Pith. sign in

REVIEW 4 major objections 6 minor 48 references

Preserving Item Semantics for Free: Rethinking Token Initialization in LLM-Based Generative Recommendation

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that random Gaussian initialization of semantic-ID tokens in LLM-based generative recommenders discards the item embeddings' geometry and steers models toward popularity, while a parameter-free centroid initialization…

desk verdict Simple, parameter-free centroid initialization for SID tokens is a good idea, but the experiments appear to compare against a collapsed mean baseline that the paper misdescribes as random Gaussian, so the headline gains are not yet trustworthy. read the letter →

arxiv 2608.07816 v1 pith:FOUCEPES submitted 2026-08-07 cs.IR cs.LG

Donald Loveland , Liam Collins , Bhuvesh Kumar , Danai Koutra , Neil Shah This is my paper · ORCID
classification cs.IRcs.LG
keywords generativerecommendationsemanticIDstokeninitializationLLMvocabularyexpansioncentroidpopularitybiasMatryoshkarepresentationlearningcold-item
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

Generative recommenders represent each item as a short sequence of discrete semantic codes, added to an LLM's vocabulary as special tokens. This paper argues that the usual way of initializing those tokens—sampling random Gaussian vectors that match the pretrained vocabulary's mean and covariance—throws away the geometry of the item embeddings from which the codes were built. As a result, the model reorganizes the token space around item popularity instead of semantics, and even the continual-pretraining phase designed to align tokens with language only partially fixes this. The proposed fix is direct: initialize each semantic-ID token from its cluster centroid in the semantic embedding space, mean-shifted into the LLM's token distribution and truncated to the LLM's hidden dimension using the nesting property of modern text embeddings. This one-time, parameter-free initialization is reported to improve pure supervised-fine-tuning Recall@5 by 6.5% on average (up to 16%), reach peak performance with 27% fewer SFT steps (up to 40%), and improve cold-item Recall@5 by 37% on average (up to 60%), with no additional training or inference overhead.

What carries the argument

The carrying mechanism is centroid initialization: a one-time, parameter-free assignment that writes each semantic SID token's embedding row as $e^{\mathrm{cent}}_{\ell,k} = \Phi_{d_{\mathrm{model}}}(c_{\ell,k} - \bar{c}) + \mu_{\mathrm{base}}$, where $c_{\ell,k}$ is the residual-$k$-means centroid for level $\ell$, code $k$, $\bar{c}$ is the mean over all semantic centroids, $\mu_{\mathrm{base}}$ is the mean of the pretrained token embeddings, and $\Phi_{d_{\mathrm{model}}}$ is a truncation to the first $d_{\mathrm{model}}$ dimensions supplied by Matryoshka representation learning, the nesting property that concentrates meaning in the leading dimensions of modern text embeddings. This injection gives the LLM a semantically organized starting space. A second mechanism explains the failure it fixes: Theorem 3.1 decomposes the gradient update on a token row and isolates the term $\eta \tilde{n}_v \sum_{t=0}^{S-1} \bar{h}^{(t)}$, which shifts every token along a shared direction by an amount proportional to its relative supervision count, making popularity a geometric axis of the learned space. The diagnostic metrics $P_{10}$ (neighborhood category purity) and $R^2_q(T_{\mathrm{pop}})$ versus $R^2_q(T_{\mathrm{sem}})$ (variance of popularity or category targets captured by the top $q$ principal components) carry the empirical comparison.

What would settle it

Run the identical pure-SFT comparison with a non-Matryoshka 4096-dimensional item encoder, truncating centroids to the LLM's 2048-dimension space; if the cold-item Recall@5 gains and $P_{10}$ purity gains over random initialization largely disappear, the Matryoshka bridge is load-bearing and the method does not transfer to arbitrary semantic embeddings.

