Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Act-With-Think: Chunk Auto-Regressive Modeling for Generative Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read CAR predicts each item as one 'think-then-act' chunk, instead of token-by-token, and reports Recall@5 gains of 7.9%-22.3% over prior generative recommenders.

desk verdict Plausible idea, shaky evidence; worth refereeing after the tables and baselines are fixed. read the letter →

arxiv 2506.23643 v1 pith:SGVIPSX6 submitted 2025-06-30 cs.IR

classification cs.IR
keywords generativerecommendationchunkautoregressivemodelingsemanticIDsuniquesequentialact-with-thinkresidualk-meansslowthinking
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

The paper argues that the two kinds of item tokens used by generative recommenders—semantic IDs that capture what an item is, and unique IDs that capture which item it is—should be generated together as one chunk rather than one token at a time. It packages each item as a SIDs-UID chunk and trains a transformer to predict the whole next chunk from the last token of the current chunk, a design it calls 'act-with-think' because the semantic tokens give the reason and the unique ID gives the action. On three product categories, CAR reports Recall@5 improvements of 7.93%-22.30% over existing generative and sequential baselines, plus large inference speedups because beam search is no longer needed. If these numbers hold under uniform evaluation, the paper would establish chunk-level autoregression as a practical alternative to token-level generation in recommendation.

What carries the argument

The load-bearing object is the 'act-with-think' chunk: a tuple in which hierarchical semantic IDs (SIDs) represent the 'why' and a unique ID (UID) represents the 'what'. The operative mechanism is chunk-level autoregression—each decoding step emits all tokens of the next chunk at once from the current chunk's last token—with a dual-branch loss that separately scores semantic reconstruction and item-level identity. A progressive context-fusion layer augments each token embedding with semantic embeddings from the preceding positions, letting coarse-to-fine semantic levels share context. Together these pieces remove the sequential dependency that standard chain decomposition enforces between SIDs and UID.

What would settle it

Re-run every baseline and the AR ablation in one codebase with identical splits, candidate sets, and metric implementations; if CAR's Recall@5 margin over the best baseline disappears or the corrected AR row approaches CAR's numbers, the central claim is false.

Watch

Extended reading notes

Core claim

The central claim is that semantic and collaborative information should be co-generated as a composite chunk, not decoded in a fixed hierarchical order. Concretely, CAR forms an 'act-with-think' chunk $(S_1,\dots,S_k,\mathrm{UID})$ per item, factorizes the joint probability as $P(S_1,\dots,S_k,\mathrm{UID}\mid H)=P(S_1\mid H)\cdots P(\mathrm{UID}\mid H)$, and predicts the entire next chunk in parallel from the previous chunk's final token. The think branch supervises the semantic IDs and the act branch supervises the unique ID, with a context-fusion step that feeds prefix semantics into each token's embedding. Experiments on three product categories report consistent gains over baselines and a scaling trend: Recall@5 rises 25.8%-30.5% as the number of SID bits increases from one to four.

Load-bearing premise

The headline gains rest on the assumption that all baselines were evaluated uniformly and correctly, which is called into question by the paper's own statement that only TIGER was reproduced and by the internally inconsistent AR ablation row for Sports.

Editorial extensions

If this is right

  • CAR removes the need for sequential beam search over SID levels; reported inference speed on the Toys dataset is 34 to 146 times faster than token-level AR as beam width grows from 5 to 20.
  • Recall@5 improves by 25.8% on Beauty and 30.5% on Toys when the SID bit count grows from one to four, evidence for a scaling effect between semantic granularity and recommendation quality.
  • Ablations show that both the progressive context fusion module and the think-branch loss contribute to the result: removing either one reduces performance.
  • Modeling SIDs and UID as relatively independent but jointly predicted branches avoids the strict conditional dependency assumed by the chain-rule factorization of standard autoregressive generation.

