Pith. sign in

REVIEW 3 major objections 5 minor 56 references

From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read Recommender systems are heading toward a planning stage that decides what an exposure should accomplish before choosing an item, this paper argues.

desk verdict A well-written vision paper that names a real future direction—semantic planning—but its load-bearing assumption (semantic IDs as a compositional planning vocabulary) is asserted, not shown; still worth a serious referee. read the letter →

arxiv 2607.09540 v4 pith:7KMVD46E submitted 2026-07-10 cs.IR

classification cs.IR
keywords recommendersystemssemanticIDsplanningmulti-stakeholderrecommendationgenerativetargetrawinformationutilizationatscale
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 reads two decades of recommender-system research as a shift in how systems use information at scale: industrial systems long relied on raw, opaque item IDs for exact lookup and memorization; more recently, richer semantic signals are being folded into the IDs themselves, turning identity into a structured, model-facing object. The authors argue that this shift is not just about generative retrieval but sets up a next stage they call semantic planning, in which a system first decides what the next exposure should accomplish (a semantic target) and only then instantiates that target as an item, offer, or generated creative. A sympathetic reader would care because this is a concrete proposal for making multi-stakeholder objectives—user, platform, provider—explicit decision variables rather than implicit score adjustments.

What carries the argument

The semantic target: an intermediate, explicitly represented planning variable specifying what the next exposure should accomplish, expressed in the structured vocabulary of semantic IDs. It carries the argument by giving the system a legible layer between user context and item selection, so that multi-stakeholder objectives can be reasoned about directly and unmet demand can be surfaced when no catalog item realizes the target.

What would settle it

A study that takes a real catalog, learns semantic IDs from item attributes, and asks whether the target 'a quiet room with free cancellation near a train station for an anxious traveler' can be composed from the learned token vocabulary; if such targets cannot be expressed in the ID space, the planning layer loses its grounding and collapses into latent intent modeling.

Watch

Extended reading notes

Core claim

The central claim is that recommender systems are evolving along an axis of information utilization: raw IDs, then semantic information around IDs, then semantic IDs. Semantic IDs, discrete token sequences that encode item content and structure, move part of the semantic information into the identity layer itself, providing a shared vocabulary across domains, modalities, and tasks. Once such a vocabulary exists, the paper argues, the system can separate the question 'what should the next exposure accomplish?' from 'which item should be shown?' The intermediate output—a semantic target representing intent, need, or desired properties—becomes an explicit multi-stakeholder decision variable, an

Load-bearing premise

Semantic IDs can serve as a legible, transferable vocabulary for expressing abstract planning targets even when those targets—like 'low-risk city-center accommodation for hesitant bookers'—are not directly present as attributes of any existing catalog item.

Editorial extensions

If this is right

  • Semantic IDs become the substrate not just for retrieval but for an intermediate control layer that predicts exposure goals before item selection.
  • Recommendations can be instantiated flexibly: the same semantic target may be realized by an item, a sponsored offer, a message, or a generated creative, changing how advertising and content are integrated.
  • Inventory gaps become explicit signals of unmet demand, feedable back to providers rather than absorbed into item scores.
  • Evaluation must move from fixed relevance judgments toward 'planning resonance'—tracking whether sequences of exposures move users toward fulfilled intent.
  • Semantic targets can serve as a shared interface between ranking, advertising, and content systems, coordinating heterogeneous platform components around one objective.

Reading between the lines

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

  • If semantic targets become explicit, the paper's framework implies a new division of labor in industrial pipelines: a planning model trained on objectives, and an instantiation model that realizes targets—an architecture that does not yet exist publicly.
  • The same logic could extend to conversational and agentic recommenders, where semantic targets are natural objects for user confirmation or negotiation, something the paper touches on only in passing.
  • A testable extension would be to benchmark retrieval vs planning on the hotel-style scenario: does a planning layer measurably change exposure outcomes, or does it reduce to re-ranking under a different loss?
  • The target-grounding challenge suggests that semantic IDs may need to be generative rather than purely catalog-derived, otherwise planning targets are constrained to combinations of existing item attributes.
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

3 major / 5 minor

Summary. This position paper proposes a historical and forward-looking narrative for recommender systems organized around how they utilize information at scale. It argues that the field has progressed from raw, semantically opaque IDs as operational anchors, through a stage in which rich semantic signals (content, context, multimodality, cross-domain structure) are used alongside raw IDs, to the current stage in which semantic information is increasingly encapsulated in semantic IDs. The paper then introduces 'semantic planning' as a possible next stage: instead of directly retrieving an existing item, the system would first predict a semantic target—a representation of what the next exposure should accomplish for users, platforms, and providers—and only then instantiate that target as an item, offer, message, or generated creative. The paper discusses implications for model design, evaluation (proposing 'planning resonance' as a new gold standard), and industrial deployment, and it explicitly lists open challenges including ID stability, target grounding, and objective observability.

