Pith.

Information Retrieval · arXiv:2609.03522 v1 · submitted 3 Sep 2026 · 37 references

EPIC: Explicit Posterior Item Conditioning for Semantic ID Diffusion Recommendation

Tuan-Binh Tran, Thanh Tam Nguyen, Quoc Viet Hung Nguyen, Dung D. Le, Tung Kieu, Thanh Trung Huynh

The paper says

“Semantic ID (SID) generative recommendation predicts the next item by generating a short tuple of discrete tokens. Recent masked-diffusion methods improve this process through bidirectional context and flexible decoding, yet recommendation ultimately requires selecting among complete catalog items.”

more of the abstract

At each denoising step, a partial SID can correspond to multiple feasible items, while existing methods primarily reason through position-wise token predictions. We propose Explicit Posterior Item Conditioning (EPIC), which introduces explicit item-level competition into SID denoising. EPIC constructs a personalized posterior over feasible candidate items using the current generation context and the user's recent interactions, then projects this distribution back to unresolved SID positions to guide subsequent token decisions. The pretrained backbone remains frozen and requires no additional decoder forward pass. Experiments on four Amazon benchmarks show consistent improvements over strong baselines, while diagnostic analyses indicate that the gains primarily arise from personalized transition evidence that preserves promising item hypotheses during denoising.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 8 Sep 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.

EPIC introduces a well-motivated item-level posterior during masked diffusion denoising, with consistent gains, but the stress-test concern about target exclusion from the candidate support is a real load-bearing assumption that needs measurement. This line is wrong

1 major objection · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 37 references say.

The paper says, at §4.2.1–4.2.4 (Eqs. 8–9, 20)

The review did not quote this passage.

Pith says: Objection 1 of 1 · major

The assertion in §4.2.1 that token likelihood is “a reliable instrument for pruning hypotheses but a poor one for ranking the survivors” is load-bearing for the design, but target retention in the retained support S_t is never measured. If the backbone frequently excludes the ground-truth item from S_t when |C_t| > M, the posterior of Eq.

more of objection 1

(17) cannot score that item and item evidence cannot act on it at that step; recovery remains possible in principle through the backbone's own token logits or at later states with |C_t| ≤ M, so this is not an absolute failure mode, but the frequency of the disabling case is exactly what is unquantified. The concern is sharpened by a training/inference asymmetry: Eq. (21)'s token loss on the refined logits eℓ does give the adapter indirect gradient in truncated states, but the direct item-level supervision of Eq. (20) is active only when 2 ≤ |C_t| ≤ M, i.e., only when S_t = C_t and no pruning occurs. At inference, every test user passes through heavily masked states with |C_t| on the order of the full catalog (12,101–24,587 items, Table 1) and S_t is a hard top-M truncation with M ≤ 256, a regime in which the posterior's ranking is never directly supervised. The trajectory analyses of §6.2 (Fig. 3b) and §6.5 quantify recoveries but not the rate at which the target is already absent from S_t before feedback can act. I request a direct diagnostic — for example, the per-step target-retention rate in S_t given |C_t| > M, for both control and EPIC trajectories — together with an analysis of gate behavior in the high-ambiguity regime, and, if feasible, an M-ablation or an untruncated (M = |C_t|) upper bound to calibrate how much of the reported gain depends on the quality of the token-likelihood proposal.

This objection may describe the copy Pith gave the reviewer, not the paper.

6 minor objections

The paper says, at §5.3

Pith says: minor objection 1 of 6

The shared-settings list omits hyperparameters that directly control the magnitude of item feedback: the clip bound c and the floor ε in Eq. (23), the smoothing coefficient η in Eq. (20), and the SID-kernel temperature τ_S in Eq. (19); please report these values or state explicitly that they are fixed defaults.

The paper says, at §6.2 and §6.5

Pith says: minor objection 2 of 6

The aggregate “hard rescue” criterion used in §6.2 is described only informally; please define it precisely (for example, target absent from all final control beams and from the control top-128 pool) and confirm that the §6.5 case study is selected under the same criterion, so that the reported counts can be interpreted.

show 4 more

The paper says, at §6.3 (Fig. 4b)

Pith says: minor objection 3 of 6

The component-removal study does not include a w/o-gate ablation of the feedback gate in Eqs. (23)–(25), even though the gate is the component that determines how much item evidence enters token decoding; a removal result, or a short analysis of gate sign and magnitude across ambiguity bins, would complete the ablation picture.