Watch

Extended reading notes

Core claim

The paper's central claim is twofold: standard random initialization leaves SID token embeddings collapsed in scale and organized primarily by popularity, and the expensive continual-pretraining stage does not reliably recover their original semantic geometry. The evidence is that the leading principal components of the learned embedding table capture 77–93% of code-level log-popularity variance while categorical structure stays near chance, and a gradient analysis shows how relative token frequency accumulates as a shared popularity axis during training. The constructive claim is that writing each residual-k-means centroid into its token row, after centering and a mean shift into the pretrained token distribution, preserves the semantic prior through training and produces better recommendation, faster convergence, and markedly better cold-item recall. On the paper's reading, the semantic geometry of SIDs should be an input to the LLM rather than something the LLM must rediscover from interaction data.

Load-bearing premise

The load-bearing premise is that the leading 2048 dimensions of a 4096-dimensional item embedding still carry the item's meaning, so truncating centroids to the LLM's width is a faithful semantic bridge; if that truncation drops semantic information for a given encoder, the injected geometry is distorted and the reported cold-item gains would not transfer.

Editorial extensions

If this is right

  • Pure supervised fine-tuning can replace much of the continual-pretraining warm-up: centroid initialization lifts average Recall@5 by 6.5% and reaches peak validation Recall@5 27% sooner.
  • On datasets that still benefit from continual pretraining, centroid initialization matches or exceeds the random baseline's peak with 1–4 CPT epochs instead of 6–10, roughly halving alignment compute.
  • Cold-item Recall@5 improves by 37% on average (60% on MovieLens, 50% on Steam) in the pure-SFT regime, showing the popularity-bias diagnostic predicts real tail-item gains.
  • The intervention is a drop-in replacement: a few lines of code, no changes to architecture, training objective, or inference, and no added parameters or compute.
  • The diagnostic result implies current multi-stage pipelines underutilize SID semantics: from random initialization the learned embeddings encode code-level popularity in their top principal components, and CPT only partially corrects this.

Reading between the lines

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

  • A natural extension is to test centroid initialization with a non-Matryoshka text encoder; if the leading-dimension truncation is not semantically faithful, the cold-item gains should shrink or disappear, which would confirm that the Matryoshka bridge, not a generic warm-start effect, is what carries the method.
  • The popularity-accumulation mechanism in Theorem 3.1 is not obviously limited to recommenders: any vocabulary expansion whose new tokens receive skewed supervision counts may develop the same frequency axis, so semantic-centroid initialization could be audited in general LLM adaptation tasks.
  • Because the paper evaluates only CPT and SFT, centroid initialization's interaction with later post-training stages is open; the geometric prior should also reduce the alignment burden in preference or reasoning training, but that is an extrapolation from the reported results.
  • Practitioners could reuse the paper's diagnostics ($P_{10}$ and $R^2_q$ of popularity or category targets) as an audit on deployed SID embedding tables to detect whether a model is organizing around popularity before running expensive cold-item evaluations.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This paper studies how semantic-ID (SID) token embeddings should be initialized when expanding an LLM vocabulary for generative recommendation. The authors argue that the standard practice of random Gaussian initialization discards the continuous geometry of the item embeddings used to construct SIDs; through diagnostic analyses (neighborhood purity, spectral scale, latent-signal R-squared) they report that trained SID embeddings under this initialization encode popularity rather than category, and that continual pretraining only partially mitigates this. They propose centroid initialization (Eq. 5): mean-shift residual k-means centroids into the pretrained token space via Matryoshka truncation, overwrite the new SID rows, and leave the training pipeline unchanged. On five datasets with Qwen3 backbones they report Recall@5 gains in pure SFT (6.5% average, up to 16%), faster convergence (27% fewer steps on average), and larger cold-item gains (37% average, up to 60%), plus CPT-epoch savings. A shuffled-centroid ablation is used to separate the benefit of non-degenerate row geometry from semantic code-to-centroid alignment.

