Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

A return-conditioned policy lifted notification sessions 0.72% over offline value-based RL

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A Decision Transformer with quantile-regression return prompts improved notification decisions at LinkedIn, boosting sessions by 0.72% over the deployed CQL baseline in a live A/B test.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A credible production DT-vs-CQL engineering story whose headline lift is fragile; the real value is the quantile-regression RTG prompting and the operational evidence. the 3 major comments →

arxiv 2509.02458 v1 pith:UXIIHU4P submitted 2025-09-02 cs.LG cs.AI

Generative Sequential Notification Optimization via Multi-Objective Decision Transformers

Borja Ocejo , Ruofan Wang , Ke Liu , Rohit K. Patra , Haotian Shen , David Liu , Yiwen Yuan , Gokulraj Mohanasundaram
show 2 more authors
Fedor Borisyuk Prakruthi Prabhakar
This is my paper
classification cs.LG cs.AI
keywords Decision Transformeroffline reinforcement learningreturn-conditioned supervised learningnotification optimizationmulti-objective rewardsquantile regressionrecommender systemsproduction deployment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 argues that a Decision Transformer—an offline reinforcement learning method that treats policy learning as supervised sequence modeling conditioned on desired future returns—is a more practical replacement for value-based approaches like conservative Q-learning in large-scale notification decision-making. It establishes this with a multi-reward formulation where the return-to-go prompt is learned, per state, via quantile regression, so the model can be queried for a target outcome level at serving time. In offline evaluation the best policy reached 96.7% next-action accuracy, and in a live A/B test it increased user sessions by 0.72% relative to the value-based baseline while reducing notification volume and leaving click-through rate statistically unchanged. The authors also contribute a production architecture using a circular buffer for sequence persistence that supports transformer inference at 150K queries per second. The larger claim is that sequence-modeling RL can be stable, scalable, and tunable enough for recommender control at industrial scale.

Core claim

The central claim is that notification send/drop decisions can be framed as return-conditioned supervised learning rather than value estimation, and that this shift fixes the practical failure modes of temporal-difference-based offline reinforcement learning. Instead of learning a Q-function, the model learns the conditional quantile distribution of the return-to-go given the history and current state, and then predicts actions autoregressively conditioned on a target quantile. Because the prompt is derived from the data distribution rather than hand-set, inference can directly request a desired level of each reward objective; this gives transparent prompt tuning over outcomes such as click-

What carries the argument

The central object is a Decision Transformer with a multi-objective reward vector and quantile-regression return-to-go (RTG) prompting. The transformer consumes trajectories of the form (state, RTG, action, reward), but rather than using the classic manual RTG update, it learns the RTG distribution from historical states. A pinball-loss output head predicts quantiles of each reward dimension's return, and at inference the target quantile is linearly interpolated from the learned quantile levels. This lets the policy be prompted at arbitrary quantile settings, enabling direct control over reward-seeking behavior. Supporting this is a circular-buffer sequence store that persists user-specific

Load-bearing premise

The policy is trained on predicted notification-value scores as rewards; if those predicted scores drift from what actually drives user sessions, the learned policy optimizes the wrong return and the measured session gain may not persist.

What would settle it

Run the same decision-transformer policy against the value-based baseline for eight consecutive weeks without retraining the reward models. If the +0.72% session advantage appears in some weeks and reverses in others in a way that tracks reward-model prediction error, the transfer claim fails. A second check: measure calibration of the learned RTG quantiles against realized future rewards; large calibration error means the policy is prompted with a phantom signal.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, notification policies can be trained with the stability of supervised learning while still optimizing long-horizon, multi-objective returns.
  • Learned return-to-go prompts maintain session activity while reducing notification volume, a direct efficiency gain over the deployed value-based baseline.
  • Longer transformer context improves next-action prediction accuracy, with a context length of 4 chosen as the production trade-off between quality and serving cost.
  • Prompt tuning gives fine-grained, interpretable control over policy objectives: raising the target quantile for click-through return increases online click-through rate in A/B tests.
  • The circular-buffer serving design demonstrates that transformer-based RL policies can run at 100K-150K queries per second in a real recommender system.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same quantile-prompted Decision Transformer recipe could transfer to other continuous recommender controls—such as feed pacing, email frequency, or ad delivery—where rewards are also predicted scores rather than ground-truth utility, as long as the predicted-reward bias remains bounded.
  • Quantile prompting may function as a dynamic exploration knob: different users or cohorts could be assigned different target quantiles, personalizing the aggressiveness of notification delivery without retraining.
  • The paper's reported instability asymmetry suggests that value-based offline RL may be a poor default for high-dimensional industrial state spaces, and that sequence-conditioned policies deserve serious consideration there.
  • A testable extension is to measure whether the learned RTG quantiles are calibrated against realized returns; if they are, the prompt could be set per user to directly target session growth.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper describes a production Decision Transformer (DT) framework for notification send/drop decisions at LinkedIn. The authors frame notification delivery as a multi-objective sequential decision problem, preserve the reward vector rather than combining it with fixed weights, and learn conditional quantiles of the return-to-go (RTG) via quantile regression to serve as prompts at inference time. They also present a circular-buffer-based serving architecture for near-real-time inference at 100–150K QPS. Offline and online experiments compare the DT policy to a multi-objective CQL baseline. The headline result is a +0.72% increase in user sessions with a -1.68% reduction in notification volume and no statistically significant CTR change (Table 1). The authors further report ablations on context length and prompt tuning, and argue that DT is more stable, scalable, and interpretable than CQL in this setting.

