Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Restoring item boundaries and suppressing hierarchical drift lets a generative recommender beat all baselines on every metric and in a live test.

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 ↗

BARGE improves generative sequential recommendation by restoring item boundaries in the encoder and suppressing hierarchical semantic drift in decoding, outperforming prior generative baselines on public and industrial benchmarks.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A serious industrial GR paper with plausible gains, but the headline comparison rests on borrowed baseline numbers and no error bars; the architecture and drift analysis are worth referee time. the 3 major comments →

arxiv 2607.21028 v2 pith:INCQIDEF submitted 2026-07-23 cs.IR

Bridging the Structural Gap: Adapting Autoregressive Generation for Recommendation

Junchao Zeng , Junzhang Zhu , Junyang Chen , Yudong Li , Wei Liu , Chengxiang Zhuo , Zang Li This is my paper
classification cs.IR
keywords generative recommendationsequential recommendationsemantic IDsresidual quantizationhierarchical decodingsemantic driftbeam searchdual-path decoding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Generative recommendation predicts the next item by generating its hierarchical semantic-ID token sequence, but this paper argues that the formulation itself creates two structural problems: flattening each item's tokens hides item boundaries in the encoder, and an error at one level of the hierarchy pushes later tokens into a wrong subtree. BARGE is proposed as a fix. Its encoder-side module aggregates each item's tokens into an item context and gates it back into the token stream; its decoder-side modules rerank partial paths with a contrastive scorer and decode through two orthogonal channels whose candidate lists are OR-fused. The paper reports the best scores on every metric for two public benchmarks, with the largest gains (+19.6% R@10 on the Beauty set) over the strongest baseline, and a large online A/B test showing higher click-through rate, click unique visitors, and total reading time. The three modules are shown to be independently useful and largely additive, which is the paper's main evidence that the identified gaps are the real bottleneck.

Core claim

The central claim is that the two structural gaps—loss of item-level structure and hierarchical semantic drift—are the dominant, repairable sources of error in generative recommendation. BARGE repairs them without enlarging the beam or the candidate budget: ICA computes a per-item context via cross-attention and fuses it into each token through a learned gate; HPR augments beam search with per-layer dual-tower contrastive scores trained by a symmetric contrastive objective; DPD rotates the item embedding by a learnable orthogonal matrix, splits it into two halves, quantizes each half with its own codebook, decodes with two towers, and OR-fuses the two ranked item lists. The paper reports tha

What carries the argument

The load-bearing mechanism is the pair of probability identities that separate the two failure modes. For HPR, ε_van_l − ε_HPR_l = Pr[Rescue_l] − Pr[Damage_l] states that reranking helps exactly when it rescues more ground-truth codewords than it damages. For DPD, Pr[E(A)] − Pr[E(A)∩E(B)] = (1−κ)Pr[E(A)] states that OR-fusion's gain over one channel is controlled by the conditional miss rate of the second channel. These identities reduce the design questions ('what weight λ?' and 'does the orthogonal split matter?') to measurable quantities, and the paper reports measurements consistent with them: an inverted-U sensitivity curve and a low overlap (about 0.18) between the two channels' top-K

Load-bearing premise

The headline margins over the strongest baseline assume that the baseline results quoted from earlier papers were produced under exactly the same evaluation protocol, preprocessing, seed handling, and metric computation as BARGE's experiments, since those baselines were not re-run by the authors.

What would settle it

Re-run the strongest baseline on the same code, seeds, early-stopping rule, and metric script as BARGE; if its numbers shift enough to close the reported +19.6% R@10 gap or reverse the ranking, the central claim of superiority collapses. Separately, a direct per-layer measurement showing Pr[Damage] > Pr[Rescue] at the chosen λ=0.25 would falsify HPR's benefit at that layer.

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

If this is right

  • Item boundaries should be treated as first-class structure in generative recommenders; the identity-preserving gated injection used here adds item context without overwriting token representations.
  • Semantic drift can be corrected with a lightweight reranker rather than by widening the beam, preserving the computational advantage of generative retrieval over full-candidate scoring.
  • OR-fusion across orthogonal channels improves item recovery even though each channel alone performs comparably; the reported overlap of about 0.18 indicates the channels make complementary mistakes.
  • The gains are consistent from public benchmarks to a large offline test to an online A/B test, so the repair is not an artifact of one evaluation setting.
  • Because the three modules are shown to be independent and additive, further improvements may be obtained by stacking additional orthogonal failure-dimension fixes.

