beyond m=2, so the injector may be compressible further.
  • The 85–100% retention result suggests a training–inference asymmetry that could be exploited for cheaper continual updating of the injector as new user history arrives.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. REPREC proposes a parameter-efficient LLM-based sequential recommender. A frozen SASRec or BERT4Rec encoder produces a user embedding; a lightweight MLP injector maps that embedding to m=6 soft tokens prepended to binary Yes/No prompts for a frozen LLaMA or Qwen LLM. Only the injector parameters (about 2.39M) are trained. On five Amazon datasets the paper reports HIT@5/HIT@10 for REPREC, LoRA-FT, LLaRA, and standalone recommenders, along with user-regime analyses and short-to-long prompt generalization experiments. The central claim is that, at matched parameter budgets, this fully frozen-frozen setup matches or beats LoRA and LLaRA, with the largest gains for casual and core users and an approximate 1.51x per-epoch training speedup in the 'cheap' setting.

Significance. If established, the result is practically significant: it would show that LoRA-level recommendation quality can be obtained without adapting either the sequential encoder or the LLM, simplifying deployment and allowing modular swaps of both backbones. The paper contributes useful ablations (random frozen injector, soft-token-count scaling, user persona breakdown), five-seed averaging for the main methods, matched-budget parameter-count tables, and openly released code. However, the headline comparison rests on a 200-negative binary decision protocol with close margins, and the efficiency and parameter-budget reporting contain inconsistencies. These issues prevent me from treating the abstract's 'consistently outperforms LoRA' claim as fully supported without further evidence.

major comments (4)
  1. [Appendix B; Tables 2 and 3] All headline comparisons rank the held-out item against K=200 fixed negatives (100 popularity-weighted, 100 uniform) drawn from the full item pool without excluding previously interacted items, scored by the Yes-token probability. This is a standard sampled proxy, but the margins in Table 2 are small (e.g., Toys HIT@10 REPREC-B 0.415±0.005 vs LoRA 0.421±0.003; Tools HIT@10 REPREC-S 0.319±0.003 vs LoRA 0.319±0.001) and no significance tests are reported. Because REPREC's injected user vector is trained by dot-product next-item prediction, it may separate true items from popular or uniform negatives more readily than LoRA's text-only scoring; the comparison could shift under full-corpus ranking. Please report full-corpus metrics or at least vary K and the popularity ratio ρ, and add significance tests or effect sizes, before claiming consistent superiority.
  2. [Table 3 and §3.4.3] The efficiency claim is not directly supported by the reported table. Table 3 gives train times for 'REPREC' (ℓ=50) and LoRA, but not for REPRECcheap (train ℓ=10). The 'Saved (min)' column equals REPREC time minus LoRA time (430−314=116, etc.), and the 'Speedup' column equals REPREC/LoRA (1.37×–1.81×), so it compares REPREC to LoRA, not short-prompt to long-prompt training. The sentence 'training with shorter histories reduces per-epoch training time by 1.43×–1.81×' needs actual per-epoch times for REPREC trained with ℓ=10. Without that column, the central efficiency claim is unverifiable.
  3. [§3.3, Appendix A, Table 4] The parameter budgets are only approximately matched: m=6 gives 2,392,192 trainable parameters vs LoRA r=8 with 2,293,760, a 4.3% difference. More importantly, LLaRA is said to be compared under the same LoRA rank, but its additional item-level projection parameters are not counted; if LLaRA's trainable count is larger, the 'outperforms or matches LLaRA' claim is not under matched budgets. Please report exact trainable-parameter counts for every method or use exact matching (e.g., m=5 for REPREC gives 1,995,904 parameters), and correct the §3.3 cross-reference to 'Appendix C' (the available appendices are A and B).
  4. [Abstract; §3.4.1, Table 2] The claim 'consistently outperforms LoRA' in the abstract is contradicted by Table 2: Toys REPREC-B is below LoRA on both HIT@5 (0.339 vs 0.346) and HIT@10 (0.415 vs 0.421), and Tools HIT@10 for REPREC-S is a tie. The body's weaker phrase 'outperforms or matches' also needs qualification because some REPREC-B cells lose. Please align the abstract and §3.4.1 with the actual table results or provide a setting in which the claimed consistency holds.
minor comments (4)
  1. [§3.4.3, Table 3] The performance-ratio range '0.85×–1.00×' in the text is inconsistent with Table 3, where Beauty shows 0.84×. The speedup range should also include 1.37× if that is the intended metric.
  2. [Figure 2] The caption uses 'prompt=0, prompt=10, prompt=50' but the text and figure description discuss 'Using no textual interaction history (prompt=0)'. Please define what these numbers mean (number of recent interaction items included as text) and make the caption self-contained.
  3. [§2.2.3] In the equation for p(y=1|u,i), the notation 'softmax(W_llm h_T)_Yes' is not fully defined; W_llm is not introduced and the relationship of h_T to the final-token hidden state should be stated more precisely.
  4. [Table 3] The table header is garbled in the current formatting; the columns need to be clearly separated so that 'REPREC cheap vs REPREC' and 'REPREC cheap vs LoRA' are distinguishable from train-time columns.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: REPREC is an empirical system evaluated against external baselines; no headline result reduces to a fitted input or self-citation chain.

