Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Hierarchical Tree Search-based User Lifelong Behavior Modeling on Large Language Model

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 lifelong user behavior can be modeled by an LLM through chunked extraction, hierarchical tree search, and temporal fusion, producing a representation that improves any recommendation model.

desk verdict Solid industrial paper with a genuinely new rating-guided tree search over chunked LLM interest summaries, but the rating labels are noisy enough that the core HTS claim needs stronger evidence. read the letter →

arxiv 2505.19505 v1 pith:JDDF2U43 submitted 2025-05-26 cs.IR cs.AI

Yu Xia , Rui Zhong , Hao Gu , Wei Yang , Chi Lu , Peng Jiang , Kun Gai This is my paper
classification cs.IRcs.AI
keywords largelanguagemodelsrecommendersystemslifelonguserbehaviormodelingchunkedextractionhierarchicaltreesearchprocessratingmodeltemporalinterestfusionclick-throughrateprediction
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's HiT-LBM framework claims that lifelong user behavior can be modeled by a large language model without hitting context-length limits: the user's history is split into time-ordered chunks, and interests are extracted chunk by chunk in a cascading prompt. To stop errors from accumulating across chunks, a hierarchical tree search generates several candidate interest summaries for each chunk and uses two trained rating models to keep the one that adds the most information. A temporal fusion module then turns the chosen interest path into a dense vector that can be injected into any recommendation model as side features. If this works as claimed, recommender systems could combine an LLM's semantic understanding with a user's full history rather than a short recent window. The paper reports consistent ranking-accuracy (AUC) and loss gains on eight CTR backbones, wins over other LLM-enhanced baselines, and online revenue and conversion lifts.

What carries the argument

The load-bearing object is the hierarchical interest tree coupled with two process rating models. Each tree layer corresponds to one behavior chunk; each node is a candidate LLM interest summary generated by expanding the parent node with Best-of-N sampling. SRM and PRM are trained on labels built from AUC comparisons: a candidate is labeled continuous if using it to predict the next chunk's items beats using the previous K interests, and effective if using it beats using the raw current chunk. The final score $S_{\text{final}}=\alpha S_c+(1-\alpha)S_e$ ranks sibling nodes, and the greedy maximum selects one node per layer, making the path the sequence of interests with locally maximal information gain. Temporal fusion then converts the path into a vector using search-score weighting, positional encoding, masked self-attention, and item-guided cross-attention.

What would settle it

Re-run HTS with different evaluation sample sizes (e.g., N=6 vs N=20) and different numbers of previous interests K, and check whether the selected optimal interest paths stay stable: if most selected nodes flip, the rating labels are noise and the search is not measuring information gain. A sharper check is to replace SRM and PRM with random selection among the Best-of-N children and see whether downstream AUC loses the entire reported HTS gain.

Watch

Extended reading notes

Core claim

The central claim is that a user's lifelong behavior is best represented as an ordered list of LLM-generated interest descriptions, one per time chunk, selected by search rather than by a single sequential pass. For each chunk the LLM samples N candidate interests from the parent node; a Sequence Rating Model scores continuity with the previous K interests, a Point Rating Model scores how well the candidate alone captures the chunk's value, and the weighted score $S_{\text{final}}=\alpha S_c+(1-\alpha)S_e$ chooses the node that expands next. The resulting path is called the optimal interest path, and its nodes supply the final user interests. Those interests are embedded, weighted by their search scores, position-encoded, fused with masked self-attention, and combined with item knowledge through cross-attention to produce a lifelong user representation. The paper argues this representation can be inserted into any ID-based CTR or CVR model and presents experiments showing it outperforms traditional sequential models and LLM-enhanced behavior modeling baselines.

Load-bearing premise

Everything rests on the assumption that the LLM's own 'Yes' probabilities, compared as AUC on a small sample of the next chunk's items, give reliable labels for whether a candidate interest summary is continuous and effective.

Editorial extensions

If this is right

  • Any ID-based click-through-rate or conversion-rate model can accept the fused interest vector and item-knowledge vector as extra features; the paper reports relative AUC gains from 3.60% to 4.35% on MovieLens-1M across eight backbones.
  • Larger interest-generating LLMs improve downstream recommendation accuracy, so the quality of the extracted interest text itself is a performance lever.
  • The chunked cascading design removes the LLM context-length barrier, allowing modeling of histories much longer than one prompt can hold.
  • Online, the framework lifted revenue by 3.5% and conversion rate by 2.3% in a 14-day A/B test, with larger relative gains on long-tail users.