The paper says, at §6.2–6.3 (Figs. 3 and 5)

Pith says: minor objection 4 of 6

The trajectory and history-perturbation analyses cover only three of the four datasets; please state whether Musical Instruments was excluded for space or for a technical reason, since otherwise readers cannot tell whether the mechanism generalizes to the smallest catalog.

The paper says, at §5.2 (Table 2)

Pith says: minor objection 5 of 6

The per-column ranking mixes locally reproduced results (†) with paper-reported values (‡) consolidated from four external sources; the significance tests correctly rely only on locally reproduced second-best methods, but one sentence noting that no ‡ cell would reverse a column's ordering under a matched protocol would strengthen the “strong baselines” claim.

The paper says, at §4.2.1

Pith says: minor objection 6 of 6

The sentence “Intuitively, token likelihood is a reliable instrument for pruning hypotheses” reads as an unsupported assertion; given the measurement gap raised in the major comment, please reword it as a hypothesis and point to evidence, including the new diagnostic if one is added.

What would settle it

For each denoising step t (T=4 down to 1) on the test set, compute the hit-rate of S_t: fraction of users where the ground-truth item i* is in S_t (Eq 9 with M=256 or chosen value). Also compute the fraction of users where i* is never in any S_t across all steps. Report these for each dataset. If the hit-rate at each step is above, say, 95%, the concern is largely mitigated. If it is lower, especially at early steps, it indicates that the backbone's pruning discards relevant items, and the gains are likely coming from cases where pruning is benign, not from the posterior itself recovering excluded items. This test directly measures the reliance on backbone quality as a proposal mechanism.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 8 Sep 2026

  1. Reader holds up confidence moderate
  2. Skeptic agrees 36 s · 17,385 tokens read, 4,022 written
  3. Editor wrote the verdict above 9 s · 18,329 tokens read, 821 written
  4. Referee major revision: 1 major objection, 6 minor 3 min 23 s · 17,991 tokens read, 25,748 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper does what it says: it builds an explicit item-level posterior over feasible candidates during each denoising step of a masked-diffusion SID recommender and feeds it back into token predictions. That is genuinely new relative to earlier work, which either reranks after generation or guides prefix autoregressive decoding with a pre-computed ranking. The core idea is simple and the paper executes it cleanly. The experiments are thorough—16 metric–dataset combinations, all improved, with statistical tests and ablations that convincingly attribute gains to the personalized transition evidence rather than generic scoring. Code is provided. I would bring this to a reading group as a solid example of how to add item-level reasoning into a generative decoding pipeline.

The soft spot is exactly the one the stress-test flags, and I think it is real. The retained candidate support St (Eq 9) is built by taking the top M items from the feasible set using the backbone's token log-likelihood. The paper asserts that token likelihood is a good pruner but a bad ranker, and then builds the learned posterior only over the survivors. But if the backbone's log-likelihood frequently prunes the ground-truth item before the posterior can act, the posterior cannot help. The diagnostic analysis in Fig 3b shows that inline feedback rescues items that post-hoc cannot reach, but it does not quantify the rate at which the target falls outside the top M in the first place. The gate module could partially compensate, but there is no ablation or analysis of gate behavior in high-ambiguity regimes where |Ct| is large. This does not invalidate the empirical results on these four Amazon datasets, but it means the method's generalizability depends on an unmeasured property of the backbone. I also note the evaluation is limited to four Amazon categories, which is a standard but narrow test bed.

Overall, the paper is a legitimate contribution within the generative recommendation subfield. The central claim that explicit item conditioning during denoising improves accuracy is supported. The soft spot is a caution, not a fatal flaw. I would send this to a serious peer review; the authors should be asked to measure how often the target falls outside St and to discuss regimes where the backbone's log-likelihood might be unreliable.

