Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Bridging Search and Recommendation through Latent Cross Reasoning

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

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

Pith's one-line read The paper claims that iteratively reasoning across search and recommendation histories — with cross-attention and GRPO-based reinforcement learning — lets a recommender selectively use search signals and beat strong baselines on Qilin and b

desk verdict Consistent gains from latent cross reasoning for search-aware recommendation, but the GRPO policy is never defined, making the headline RL results unreproducible as written. read the letter →

arxiv 2508.04152 v1 pith:VIDOKASV submitted 2025-08-06 cs.IR

Teng Shi , Weicong Qin , Weijie Yu , Xiao Zhang , Ming He , Jianping Fan , Jun Xu This is my paper
classification cs.IR
keywords search-enhancedrecommendationlatentreasoningcross-attentioncontrastivelearningGRPOreinforcementsequentialrecommendersystems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a recommender can decide which parts of a user's search history actually help predict the next item, instead of treating all search behaviors as equally useful. It proposes LCR-SER, which encodes search and recommendation histories separately, then runs iterative latent cross reasoning: at each step, the search representation reads from the recommendation history and vice versa, refining both. The authors claim this selective reasoning consistently outperforms strong baselines on three public benchmarks, and that contrastive learning plus GRPO-based reinforcement learning add further gains. If right, the contribution is a mechanism — not a new model family — for turning noisy cross-behavior signals into recommendation signals.

What carries the argument

Latent cross reasoning: the hidden states of the search and recommendation encoders are fed back into the same encoders for K steps, with cross-attention letting each behavior stream read the other's history. The mechanism is what does the selecting: it lets the model iteratively refine which search signals are useful, with the final-step states aggregated by target-aware attention for prediction.

What would settle it

Compute the exact gradient of Eq. (15) with respect to the MLP parameters in Eq. (10). If $\pi_\Phi(v_{L+1}\mid T_i)$ is not a well-defined, normalized probability over candidate items, the GRPO update cannot be executed and the RL gains in Table 3 row 5 would not be reproducible; alternatively, replace the ratio with a softmax-normalized score and compare against the reported gains.

Watch

Extended reading notes

Core claim

LCR-SER's central claim is that the reason search-enhanced recommendation underperforms is not lack of search data but lack of selection: irrelevant queries such as "Visa Apply" dilute the useful electronics signals. The framework attacks this with a latent reasoning loop. Starting from the final hidden states of two Transformer encoders, it alternates self-attention over each history with cross-attention from the other history, letting each side extract what is relevant to the other. A target-aware attention aggregation then weights the reasoning steps by similarity to the candidate item, a margin-based contrastive loss pulls cross-attentive states toward the target, and GRPO optimizes rank

Load-bearing premise

The GRPO update in Eq. (15) treats $\pi_\Phi(v_{L+1}\mid T_i)$ as a well-defined probability of the target item given a reasoning trajectory, but the paper never defines this distribution in terms of the MLP score in Eq. (10); if that likelihood is not properly normalized and differentiable, the reinforcement learning objective cannot be implemented as written.

Editorial extensions

If this is right

  • Search histories can be used without manual filtering: the model learns which search behaviors matter for a given target item.
  • Latent reasoning techniques developed for LLMs transfer to a non-textual, dual-behavior recommendation setting.
  • Reinforcement learning with ranking-metric rewards improves recommendation beyond supervised pre-training.
  • Reasoning depth can be chosen per sample through target-aware aggregation rather than a fixed depth.
  • The distance-to-target diagnostic gives a direct check that reasoning steps are moving representations toward the item.