Reading between the lines

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

  • The same chunk-search-fuse recipe could transfer to other long-context LLM tasks, such as long-document summarization or dialogue memory, wherever a cascading prompt would otherwise accumulate errors.
  • The AUC-based information-gain labels are a general way to measure whether a compressed summary adds signal; they could be reused as a reward signal to fine-tune the interest generator itself rather than only selecting among its samples.
  • The greedy one-best expansion is a design choice; a beam search over the tree would be a direct extension that might recover from an early wrong selection, at higher inference cost.
  • Practical deployment may require distilling the per-chunk LLM calls into a smaller student model, since the pipeline as described invokes the LLM for expansion and rating many times per user.
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. This paper proposes HiT-LBM, a model-agnostic framework for lifelong user behavior modeling with LLMs. User histories are split into chunks (CUBE); for each chunk, candidate interest summaries are generated by LLM best-of-N sampling and scored by two learned process rating models, SRM (continuity) and PRM (effectiveness), trained on binary labels derived from AUC comparisons of LLM "Yes" probabilities on a 12-item sample from the next chunk (Section 3.3.1, Eqs. 3-5); a greedy tree search (Algorithm 1) selects the highest-scoring node at each layer; finally, temporal-aware interest fusion (TIF) weights and encodes the selected interests and item knowledge into dense side information injected into arbitrary CTR/CVR backbones (Section 3.4). Experiments on MovieLens-1M and Amazon Book across eight backbones, comparisons with DIN/SIM/KAR/TRSR/LIBER, ablations, and an online A/B test report consistent AUC/LogLoss improvements.

Significance. If the reported results are robust, HiT-LBM would be a useful contribution to LLM-enhanced recommendation: it offers a concrete way to sidestep context-window limits via chunking, attempts to control cascading degradation via a search procedure, and demonstrates transfer to multiple backbone architectures and an online setting. The experimental breadth is a genuine strength: eight backbone models, two public datasets, systematic ablations across modules (CUBE/TIF/HTS), LLM scales, and interest encoders, plus a 14-day online A/B test. The framework is also clearly described, with explicit prompts and algorithm pseudocode. The main weakness is that the core novelty—HTS—is supervised by labels that are themselves noisy comparisons of LLM AUCs computed on 6 positive and 6 negative items, and the paper provides no uncertainty or robustness analysis for these labels, so the central mechanism is not yet established.

major comments (4)
  1. [Section 3.3.1, Eqs. (3)-(5), and Algorithm 1] The SRM/PRM training labels are derived from AUC comparisons on only N=6 positive and N=6 negative items sampled from B_{t+1}. With 36 positive-negative pairs, the estimated AUC has a coarse support and a standard error on the order of 0.15-0.17 under standard approximations; the paper provides no bootstrap confidence intervals, no repeated sampling analysis, and no held-out accuracy of SRM/PRM themselves. The inequality comparisons in Eq. (5) that define the continuity and effectiveness labels are therefore likely to be within noise for a substantial fraction of chunks. This matters because Algorithm 1 greedily commits to the child node with the highest S_final at each layer, so a single erroneous label can eliminate the true interest path. The claim that HTS "ensures information gain for each behavior chunk" (Section 3.3) is thus not supported, and the reported gains could in principle come from CUBE/TIF plus item knowledge rather than from HTS. I would like to see (i) the distribution of AUC differences underlying Eq. (5), (ii) SRM/PRM classification accuracy on held-out chunks, (iii) an ablation that replaces HTS with random child selection or with N=1 expansion, and (iv) a label-noise analysis such as bootstrapping over sampled items. The fact that the labels are produced by the same LLM family that generates the candidate interests further reinforces the need for such validation.
  2. [Table 2 and Section 4.2.1] The asterisks in Table 2 claim "statistically significant improvement (measured by t-test with p-value<0.001)", but the paper reports only a single AUC and LogLoss value per model and dataset, with no standard deviations, no number of runs or seeds, and no paired-test details. With one value per condition, a t-test is not defined. This is load-bearing because the average improvements of 4.12% on MovieLens-1M and 1.35% on Amazon Book are highlighted as the main offline evidence. Please report mean and variance over multiple runs/random seeds, specify the exact test used (paired or unpaired, number of observations), or remove the significance claim. The same issue applies to the comparisons in Table 3 and Table 5, where no significance information is given.
  3. [Section 4.1.1 and Eq. (3)] For Amazon Book, samples with rating 5 are treated as positive. Since Eq. (3) requires randomly selecting N=6 positive items from the next behavior chunk B_{t+1}, many chunks of length 50 will contain fewer than six rating-5 items. The paper does not state how such chunks are handled. If they are filtered out, the SRM/PRM training set is biased toward chunks with many positive behaviors; if they are handled by substituting lower-rated items or by downsampling, the labels are corrupted. This directly affects the training of the process rating models and therefore the tree search. Please specify the exact sampling/filtering procedure and report how many chunks were excluded or adjusted on each dataset.
  4. [Section 3.4.2, Eq. (12)] The causal mask in the Masked Self-Attention block is specified incorrectly. The text defines Mask as an upper-triangular matrix filled with 1 above the diagonal and 0 elsewhere, and Eq. (12) adds this Mask inside the softmax. Adding a constant of 1 does not prevent attention to future positions; causal masking requires adding -infinity (or an equivalent multiplicative mask) to the upper triangle. As written, the temporal fusion cannot ensure that "the representation at each position is only aware of the interest representations preceding it." Please correct the equation and clarify the actual implementation, since the temporal ordering is the distinguishing property of TIF.
minor comments (5)
  1. [Throughout] There are several typos: "Bahavior" in Section 3.2, "Temporal-Ware" should be "Temporal-Aware" in headings, and Algorithm 1 as well as Section 4.1.4 refer to "SPM" where the text elsewhere uses "SRM".
  2. [Section 3.3.1 and Algorithm 1] The symbol N is overloaded: in Eq. (3) it is the number of positive/negative samples per chunk (N=6), while in Algorithm 1 it is the number of expansions per node (N=10). Please use distinct symbols to avoid confusion.
  3. [Section 3.3.1] The text says SRM and PRM are trained on datasets D_conf and D_eff, but D_cont was defined earlier; this inconsistency should be fixed.
  4. [Section 4.3.1 and Table 4] The online A/B test reports revenue and CVR lifts without confidence intervals, p-values, or details on user assignment and overlap between buckets. Please provide at least the number of users per bucket and the statistical significance of the differences.
  5. [Section 4.1.4] The statement that batch size and learning rate were determined by grid search is not accompanied by the search ranges or the selected values; this limits reproducibility. The paper also does not mention whether code or extracted interest representations will be released.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; HTS rating models are trained on proxy labels, but final recommendation evaluation is on a separate test split and does not reduce to those labels.

full rationale

The paper's derivation chain is self-contained. CUBE generates per-chunk interest summaries via LLM prompting (Eq. 2). HTS trains SRM/PRM as binary classifiers on labels constructed from LLM AUC comparisons over a 12-item sample from the next chunk (Eqs. 3-5), then uses these ratings to greedily select nodes (Algorithm 1). This is a standard train-a-proxy-then-optimize-proxy design, not a definitional reduction: the rating labels are not the same as the final CTR/CVR ground truth, and the final evaluation uses a separate timestamp-based test split (Section 4.1.1) with AUC/LogLoss on held-out interactions. The only self-referential aspect is that the same fixed LLM (Qwen2.5-7B) both generates candidate interests and supplies the 'Yes' probabilities that define the rating labels, which is a methodological reliability concern (small-sample AUC, potential self-preference bias) rather than a circularity: the SRM/PRM are fitted to those labels, but the central claim of improved recommendation performance is verified externally on test data and online A/B traffic. No load-bearing self-citation, uniqueness theorem, or renamed-ansatz step was found. Concerns about label noise with N=6 per class and greedy one-best expansion are correctness-risk issues, not circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The framework rests on seven hand-chosen hyperparameters and five domain assumptions, the most fragile being the AUC-comparison label heuristic and the greedy one-best-path search. The ledger shows that HiT-LBM is a system engineering contribution whose gains come from a specific configuration of known neural and LLM components.

free parameters (7)
  • chunk_length_L = 50
    Sets how many behaviors each LLM prompt covers; chosen in parameter config without sensitivity analysis (Sec 4.1.4).
  • eval_sample_size_N = 6
    Number of positive and negative items sampled from the next chunk to compute AUC labels; small sample makes labels noisy (Sec 3.3.1, Eq. 3).
  • previous_interest_count_K = 1
    Number of past interests used as context for the Sequence Rating Model; fixed without analysis (Sec 4.1.4).
  • fusion_weight_alpha = 0.5
    Weight balancing continuity and effectiveness scores in Eq. (7); no sensitivity study reported.
  • num_expansions_per_node = 10
    Best-of-N sampling count during tree expansion (Algorithm 1, Sec 4.1.4).
  • interest_encoder = BGE
    Chosen after comparing BERT, Longformer, and BGE; final choice is BGE (Sec 4.4.4).
  • interest_generator_LLM = Qwen2.5-7B
    LLM used to generate interest summaries; scale is varied in Sec 4.4.3, but the model family is fixed.
assumptions (5)
  • domain assumption The LLM's P('Yes') token probability is a valid preference score for computing AUC on the evaluation set.
    Sec 3.3.1 Eq. (4) uses these probabilities to build Y_seq, Y_point, Y_hist and the AUC labels that train SRM/PRM; if the probabilities are uncalibrated or prompt-sensitive, the labels are biased.
  • domain assumption A text summary of a behavior chunk preserves enough information about user interests for downstream recommendation.
    CUBE compresses each chunk into a single interest summary and the pipeline discards the raw behaviors at fusion time; information loss inside a chunk is not modeled.
  • ad hoc to paper The AUC-comparison heuristic correctly identifies information gain and interest validity.
    Sec 3.3.1 labels an interest positive for continuity only if AUC_point > AUC_seq and positive for effectiveness only if AUC_point > AUC_hist; this operational definition is assumed, not proven, to capture interest quality.
  • ad hoc to paper Greedy expansion of a single best node per layer finds an optimal or sufficiently good interest path.
    Algorithm 1 keeps only the argmax child at each layer, so errors cannot be corrected later, yet the paper calls the result the optimal search path (Eq. 8).
  • domain assumption Rating models trained on training-set chunks generalize to test users and new interest texts.
    SRM/PRM are trained on one sample per chunk of the training set and then applied to all users; no user-level split is used to validate this transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Tree Search-based User Lifelong Behavior Modeling on Large Language Model." pith.science (2026). https://pith.science/paper/JDDF2U43

@misc{pith2026250519505,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Tree Search-based User Lifelong Behavior Modeling on Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JDDF2U43}},
  note         = {Machine review of arXiv:2505.19505}
}
read the original abstract

Large Language Models (LLMs) have garnered significant attention in Recommendation Systems (RS) due to their extensive world knowledge and robust reasoning capabilities. However, a critical challenge lies in enabling LLMs to effectively comprehend and extract insights from massive user behaviors. Current approaches that directly leverage LLMs for user interest learning face limitations in handling long sequential behaviors, effectively extracting interest, and applying interest in practical scenarios. To address these issues, we propose a Hierarchical Tree Search-based User Lifelong Behavior Modeling framework (HiT-LBM). HiT-LBM integrates Chunked User Behavior Extraction (CUBE) and Hierarchical Tree Search for Interest (HTS) to capture diverse interests and interest evolution of user. CUBE divides user lifelong behaviors into multiple chunks and learns the interest and interest evolution within each chunk in a cascading manner. HTS generates candidate interests through hierarchical expansion and searches for the optimal interest with process rating model to ensure information gain for each behavior chunk. Additionally, we design Temporal-Ware Interest Fusion (TIF) to integrate interests from multiple behavior chunks, constructing a comprehensive representation of user lifelong interests. The representation can be embedded into any recommendation model to enhance performance. Extensive experiments demonstrate the effectiveness of our approach, showing that it surpasses state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2505.19505 by the authors.

Figure 1
Figure 1. Overview of the HiT-LBM framework. enhance the recommendation performance of LLMs by incorpo￾rating ID feature during the fine-tuning process. To address the challenge of long-sequence recommendation, Rella [16] proposes a retrieval-based approach, where the 𝐾 most relevant items to the target item are retrieved and used as input for LLM fine-tuning. In the LLM enhance Rec paradigm, the capabilities of LLMs are harn… view at source ↗
Figure 2
Figure 2. Example prompts for User Interest Learning and Item Knowledge Extraction, denoted as [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The prompts we construct for the recommendation task incorporate the user’s previous interest (𝑝 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The detailed workflow of how the Process Rating [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: The impact of different LLM scale. 4.4.3 effect of different LLM scale (RQ6). To validate the impact of different scales of LLMs on modeling interests, we implement HiT-LBM using various scales of the Qwen2.5 series LLMs. We integrate the interests generated by LLMs of…
Figure 7
Figure 7. Figure 7: The impact of different interest encoder. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

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. Full citation record

  1. RecoWorld: Building Simulated Environments for Agentic Recommender Systems

    cs.IR 2025-09 conditional novelty 5.0 of 10

    A design proposal, not a tested system: a dual-view simulation loop in which an LLM-simulated user issues reflective instructions when about to disengage, and an instruction-following recommender adapts to maximize si...

  2. R4ec: A Reasoning, Reflection, and Refinement Framework for Recommendation Systems

    cs.IR 2025-07 conditional novelty 5.0 of 10

    R4ec trains a small reflection model to critique and refine LLM-generated user and item knowledge, which then improves downstream recommendation accuracy.

Reference graph

Works this paper leans on

43 extracted references · 19 canonical work pages · cited by 2 Pith papers

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Amos Azaria, Avinatan Hassidim, Sarit Kraus, Adi Eshkol, Ofer Weintraub, and Irit Netanely. 2013. Movie Recommender System for Profit Maximization. In Proceedings of the 7th ACM Conference on Recommender Systems . 121–128

    work page 2013

  3. [3]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He

  4. [4]

    Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long- document transformer. arXiv preprint arXiv:2004.05150 (2020)

    arXiv 2020

  5. [5]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216 (2024)

    arXiv 2024

  6. [6]

    Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

    arXiv 2018

  7. [7]

    Rahul Dey and Fathi M Salem. 2017. Gate-Variants of Gated Recurrent Unit (GRU) Neural Networks. In 2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS). IEEE, 1597–1600

    work page 2017

  8. [8]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine Based Neural Network for CTR Prediction. arXiv preprint arXiv:1703.04247 (2017)

    arXiv 2017

Show all 43 references
  1. [9]

    Ruining He and Julian McAuley. 2016. VBPR: Visual Bayesian Personalized Rank- ing from Implicit Feedback. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 30

    2016

  2. [10]

    Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. 2022. Cascaded Diffusion Models for High Fidelity Image Generation. Journal of Machine Learning Research 23, 47 (2022), 1–33

    2022

  3. [11]

    Yutong Hu, Quzhe Huang, Mingxu Tao, Chen Zhang, and Yansong Feng. 2024. Can Perplexity Reflect Large Language Model’s Ability in Long Text Understand- ing? arXiv preprint arXiv:2405.06105 (2024)

    2024 arXiv

  4. [12]

    Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for Click-Through Rate Prediction. In Proceedings of the 13th ACM Conference on Recommender Systems . 169–177

    2019

  5. [13]

    Jiaqi Li, Mengmeng Wang, Zilong Zheng, and Muhan Zhang. 2023. LooGLE: Can Long-Context Language Models Understand Long Contexts? arXiv preprint arXiv:2311.04939 (2023)

    2023 arXiv

  6. [14]

    Zekun Li, Zeyu Cui, Shu Wu, Xiaoyu Zhang, and Liang Wang. 2019. Fi-gnn: Modeling feature interactions via graph neural networks for ctr prediction. In Proceedings of the 28th ACM international conference on information and knowledge management. 539–548

    2019

  7. [15]

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mi...

    2018

  8. [16]

    Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. Rella: Retrieval-Enhanced Large Language Models for Lifelong Sequential Behavior Comprehension in Recommendation. In Proceedings of the ACM on Web Conferen...

    2024

  9. [17]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 Technical Report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  10. [18]

    Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee

  11. [19]

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-Based User Interest Modeling with Lifelong Sequential Behavior Data for Click-Through Rate Prediction. In Proceedings of the 29th ACM International Conference on Informat...

    2020

  12. [20]

    Mathieu Ravaut, Aixin Sun, Nancy Chen, and Shafiq Joty. 2024. On Context Utilization in Summarization with Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2764–2781

    2024

  13. [22]

    Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. AutoInt: Automatic Feature Interaction Learning via Self- Attentive Neural Networks. In Proceedings of the 28th ACM International Confer- ence on Information and Knowledge Manageme...

    2019

  14. [23]

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

  15. [24]

    Zhu Sun, Hongyang Liu, Xinghua Qu, Kaidong Feng, Yan Wang, and Yew Soon Ong. 2024. Large Language Models for Intent-Driven Session Recommendations. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, W...

    2024

  16. [25]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  17. [26]

    Aaron Van den Oord, Sander Dieleman, and Benjamin Schrauwen. 2013. Deep Content-Based Music Recommendation. Advances in Neural Information Process- ing Systems 26 (2013)

    2013

  18. [27]

    A Vaswani. 2017. Attention is All You Need. Advances in Neural Information Processing Systems (2017)

    2017

  19. [29]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & Cross Network for Ad Click Predictions. In Proceedings of the ADKDD’17 . 1–7

    2017

  20. [30]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-Scale Learning to Rank Systems. In Proceedings of the Web Conference 2021. 1785–1797

    2021

  21. [31]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Llmrec: Large Language Models with Graph Augmentation for Recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining . 806–815

    2024

  22. [32]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models. In Proceedings of the 18th ACM Conference on Recommender ...

    2024

  23. [33]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al . 2024. Qwen2.5 Technical Report. arXiv preprint arXiv:2412.15115 (2024)

    2024 arXiv

  24. [34]

    Shenghao Yang, Weizhi Ma, Peijie Sun, Qingyao Ai, Yiqun Liu, Mingchen Cai, and Min Zhang. 2024. Sequential Recommendation with Latent Relations based on Large Language Model. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Informati...

    2024

  25. [35]

    Yang Zhang, Keqin Bao, Ming Yan, Wenjie Wang, Fuli Feng, and Xiangnan He

  26. [36]

    Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He

  27. [37]

    Zhi Zheng, Wenshuo Chao, Zhaopeng Qiu, Hengshu Zhu, and Hui Xiong. 2024. Harnessing large language models for text-rich sequential recommendation. In Proceedings of the ACM on Web Conference 2024 . 3207–3216

    2024

  28. [38]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep Interest Evolution Network for Click-Through Rate Prediction. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 33. 5941–5948

    2019

  29. [39]

    arXiv preprint arXiv:2406.03210 (2024)

    Text-Like Encoding of Collaborative Information in Large Language Models for Recommendation. arXiv preprint arXiv:2406.03210 (2024)

    2024 arXiv

  30. [40]

    Chenxu Zhu, Shigang Quan, Bo Chen, Jianghao Lin, Xiaoling Cai, Hong Zhu, Xiangyang Li, Yunjia Xi, Weinan Zhang, and Ruiming Tang. 2024. LIBER: Life- long User Behavior Modeling Based on Large Language Models. arXiv preprint arXiv:2411.14713 (2024)

    2024 arXiv

  31. [41]

    arXiv preprint arXiv:2310.19488 (2023)

    Collm: Integrating Collaborative Embeddings into Large Language Models for Recommendation. arXiv preprint arXiv:2310.19488 (2023)

    2023 arXiv

  32. [44]

    Guorui Zhou, Xiaoqiang Zhu, Chenru 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. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining...

    2018

  33. [2019]

    In Proceedings of the 28th ACM International Conference on Information and Knowledge Management

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management . 1441–1450

  34. [2023]

    In Proceedings of the 17th ACM Conference on Recommender Systems

    Tallrec: An Effective and Efficient Tuning Framework to Align Large Lan- guage Model with Recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

  35. [2024]

    arXiv preprint arXiv:2403.03900 (2024)

    Mamba4rec: Towards Efficient Sequential Recommendation with Selective State Space Models. arXiv preprint arXiv:2403.03900 (2024)

    2024 arXiv

Pith tools

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