Reading between the lines

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

  • If the chunk-level factorization is the real driver of the gains, the same act-with-think packaging could transfer to other sequence tasks where tokens come from distinct but correlated vocabularies, such as generating an entity together with its descriptive tags.
  • A stronger test of the 'slow thinking' analogy would hold the model size fixed and vary only the number of semantic levels beyond four; if gains saturate, the bit-count scaling is a token-granularity effect rather than a reasoning-depth effect.
  • Because CAR decouples SIDs from UID at prediction time, it opens a natural way to control recommendations by editing the semantic prefix, such as forcing a category or brand token, something a SIDs-first token-by-token model cannot cleanly do.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Chunk AutoRegressive Modeling (CAR) for generative recommendation. Items are represented as chunks composed of hierarchical semantic IDs (SIDs) plus a unique ID (UID); the model predicts all tokens of the next chunk in parallel from the last token of the current chunk, using a dual-branch loss with a think loss for SIDs and an act loss for the UID. A progressive context-fusion mechanism enriches token representations with prefix information. Experiments on three Amazon categories report consistent Recall@K and NDCG@K improvements over several sequential and generative baselines, plus an inference-speed advantage over standard autoregressive decoding. The paper also reports a scaling effect of SID bit count, which it interprets as evidence of a slow-thinking-like mechanism.

Significance. If the reported results are validated, CAR would be a simple and potentially useful modification to token-level autoregressive generative recommenders: packing SIDs and UID into a chunk and predicting them in parallel is architecturally straightforward, removes sequential decoding dependencies, and yields large speedups in the reported inference-speed comparison. The paper also provides a clear, falsifiable comparison of residual KMeans against LSH and RQ-VAE for semantic ID construction. The main weakness is empirical: the evaluation tables contain an impossible entry and the headline improvement range is inconsistent with the reported best result, and baseline numbers come from heterogeneous sources without significance testing. As presented, the central quantitative claim is not yet established.

major comments (4)
  1. [§4.2, Table 3] In Table 3, the AR row for Sports reports Recall@5=0.0164 and Recall@10=0.0105, and NDCG@5=0.0251 and NDCG@10=0.0133. Recall@K and NDCG@K are monotone nondecreasing in K, so these values are impossible for any valid ranking; the row appears to have the @5 and @10 columns swapped. As printed, the AR baseline for Sports cannot be used to support the claim that CAR significantly outperforms standard autoregressive modeling. Please correct the table and confirm the values with a rerun or a precise explanation.
  2. [Abstract and §4.1, Table 2] The abstract states that CAR improves Recall@5 by 7.93% to 22.30%, but Table 2 reports a Sports Recall@5 improvement of 28.16% (0.0405 vs 0.0316) and a Toys improvement of 20.19%. The upper bound in the abstract does not match the paper's own best reported gain. Please reconcile the reported range and ensure all quantitative claims in the abstract are consistent with the tables.
  3. [§4.1, evaluation protocol] Section 4.1 says that only TIGER was reproduced by the authors, HSTU numbers are taken from the ActionPiece paper, and the remaining baselines come from their original publications or third-party code. Because the central claim is a set of relative percentage improvements over these baselines, the evaluation needs either a single unified protocol applied to all methods or an explicit per-baseline statement of dataset filtering, sequence truncation, validation split, and metric implementation. As it stands, the headline gains may not be apples-to-apples, and no error bars or significance tests are reported to support the word 'significant' in the abstract and Section 4.2.
  4. [§4.4, Table 5] The scaling effect of SID bit number is presented as evidence of a slow-thinking-style mechanism. However, increasing the number of SID bits also increases the amount of semantic information available to the decoder, so the observed monotone improvement may simply reflect richer item descriptions rather than deeper intermediate computation. Please provide a control that separates increased information content from increased number of reasoning steps (for example, fixing the total available information while varying the number of decoding steps), or soften the claim that CAR emulates slow thinking in the LLM sense.
minor comments (5)
  1. [§4.2] The text says 'The results in Table 1 demonstrate that CAR significantly outperforms AR,' but the relevant results are in Table 3, not Table 1; please fix the cross-reference.
  2. [§3.1] The residual KMeans method is described as 'parameter-free,' but the number of SID levels and the codebook cardinality per level are hyperparameters chosen by the experimenter (as evidenced by Table 5); please qualify this terminology.
  3. [§4.5, Table 6] In Table 6, CAR is said to avoid beam search, yet the table lists 'num_beams' values for both AR and CAR; please clarify whether the beam count is varied for CAR and what the column means in that case.
  4. [References] References [28] and [29] are identical (both are 'Transformer memory as a differentiable search index'); please remove the duplicate.
  5. [§1] There is a typo in the phrase 'mechanism' spelled as 'machanism' in the Introduction; please proofread the manuscript for similar typos.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CAR's chunk construction, factorized objective, and losses define the method rather than being fitted to the evaluation, and the held-out experiments are self-contained.

