Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Representing each item as a single dedicated embedding in an LLM, instead of a multi-token text string, improves recommendation quality and cuts inference latency by an estimated 5x–14x.

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

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

T0 review · deepseek-v4-flash

2026-08-05 10:42 UTC pith:454HYVNY

load-bearing objection Single-token item decoding is a real, mostly-supported advance; the 5-14x speedup is an estimate, not a measurement, and the quality claim needs a footnote on Arts. the 3 major comments →

arxiv 2509.03746 v1 pith:454HYVNY submitted 2025-09-03 cs.IR

Efficient Item ID Generation for Large-Scale LLM-based Recommendation

Anushya Subbiah , Vikram Aggarwal , James Pine , Steffen Rendle , Krishna Sayana , Kun Su This is my paper
classification cs.IR
keywords LLM-based recommendationsingle-token item representationTwo-Level Softmaxgenerative retrievalsequential recommendationinference latencyitem catalogAmazon Review dataset
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.

Most LLM-based recommenders represent an item as several text tokens, forcing the model to decode a title or code one token at a time and to read long prompts built from multi-token histories. This paper argues the opposite: each item gets one direct embedding vector, treated like a first-class token alongside text, and the model predicts the next item in a single decoding step. With a Two-Level Softmax that splits the item catalog into roughly the square root of its size in clusters, training stays tractable at millions of items and inference uses approximate nearest neighbor search or cluster pruning. On four Amazon shopping datasets the method matches or beats the multi-token baselines on Recall and NDCG, while a latency model estimates 5x–14x faster inference. The practical import is that real-time, large-catalog LLM recommendation may not require abandoning LLMs or accepting slow multi-step generation.

Core claim

The central claim is that item IDs should be first-class tokens of the LLM, with their own trainable embedding table, and that single-step decoding from one output embedding is both feasible and superior to the multi-token, multi-step pattern dominant in the literature. The paper introduces a training paradigm in which 25% of examples use only item IDs for input and label, so the model learns to predict from IDs alone at inference, removing the need for separate alignment or item-pretraining stages. For catalogs where a full softmax over all items is impractical, the Two-Level Softmax factorizes the output distribution as cluster probability times within-cluster probability, lowering trainin

What carries the argument

The load-bearing mechanism is the Two-Level Softmax over an extended vocabulary of text tokens plus items. Items get dedicated embedding vectors trained end-to-end and are assigned to clusters with centroid embeddings; the output probability factorizes as P(w|H) = P(c(w)|H) · P(w|c(w),H). This factorization does two jobs: training only needs the target's cluster, cutting the softmax denominator to O(|V|+sqrt(|I|)), and inference can either search via approximate nearest neighbors over indexed vectors e_c(w)+e_w, or prune whole clusters using the upper bound P(w|H) ≤ P(c(w)|H). The companion trick is the training schedule that feeds 25% ID-only examples, making single-token ID-only inference

Load-bearing premise

The efficiency advantage rests on a latency model that assumes prefill cost grows linearly with token count and uses published per-token prefill/decode times from Mistral-7B and PaLM; real serving with batching, prefill caching, or amortized decoding could shrink the estimated 5x–14x speedup.

What would settle it

Run both designs on the same production-style GPU serving stack with realistic batching and prefill caching and measure end-to-end wall-clock latency per recommendation on a million-item catalog; if the single-token method is less than roughly 2x faster than a comparable multi-token LLM recommender, the paper's core efficiency claim fails. On the quality side, train a strong multi-token baseline with the same compute and data and compare Recall@10 and NDCG@10; the paper predicts the single-token method will still match or beat it.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • LLM-based recommenders can serve real-time requests: one decode step per recommendation instead of one per token of a title or category.
  • Small-to-medium catalogs can use Full Softmax directly, while catalogs of millions of items remain tractable with the Two-Level Softmax plus ANN or hierarchical pruning.
  • The 25% ID-only training mix lets the model exploit rich metadata during training while depending only on item IDs at inference, eliminating separate alignment stages.
  • Larger backbone LLMs improve recommendation quality, so the efficiency gain matters more as these models scale.
  • If the latency model holds, deployments with published prefill and decode characteristics would see 5x–14x end-to-end speedups for comparable or better recommendation quality.