Reading between the lines

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

  • I infer that if the selection mechanism is what matters, a natural stress test is to replace cross-attention with a learned gate over the search history; if the gate matches LCR-SER's gains, iterative cross-reasoning is sufficient but not necessary.
  • I infer the same latent cross-reasoning loop could be applied to other paired behavior streams, such as browsing and purchasing or impression and click.
  • I infer that a direct comparison of the GRPO variant against simpler reward-weighted fine-tuning would separate the effect of the policy constraint from the ranking reward, since the paper leaves the RL likelihood implicit.
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 LCR-SER, a search-enhanced recommendation model built around iterative latent cross reasoning. Separate Transformer encoders process the user's search and recommendation histories; a multi-step reasoning module applies self- and cross-attention to refine the two representations, a target-aware attention layer aggregates the reasoning states, and the final preference score is produced by an MLP. The model is trained with a binary cross-entropy loss, a target-aware contrastive loss (TCL), and a GRPO-style reinforcement-learning stage intended to optimize ranking metrics. Experiments on Qilin, KuaiSAR-Small, and KuaiSAR-Large compare against 15 baselines; Table 2 reports consistent improvements in HR@1/5/10 and NDCG@5/10, and ablations and analyses probe the contribution of each module, the effect of reasoning steps, hyperparameters, and inference latency.

Significance. If the empirical results hold, LCR-SER is a competitive search-enhanced recommender, and the idea of iteratively selecting search signals through cross-history reasoning is a useful contribution. The evaluation is broad, uses public benchmarks, and the central comparison with external baselines is not circular. The paper also reports an inference-latency analysis, which is helpful for practical assessment. However, the central claim currently rests on an underspecified and, as written, apparently unimplementable RL objective in Section 3.4; the statistical significance claims lack supporting variance information; and the mechanism analyses in Figures 4–5 partly restate the contrastive objective rather than independently validating the reasoning mechanism. These issues are fixable, but they are load-bearing for the reproducibility of the headline results.

major comments (4)
  1. [Section 3.4, Eqs. (13)–(16)] The GRPO objective is not well-defined. Eq. (15) uses a policy ratio πΦ(v_{L+1}|T_i)/πold(v_{L+1}|T_i), but πΦ is never defined from the architecture. The model outputs a scalar score via Eq. (10); no action distribution, normalization, or generation process is specified. The trajectories T_i are generated by a deterministic transformer with Gaussian noise injected into the initial states (Eq. (14)), so there is no stochastic policy with well-defined conditional probabilities. In addition, Eq. (16) is not the standard KL divergence between two probability distributions: for r = πref/πΦ it gives r − log r − 1, not p log(p/q)+(1−p)log((1−p)/(1−q)) for Bernoulli probabilities or the categorical KL. The reward is a ranking metric over a candidate list, while the policy ratio refers to a single target item, so credit assignment is unclear. Consequently, the RL gradient, the KL constraint, the
  2. [Section 4.2, Table 2] The statistical significance claim is unverifiable as reported. The table marks the best result with an asterisk and says a t-test gives p<0.05, but the paper reports no error bars, no number of seeds, no standard deviations, and no specification of what units the t-test is over (e.g., independent test users, repeated runs, or sampled item sets). Without this information, the claim that LCR-SER significantly outperforms the second-best method cannot be checked. Please report mean±std over multiple random seeds or a paired test with a clear description of the pairing, and make the test details reproducible. Key hyperparameters (K, N, γ, σ², λTCL, λKL, λReg) are also deferred to the supplementary material and should be stated in the main text or a public implementation.
  3. [Section 4.4.1, Figures 4 and 5] The evidence that reasoning 'extracts useful information' is partly self-confirming. The TCL loss in Eq. (9) explicitly minimizes d(e_{v_{L+1}}, h_s^(K)) − d(e_{v_{L+1}}, h~_s^(K)) + m, i.e., it directly pulls the final reasoning state toward the target embedding relative to the no-cross-attention state. Therefore, observing in Figure 4 that distances decrease with more reasoning steps, and in Figure 5 that MCA plus LTCL yields smaller distances, largely confirms that the optimizer minimized the quantity being plotted. This does not independently validate that latent cross reasoning selects beneficial search signals. Please present these figures explicitly as optimization diagnostics, or add an independent probe (e.g., attention weights against human relevance judgments, transfer/held-out performance without the auxiliary loss, or a qualitative case study) to support the mechanism claim.
  4. [Section 4.3.3, Table 3] The claim that adding LTCL 'further enhances performance' is not uniformly supported by the table. On Qilin, HR@5 decreases from 0.3804 (row ③) to 0.3793 (row ④) when LTCL is added, while NDCG@5 increases slightly. The text does not acknowledge this tradeoff. The authors should either explain the discrepancy (e.g., why NDCG@5 is the primary criterion) or adjust the claim so it is consistent with the reported numbers.
minor comments (5)
  1. [Eq. (5)] In the symmetric recommendation reasoning, the second output is written as h_s^(k) = FFN_r(...); this should presumably be h_r^(k), since h_s^(k) was already defined in Eq. (4). Please correct the notation.
  2. [Eq. (14)] The indicator I[i>1] makes the first trajectory deterministic. It would help to state explicitly why this is treated as a draw from πold, and how γ and σ² are chosen, since these control the effective exploration distribution.
  3. [Table 4] The latency analysis reports times, but not the batch size, number of users in the test set, or whether the reported time includes only the online part or the full inference including offline precomputation. Please clarify.
  4. [Figure 8] The labels in Figure 8(b) appear garbled or unreadable in the manuscript; please regenerate the figure so that axis labels and the legend are legible.
  5. [Section 4.1.1] The description of the data split is brief: Qilin uses leave-one-out following [23,25], while KuaiSAR uses chronological splitting following [19]. Please state exactly how training/validation/test sets are formed for each dataset, including whether the same users appear in multiple splits.

