Pith. sign in

REVIEW 4 major objections 5 minor 6 cited by

LARES: Latent Reasoning for Sequential Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a depth-recurrent latent reasoning module, which refines all item representations at each of several reasoning steps, lets sequential recommenders scale test-time computation without adding parameters, and that this…

desk verdict Solid empirical transfer of recurrent-depth reasoning to sequential recommendation, but the central test-time-scaling claim is missing its key control: same-FLOPs accuracy against a plain deeper transformer. read the letter →

arxiv 2505.16865 v2 pith:WFREHWVD submitted 2025-05-22 cs.IR

Enze Liu , Bowen Zheng , Xiaolei Wang , Wayne Xin Zhao , Jinpeng Wang , Sheng Chen , Ji-Rong Wen This is my paper
classification cs.IR
keywords sequentialrecommendationlatentreasoningtest-timecomputationscalingrecurrentdepthself-supervisedpre-trainingreinforcementpost-trainingcontrastivelearningnext-itemprediction
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

Sequential recommenders predict the next item from a user's past interactions, and their accuracy is normally limited by model size because the data are too sparse to train very large models. LARES tries a different route: keep the parameters fixed and let the model 'think' longer in a continuous latent space before predicting. It does this with a small recurrent block applied several times to the representations of all items in the sequence, so every token is refined at each reasoning step. A two-stage training scheme, self-supervised alignment of reasoning trajectories and steps followed by reinforcement post-training with recommendation metrics as rewards, teaches the model to reason in this latent space. On four Amazon categories, the paper reports that LARES outperforms all non-reasoning and reasoning baselines on every metric and, used as a plug-in, improves four existing recommenders. The payoff, if correct, is a parameter-free way to spend more computation at serving time for better recommendations.

What carries the argument

The central object is the depth-recurrent core-block $\mathcal{C}$: a small Transformer block (one or two layers) applied iteratively $k$ times to the latent representations of all input tokens, so the entire sequence is refined at every reasoning step. The reasoning process is written as $\prod_{i=1}^{k} P(S_i^u \mid S_{i-1}^u, S_0^u)$ with the initial hidden state $\boldsymbol{H}$ as a residual connection, giving effective depth $L_1 + L_2 k$ without adding parameters. The depth $k$ is sampled from a log-normal Poisson distribution during training and fixed at $\bar{k}+1$ at inference. Two contrastive (InfoNCE) alignment losses, one aligning final representations of positive pairs with the same target and same step count and one aligning an intermediate thought against the final thought, shape the reasoning path, and GRPO with NDCG/Recall rewards then explores better trajectories. This machinery carries the paper's empirical claim: the gain comes from spending more computation on the same parameters by refining every token, not from generating extra tokens.

What would settle it

Run the same LARES pipeline without the reinforcement-training data filter, or evaluate it on the sequences the filter discards, where the target item was never ranked in the top 100 by the pre-trained model across three trials. If the gains over the pre-training-only model vanish or reverse on that hard subset, the claim that depth-recurrent latent reasoning itself drives the reported improvements is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery claim is that depth-recurrent latent reasoning, iteratively refining the hidden states of all input tokens with the same Transformer block, is an effective and scalable paradigm for sequential recommendation. Expressed probabilistically, LARES models multi-step thinking as $P(T_u \mid S_u;\Theta)=\prod_{i=1}^{k}P(S_i^u \mid S_{i-1}^u,S_0^u;\Theta)$, with the last thought state used for prediction. Unlike the prior latent-reasoning recommender ReaRec, which appends one new token per step, LARES refines every token at every step, increasing computation density without adding parameters. The paper reports that this design, combined with self-supervised pre-training (trajectory-level and step-level alignment) and reinforcement post-training (GRPO with Recall/NDCG rewards), outperforms all listed baselines across four Amazon 2023 datasets on all Recall@K and NDCG@K metrics under full-item ranking with $p<0.01$, and that the same module improves four different backbones. The authors interpret this as evidence that scaling computation depth, rather than parameter count, is what matters for sequential recommendation.

Load-bearing premise

The load-bearing premise is the data filter used in the reinforcement post-training stage, which keeps only sequences where the pre-trained model already ranks the target item in the top 100 across three trials; if the discarded hard sequences are numerous or systematically different, the reported gains could come from an easier training distribution rather than from latent reasoning.