Significance. If the semantic-planning thesis is accepted, the paper provides a coherent synthesis connecting several active lines of research—semantic-ID tokenization, generative retrieval, cross-domain and multimodal recommendation, and multi-stakeholder optimization—under a single narrative. Its main value is as a research agenda: it articulates a concrete intermediate representation (the semantic target) that could guide future work, and it does so with appropriate hedging, framing semantic planning as 'a possible direction' and acknowledging substantial open challenges in Section 4.3. The paper makes no empirical claims and does not overclaim certainty; its strength is the clarity of the historical synthesis and the novelty of the proposed planning layer. The main risk is that the argument's load-bearing assumption—that catalog-derived semantic IDs can serve as a compositional, legible vocabulary for abstract planning targets—is asserted rather than demonstrated.

major comments (3)
  1. [§4.1] The central claim that 'structured semantic IDs can provide a more grounded representation through which such targets can be expressed' is the key premise of the semantic-planning thesis, but no evidence is provided. Existing semantic-ID methods (e.g., RQ-VAE tokenization in [28], [12]) are trained to reconstruct or index existing catalog items; they are not designed to form a compositional language in which novel, inventory-absent targets such as 'low-risk city-center accommodation for hesitant bookers' can be expressed. If semantic targets cannot be represented in semantic-ID space, the planning layer must fall back on free-text or continuous representations, and the claimed distinction from latent intent modeling (which Section 4.1 explicitly contrasts with semantic planning) collapses. Section 4.3 lists 'target grounding' as an open challenge, which is an admission that this gap is u
  2. [§4.1, Figure 1] The 'semantic target' is not defined at a level that makes the proposal testable or even crisply separable from existing latent intent models. Figure 1 labels the target as '(Intent, Need, Attribute)', but the paper does not specify what space the target lives in: is it a sequence of semantic-ID tokens, a vector, a structured set of attributes, or a natural-language utterance? What is the training objective for the planning layer, and how is it supervised given that targets are not directly observed? Without at least one concrete representation and a sketch of the learning problem, the distinction between 'predicting a semantic target' and 'directly predicting item scores from user intent' remains a matter of framing rather than a substantive architectural claim. The authors should provide a formal or at least operational definition of a semantic target, or explicitly state that the prop
  3. [§4.3] The proposed evaluation concept 'planning resonance' is introduced as the new gold standard, but it is not operationalized. The definition given—'the degree to which a system's trajectory of decisions navigates users toward fulfilled intent'—is not measurable as stated, and the citation [46] is about LLM-based user simulators rather than a formal evaluation metric. For a position paper this could be acceptable if framed as an open research direction, but the paper presents it as a consequence of the semantic-planning shift. The authors should clarify whether planning resonance is intended as an evaluation metric, a qualitative ideal, or a research challenge, and if the former, outline how it could be quantified (e.g., via user simulation, long-term outcome measures, or multi-stakeholder utility).
minor comments (5)
  1. [Front matter] The ACM Reference Format block and the running header contain placeholder text ('Conference acronym ’XX', 'Trovato et al.', year 2018, fake DOI). These should be replaced with correct venue, year, and DOI before publication.
  2. [Figure 1] The diagram is dense and the arrows labeled 'guide' and 'signal unmet demand' are not fully explained in the caption. Consider clarifying the flow and defining all symbols, or simplifying the figure to improve readability.
  3. [§2] Minor style: 'Firstly' and 'Secondly' in the second paragraph are slightly informal; 'First' and 'Second' would be more standard in a technical paper.
  4. [References] Some references are incomplete (e.g., [22] appears to be missing venue/page details) and several 2026 entries may need verification. Please ensure all citations are fully resolved, particularly since the paper's argument relies on the timeliness of semantic-ID work.
  5. [§4.3] The phrase 'evaluation agents' is introduced with quotes but not defined; it would help to specify what distinguishes an evaluation agent from a standard user simulator, or to link it more concretely to the cited work.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a forward-looking position/vision piece with no fitted parameters or derivation that reduces to its inputs.

full rationale

The manuscript contains no equations, no fitted parameters, and no quantitative predictions; its central claim (semantic planning as the next stage after semantic IDs) is introduced as 'one possible direction' (Section 1) rather than derived from the cited methods. The paper explicitly lists target grounding, ID stability, and objective observability as open challenges (Section 4.3), which shows it is not claiming a closed or forced result. The authors' self-citations ([5], [29], [30], [36]) are used only as background examples of industrial, cross-domain, or contextual techniques; none is invoked as a uniqueness theorem, a forcing axiom, or the source of the semantic-planning proposal. It also explicitly contrasts semantic planning with intent modeling and conversational recommendation, so the possible worry that the proposal is merely a renamed existing idea is addressed rather than concealed. The concern that catalog-derived semantic IDs may not compose into abstract targets is a support gap acknowledged by the paper, but it is not circularity: no 'prediction' is equivalent by construction to an input, and no load-bearing conclusion is imported solely from the authors' prior work. Therefore the honest finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 2 invented entities

The paper is a position paper and introduces no fitted parameters. Its load-bearing premise—that semantic IDs can express abstract planning targets—is an untested domain/postulated assumption. Two new conceptual entities (semantic target, planning resonance) are introduced without operational handles.

assumptions (3)
  • domain assumption Industrial recommender systems are multi-stakeholder systems with often conflicting user, platform, and provider objectives.
    Adopted from [1],[14],[19] and used in Section 1 to motivate why the field must move beyond user-item matching toward semantic planning.
  • ad hoc to paper Semantic IDs can serve as a legible, transferable vocabulary for intermediate planning targets.
    Central technical premise stated in Section 4.1: semantic IDs 'provide a more grounded representation' for expressing targets. No evidence is provided; Section 4.3 lists target grounding as an open challenge.
  • domain assumption The Cranfield-style static evaluation paradigm is insufficient for evaluating goal-oriented recommendation trajectories.
    Invoked in Section 4.3 to justify the proposed 'planning resonance' evaluation; a methodological value judgment, not an empirically established fact.
invented entities (2)
  • Semantic target (planning layer)
    purpose: An intermediate decision variable representing what the next exposure should accomplish (user need, platform objective, provider content) before item selection or creative generation.
    Introduced as a new abstraction in Sections 4.1–4.2; no operational definition, implementation, or falsifiable prediction is provided.
  • Planning resonance
    purpose: A proposed evaluation gold standard: the degree to which a trajectory of exposures navigates users toward fulfilled intent.
    Named in Section 4.3 as the evaluation counterpart of semantic planning; not formally defined or measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale." pith.science (2026). https://pith.science/paper/7KMVD46E

@misc{pith2026260709540,
  author       = {Pith},
  title        = {Pith review of: From Raw IDs to Semantic Planning: How Recommender Systems Utilize Information at Scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7KMVD46E}},
  note         = {Machine review of arXiv:2607.09540}
}
read the original abstract

The evolution of recommender systems can be explored by asking how they utilize information at scale. Throughout most of the historical period under consideration during the past two decades, industrial systems have relied on raw IDs, which are discrete, globally unique, and semantically opaque identifiers that enable exact lookup, logging, and item-specific memorization at scale. Over time, however, recommender systems have sought to utilize richer sources of information, including item content, context, multimodal signals, and cross-domain structure. This development has led to a new stage in which part of such information is no longer used solely as auxiliary features around item identity, but is increasingly encapsulated in semantic IDs that provide a more structured, model-facing form of identity. We argue that this shift goes beyond the rise of generative recommendation over traditional methods. Indeed, it reflects a broader evolution in how recommender systems utilize information under industrial-scale constraints. This paper looks at the past, present, and future to examine three connected questions: why raw IDs dominated the early development of recommender systems, why semantic information is increasingly being encapsulated in IDs today, and what may come next once recommendations move beyond semantic retrieval. In particular, we introduce semantic planning as a possible future direction in which the system first predicts the semantic target of the next exposure, and only then instantiates that target as a specific item or generated creative. We further argue that such a shift may require changes not only in model design but also in evaluation and in the way recommender systems coordinate the objectives of users, platforms, and providers.

Figures

Figures reproduced from arXiv: 2607.09540 by the authors.

Figure 1
Figure 1. Paradigm shift from semantic retrieval to semantic plan￾ning. (a) Semantic Retrieval maps the context directly to a semantic ID to fetch an existing catalog item; (b) Semantic Planning separates the semantic target from the instantiation. Semantic retrieval maps the current user state to an identifier that resolves to a specific item, while semantic planning aligns the current context with a semantic target specifyi… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 10 linked inside Pith

  1. [28]

    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. InNeurIPS

  2. [12]

    McAuley, and Wayne Xin Zhao

    Yupeng Hou, Zhankui He, Julian J. McAuley, and Wayne Xin Zhao. 2023. Learn- ing Vector-Quantized Item Representation for Transferable Sequential Recom- menders. InWWW. ACM, 1162–1171

  3. [46]

    Se-eun Yoon, Zhankui He, Jessica Maria Echterhoff, and Julian J. McAuley. 2024. Evaluating Large Language Models as Generative User Simulators for Conversa- tional Recommendation. InNAACL-HLT. Association for Computational Linguis- tics, 1490–1504

  4. [1]

    Himan Abdollahpouri, Gediminas Adomavicius, Robin Burke, Ido Guy, Dietmar Jannach, Toshihiro Kamishima, Jan Krasnodebski, and Luiz Augusto Pizzato. 2020. Multistakeholder recommendation: Survey and research directions.User Model. User Adapt. Interact.30, 1 (2020), 127–158

  5. [2]

    Gediminas Adomavicius, Bamshad Mobasher, Francesco Ricci, and Alexander Tuzhilin. 2011. Context-Aware Recommender Systems.AI Mag.32, 3 (2011), 67–80

  6. [3]

    Honghui Bao, Wenjie Wang, Xinyu Lin, Fengbin Zhu, Teng Sun, Fuli Feng, and Tat- Seng Chua. 2025. Heterogeneous User Modeling for LLM-based Recommendation. InRecSys. ACM, 145–154

  7. [4]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Anil, Zakaria Haque, Lichan Hong, Vihan Jain, Xiaobing Liu, and Hemal Shah

  8. [5]

    Akira Fukumoto, Aghiles Salah, Sarthak Shrivastava, Alexandru Tatar, Yannick Schwartz, Vincent Michel, and Lee Xiong. 2025. Contrastive Conditional Embed- dings for Item-based Recommendation at E-commerce Scale. InRecSys. ACM, 931–934

Show all 56 references
  1. [6]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In IJCAI. ijcai.org, 1725–1731

  2. [7]

    Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, Xinyang Yi, Lexi Baugher, Baykal Cakici, Ed H

    Ruining He, Lukasz Heldt, Lichan Hong, Raghunandan H. Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, Xinyang Yi, Lexi Baugher, Baykal Cakici, Ed H. Chi, Cristos Goodrow, Ningren Han, He Ma, Rómer Rosales, Abby Van Soest, Devansh Tan...

  3. [8]

    Ruining He and Julian J. McAuley. 2016. VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback. InAAAI. AAAI Press, 144–150

  4. [9]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. InSIGIR. ACM, 639–648

  5. [10]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural Collaborative Filtering. InWWW. ACM, 173–182

  6. [11]

    Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian J. McAuley. 2023. Large Language Models as Zero-Shot Conversational Recommenders.CoRR abs/2308.10053 (2023)

  7. [13]

    Peiyu Hu, Wayne Lu, and Jia Wang. 2026. From IDs to Semantics: A Genera- tive Framework for Cross-Domain Recommendation with Adaptive Semantic Tokenization. InAAAI. AAAI Press, 14874–14882

  8. [14]

    Dietmar Jannach and Himan Abdollahpouri. 2023. A survey on multi-objective recommender systems.Frontiers in Big Data6 (2023), 1157899. doi:10.3389/fdata. 2023.1157899

  9. [15]

    2010.Recommender systems: an introduction

    Dietmar Jannach, Markus Zanker, Alexander Felfernig, and Gerhard Friedrich. 2010.Recommender systems: an introduction. Cambridge University Press

  10. [16]

    Jie Jiang, Xinxun Zhang, Enming Zhang, Yuling Xiong, Jun Zhang, Jingwen Wang, Huan Yu, Yuxiang Wang, Hao Wang, Xiao Yan, and Jiawei Jiang. 2026. End-to-End Semantic ID Generation for Generative Advertisement Recommendation.CoRR abs/2602.10445 (2026)

  11. [17]

    Jinqiu Jin, Yang Zhang, Fuli Feng, and Xiangnan He. 2025. Generative Multi- Target Cross-Domain Recommendation.CoRRabs/2507.12871 (2025)

  12. [18]

    Kyungho Kim, Sunwoo Kim, Geon Lee, Jinhong Jung, and Kijung Shin. 2025. Multi-behavior Recommender Systems: A Survey. InPAKDD (4) (Lecture Notes in Computer Science). Springer, 435–452

  13. [19]

    Roger Zhe Li. 2023. Metric Optimization and Mainstream Bias Mitigation in Recommender Systems. arXiv:2311.06689 [cs.IR] https://arxiv.org/abs/2311.06689

  14. [20]

    Junling Liu, Chao Liu, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is ChatGPT a Good Recommender? A Preliminary Study.CoRRabs/2304.10149 (2023)

  15. [21]

    Jiawei Liu, Chuan Shi, Cheng Yang, Zhiyuan Lu, and Philip S. Yu. 2022. A survey on heterogeneous information network based recommender systems: Concepts, methods, applications and resources.AI Open3 (2022), 40–57

  16. [22]

    Ruochen Liu, Hao Chen, Yuanchen Bei, Qijie Shen, Fangwei Zhong, Senzhang Wang, and Jianxin Wang. 2024. Fine Tuning Out-of-Vocabulary Item Recommen- dation with User Sequence Imagination. InNeurIPS

  17. [23]

    Pasquale Lops, Marco de Gemmis, and Giovanni Semeraro. 2011. Content-based Recommender Systems: State of the Art and Trends. InRecommender Systems Handbook. Springer, 73–105

  18. [24]

    Jianxin Ma, Chang Zhou, Peng Cui, Hongxia Yang, and Wenwu Zhu. 2019. Learn- ing Disentangled Representations for Recommendation. InNeurIPS. 5712–5723

  19. [25]

    Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-Domain Recommendation: An Embedding and Mapping Approach. InIJCAI. ijcai.org, 2464–2470

  20. [26]

    Maxim Naumov, Dheevatsa Mudigere, Hao-Jun Michael Shi, Jianyu Huang, Narayanan Sundaraman, Jongsoo Park, Xiaodong Wang, Udit Gupta, Carole-Jean Wu, Alisson G. Azzolini, Dmytro Dzhulgakov, Andrey Mallevich, Ilia Cherni- avskii, Yinghai Lu, Raghuraman Krishnamoorthi, Ansha Yu, V...

  21. [27]

    Gustavo Penha, Edoardo D’Amico, Marco De Nadai, Enrico Palumbo, Alexandre Tamborrino, Ali Vardasbi, Max Lefarov, Shawn Lin, Timothy Christopher Heath, Francesco Fabbri, and Hugues Bouchard. 2025. Semantic IDs for Joint Generative Conference acronym ’XX, June 03–05, 2018, Woods...

  22. [29]

    Aghiles Salah and Hady W. Lauw. 2018. A Bayesian Latent Variable Model of User Preferences with Item Context. InIJCAI. ijcai.org, 2667–2674

  23. [30]

    Aghiles Salah, Thanh-Binh Tran, and Hady W. Lauw. 2021. Towards Source- Aligned Variational Models for Cross-Domain Recommendation. InRecSys. ACM, 176–186

  24. [31]

    Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Diet- mar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Dennison

    D. Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Diet- mar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Dennison. 2015. Hidden Technical Debt in Machine Learning Systems. InNIPS. 2503–2511

  25. [32]

    Chuan Shi, Yitong Li, Jiawei Zhang, Yizhou Sun, and Philip S. Yu. 2017. A Survey of Heterogeneous Information Network Analysis.IEEE Trans. Knowl. Data Eng. 29, 1 (2017), 17–37

  26. [33]

    Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Enyun Yu. 2025. GenSAR: Unifying Balanced Search and Recommendation with Gener- ative Retrieval. InRecSys. ACM, 124–134

  27. [34]

    Alex Shtoff, Michael Viderman, Naama Haramaty-Krasne, Oren Somekh, Ariel Raviv, and Tularam Ban. 2024. Low Rank Field-Weighted Factorization Machines for Low Latency Item Recommendation. InRecSys. ACM, 238–246

  28. [35]

    Cohen, and Donald Metzler

    Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Prakash Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer Memory as a Differentiable Search Index. In NeurIPS

  29. [36]

    Quoc-Tuan Truong, Aghiles Salah, and Hady W. Lauw. 2021. Multi-Modal Rec- ommender Systems: Hands-On Exploration. InRecSys. ACM, 834–837

  30. [37]

    Moshe Unger, Alexander Tuzhilin, and Amit Livne. 2020. Context-Aware Recom- mendations Based on Deep Learning Frameworks.ACM Trans. Manag. Inf. Syst. 11, 2 (2020), 8:1–8:15