Significance. If the main empirical claim holds, the contribution is practically valuable: a parameter-free, drop-in initialization that improves recommendation quality and training efficiency with no architectural changes. The paper is well structured, has careful diagnostic metrics, a useful shuffled-centroid control, five datasets, and several internal consistency checks (e.g., deeper CPT reducing the head/tail gap). The claim of "free" gains is appropriately scoped to the training pipeline. However, the reliability of the headline numbers currently depends on a baseline mismatch and on single-run estimates, so the significance is conditional until those points are resolved.

major comments (4)
  1. [Section B / Table 6 / Algorithm 1 / Eq. (2)] The implemented "random" baseline is not the baseline defined in Eq. (2). The text says the baseline is "the standard HuggingFace mean-resizing default, initializing each new row from the scaled Gaussian defined in Equation (2)", but HuggingFace's mean_resizing=True fills every new row with the identical mean vector of the base embedding table; it does not draw independent Gaussian samples. A collapsed identical-row start is a degenerate baseline, and the abstract's first claim, the motivation in Section 2, and the keyword "random initialization" all describe an initialization that is not what was run. The shuffled-centroid ablation in Appendix H preserves row norms and the multiset of rows, but it does not control for collapsed versus non-degenerate initialization, and it does not compare shuffled centroids against a true Gaussian baseline on most datasets. Please rerun the comparison with independent Gaussian draws with a stated epsilon, or reframe the paper as a comparison against mean-resized initialization and adjust Eq. (2), the abstract, and the motivation accordingly.
  2. [Tables 1-4 / Figures 3-5 / Eq. (2)] All main quantitative claims are single-run point estimates with no error bars, no seed information, and no value reported for the variance-scaling factor epsilon in Eq. (2). The headline numbers (6.5%, 16%, 27%, 37%, 60%) are therefore not accompanied by any estimate of run-to-run variability, which is particularly important because the proposed method is intended to beat a stochastic initialization. Please report mean and standard deviation over at least three seeds for the main pure-SFT and cold-item results, state the seed(s) and epsilon, and indicate whether checkpoint selection itself is stable across seeds.
  3. [Eq. (5) / Appendix B / Section 4.1] The bridge from semantic centroids to the LLM space rests entirely on the assertion that truncating Qwen3-Embedding-8B's 4096-dimensional vectors to the first d_model=2048 dimensions preserves the semantics relevant for recommendation. The Matryoshka property guarantees nested representations at training time, but it does not automatically guarantee that the first half of the coordinate space is a faithful semantic subspace for this item corpus or for the centroid structure produced by residual k-means. This assumption is load-bearing because without it the parameter-free projection in Eq. (5) has no empirical justification, and the reported cold-item gains would not be expected to transfer to other embedding models or backbone sizes. Please add a direct validation, e.g., category purity or SID-code agreement of truncated versus full centroids, and ideally the same diagnostic with a second embedding model.
  4. [Section 3.3 / Section F] Theorem 3.1 is derived under full-batch gradient descent with untied output embeddings, while the experiments use mini-batch SFT with tied input/output embeddings. The paper's Section F describes the proof as providing intuition, and the empirical popularity-encoding claim is separately supported by the R-squared diagnostics, so this is not fatal. However, the main text currently presents the theorem as an explanation of the experimental phenomenon without stating these mismatches; please add a sentence in Section 3.3 making clear that the theorem models a simplified untied/full-batch setting and that the residual term may absorb the omitted input-side and stochastic effects.
minor comments (6)
  1. [Abstract / Section 2 / Table 6] The terminology is inconsistent: the abstract says "random Gaussian vectors", Eq. (2) defines a Gaussian draw, and Table 6 says the baseline is "mean resizing". This should be unified even if the experiment is rerun.
  2. [Figure 2 caption] The labels "Random CPT+SFT Learned Neighbors" and "Random Neighbors" are easy to confuse; please rename them to, e.g., "Random init + CPT+SFT" and "Expected under random neighbors".
  3. [Table 2] The green/red color encoding is not legible in grayscale; please add explicit plus/minus signs or arrows in addition to color.
  4. [Appendix H / Table 9] The notation d1 and d10 is used before it is defined in the text; please define it at first use as CPT depth 1 and CPT depth 10.
  5. [Abstract / Section 5.1 / Appendix G] The abstract's "half as many CPT epochs" phrasing is stronger than the evidence for Sports, where Appendix G reports that centroid initialization needs additional CPT epochs to match the random baseline's peak; please qualify or soften the claim.
  6. [Overall] No code or data availability statement is included; please provide a reproducibility statement with configuration details (epsilon, seeds, dataset splits, and checkpoint-selection protocol).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline result is a held-out empirical comparison between two parameter-free initializations, and the supporting theorem is an explanatory identity, not a fitted prediction.