Where Pith is reading between the lines

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

  • The rescue/damage identity suggests a per-layer diagnostic that the paper does not run: if measured Pr[Damage] > Pr[Rescue] at any layer under the chosen λ, the reranker should be weakened or disabled at that layer. The identity makes this test immediate.
  • The two orthogonal channels may specialize along interpretable semantic facets; a natural extension is to inspect which items are rescued exclusively by each channel and test whether they cluster by category, brand, or price tier, which the paper does not report.
  • The OR-fusion identity generalizes to more than two channels; as long as each additional channel has conditional miss rate below 1, the miss probability falls multiplicatively, suggesting a possible scaling direction beyond the two-channel design.
  • Because the quantitative comparison with several baselines rests on numbers taken from previous papers, re-running those baselines under identical seeds and preprocessing is the direct way to verify the reported margins; the paper's own design claims do not depend on this, but the headline superiority claim does.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper argues that autoregressive generative recommendation loses item-level structure when multi-token semantic IDs are flattened and suffers from semantic drift when decoding over hierarchical codebooks. It proposes BARGE, consisting of three modules: Item Context-Aware Attention (ICA) to restore item boundaries, Hierarchical Path Reranking (HPR) to rerank beam candidates with path-level dual-tower scoring, and Dual-Path Decoding (DPD) with an orthogonal two-channel OSQ-VAE tokenizer and OR-fusion. On Amazon Beauty and Sports and Outdoors, BARGE is reported to achieve the best score on every metric, with representative gains of +19.6% R@10 on Beauty and +8.8%/+16.7% R@10/N@10 on Sports over the strongest baseline. Additional results include a Tencent offline test and an online A/B test reporting +0.60% CTR, +1.34% click unique visitors, and +1.70% total reading time. The paper also includes component ablations, hyperparameter sensitivity analyses, codebook configuration studies, and post-hoc diagnostics of drift recovery and channel complementarity.

Significance. If the empirical claims are reproducible, the paper makes a useful contribution to generative recommendation: the three modules are lightweight, the ablation isolating BARGE-base from the codebook design is a strong experimental practice, the hyperparameter sensitivity analysis is informative, and the online A/B test demonstrates industrial applicability. The conceptual framing of encoder-side and decoder-side structural gaps is clear. However, the central comparative claim rests on baseline numbers borrowed from prior work under an unverified 'same evaluation protocol' assumption, and no variance or significance information is provided for the main tables. The design-rationale identities in Section IV-E are definitional rather than predictive, and the evidence that orthogonal rotation yields complementary channels is only post hoc and lacks statistical quantification. These issues do not necessarily invalidate the approach, but they currently make the headline superiority claim conditional.

major comments (3)
  1. [Section V-A, Baselines; Table II] The central claim that BARGE is best on every metric on both Amazon datasets is supported by Table II, but Section V-A states that baseline results are 'taken from [5], [8] under the same evaluation protocol' and that HSTU and ActionPiece are 'directly sourced from [5]'. The manuscript does not provide evidence that the original pipelines used identical preprocessing, candidate scoring (full-item vs. sampled), beam width, early stopping, seed handling, or metric computation. If any of these differ, the reported margins (e.g., +19.6% R@10 on Beauty) could change substantially. No error bars or confidence intervals are reported even though results are averaged over three seeds. This is load-bearing for the headline result. The authors should either re-run all generative baselines under their own protocol, or provide explicit protocol-compatibility evidence, per-method hyperparameters, and
  2. [Section IV-E, Eq. (18)-(19); Section V-G] Equations (18) and (19) are algebraic identities: Eq. (18) restates the rescue/damage balance and Eq. (19) restates the chain rule for two channel-miss events. As written, they do not by themselves provide a 'verifiable condition' with predictive content; they merely name quantities that could be measured. The paper further states that the orthogonal rotation R in OSQ-VAE drives the conditional miss rate κ down, but orthogonality of subspaces (S_A ⟂ S_B) does not imply that the top-K retrieval failures of the two decoders are complementary, since both decoders share the same encoder and history. The empirical complementarity measurements in Table VIII (Jaccard 0.18/0.17, exclusive-hit fractions 15-24%) are suggestive but are reported without confidence intervals, and κ is not directly reported. Please provide a direct estimate of κ, its variability across seeds, and a statistical test th
  3. [Section V-J, Online A/B Test] The online A/B test is reported in a single sentence: 6% of live traffic, three metric improvements, and 'statistically significant'. For an industrial-scale claim, this is insufficient detail. The manuscript should report the duration of the test, the number of users in each arm, the exact control system, how significance was computed (test type, confidence level, whether metrics were corrected for multiple comparisons), and the variance of the metrics. Without these, the practical value claim cannot be assessed.
minor comments (5)
  1. [Section V-A, Implementation details] The commitment weight β in Eq. (14) is never given a value. Since the codebook configuration and training details are discussed, please report β and any other OSQ-VAE hyperparameters (e.g., codebook initialization, number of Householder reflections).
  2. [Section II.B vs. Section IV.C] HPR is described as 'label-free' in the related work, but Section IV.C trains it with positive pairs consisting of the ground-truth cumulative path embedding. This is direct supervision on ground-truth semantic IDs. Please reword the 'label-free' claim or clarify that it means no additional user-item labels beyond the existing next-item targets.
  3. [Section V-C, Table IV] The efficiency comparison between BARGE and TIGER is not apples-to-apples because BARGE uses a 2-layer encoder while TIGER uses a 4-layer encoder. Please state this explicitly in the table caption or provide a matched-encoder variant so the parameter savings are not overstated.
  4. [Section V-F, Table VI] The column headers 'Mismatch↓', 'Rank↓', 'Prob↑' are not defined in the text. Please define these metrics and clarify how 'Mismatch' is computed.
  5. [Section IV.C, Eq. (8)] The notation 'cos(·,·) · e^{τ_l}' is an unusual way to express a temperature-scaled score. If τ_l is a log-temperature, the standard form would be a division or multiplication by the temperature after exponentiating; please clarify the exact functional form.