The strongest objection

EPIC constructs the retained candidate support S_t (Eq 9) by selecting the top M items from the feasible set C_t using the backbone's token log-likelihood sum (Eq 8). The paper asserts that token likelihood is 'a reliable instrument for pruning hypotheses but a poor one for ranking the survivors.' However, if the pruning itself eliminates the ground-truth item, the item posterior built over S_t cannot recover it at that or any subsequent step. The method's diagnostic analyses (Fig 3b, hard rescue) show recoveries, but they do not quantify how often the target is already excluded from S_t before feedback can act. This matters because frontier-aware learning (Eq 20) restricts direct item supervision to states where |C_t| ≤ M, i.e., when S_t = C_t and no pruning occurs. At inference, when |C_t| > M, pruning is active, but the posterior has never been trained to handle cases where the target lies outside the top M by token likelihood. The gate module (Eq 23–25) could in principle mitigate this, but no ablation or analysis of gate behavior in the high-ambiguity regime (large feasible set) is provided. Consequently, the method's effectiveness is contingent on the backbone's token log-likelihood being a sufficiently faithful proposal distribution—a property that is neither measured nor argued for. If the backbone systematically underestimates some relevant items (e.g., long-tail or collaborative-weak items), those items are silently removed from consideration, and the reported gains may reflect only a subset of users where the backbone already performs well. This does not invalidate the empirical results on the four Amazon benchmarks, but it is a load-bearing assumption about the interaction of pruning and posterior construction that limits understanding and generalization.

The reader's weakest_assumption is that S_t always contains the target. This is indeed a load-bearing assumption for the mechanism: if the target is pruned, the posterior cannot recover it. However, the paper's empirical results show consistent improvements, so the assumption appears to hold often enough on these datasets with the chosen backbone. My analysis does not find a fatal flaw; the concern is about generalization and unmeasured reliance. The reader's verict ACCEPT remains appropriate because the central claim of improved performance on the four benchmarks is supported by the evidence. The concrete test I propose would further validate the mechanism but is not necessary to accept the main claim. Thus I recommend no change to the verdict.

The referee's report

The contribution is solid and timely if the results hold: it provides evidence, with unusually careful controls, that item-level competition must act during denoising rather than only at final ranking, and it supplies a lightweight mechanism for doing so (0.391M trainable parameters, +34.2% inference latency, no additional backbone forward pass, exact initial no-op through zero-initialized residuals). The evaluation deserves explicit credit: matched inline-versus-post-hoc arms (Fig. 3a) isolate trajectory effects from reranking; outside-pool recoveries (Fig. 3b) refute a pure reranking explanation; wrong-user and reversed-history controls (Fig. 5a) rule out a generic popularity prior; each per-column second-best method was locally reproduced under a matched protocol, with Holm-corrected paired tests over users; and reproducibility scripts are provided. The main limits are scope and a mechanism gap: the experiments cover four Amazon 5-core categories with catalogs up to 24,587 items, a single backbone (LLaDA-Rec), and a single tokenizer, and the method's operation in the truncated-support regime (|C_t| > M) is asserted rather than measured, as detailed in the major comment.

Pith's reading of the paper (a machine's summary, not the authors')

EPIC claims that masked-diffusion Semantic ID recommenders lose accuracy because they only reason about individual SID tokens, leaving competition among complete catalog items implicit. The paper introduces an additive module atop a frozen backbone that, at every denoising step, builds an explicit posterior over the feasible complete items consistent with the partially resolved identifier, personalizes it with the user's recent interactions, and projects that item posterior back onto unresolved token positions. EPIC reports consistent improvements over strong baselines on four Amazon benchmarks across all 16 dataset--metric combinations, with relative gains between 1.2% and 16.8%, and argues the gains come from preserving promising item hypotheses during denoising rather than reranking final candidates.

What carries the argument. The central object is an explicit, learned item posterior $\pi_t$ over a retained candidate support $S_t$ of feasible items, computed at every denoising step from the current partial SID, the frozen backbone's hidden states, and the user's recent interaction history. The mechanism is: (1) the feasible set $C_t$ is computed exactly by matching resolved SID positions against the fixed catalog; (2) the backbone's token log-likelihood selects a top-$M$ retained support $S_t$ as proposal, deliberately not contributing to the posterior energy; (3) a pairwise item encoder plus history attention produces candidate-specific transition evidence $T_t(i,H_u)$, giving the personalized posterior; (4) this posterior is marginalized to unresolved positions and added as a clipped log-probability residual to the backbone logits, gated by an ambiguity-aware, zero-initialized gate that makes EPIC an exact initial no-op. Training uses frontier-aware item supervision, applied only when $2 \le |C_t| \le M$.