Circularity Check

1 steps flagged · score 4.0 of 10

Main benchmark results are externally validated, but the distance-to-target analyses in §4.4.1/§4.4.3 relabel the Eq. (9) training objective as evidence for reasoning; GRPO is under-specified but not circular.

  1. fitted input called prediction [Section 4.4.1 / Fig. 4 and Section 4.4.3 / Fig. 5 vs. Eq. (9), §3.3.1]
    "LTCL-S = max{d(e_{v_{L+1}}, h_s^{(K)}) − d(e_{v_{L+1}}, \tilde h_s^{(K)}) + m, 0} ... This objective explicitly promotes the extraction of complementary information relevant to the target ... As the number of reasoning steps increases, the hidden representations become progressively closer to the target ... These results demonstrate that the reasoning process effectively distills useful information from search history to benefit recommendation."

    The evidence offered for 'reasoning distills target-relevant information' is the Euclidean distance between hidden states and target embeddings. But that distance is exactly what the training objective in Eq. (9) minimizes (a margin loss forcing cross-attended states closer to e_{v_{L+1}} than no-cross-attention states), and Eq. (11) additionally trains representations to predict the target. Observing that trained representations are closer to the target is therefore a consequence of the loss, not an independent confirmation of the reasoning mechanism. Figure 5 makes the same move: it credits LTCL with reducing the distance that LTCL was defined to reduce. This is a self-confirming analysis; it does not by itself invalidate the external benchmark comparisons in Table 2.

full rationale

The headline comparison (Table 2) is against external baselines on public datasets (Qilin, KuaiSAR), so the main claim of improved search-enhanced recommendation is not circular. The incremental ablations in Table 3 are measured by ranking metrics and provide independent evidence for the components. The only genuine circularity I can pin to the paper's equations is the internal 'reasoning effectiveness' analysis: Fig. 4 and Fig. 5 use distance-to-target as the success measure, while Eq. (9) is a margin loss that explicitly minimizes that same distance. Thus those figures reduce, by construction, to the training objective and should not be cited as evidence that reasoning extracts useful signals. This is partial circularity in supporting analyses, not in the central external evaluation. Separately, Eq. (15) never defines the policy likelihood pi_Phi(v_{L+1}|T_i) in terms of the MLP score of Eq. (10); that is a reproducibility/correctness gap in the RL stage, but it is not a circular derivation, so it does not raise the circularity score. I found no load-bearing self-citation chain or imported uniqueness theorem; citations to UniSAR and others are baselines, data splits, or architectural precedents.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

No externally benchmarked constants are derived; the contribution is architectural. The main support for the contribution comes from the paper's own experiments, with the RL and reasoning claims resting on several under-specified assumptions.