Circularity Check

0 steps flagged

No significant circularity: BARGE's gains are empirically validated, and the design-rationale identities are explicit tautologies rather than loaded derivations.

full rationale

BARGE's central claims rest on controlled experiments (component ablations, hyperparameter sweeps, an industrial offline test, and an online A/B test), not on a chain of derivations that reduce to their inputs. The design-rationale identities in Section IV-E are explicitly tautological—Eq. (18) restates the rescue/damage decomposition of two beam-search error probabilities, and Eq. (19) restates the chain rule for the OR-fusion miss event—and the paper uses them only to convert design questions into measurable quantities (Pr[Rescue]>Pr[Damage]; kappa<1), which it then measures directly. This is post-hoc framing, not a prediction derived from the model's own outputs. HPR is trained with ground-truth path positives via InfoNCE and evaluated on held-out data; measuring rescue counts afterward is a diagnostic, not a circular reduction. The comparison against baselines borrows numbers from [5], [8] under an unverified 'same protocol' assumption; this is a genuine external-validity risk, but it is not a form of circularity because the borrowed numbers are not derived from BARGE's own parameters or definitions. No load-bearing self-citation chain or imported uniqueness theorem appears in the derivation. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 4 invented entities

The paper's central claims rest on a standard generative-recommendation pipeline (RQ-VAE semantic IDs, Transformer encoder-decoder), plus new modules (ICA, HPR, DPD) and tuned hyperparameters (λ, Top-N, codebook sizes, β). The main assumptions are the faithfulness of semantic IDs, the comparability of borrowed baselines, and the post-hoc-validated complementarity of the orthogonal channels. No new physical or external entities are invented; the invented entities are architectural components whose only evidence is the paper's own experiments.