Where Pith is reading between the lines

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

  • Editorial inference: the 5x–14x figure is an estimate built on a linear prefill model and published per-token latencies; real serving with prefill caching, batching, or shared KV caches could narrow or widen the gap.
  • Editorial inference: the same single-token treatment should transfer to other discrete entities—users, documents, videos, knowledge-graph nodes—whenever an LLM must emit one object from a huge catalog.
  • Editorial inference: the ANN variant drops the per-cluster log-normalizer, approximating the two-level score by the dot product of the output embedding with e_c(w)+e_w; the ablation showing random clustering nearly matches KMeans for ANN suggests the approximation depends mainly on cluster balance, not semantic structure.
  • Editorial inference: the Amazon datasets used here have 10k–38k items, so the paper does not directly demonstrate the million-item regime; a natural test would be to measure ANN retrieval latency and quality on a genuinely million-item catalog.

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

3 major / 5 minor

Summary. The paper challenges the common practice of representing item IDs in LLM-based recommenders as multi-token text sequences. It proposes to treat each item as a single learnable embedding, to be projected into the LLM's embedding space, and to generate the next item in a single decoding step. For scalable training and inference, a Two-Level softmax is introduced: items are partitioned into clusters, a first softmax selects a cluster and a second softmax selects an item within the cluster. Two inference methods are proposed: hierarchical structure-based pruning and ANN search over cluster-centroid-plus-item vectors. A training paradigm mixes full metadata with a fixed fraction (25%) of ID-only examples, so that inference can rely on IDs alone. Experiments on four Amazon review datasets report quality comparable or better than CALRec and other baselines, and a latency model is used to claim a 5x-14x inference speedup over multi-token decoding.

Significance. If the empirical claims hold, the paper makes a useful and potentially influential point: the dominant multi-token approach to item IDs in LLM recommenders is not necessary, and direct single-vector item representations can improve both quality and latency. The Two-Level softmax with ANN inference is a sensible adaptation of hierarchical softmax to large catalogs, and the paper is transparent that the speedup analysis is based on an assumed latency model. The release of code is a plus. However, the central quality claim is only partially supported by the reported numbers, and the headline efficiency gain is not measured. These gaps are load-bearing because the paper's main contributions are exactly the quality and efficiency advantages over multi-token methods. The empirical core is plausible but needs strengthening before the claims can be accepted as stated.

major comments (3)
  1. [§6.2.1, Table 4] The claim that "Full Softmax outperforms" the multi-token LLM baselines is contradicted on the Arts dataset: Full Softmax Recall@1 is 0.0626 vs. CALRec's 0.0636, and MRR is 0.0780 vs. 0.0815. On Instruments, Recall@1 ties at 0.0718. Since the baselines' numbers are taken from CALRec [21] rather than re-run in the same harness, and since no error bars, seeds, or significance tests are reported, the conclusion that single-token decoding improves quality is not established uniformly. Please re-run the baselines under the same protocol or substantially temper the quality claim.
  2. [§5.1, Figure 3] The headline 5x-14x speedup is a synthetic estimate, not a measured latency improvement. It depends on the linear-prefill assumption l_prefill(n)=n·l_prefill and on external per-token latency constants from Mistral-7B and PaLM deployments. Modern serving stacks with chunked prefill, batching, or amortized prefill can make the actual speedup much smaller. The abstract and conclusion state 5x-14x without direct latency validation. I recommend either reporting wall-clock end-to-end latency measurements under a realistic serving configuration, or clearly labeling the range as a model-based upper-bound estimate and removing it from the abstract-level claims.
  3. [§3.4, §6.2.5] The training paradigm is a central contribution, but the 25% ID-only mixing ratio is described as "crucial" without any sensitivity analysis. No ablation varies this ratio or the metadata sampling probability, so the reader cannot tell how robust the ID-only inference behavior is to this hyperparameter. Since the paper argues that this paradigm removes multi-stage pretraining, a small ablation or at least a statement of the ratio's effect is needed to support that claim.
minor comments (5)
  1. [Figure 3] The right panel's labels (e.g., "1.9 16 77") are visually crowded and hard to read; the caption also lists "LLM: Upper bound" and "LLM: Lower bound" before naming PaLM and Mistral, which is confusing. Please clarify the legend and the bound notation.
  2. [§6.1.2] The paper does not state how many random seeds are averaged for the reported numbers, nor the variance. Reporting standard deviations or at least multiple seeds would strengthen the quality comparisons.
  3. [Table 4 caption] The caption says baseline results are taken from CALRec [21]. This is transparent, but please also specify whether the same data split and evaluation code were used; otherwise the comparison is difficult to interpret.
  4. [§4.3.2] The ANN approximation is introduced as dropping the cluster-dependent log-partition term. It would be helpful to state explicitly that this is an approximation and to quantify its effect, perhaps by reporting the agreement rate between ANN and exact Structure inference on the test sets. Table 5 partially addresses this, but not in an exact-match form.
  5. [§6.1.1] Minor typo: "The Amazon Review dataset is popular and widely[19, 21, 26] used" is missing a space before the citation bracket.