full rationale

The central claim is that centroid initialization (Eq. 5) improves pure-SFT Recall@5, convergence speed, and cold-item recall relative to the standard random/mean-resized baseline. This is evaluated on held-out test targets via constrained beam search; the proposed initialization is a deterministic, parameter-free function of precomputed semantic centroids and the base-vocabulary mean, so the reported gains are not constructed from the evaluation metric or from the training labels. Theorem 3.1 and its proof in Appendix F are exact gradient-decomposition identities: they isolate a popularity-dependent term e_n_v sum_t hbar^(t) while explicitly retaining a residual Delta^(t)_v, and the paper itself states the proof 'is meant to provide intuition on how popularity can become encoded, not state that it will become encoded.' Thus the theorem explains, rather than fits, the observed embedding diagnostics. The Matryoshka truncation assumption behind Eq. (5) is load-bearing but is supported by external references to Matryoshka representation learning and embedding model documentation ([18], [20], [29]), not by a self-citation chain; whether truncation is semantically faithful is a correctness/robustness risk, not circularity. The self-citations present ([10], [16], [26]) are contextual related-work and handbook references and do not carry the derivation. A possible concern that the HuggingFace 'mean resizing' baseline may not exactly match the Gaussian covariance model of Eq. (2) is an experimental-fidelity issue about the fairness of the baseline, not a circularity: even if the baseline were degenerate, centroid initialization would still be empirically compared rather than derived from its own inputs. No load-bearing step reduces, by the paper's own equations or by self-citation, to its own inputs, so no circular step can be exhibited.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central method has no learned parameters and introduces no new entities. Its load-bearing assumptions are the Matryoshka truncation and the centroid-as-proxy faithfulness. The only uncontrolled continuous knob in the comparison is the scale factor epsilon in the random-init baseline, which is not reported.

free parameters (1)
  • epsilon (random-init variance scale) = unreported
    Equation (2) scales the base-vocabulary covariance by epsilon for the random initialization baseline. The paper states epsilon is small but never reports its value; since the baseline uses the HuggingFace mean-resizing default, the comparability of random vs centroid init depends on an uncontrolled implementation choice.
assumptions (2)
  • domain assumption Matryoshka property: for Qwen3-Embedding-8B, the leading d_model dimensions of the 4096-d item embedding retain the semantic information needed for the SID geometry, so truncation in Equation (5) preserves semantics.
    Invoked in Section 4.1 and Appendix B to justify dimension matching without a learned projection; not verified for this encoder beyond the MRL citation.
  • domain assumption Residual k-means centroids faithfully represent the semantic content of the code region for each SID code, so using the centroid as the token embedding is a valid semantic prior.
    Sections 2 and 4.1; dedup codes are excluded, but the assumption that the centroid is the best proxy for the code's semantics is not empirically validated against, for example, random item embedding within the region.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Preserving Item Semantics for Free: Rethinking Token Initialization in LLM-Based Generative Recommendation." pith.science (2026). https://pith.science/paper/FOUCEPES

@misc{pith2026260807816,
  author       = {Pith},
  title        = {Pith review of: Preserving Item Semantics for Free: Rethinking Token Initialization in LLM-Based Generative Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FOUCEPES}},
  note         = {Machine review of arXiv:2608.07816}
}
read the original abstract