free parameters (8)
  • Reasoning steps K = not reported; Figure 6 shows per-dataset tuning, with K=5 the maximum evaluated
    Chosen by hand or validation; affects accuracy and latency.
  • TCL margin m = not reported
    Required positive margin in the target-aware contrastive loss, Eq. (9).
  • lambda_TCL = not reported; swept from 1e-4 to 1 in Figure 7
    Weight of the contrastive loss in Eq. (12).
  • lambda_Reg = not reported
    L2 regularization weight in Eq. (12).
  • lambda_KL = not reported; swept from 1e-2 to 1e2 in Figure 8
    KL penalty in the GRPO objective, Eq. (15).
  • Noise scale gamma and sigma^2 = not reported
    Control Gaussian perturbation of initial reasoning states in Eq. (14).
  • Trajectory count N = not reported
    Number of sampled trajectories in the GRPO objective, Eq. (15).
  • Transformer dimension d, layers, heads = not reported
    Model size parameters left to the supplementary material.
assumptions (5)
  • domain assumption Separate encoders with causal masks preserve temporal order in both histories
    Section 3.2.2 assumes sequential histories can be modeled with causal self-attention.
  • ad hoc to paper The GRPO policy pi_Phi(v|T) is well-defined and differentiable
    Eqs. (15) and (16) use this likelihood without defining it; central to the RL training claim.
  • ad hoc to paper Ranking-metric rewards computed on candidate items do not leak test labels
    Section 3.4.2 uses HR@1 as reward during training; leakage safeguards are not stated.
  • ad hoc to paper Gaussian noise in initial reasoning states explores useful trajectories
    Eq. (14) perturbs reasoning but the exploration benefit is asserted, not demonstrated independently.
  • domain assumption Search and recommendation histories jointly contain signal for the next item
    Task formulation in Section 3.1 assumes the next item can be predicted from both histories.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Search and Recommendation through Latent Cross Reasoning." pith.science (2026). https://pith.science/paper/VIDOKASV

@misc{pith2026250804152,
  author       = {Pith},
  title        = {Pith review of: Bridging Search and Recommendation through Latent Cross Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VIDOKASV}},
  note         = {Machine review of arXiv:2508.04152}
}
read the original abstract

Search and recommendation (S&R) are fundamental components of modern online platforms, yet effectively leveraging search behaviors to improve recommendation remains a challenging problem. User search histories often contain noisy or irrelevant signals that can even degrade recommendation performance, while existing approaches typically encode S&R histories either jointly or separately without explicitly identifying which search behaviors are truly useful. Inspired by the human decision-making process, where one first identifies recommendation intent and then reasons about relevant evidence, we design a latent cross reasoning framework that first encodes user S&R histories to capture global interests and then iteratively reasons over search behaviors to extract signals beneficial for recommendation. Contrastive learning is employed to align latent reasoning states with target items, and reinforcement learning is further introduced to directly optimize ranking performance. Extensive experiments on public benchmarks demonstrate consistent improvements over strong baselines, validating the importance of reasoning in enhancing search-aware recommendation.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. S$^2$GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation

    cs.IR 2026-01 unverdicted novelty 7.0 of 10

    S²GR adds stepwise thinking tokens with contrastive supervision on codebook clusters to balance computational focus and ground reasoning paths in generative recommendation.

  2. 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.

Reference graph

Works this paper leans on

58 extracted references · 34 canonical work pages · cited by 2 Pith papers

  1. [1]

    Jia Chen, Qian Dong, Haitao Li, Xiaohui He, Yan Gao, Shaosheng Cao, Yi Wu, Ping Yang, Chen Xu, Yao Hu, et al. 2025. Qilin: A Multimodal Information Retrieval Dataset with APP-level User Sessions. arXiv preprint arXiv:2503.00501 (2025)

    work page Pith review arXiv 2025

  2. [2]

    Sunhao Dai, Ninglu Shao, Jieming Zhu, Xiao Zhang, Zhenhua Dong, Jun Xu, Quanyu Dai, and Ji-Rong Wen. 2024. Modeling user attention in music recom- mendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 761–774

    work page 2024

  3. [3]

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

  4. [4]

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

    arXiv 2025

  5. [5]

    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. arXiv preprint arXiv:2412.06769 (2024)

    arXiv 2024

  6. [6]