free parameters (4)
  • HPR reranking weight λ = 0.25
    Chosen from the test-set NDCG@10 sensitivity curves in Fig. 3 (Section V-E); λ=0 recovers vanilla beam search, so HPR's gain depends on this tuning.
  • HPR scoring pool size Top-N = 400
    Chosen from the test-set NDCG@10 plateau in Fig. 3 (Section V-E); affects the candidate pool before rescoring.
  • Codebook configuration (|C1|,|C2|,|C3|,|C4|) = (512,256,128,64)
    Selected via the codebook configuration analysis in Fig. 4 (Section V-H) on test sets; the depth and sizes affect the semantic ID structure that ICA/HPR/DPD operate on.
  • OSQ-VAE commitment weight β
    Introduced in Eq. (14) as the commitment loss weight, but its value is never reported; it controls the OSQ-VAE training and is a hand-chosen hyperparameter.
axioms (6)
  • domain assumption RQ-VAE/OSQ-VAE semantic IDs faithfully represent item semantics through residual quantization.
    Invoked in Section III and IV-D; the entire generative recommendation framework and BARGE's modules depend on the meaningfulness of hierarchical semantic IDs.
  • domain assumption The flattened token sequence X is an adequate input representation for the encoder, and item boundaries can be recovered by the proposed ICA.
    Section III defines X as the concatenation of item IDs; the paper argues flattening is suboptimal, but still uses it as the base representation for ICA.
  • standard math InfoNCE is a valid surrogate for mutual information between user context and cumulative path.
    Used in Section IV-C, Eq. (10), citing Oord et al. [48]; the HPR training relies on this standard result.
  • domain assumption Prior baselines' reported numbers under the 'same evaluation protocol' are directly comparable to BARGE's results.
    Section V-A states baseline results are taken from [5], [8]; the correctness of the comparison table hinges on this assumption.
  • domain assumption Leave-one-out evaluation on the full item set is a valid proxy for recommendation quality.
    Section V-A adopts the leave-one-out protocol and full-item-set metrics, standard in the field but still a modeling choice.
  • ad hoc to paper The learned orthogonal rotation yields two channels whose top-K candidate sets are complementary enough for OR-fusion to help.
    This is a design hypothesis, not derived; the paper measures κ and Jaccard overlap post hoc in Section V-G (Eq. 19 frames it as a measurable condition).
invented entities (4)
  • OSQ-VAE (Orthogonal Split-and-Quantize VAE) no independent evidence
    purpose: Produces two orthogonal semantic-ID channels per item by rotating and splitting the item embedding.
    A new tokenizer module; its utility is evidenced only by in-paper reconstruction loss and downstream metrics, not by any prediction outside the paper.
  • Dual-Decoder no independent evidence
    purpose: Two independent decoder towers that each predict one channel's semantic ID.
    A new architectural component; no falsifiable handle outside the reported experiments.
  • HPR scorer (per-layer dual-tower reranker) no independent evidence
    purpose: Rescores beam-search candidates using contrastively trained path-level coherence scores.
    A trained reranker component; its value is measured through ablation and sensitivity curves, not predicted a priori.
  • ICA module (Item Context-Aware Attention) no independent evidence
    purpose: Injects item-level context into each token before the encoder.
    A gated cross-attention module; in-paper ablations and gate activation distributions support it, but no external evidence is provided.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging the Structural Gap: Adapting Autoregressive Generation for Recommendation." pith.science (2026). https://pith.science/paper/INCQIDEF

@misc{pith2026260721028,
  author       = {Pith},
  title        = {Pith review of: Bridging the Structural Gap: Adapting Autoregressive Generation for Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INCQIDEF}},
  note         = {Machine review of arXiv:2607.21028}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generative Recommendation (GR) has emerged as a new paradigm for sequential recommendation, in which a representative line of work encodes items into hierarchical semantic IDs via residual quantization and predicts the IDs token by token. However, this generative formulation still exhibits structural gaps with respect to the recommendation task: flattening multi-token IDs into a single sequence destroys item-level structure, and the inconsistency between training and inference over a hierarchical codebook gives rise to semantic drift. To bridge these two gaps, we propose BARGE, which employs Item Context-Aware Attention (ICA) to restore item-level structure during encoding, and Hierarchical Path Reranking (HPR) together with Dual-Path Decoding (DPD) to suppress semantic drift from two complementary angles during decoding. Extensive experiments and analytical studies on public benchmarks and a large-scale offline test demonstrate that BARGE achieves superior recommendation performance. An online A/B test on a Tencent platform yields improvements of 0.60% in click-through rate, 1.34% in click unique visitors, and 1.70% in total reading time, confirming the practical value of BARGE in industrial-scale recommendation.

Figures

Figures reproduced from arXiv: 2607.21028 by Chengxiang Zhuo, Junchao Zeng, Junyang Chen, Junzhang Zhu, Wei Liu, Yudong Li, Zang Li.