Recent advances in generative recommendation (GR) leverage large language models (LLMs) as recommender backbones, enabling LLMs to directly generate recommendations conditioned on item-interaction histories. In these systems, items are often represented through semantic IDs (SIDs) added to the LLM vocabulary as special tokens. Ideally, SIDs imbue item token representations with semantic priors, thereby improving model generalization. However, standard vocabulary expansion typically initializes these tokens as random Gaussian vectors, discarding the SIDs' underlying continuous geometry and forcing the LLM to relearn token relationships from interaction data. To demonstrate the consequences of this design, we first show that training from this initialization tends to organize SID embeddings around item popularity rather than semantics. We further show that, despite partially reducing the reliance on popularity and improving cold item performance, the computationally expensive process of continual pretraining (CPT) fails to reliably recover the original semantic geometry. To address these findings, we propose a simple, parameter-free intervention that initializes SID token embeddings directly from their corresponding centroids in the semantic embedding space. Requiring only a few lines of code and no additional training or inference overhead, this drop-in approach improves pure-SFT Recall@5 by up to 16%, reaches peak performance with up to 40% fewer SFT steps, and improves cold-item Recall@5 by up to 60%. Moreover, on datasets that benefit from additional CPT, centroid initialization reaches comparable performance while requiring half as many CPT epochs. Together, our findings show that preserving SID geometry, beyond shared-prefix structure, provides a simple and effective semantic prior for LLM-based GR.

Figures

Figures reproduced from arXiv: 2608.07816 by the authors.

Figure 1
Figure 1. Learned embedding geometry for Amazon Beauty [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Neighborhood purity (𝑃10) of learned SID embed￾dings under CPT+SFT, compared to a random baseline (gray) and the original semantic embeddings (green). While the semantic embeddings exhibit strong categorical coherence, standard random initialization fails to preserve this struc￾ture. nearest neighbors (via L2 distance) of code index 𝑘 ∈ {0, . . . , 𝐾 − 1} within the learned embedding space. We then compute the fract… view at source ↗
Figure 3
Figure 3. Neighborhood purity (𝑃10) of SID embeddings under centroid initialization and CPT+SFT, vs. random initializa￾tion. Values above bars indicate the increase in purity. Cen￾troid initialization enhances the semantic structure of the latent space. let SID code index 𝑘 at level ℓ correspond to a centroid cℓ,𝑘 ∈ R 𝑑 in the semantic embedding space. Additionally, assume the LLM has embedding matrix E ∈ R |V |×𝑑model . To e… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Recall@5 across varying CPT depths. Centroid-init can reach the random-init performance with fewer CPT epochs. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Relative Recall@5 gap (centroid − random) on cold targets (train-target freq. ≤ 1) versus the number of CPT epochs. With fewer CPT epochs, centroid initialization tends to produce larger cold-item recall gains. Robustness Across Scale. To investigate whether benefits p…
Figure 6
Figure 6. Figure 6: Input to the embedding model. For Beauty, this [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 8
Figure 8. Figure 8: SFT example for Amazon Beauty: Uses ChatML [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Token frequency vs. average item popularity for the level-0 codebook [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Relative metric gaps (centroid − random) on cold targets (train-target freq. ≤ 1) versus the number of CPT epochs. With fewer CPT epochs, centroid initialization tends to produce larger cold-item metric gains. Recall@10 and NDCG Across CPT Epochs. In [PITH_FULL_IMAGE…
Figure 11
Figure 11. Figure 11: Recall@10, NDCG@5, and NDCG@10 across numbers of CPT epochs. Centroid initialization often reaches the [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 16 canonical work pages

  1. [1]