  31. [38]

    Chong Wang and David M. Blei. 2011. Collaborative topic modeling for recom- mending scientific articles. InKDD. ACM, 448–456

  32. [39]

    Hao Wang, Naiyan Wang, and Dit-Yan Yeung. 2015. Collaborative Deep Learning for Recommender Systems. InKDD. ACM, 1235–1244

  33. [40]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & Cross Network for Ad Click Predictions. InADKDD@KDD. ACM, 12:1–12:7

  34. [41]

    Wenjie Wang, Xinyu Lin, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2023. Generative Recommendation: Towards Next-generation Recommender Paradigm. CoRRabs/2304.03516 (2023)

  35. [42]

    Xiaolei Wang, Xinyu Tang, Xin Zhao, Jingyuan Wang, and Ji-Rong Wen. 2023. Rethinking the Evaluation for Conversational Recommendation in the Era of Large Language Models. InEMNLP. Association for Computational Linguistics, 10052–10065

  36. [43]

    Yancheng Wang, Ziyan Jiang, Zheng Chen, Fan Yang, Yingxue Zhou, Eunah Cho, Xing Fan, Yanbin Lu, Xiaojiang Huang, and Yingzhen Yang. 2024. RecMind: Large Language Model Powered Agent For Recommendation. InNAACL-HLT (Findings) (Findings of ACL). Association for Computational Lin...

  37. [44]

    Yuhao Wang, Junwei Pan, Xinhang Li, Maolin Wang, Yuan Wang, Yue Liu, Dapeng Liu, Jie Jiang, and Xiangyu Zhao. 2025. Empowering Large Language Model for Sequential Recommendation via Multimodal Embeddings and Semantic IDs. In CIKM. ACM, 3209–3219

  38. [45]

    Yi Xu, Moyu Zhang, Chenxuan Li, Zhihao Liao, Haibo Xing, Hao Deng, Jinxin Hu, Yu Zhang, Xiaoyi Zeng, and Jing Zhang. 2026. MMQ: Multimodal Mixture- of-Quantization Tokenization for Semantic ID Generation and User Behavioral Adaptation. InWSDM. ACM, 788–797

  39. [47]

    Feng Yuan, Lina Yao, and Boualem Benatallah. 2019. DARec: Deep Domain Adaptation for Cross-Domain Recommendation via Transferring Rating Patterns. InIJCAI. ijcai.org, 4227–4233

  40. [48]

    Tianzi Zang, Yanmin Zhu, Haobing Liu, Ruohan Zhang, and Jiadi Yu. 2023. A Survey on Cross-domain Recommendation: Taxonomies, Methods, and Future Directions.ACM Trans. Inf. Syst.41, 2 (2023), 42:1–42:39

  41. [49]

    Weizhi Zhang, Yuanchen Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, Feiran Huang, Sheng Zhou, Jiajun Bu, Allen Lin, James Caverlee, Fakhri Karray, Irwin King, and Philip S. Yu

  42. [50]

    Jujia Zhao, Zihan Wang, Shuaiqun Pan, Suzan Verberne, and Zhaochun Ren. 2026. Unifying Search and Recommendation in LLMs via Gradient Multi-Subspace Tuning.CoRRabs/2601.09496 (2026)

  43. [51]

    Lei Zheng, Vahid Noroozi, and Philip S. Yu. 2017. Joint Deep Modeling of Users and Items Using Reviews for Recommendation. InWSDM. ACM, 425–434

  44. [52]

    Guorui Zhou, Xiaoqiang Zhu, Chengru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. InKDD. ACM, 1059–1068

  45. [53]

    Jing Zhu, Mingxuan Ju, Yozen Liu, Danai Koutra, Neil Shah, and Tong Zhao. 2025. Beyond Unimodal Boundaries: Generative Recommendation with Multimodal Semantics.CoRRabs/2503.23333 (2025)

  46. [54]

    Lixi Zhu, Xiaowen Huang, and Jitao Sang. 2025. A LLM-based Controllable, Scalable, Human-Involved User Simulator Framework for Conversational Rec- ommender Systems. InWWW. ACM, 4653–4661

  47. [2016]

    InDLRS@RecSys

    Wide & Deep Learning for Recommender Systems. InDLRS@RecSys. ACM, 7–10

  48. [2025]

    Cold-Start Recommendation towards the Era of Large Language Models (LLMs): A Comprehensive Survey and Roadmap.CoRRabs/2501.01945 (2025)

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.