Circularity Check

0 steps flagged

No significant circularity: quality results come from a standard trained head with external baselines, and the speedup is an explicitly modeled estimate, not a fit.

full rationale

The paper's central claims are not circular. The single-token item representation, Full Softmax (Eq. 1), and Two-Level softmax (Eqs. 2-4) are explicit model definitions, not quantities fitted to the reported outcomes. The ANN inference is presented as an approximation and validated by held-out quality ablations (Table 5), not by construction. The 5x-14x speedup is computed in Section 5.1 from an explicit latency model with external per-token constants from Sarathi-Serve/Mistral and PaLM, using measured prefill/decode token counts; it is transparently an estimate, not a statistic fitted to the claimed speedup. Baseline results are taken from CALRec [21], an external non-overlapping work, and the only self-citation of note (Rendle et al. [27]) is used for an optional K-means initialization that is ablated against Frequency and Random clustering, both of which remain competitive, so no load-bearing claim depends on it. No uniqueness theorem, ansatz, or definitional equivalence is imported from the authors' prior work. The derivation chain is self-contained or externally grounded.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claim rests on the two-level softmax factorization, fixed precomputed clusters, a linear prefill latency model, and an empirical approximation in ANN inference. No new physical entities are introduced; the free parameters are standard hyperparameters plus the ID-only training ratio and cluster count.

free parameters (6)
  • Item embedding dimension k = 512
    Chosen by hand; separate from LLM text embedding dimension; balances capacity and storage.
  • ID-only training ratio = 25%
    Chosen by hand; deliberately forces model to handle ID-only inference; critical to quality at test time.
  • Number of clusters = approximately sqrt(|I|)
    Chosen to trade off cluster-level and item-level softmax cost; controls training/inference complexity.
  • Learning rate = 5e-3 (weight decay 1e-5; 1e-2 for PaLM 2 XS)
    Standard hyperparameter tuned for the setup; affects convergence and final quality.
  • Batch size = 512
    Standard training hyperparameter; chosen for throughput.
  • Max training steps = 50000
    Stopping criterion; quality may vary with more or less training.
axioms (6)
  • standard math Hierarchical softmax factorization P(w|H)=P(c(w)|H) P(w|c(w),H) is a valid probability decomposition.
    Used in Eq. 2-4; follows from the chain rule of probability.
  • domain assumption Cluster assignments c are precomputed and fixed during training; item embeddings change but clusters do not.
    Section 4.2 and 6.1.2; if clusters drift from item embeddings, the factorization becomes a poor approximation.
  • domain assumption The iALS item embeddings used for Kmeans clustering are trained without test-interaction leakage.
    Section 6.2.3; not stated whether the external MF model sees the held-out last interaction, which could leak test labels into cluster structure.
  • domain assumption Prefill latency is linear in token count for the speedup bounds (l_prefill(n)=n*l_prefill).
    Section 5.1; used to derive upper/lower bounds and the 5x-14x numbers; actual prefill may be nonlinear under batching or KV-cache reuse.
  • domain assumption ANN search (ScaNN/FAISS) returns near-exact top-k with negligible quality loss.
    Section 4.3.2; approximated log-partition term and ANN indexing; validated only empirically on the four Amazon subsets.
  • domain assumption A small projection head can align randomly initialized item embeddings with the pretrained LLM embedding space without damaging the learned text representation.
    Section 3.2; end-to-end training assumption, not proven.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 16626 in / 15803 out tokens · 143376 ms · 2026-08-05T10:42:05.189213+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Efficient Item ID Generation for Large-Scale LLM-based Recommendation." pith.science (2026). https://pith.science/paper/454HYVNY