full rationale

REPREC's central claims are empirical: a frozen SASRec/BERT4Rec encoder embedding is projected by a trainable MLP injector into m soft tokens that condition a frozen LLM, and the model is compared with LoRA-FT, LLaRA, and standalone sequential recommenders on five Amazon datasets using a shared held-out next-item protocol. There is no analytic derivation that folds the target result back into the inputs. The injector parameters are trained by the cross-entropy loss defined in Section 2.2.3, and evaluation uses fixed 200-negative sampling (Appendix B) applied uniformly to all compared methods, so the comparison is not a fitted-input-called-prediction. The matched-parameter-budget choice (m=6 soft tokens approx. LoRA r=8; 2.39M vs 2.29M parameters) is a deliberate design calibration, not a constructed prediction. The only self-citation is [11] (FinTRec, same authors) in a related-work list of 'prior industry-focused systems'; it is not load-bearing for any result. The paper contains no uniqueness theorem imported from the authors' prior work and no ansatz smuggled in via self-citation. The closest caveat is that all headline numbers depend on the Appendix B sampled-negative protocol; this is an external-validity risk of the metric, not circularity, and it applies equally to every baseline. Overall: no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The contribution is empirical, not derivational. The load-bearing inputs are assumptions inherited from multimodal soft-prompt literature and the standard sequential-recommendation evaluation scheme. No new physical or architectural entity is invented; the 'soft tokens' are ordinary learned parameters. The free parameters listed above are mostly hyperparameters chosen by hand or matched to the LoRA budget.

free parameters (5)
  • number of soft tokens m = 6 (swept over 2,4,6,8 in Fig. 2)
    Main configuration chosen to approximate LoRA r=8 parameter budget; the paper varies it in Fig. 2 but does not describe a held-out model-selection procedure.
  • injector hidden width h_d = 128
    MLP hidden dimension is fixed at 128 without a reported sensitivity sweep or justification.
  • training prompt history length ℓ = 50 (main), 10 (cheap)
    RQ3 explicitly varies training prompt length between 10 and 50; the main experiments use 50, and the efficiency claims depend on this choice.
  • negative sampling count K and popularity ratio ρ = K=200, ρ=0.5
    Evaluation protocol hyperparameters: 100 popularity-weighted and 100 uniform negatives per user. These affect all reported Hit@K numbers.
  • user persona thresholds = casual 0–5, core 6–20, power >20 interactions
    The user-regime conclusions (RQ2) depend on these post hoc thresholds; no sensitivity analysis is provided.
assumptions (4)
  • domain assumption Frozen LLMs can be effectively conditioned by prepended learned soft tokens from a foreign representation space
    Inherited from CLIPCap/BLIP-2 (cited in §1). If the frozen LLM embedding space is not receptive to foreign soft tokens, the injector cannot transfer user-level signal. Invoked throughout §2.2.3.
  • domain assumption SASRec/BERT4Rec user embeddings trained by next-item prediction capture sufficient collaborative/sequential signal for next-item ranking
    REPREC's gains depend on encoder quality; BERT4Rec is much weaker standalone (e.g., HIT@5 .080 on Beauty), yet REPREC-B still improves, suggesting the LLM carries much of the burden. Stated in §2.2.1.
  • domain assumption Binary Yes/No decision prompting with sampled negatives is a valid surrogate for ranking next items
    All LLM baselines are scored by the probability of a Yes token on a binary prompt, and ranking compares the ground-truth item against 200 sampled negatives (Appendix B). This constrains the meaning of Hit@K.
  • domain assumption Backpropagation through a frozen LLM into the injector is sufficient to align the two representation spaces without adapting either backbone
    This is the central design premise of REPREC: gradients flow through LLM activations into injector parameters only. Whether this reliably aligns arbitrary encoder–LLM pairs is an empirical assumption, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REPREC: Representation Driven Parameter-Efficient Recommendation System." pith.science (2026). https://pith.science/paper/YOSNOOHA