full rationale

The derivation chain does not reduce to its inputs. The SIDs are produced by a parameter-free residual KMeans applied to Sentence-T5 embeddings, and both SIDs and UID are supervised targets in a held-out last-item evaluation; no parameter is fitted to the test set and then renamed a prediction. The factorization in Eq. (2) is a stated modeling choice, implemented through the dual loss in Eqs. (3)-(5), rather than a quantity derived from the data it aims to predict, so the 'act-with-think' framing is definitional to the architecture, not circular. The only author-overlapping citation is to EAGER (Ref. [32], whose authors include Jieming Zhu, Zhenhua Dong, and Ruiming Tang), but EAGER is cited only as background and motivation and is not invoked as proof of CAR's correctness, uniqueness, or superiority. The claimed scaling effect from longer SIDs is an empirical trend on held-out Recall/NDCG, and the slow-thinking analogy is interpretive, not a load-bearing derivation. Concerns raised by the skeptical summary, such as the mixed-source baselines in Table 2 and the impossible AR ablation row in Table 3 (Sports Recall@10=0.0105 below Recall@5=0.0164), are evaluation-protocol and internal-consistency issues, not circularity: a wrong or non-uniform baseline does not make the model's output equivalent to its input by construction. On circularity grounds, therefore, the paper is self-contained; any criticism belongs to correctness and reproducibility risk.

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

CAR's central empirical result rests on hand-chosen hyperparameters (alpha, SID bit count, codebook size, history length), on the unverified factorization in Eq. (2), and on the assumption that semantic IDs from residual KMeans capture decision-relevant semantics. The slow thinking interpretation is an analogy, not a mechanism. None of these premises are tested independently, so the contribution is a configuration rather than a first-principles result.

free parameters (5)
  • alpha (think loss weight)
    Equation (5) combines think and act losses with weight alpha; the paper never reports its value or tuning procedure, and results depend on it.
  • Number of SID bits/levels = 4 (default)
    Main results use 4 semantic ID levels with cardinality 256; Table 5 shows performance scales with this number, so it is a chosen hyperparameter that partly drives the reported gains.
  • Codebook cardinality per semantic level = 256
    Section 4.3 states all SID methods use four layers with cardinality 256; this granularity is chosen by hand and affects semantic fidelity and the scaling curve.
  • Max history length = 20
    Training limits each user's historical items to 20; longer histories might change results, so this truncation is a modeling choice.
  • GPT-2 architecture hyperparameters = 1 layer, 8 heads, hidden 1024, embedding 128, lr 1e-5, batch 256
    These are standard but arbitrary choices; no sensitivity analysis is given, so the reported improvements are tied to this configuration.
assumptions (5)
  • domain assumption Residual KMeans on Sentence-T5 embeddings yields meaningful hierarchical semantic IDs.
    Section 3.1 assumes clustering residuals preserves shared semantics; no evaluation of codebook quality is provided beyond downstream recommendation.
  • domain assumption Users make decisions by thinking about item semantics (brand, category) before acting (clicking or buying).
    Section 1 motivates act-with-think with this cognitive claim; it is not tested.
  • ad hoc to paper Equation (2): P(S1,S2,UID|H) factorizes into P(S1|H)P(S2|H)P(UID|H).
    Section 3.2 replaces the chain rule with a conditional independence assumption to enable parallel chunk decoding; no evidence shows this independence holds.
  • ad hoc to paper More SID bits correspond to more intermediate thinking steps, so bit-count scaling is evidence of slow-thinking behavior.
    Section 4.4 draws this analogy; bit count only increases input and output vocabulary size and does not imply deeper computation.
  • domain assumption Item texts encoded by Sentence-T5 capture the semantics relevant to user choices.
    Section 3.1 relies on the pretrained encoder without verifying that its embedding space aligns with purchase behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Act-With-Think: Chunk Auto-Regressive Modeling for Generative Recommendation." pith.science (2026). https://pith.science/paper/SGVIPSX6