    Daiwei Chen, Zhoutong Fu, Chengming Jiang, Haichao Zhang, Ran Zhou, Tan Wang, Chunnan Yao, Guoyao Li, Rui Cai, Yihan Cao, Ruijie Jiang, Fedor Borisyuk, Jianqiang Shen, Jingwei Wu, and Ramya Korlakai Vinayak. 2026. Grounded Token Initialization for New Vocabulary in LMs for Generative Recommendation. arXiv:2604.02324 [cs.CL] https://arxiv.org/abs/2604.02324

    arXiv 2026

  2. [2]

    Xiangyi Chen, Zelun Wang, Xinyi Li, Yi-Ping Hsu, Jaewon Yang, and Jiajing Xu

    work page

  3. [3]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. OneRec: Unifying Retrieve and Rank with Genera- tive Recommender and Iterative Preference Alignment. arXiv:2502.18965 [cs.IR] https://arxiv.org/abs/2502.18965

    arXiv 2025

  4. [4]

    Konstantin Dobler and Gerard de Melo. 2023. FOCUS: Effective Embedding Ini- tialization for Monolingual Specialization of Multilingual Models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Asso- ciation for Computational Linguistics, 13440–13454. doi:10.18653/v1/2023.emnlp- main.829

    doi:10.18653/v1/2023.emnlp- 2023

  5. [5]

    Zhiang Dong, Liya Hu, Jingyuan Chen, Zhihua Wang, and Fei Wu. 2025. Com- prehend 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

    doi:10.1145/3716499 2025

  6. [6]

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin

  7. [7]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InProceedings of the 16th ACM conference on recommender systems. 299–315

    2022

  8. [8]

    Gomez-Uribe and Neil Hunt

    Carlos A. Gomez-Uribe and Neil Hunt. 2016. The Netflix Recommender System: Algorithms, Business Value, and Innovation.ACM Trans. Manage. Inf. Syst.6, 4, Article 13 (Dec. 2016), 19 pages. doi:10.1145/2843948

    doi:10.1145/2843948 2016

Show all 48 references
  1. [9]

    Ruining He, Lukasz Heldt, Lichan Hong, Raghunandan Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, Xinyang Yi, Lexi Baugher, Baykal Cakici, Ed Chi, Cristos Goodrow, Ningren Han, He Ma, Romer Rosales, Abby Van Soest, Devansh Tandon, S...

    2026

  2. [10]

    Yinhan He, Liam Collins, Bhuvesh Kumar, Jundong Li, Neil Shah, and Donald Loveland. 2026. Implicit Reasoning for Large Language Model-based Generative Recommendation. arXiv:2606.14142 [cs.CL] https://arxiv.org/abs/2606.14142

    2026 arXiv

  3. [11]

    Yingzhi He, Yan Sun, Junfei Tan, Yuxin Chen, Xiaoyu Kong, Chunxu Shen, Xiang Wang, An Zhang, and Tat-Seng Chua. 2026. Reasoning over semantic ids enhances generative recommendation.arXiv preprint arXiv:2603.23183(2026)

    2026 arXiv

  4. [12]

    Min Hou, Le Wu, Yuxin Liao, Yonghui Yang, Zhen Zhang, Yu Wang, Chang- long Zheng, Han Wu, and Richang Hong. 2026. A Survey on Generative Recommendation: Data, Model, and Tasks. arXiv:2510.27157 [cs.IR] https: //arxiv.org/abs/2510.27157

    2026 arXiv

  5. [13]

    Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learning vector-quantized item representation for transferable sequential recommenders. InProceedings of the ACM Web Conference 2023. 1162–1171

    2023

  6. [14]

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large language models are zero-shot rankers for recommender systems. InEuropean conference on information retrieval. Springer, 364–381

    2024

  7. [15]

    Wei Jin, Haitao Mao, Zheng Li, Haoming Jiang, Chen Luo, Hongzhi Wen, Haoyu Han, Hanqing Lu, Zhengyang Wang, Ruirui Li, et al . 2024. Amazon-m2: A multilingual multi-locale shopping session dataset for recommendation and text generation.Advances in Neural Information Processing...