Core claim. The central claim is that making item-level competition explicit during SID denoising improves next-item recommendation. EPIC defines the feasible candidate set $C_t$ of catalog items consistent with the currently resolved SID positions, constructs a normalized item posterior over a retained support using a candidate-conditioned transition field over the user's recent history, and marginalizes this posterior back to unresolved positions, fusing it into the backbone token logits through an ambiguity-aware gate. The paper argues this is categorically different from post-hoc reranking: item evidence applied inline changes the denoising trajectory itself, preserving targets that would otherwise be pruned from the feasible set. The evidence includes matched inline-versus-post-hoc comparisons where inline feedback improves NDCG@5 by 15.7--26.8%, and hard-rescue cases where targets absent from the final control beam are recovered only by inline feedback. Ablations attribute the gain primarily to personalized transition evidence, with 12.4--17.9% degradation when the user's history is replaced by another user's history.

The premise it rests on. The paper assumes that the retained candidate support $S_t$, chosen as the top $M$ feasible items by the backbone's token log-likelihood, always contains the target item; if the target falls outside this support, the learned item posterior cannot recover it.

If the paper is right (Pith's inferences, not the authors')

EPIC shows that a frozen pretrained diffusion backbone plus a lightweight item-level adapter can outperform full fine-tuning of the backbone for next-item recommendation.

Applying item evidence during denoising is far more effective than using the same learned scorer only on final candidates: inline inference beats matched post-hoc inference by 15.7--26.8% NDCG@5 on Beauty, Sports, and Toys.

The gains depend on personalized transition evidence: replacing the user's own recent history with another user's history degrades NDCG@5 by 12.4--17.9%, whereas merely reversing the order degrades only 1.5--4.1%.

The choice of posterior construction matters: folding token likelihood into the posterior energy (product-of-experts or cavity messages) is consistently worse than keeping the learned item posterior independent of the proposal scores.

Partial resolution is a favorable cost--quality point: resolving only two of four SID positions gives the best Beauty NDCG@5 and is 1.72x faster than resolving all four.

EPIC's support-based scheme has a natural extension to approximate candidate construction where the dense catalog scan is replaced by an inverted SID index; the paper notes this possibility but its effectiveness under such approximation is untested.

The personalized transition field is effectively a differentiable item-to-item relevance model conditioned on current uncertainty, so EPIC could plausibly transfer to other generative interfaces where feasible sets over complete outputs can be formed, even without discrete SIDs.

The largest relative gains occur on ranking-sensitive metrics (16.8% NDCG@5 on Toys, 14.1% NDCG@10 on Toys), suggesting EPIC primarily improves the quality of the ranked head of the list rather than merely broad recall.

A concrete testable extension for future generative recommenders is to report how many top-$k$ misses are absent from all final beams and recoverable by inline item feedback, since that number directly separates trajectory preservation from reranking.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Semantic ID (SID) generative recommendation predicts the next item by generating a short tuple of discrete tokens. Recent masked-diffusion methods improve this process through bidirectional context and flexible decoding, yet recommendation ultimately requires selecting among complete catalog items. At each denoising step, a partial SID can correspond to multiple feasible items, while existing methods primarily reason through position-wise token predictions. We propose Explicit Posterior Item Conditioning (EPIC), which introduces explicit item-level competition into SID denoising. EPIC constructs a personalized posterior over feasible candidate items using the current generation context and the user's recent interactions, then projects this distribution back to unresolved SID positions to guide subsequent token decisions. The pretrained backbone remains frozen and requires no additional decoder forward pass. Experiments on four Amazon benchmarks show consistent improvements over strong baselines, while diagnostic analyses indicate that the gains primarily arise from personalized transition evidence that preserves promising item hypotheses during denoising.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 1.00

“Experiments on four Amazon benchmarks show consistent improvements over strong baselines, while diagnostic analyses indicate that the gains primarily arise from personalized transition evidence that preserves promising item hypotheses during denoising.”