    Zhankui He, Handong Zhao, Zhaowen Wang, Zhe Lin, Ajinkya Kale, and Julian Mcauley. 2022. Query-Aware Sequential Recommendation. In Proceedings of the 31st ACM International Conference on Information &amp; Knowledge Management (Atlanta, GA, USA) (CIKM ’22). Association for Computing Machinery, New York, NY, USA, 4019–4023

    work page 2022

  7. [7]

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

  8. [8]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In 2018 IEEE International Conference on Data Mining (ICDM) . IEEE, 197–206

    2018

Show all 58 references
  1. [9]

    Enze Liu, Bowen Zheng, Xiaolei Wang, Wayne Xin Zhao, Jinpeng Wang, Sheng Chen, and Ji-Rong Wen. 2025. LARES: Latent Reasoning for Sequential Recom- mendation. arXiv preprint arXiv:2505.16865 (2025)

    2025 arXiv

  2. [10]

    Gustavo Penha, Ali Vardasbi, Enrico Palumbo, Marco De Nadai, and Hugues Bouchard. 2024. Bridging Search and Recommendation in Generative Retrieval: Does One Task Help the Other?. In Proceedings of the 18th ACM Conference on Recommender Systems. 340–349

    2024

  3. [11]

    Weicong Qin, Zelin Cao, Weijie Yu, Zihua Si, Sirui Chen, and Jun Xu. 2024. Explicitly integrating judgment prediction with legal document retrieval: a law- guided generative approach. In Proceedings of the 47th international ACM SIGIR conference on research and development in ...

    2024

  4. [12]

    Weicong Qin, Yi Xu, Weijie Yu, Chenglei Shen, Ming He, Jianping Fan, Xiao Zhang, and Jun Xu. 2025. MAPS: Motivation-Aware Personalized Search via LLM-Driven Consultation Alignment. arXiv preprint arXiv:2503.01711 (2025)

    2025 arXiv

  5. [13]

    Weicong Qin, Yi Xu, Weijie Yu, Chenglei Shen, Xiao Zhang, Ming He, Jianping Fan, and Jun Xu. 2024. Enhancing sequential recommendations through multi- perspective reflections and iteration. arXiv preprint arXiv:2409.06377 (2024)

    2024 arXiv

  6. [14]

    Weicong Qin, Yi Xu, Weijie Yu, Teng Shi, Chenglei Shen, Ming He, Jianping Fan, Xiao Zhang, and Jun Xu. 2025. Similarity= Value? Consultation Value Assessment and Alignment for Personalized Search. arXiv preprint arXiv:2506.14437 (2025)

    2025 arXiv

  7. [15]

    Weicong Qin, Weijie Yu, Kepu Zhang, Haiyuan Zhao, Jun Xu, and Ji-Rong Wen

  8. [16]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al . 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    2024 arXiv

  9. [17]

    Chenglei Shen, Xiao Zhang, Teng Shi, Changshuo Zhang, Guofu Xie, and Jun Xu

  10. [18]

    Information Sciences 702 (2025), 121915

    Uncertainty-aware evidential learning for legal case retrieval with noisy correspondence. Information Sciences 702 (2025), 121915

    2025

  11. [19]

    Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. UniSAR: Modeling User Transition Behaviors between Search and Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development ...

    2024

  12. [20]

    Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Han Li. 2025. Retrieval Augmented Generation with Collaborative Filtering for Per- sonalized Text Generation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Info...

    2025

  13. [21]

    Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Enyun Yu. 2025. Unified Generative Search and Recommendation. arXiv preprint arXiv:2504.05730 (2025)

    2025 arXiv

  14. [22]

    Chenglei Shen, Jiahao Zhao, Xiao Zhang, Weijie Yu, Ming He, and Jianping Fan. 2024. Generating Model Parameters for Controlling: Parameter Diffusion for Controllable Multi-Task Recommendation. arXiv preprint arXiv:2410.10639 (2024)

    2024 arXiv

  15. [23]

    Zihua Si, Zhongxiang Sun, Xiao Zhang, Jun Xu, Xiaoxue Zang, Yang Song, Kun Gai, and Ji-Rong Wen. 2023. When Search Meets Recommendation: Learning Dis- entangled Search Representation for Recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research ...

