Pith. sign in

REVIEW 6 major objections 5 minor 2 cited by

BehaveGPT: A Foundation Model for Large-scale User Behavior Modeling

T0 review · 6 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 proposes BehaveGPT, a transformer foundation model for user behavior, claiming over 10% gains in macro and weighted recall and a first scaling-law fit for behavior data.

desk verdict Useful empirical study undermined by an undefined DRO loss and a self-contradictory scaling-law argument; worth peer review but not acceptance as is. read the letter →

arxiv 2505.17631 v1 pith:MA3AMJP4 submitted 2025-05-23 cs.IR cs.AI

Jiahui Gong , Jingtao Ding , Fanjin Meng , Chen Yang , Hong Chen , Zuojian Wang , Haisheng Lu , Yong Li This is my paper
classification cs.IRcs.AI
keywords userbehaviormodelingfoundationmodelsnextpredictionlong-taillearningdistributionallyrobustoptimizationscalinglawsequentialrecommendationcross-domainadaptation
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

BehaveGPT is an attempt to build the analogue of a language foundation model for user behavior: a transformer trained from scratch on raw behavior logs, with embeddings for weekday, time slot, location, and event, and a pretraining objective based on distributionally robust optimization (DRO) that deliberately gives tail behaviors a larger uncertainty set. The paper claims this model beats state-of-the-art recommendation, long-tail, and LLM-based baselines on three real-world datasets—mobile phone logs, app check-ins, and mobility trajectories—with more than 10% improvement in macro and weighted recall on next-behavior prediction. It also reports the first scaling-law measurement for the user behavior domain, fitting the language-model loss form $L(N,D)=C_N N^{-\alpha}+C_D D^{-\beta}+L_0$ on the Honor dataset with $\alpha=0.51$ and $\beta=0.23$. If these results hold, behavior prediction would no longer need language pretraining, and the fitted data-to-model ratio near 5 would suggest that behavior foundation models can be trained far more cheaply than large language models.

What carries the argument

The load-bearing mechanism is the DRO-based pretraining paradigm. Distributionally robust optimization replaces the standard cross-entropy objective with a worst-case expected loss over an uncertainty set of behavior distributions; classes with few samples receive larger uncertainty sets, so the model is forced to stay accurate on rare behaviors. The architecture around it is a transformer with four embedding streams (weekday, time slot, location, event), FlashAttention for efficiency, and an MLP prediction layer. The scaling-law analysis is carried by the functional form $L(N,D)=C_N N^{-\alpha}+C_D D^{-\beta}+L_0$, fit to eighteen checkpoints spanning 0.4M to 24M parameters and data sizes up to 600M logs.

What would settle it

A direct test is to fit the same scaling form on a second behavior dataset with models larger than 24M parameters: if the fitted exponents shift substantially or the loss stops falling as data grows, the claimed scaling law and the data-to-model ratio of about 5 do not hold. A second test is to train the same transformer with plain cross-entropy and no DRO objective; if macro and weighted recall still match or beat the baselines, the pretraining paradigm is not the cause of the gains.

Watch

Extended reading notes

Core claim

BehaveGPT is trained on hundreds of millions of user behavior logs and uses a transformer with separate embeddings for weekday, time slot, location, and event; the prediction head is an MLP, and FlashAttention keeps long sequences tractable. The pretraining objective is DRO-based: instead of plain cross-entropy, the model minimizes the worst-case expected loss over an uncertainty set that is larger for rare tail behaviors, which the paper argues is why it can model head and tail behaviors without sacrificing either. Across next-behavior prediction, the paper reports that BehaveGPT beats ten baselines on the Honor, Mobile, and Tencent datasets, with the largest per-metric gains around 19% and double-digit gains on most weighted and ranking metrics. The paper also claims strong few-shot new-behavior prediction, better long-term generation measured by distributional and diversity metrics, and cross-domain transfer from Honor to the other two datasets. The scaling-law section fits the language-model curve to models from 0.4M to 24M parameters and concludes that user-behavior data favors a data-to-model ratio near 5, an order of magnitude below LLM practice.