How the main claim is stated: asserted 0.78

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.15 · human participants 0.06 · controlled comparison 0.88 · randomized assignment 0.01 · states quantitative result 0.12 · compares with prior results 0.97 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Semantic ID (SID) generative recommendation predicts the next item by generating a short tuple of discrete tokens. background 1.00 0.07 0.02 asserted 1.00
Recent masked-diffusion methods improve this process through bidirectional context and flexible decoding, yet recommendation ultimately requires selecting among complete catalog items. background 0.86 0.05 0.03 asserted 1.00
At each denoising step, a partial SID can correspond to multiple feasible items, while existing methods primarily reason through position-wise token predictions. gap 0.88 0.08 0.03 asserted 0.99
We propose Explicit Posterior Item Conditioning (EPIC), which introduces explicit item-level competition into SID denoising. contribution 0.98 0.90 0.12 asserted 0.99
EPIC constructs a personalized posterior over feasible candidate items using the current generation context and the user's recent interactions, then projects this distribution back to unresolved SID positions to guide subsequent token decisions. method 0.98 0.91 0.16 asserted 1.00
The pretrained backbone remains frozen and requires no additional decoder forward pass. method 0.93 0.76 0.03 asserted 1.00
Experiments on four Amazon benchmarks show consistent improvements over strong baselines, while diagnostic analyses indicate that the gains primarily arise from personalized transition evidence that preserves promising item hypotheses during denoising. result 0.85 0.96 0.73 asserted 0.78

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.95

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,525 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (7)
Figure 1 of the paper
Figure 1. Why item evidence must act during denoising. A
Figure 2 of the paper
Figure 2. EPIC within iterative SID denoising. At state x𝑡 , the frozen backbone provides token evidence, while the item-level branch encodes a retained support of feasible candidates and the user’s recent history. Personalized transition evidence yields an item posterior, which is marginalized to unresolved SID positions and fused with the backbone logits through an ambiguity￾aware gate. The refined token posterior produces x𝑡−1, and the procedure repeats. with a zero-initialized scale. The initial model therefore recovers backbone inference exactly and departs from it only to the extent…
Figure 3 of the paper
Figure 3. Inline item evidence changes denoising. Delayed
Figure 5 of the paper
Figure 5. EPIC depends on user-specific history; learned po
Figure 6 of the paper
Figure 6. EPIC helps across ambiguity regimes, while partial
Figure 7 of the paper
Figure 7. To avoid selecting an example based on semantic appear

Showing 6 of the paper's 7 figures. Figures are the authors', reproduced from the paper.