Figure 1
Figure 1. Figure 1: Two structural gaps in generative recommendation. Left (Item-Boundary Gap, P1): after each item is tokenized into L hierarchical semantic IDs and flattened into a single sequence, item boundaries vanish from the input to the GR model, and the encoder can no longer tell which tokens belong to the same item (“Where is the item?”). Right (Semantic Drift, P2): on the hierarchical codebook tree, an error at any… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of BARGE. (Top-left) The pretrained OSQ-VAE tokenizes each item into two channel-specific semantic-ID tuples (cbook A and cbook B) via orthogonal rotation and dual codebook stacks; the dual IDs are projected by an MLP into the flattened token sequence. (Top￾middle) The user history passes through the ICA before the shared encoder. (Top-right) Two independent decoders (Channel A and Channel B) each… view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity of HPR hyperparameters (NDCG@10). grained signal than rank-based fusion when channel scores are well-calibrated by the per-channel HPR. • Rotation results. Replacing the learned R with a frozen random orthogonal matrix consistently degrades performance on both datasets, since a random R splits the embedding space along directions unrelated to the recommendation objective. This confirms that the… view at source ↗
Figure 5
Figure 5. Figure 5: Cumulative hit rate comparison with and without ICA across semantic ID layers. The advantage of ICA grows as layers accumulate, demonstrating stronger resistance to semantic drift at deeper layers. To support these two choices empirically, we compare four configurations: (1) (512, 256, 128, 64), our final setting; (2) (256, 256, 256) plus a random collision-resolving ID, which is the original TIGER configu… view at source ↗
Figure 4
Figure 4. Figure 4: Codebook configuration analysis on Beauty and Sports. Four fully-learned configurations (no random collision ID) are compared across R@5, N@5, R@10, and N@10. H. Codebook Configuration Analysis BARGE adopts a fully learned 4-layer codebook (L=4) with layer-wise decreasing sizes (|C1|, . . . , |C4|) = (512, 256, 128, 64), which differs from the TIGER convention [3] of a 3-layer uniform codebook plus a colli… view at source ↗
Figure 6
Figure 6. Figure 6: Gate activation value distribution across semantic ID layers. Gate values are stably concentrated around 0.35–0.38, indicating selective and consistent injection of item-level context. layers (c1 through c4), which suggests that item-level context is stably beneficial at all levels of the hierarchy. HPR: Drift-recovery case study. To examine when HPR is most beneficial, we sweep the entire Sports test set … 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

53 extracted references · 15 linked inside Pith

  1. [1]

    Autoregressive entity retrieval,

    N. De Cao, G. Izacard, S. Riedel, and F. Petroni, “Autoregressive entity retrieval,”arXiv preprint arXiv:2010.00904, 2020

    Pith/arXiv arXiv 2010

  2. [2]

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5),

    S. Geng, S. Liu, Z. Fu, Y . Ge, and Y . Zhang, “Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5),” inProceedings of the 16th ACM Conference on Recommender Systems, 2022, pp. 299–315

    2022

  3. [3]

    Recommender systems with generative retrieval,

    S. Rajput, N. Mehta, A. Singh, R. Hulikal Keshavan, T. Vu, L. Heldt, L. Hong, Y . Tay, V . Tran, J. Samostet al., “Recommender systems with generative retrieval,”Advances in Neural Information Processing Systems, vol. 36, pp. 10 299–10 315, 2023

    2023

  4. [4]

    Higr: Efficient generative slate recommendation via hierarchical planning and multi-objective preference alignment,

    Y . Pang, Z. Liu, Y . Li, S. Zhu, Z. Luo, C. Yu, S. Wu, S. Shen, C. Xu, B. Wanget al., “Higr: Efficient generative slate recommendation via hierarchical planning and multi-objective preference alignment,”arXiv preprint arXiv:2512.24787, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Actionpiece: Contextually tokeniz- ing action sequences for generative recommendation,

    Y . Hou, J. Ni, Z. He, N. Sachdeva, W.-C. Kang, E. H. Chi, J. McAuley, and D. Z. Cheng, “Actionpiece: Contextually tokeniz- ing action sequences for generative recommendation,”arXiv preprint arXiv:2502.13581, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Onesug: The unified end-to-end generative framework for e- commerce query suggestion,

    X. Guo, B. Chen, S. Wang, Y . Yang, M. Cheng, C. Lei, Y . Ding, and H. Li, “Onesug: The unified end-to-end generative framework for e- commerce query suggestion,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 17, 2026, pp. 14 774–14 782

    2026

  7. [7]

    Learnable item tokenization for generative recommendation,

    W. Wang, H. Bao, X. Lin, J. Zhang, Y . Li, F. Feng, S.-K. Ng, and T.-S. Chua, “Learnable item tokenization for generative recommendation,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 2400–2409

    2024

  8. [8]

    Sparse meets dense: Unified generative recom- mendations with cascaded sparse-dense representations,

    Y . Yang, Z. Ji, Z. Li, Y . Li, Z. Mo, Y . Ding, K. Chen, Z. Zhang, J. Li, S. Liet al., “Sparse meets dense: Unified generative recom- mendations with cascaded sparse-dense representations,”arXiv preprint arXiv:2503.02453, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Order-agnostic identifier for large language model-based generative recommendation,