@misc{pith2026250623643,
  author       = {Pith},
  title        = {Pith review of: Act-With-Think: Chunk Auto-Regressive Modeling for Generative Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SGVIPSX6}},
  note         = {Machine review of arXiv:2506.23643}
}
abstract

Generative recommendation (GR) typically encodes behavioral or semantic aspects of item information into discrete tokens, leveraging the standard autoregressive (AR) generation paradigm to make predictions. However, existing methods tend to overlook their intrinsic relationship, that is, the semantic usually provides some reasonable explainability "$\textbf{why}$" for the behavior "$\textbf{what}$", which may constrain the full potential of GR. To this end, we present Chunk AutoRegressive Modeling (CAR), a new generation paradigm following the decision pattern that users usually think semantic aspects of items (e.g. brand) and then take actions on target items (e.g. purchase). Our CAR, for the $\textit{first time}$, incorporates semantics (SIDs) and behavior (UID) into a single autoregressive transformer from an ``act-with-think'' dual perspective via chunk-level autoregression. Specifically, CAR packs SIDs and UID into a conceptual chunk for item unified representation, allowing each decoding step to make a holistic prediction. Experiments show that our CAR significantly outperforms existing methods based on traditional AR, improving Recall@5 by 7.93% to 22.30%. Furthermore, we verify the scaling effect between model performance and SIDs bit number, demonstrating that CAR preliminary emulates a kind of slow-thinking style mechanism akin to the reasoning processes observed in large language models (LLMs).

Figures

Figures reproduced from arXiv: 2506.23643 by the authors.

Figure 1
Figure 1. An overview of the Chunk AutoRegressive Modeling framework. First, each item is [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. CAR consists of two independent training stages. Stage 1: Semantic IDs (SIDs) are [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 12 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

  2. [2]

    A neural probabilistic language model

    Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. A neural probabilistic language model. Journal of machine learning research, 3(Feb):1137–1155, 2003

  3. [3]

    N-gram-based text categorization

    William B Cavnar, John M Trenkle, et al. N-gram-based text categorization. In Proceedings of SDAIR-94, 3rd annual symposium on document analysis and information retrieval, volume 161175, page 14. Ann Arbor, Michigan, 1994

  4. [4]

    Similarity estimation techniques from rounding algorithms

    Moses S Charikar. Similarity estimation techniques from rounding algorithms. In Proceedings of the thiry-fourth annual ACM symposium on Theory of computing, pages 380–388, 2002

  5. [5]

    An empirical study on eliciting and improving r1-like reasoning models

    Zhipeng Chen, Yingqian Min, Beichen Zhang, Jie Chen, Jinhao Jiang, Daixuan Cheng, Wayne Xin Zhao, Zheng Liu, Xu Miao, Yang Lu, et al. An empirical study on eliciting and improving r1-like reasoning models. arXiv preprint arXiv:2503.04548, 2025

  6. [6]

    Recommen- dation with generative models

    Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, Rene Vidal, Maheswaran Sathiamoorthy, Atoosa Kasrizadeh, Silvia Milano, et al. Recommen- dation with generative models. arXiv preprint arXiv:2409.15173, 2024

  7. [7]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

  8. [8]

    Deepseek-coder: When the large language model meets programming–the rise of code intelligence

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming–the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024

Show all 40 references
  1. [9]

    Actionpiece: Contextually tokenizing action sequences for generative recommendation

    Yupeng Hou, Jianmo Ni, Zhankui He, Noveen Sachdeva, Wang-Cheng Kang, Ed H Chi, Julian McAuley, and Derek Zhiyuan Cheng. Actionpiece: Contextually tokenizing action sequences for generative recommendation. arXiv preprint arXiv:2502.13581, 2025

  2. [10]

    Approximate nearest neighbors: towards removing the curse of dimensionality

    Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: towards removing the curse of dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing, pages 604–613, 1998

  3. [11]

    Locality-preserving hashing in multidimensional spaces

    Piotr Indyk, Rajeev Motwani, Prabhakar Raghavan, and Santosh Vempala. Locality-preserving hashing in multidimensional spaces. In Proceedings of the twenty-ninth annual ACM symposium on Theory of computing, pages 618–625, 1997

  4. [12]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

  5. [13]

    Self-attentive sequential recommendation

    Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), pages 197–206. IEEE, 2018