    2024

  8. [16]

    Clark Mingxuan Ju, Liam Collins, Leonardo Neves, Bhuvesh Kumar, Louis Yufeng Wang, Tong Zhao, and Neil Shah. 2025. Generative Recommendation with Seman- tic IDs: A Practitioner’s Handbook. InProceedings of the 34th ACM International Conference on Information and Knowledge Mana...

    2025

  9. [17]

    Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, and Xiangnan He. 2025. MiniOneRec: An Open-Source Framework for Scaling Generative Recommendation. arXiv:2510.24431 [cs.IR]

    2025

  10. [18]

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

    2022

  11. [19]

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2024. Large language models for generative recommendation: A survey and visionary discussions. InPro- ceedings of the 2024 joint international conference on computational linguistics, language resources and evaluation (LREC-COLI...

    2024

  12. [20]

    Mingxin Li, Yanzhao Zhang, Dingkun Long, Keqin Chen, Sibo Song, Shuai Bai, Zhibo Yang, Pengjun Xie, An Yang, Dayiheng Liu, et al. 2026. Qwen3-VL- Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking.arXiv preprint arXiv:260...

    2026 arXiv

  13. [21]

    Fake Lin, Binbin Hu, Zhi Zheng, Xi Zhu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, and Tong Xu. 2026. Token-level Collaborative Alignment for LLM-based Generative Recommendation. InProceedings of the ACM Web Conference 2026(United Arab Emirates)(WWW ’26). Association for Computing Ma...

    2026

  14. [22]

    Siyi Lin, Chongming Gao, Jiawei Chen, Sheng Zhou, Binbin Hu, Yan Feng, Chun Chen, and Can Wang. 2025. How Do Recommendation Models Amplify Pop- ularity Bias? An Analysis from the Spectral Perspective. InProceedings of the Eighteenth ACM International Conference on Web Search a...

    2025

  15. [23]

    Huafeng Liu, Jingxuan Wen, Liping Jing, and Jian Yu. 2019. Deep gener- ative ranking for personalized recommendation. InProceedings of the 13th ACM Conference on Recommender Systems(Copenhagen, Denmark)(RecSys ’19). Association for Computing Machinery, New York, NY, USA, 34–42...

    2019

  16. [24]

    Zhenghao Liu, Sen Mei, Chenyan Xiong, Xiaohua Li, Shi Yu, Zhiyuan Liu, Yu Gu, and Ge Yu. 2023. Text matching improves sequential recommendation by reducing popularity biases. InProceedings of the 32nd ACM international Loveland et al. conference on information and knowledge ma...

    2023

  17. [25]

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, Yifei Hu, Qigen Hu, Xinchen Luo, Lejian Ren, Zixing Zhang, Qianqian Wang, Kuo Cai, Yunfan Wu, Hongtao Cheng, Zexuan Cheng, Lu Ren, Huanjie Wang...

    2025

  18. [26]

    Donald Loveland, Mingxuan Ju, Tong Zhao, Neil Shah, and Danai Koutra. 2025. On the Role of Weight Decay in Collaborative Filtering: A Popularity Perspective. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 1975–1986

    2025

  19. [27]

    Masoud Mansoury, Himan Abdollahpouri, Mykola Pechenizkiy, Bamshad Mobasher, and Robin Burke. 2020. Feedback Loop and Bias Amplification in Recommender Systems. InProceedings of the 29th ACM International Confer- ence on Information & Knowledge Management(Virtual Event, Ireland...

    2020

  20. [28]

    Nandini Mundra, Aditya Nanda Kishore Khandavally, Raj Dabre, Ratish Pudup- pully, Anoop Kunchukuttan, and Mitesh M Khapra. 2024. An Empirical Com- parison of Vocabulary Expansion and Initialization Approaches For Language Models. InProceedings of the 28th Conference on Computa...