Editorial extensions

If this is right

  • Test-time computation can be scaled in sequential recommendation without growing the model: increasing the reasoning depth raises effective depth and accuracy up to an optimum, at roughly 10 percent extra inference latency over SASRec at comparable FLOPs.
  • LARES acts as a plug-in: SASRec, FMLP-Rec, BSARec, and TedRec all improve, with the text-enhanced TedRec gaining on average about 10.9 percent and 7.9 percent on the two reported datasets.
  • The full pipeline outperforms the pre-training-only variant, and removing any component degrades results, so both alignment objectives and reinforcement post-training each contribute to the gain.
  • The optimal reasoning depth is dataset-dependent, four for Instrument and three for Scientific, and too many steps hurt performance, indicating an overthinking regime.

Reading between the lines

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

  • Inference: the paper does not isolate the effect of the reinforcement-training data filter, which keeps only sequences where the pre-trained model already ranks the target item in the top 100 across three trials; if the discarded hard sequences are numerous or systematically different, part of the reported gain could come from an easier training distribution rather than from latent reasoning itsel
  • Inference: the same all-token refinement recipe should transfer to other sparse-sequence tasks such as session-based recommendation or retrieval-style candidate generation, wherever the input is a short sequence of IDs and the output is a ranking over a large item table.
  • Inference: adaptive per-sequence depth, which the paper lists as future work, could be implemented by early stopping based on the entropy or confidence of the predicted distribution at each step, so simple sequences would cost fewer FLOPs on average.
  • Inference: because the method adds no parameters, LARES-style depth scaling could ride on top of large text-enhanced recommenders, letting practitioners trade compute at serving time rather than retraining bigger models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes LARES, a depth-recurrent latent reasoning framework for sequential recommendation. The architecture consists of a pre-block that maps item embeddings into a latent space and a core-block that is iteratively applied to refine all input token representations for a variable number of reasoning steps, enabling test-time compute scaling without adding parameters. Training proceeds in two stages: self-supervised pre-training with trajectory-level and step-level alignment objectives, followed by reinforcement post-training based on GRPO with ranking-metric rewards. The authors evaluate LARES on four Amazon 2023 subsets, reporting large improvements over non-reasoning and reasoning baselines, ablations of the training components, compatibility experiments with four backbones, and an inference-time analysis.

Significance. If the reported gains hold, LARES would be a useful contribution: it offers a way to increase the effective depth of sequential recommenders at test time with no parameter increase, and it demonstrates a two-stage training recipe (self-supervised alignment plus reinforcement post-training) for latent reasoning in item-ID spaces. The paper includes several strong practices: full-ranking evaluation over the entire item pool, leave-one-out evaluation, compatibility tests with multiple backbones, ablation studies, and a public anonymous code link. The empirical claims are assessed against external benchmarks on standard datasets, so there is no circularity in the central evaluation. However, the headline comparison and the RL stage contain methodological gaps that must be addressed before the claims can be accepted.