    X. Lin, H. Shi, W. Wang, F. Feng, Q. Wang, S.-K. Ng, and T.-S. Chua, “Order-agnostic identifier for large language model-based generative recommendation,” inProceedings of the 48th international ACM SIGIR conference on research and development in information retrieval, 2025, pp. 1923–1933

    2025

  10. [10]

    Generative recommender with end-to-end learnable item tokenization,

    E. Liu, B. Zheng, C. Ling, L. Hu, H. Li, and W. X. Zhao, “Generative recommender with end-to-end learnable item tokenization,” inProceed- ings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2025, pp. 729–739

    2025

  11. [11]

    Universal item tokenization for transferable generative recommendation,

    B. Zheng, H. Lu, Y . Chen, W. X. Zhao, and J.-R. Wen, “Universal item tokenization for transferable generative recommendation,”arXiv preprint arXiv:2504.04405, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Vqrae: Representation quantization autoencoders for mul- timodal understanding, generation and reconstruction,

    S. Du, J. Guo, B. Li, S. Cui, Z. Xu, Y . Luo, Y . Wei, K. Gai, X. Wang, K. Wuet al., “Vqrae: Representation quantization autoencoders for mul- timodal understanding, generation and reconstruction,”arXiv preprint arXiv:2511.23386, 2025

    arXiv 2025

  13. [13]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyalset al., “Neural discrete representation learning,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  14. [14]

    Autoregressive image generation using residual quantization,

    D. Lee, C. Kim, S. Kim, M. Cho, and W.-S. Han, “Autoregressive image generation using residual quantization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 523–11 532

    2022

  15. [15]

    Session- based recommendations with recurrent neural networks,

    B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk, “Session- based recommendations with recurrent neural networks,”arXiv preprint arXiv:1511.06939, 2015

    Pith/arXiv arXiv 2015

  16. [16]

    Personalized top-n sequential recommendation via convolutional sequence embedding,

    J. Tang and K. Wang, “Personalized top-n sequential recommendation via convolutional sequence embedding,” inProceedings of the Eleventh ACM International Conference on Web Search and Data Mining, 2018, pp. 565–573

    2018

  17. [17]

    Self-attentive sequential recommenda- tion,

    W.-C. Kang and J. McAuley, “Self-attentive sequential recommenda- tion,” in2018 IEEE international conference on data mining (ICDM). IEEE, 2018, pp. 197–206

    2018

  18. [18]

    Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer,

    F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang, “Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer,” inProceedings of the 28th ACM international confer- ence on information and knowledge management, 2019, pp. 1441–1450

    2019

  19. [19]

    Feature-level deeper self-attention network for sequential recom- mendation

    T. Zhang, P. Zhao, Y . Liu, V . S. Sheng, J. Xu, D. Wang, G. Liu, X. Zhou et al., “Feature-level deeper self-attention network for sequential recom- mendation.” inIJCAI, 2019, pp. 4320–4326

    2019

  20. [20]

    S3-rec: Self-supervised learning for sequential recom- mendation with mutual information maximization,

    K. Zhou, H. Wang, W. X. Zhao, Y . Zhu, S. Wang, F. Zhang, Z. Wang, and J.-R. Wen, “S3-rec: Self-supervised learning for sequential recom- mendation with mutual information maximization,” inProceedings of the 29th ACM international conference on information & knowledge management, 2020, pp. 1893–1902

    2020

  21. [21]

    Diffurec: A diffusion model for sequential recommendation,

    Z. Li, A. Sun, and C. Li, “Diffurec: A diffusion model for sequential recommendation,”ACM Transactions on Information Systems, vol. 42, no. 3, pp. 1–28, 2023

    2023

  22. [22]

    Strec: Sparse transformer for sequential recommendations,

    C. Li, Y . Wang, Q. Liu, X. Zhao, W. Wang, Y . Wang, L. Zou, W. Fan, and Q. Li, “Strec: Sparse transformer for sequential recommendations,” inProceedings of the 17th ACM conference on recommender systems, 2023, pp. 101–111

    2023

  23. [23]

    Zero-shot recommendation: Towards class semantic relation learning for inferring labels of unseen micro-videos,

    J. Chen, H. Wang, Y . Wu, Q. Lin, Y . Diao, and J. Ji, “Zero-shot recommendation: Towards class semantic relation learning for inferring labels of unseen micro-videos,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 24, 2026, pp. 20 163–20 171

    2026

  24. [24]