    2024 doi

  21. [29]

    Zach Nussbaum, John Xavier Morris, Andriy Mulyar, and Brandon Duderstadt

  22. [30]

    Gustavo Penha, Edoardo D’Amico, Marco De Nadai, Enrico Palumbo, Alexandre Tamborrino, Ali Vardasbi, Max Lefarov, Shawn Lin, Timothy Heath, Francesco Fabbri, et al. 2025. Semantic ids for joint generative search and recommendation. InProceedings of the Nineteenth ACM Conference...

    2025

  23. [31]

    Haohao Qu, Wenqi Fan, Zihuai Zhao, and Qing Li. 2025. Tokenrec: Learning to tokenize id for llm-based generative recommendations.IEEE Transactions on Knowledge and Data Engineering(2025)

    2025

  24. [32]

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al

  25. [33]

    Ben Schafer, Joseph Konstan, and John Riedl

    J. Ben Schafer, Joseph Konstan, and John Riedl. 1999. Recommender systems in e-commerce. InProceedings of the 1st ACM Conference on Electronic Commerce (Denver, Colorado, USA)(EC ’99). Association for Computing Machinery, New York, NY, USA, 158–166. doi:10.1145/336992.337035

    1999

  26. [34]

    Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, et al. 2024. Better generalization with semantic ids: A case study in ranking for recommendations. InProceedings of the 18th ACM...

    2024

  27. [35]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable item tokenization for generative recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2400–2409

    2024

  28. [36]

    Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, et al . 2024. Eager: Two-stream generative recommender with behavior-semantic collaboration. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and ...

    2024

  29. [37]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender ...

    2024

  30. [38]

    Atsuki Yamaguchi, Aline Villavicencio, and Nikolaos Aletras. 2026. How Can We Effectively Expand the Vocabulary of LLMs with 0.01GB of Target Language Text? Computational Linguistics52, 1 (March 2026), 295–330. doi:10.1162/coli.a.581

    2026 doi

  31. [39]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jia...

    2025 arXiv

  32. [40]

    Liu Yang, Fabian Paischer, Kaveh Hassani, Jiacheng Li, Shuai Shao, Zhang Gabriel Li, Yun He, Xue Feng, Nima Noorshams, Sem Park, Bo Long, Robert D Nowak, Xiaoli Gao, and Hamid Eghbalzadeh. 2025. Unifying Generative and Dense Retrieval for Sequential Recommendation.Transactions...

    2025

  33. [41]

    Qihang Yu, Kairui Fu, Zheqi Lv, Shengyu Zhang, Xinhui Wu, Chen Lin, Feng Wei, Bo Zheng, and Fei Wu. 2026. ThinkRec: Thinking-based recommendation via LLM. InProceedings of the ACM Web Conference 2026. ACM, 5698–5709. doi:10.1145/3774904.3792070

    2026

  34. [42]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, et al. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommenda- tions. InInternational Conference on Machine Learn...

    2024

  35. [43]

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

    2024

  36. [44]

    Hair Care > Conditioners

    Guorui Zhou, Honghui Bao, Jiaming Huang, Jiaxin Deng, Jinghao Zhang, Junda She, Kuo Cai, Lejian Ren, Lu Ren, Qiang Luo, et al. 2025. OpenOneRec Technical Report.arXiv preprint arXiv:2512.24762(2025). A Additional Related Work In this section, we expand on the related work intr...

    2025

  37. [2019]

    InThe World Wide Web Conference(San Francisco, CA, USA)(WWW ’19)

    Graph Neural Networks for Social Recommendation. InThe World Wide Web Conference(San Francisco, CA, USA)(WWW ’19). Association for Computing Machinery, New York, NY, USA, 417–426. doi:10.1145/3308558.3313488

  38. [2023]

    Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

    2023

  39. [2024]

    Transactions on Machine Learning Research(2024)

    Nomic Embed: Training a Reproducible Long Context Text Embedder. Transactions on Machine Learning Research(2024)

    2024

  40. [2026]

    LLMs Need Encoders for Semantic IDs Too.arXiv preprint arXiv:2606.00324 (2026)

    2026 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.