major comments (4)
  1. [§3.4.4, Table 4 and Figure 2] The headline comparison is confounded with total test-time compute. LARES with a 2-layer pre-block, a 2-layer core-block, and 4 reasoning steps has effective depth 10 (2+2×4), while the Base backbones in Figure 2 and the baselines in Table 2 are standard shallow models. Table 4 reports only inference time for SASRec at matched FLOPs, not recommendation accuracy for those deeper SASRec variants. To support the claim that recurrent latent reasoning, rather than simply more computation, produces the reported gains, the authors need an accuracy comparison at matched FLOPs or matched effective depth, for example by adding the accuracy of a 10-layer SASRec and other stacked-Transformer baselines to Figure 2 or Table 4.
  2. [§2.4.3, Data Selection] The RPT data selection discards all instances where the pre-trained model cannot rank the target item in the top 100 across three inference trials. The paper does not report how many training sequences are removed on each dataset, nor whether the evaluation set is filtered in the same way. If a substantial fraction of hard sequences is removed, the RPT gains reported in Table 3 could reflect an easier training distribution rather than improved latent reasoning. Please report the filtering statistics and evaluate LARES both on the unfiltered test set and on the hard subset (e.g., the instances that the pre-trained model fails to rank in the top 100), so the reader can separate the effect of data selection from the effect of reinforcement learning.
  3. [§2.4.1, Eqs. (16)–(17)] The adaptation of GRPO to the latent reasoning setting is not derived. The paper states that π(y_i|x) is reformulated as a product over reasoning steps of the probability of recommending the target item at each step. This is not the generation probability of the latent thought trajectory that the policy actually produces, and no argument is given for why this product is a valid surrogate policy probability for the GRPO objective. Since the clipped importance ratio and the KL regularization term both depend on this quantity, the RL objective is not formally grounded as stated. Please provide a derivation or an explicit justification of this surrogate objective and state what approximation is being made.
  4. [§3.2 and Table 2] All reported results are point estimates with no standard deviations, confidence intervals, or number of seeds. The claim of statistical significance (p<0.01) in Table 2 is not verifiable without information about the number of repeated runs and the variance of the metrics. Please add repeated-run statistics for the main comparison and for the ablation table, or at least state how many seeds were used and report the spread.
minor comments (5)
  1. [Figure 1 caption] There is a typo: 'repsectively' should read 'respectively'.
  2. [References] References [5] and [6] are the same paper and should be merged into a single entry.
  3. [§2.2, Eqs. (4)–(7)] The notation is inconsistent: S_u is the input sequence, S^i_u is used for the thought at step i in Eq. (4), while T_0 and T_i denote latent thoughts in Eqs. (6)–(7). Please align the notation so the initial input, the initial latent state, and the intermediate thoughts are distinguishable.
  4. [Table 4 caption] The meaning of the 'Step' column for SASRec and the 'Params'/'FLOPs' multiplier notation (e.g., 4x, 2x+2x) is not defined in the caption. Please clarify how the matched-FLOPs configuration was constructed.
  5. [§3.4.1, Figure 2] For TedRec, the text-enhanced backbone, the paper does not describe how the latent reasoning module is attached to the text and ID fusion architecture. Please add a sentence explaining the adaptation for text-enhanced backbones.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LARES' central claims are empirical, evaluated on held-out full-ranking benchmarks, and no prediction reduces to a fitted input or self-citation chain by construction.

full rationale

The paper's central claims—that LARES outperforms all baselines (Section 3.2) and that it consistently improves four backbones (Section 3.4.1)—are empirical statements tested against external Amazon 2023 benchmarks using full-ranking Recall@K and NDCG@K. No equation defines the predicted metric in terms of the training objective or a fitted parameter: the recurrent architecture in Eqs. (5)-(9) is a model definition, the SPT losses in Eqs. (11)-(15) are auxiliary training objectives, and the RPT reward in Eq. (18) reuses the evaluation metrics as a training signal, which is standard practice and not a by-construction identity with the held-out evaluation. The data-selection step in Section 2.4.3 ('we exclude instances where the pre-trained model fails to rank the target item within the top 100 positions across three independent inference trials') is self-referential in choosing the RPT training distribution, but it is a data-filtering choice rather than a fitted parameter renamed as a prediction, and it does not make the reported test results equivalent to the filter's output. The matched-FLOPs discussion in Table 4 compares inference time but not matched-FLOPs accuracy; that is an experimental-comparability or correctness concern, not a circularity. The paper's citations to prior work by the same group (e.g., TedRec, FMLP-Rec, scaling-law studies) are contextual or baseline-related and are not load-bearing support for any derivation. No uniqueness theorem or ansatz is imported from the authors' own prior work to force the chosen architecture, so the derivation chain is self-contained rather than circular.

Assumptions & free parameters 9 free parameters · 5 assumptions · 1 invented entities

The central claim is empirical; most 'free parameters' are training hyperparameters tuned on validation, which is standard ML practice. The most consequential choices are the top-100 data selection threshold and the RL reward metric, both of which directly influence the measured gains and are not reported with sensitivity analysis. The domain assumptions about latent reasoning and the GRPO adaptation are the main unverified premises.