Load-bearing premise

The scaling-law conclusions assume that user-behavior loss follows the same smooth power-law curve in model and data size that language models do, and that the fitted exponents stay meaningful in the regime the paper itself marks as 'Data is insufficient!' because larger models overfit.

Editorial extensions

If this is right

  • Next-behavior prediction on app usage, mobility, and intent datasets would improve by more than 10% in macro and weighted recall over the strongest baselines.
  • A model pretrained on one behavior domain can be adapted to another with a small fine-tuning budget, reusing transformer and prediction-layer parameters.
  • New behaviors can be learned from a small fine-tuning set, roughly 0.25% of the pretraining data, after adding embeddings and output weights for the new behavior.
  • Long-term generation produces sequences closer to the real behavior distribution and more diverse, according to KS, Wasserstein distance, JSD, BLEU, and Distinct-2.
  • Scaling data and model size together lowers loss, with the fitted exponents suggesting that data scaling should receive slightly more emphasis than model scaling.

Reading between the lines

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

  • If the data-to-model ratio near 5 transfers beyond the Honor dataset, user-behavior foundation models could be trained with far less data than language models need; the paper's own 'Data is insufficient!' label warns that the fit sits partly in an overfitting regime.
  • The DRO objective is a generic answer to long-tail sequential data, so the same uncertainty-set pretraining could be tested on other imbalanced behavior streams, such as financial transactions or health logs, where BehaveGPT is not evaluated.
  • Because the fitted exponents have $\alpha > \beta$, the fit itself says that adding data should help more than adding parameters at the margin; a clean check would compare equal-compute checkpoints that scale data versus parameters.
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

6 major / 5 minor

Summary. The paper proposes BehaveGPT, a transformer-based model for user behavior prediction trained on mobile-phone, app-usage, and trajectory datasets, with a distributionally robust optimization (DRO) pretraining paradigm intended to balance head and tail behaviors. The authors report experiments for next-behavior prediction, new-behavior prediction, long-term generation, and cross-domain adaptation, claiming more than 10% improvement over ten baselines, and they present a Chinchilla-style scaling-law fit on the Honor dataset with exponents alpha=0.51 and beta=0.23, leading to a recommended data-to-model ratio around 5.

Significance. If the claims were substantiated, BehaveGPT would be a notable early step toward a behavior-domain foundation model: the benchmark coverage is broad (four downstream tasks, three real-world datasets, ten baselines), and the scaling-law analysis addresses a genuinely underexplored question for behavior data. The paper also deserves credit for explicitly attempting to handle long-tail behaviors and for comparing against a wide range of modern baselines. However, the central DRO formulation is never actually specified, the scaling-law fit is made in a regime that the paper's own text identifies as overfitting, and the data counts are mutually contradictory; these issues currently prevent the results from being reproducible and the headline claims from being verifiable.

major comments (6)
  1. [Section 3.2, Eq. (4)] The DRO-based pretraining paradigm is the paper's central methodological novelty, but Eq. (4) is not a well-posed training objective. The set definition uses p_b both as a distribution and as an atom probability, the constraint 'p_b = epsilon p_b(b) <= p_b^train(b) for all b' is not formally meaningful as written, and the displayed expression contains only a supremum over p_b, with no minimization over the model parameters and no computable loss or algorithm. Consequently, the 'DRO loss' referenced in Section 4.6 cannot be implemented or evaluated from the manuscript as written. Please provide the full min-max objective, the definition of epsilon, and the optimization procedure, or the main methodological contribution is unsupported.
  2. [Abstract and Table 2] The abstract states that BehaveGPT achieves 'more than a 10% improvement in macro and weighted recall', but Table 2 reports several improvements below 10%, including 9.59% for Honor Rec_m, 7.84% for Tencent Rec_m, 6.05% for Tencent Rec_w, and 8.51% for Tencent N@3. The claim is therefore factually inaccurate as written and should be restricted to the specific metrics and datasets for which it actually holds.
  3. [Section 4.1.1, Table 1, and Section 4.7] The reported data sizes are mutually inconsistent. The abstract and Section 4.7 say 'more than 600 million behavior logs' and 'data sizes ranging up to 600 million', Section 4.1.1 says the Honor dataset contains 'over 2 million logs', but Table 1 reports 205,605,167 logs for Honor and the three datasets sum to about 210 million. These differences change the experimental scale by orders of magnitude and make the scaling-law fit ungrounded; the paper must state the exact per-run dataset sizes and reconcile the totals.
  4. [Section 4.7, Eq. (5), and Fig. 7(b)] The Chinchilla-form fit is applied in a regime that the paper itself says is not power-law: the text states that beyond a threshold model size the model 'begins to overfit the training set... causing a sharp decline in performance', and Fig. 7(b) is labeled 'Data is insufficient!'. Since Eq. (5) is strictly decreasing in both N and D, it cannot represent the non-monotonic loss described, so the fitted exponents alpha=0.51 and beta=0.23 are not supported. The authors should either restrict the fit to the power-law regime with per-point N and D values, residuals, and confidence intervals, or remove the scaling-law claim.
  5. [Section 4.7, Fig. 8] The derivation of the recommended data-to-model ratio of about 5 is also not validated. The description of the scaling runs ('from 1N to 10N, where N denotes the model size') is ambiguous, no table of losses per (N, D) point is provided, and the recommendation is derived from the same fitted curve rather than checked on held-out data or an independent predictor. As a result, the comparison to Llama2's ratio of 70-100 and MiniCPM's ratio of 192 cannot be assessed.
  6. [Tables 2-5] The central performance comparisons are reported without standard deviations, confidence intervals, or significance tests. Given that some claimed margins are as small as 6.05%, the reader cannot distinguish signal from noise. Please report multiple-seed results or otherwise justify the precision of the reported values.
minor comments (5)
  1. [Eq. (5)] The scaling-law equation should be written with proper exponents, e.g., L(N,D) = C_N N^{-alpha} + C_D D^{-beta} + L0, rather than the malformed notation N*(-alpha) and D*(-beta).
  2. [Figures 2, 5, 6, 7, and 8] The figures in the submitted PDF contain unparsed Unicode/PDF glyph sequences, so the axis labels and legends are unreadable; they must be regenerated before the paper can be properly evaluated.
  3. [Section 2.1 and Appendix B] There are several typos that should be fixed, including 'Data Ananlysis' in Section 2.1, 'Distrinct-2' in Appendix B, the broken citation 'cite pituning' in Section 1, and the stray character in the phrase 'user behavior„' in Section 5.1.
  4. [Section 4.2] The text says the model 'gets superior results on both datasets' although Table 2 reports results on three datasets; this wording should be corrected.
  5. [Section 3.1] The claim that Flash Attention operates 2-4 times faster and reduces memory usage by 5-20 times is stated without a benchmark or detailed citation; if these numbers are not essential to the architecture description, they should be removed or supported.

Circularity Check

1 steps flagged · score 4.0 of 10

The scaling-law recommendation is a fitted consequence of Eq. (5); the benchmark claim is grounded in external baselines.

  1. fitted input called prediction [Section 4.7, Eq. (5), Figures 7 and 8]
    "Then we fit the lossed with model size N and data size D following the curvefit function [14]: L(N,D)=C_N N^(−α)+ C_D D^(−β)+ L0. (5) In our experiments, the fitted relationship between loss and N,D is shown in Figure 8. Specifically, we have α = 0.51, β = 0.23. Since α is slightly larger than β, this result shows that as the computation scale, we should slightly emphasize more on data scaling than model scaling. Based on the points in Figure 8(a), our data-to-model ratio is approximately 5, while Llama2's ratio ranges from 70 to 100."

    The data-to-model ratio and the advice to emphasize data scaling are not measured independently; they are algebraic consequences of fitting Eq. (5) to the authors' own loss curves. The exponents α and β are free parameters of the assumed Chinchilla form, so the recommendation to emphasize data scaling (α > β) and the ratio ≈ 5 are determined by the fitted constants by construction. No held-out loss point or external dataset validates the fitted curve; the paper's own Figure 7(b) labels part of the regime 'Data is insufficient!' and states that the model 'begins to overfit the training set... causing a sharp decline in performance,' which the strictly decreasing power law cannot represent.

full rationale

The paper's primary performance claim (Section 4.2 and Tables 2-5) is evaluated against ten external baselines on three datasets, so it does not reduce to the model's own assumptions. The DRO pretraining paradigm is validated by ablation-style comparisons (Figure 6), which is an empirical claim rather than a circular derivation. The only step that approaches circularity is Section 4.7: the 'scaling law in the user behavior domain' is produced by fitting the Chinchilla form Eq. (5) to the authors' own loss curves, and the headline outputs (α=0.51, β=0.23, data-to-model ratio ≈ 5) are direct algebraic functions of those fitted parameters. The conclusion that 'we should slightly emphasize more on data scaling than model scaling' is likewise a restatement of α > β. This is a fitted-input-called-prediction pattern confined to the scaling-law contribution. The paper's own Figure 7(b) even labels a portion of the fitted regime 'Data is insufficient!' and describes overfitting with a 'sharp decline in performance,' which the strictly decreasing power law cannot represent; that inconsistency is a correctness risk rather than additional circularity. Self-citations such as [11] for metrics and baselines are present but not load-bearing: the metrics and baselines are standard and independently defined. Overall, the central 'outperforms SOTA' claim is self-contained against external benchmarks, so the circularity score is moderate, not high.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on fitted scaling-law parameters whose values are only partially reported, on an ill-defined DRO uncertainty set, and on the assumption that a Chinchilla-style scaling law holds for behavior data. No new physical or conceptual entities are introduced.

free parameters (4)
  • alpha (scaling exponent for model size) = 0.51
    Fitted via Eq. (5) to Honor scaling-law data; no confidence intervals are reported.
  • beta (scaling exponent for data size) = 0.23
    Fitted via Eq. (5) to Honor scaling-law data; no confidence intervals are reported.
  • C_N, C_D, L0 (scaling law coefficients) = not reported
    Fitted constants in Eq. (5) are not reported, so the fit is not reproducible.
  • epsilon (DRO uncertainty set size) = not reported
    Introduced in Eq. (4) to define the uncertainty set; no value or tuning procedure is given.
assumptions (4)
  • domain assumption Chinchilla scaling-law form L(N,D)=C_N N^(-alpha)+C_D D^(-beta)+L0 applies to user behavior data at the tested model sizes.
    The paper fits this form from [14] to Honor data; if the form does not hold, the scaling conclusions fail. See Section 4.7.
  • domain assumption The behavior label set B is fixed and the event-to-behavior mapping is known and consistent.
    Problem definition in Section 2.2; next-behavior prediction assumes these labels are well-defined.
  • ad hoc to paper The DRO uncertainty set defined in Eq. (4) is well-posed, with a non-empty set and finite supremum.
    Eq. (4) is the only formal statement of the DRO paradigm and as written it is ill-typed; the authors proceed as if a valid DRO loss exists.
  • domain assumption The three datasets share enough structure that pretraining on Honor transfers to Mobile and Tencent.
    Cross-domain experiments in Section 4.5 rely on this assumption to claim generalizable behavior representations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BehaveGPT: A Foundation Model for Large-scale User Behavior Modeling." pith.science (2026). https://pith.science/paper/MA3AMJP4

@misc{pith2026250517631,
  author       = {Pith},
  title        = {Pith review of: BehaveGPT: A Foundation Model for Large-scale User Behavior Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MA3AMJP4}},
  note         = {Machine review of arXiv:2505.17631}
}
read the original abstract

In recent years, foundational models have revolutionized the fields of language and vision, demonstrating remarkable abilities in understanding and generating complex data; however, similar advances in user behavior modeling have been limited, largely due to the complexity of behavioral data and the challenges involved in capturing intricate temporal and contextual relationships in user activities. To address this, we propose BehaveGPT, a foundational model designed specifically for large-scale user behavior prediction. Leveraging transformer-based architecture and a novel pretraining paradigm, BehaveGPT is trained on vast user behavior datasets, allowing it to learn complex behavior patterns and support a range of downstream tasks, including next behavior prediction, long-term generation, and cross-domain adaptation. Our approach introduces the DRO-based pretraining paradigm tailored for user behavior data, which improves model generalization and transferability by equitably modeling both head and tail behaviors. Extensive experiments on real-world datasets demonstrate that BehaveGPT outperforms state-of-the-art baselines, achieving more than a 10% improvement in macro and weighted recall, showcasing its ability to effectively capture and predict user behavior. Furthermore, we measure the scaling law in the user behavior domain for the first time on the Honor dataset, providing insights into how model performance scales with increased data and parameter sizes.

Figures

Figures reproduced from arXiv: 2505.17631 by the authors.

Figure 1
Figure 1. The paradigm of foundation model. [29, 36], as it allows them to optimize their actions to achieve specific goals. These agents must be capable of perceiving their environment, learning from behaviors, and predicting future behaviors to adapt and make intelligent decisions. User behavior prediction is a critical first step toward achieving behavioral intelligence, as it enables models to understand and predict futur… view at source ↗
Figure 2
Figure 2. Distribution of behaviors in different datasets. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The BehaveGPT framework. • Prediction Layer We use a Multilayer Perception (MLP) to be the prediction layer, which can be formed as, m = W2 (𝜎(W1H𝑡 + 𝜀1)) + 𝜀2, (3) where W, 𝜀 are the trainable weight matrix and the bias matrix. The output of the MLP m is the predicted user behavior distribution. Manuscript submitted to ACM [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The illustration of DRO-based pretrain paradigm. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of cross-domain method. 4.6 Effectiveness of the DRO-based Pretraining paradigm The DRO-based pretraining paradigm, enhanced with DRO loss, ensures balanced performance across head and tail classes while improving generalization in the user behavior domain. …
Figure 6
Figure 6. Figure 6: Effectiveness of the DRO-based pretraining paradigm. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Scalability of BehaveGPT on the Honor datasets. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: The fitted scaling law plotted along the data size. [PITH_FULL_IMAGE:figures/full_fig_p017_8.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. Clinical Audit Logs as Multi-Axial Traces of Care Delivery

    cs.CY 2026-07 conditional novelty 5.0 of 10

    EHR audit log entries should be read as multi-axial events — clinician, patient, team, and process orderings at once — motivating shared foundation-model representations and a cross-axis benchmark.

  2. A Foundation Model for Multimodal Event Sequences in Financial Applications

    cs.LG 2026-07 conditional novelty 4.5 of 10

    Early-fusion next-event pretraining on multimodal bank event sequences yields reusable user embeddings that, combined with engineered features, improve multi-task financial predictions and production NPV.

Reference graph

Works this paper leans on

62 extracted references · 36 canonical work pages · cited by 2 Pith papers

  1. [1]

    Albert Bandura. 1977. Social learning theory. Englewood Cliffs (1977)

    work page 1977

  2. [2]

    Tom B Brown. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165 (2020)

    arXiv 2020

  3. [3]

    Lei Chen, Chen Gao, Xiaoyi Du, Hengliang Luo, Depeng Jin, Yong Li, and Meng Wang. 2024. Enhancing ID-based Recommendation with Large Language Models. ACM Transactions on Information Systems (2024)

    work page 2024

  4. [4]

    Wayne W Daniel. 1990. Kolmogorov–Smirnov one-sample test. Applied nonparametric statistics 2 (1990)

    work page 1990

  5. [5]

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems 35 (2022), 16344–16359

    work page 2022

  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]

    Jingtao Ding, Fuli Feng, Xiangnan He, Guanghui Yu, Yong Li, and Depeng Jin. 2018. An improved sampler for bayesian personalized ranking by leveraging view data. In Companion proceedings of the the web conference 2018 . 13–14

    work page 2018

  8. [8]

    Jingtao Ding, Yuhan Quan, Quanming Yao, Yong Li, and Depeng Jin. 2020. Simplify and robustify negative sampling for implicit collaborative filtering. Advances in Neural Information Processing Systems 33 (2020), 1094–1105

    work page 2020

Show all 62 references
  1. [9]

    Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  2. [10]

    Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. 2024. Large language models empowered agent-based modeling and simulation: A survey and perspectives. Humanities and Social Sciences Communications 11, 1 (2024), 1–24

    2024

  3. [11]

    Jiahui Gong, Jingtao Ding, Fanjin Meng, Guilong Chen, Hong Chen, Shen Zhao, Haisheng Lu, and Yong Li. 2024. A Population-to-individual Tuning Framework for Adapting Pretrained LM to On-device User Intent Prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge...

    2024

  4. [12]

    Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. 2022. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933 (2022)

    2022 arXiv

  5. [13]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. 2022. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 16000–16009

    2022

  6. [14]

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. 2022. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022)

    2022 arXiv

  7. [15]

    Jakob Hohwy. 2013. The predictive mind. OUP Oxford

    2013

  8. [16]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  9. [17]

    Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. 2024. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395 (2024)

    2024 arXiv

  10. [18]

    De Yong Jiang, Hong Zhang, Harish Kumar, Quadri Noorulhasan Naveed, Chandan Takhi, Vishal Jagota, and Rituraj Jain. 2022. Automatic control model of power information system Access based on artificial intelligence technology.Mathematical Problems in Engineering 2022, 1 (2022), 5677634

    2022

  11. [19]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

    2018

  12. [20]

    Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2015. A diversity-promoting objective function for neural conversation models. arXiv preprint arXiv:1510.03055 (2015)

    2015 arXiv

  13. [21]

    Jiayu Li, Peijie Sun, Zhefan Wang, Weizhi Ma, Yangkun Li, Min Zhang, Zhoutian Feng, and Daiyue Xue. 2023. Intent-aware Ranking Ensemble for Personalized Recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Ret...

    2023

  14. [22]

    Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian McAuley. 2023. Text is all you need: Learning language representations for sequential recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1258–1267

    2023

  15. [23]

    Fengming Lin, Xiaolei Fang, and Zheming Gao. 2022. Distributionally robust optimization: A review on theory and applications. Numerical Algebra, Control and Optimization 12, 1 (2022), 159–212

    2022

  16. [24]

    Fan Liu, Zhiyong Cheng, Huilin Chen, Yinwei Wei, Liqiang Nie, and Mohan Kankanhalli. 2022. Privacy-preserving synthetic data generation for recommendation systems. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieva...

    2022

  17. [25]

    Qidong Liu, Xian Wu, Xiangyu Zhao, Yejing Wang, Zijian Zhang, Feng Tian, and Yefeng Zheng. 2024. Large Language Models Enhanced Sequential Recommendation for Long-tail User and Item. arXiv preprint arXiv:2405.20646 (2024)

    2024 arXiv

  18. [26]

    Siyi Liu and Yujia Zheng. 2020. Long-tail session-based recommendation. In Proceedings of the 14th ACM conference on recommender systems . 509–514

    2020

  19. [27]

    Yuli Liu, Christian Walder, Lexing Xie, and Yiqun Liu. 2024. Probabilistic Attention for Sequential Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1956–1967

    2024

  20. [28]

    Jing Long, Guanhua Ye, Tong Chen, Yang Wang, Meng Wang, and Hongzhi Yin. 2024. Diffusion-based cloud-edge-device collaborative learning for next POI recommendations. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2026–2036

    2024

  21. [29]

    Xiaojuan Ma, Emily Yang, and Pascale Fung. 2019. Exploring perceived emotional intelligence of personality-driven virtual agents in handling user challenges. In The World Wide Web Conference. 1222–1233

    2019

  22. [30]

    María Luisa Menéndez, JA Pardo, L Pardo, and MC Pardo. 1997. The jensen-shannon divergence. Journal of the Franklin Institute 334, 2 (1997), 307–318

    1997

  23. [31]

    Homogenizing Effect of Large Language Model (LLM) on Creative Diversity: An Empirical Comparison

    Kibum Moon, Adam Green, and Kostadin Kushlev. Homogenizing Effect of Large Language Model (LLM) on Creative Diversity: An Empirical Comparison. (????)

  24. [32]

    Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. 2023. Levels of AGI: Operationalizing Progress on the Path to AGI. arXiv preprint arXiv:2311.02462 (2023)

    2023

  25. [33]

    Devaki Nadkarni. 2016. SUPERFORECASTING: The Art & Science of Forecasting. Anvesha 9, 1 (2016), 61

    2016

  26. [34]

    Victor M Panaretos and Yoav Zemel. 2019. Statistical aspects of Wasserstein distances. Annual review of statistics and its application 6, 1 (2019), 405–431

    2019

  27. [35]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics . 311–318

    2002

  28. [36]

    Pat Pataranutaporn, Valdemar Danry, Joanne Leong, Parinya Punpongsanon, Dan Novy, Pattie Maes, and Misha Sra. 2021. AI-generated characters for supporting personalized learning and well-being. Nature Machine Intelligence 3, 12 (2021), 1013–1022

    2021

  29. [37]

    Yingtao Peng, Chen Gao, Yu Zhang, Tangpeng Dan, Xiaoyi Du, Hengliang Luo, Yong Li, and Xiaofeng Meng. 2025. Denoising alignment with large language model for recommendation. ACM Transactions on Information Systems 43, 2 (2025), 1–35

    2025

  30. [38]

    Chaoyi Pu, Zhiang Wu, Hui Chen, Kai Xu, and Jie Cao. 2018. A Sequential Recommendation for Mobile Apps: What Will User Click Next App?. In 2018 IEEE International Conference on Web Services (ICWS) . 243–248. https://doi.org/10.1109/ICWS.2018.00038

    2018

  31. [39]

    Yuhan Quan, Jingtao Ding, Chen Gao, Nian Li, Lingling Yi, Depeng Jin, and Yong Li. 2023. Alleviating video-length effect for micro-video recommendation. ACM Transactions on Information Systems 42, 2 (2023), 1–24

    2023

  32. [40]

    Alec Radford. 2018. Improving language understanding by generative pre-training. (2018)

    2018

  33. [41]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9

    2019

  34. [42]

    Jiang-Xin Shi, Tong Wei, Zhi Zhou, Jie-Jing Shao, Xin-Yan Han, and Yu-Feng Li. 2024. Long-Tail Learning with Foundation Model: Heavy Fine-Tuning Hurts. In Forty-first International Conference on Machine Learning

    2024

  35. [43]

    Luc Steels. 2018. Building agents out of autonomous behavior systems. In The artificial life route to artificial intelligence . Routledge, 83–121

    2018

  36. [44]

    Karen Tatarian, Rebecca Stower, Damien Rudaz, Marine Chamoux, Arvid Kappas, and Mohamed Chetouani. 2022. How does modality matter? investigating the synthesis and effects of multi-modal robot behavior on social intelligence. International Journal of Social Robotics 14, 4 (2022...

    2022

  37. [45]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  38. [46]

    Chen Wang, Ziwei Fan, Liangwei Yang, Mingdai Yang, Xiaolong Liu, Zhiwei Liu, and Philip Yu. 2024. Pre-Training with Transferable Attention for Addressing Market Shifts in Cross-Market Sequential Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Disc...

    2024

  39. [47]

    Jianling Wang, Kaize Ding, Ziwei Zhu, and James Caverlee. 2021. Session-based Recommendation with Hypergraph Attention Networks. ArXiv abs/2112.14266 (2021). https://api.semanticscholar.org/CorpusID:232073844

    2021 arXiv

  40. [48]

    Peng Wang, Jiang Xu, Chunyi Liu, Hao Feng, Zang Li, and Jieping Ye. 2020. Masked-field Pre-training for User Intent Prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM ’20) . Association for Computing Machinery, New Y...

    2020

  41. [49]

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. 2022. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682 (2022)

    2022 arXiv

  42. [50]

    Bruce Croft, and Haiqing Chen

    Liu Yang, Minghui Qiu, Chen Qu, Cen Chen, Jiafeng Guo, Yongfeng Zhang, W. Bruce Croft, and Haiqing Chen. 2020. IART: Intent-aware Response Ranking with Transformers in Information-seeking Conversation Systems. In Proceedings of The Web Conference 2020 (WWW ’20) . Association f...

    2020

  43. [51]

    Fuxun Yu, Zhuwei Qin, Chenchen Liu, Di Wang, and Xiang Chen. 2020. REIN the RobuTS: Robust DNN-based image recognition in autonomous driving systems. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 40, 6 (2020), 1258–1271. Manuscript submitted to ...

    2020

  44. [52]

    Lantao Yu, Weinan Zhang, Jun Wang, and Yong Yu. 2017. Seqgan: Sequence generative adversarial nets with policy gradient. In Proceedings of the AAAI conference on artificial intelligence , Vol. 31

    2017

  45. [53]

    Yuan Yuan, Jingtao Ding, Jie Feng, Depeng Jin, and Yong Li. 2024. Unist: A prompt-empowered universal model for urban spatio-temporal prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4095–4106

    2024

  46. [54]

    Yuan Yuan, Jingtao Ding, Depeng Jin, and Yong Li. 2025. Learning the complexity of urban mobility with deep generative network. PNAS nexus 4, 5 (2025), pgaf081

    2025

  47. [55]

    Yuan Yuan, Jingtao Ding, Huandong Wang, and Depeng Jin. 2024. Generating daily activities with need dynamics. ACM Transactions on Intelligent Systems and Technology 15, 2 (2024), 1–28

    2024

  48. [56]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi

  49. [57]

    Buyun Zhang, Liang Luo, Yuxin Chen, Jade Nie, Xi Liu, Shen Li, Yanli Zhao, Yuchen Hao, Yantao Yao, Ellie Dingqiao Wen, Jongsoo Park, Maxim Naumov, and Wenlin Chen. 2024. Wukong: Towards a Scaling Law for Large-Scale Recommendation. In Proceedings of the 41st International Conf...

    2024

  50. [58]

    Yu Zhang, Shutong Qiao, Jiaqi Zhang, Tzu-Heng Lin, Chen Gao, and Yong Li. 2025. A Survey of Large Language Model Empowered Agents for Recommendation and Search: Towards Next-Generation Information Retrieval. arXiv preprint arXiv:2503.05659 (2025)

    2025 arXiv

  51. [59]

    Yin Zhang, Ruoxi Wang, Derek Zhiyuan Cheng, Tiansheng Yao, Xinyang Yi, Lichan Hong, James Caverlee, and Ed H Chi. 2023. Empowering Long-tail Item Recommendation through Cross Decoupling Network (CDN). In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and ...

    2023

  52. [60]

    Yujia Zheng, Siyi Liu, Zekun Li, and Shu Wu. 2021. Cold-start Sequential Recommendation via Meta Learner. Proceedings of the AAAI Conference on Artificial Intelligence 35, 5 (May 2021), 4706–4713. https://doi.org/10.1609/aaai.v35i5.16601

    2021 doi

  53. [61]

    Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al . 2023. One fits all: Power general time series analysis by pretrained lm. Advances in neural information processing systems 36 (2023), 43322–43355. Manuscript submitted to ACM

    2023

  54. [2024]

    Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research) , Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, ...

Pith tools

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