    Wide & deep learning for recommender systems,

    H.-T. Cheng, L. Koc, J. Harmsen, T. Shaked, T. Chandra, H. Aradhye, G. Anderson, G. Corrado, W. Chai, M. Ispiret al., “Wide & deep learning for recommender systems,” inProceedings of the 1st workshop on deep learning for recommender systems, 2016, pp. 7–10

    2016

  25. [25]

    Deep neural networks for youtube recommendations,

    P. Covington, J. Adams, and E. Sargin, “Deep neural networks for youtube recommendations,” inProceedings of the 10th ACM conference on recommender systems, 2016, pp. 191–198

    2016

  26. [26]

    Deepfm: a factorization- machine based neural network for ctr prediction,

    H. Guo, R. Tang, Y . Ye, Z. Li, and X. He, “Deepfm: a factorization- machine based neural network for ctr prediction,”arXiv preprint arXiv:1703.04247, 2017

    Pith/arXiv arXiv 2017

  27. [27]

    One model to rank them all: Unifying online advertising with end-to-end learning,

    J. Qiu, Z. Wang, F. Zhang, Z. Zheng, J. Zhu, J. Fan, T. Zhang, H. Wang, and X. Wang, “One model to rank them all: Unifying online advertising with end-to-end learning,”arXiv e-prints, pp. arXiv–2505, 2025

    2025

  28. [28]

    M6-rec: Generative pretrained language models are open-ended recommender systems,

    Z. Cui, J. Ma, C. Zhou, J. Zhou, and H. Yang, “M6-rec: Generative pretrained language models are open-ended recommender systems,” inProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022

    2022

  29. [29]

    A survey of generative search and recommendation in the era of large language models,

    Y . Li, X. Lin, W. Wang, F. Feng, L. Pang, W. Li, L. Nie, X. He, and T.-S. Chua, “A survey of generative search and recommendation in the era of large language models,”arXiv preprint arXiv:2404.16924, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    A bi-step grounding paradigm for large language mod- els in recommendation systems,

    K. Bao, J. Zhang, W. Wang, Y . Zhang, Z. Yang, Y . Luo, C. Chen, F. Feng, and Q. Tian, “A bi-step grounding paradigm for large language mod- els in recommendation systems,”ACM Transactions on Recommender Systems, vol. 3, no. 4, pp. 1–27, 2025

    2025

  31. [31]

    Reinforced latent reasoning for llm-based recommendation,

    Y . Zhang, W. Xu, X. Zhao, W. Wang, F. Feng, X. He, and T.-S. Chua, “Reinforced latent reasoning for llm-based recommendation,”arXiv preprint arXiv:2505.19092, 2025

    arXiv 2025

  32. [32]

    Mviger: Multi- view variational integration of complementary knowledge for generative recommender,

    T. Kim, S. Yoon, S. Kang, J. Yeo, and D. Lee, “Mviger: Multi- view variational integration of complementary knowledge for generative recommender,”arXiv preprint arXiv:2408.08686, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    How to index item ids for recommendation foundation models,

    W. Hua, S. Xu, Y . Ge, and Y . Zhang, “How to index item ids for recommendation foundation models,” inProceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, 2023, pp. 195–204

    2023

  34. [34]

    Learning to tokenize for generative retrieval,

    W. Sun, L. Yan, Z. Chen, S. Wang, H. Zhu, P. Ren, Z. Chen, D. Yin, M. Rijke, and Z. Ren, “Learning to tokenize for generative retrieval,”Ad- vances in Neural Information Processing Systems, vol. 36, pp. 46 345– 46 361, 2023

    2023

  35. [35]

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

    J. Zhai, L. Liao, X. Liu, Y . Wang, R. Li, X. Cao, L. Gao, Z. Gong, F. Gu, M. Heet al., “Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations,”arXiv preprint arXiv:2402.17152, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment,

    J. Deng, S. Wang, K. Cai, L. Ren, Q. Hu, W. Ding, Q. Luo, and G. Zhou, “Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment,”arXiv preprint arXiv:2502.18965, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Onerec-think: In-text reasoning for generative recommendation,

    Z. Liu, S. Wang, X. Wang, R. Zhang, J. Deng, H. Bao, J. Zhang, W. Li, P. Zheng, X. Wuet al., “Onerec-think: In-text reasoning for generative recommendation,”arXiv preprint arXiv:2510.11639, 2025

    arXiv 2025

  38. [38]

