Pith. sign in

REVIEW 4 major objections 5 minor 57 references

RecBase: a 1.5B model pretrained on item concept IDs matches or beats 7B LLMs in zero-shot recommendation

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

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

A from-scratch model that tokenizes items into hierarchical codes and predicts next-item codes reaches higher average zero-shot AUC on 8 datasets than LLM recommenders up to 7B parameters.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection RecBase has a plausible and promising recipe for recommendation-oriented pretraining, but the headline comparison rests on an unspecified evaluation protocol and one overclaimed consistency sentence. the 4 major comments →

arxiv 2509.03131 v1 pith:BYQUCIWM submitted 2025-09-03 cs.IR cs.LG

RecBase: Generative Foundation Model Pretraining for Zero-Shot Recommendation

Sashuai Zhou , Weinan Gan , Qijiong Liu , Ke Lei , Jieming Zhu , Hai Huang , Yan Xia , Ruiming Tang
show 2 more authors
Zhenhua Dong Zhou Zhao
This is my paper
classification cs.IR cs.LG
keywords zero-shot recommendationcross-domain generalizationautoregressive pretrainingitem tokenizationresidual quantizationcurriculum learningfoundation modelrecommendation ranking
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper argues that the standard recipe for foundation-model recommendation is pointing the wrong way: pretraining on natural language leaves a knowledge gap when the actual task is ranking items. RecBase instead pretrains a transformer from scratch on cross-domain user histories, where each item is replaced by a compact hierarchical concept ID learned from its text description. The result, if the paper's evaluation is accepted, is that a 1.5B model reaches an overall zero-shot AUC of 0.6063 on eight unseen datasets, above the strongest large-language-model baseline (Mistral, 0.5941), and a 0.3B version already beats several smaller LLMs. The broader stake is that recommendation ability can be trained directly from item co-occurrence signals and semantic embeddings, without requiring language understanding or item IDs limited to one domain.

Core claim

The paper's central claim is that a domain-agnostic recommendation foundation model can be obtained by: (1) mapping every item's text description into a shared hierarchy of four concept IDs via CL-VAE, a curriculum-learning-enhanced RQ-VAE; (2) training an autoregressive transformer (RecBase) to predict the next item's concept IDs in a user's chronological interaction history; and (3) ranking candidate items by the probability assigned to their concept IDs. On eight datasets not seen during training, the 1.5B version outperforms LLMs up to 7B in zero-shot ranking AUC, and fine-tuning on target-domain data further improves the same model. The paper treats the gap between language-centric pret

What carries the argument

CL-VAE (Curriculum Learning Enhanced RQ-VAE): a residual-quantized variational autoencoder whose four level-specific codebooks (size 2048 each) are trained stage by stage, with first-level reinitialization when codebook usage is low; it converts text embeddings of item descriptions into unified hierarchical concept IDs across domains. The second piece is the autoregressive transformer over those IDs, which predicts each item's ID bit by bit from the user's history. Together they create a shared compact vocabulary (20,000 tokens) for recommendation rather than language.

Load-bearing premise

The headline comparison assumes both RecBase and the LLM baselines rank the same candidate items under the same scoring protocol; if the negative candidates or scoring normalizations differ, the reported AUC gap is not established.

What would settle it

Run a controlled zero-shot comparison with identical candidate pools per user: score every candidate with RecBase's next-concept-ID probability and with Mistral's Yes-token softmax under the same history truncation. If RecBase's overall AUC does not exceed Mistral's, the central claim is refuted.

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

If this is right

  • Pretraining a transformer from scratch on item IDs can beat prompting a 7B LLM, so language fluency is not the source of zero-shot recommendation ability.
  • A small, recommendation-specialized vocabulary yields substantially lower inference cost than natural-language tokenization: the paper reports 295 seconds for RecBase-1.5B versus 946 seconds for Mistral-7B on 20,000 test interactions.
  • In-domain fine-tuning of the same base gives large gains, e.g., Steam AUC rises from 0.7450 to 0.9173, making the pretrained model a viable starting point for domain adaptation.
  • Codebook design matters: four levels of 2048 codewords balance collision and utilization, and adding levels beyond four brings diminishing returns while increasing decoding cost.

Where Pith is reading between the lines

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

  • Because the pretraining corpus includes news, book, and short-video interactions, three of the eight 'unseen' test datasets overlap in domain with pretraining; a truly unseen-domain test (e.g., music or finance) would test the domain-agnostic claim more strictly.
  • If the reported advantage survives matched candidate sets, it suggests item-level co-occurrence plus semantic embeddings capture most of the signal, and natural-language prompts add little for ranking.
  • Text-only preprocessing discards visual features in video and fashion datasets; quantizing multimodal embeddings into the same concept-ID space is a natural, testable extension.
  • Hierarchical concept IDs could enable constrained decoding or retrieval-style candidate filtering at inference, further cutting cost beyond what the paper measures.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes RecBase, a generative foundation model for recommendation. Item descriptions are first embedded with NV-Embed-v2 and discretized into hierarchical 4-level concept IDs via a curriculum-learned RQ-VAE (CL-VAE). A decoder-only transformer is then pretrained from scratch on user interaction sequences, with the objective of predicting the next concept ID. Pretraining uses 15 domains; zero-shot evaluation is conducted on eight RecBench datasets using AUC. The main claim is that RecBase-large (1.3B parameters) matches or exceeds LLM-based recommenders up to 7B parameters, with an overall AUC of 0.6063 versus Mistral's 0.5941 and with lower inference latency.

Significance. If the evaluation protocol is confirmed, the result is significant: it demonstrates that a compact, recommendation-specific tokenizer plus autoregressive pretraining can outperform much larger prompt-based LLM recommenders on zero-shot ranking. The CL-VAE contributions—curriculum learning over codebook levels and codebook reinitialization—are well motivated and supported by ablations. The paper also provides useful evidence on codebook size/level trade-offs and on the gains from in-domain fine-tuning. The main significance is conditional on the comparability of the AUC numbers, which the current manuscript does not fully establish.

major comments (4)
  1. [§4.1 / Table 1] The headline comparison requires that RecBase and the LLM baselines are evaluated on the same candidate-item set per user, but the paper never specifies how negative candidate items are drawn for RecBase's ranking. Equation (10) defines only the LLM score for a given user–item prompt; RecBase's inference is described as comparing probabilities of predicted concept IDs, without stating the candidate pool. If RecBase ranks all test items while the LLM baselines score a smaller, RecBench-sampled candidate set (or vice versa), the AUC values in Table 1 are not commensurable. This is a load-bearing omission for the central claim of superiority over LLM baselines. Please document the exact candidate construction for both RecBase and all baselines, or re-run the comparison under one identical protocol.
  2. [§4.1 / RecBench references] The baseline AUC numbers in Table 1 appear to be taken from RecBench (Liu et al., 2025a,b), which is co-authored by the present authors. The manuscript does not state whether these numbers were re-run for this paper or copied from the cited benchmark, nor does it give the split/user filtering details for each dataset. To make the comparison reproducible and verifiable, the authors should either report that the baseline numbers are quoted from the specific RecBench configuration (with dataset/split identifiers) or provide the commands/settings used to reproduce them. This is a reproducibility requirement, not an accusation.
  3. [Appendix A / §4.2 / Abstract] Several 'zero-shot' test domains overlap with pretraining domains: MIND (news) overlaps EBNeRD/PENS, MicroLens (short video) overlaps PixelRec/KuaiRec, and Goodreads (books) overlaps Amazon Books. Thus the evaluation is zero-shot with respect to datasets, but not necessarily with respect to domains. The abstract's 'cross-domain' claim should be qualified to state which evaluations are genuinely unseen-domain. In addition, RecBase uses NV-Embed-v2 to obtain item embeddings; the parameters of this external encoder are not counted in the reported 1.3B model size. The paper should state that the parameter comparison is for the RecBase core only, or provide a system-level comparison that includes the embedding model.
  4. [§5.1 / Table 4] The ablation table and the accompanying text require a clearer statement of the subset used to compute 'Overall.' In Table 1, RecBasebase's overall AUC is 0.5601; in Table 4 the same label has an overall value of 0.5879. If Table 4's Overall is computed only over Yelp, Steam, H&M, and HotelRec, this should be stated explicitly. This matters because the ablation conclusions are drawn from that Overall value and because the table includes the counterintuitive result that removing formatted text improves HotelRec (0.4966 vs. 0.4874), which also conflicts with the text's claim that RecBasebase achieves the best performance across all datasets.
minor comments (5)
  1. [§4.3 / Table 1] The sentence 'RecBase consistently outperforms advanced LLM-based models such as Llama-3 and Qwen-2 across all datasets' is contradicted by the HotelRec row: RecBaselarge is 0.5124, while Llama-3 is 0.5342 and Qwen-2 is 0.5311. Please qualify the claim, e.g., 'on the majority of datasets' or 'in terms of overall AUC.'
  2. [§5.2 / Figure 2] The text says 'As shown in Figure 2, with an increase in size...', but Figure 2 is the t-SNE visualization. The codebook size/level discussion refers to Figure 5 (a) and (b). Please correct the cross-reference.
  3. [Algorithm 1] The curriculum criterion is only 'if total loss converges,' and the codebook reinitialization threshold is not defined. These are important hyperparameters for reproducibility; please specify the convergence criterion and the usage-rate threshold for reinitialization.
  4. [§5.4 / Figure 6] The inference-time comparison lacks hardware, batch size, sequence-length distribution, and whether the reported time includes prompt encoding for the LLM baselines. Adding these details would make the efficiency claim reproducible.
  5. [§4.2 / Table 2] The vocabulary size is stated as 20,000, while the text elsewhere describes a 4-level codebook with 2048 codes per level. It would help to state explicitly that the 20,000 vocabulary is the union of the four codebooks plus special tokens, since the relationship is otherwise unclear.

Circularity Check

0 steps flagged

No significant circularity: the central comparison is against external held-out datasets and the modeling chain does not reduce to its own inputs.

full rationale

RecBase's derivation chain is: (1) map item text to embeddings with NV-Embed-v2; (2) quantize embeddings into hierarchical concept IDs via CL-VAE; (3) autoregressively predict next concept-ID bits; (4) rank test items by predicted ID probabilities and evaluate AUC on RecBench. None of these steps defines a target quantity in terms of the fitted model's own output: the tokenizer is trained on reconstruction/entropy losses, and the pretraining objective (Eq. 9) is a standard next-token NLL. The headline zero-shot numbers are measured on eight externally defined datasets; the fact that some test domains (news, video, books) overlap with pretraining domains weakens the 'zero-shot' label but is a data-splitting validity concern, not a circular derivation. RecBench is a published benchmark with fixed prompts and scoring (Eq. 10), so its baseline figures are independently reproducible; co-authorship of the benchmark and the present paper does not make the comparison circular. The paper's own internal evidence (Figure 2, Figure 5c) is self-referential but is not used to derive the AUC numbers. The only near-issue is an overstatement in Section 4.3 ('consistently outperforms ... across all datasets') contradicted by Table 1 on HotelRec, but that is a correctness slip, not circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 1 invented entities

The headline zero-shot result rests on a chain of learned components (NV-Embed-v2 embeddings, CL-VAE codebooks, roughly 1.3B transformer) whose training has at least six unstated or ablation-chosen hyperparameters (gamma, beta, curriculum schedule, reinit threshold, K, m), and the evaluation's candidate set is unspecified. This is neither circular nor fabricated, but it means the contribution is not reproducible from the text alone.

free parameters (8)
  • codebook size K = 2048 per level
    Selected by ablation (Figure 5a) on test datasets Yelp/Steam/H&M/HotelRec, not derived.
  • number of codebook levels m = 4
    Selected by ablation (Figure 5b) on the same test datasets; trades representation capacity against vocabulary utilization.
  • entropy loss weight gamma = not reported
    Appears in Eq. 4 (L = L_R + L_Q + gamma L_E) but no value is given; exact training unreproducible.
  • commitment loss weight beta = not reported
    Appears in Eq. 6 but no value is given.
  • curriculum stage schedule = not reported
    Algorithm 1 adds codebook levels when "Total loss converges"; convergence criterion and per-stage epochs undefined.
  • codebook reinitialization threshold = not reported
    Algorithm 1 reinitializes C0 with KMeans when "codebook usage rate is low"; "low" is undefined.
  • user history filters = remove users with <15 interactions; truncate histories >2500
    Data preprocessing choices (Appendix B.2) that shape the training distribution.
  • evaluation candidate set for AUC = unspecified
    The number and selection of negative candidates for the zero-shot AUC is never stated; the headline comparison depends on it.
axioms (6)
  • domain assumption NV-Embed-v2 text embeddings capture item-level semantics sufficient for cross-domain recommendation
    Section 4.2 uses these embeddings as the sole item representation; if text fails to model dynamic item-level user interests, the pipeline loses its substrate.
  • domain assumption RQ-VAE residual quantization preserves recommendation-relevant item information
    Section 3.2 maps continuous embeddings to 4 concept IDs; information loss at this step bounds what the autoregressive model can predict.
  • domain assumption Autoregressive next-code prediction on concept ID sequences is a valid objective for learning user interest
    Section 3.3 defines the only training signal; no evidence is given that code-level next-token likelihood ranks items by user interest.
  • domain assumption Codebooks and token distributions trained on pretraining domains transfer to test items without adaptation
    The zero-shot evaluation (Section 4.3) assumes new-domain items map to well-used codes and are rankable by the pretrained LM.
  • domain assumption RecBench AUC scores across methods are commensurable (same candidate items, same protocol)
    Section 4.1 describes per-method scoring but never states the candidate construction shared by RecBase and LLM baselines.
  • standard math Standard ML machinery (softmax, stop-gradient, VQ losses) behaves as expected
    Incidental to the method; not load-bearing beyond standard practice.
invented entities (1)
  • Unified hierarchical concept ID space (CL-VAE codebook, 4 levels of 2048 codes) no independent evidence
    purpose: Shared item tokenization across domains enabling a compact vocabulary and autoregressive next-code prediction
    The concept ID space exists only inside this paper's pipeline; its quality is evidenced by the paper's own AUC results and internal t-SNE/usage plots, with no external falsifiable handle.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RecBase: Generative Foundation Model Pretraining for Zero-Shot Recommendation." pith.science (2026). https://pith.science/paper/BYQUCIWM

@misc{pith2026250903131,
  author       = {Pith},
  title        = {Pith review of: RecBase: Generative Foundation Model Pretraining for Zero-Shot Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BYQUCIWM}},
  note         = {Machine review of arXiv:2509.03131}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advances in LLM-based recommendation have shown promise, yet their cross-domain generalization is hindered by a fundamental mismatch between language-centric pretraining and the recommendation task. Existing methods, relying on language-level knowledge, fail to capture dynamic, item-level user interests across domains. To bridge this gap, we propose RecBase, a domain-agnostic foundational model pretrained with a recommendation-oriented objective. RecBase leverages a large-scale, heterogeneous, cross-domain corpus with unified textual representations and feature mappings to enhance cross-domain generalization. To further align item semantics across domains, we introduce a unified item tokenizer that encodes items into hierarchical concept identifiers, enabling structured representation and efficient vocabulary sharing. The model is trained using an autoregressive objective to capture complex item-level sequential patterns. On eight real-world datasets, our 1.5B-parameter model matches or surpasses the performance of LLM baselines up to 7B parameters in zero-shot and cross-domain recommendation tasks.

Figures

Figures reproduced from arXiv: 2509.03131 by Hai Huang, Jieming Zhu, Ke Lei, Qijiong Liu, Ruiming Tang, Sashuai Zhou, Weinan Gan, Yan Xia, Zhenhua Dong, Zhou Zhao.

Figure 1
Figure 1. Figure 1: Overview of the RecBase model. (a) illustrates the use of discrete representation techniques to transform [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of t-SNE Clustering in the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the zero-shot transfer and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training and test datasets distribution [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Codebook structure vs. Collision rate and Uti [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of Inference Latency. 6 Conclusion In this paper, we introduce RecBase, a foundation model tailored to the challenges of zero-shot and multi-domain recommendation. By pretraining on a large-scale cross-domain corpus with structured text representations and unified feature mappings, RecBase demonstrates strong generalization across heterogeneous recommendation tasks. The incor￾poration of curricu… view at source ↗

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

57 extracted references · 28 canonical work pages · 5 internal anchors

  1. [1]

    Diego Antognini and Boi Faltings. 2020. https://aclanthology.org/2020.lrec-1.605/ H otel R ec: a novel very large-scale hotel recommendation dataset . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4917--4923, Marseille, France. European Language Resources Association

    work page 2020

  2. [2]

    Xiang Ao, Xiting Wang, Ling Luo, Ying Qiao, Qing He, and Xing Xie. 2021. https://doi.org/10.18653/v1/2021.acl-long.7 PENS : A dataset and generic framework for personalized news headline generation . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Pr...

    work page doi:10.18653/v1/2021.acl-long.7 2021

  3. [3]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. https://doi.org/10.1145/3604915.3608857 Tallrec: An effective and efficient tuning framework to align large language model with recommendation . In Proceedings of the 17th ACM Conference on Recommender Systems, RecSys 2023, Singapore, Singapore, September 18-22, 2023 , page...

    arXiv 2023

  4. [4]

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41--48

    work page 2009

  5. [5]

    Brown, Benjamin Mann, Nick Ryder, and et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, and et al. 2020. https://proceedings.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html Language models are few-shot learners . In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

    work page 2020

  6. [6]

    Yu Cheng, Yunzhu Pan, Jiaqi Zhang, Yongxin Ni, Aixin Sun, and Fajie Yuan. 2024. An image dataset for benchmarking recommender systems with raw pixels. In Proceedings of the 2024 SIAM International Conference on Data Mining (SDM), pages 418--426. SIAM

    work page 2024

  7. [7]

    Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. https://doi.org/10.48550/ARXIV.2205.08084 M6-rec: Generative pretrained language models are open-ended recommender systems . CoRR, abs/2205.08084

    Pith/arXiv arXiv doi:10.48550/arxiv.2205.08084 2022

  8. [8]

    Jacob Devlin, Ming - Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/V1/N19-1423 BERT: pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, ...

    doi:10.18653/v1/n19-1423 2019

  9. [9]

    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

    Pith/arXiv arXiv 2024

  10. [10]

    Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, and Tat-Seng Chua. 2022. Kuairec: A fully-observed dataset and insights for evaluating recommender systems. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 540--550

    work page 2022

  11. [11]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. https://doi.org/10.1145/3523227.3546767 Recommendation as language processing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5) . In RecSys '22: Sixteenth ACM Conference on Recommender Systems, Seattle, WA, USA, September 18 - 23, 2022 , pages 299--315. ACM

    arXiv 2022

  12. [12]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. 2015. https://doi.org/10.1145/2827872 The movielens datasets: History and context . ACM Trans. Interact. Intell. Syst., 5(4)

    doi:10.1145/2827872 2015

  13. [13]

    Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David A. Sontag. 2023. https://proceedings.mlr.press/v206/hegselmann23a.html Tabllm: Few-shot classification of tabular data with large language models . In International Conference on Artificial Intelligence and Statistics, 25-27 April 2023, Palau de Congressos, Valencia...

    work page 2023

  14. [14]

    Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, Ye Wang, Sihang Cai, Xiaoda Yang, Quanyu Dai, Zhenhua Dong, Zhimeng Zhang, and Zhou Zhao. 2025. EAGER-LLM: enhancing large language models as recommenders through exogenous behavior-semantic integration. In WWW, pages 2754--2762

    work page 2025

  15. [15]

    Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. 2024 a . Bridging language and items for retrieval and recommendation. arXiv preprint arXiv:2403.03952

    Pith/arXiv arXiv 2024

  16. [16]

    McAuley, and Wayne Xin Zhao

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian J. McAuley, and Wayne Xin Zhao. 2024 b . Large language models are zero-shot rankers for recommender systems. In Advances in Information Retrieval - 46th European Conference on Information Retrieval (ECIR), volume 14609, pages 364--381

    work page 2024

  17. [17]

    Chengkai Huang, Tong Yu, Kaige Xie, Shuai Zhang, Lina Yao, and Julian J. McAuley. 2024. Foundation models for recommender systems: A survey and new perspectives. CoRR, abs/2402.11143

    Pith/arXiv arXiv 2024

  18. [18]

    Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. 2023. https://doi.org/10.48550/ARXIV.2307.00457 Genrec: Large language model for generative recommendation . CoRR, abs/2307.00457

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2307.00457 2023

  19. [19]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L \' e lio Renard Lavaud, Marie - Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timoth \' e e Lacroix, and William El Sayed. 2023. https://doi.org/...

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.06825 2023

  20. [20]

    Johannes Kruse, Kasper Lindskow, Saikishore Kalloori, Marco Polignano, Claudio Pomo, Abhishek Srivastava, Anshuk Uppal, Michael Riis Andersen, and Jes Frellsen. 2024. https://doi.org/10.1145/3687151.3687152 Eb-nerd a large-scale dataset for news recommendation . In Proceedings of the Recommender Systems Challenge 2024, RecSys Challenge ’24, page 1–11. ACM

    work page arXiv 2024

  21. [21]

    Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. 2024. https://doi.org/10.48550/ARXIV.2405.17428 Nv-embed: Improved techniques for training llms as generalist embedding models . CoRR, abs/2405.17428

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.17428 2024

  22. [22]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. https://arxiv.org/abs/2203.01941 Autoregressive image generation using residual quantization . Preprint, arXiv:2203.01941

    Pith/arXiv arXiv 2022

  23. [23]

    Xingchen Li, Xiang Wang, Xiangnan He, Long Chen, Jun Xiao, and Tat - Seng Chua. 2020. https://doi.org/10.1145/3397271.3401080 Hierarchical fashion graph network for personalized outfit recommendation . In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July ...

    arXiv 2020

  24. [24]

    Yu, and Kai Shu

    Yueqing Liang, Liangwei Yang, Chen Wang, Xiongxiao Xu, Philip S. Yu, and Kai Shu. 2025. Taxonomy-guided zero-shot recommendations with llms. In Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025 , pages 1520--1530

    work page 2025

  25. [25]

    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. In Proceedings of the ACM on Web Conference 2024 (WWW) , pages 3319--3330

    work page 2024

  26. [26]

    Junling Liu, Chao Liu, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is chatgpt a good recommender? A preliminary study. CoRR, abs/2304.10149

    Pith/arXiv arXiv 2023

  27. [27]

    Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024 a . Once: Boosting content-based recommendation with both open- and closed-source large language models. In WSDM

    work page 2024

  28. [28]

    Qijiong Liu, Jieming Zhu, Lu Fan, Kun Wang, Hengchang Hu, Wei Guo, Yong Liu, and Xiao-Ming Wu. 2025 a . Benchmarking llms in recommendation tasks: A comparative evaluation with conventional recommenders. arXiv preprint arXiv:2503.05493

    arXiv 2025

  29. [29]

    Qijiong Liu, Jieming Zhu, Yingxin Lai, Xiaoyu Dong, Lu Fan, Zhipeng Bian, Zhenhua Dong, and Xiao-Ming Wu. 2025 b . Evaluating recabilities of foundation models: A multi-domain, multi-dataset benchmark. arXiv preprint arXiv:2508.21354

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Qijiong Liu, Jieming Zhu, Yanting Yang, Quanyu Dai, Zhaocheng Du, Xiao - Ming Wu, Zhou Zhao, Rui Zhang, and Zhenhua Dong. 2024 b . Multimodal pretraining, adaptation, and generation for recommendation: A survey. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) , pages 6566--6576

    work page 2024

  31. [31]

    Yunshan Ma, Yingzhi He, An Zhang, Xiang Wang, and Tat - Seng Chua. 2022. https://doi.org/10.1145/3534678.3539229 Crosscbr: Cross-view contrastive learning for bundle recommendation . In KDD '22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022 , pages 1233--1241. ACM

    work page arXiv 2022

  32. [32]

    Shervin Minaee, Tom \' a s Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. 2024. Large language models: A survey. CoRR, abs/2402.06196

    Pith/arXiv arXiv 2024

  33. [33]

    Sanmit Narvekar, Bei Peng, Matteo Leonetti, Jivko Sinapov, Matthew E Taylor, and Peter Stone. 2020. Curriculum learning for reinforcement learning domains: A framework and survey. Journal of Machine Learning Research, 21(181):1--50

    2020

  34. [34]

    Yongxin Ni, Yu Cheng, Xiangyan Liu, Junchen Fu, Youhua Li, Xiangnan He, Yongfeng Zhang, and Fajie Yuan. 2023. A content-driven micro-video recommendation dataset at scale. arXiv preprint arXiv:2309.15379

    Pith/arXiv arXiv 2023

  35. [35]

    Bo Peng, Srinivasan Parthasarathy, and Xia Ning. 2022. https://doi.org/10.48550/ARXIV.2209.07997 Recursive attentive methods with reused item representations for sequential recommendation . CoRR, abs/2209.07997

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2209.07997 2022

  36. [36]

    Petrov and Craig Macdonald

    Aleksandr V. Petrov and Craig Macdonald. 2023. https://www.paperdigest.org/paper/?paper_id=arxiv-2306.11114 Generative sequential recommendation with gptrec . ARXIV-CS.IR

    Pith/arXiv arXiv 2023

  37. [37]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/20dcab0f14046a5c6b02b61da9f13229-Abstract-Conference.html Recommender systems with generative retrieval...

    work page 2023

  38. [38]

    Hugo Touvron, Louis Martin, Kevin Stone, and et al. 2023. https://doi.org/10.48550/ARXIV.2307.09288 Llama 2: Open foundation and fine-tuned chat models . CoRR, abs/2307.09288

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.09288 2023

  39. [39]

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2018. https://arxiv.org/abs/1711.00937 Neural discrete representation learning . Preprint, arXiv:1711.00937

    Pith/arXiv arXiv 2018

  40. [40]

    Mengting Wan, Rishabh Misra, Ndapa Nakashole, and Julian McAuley. 2019. https://doi.org/10.18653/v1/P19-1248 Fine-grained spoiler detection from large-scale review corpora . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2605--2610, Florence, Italy. Association for Computational Linguistics

    doi:10.18653/v1/p19-1248 2019

  41. [41]

    Shoujin Wang, Longbing Cao, Liang Hu, Shlomo Berkovsky, Xiaoshui Huang, Lin Xiao, and Wenpeng Lu. 2021 a . https://doi.org/10.1109/MIS.2020.2997362 Hierarchical attentive transaction embedding with intra- and inter-transaction dependencies for next-item recommendation . IEEE Intell. Syst. , 36(4):56--64

    work page arXiv 2021

  42. [42]

    Xin Wang, Yudong Chen, and Wenwu Zhu. 2021 b . A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555--4576

    work page 2021

  43. [43]

    Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, and Zhenhua Dong. 2024 a . EAGER: two-stream generative recommender with behavior-semantic collaboration. In KDD, pages 3245--3254

    work page 2024

  44. [44]

    Yidan Wang, Zhaochun Ren, Weiwei Sun, Jiyuan Yang, Zhixiang Liang, Xin Chen, Ruobing Xie, Su Yan, Xu Zhang, Pengjie Ren, Zhumin Chen, and Xin Xin. 2024 b . https://doi.org/10.1145/3627673.3679692 Content-based collaborative generation for recommender systems . In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management,...

    work page arXiv 2024

  45. [45]

    Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, and Ming Zhou. 2020. https://doi.org/10.18653/v1/2020.acl-main.331 MIND : A large-scale dataset for news recommendation . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3597--3606, Onli...

    doi:10.18653/v1/2020.acl-main.331 2020

  46. [46]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A survey on large language models for recommendation. World Wide Web (WWW) , 27(5):60

    work page 2024

  47. [47]

    An Yang, Baosong Yang, Binyuan Hui, and et al. 2024. https://doi.org/10.48550/ARXIV.2407.10671 Qwen2 technical report . CoRR, abs/2407.10671

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.10671 2024

  48. [48]

    Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T. Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2022. https://doi.org/10.48550/ARXIV.2205.01068 OPT: open pre-trained transfor...

    Pith/arXiv arXiv doi:10.48550/arxiv.2205.01068 2022

  49. [49]

    Yabin Zhang, Wenhui Yu, Erhan Zhang, Xu Chen, Lantao Hu, Peng Jiang, and Kun Gai. 2024. https://doi.org/10.48550/ARXIV.2404.08675 Recgpt: Generative personalized prompts for sequential recommendation via chatgpt training paradigm . CoRR, abs/2404.08675

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2404.08675 2024

  50. [50]

    Zizhuo Zhang and Bang Wang. 2023. https://doi.org/10.1145/3539618.3591752 Prompt learning for news recommendation . In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023 , pages 227--237. ACM

    arXiv 2023

  51. [51]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji - Rong Wen. 2024. https://doi.org/10.1109/ICDE60146.2024.00118 Adapting large language models by integrating collaborative semantics for recommendation . In 40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, 2024 , pages 1435...

    arXiv 2024

  52. [52]

    Zhi Zheng, Zhaopeng Qiu, Xiao Hu, Likang Wu, Hengshu Zhu, and Hui Xiong. 2023. https://www.paperdigest.org/paper/?paper_id=arxiv-2307.02157 Generative job recommendations with large language model . ARXIV-CS.IR

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  53. [53]

    Jieming Zhu, Quanyu Dai, Liangcai Su, Rong Ma, Jinyang Liu, Guohao Cai, Xi Xiao, and Rui Zhang. 2022. BARS: towards open benchmarking for recommender systems. In The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) , pages 2912--2923

    work page 2022

  54. [54]

    Jieming Zhu, Mengqun Jin, Qijiong Liu, Zexuan Qiu, Zhenhua Dong, and Xiu Li. 2024 a . Cost: Contrastive quantization based semantic tokenization for generative recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems ( RecSys) , pages 969--974

    work page 2024

  55. [55]

    Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024 b . https://www.paperdigest.org/paper/?paper_id=www-3589334.3645347-2024-05-13 Collaborative large language model for recommender systems . WWW

    arXiv 2024

  56. [56]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  57. [57]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.