free parameters (9)
  • alpha (TLA weight) = 0.1, 0.2, 0.3, or 0.4 (tuned)
    Weight for trajectory-level alignment in Eq. (15). Chosen by validation performance; Figure 4 shows sensitivity.
  • gamma (SLA weight) = 0.1, 0.3, 0.5, or 0.7 (tuned)
    Weight for step-level alignment in Eq. (15). Tuned per dataset.
  • beta (KL regularization in GRPO) = 0.5 or 1.0
    Controls divergence from reference policy in Eq. (16).
  • sigma2 (log-normal Poisson variance) = 0.5
    Variance of the sampling distribution for reasoning depth k in Eq. (8). Fixed by hand.
  • kbar (mean reasoning steps) = 3, 4, or 6 (tuned)
    Mean iteration count for the depth-recurrent core block. Selected per dataset.
  • sigma1 (initial latent state std) = 1.0
    Standard deviation of the Gaussian initialization for T0 in Eq. (7).
  • G (rollout number in GRPO) = 4
    Number of sampled trajectories per input in the RL stage.
  • reward metric choice = Recall@5 or Recall@10
    Which metric is used as the RL reward in Eq. (18); selected from two options.
  • top-100 data selection threshold = 100
    Instances where the target is not ranked in the top 100 by the pre-trained model are excluded from RPT training (Section 2.4.3). A hand-chosen threshold that directly affects the training distribution.
assumptions (5)
  • domain assumption The factorization P(v_{n+1}|S_u) = P(v_{n+1}|T_u, S_u) * P(T_u|S_u) holds (Eq. 2).
    Assumes the latent reasoning trajectory T_u is a sufficient summary for next-item prediction.
  • domain assumption Repeated application of the same core-block on all tokens yields better representations as depth increases.
    The core mechanism of LARES; the paper provides empirical evidence but no theoretical justification.
  • standard math InfoNCE is a valid objective for aligning trajectory-level and step-level representations.
    Uses the standard contrastive loss from Oord et al.; assumes it encourages the intended invariance.
  • domain assumption The GRPO objective can be applied with a policy probability redefined as the product of per-step next-item probabilities.
    Section 2.4.1 replaces the token-generation probability with product over latent steps of target-item probabilities; this is an unproven adaptation.
  • domain assumption The final latent state at the last position represents the user's preference.
    Common in sequential recommendation; the paper uses T_k[-1] as the user representation without justification.
invented entities (1)
  • Latent thought trajectory T_u
    purpose: A sequence of hidden states that iteratively refine the user representation during 'reasoning'.
    Internal model construct with no external measurement or falsifiable handle; it is defined by the architecture and training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LARES: Latent Reasoning for Sequential Recommendation." pith.science (2026). https://pith.science/paper/WFREHWVD

@misc{pith2026250516865,
  author       = {Pith},
  title        = {Pith review of: LARES: Latent Reasoning for Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFREHWVD}},
  note         = {Machine review of arXiv:2505.16865}
}
read the original abstract

Sequential recommender systems have become increasingly important in real-world applications that model user behavior sequences to predict their preferences. However, existing sequential recommendation methods predominantly rely on non-reasoning paradigms, which may limit the model's computational capacity and result in suboptimal recommendation performance. To address these limitations, we present LARES, a novel and scalable LAtent REasoning framework for Sequential recommendation that enhances model's representation capabilities through increasing the computation density of parameters by depth-recurrent latent reasoning. Our proposed approach employs a recurrent architecture that allows flexible expansion of reasoning depth without increasing parameter complexity, thereby effectively capturing dynamic and intricate user interest patterns. A key difference of LARES lies in refining all input tokens at each implicit reasoning step to improve the computation utilization. To fully unlock the model's reasoning potential, we design a two-phase training strategy: (1) Self-supervised pre-training (SPT) with dual alignment objectives; (2) Reinforcement post-training (RPT). During the first phase, we introduce trajectory-level alignment and step-level alignment objectives, which enable the model to learn recommendation-oriented latent reasoning patterns without requiring supplementary annotated data. The subsequent phase utilizes reinforcement learning (RL) to harness the model's exploratory ability, further refining its reasoning capabilities. Comprehensive experiments on real-world benchmarks demonstrate our framework's superior performance. Notably, LARES exhibits seamless compatibility with existing advanced models, further improving their recommendation performance. Our code is available at https://anonymous.4open.science/r/LARES-E458/.

Figures

Figures reproduced from arXiv: 2505.16865 by the authors.

Figure 1
Figure 1. Overall framework of LARES. A and C denote the pre-block and core-block, repsectively. “TLA” and “SLA” represent the Trajectory￾Level Alignment and Step-Level Alignment. represented as follows: 𝑃 (𝑇𝑢 |𝑆𝑢; Θ) = Ö 𝑘 𝑖=1 𝑃 (𝑡𝑖 |𝑆𝑢, 𝑡<𝑖 ; Θ), (3) where 𝑘 is the number of reasoning steps and 𝑡𝑖 and 𝑡<𝑖 denote the generated thoughts at the 𝑖-th step and preceding the 𝑖-th step, respectively. Scaling Latent Reasoning With … view at source ↗
Figure 3
Figure 3. Performance of different reasoning steps on Instru [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Performance of different alignment coefficients [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. BEAR: Towards Beam-Search-Aware Optimization for Recommendation with Large Language Models

    cs.IR 2026-01 reject novelty 7.0 of 10

    BEAR is a cheap token-level top-B regularizer for LLM-based recommendation, but its central claim that this condition is necessary for beam-search survival is incorrect.

  2. From Understanding to Action: Feedback-Grounded Policy Discovery for Generative Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    A feedback-grounded framework discovers recommendation policies by their measured advantage over intent-only baselines and distills them into two latent tokens of a lightweight Semantic-ID recommender.

  3. Hierarchical Latent Reasoning for LLM-based Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    HiLaR aligns LLM latent reasoning states with temporally quantized user preference hierarchies and optimizes them with layer-aware process rewards, improving recommendation accuracy on four Amazon datasets.

  4. EvoReason: Self-Evolving Reasoning Primitive-Guided On-Policy Distillation for Latent Reasoning in Generative Recommendation

    cs.IR 2026-07 conditional novelty 5.0 of 10

    A self-evolving, primitive-guided distillation method for latent reasoning in generative recommendation reports consistent ranking gains over existing latent-reasoning baselines.

  5. The Latent Space: Foundation, Evolution, Mechanism, Ability, and Outlook

    cs.AI 2026-04 accept novelty 5.0 of 10

    A large survey organizes latent-space work in language-based models by foundation, evolution, four mechanisms, seven abilities, and open challenges.

  6. Bridging Search and Recommendation through Latent Cross Reasoning

    cs.IR 2025-08 conditional novelty 5.0 of 10

    A latent cross reasoning model with contrastive learning and GRPO reinforcement learning improves search-enhanced recommendation on Qilin and KuaiSAR.

Reference graph

Works this paper leans on

60 extracted references · 9 canonical work pages · cited by 6 Pith papers

  1. [1]

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Pi- otr Nyczyk, and Torsten Hoefler. 2024. Graph of Thoughts: Solving Elaborate Problems with Large Language Models. InThirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Confer...

    2024

  2. [2]

    Jianxin Chang, Chen Gao, Yu Zheng, Yiqun Hui, Yanan Niu, Yang Song, Depeng Jin, and Yong Li. 2021. Sequential Recommendation with Graph Neural Networks. InSIGIR ’21: The 44th International ACM SIGIR Conference on Research and Devel- opment in Information Retrieval, Virtual Event, Canada, July 11-15, 2021, Fernando Diaz, Chirag Shah, Torsten Suel, Pablo Ca...

    arXiv 2021

  3. [3]

    Haolin Chen, Yihao Feng, Zuxin Liu, Weiran Yao, Akshara Prabhakar, Shelby Heinecke, Ricky Ho, Phil Mui, Silvio Savarese, Caiming Xiong, and Huan Wang

    work page

  4. [4]

    Huiyuan Chen, Yusan Lin, Menghai Pan, Lan Wang, Chin-Chia Michael Yeh, Xiaoting Li, Yan Zheng, Fei Wang, and Hao Yang. 2022. Denoising Self-Attentive Sequential Recommendation. InRecSys ’22: Sixteenth ACM Conference on Rec- ommender Systems, Seattle, W A, USA, September 18 - 23, 2022, Jennifer Golbeck, F. Maxwell Harper, Vanessa Murdock, Michael D. Ekstra...

    arXiv 2022

  5. [6]

    Zhipeng Chen, Yingqian Min, Beichen Zhang, Jie Chen, Jinhao Jiang, Daixuan Cheng, Wayne Xin Zhao, Zheng Liu, Xu Miao, Yang Lu, Lei Fang, Zhongyuan Wang, and Ji-Rong Wen. 2025. An Empirical Study on Eliciting and Improving R1-like Reasoning Models.CoRRabs/2503.04548 (2025). doi:10.48550/ARXIV. 2503.04548 arXiv:2503.04548

    arXiv doi:10.48550/arxiv.2503.04548 2025

  6. [7]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai D...

    arXiv doi:10.48550/arxiv.2501.12948 2025

  7. [8]

    Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein

    Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein

    work page

  8. [9]

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, and Wen- feng Liang. 2024. DeepSeek-Coder: When the Large Language Model Meets Programming - The Rise of Code Intelligence.CoRRabs/2401.14196 (2024). doi:10.48550/ARXIV.2401.14196 arXiv:2401.14196

    arXiv doi:10.48550/arxiv.2401.14196 2024

Show all 60 references
  1. [10]

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training Large Language Models to Reason in a Continuous Latent Space.CoRRabs/2412.06769 (2024). doi:10.48550/ARXIV.2412. 06769 arXiv:2412.06769

    2024 doi arXiv

  2. [11]

    Sheng, Jiajie Xu, Guanfeng Liu, and Xiaofang Zhou

    Yongjing Hao, Tingting Zhang, Pengpeng Zhao, Yanchi Liu, Victor S. Sheng, Jiajie Xu, Guanfeng Liu, and Xiaofang Zhou. 2023. Feature-Level Deeper Self-Attention Network With Contrastive Learning for Sequential Recommendation.IEEE Trans. Knowl. Data Eng.35, 10 (2023), 10112–1012...

    2023

  3. [12]

    Ruining He and Julian J. McAuley. 2016. Fusing Similarity Models with Markov Chains for Sparse Sequential Recommendation. InIEEE 16th International Confer- ence on Data Mining, ICDM 2016, December 12-15, 2016, Barcelona, Spain, Francesco Bonchi, Josep Domingo-Ferrer, Ricardo B...

    2016

  4. [13]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  5. [14]

    Balázs Hidasi, Massimo Quadrana, Alexandros Karatzoglou, and Domonkos Tikk. 2016. Parallel Recurrent Neural Network Architectures for Feature-rich Session-based Recommendations. InProceedings of the 10th ACM Conference on Recommender Systems, Boston, MA, USA, September 15-19, ...

    2016

  6. [15]

    Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards Universal Sequence Representation Learning for Recom- mender Systems. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14...

    2022

  7. [16]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, Alex If- timie, Alex Karpenko, Alex Tachard Passos, Alexander Neitz, Alexander Prokofiev, Alexander Wei, Allison Tam, Ally Bennett, Ananya...

    2024 doi arXiv

  8. [17]

    Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InIEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018. IEEE Computer Society, 197–206. doi:10.1109/ ICDM.2018.00035 LARES: Latent Reasoning for Sequent...

    2018

  9. [18]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s Verify Step by Step. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, Ma...

    2024

  10. [19]

    Enze Liu, Bowen Zheng, Wayne Xin Zhao, and Ji-Rong Wen. 2025. Bridg- ing Textual-Collaborative Gap through Semantic Codes for Sequential Rec- ommendation.CoRRabs/2503.12183 (2025). doi:10.48550/ARXIV.2503.12183 arXiv:2503.12183

    2025 doi arXiv

  11. [20]

    Mingrui Liu, Sixiao Zhang, and Cheng Long. 2024. Facet-Aware Multi-Head Mixture-of-Experts Model for Sequential Recommendation.CoRRabs/2411.01457 (2024). doi:10.48550/ARXIV.2411.01457 arXiv:2411.01457

    2024 doi arXiv

  12. [21]

    Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, Jiao Sun, and Abhinav Rastogi. 2024. Improve Mathematical Reasoning in Language Models by Automated Process Supervision. CoRRabs/2406.06592 (2024). doi:10.48550/ARXIV.2...

    2024 doi arXiv

  13. [23]

    Massimo Quadrana, Alexandros Karatzoglou, Balázs Hidasi, and Paolo Cremonesi

  14. [24]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. InAdvances in Neural Inform...

    2023

  15. [25]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized Markov chains for next-basket recommendation. InProceedings of the 19th International Conference on World Wide Web, WWW 2010, Raleigh, North Carolina, USA, April 26-30, 2010, Mic...

    2010

  16. [26]

    Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, and Sashank J. Reddi. 2025. Reasoning with Latent Thoughts: On the Power of Looped Transformers.CoRRabs/2502.17416 (2025). doi:10.48550/ARXIV.2502.17416 arXiv:2502.17416

    2025 doi arXiv

  17. [27]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.CoRRabs/2402.03300 (2024). doi:10.48550/ARXIV.2402.03300 arXiv:2402.03300

    2024 doi arXiv

  18. [28]

    Yehjin Shin, Jeongwhan Choi, Hyowon Wi, and Noseong Park. 2024. An Atten- tive Inductive Bias for Sequential Recommendation beyond the Self-Attention. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of ...

    2024

  19. [29]

    Uriel Singer, Haggai Roitman, Yotam Eshel, Alexander Nus, Ido Guy, Or Levi, Idan Hasson, and Eliyahu Kiperwasser. 2022. Sequential Modeling with Multiple Attributes for Watchlist Recommendation in E-Commerce. InWSDM ’22: The Fifteenth ACM International Conference on Web Search...

    2022

  20. [30]

    Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. 2025. Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs. arXiv:2505.00127 [cs.CL] https://arxiv.org/abs/2505.00127

    2025 arXiv

  21. [31]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  22. [32]

    Yong Kiam Tan, Xinxing Xu, and Yong Liu. 2016. Improved Recurrent Neural Networks for Session-based Recommendations. InProceedings of the 1st Workshop on Deep Learning for Recommender Systems, DLRS@RecSys 2016, Boston, MA, USA, September 15, 2016. ACM, 17–22. doi:10.1145/29884...

    2016

  23. [33]

    Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Wu Jian, and Yuning Jiang. 2025. Think Before Recommend: Unleashing the Latent Reasoning Power for Sequential Recommendation.CoRRabs/2503.22675 (2025). doi:10. 48550/ARXIV.2503.22675 arXiv:2503.22675

    2025 doi arXiv

  24. [34]

    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, WSDM 2018, Marina Del Rey, CA, USA, February 5-9, 2018, Yi Chang, Chengxia...

    2018

  25. [35]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Con- gcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding...

    2025 doi arXiv

  26. [36]

    Hao Wang, Jianxun Lian, Mingqi Wu, Haoxuan Li, Jiajun Fan, Wanyue Xu, Chaozhuo Li, and Xing Xie. 2023. ConvFormer: Revisiting Transformer for Sequential User Modeling.CoRRabs/2308.02925 (2023). doi:10.48550/ARXIV. 2308.02925 arXiv:2308.02925

    2023 doi arXiv

  27. [37]

    Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan

  28. [38]

    Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive Learning for Sequential Recommendation. In 38th IEEE International Conference on Data Engineering, ICDE 2022, Kuala Lumpur, Malaysia, May 9-12, 2022. IEEE, 1259–12...

    2022

  29. [39]

    Lanling Xu, Zhen Tian, Bingqian Li, Junjie Zhang, Daoyuan Wang, Hongyu Wang, Jinpeng Wang, Sheng Chen, and Wayne Xin Zhao. 2024. Sequence-level Semantic Representation Fusion for Recommender Systems. InProceedings of the 33rd ACM International Conference on Information and Kno...

    2024

  30. [40]

    Songpei Xu, Shijia Wang, Da Guo, Xianwen Guo, Qiang Xiao, Fangjian Li, and Chuanjiang Luo. 2025. An Efficient Large Recommendation Model: Towards a Resource-Optimal Scaling Law.CoRRabs/2502.09888 (2025). doi:10.48550/ARXIV. 2502.09888 arXiv:2502.09888

    2025 doi arXiv

  31. [41]

    Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025. SoftCoT: Soft Chain- of-Thought for Efficient Reasoning with LLMs. arXiv:2502.12134 [cs.CL] https: //arxiv.org/abs/2502.12134

    2025 arXiv

  32. [42]

    An Yan, Shuo Cheng, Wang-Cheng Kang, Mengting Wan, and Julian J. McAuley

  33. [43]

    Session-Based Recommendation with Graph Neural Networks. InThe Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty- First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Ar...

    2019 doi

  34. [44]

    Ling Yang, Zhaochen Yu, Bin Cui, and Mengdi Wang. 2025. ReasonFlux: Hier- archical LLM Reasoning via Scaling Thought Templates.CoRRabs/2502.06772 (2025). doi:10.48550/ARXIV.2502.06772 arXiv:2502.06772

    2025 doi arXiv

  35. [45]

    Gonzalez, and Bin Cui

    Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E. Gonzalez, and Bin Cui. 2024. Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models. InAdvances in Neural Information Pro- cessing Systems 38: Annual Conference on Neural I...

    2024

  36. [46]

    Jose, and Xiangnan He

    Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M. Jose, and Xiangnan He. 2019. A Simple Convolutional Generative Network for Next Item Recommendation. InProceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM 2019, Melbourne, V...

    2019

  37. [47]

    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 Generative Recommendations. InForty-first International ...

    2024

  38. [48]

    Gaowei Zhang, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, and Ji- Rong Wen. 2024. Scaling Law of Large Sequential Recommendation Models. InProceedings of the 18th ACM Conference on Recommender Systems, RecSys 2024, Bari, Italy, October 14-18, 2024, Tommaso Di Noia, Pasqual...

    2024

  39. [49]

    CosRec: 2D Convolutional Neural Networks for Sequential Recommenda- tion. InProceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019, Wenwu Zhu, Dacheng Tao, Xueqi Cheng, Peng Cui, Elke A. Runde...

    2019

  40. [50]

    Bencheng Yan, Shilei Liu, Zhiyuan Zeng, Zihao Wang, Yizhen Zhang, Yujin Yuan, Langming Liu, Jiaqi Liu, Di Wang, Wenbo Su, Pengjie Wang, Jian Xu, and Bo Zheng. 2025. Unlocking Scaling Law in Industrial Recommendation Systems with a Three-step Paradigm based Large User Model.CoR...

    2025 doi

  41. [51]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji-Rong Wen. 2021. RecBole: Towards a Unified, Com...

    2021

  42. [52]

    Bowen Zheng, Zihan Lin, Enze Liu, Chen Yang, Enyang Bai, Cheng Ling, Wayne Xin Zhao, and Ji-Rong Wen. 2024. A Large Language Model Enhanced Sequential Recommender for Joint Video and Comment Recommendation.CoRR abs/2403.13574 (2024). doi:10.48550/ARXIV.2403.13574 arXiv:2403.13574

    2024 doi arXiv

  43. [53]

    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. InCIKM ’20: The 29th ACM International Conference on Informatio...

    2020

  44. [54]

    Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. 2022. Filter-enhanced MLP is All You Need for Sequential Recommendation. InWWW ’22: The ACM Web Conference 2022, Virtual Event, Lyon, France, April 25 - 29, 2022, Frédérique Laforest, Raphaël Troncy, Elena Simperl, Deepak Agar...

    2022

  45. [55]

    Peilin Zhou, Qichen Ye, Yueqi Xie, Jingqi Gao, Shoujin Wang, Jae Boum Kim, Chenyu You, and Sunghun Kim. 2023. Attention Calibration for Transformer- based Sequential Recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management,...

    2023

  46. [56]

    Junjie Zhang, Beichen Zhang, Wenqi Sun, Hongyu Lu, Wayne Xin Zhao, Yu Chen, and Ji-Rong Wen. 2025. Slow Thinking for Sequential Recommendation.CoRR abs/2504.09627 (2025). doi:10.48550/ARXIV.2504.09627 arXiv:2504.09627

    2025 doi arXiv

  47. [57]

    Wayne Xin Zhao, Yupeng Hou, Xingyu Pan, Chen Yang, Zeyu Zhang, Zihan Lin, Jingsen Zhang, Shuqing Bian, Jiakai Tang, Wenqi Sun, Yushuo Chen, Lanling Xu, Gaowei Zhang, Zhen Tian, Changxin Tian, Shanlei Mu, Xinyan Fan, Xu Chen, and Ji-Rong Wen. 2022. RecBole 2.0: Towards a More U...

    2022

  48. [2016]

    In4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, Yoshua Bengio and Yann LeCun (Eds.)

    Session-based Recommendations with Recurrent Neural Networks. In4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, Yoshua Bengio and Yann LeCun (Eds.). http://arxiv.org/abs/1511.06939

    2016 arXiv

  49. [2017]

    Personalizing Session-based Recommendations with Hierarchical Recurrent Neural Networks. InProceedings of the Eleventh ACM Conference on Recommender Systems, RecSys 2017, Como, Italy, August 27-31, 2017, Paolo Cremonesi, Francesco Ricci, Shlomo Berkovsky, and Alexander Tuzhili...

    2017

  50. [2019]

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformer. InProceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019, Wenwu Zhu, Dacheng Tao, Xueqi Che...

    2019

  51. [2024]

    doi:10.48550/ ARXIV.2411.04282 arXiv:2411.04282

    Language Models are Hidden Reasoners: Unlocking Latent Reasoning Capabilities via Self-Rewarding.CoRRabs/2411.04282 (2024). doi:10.48550/ ARXIV.2411.04282 arXiv:2411.04282

    2024 doi arXiv

  52. [2025]

    doi:10.48550/ARXIV.2502.05171 arXiv:2502.05171

    Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach.CoRRabs/2502.05171 (2025). doi:10.48550/ARXIV.2502.05171 arXiv:2502.05171

    2025 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.