    Generating long semantic ids in parallel for recommendation,

    Y . Hou, J. Li, A. Shin, J. Jeon, A. Santhanam, W. Shao, K. Hassani, N. Yao, and J. McAuley, “Generating long semantic ids in parallel for recommendation,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 2, 2025, pp. 956–966

    2025

  39. [39]

    Drift- aware continual tokenization for generative recommendation,

    Y . Feng, J. Liu, M. Han, D. Li, H. Gu, P. Zhang, T. Lu, and N. Gu, “Drift- aware continual tokenization for generative recommendation,”arXiv preprint arXiv:2603.29705, 2026

    arXiv 2026

  40. [40]

    Delrec: Distilling sequential pattern to enhance llms-based sequential recommendation,

    H. Zhang, G. Sun, J. Lu, G. Liu, and X. S. Fang, “Delrec: Distilling sequential pattern to enhance llms-based sequential recommendation,” in 13 2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE, 2025, pp. 1–14

    2025

  41. [41]

    Reg4rec: Reasoning-enhanced genera- tive model for large-scale recommendation systems,

    H. Xing, H. Deng, Y . Mao, L. Mu, J. Hu, Y . Xu, H. Zhang, J. Wang, S. Wang, Y . Zhanget al., “Reg4rec: Reasoning-enhanced genera- tive model for large-scale recommendation systems,”arXiv preprint arXiv:2508.15308, 2025

    arXiv 2025

  42. [42]

    Train short, test long: Attention with linear biases enables input length extrapolation,

    O. Press, N. A. Smith, and M. Lewis, “Train short, test long: Attention with linear biases enables input length extrapolation,”arXiv preprint arXiv:2108.12409, 2021

    Pith/arXiv arXiv 2021

  43. [43]

    Longformer: The long- document transformer,

    I. Beltagy, M. E. Peters, and A. Cohan, “Longformer: The long- document transformer,”arXiv preprint arXiv:2004.05150, 2020

    Pith/arXiv arXiv 2004

  44. [44]

    Trie-aware transformers for generative recommendation,

    Z. Xu, J. Chen, S. Chen, Y . He, J. Yang, C. Yuan, K. Ding, and C. Wang, “Trie-aware transformers for generative recommendation,” arXiv preprint arXiv:2602.21677, 2026

    arXiv 2026

  45. [45]

    Promise: Process reward models unlock test-time scaling laws in generative recommendations,

    C. Guo, K. Cai, Y . Zhou, Q. Luo, R. Tang, H. Li, K. Gai, and G. Zhou, “Promise: Process reward models unlock test-time scaling laws in generative recommendations,”arXiv preprint arXiv:2601.04674, 2026

    arXiv 2026

  46. [46]

    Apao: Adaptive prefix- aware optimization for generative recommendation,

    Y . Yu, Y . Wang, W. Ma, Z. Guo, and M. Zhang, “Apao: Adaptive prefix- aware optimization for generative recommendation,”arXiv preprint arXiv:2603.02730, 2026

    Pith/arXiv arXiv 2026

  47. [47]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  48. [48]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  49. [49]

    Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering,

    R. He and J. McAuley, “Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering,” inproceedings of the 25th international conference on world wide web, 2016, pp. 507– 517

    2016

  50. [50]

    Hierarchical gating networks for sequential recommendation,

    C. Ma, P. Kang, and X. Liu, “Hierarchical gating networks for sequential recommendation,” inProceedings of the 25th ACM SIGKDD interna- tional conference on knowledge discovery & data mining, 2019, pp. 825–833

    2019

  51. [51]

    Plum: Adapting pre-trained lan- guage models for industrial-scale generative recommendations,

    R. He, L. Heldt, L. Hong, R. Keshavan, S. Mao, N. Mehta, Z. Su, A. Tsai, Y . Wang, S.-C. Wanget al., “Plum: Adapting pre-trained lan- guage models for industrial-scale generative recommendations,”arXiv preprint arXiv:2510.07784, 2025

    arXiv 2025

  52. [52]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,”Advances in neural information processing systems, vol. 30, 2017

    2017

  53. [53]

    Approximate nearest neighbor search under neural similarity metric for large-scale recommendation,

    R. Chen, B. Liu, H. Zhu, Y . Wang, Q. Li, B. Ma, Q. Hua, J. Jiang, Y . Xu, H. Denget al., “Approximate nearest neighbor search under neural similarity metric for large-scale recommendation,” inProceedings of the 31st ACM International Conference on Information & Knowledge Management, 2022, pp. 3013–3022. 14

    2022

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.