@misc{pith2026250903746,
  author       = {Pith},
  title        = {Pith review of: Efficient Item ID Generation for Large-Scale LLM-based Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/454HYVNY}},
  note         = {Machine review of arXiv:2509.03746}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Integrating product catalogs and user behavior into LLMs can enhance recommendations with broad world knowledge, but the scale of real-world item catalogs, often containing millions of discrete item identifiers (Item IDs), poses a significant challenge. This contrasts with the smaller, tokenized text vocabularies typically used in LLMs. The predominant view within the LLM-based recommendation literature is that it is infeasible to treat item ids as a first class citizen in the LLM and instead some sort of tokenization of an item into multiple tokens is required. However, this creates a key practical bottleneck in serving these models for real-time low-latency applications. Our paper challenges this predominant practice and integrates item ids as first class citizens into the LLM. We provide simple, yet highly effective, novel training and inference modifications that enable single-token representations of items and single-step decoding. Our method shows improvements in recommendation quality (Recall and NDCG) over existing techniques on the Amazon shopping datasets while significantly improving inference efficiency by 5x-14x. Our work offers an efficiency perspective distinct from that of other popular approaches within LLM-based recommendation, potentially inspiring further research and opening up a new direction for integrating IDs into LLMs. Our code is available here https://drive.google.com/file/d/1cUMj37rV0Z1bCWMdhQ6i4q4eTRQLURtC

Figures

Figures reproduced from arXiv: 2509.03746 by Anushya Subbiah, James Pine, Krishna Sayana, Kun Su, Steffen Rendle, Vikram Aggarwal.

Figure 1
Figure 1. Figure 1: This figure highlights the key distinction between baselines and our technique’s [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: This figure illustrates the efficiency gains of our Two-Level softmax during training and inference. For training, only the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Latency comparison: Left and middle plots show the number of decode steps / number of prefill tokens for single-token [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] 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

40 extracted references · 13 canonical work pages

  1. [1]

    Amey Agrawal, Nitin Kedia, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav Gulavani, Alexey Tumanov, and Ramachandran Ramjee. 2024. Taming {Throughput-Latency} tradeoff in{LLM} inference with{Sarathi-Serve}. In 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24). 117–134. https://dl.acm.org/doi/10.5555/3691938.3691945

    work page arXiv 2024

  2. [2]

    Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023). arXiv:2305.10403 [cs.CL] http://arxiv.org/abs/2305.10403

    Pith/arXiv arXiv 2023

  3. [3]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems (RecSys ’16) . Association for Computing Machinery, New York, NY, USA, 191–198. doi:10.1145/2959100.2959190

    arXiv 2016

  4. [4]

    Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. 2024. A Review of Modern Recommender Systems Using Generative Models (Gen-RecSys). In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24) . Association...

    arXiv 2024

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. doi:10.48550/ARXIV.1810.04805

    Pith/arXiv arXiv doi:10.48550/arxiv.1810.04805 2018

  6. [6]

    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). doi:10.48550/ARXIV.2203.13366

    Pith/arXiv arXiv doi:10.48550/arxiv.2203.13366 2022

  7. [7]

    Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, and Sanjiv Kumar. 2020. Accelerating Large-Scale Inference with Anisotropic Vector Quantization. doi:10.48550/arXiv.1908.10396 arXiv:1908.10396

    Pith/arXiv arXiv doi:10.48550/arxiv.1908.10396 2020

  8. [8]

    Jesse Harte, Wouter Zorgdrager, Panos Louridas, Asterios Katsifodimos, Dietmar Jannach, and Marios Fragkoulis. 2023. Leveraging Large Language Models for Sequential Recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems (RecSys ’23) . Association for Computing Machinery, New York, NY, USA, 1096–1102. doi:10.1145/3604915.3610639

    work page arXiv 2023

  9. [9]

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

  10. [10]

    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. In Proceedings of the 28th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining (KDD ’22). Association for Computing Machinery, New York, NY, USA, 585–593. doi:10.1145/3534678.3539381

    arXiv 2022

  11. [11]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. LoRA: Low-Rank Adaptation of Large Language Models. doi:10.48550/arXiv.2106.09685 arXiv:2106.09685

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.09685 2021

  12. [12]

    Wenyue Hua, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. 2023. How to Index Item IDs for Recommendation Foundation Models. In Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region . ACM, Beijing China, 195–204. doi:10.1145/3624918.3625339

    arXiv 2023

  13. [13]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2017. Billion-scale similarity search with GPUs. doi:10.48550/arXiv.1702.08734 arXiv:1702.08734

    Pith/arXiv arXiv doi:10.48550/arxiv.1702.08734 2017

  14. [14]

    Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of Tricks for Efficient Text Classification. In Proceedings of the 15th Confer- ence of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers , Mirella Lapata, Phil Blunsom, and Alexander Koller (Eds.). Association for Computational Linguis...

    work page 2017

  15. [15]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Rec- ommendation. In 2018 IEEE International Conference on Data Mining (ICDM) . 197–206. doi:10.1109/ICDM.2018.00035 ISSN: 2374-8486

    arXiv 2018

  16. [16]

    Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The Efficient Transformer. doi:10.48550/ARXIV.2001.04451

    Pith/arXiv arXiv doi:10.48550/arxiv.2001.04451 2020

  17. [17]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix Factorization Techniques for Recommender Systems. Computer 42, 8 (Aug. 2009), 30–37. doi:10.1109/MC.2009.263

    doi:10.1109/mc.2009.263 2009

  18. [18]

    Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R. Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, Yi Luan, Sai Meher Karthik Duddu, Gustavo Hernandez Abrego, Weiqiang Shi, Nithi Gupta, Aditya Kusupati, Prateek Jain, Siddhartha Reddy Jonnalagadda, Ming-Wei Chang, and Iftekhar Naim. 2024. Gecko: Versatile Text Embeddings Distil...

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.20327 2024

  19. [19]

    Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian McAuley. 2023. Text Is All You Need: Learning Language Representations for Sequential Recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23) . Association for Computing Machinery, New York, NY, USA, 1258–1267. doi:10.11...

    work page arXiv 2023

  20. [20]

    Jinming Li, Wentao Zhang, Tiantian Wang, Guanglei Xiong, Alan Lu, and Gérard G. Medioni. 2023. GPT4Rec: A Generative Framework for Personal- ized Recommendation and User Interests Interpretation. ArXiv abs/2304.03879 (2023). https://api.semanticscholar.org/CorpusID:258048480

    Pith/arXiv arXiv 2023

  21. [21]

    Yaoyiran Li, Xiang Zhai, Moustafa Alzantot, Keyi Yu, Ivan Vulić, Anna Korhonen, and Mohamed Hammad. 2024. CALRec: Contrastive Alignment of Generative LLMs for Sequential Recommendation. In 18th ACM Conference on Recommender Systems. ACM, Bari Italy, 422–432. doi:10.1145/3640457.3688121

    arXiv 2024

  22. [22]

    Xiao Luo, Haixin Wang, Daqing Wu, Chong Chen, Minghua Deng, Jianqiang Huang, and Xian-Sheng Hua. 2023. A Survey on Deep Hashing Methods. ACM Transactions on Knowledge Discovery from Data 17, 1 (Feb. 2023), 1–50. doi:10. 1145/3532624

    work page 2023

  23. [23]

    Frederic Morin and Yoshua Bengio. 2005. Hierarchical Probabilistic Neural Network Language Model. In International Workshop on Artificial Intelligence and Statistics. PMLR, 246–252. https://proceedings.mlr.press/r5/morin05a.html

    work page 2005

  24. [24]

    Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , Kentaro Inui, Jing Jiang, Vincent Ng, and...

    doi:10.18653/v1/d19- 2019

  25. [25]

    Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Brad- bury, Anselm Levskaya, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. 2022. Efficiently Scaling Transformer Inference. (2022). doi:10.48550/ ARXIV.2211.05102

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.05102 2022

  26. [26]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Keshavan, Trung Vu, Lukasz Heidt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender systems with generative retrieval. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS ’23) . Curra...

    work page 2023

  27. [27]

    Steffen Rendle, Walid Krichene, Li Zhang, and Yehuda Koren. 2022. Revisiting the Performance of iALS on Item Recommendation Benchmarks. In Proceedings of the 16th ACM Conference on Recommender Systems (RecSys ’22) . Association for Computing Machinery, New York, NY, USA, 427–435. doi:10.1145/3523227. 3548486

    doi:10.1145/3523227 2022

  28. [28]

    Noam Shazeer. 2019. Fast Transformer Decoding: One Write-Head is All You Need. doi:10.48550/ARXIV.1911.02150

    Pith/arXiv arXiv doi:10.48550/arxiv.1911.02150 2019

  29. [29]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  30. [30]

    Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W

    Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Proceedings of the 36th International Conference on Neural Information Processing Systems (NIPS ’22). Curran Associates Inc., R...

    work page 2022

  31. [31]

    Kilian Weinberger, Anirban Dasgupta, John Langford, Alex Smola, and Josh Attenberg. 2009. Feature hashing for large scale multitask learning. InProceedings of the 26th Annual International Conference on Machine Learning . ACM, Montreal Quebec Canada, 1113–1120. doi:10.1145/1553374.1553516

    work page arXiv 2009

  32. [32]

    Shuyuan Xu, Wenyue Hua, and Yongfeng Zhang. 2024. OpenP5: An Open-Source Platform for Developing, Training, and Evaluating LLM-based Recommender Sys- tems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’24) . Association for Computing Machinery, New York, NY, USA, 386–394. doi:10....

    arXiv 2024

  33. [33]

    Xinyang Yi, Ji Yang, Lichan Hong, Derek Zhiyuan Cheng, Lukasz Heldt, Aditee Kumthekar, Zhe Zhao, Li Wei, and Ed Chi. 2019. Sampling-bias-corrected neural modeling for large corpus item recommendations. InProceedings of the 13th ACM Conference on Recommender Systems (RecSys ’19) . Association for Computing Machinery, New York, NY, USA, 269–277. doi:10.1145...

    arXiv 2019

  34. [34]

    Shusheng Zhang, Lina Yao, Xiang Xu, Chengan Wang, and Xu Wang. 2019. Next item recommendation with self-attentive metric learning. In Proceedings of The 30th International Joint Conference on Artificial Intelligence . 4316–4322. doi:10.24963/ijcai.2019/599

    work page doi:10.24963/ijcai.2019/599 2019

  35. [35]

    Sheng, Jiajie Xu, De- qing Wang, Guanfeng Liu, and Xiaofang Zhou

    Tingting Zhang, Pengpeng Zhao, Yanchi Liu, Victor S. Sheng, Jiajie Xu, De- qing Wang, Guanfeng Liu, and Xiaofang Zhou. 2019. Feature-level deeper self- attention network for sequential recommendation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI’19) . AAAI Press, Macao, China, 4320–4326

    work page 2019

  36. [36]

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-Rec: Self-Supervised Learning for Sequential Recommendation with Mutual Information Maximization. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM ’20) . Association for Computing Machinery,...

    work page arXiv 2020

  37. [37]

    Jing Zhu, Mingxuan Ju, Yozen Liu, Danai Koutra, Neil Shah, and Tong Zhao. 2025. Beyond Unimodal Boundaries: Generative Recommendation with Multimodal Semantics. doi:10.48550/ARXIV.2503.23333

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.23333 2025

  38. [2015]

    arXiv:1511.06939 [cs.IR] https://arxiv.org/abs/1511.06939

    Pith/arXiv arXiv

  39. [2016]

    In Pro- ceedings of the 3rd International Conference on Learning Representations, ICLR

    Session-based recommendations with recurrent neural networks. In Pro- ceedings of the 3rd International Conference on Learning Representations, ICLR

    work page

  40. [2019]

    InProceedings of the 28th ACM International Confer- ence on Information and Knowledge Management (CIKM ’19)

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformer. InProceedings of the 28th ACM International Confer- ence on Information and Knowledge Management (CIKM ’19). Association for Com- puting Machinery, New York, NY, USA, 1441–1450. doi:10.1145/3357384.3357895

    arXiv