  6. [14]

    Matrix factorization techniques for recom- mender systems

    Yehuda Koren, Robert Bell, and Chris V olinsky. Matrix factorization techniques for recom- mender systems. Computer, 42(8):30–37, 2009

  7. [15]

    Bert: a review of applications in natural language processing and under- standing

    Mikhail V Koroteev. Bert: a review of applications in natural language processing and under- standing. arXiv preprint arXiv:2103.11943, 2021

  8. [16]

    Controlar: Controllable image generation with autoregressive models

    Zongming Li, Tianheng Cheng, Shoufa Chen, Peize Sun, Haocheng Shen, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Controlar: Controllable image generation with autoregressive models. In International Conference on Learning Representations, 2025

  9. [17]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 10

  10. [18]

    Roberta: A robustly optimized bert pretraining approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019

  11. [19]

    Hierarchical gating networks for sequential recommen- dation

    Chen Ma, Peng Kang, and Xue Liu. Hierarchical gating networks for sequential recommen- dation. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 825–833, 2019

  12. [20]

    Image-based recommendations on styles and substitutes

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43–52, 2015

  13. [21]

    Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models

    Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith B Hall, Daniel Cer, and Yinfei Yang. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. arXiv preprint arXiv:2108.08877, 2021

  14. [22]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020

  15. [23]

    Recommender systems with generative retrieval

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al. Recommender systems with generative retrieval. Advances in Neural Information Processing Systems, 36:10299–10315, 2023

  16. [24]

    Better generalization with semantic ids: A case study in ranking for recommendations

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

  17. [25]

    Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. InPro- ceedings of the 28th ACM international conference on information and knowledge management, pages...

  18. [26]

    Think before recommend: Unleashing the latent reasoning power for sequential recommendation

    Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Wu Jian, and Yuning Jiang. Think before recommend: Unleashing the latent reasoning power for sequential recommendation. arXiv preprint arXiv:2503.22675, 2025

  19. [27]

    Personalized top-n sequential recommendation via convolutional sequence embedding

    Jiaxi Tang and Ke Wang. Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining, pages 565–573, 2018

  20. [29]

    Transformer memory as a differentiable search index

    Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. Transformer memory as a differentiable search index. Advances in Neural Information Processing Systems, 35:21831–21843, 2022

  21. [30]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37:84839–84865, 2024

  22. [31]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

  23. [32]

    Eager: Two-stream generative recommender with behavior-semantic collaboration

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

  24. [33]

    A neural corpus indexer for document retrieval

    Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, et al. A neural corpus indexer for document retrieval. Advances in Neural Information Processing Systems, 35:25600–25614, 2022

  25. [34]

    Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations

    Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, et al. Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations. arXiv preprint arXiv:2503.02453, 2025

  26. [35]

    Sampling-bias-corrected neural modeling for large corpus item recommendations

    Xinyang Yi, Ji Yang, Lichan Hong, Derek Zhiyuan Cheng, Lukasz Heldt, Aditee Kumthekar, Zhe Zhao, Li Wei, and Ed Chi. Sampling-bias-corrected neural modeling for large corpus item recommendations. In Proceedings of the 13th ACM conference on recommender systems, pages 269–277, 2019

  27. [36]

    Soundstream: An end-to-end neural audio codec

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495–507, 2021

  28. [37]

    Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Michael He, et al. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations. arXiv preprint arXiv:2402.17152, 2024

  29. [38]

    Slow thinking for sequential recommendation

    Junjie Zhang, Beichen Zhang, Wenqi Sun, Hongyu Lu, Wayne Xin Zhao, Yu Chen, and Ji-Rong Wen. Slow thinking for sequential recommendation. arXiv preprint arXiv:2504.09627, 2025

  30. [39]

    What, how, where, and how well? a survey on test-time scaling in large language models

    Qiyuan Zhang, Fuyuan Lyu, Zexu Sun, Lei Wang, Weixu Zhang, Zhihan Guo, Yufei Wang, Irwin King, Xue Liu, and Chen Ma. What, how, where, and how well? a survey on test-time scaling in large language models. arXiv preprint arXiv:2503.24235, 2025

  31. [40]

    A survey of large language models

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2), 2023

  32. [41]

    S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization. In Proceedings of the 29th ACM international conference on informatio...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.