    2023

  16. [24]

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, et al. 2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419 (2025). Bridging Searc...

    2025 arXiv

  17. [25]

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

  18. [26]

    Zihua Si, Xueran Han, Xiao Zhang, Jun Xu, Yue Yin, Yang Song, and Ji-Rong Wen. 2022. A Model-Agnostic Causal Learning Framework for Recommendation Using Search Data. In Proceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France) (WWW ’22). Association for Computin...

    2022

  19. [27]

    Zhongxiang Sun, Kepu Zhang, Weijie Yu, Haoyu Wang, and Jun Xu. 2024. Logic rules as explanations for legal case retrieval. arXiv preprint arXiv:2403.01457 (2024)

    2024 arXiv

  20. [28]

    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. arXiv preprint arXiv:2503.22675 (2025)

    2025 arXiv

  21. [29]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)

    2008

  22. [30]

    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 systems 30 (2017)

    2017

  23. [31]

    Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Dewei Leng, Yanan Niu, Yang Song, Xiao Zhang, and Jun Xu. 2023. KuaiSAR: A Unified Search And Recommendation Dataset. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 5407–5411

    2023

  24. [32]

    Chuhan Wu, Fangzhao Wu, Mingxiao An, Tao Qi, Jianqiang Huang, Yongfeng Huang, and Xing Xie. 2019. Neural News Recommendation with Heterogeneous User Behavior. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Jo...

    2019

  25. [33]

    Jiayi Xie, Shang Liu, Gao Cong, and Zhenzhong Chen. 2024. UnifiedSSR: A Unified Framework of Sequential Search and Recommendation. In Proceedings of the ACM on Web Conference 2024 . 3410–3419

    2024

  26. [34]

    Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive learning for sequential recommendation. In 2022 IEEE 38th international conference on data engineering (ICDE) . IEEE, 1259– 1273

    2022

  27. [35]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

    2025 arXiv

  28. [36]

    Yuening Wang, Man Chen, Yaochen Hu, Wei Guo, Yingxue Zhang, Huifeng Guo, Yong Liu, and Mark Coates. 2024. Enhancing Click-through Rate Prediction in Recommendation Domain with Search Query Representation. In Proceedings of the 33rd ACM International Conference on Information a...

    2024

  29. [37]

    Bruce Croft

    Hamed Zamani and W. Bruce Croft. 2018. Joint Modeling and Optimization of Search and Recommendation. In Proceedings of the First Biennial Conference on Design of Experimental Search & Information Retrieval Systems, Bertinoro, Italy, August 28-31, 2018 (CEUR Workshop Proceeding...

    2018

  30. [38]

    Bruce Croft

    Hamed Zamani and W. Bruce Croft. 2020. Learning a Joint Search and Recommen- dation Model from User-Item Interactions. In Proceedings of the 13th International Conference on Web Search and Data Mining (Houston, TX, USA) (WSDM ’20). Association for Computing Machinery, New York...

    2020

  31. [39]

    Changshuo Zhang, Sirui Chen, Xiao Zhang, Sunhao Dai, Weijie Yu, and Jun Xu

  32. [40]

    Changshuo Zhang, Teng Shi, Xiao Zhang, Qi Liu, Ruobing Xie, Jun Xu, and Ji-Rong Wen. 2024. Modeling domain and feedback transitions for cross-domain sequential recommendation. arXiv preprint arXiv:2408.08209 (2024)

    2024 arXiv

  33. [41]

    Jing Yao, Zhicheng Dou, Ruobing Xie, Yanxiong Lu, Zhiping Wang, and Ji-Rong Wen. 2021. USER: A Unified Information Search and Recommendation Model Based on Integrated Behavior Sequence. In Proceedings of the 30th ACM Interna- tional Conference on Information ]&amp; Knowledge M...