References (37)
  1. [1] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. 2021. Structured Denoising Diffusion Models in Discrete State-Spaces. In Advances in Neural Information Processing Systems on Pith
  2. [2] Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021. Au- toregressive Entity Retrieval. InProceedings of the International Conference on Learning Representations on Pith
  3. [3] 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 on Pith
  4. [4] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk on Pith
  5. [5] Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learn- ing Vector-Quantized Item Representation for Transferable Sequential Recom- menders. InProceedings of the ACM Web Conference 2023 on Pith
  6. [6] Yupeng Hou, Jiacheng Li, Xiangjun Fu, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. 2026. Bridging Language and Items for Retrieval and Recom- mendation: Benchmarking LLMs as Semantic Encoders. InProceedings of the 64th Annual Meeting of the Association for Computational... on Pith
  7. [7] 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 Mining V.2 on Pith
  8. [8] Yupeng Hou, Jianmo Ni, Zhankui He, Noveen Sachdeva, Wang-Cheng Kang, Ed H. Chi, Julian McAuley, and Derek Zhiyuan Cheng. 2025. ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation. InProceedings of the 42nd International Conference on Machine Learning on Pith
  9. [9] Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation. In2018 IEEE International Conference on Data Mining on Pith
  10. [10] Zhao Liu, Yichen Zhu, Yiqing Yang, Xiao Lv, Guoping Tang, Rui Huang, Qiang Luo, Ruiming Tang, Kun Gai, and Guorui Zhou. 2026. DiffGRM: Diffusion-based Generative Recommendation Model. InProceedings of the ACM Web Conference 2026 on Pith
  11. [11] Chen Ma, Peng Kang, and Xue Liu. 2019. Hierarchical Gating Networks for Se- quential Recommendation. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining on Pith
  12. [12] Lingyu Mu, Hao Deng, Haibo Xing, Jinxin Hu, Yu Zhang, Xiaoyi Zeng, and Jing Zhang. 2026. Masked Diffusion Generative Recommendation.arXiv preprint arXiv:2601.19501(2026) on PitharXiv
  13. [13] Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith B. 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 Linguistics: ACL 2022 on Pith
  14. [14] Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025. Large Language Diffusion Models. InAdvances in Neural Information Processing Systems on Pith
  15. [15] Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan H. Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. InAdvances in Neural Informa... on Pith
  16. [16] Kulin Shah, Bhuvesh Kumar, Neil Shah, and Liam Collins. 2025. Masked Diffusion for Generative Recommendation.arXiv preprint arXiv:2511.23021(2025) on PitharXiv
  17. [17] Teng Shi, Chenglei Shen, Weijie Yu, Shen Nie, Chongxuan Li, Xiao Zhang, Ming He, Yan Han, and Jun Xu. 2025. LLaDA-Rec: Discrete Diffusion for Parallel Seman- tic ID Generation in Generative Recommendation.arXiv preprint arXiv:2511.06254 (2025) on PitharXiv
  18. [18] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang on Pith
  19. [19] Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining on Pith
  20. [20] Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer Memory as a Differentiable Search Index. In Advances in Neural Information Processing Systems on Pith
show 17 more
  1. [21] Daria Tikhonovich, Oleg Sorokin, Vladislav Dodonov, Mariia Ulianova, and Ilya Murzin. 2026. Gryphon: A Unified Architecture for Semantic-ID Gener- ation and Item-Level Scoring in Industrial Recommendations.arXiv preprint arXiv:2606.08604(2026) on PithPith's reviewarXiv
  2. [22] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. Neural Discrete Representation Learning. InAdvances in Neural Information Processing Systems. EPIC on Pith
  3. [23] Junting Wang, Xinrui He, Yunzhe Li, and Hari Sundaram. 2026. Understand- ing Semantic IDs: From Item Representation to Item Selection in Generative Recommendation.arXiv preprint arXiv:2607.24995(2026) on PithPith's reviewarXiv
  4. [24] Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable Item Tokenization for Generative Recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management on Pith
  5. [25] Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua on Pith
  6. [26] Kaiwen Wei, Kejun He, Xiaomian Kang, Jie Zhang, Yuming Yang, Li Jin, Zhenyang Li, Jiang Zhong, He Bai, and Junnan Zhu. 2026. From Past To Path: Masked History Learning for Next-Item Prediction in Generative Recommendation. In Proceedings of the 64th Annual Meeting of the Assoc... on Pith
  7. [27] Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiangnan He. 2023. Generate What You Prefer: Reshaping Sequential Recom- mendation via Guided Diffusion. InAdvances in Neural Information Processing Systems on Pith
  8. [28] Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Gener- ative Recommendations. InProceedings of the 41st ... on Pith
  9. [29] Tingting Zhang, Pengpeng Zhao, Yanchi Liu, Victor S. Sheng, Jiajie Xu, Deqing Wang, Guanfeng Liu, and Xiaofang Zhou. 2019. Feature-level Deeper Self- Attention Network for Sequential Recommendation. InProceedings of the Twenty- Eighth International Joint Conference on Artifici... on Pith
  10. [30] Yuanbo Zhao, Ruochen Liu, Senzhang Wang, Jun Yin, Yuxin Dong, Huan Gong, Hao Chen, Shirui Pan, and Chengqi Zhang. 2026. SimGR: Escaping the Pit- falls of Generative Decoding in LLM-based Recommendation.arXiv preprint arXiv:2602.07847(2026) on PithPith's reviewarXiv
  11. [31] Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. In2024 IEEE 40th International Conference on Data Engineering on Pith
  12. [32] Bowen Zheng, Hongyu Lu, Yu Chen, Wayne Xin Zhao, and Ji-Rong Wen. 2026. Universal Item Tokenization for Transferable Generative Recommendation. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval on Pith
  13. [33] Qiyong Zhong, Jiajie Su, Yunshan Ma, Julian McAuley, and Yupeng Hou. 2025. Pctx: Tokenizing Personalized Context for Generative Recommendation.arXiv preprint arXiv:2510.21276(2025) on PitharXiv
  14. [34] Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-Rec: Self-Supervised Learning for Sequential Recommendation with Mutual Information Maximization. InPro- ceedings of the 29th ACM International Conference on Inf... on Pith
  15. [2016] Session-based Recommendations with Recurrent Neural Networks. In Proceedings of the International Conference on Learning Representations on Pith
  16. [2019] BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Management on Pith
  17. [2023] Diffusion Recommender Model. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval on Pith
Cited by 3 Pith papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment