Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation

Chenxi Li; Xu Yang; Yuchen Lu

REVIEW 4 major objections 5 minor 25 references

Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation

T0 review · 4 major / 5 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 Adaptive capacity allocation for semantic IDs improves generative recommendation accuracy at a fixed capacity budget.

desk verdict A modest, cleanly-controlled idea—adaptive bit allocation over semantic ID slots—whose empirical support would be stronger with error bars and full budget disclosure. read the letter →

arxiv 2608.09685 v1 pith:LPLLK2ZV submitted 2026-08-10 cs.AI

Chenxi Li , Yuchen Lu , Xu Yang This is my paper · ORCID

classification cs.AI

keywords sequentialrecommendationsemanticIDparallelgenerationproductquantizationadaptivecapacityallocationcodebooksizegenerativeretrievalaccuracy

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

This paper argues that the manually fixed, homogeneous semantic-ID structures used by parallel generative recommenders waste prediction capacity, and that adaptively spending a fixed bit budget across candidate semantic subspaces yields better retrieval accuracy without slowing inference. The proposed framework, InforID, replaces hand-chosen ID lengths and equal-size codebooks with a greedy allocation that gives each bit to the slot where it most reduces subspace reconstruction loss, then deletes slots that receive no bits. On four public review benchmarks, InforID beats or matches the parallel-semantic-ID baseline RPG using the same item embeddings, backbone, and capacity budget. If the finding holds, practitioners should treat the shape of a semantic ID as a data-dependent design decision rather than a fixed hyperparameter.

What carries the argument

The load-bearing object is the greedy marginal-gain allocation over candidate semantic slots. Starting from zero bits, InforID adds one bit at a time to the slot with the largest reduction in subspace reconstruction loss, $L_j(b_j) = \frac{1}{|\mathcal{I}|}\sum_{i \in \mathcal{I}} \min_{\mathbf{e} \in C_j(b_j)} \| \mathbf{z}^{(j)}_i - \mathbf{e} \|_2^2$, until the budget $B$ is exhausted. This determines each slot's codebook size $K_j = 2^{b_j}$, removes zero-bit slots, and defines the output dimensions of the per-slot heads used in the parallel factorization $P(\mathbf{c}_{i^+}|u) = \prod_{j \in A} P^{(j)}(c_{i^+,j}|u)$.

What would settle it

Take a dataset where one projected subspace has low variance but is known to encode a preference-relevant categorical feature, such as brand or price band, and compare InforID against a uniform-allocation baseline under the same bit budget. If InforID assigns zero bits to that slot and then underperforms the uniform baseline in retrieval, the reconstruction-loss proxy fails as an allocation signal.

Watch

Extended reading notes

Core claim

The central claim is that uniform codebook capacity across semantic slots is inefficient, and that adaptive capacity allocation can improve one-step parallel generative recommendation at the same total capacity budget. InforID formalizes semantic ID construction as minimizing summed subspace reconstruction loss under a fixed bit budget, using a greedy algorithm that assigns each bit to the slot with the largest marginal reduction, $\Delta_j = L_j(b_j) - L_j(b_j+1)$. Slots whose final bit-width is zero are dropped, so the effective ID length is data-dependent and slot codebooks become heterogeneous. The paper shows controlled gains over RPG, for instance Sports Recall@5 rising from 0.0314 to 0.0329 under identical item embeddings and backbone, and reports lower collision rates than homogeneous OPQ under the same budget.

Load-bearing premise

The framework relies on marginal reconstruction loss reduction, $L_j(b_j)-L_j(b_j+1)$, as a faithful proxy for how much an extra bit helps downstream retrieval; if a slot with small reconstruction gain actually carries strong user-preference signal, the allocation will starve it or delete it, and the accuracy gains vanish.

Editorial extensions

If this is right

Slot-level codebook sizes and effective ID length become outputs of the allocation process rather than hyperparameters, so practitioners no longer need to grid-search the ID structure.
Uniformly expanding semantic slots is shown to be an unreliable accuracy lever: performance peaks and then fluctuates as slots are added, so budget is better spent adaptively.
Because zero-bit slots are excluded, the resulting IDs have lower collision rates than homogeneous OPQ under the same bit budget, improving item discriminability.
The gains in Table 2 hold under the same item embeddings and backbone as RPG, isolating ID construction as the source of improvement.
One-step parallel prediction is preserved, so the accuracy gains do not trade away inference efficiency.

Reading between the lines

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

The paper's allocation signal is reconstruction loss, but the final criterion is retrieval; a stronger variant could allocate bits by expected retrieval gain on a validation set, which the paper does not test.
The same greedy capacity allocation could apply outside recommendation to any retrieval or indexing setup that quantizes heterogeneous subspaces, such as dense retrieval indexes or multimodal embedding compression; the paper does not make that claim.
The finding suggests that how many semantic tokens an item needs should also be item-dependent; InforID allocates per-slot capacity globally rather than per item, so per-item variable-length IDs are an untested next step.
Because InforID requires only k-means and a greedy pass over slots, it can be dropped into existing parallel generative recommenders without retraining the backbone; the paper demonstrates this by plugging into RPG.

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.

Desk Editor's Note

read the letter

Quick take: this is a small, honest paper. It takes the standard quantization trick of greedy marginal-reconstruction-loss bit allocation and applies it to semantic ID construction for parallel generative recommenders. That is a reasonable and clearly-scoped contribution, not a new algorithm. The controlled comparison against RPG—same embeddings, same backbone, only the ID construction changes—is the right design and the one that makes the result believable. The paper also deserves credit for checking that uniform slot expansion is not monotonic (Table 1) and for including a budget sensitivity curve.

The soft spots are the ones you'd expect. The gain over RPG is real but small (Sports R@5 0.0314 to 0.0329), and no variance is reported despite three runs. The central attribution—that the gain comes from replacing the homogeneous ID structure—rests on the §3.2 proxy assumption: marginal reduction in reconstruction loss is treated as marginal utility for retrieval. The text explicitly calls it a 'lightweight proxy' and never isolates it against, say, random allocation or an oracle using validation NDCG. That isn't fatal: for a heuristic target-construction method, the proxy is plausible and Table 3 gives some indirect support by showing lower reconstruction loss and higher NDCG together. But the paper would be stronger if it acknowledged this as a heuristic and showed one ablation using a different allocation signal. Also, exact bit budgets per dataset are not stated (only a range in the sensitivity figure), so 'comparable capacity budgets' is hard to verify without the code.

The citation pattern looks fine—OPQ, TIGER, RPG, RecJPQ are the right antecedents, and the paper is not claiming to invent quantization. Self-citation is not an issue here.

Who is the audience? People working on semantic-ID recommenders or parallel decoding for retrieval. It's a subfield-level improvement, not a paradigm shift. The paper deserves a serious referee: the idea is clear, the experiments are controlled, and the weaknesses are fixable with disclosure and error bars rather than being load-bearing. I'd accept it for review with the expectation of a revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes InforID, a method for constructing semantic item IDs for parallel generative recommendation. Instead of fixing a homogeneous ID structure with equal codebook sizes for every semantic slot, InforID allocates a fixed total bit budget across candidate slots by greedily assigning bits to the slot with the largest marginal reduction in subspace reconstruction loss. Slots receiving zero bits are removed, yielding a data-dependent effective ID length and slot-specific codebook sizes. The authors evaluate InforID on four Amazon review datasets against SASRec, VQ-Rec, RecJPQ, TIGER, and RPG, and report improved or comparable Recall and NDCG while preserving one-step parallel prediction. A controlled comparison against PQ and OPQ under the same bit budget is used to attribute the gains to adaptive capacity allocation.

Significance. If the central proxy assumption holds, InforID is a simple, lightweight, and plausibly general improvement over manually fixed homogeneous semantic IDs: it plugs into existing parallel generative recommenders without changing the inference paradigm. The design is transparent and the controlled comparison against RPG, holding item embeddings and prediction backbone fixed, is a genuine strength, as is the public code link. However, the paper's central claim depends on an assumption that the authors themselves label a 'lightweight proxy,' and the empirical support is weakened by small margins, missing variance reporting, and the absence of an ablation that isolates the allocation signal from the mere fact of heterogeneous capacity. The contribution is therefore potentially useful but not yet fully established.

major comments (4)

[Section 3.2 and Section 4.3] The load-bearing assumption is that the marginal reconstruction-loss reduction Δ_j = L_j(b_j) − L_j(b_j+1) is a valid proxy for the utility of an extra bit for downstream retrieval. The manuscript explicitly states that reconstruction loss 'serves only as a lightweight proxy,' but no experiment isolates this assumption. Table 3 compares PQ, OPQ, and InforID, yet all three are reconstruction-loss-oriented construction strategies; there is no comparison against random bit allocation, subspace-energy allocation, or a validation-NDCG oracle. Without such an ablation, the sentence in Section 4.2 that 'the gain mainly comes from replacing the manually fixed homogeneous ID structure' does not distinguish the effect of heterogeneous capacity from the effect of the specific greedy allocation signal.
[Section 4.1 and Table 2] The claim of evaluation 'under comparable capacity budgets' is not operationalized. The paper never states the total bit budget B used for each dataset, the number of candidate slots M, or the codebook sizes used by RPG and InforID. Since Figure 3 shows that performance depends strongly on B, the reader cannot verify that RPG and InforID are truly on equal footing. In addition, all results are reported as averages over three runs without standard deviations, confidence intervals, or significance tests. Several headline gains are small (e.g., Sports R@5 0.0314 to 0.0329; Beauty R@5 0.0550 to 0.0562), making it hard to separate the claimed improvement from seed noise. Exact budget settings and variance estimates are needed to support the central comparison.
[Section 4.3] The claim that InforID removes redundant capacity is supported only by relative reconstruction loss and a single sentence about lower collision rates. No absolute codebook sizes, number of zero-bit slots, effective ID lengths, or collision-rate numbers are reported. Zero-bit slot removal is central to the mechanism, because slots with b_j = 0 are excluded from prediction and scoring. Reporting how many slots are dropped per dataset, and ideally showing what happens when those slots are forcibly retained, would directly test the redundancy claim that motivates the paper. Without these numbers, the conclusion in Section 5 that homogeneous IDs contain redundant capacity remains only indirectly supported.
[Section 3.2 and Figure 3] The greedy allocation procedure is presented as the core method, but no analysis is given of its quality relative to the optimal allocation, nor of its sensitivity to tie-breaking or the order of bit assignment. Since the allocation objective is separable across subspaces only after the codebooks are fixed, a small-scale comparison against exhaustive allocation (for small M and B) or a discussion of suboptimality would strengthen the method. Additionally, Figure 3 plots NDCG@10 versus B from 64 to 352 without error bars, so the claimed saturation pattern may be consistent with seed noise.

minor comments (5)

[Tables 1 and 3, Figure 3] Error bars or per-seed results should be reported for the key tables and for Figure 3; the non-monotonic pattern in Table 1 is central to the motivation and would be more convincing with variance information.
[Figure 1] The figure lacks axis labels and the caption does not define the units of subspace energy or the exact meaning of λ_d; adding these would make the motivation easier to verify.
[Section 4.1] The text says TIGER uses sentence-t5-base while RPG and InforID use text-embedding-3-large, and that both embeddings were tested for TIGER, RPG, and InforID; please clarify which embedding is used for the numbers in Table 2 and where the additional embedding comparison appears.
[Section 4.2] The phrase 'best or comparable' is not defined; specifying a threshold or statistical criterion (e.g., overlapping confidence intervals) would make the claim more precise.
[ACM Reference Format] The reference-format block still contains placeholder text ('Conference acronym ’XX', '2018', 'Woodstock, NY') and should be corrected to the actual venue and year.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: InforID's capacity allocation minimizes reconstruction loss, while the reported claims concern downstream retrieval accuracy; there is no self-citation chain or fitted parameter renamed as prediction.