    2021

  34. [42]

    Changshuo Zhang, Xiao Zhang, Teng Shi, Jun Xu, and Ji-Rong Wen. 2025. Test- Time Alignment for Tracking User Interest Shifts in Sequential Recommendation. arXiv preprint arXiv:2504.01489 (2025)

    2025 arXiv

  35. [43]

    Kepu Zhang, Teng Shi, Sunhao Dai, Xiao Zhang, Yinfeng Li, Jing Lu, Xiaoxue Zang, Yang Song, and Jun Xu. 2024. SAQRec: Aligning Recommender Systems to User Satisfaction via Questionnaire Feedback. In Proceedings of the 33rd ACM International Conference on Information and Knowle...

    2024

  36. [44]

    Kepu Zhang, Zhongxiang Sun, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Jun Xu. 2025. Trigger3: Refining Query Correction via Adaptive Model Selector. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 13260–13268

    2025

  37. [45]

    In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Reinforcing Long-Term Performance in Recommender Systems with User- Oriented Exploration Policy. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1850–1860

  38. [46]

    Kepu Zhang, Weijie Yu, Zhongxiang Sun, and Jun Xu. 2025. Syler: A framework for explicit syllogistic legal reasoning in large language models. arXiv preprint arXiv:2504.04042 (2025)

    2025 arXiv

  39. [47]

    Changshuo Zhang, Teng Shi, Xiao Zhang, Yanping Zheng, Ruobing Xie, Qi Liu, Jun Xu, and Ji-Rong Wen. 2024. QAGCF: Graph Collaborative Filtering for Q&A Recommendation. arXiv preprint arXiv:2406.04828 (2024)

    2024 arXiv

  40. [48]

    Yuting Zhang, Yiqing Wu, Ruidong Han, Ying Sun, Yongchun Zhu, Xiang Li, Wei Lin, Fuzhen Zhuang, Zhulin An, and Yongjun Xu. 2024. Unified Dual-Intent Translation for Jont Modeling of Search and Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discov...

    2024

  41. [49]

    Yang Zhang, Wenxin Xu, Xiaoyan Zhao, Wenjie Wang, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2025. Reinforced Latent Reasoning for LLM-based Recommendation. arXiv preprint arXiv:2505.19092 (2025)

    2025

  42. [50]

    Jujia Zhao, Wenjie Wang, Chen Xu, Xiuying Chen, Zhaochun Ren, and Suzan Verberne. 2025. Unifying Search and Recommendation: A Generative Paradigm Inspired by Information Theory. arXiv preprint arXiv:2504.06714 (2025)

    2025

  43. [51]

    Kepu Zhang, Weijie Yu, Sunhao Dai, and Jun Xu. 2024. Citalaw: Enhancing llm with citations in legal domain. arXiv preprint arXiv:2412.14556 (2024)

    2024 arXiv

  44. [52]

    Rui-Jie Zhu, Tianhao Peng, Tianhao Cheng, Xingwei Qu, Jinfa Huang, Dawei Zhu, Hao Wang, Kaiwen Xue, Xuanliang Zhang, Yong Shan, et al. 2025. A survey on latent reasoning. arXiv preprint arXiv:2507.06203 (2025)

    2025 arXiv

  45. [53]

    Xiao Zhang, Teng Shi, Jun Xu, Zhenhua Dong, and Ji-Rong Wen. 2024. Model- agnostic causal embedding learning for counterfactually group-fair recommen- dation. IEEE Transactions on Knowledge and Data Engineering (2024)

    2024

  46. [57]

    Kai Zhao, Yukun Zheng, Tao Zhuang, Xiang Li, and Xiaoyi Zeng. 2022. Joint Learning of E-Commerce Search and Recommendation with a Unified Graph Neural Network. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining (Virtual Event, AZ, USA) (...

    2022

  47. [2016]

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

    2016

  48. [2019]

    In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19)

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19). ACM, New York, NY, USA, 1441–1450

  49. [2024]

    arXiv preprint arXiv:2407.06083 (2024)

    A survey of controllable learning: Methods and applications in information retrieval. arXiv preprint arXiv:2407.06083 (2024)

    2024 arXiv

  50. [2025]

    arXiv preprint arXiv:2502.05171 (2025)

    Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach. arXiv preprint arXiv:2502.05171 (2025)

    2025 arXiv

Pith tools

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