@misc{pith2026260724845,
  author       = {Pith},
  title        = {Pith review of: REPREC: Representation Driven Parameter-Efficient Recommendation System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOSNOOHA}},
  note         = {Machine review of arXiv:2607.24845}
}
read the original abstract

Large language models (LLMs) have been applied to sequential recommendation by formulating it as a natural language task. Previous work has improved personalization by incorporating collaborative and sequential signals through input conditioning or LLM fine-tuning. However, existing approaches often rely on one or more of the following: LLM fine-tuning, additional architectural modules, representation distillation, or item-level conditioning over long interaction histories, increasing training complexity and deployment cost. We propose REPREC, a lightweight framework that reformulates LLM-based sequential recommendation through lightweight user representation alignment. REPREC maps a fixed-size user embedding from a frozen sequential encoder into a small set of learned soft tokens through a lightweight MLP injector that conditions a frozen LLM, leaving both pretrained backbones unchanged while training only the injector. We conducted exhaustive experiments on multiple benchmark datasets and demonstrate that REPREC consistently outperforms LoRA while remaining compatible with different pretrained sequential encoders and LLM backbones, enabling a modular and production-friendly recommendation pipeline without modifying either pretrained component. The gains are particularly pronounced for casual and core users across all datasets, highlighting REPREC's effectiveness in low-data regimes. Finally, when trained on short prompt histories and evaluated with longer contexts, REPREC maintains 85-100% of LoRA's performance while reducing per-epoch training time by an average of 1.51X, demonstrating an effective balance between recommendation quality and computational efficiency for production deployment. The code is available at https://github.com/phdbotcode/REPREC

Figures

Figures reproduced from arXiv: 2607.24845 by the authors.

Figure 1
Figure 1. Overview of the REPREC architecture. The sequential encoder produces a user embedding projected into [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. HIT@5 versus the number of trainable parameters on Beauty, Sports, and Toys using LLaMA as LLM. For REPREC, we vary the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comparing REPREC-S and LoRA with LLaMA across Casual and Core users. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparing REPREC-S, LoRA, and REPREC-S variant for Power users. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 2 canonical work pages

  1. [1]

    Marco De Nadai, Francesco Fabbri, Paul Gigioli, Alice Wang, Ang Li, Fabrizio Silvestri, Laura Kim, Shawn Lin, Vladan Radosavljevic, Sandeep Ghael, David Nyhan, Hugues Bouchard, Mounia Lalmas, and Andreas Damianou. 2024. Personalized Audiobook Recommendations at Spotify Through Graph Neural Networks. InCompanion Proceedings of the ACM Web Conference 2024(S...

    arXiv 2024

  2. [2]

    Zhiang Dong, Liya Hu, Jingyuan Chen, Zhihua Wang, and Fei Wu. 2025. Comprehend Then Predict: Prompting Large Language Models for Recommendation with Semantic and Collaborative Data.ACM Trans. Inf. Syst.43, 5, Article 115 (July 2025), 26 pages. doi:10.1145/3716499

    work page doi:10.1145/3716499 2025

  3. [3]

    Ghazal Fazelnia, Sanket Gupta, Claire Keum, Mark Koh, Ian Anderson, and Mounia Lalmas. 2024. Generalized User Representations for Transfer Learning. arXiv:2403.00584 [cs.IR] https://arxiv.org/abs/2403.00584

    arXiv 2024

  4. [4]

    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(Seattle, WA, USA) (RecSys ’22). Association for Computing Machinery, New York, NY, USA, 299–315. doi:10.1...

    arXiv 2022

  5. [5]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural Collaborative Filtering. InProceedings of the 26th International Conference on World Wide Web(Perth, Australia)(WWW ’17). International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, CHE, 173–182. doi:10.1145/3038912.3052569

    arXiv 2017

  6. [6]

    Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. 2024. Bridging Language and Items for Retrieval and Recommendation. arXiv preprint arXiv:2403.03952(2024)

    arXiv 2024

  7. [7]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations

    2022

  8. [8]

    Albert Qiaochu Jiang, Alexandre Sablayrolles, Arthur Mensch, et al. 2023. Mistral 7B.arXiv preprint arXiv:2310.06825(2023)

    arXiv 2023

Show all 25 references
  1. [9]

    Wenqi Jiang, Zhenhao He, Shuai Zhang, Kai Zeng, Liang Feng, Jiansong Zhang, Tongxuan Liu, Yong Li, Jingren Zhou, Ce Zhang, and Gustavo Alonso. 2021. FleetRec: Large-Scale Recommendation Inference on Hybrid GPU-FPGA Clusters. InProceedings of the 27th ACM SIGKDD Conference on K...

    2021

  2. [10]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recommendation. In2018 IEEE International Conference on Data Mining (ICDM). 197–206. doi:10.1109/ICDM.2018.00035

    2018

  3. [11]

    Dwipam Katariya, Snehita Varma, Akshat Shreemali, Benjamin Wu, Kalanand Mishra, and Pranab Mohanty. 2025. FinTRec: Transformer Based Unified Contextual Ads Targeting and Personalization for Financial Applications.arXiv preprint arXiv:2511.14865(2025)

    2025

  4. [12]

    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

  5. [13]

    Sein Kim, Hongseok Kang, Kibum Kim, Jiwan Kim, Donghyun Kim, Minchul Yang, Kwangjin Oh, Julian McAuley, and Chanyoung Park. 2025. Lost in Sequence: Do Large Language Models Understand Sequential Recommendation?. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Disc...

    2025

  6. [14]

    Walid Krichene and Steffen Rendle. 2022. On sampled metrics for item recommendation.Commun. ACM65, 7 (June 2022), 75–83. doi:10.1145/3535335

    2022 doi

  7. [15]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. InProceedings of the 40th International Conference on Machine Learning(Honolulu, Hawaii, USA)(ICML’23). JMLR.org...

    2023

  8. [16]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assistant. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval(Washington DC, U...

    2024

  9. [17]

    Ron Mokady, Amir Hertz, and Amit Bermano. 2021. ClipCap: CLIP Prefix for Image Captioning. doi:10.48550/arXiv.2111.09734

    2021 doi arXiv

  10. [18]

    Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O’Banion, and Jun Xie. 2025. User-LLM: Efficient LLM Contextualization with User Embeddings. InCompanion Proceedings of the ACM on Web Conference 2025(Sydney NSW, Australia) (WWW ...

    2025

  11. [19]

    Nikil Pancha, Andrew Zhai, Jure Leskovec, and Charles Rosenberg. 2022. PinnerFormer: Sequence Modeling for User Representation at Pinterest. doi:10.48550/arXiv.2205.04507

    2022 doi arXiv

  12. [20]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Management (B...

    2019

  13. [21]

    Riya Widayanti, Mochamad Heru Chakim, Chandra Lukita, Untung Rahardja, and Ninda Lutfiani. 2023. Improving Recommender Systems using Hybrid Techniques of Collaborative Filtering and Content-Based Filtering.Journal of Applied Data Sciences4, 3 (2023), 289–302. doi:10.47738/jads...

    2023 doi

  14. [22]

    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 (ICDE). 1259–1273. doi:10.1109/ICDE53745.2022.00099

    2022

  15. [23]

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2025. Recommendation as Instruction Following: A Large Language Model Empowered Recommendation Approach.ACM Trans. Inf. Syst.43, 5, Article 114 (July 2025), 37 pages. doi:10.1145/3708882

    2025 doi

  16. [24]

    Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He. 2025. CoLLM: Integrating Collaborative Embeddings Into Large Language Models for Recommendation.IEEE Transactions on Knowledge and Data Engineering37, 5 (2025), 2329–2340. doi:10.1109/TKDE.2025.3540912

    2025

  17. [25]

    Xiaoxue Zhao, Weinan Zhang, and Jun Wang. 2013. Interactive collaborative filtering. InProceedings of the 22nd ACM International Conference on Information & Knowledge Management(San Francisco, California, USA)(CIKM ’13). Association for Computing Machinery, New York, NY, USA, ...

    2013

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.