full rationale

The central derivation is self-contained and does not reduce to its inputs by construction. The capacity allocation objective in Section 3.2 minimizes total subspace reconstruction loss under a fixed bit budget, with the marginal gain Δ_j = L_j(b_j) − L_j(b_j+1) used only to choose which slot receives the next bit; the reported claim is retrieval accuracy (Recall/NDCG on four Amazon datasets), which is measured after training with the resulting heterogeneous semantic IDs. The paper explicitly labels reconstruction loss as a 'lightweight proxy' rather than the evaluation criterion, and Table 3 reports both reconstruction loss and NDCG@10 as separate quantities, so the objective is not defined in terms of the downstream metric. No parameter is fitted to a subset of the retrieval data and then reported as a prediction of the same quantity: the bit budget B is a user-set hyperparameter, and the comparison with RPG and homogeneous OPQ is external, with RPG/InforID sharing the same item embeddings and backbone. The references cite external prior work (RPG, TIGER, OPQ, k-means), with no load-bearing self-citations, no imported uniqueness theorem, and no ansatz smuggled in via the authors' own prior publications. The only soft spot is the Section 3.2 assumption that reconstruction-loss reduction is a useful proxy for retrieval utility, but the paper treats this as an empirically supported heuristic with controlled ablations rather than as an identity or a fitted prediction, so it is a validity concern, not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central method rests on user-set capacity hyperparameters B and M and on several modeling assumptions inherited from prior semantic ID work, plus the heuristic that minimizing subspace reconstruction loss is a useful signal for building prediction targets. No invented entities are introduced; the codebooks, slots, and parallel heads are existing constructs.

free parameters (2)

B: total bit budget = Not reported per dataset; sensitivity sweep 64-352 in Figure 3
Manually chosen capacity constraint; the greedy allocation distributes B across slots, so it controls effective ID length and codebook sizes.
M: number of candidate semantic slots = Not reported for main experiments; Table 1 varies 13-18 on Sports
Manually chosen candidate slot count after projection; it determines the granularity of the allocation search space.

assumptions (4)

domain assumption OPQ-transformed projected subspaces are meaningful semantic units for item ID construction.
Section 3.1 partitions the projected representation into M equal-size subspaces and treats each as a candidate slot, following OPQ [3,11]; if this partition is semantically arbitrary, the allocation has no useful target structure to exploit.
ad hoc to paper Subspace reconstruction loss is a valid proxy for downstream retrieval usefulness.
Section 3.2 chooses the slot with the largest marginal reduction in reconstruction loss as the allocation signal, while the actual goal is retrieval accuracy; the paper calls it a 'lightweight proxy' and only validates it on four datasets.
domain assumption Parallel factorization over retained slots is a valid approximation of target ID probability.
Section 3.3 factorizes P(c_i+|u) as a product over slots, inherited from RPG [10]; if slot tokens are strongly dependent given user history, this factorization loses information.
standard math k-means codebooks approximate the subspace distributions well enough under each bit allocation.
Section 3.2 uses k-means clustering [13] to build codebooks and reconstruction loss; k-means is not globally optimal but is a standard heuristic.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation." pith.science (2026). https://pith.science/paper/LPLLK2ZV

@misc{pith2026260809685,
  author       = {Pith},
  title        = {Pith review of: Adaptive Semantic Capacity Allocation for Parallel Generative Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LPLLK2ZV}},
  note         = {Machine review of arXiv:2608.09685}
}

read the original abstract

Autoregressive semantic ID recommenders are constrained by expensive beam-search decoding, which limits the practical length of item identifiers. Parallel generation methods alleviate this bottleneck by predicting all semantic ID tokens simultaneously, enabling longer IDs. However, existing semantic ID methods still rely on manually predefined and homogeneous ID structures, where both the number of semantic slots and the codebook size of each slot are treated as fixed hyperparameters. This ignores the heterogeneous capacity demands of different semantic subspaces and may allocate prediction capacity to slots with limited utility. We show that uniformly expanding semantic slots can provide limited gains, indicating redundant capacity in homogeneous semantic IDs. We propose InforID, a lightweight adaptive semantic target construction framework for parallel generative recommendation. InforID allocates a fixed capacity budget across candidate semantic slots, thereby jointly determining the effective ID length and slot-specific codebook sizes. Experiments demonstrate improved recommendation accuracy under comparable capacity budgets while preserving one-step parallel prediction.

Figures

Figures reproduced from arXiv: 2608.09685 by the authors.

**Figure 2.** Overview of InforID. Adaptive capacity allocation determines heterogeneous semantic IDs, removes zero-bit slots. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Sensitivity analysis of total bit budget [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 2 canonical work pages

[1]

Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D Lee, Deming Chen, and Tri Dao. 2024. Medusa: Simple llm inference acceleration framework with multiple decoding heads.arXiv preprint arXiv:2401.10774(2024)

arXiv 2024
[2]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PmLR, 1597–1607

2020
[3]

Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. 2013. Optimized product quantization.IEEE transactions on pattern analysis and machine intelligence36, 4 (2013), 744–755

2013
[4]

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
[5]

Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask- predict: Parallel decoding of conditional masked language models. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 6112–6121

work page 2019
[6]

Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozière, David Lopez-Paz, and Gabriel Synnaeve. 2024. Better & faster large language models via multi-token prediction.arXiv preprint arXiv:2404.19737(2024)

arXiv 2024
[7]

Jiatao Gu, James Bradbury, Caiming Xiong, Victor OK Li, and Richard Socher. 2017. Non-autoregressive neural machine translation.arXiv preprint arXiv:1711.02281 (2017)

arXiv 2017
[8]

Ruining He and Julian McAuley. 2016. 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. 507–517

2016

Show all 25 references

[9]

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
[10]

Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic ids in parallel for recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Min...

2025
[11]

Herve Jegou, Matthijs Douze, and Cordelia Schmid. 2010. Product quantization for nearest neighbor search.IEEE transactions on pattern analysis and machine intelligence33, 1 (2010), 117–128

2010
[12]

Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

2018
[13]

James B McQueen. 1967. Some methods of classification and analysis of multi- variate observations. InProc. of 5th Berkeley Symposium on Math. Stat. and Prob. 281–297

1967
[14]

Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith Hall, Daniel Cer, and Yinfei Yang. 2022. Sentence-t5: Scalable sentence encoders from pre- trained text-to-text models. InFindings of the association for computational lin- guistics: ACL 2022. 1864–1874

2022
[15]

Aleksandr V Petrov and Craig Macdonald. 2023. Generative sequential recom- mendation with gptrec.arXiv preprint arXiv:2306.11114(2023)

2023 arXiv
[16]

Aleksandr V Petrov and Craig Macdonald. 2024. RecJPQ: training large-catalogue sequential recommenders. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 538–547

2024
[17]

Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al
[18]

Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks.Advances in neural information processing systems27 (2014)

2014
[19]

Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning.Advances in neural information processing systems30 (2017)

2017
[20]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

2017
[21]

Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 11897–11916

2024
[22]

Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al
[23]

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

2024 arXiv
[2016]

Google’s neural machine translation system: Bridging the gap between human and machine translation.arXiv preprint arXiv:1609.08144(2016)

2016 arXiv
[2023]

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

2023

Pith tools

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

[1] [1]

Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D Lee, Deming Chen, and Tri Dao. 2024. Medusa: Simple llm inference acceleration framework with multiple decoding heads.arXiv preprint arXiv:2401.10774(2024)

arXiv 2024

[2] [2]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PmLR, 1597–1607

2020

[3] [3]

Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. 2013. Optimized product quantization.IEEE transactions on pattern analysis and machine intelligence36, 4 (2013), 744–755

2013

[4] [4]

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

[5] [5]

Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask- predict: Parallel decoding of conditional masked language models. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 6112–6121

work page 2019

[6] [6]

Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozière, David Lopez-Paz, and Gabriel Synnaeve. 2024. Better & faster large language models via multi-token prediction.arXiv preprint arXiv:2404.19737(2024)

arXiv 2024

[7] [7]

Jiatao Gu, James Bradbury, Caiming Xiong, Victor OK Li, and Richard Socher. 2017. Non-autoregressive neural machine translation.arXiv preprint arXiv:1711.02281 (2017)

arXiv 2017

[8] [8]

Ruining He and Julian McAuley. 2016. 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. 507–517

2016