Significance. If the reported online result is trustworthy, this is a valuable industrial validation of return-conditioned sequence modeling for large-scale recommender-system control. The quantile-regression RTG head is a sensible adaptation of DT to non-episodic, multi-reward settings, and the production engineering contributions—circular-buffer sequence persistence, per-tenant model serving, and scale—are concrete and substantial. The paper is also honest about a key limitation, namely the bias in predicted reward signals (Section 3). However, the central claim rests on a single online A/B point estimate that is reported without confidence intervals, multiple-testing considerations, or a controlled comparison that isolates the DT architecture from the expanded feature set. These issues must be resolved before the headline number can be accepted.

major comments (3)
  1. [§6.3, Table 1, Figure 4] The headline +0.72% session gain is reported as the outcome of the 'best-performing DT policy' selected after prompt tuning over 'a set of quantiles centered around 0.7' (§6.3). The paper does not state how many prompt configurations were tested, how selection was made, or what the uncertainty is: no confidence intervals, standard errors, or p-values are given for any row of Table 1, and only the CTR row is marked NSS. Figure 4 shows several independent online A/B tests with no error bars. If the policy was chosen as the best among many tested online configurations, the point estimate is subject to winner's-curse inflation; under the null, the maximum of several noisy lifts is positive. To support the abstract's precise number, the authors should report the full set of tested configurations, the pre-specified selection rule, per-test sample sizes and confidence intervals, and a multiple-
  2. [§6.3] The comparison labeled 'Overall DT vs. CQL' is not an architecture-only comparison. The paper states, 'We use a richer state representation compared to the CQL baseline,' and that incorporating these signals into a CQL agent with an enlarged network did not yield stable or effective policies. Thus the DT policy differs from the CQL baseline in both the policy class and the feature set. The +0.72% session gain could be driven partly or entirely by the richer features rather than by the DT architecture. A clean comparison would require a CQL policy trained on the same state representation, or at minimum an explicit statement that a DT policy with the same feature set as CQL also outperforms CQL. The 'Basic DT' row is not that control either, because it still uses the DT class. Please provide a matched-feature CQL result or soften the attribution of the gain to the DT framework.
  3. [Table 1 (§6.3)] The incremental rows are not internally consistent as a decomposition of the overall row. The Volume column satisfies -0.31 - 1.18 - 0.19 = -1.68, but the Sessions column can sum to +0.72 only if the 'Learned Prompts' row has a point estimate of exactly 0 (since 0.51 + 0.21 = 0.72). If the sessions estimate is exactly 0, then the text's claim that learned prompts 'allowed the policy to maintain session activity with fewer notifications' is not an improvement, and the broader conclusion that the observed improvements follow from learned prompts is unsupported. If the estimate is not exactly 0, the additive interpretation is unjustified without a factorial design. Please clarify whether these rows represent staged, pre-registered increments or post hoc selected subsets, and report the actual point estimate for the 'Learned Prompts' sessions row.
minor comments (4)
  1. [§4.2.1 / §4.2.2] The symbol λ is used both as the RTG loss weight in L_DT and as the interpolation weight λ_α in the quantile approximation. This is confusing; please use distinct notation. Also, d_model and n_r are used without definitions.
  2. [§6.1 / §6.2] The values of the look-ahead horizon H, discount factor γ, RTG loss weight λ, and the exact reward components are not reported for the online experiments. Since the RTG definition depends directly on H and γ, these values are needed for reproducibility. The offline section also reports only action accuracy and pinball loss, with no CQL offline baseline; please clarify what the offline results are meant to show.
  3. [§6.3 / Table 1] The notation 'NSS' is used without a formal definition. Please specify the significance threshold, the test used, and whether confidence intervals are available for the session and volume rows as well. If the rows are not all significant, the narrative should not imply that every component contributed meaningfully.
  4. [Abstract / Title] The word 'Generative' in the title and abstract is used without an explicit definition. The model is a return-conditioned sequence model that predicts actions and RTGs; it may be worth clarifying what 'generative' means here, or removing the term to avoid overclaiming.

Circularity Check

0 steps flagged

No significant circularity: central claim is a measured online A/B outcome, not a derivation from fitted inputs; self-citations are not load-bearing.

full rationale

The paper's central claim (+0.72% sessions for DT vs. CQL) is an empirical result from an online A/B test, not an analytic derivation. The training rewards include predicted measures of notification value, but the evaluation metric (sessions) is measured directly on real user behavior, so there is no self-definitional identity between the two. The quantile-regression RTG head is trained and evaluated on held-out validation data via pinball loss, and the prompt-tuning procedure selects alpha based on online A/B outcomes; while this raises statistical selection concerns, it does not make the reported lift equivalent to the inputs by construction. The self-citation [27] for the CQL baseline is used as a deployed comparison system, and the comparison is directly tested in the experiment rather than assumed from the citation. No equation in the paper reduces one target to another by definition; therefore no circular step is present.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central empirical claim depends on several unreported hyperparameters and on the validity of the reward and state modeling choices. No new entities are introduced; the framework reuses DT, RTG, and quantile regression.

free parameters (7)
  • Look-ahead horizon H
    Used to compute return-to-go R_t = sum_{l=0}^H gamma^l r_{t+l}; value not reported, but directly controls the temporal objective.
  • Discount factor gamma
    Used in RTG computation; not reported.
  • RTG loss weight lambda
    Weight combining action and RTG losses in L_DT = L_action + lambda * L_RTG; not reported.
  • Prompt quantile alpha (per reward) = tuned around 0.7; case study uses 0.5, 0.75, 0.95
    The return-to-go prompt is set by selecting a target quantile; this is a free parameter tuned via online A/B experiments (Sections 6.3, 6.5), effectively fit to the measured session/CTR outcomes.
  • Number of quantile levels M = 3 in offline experiments; production set unspecified
    Offline experiments use quantiles 0.25, 0.5, 0.75; controls the granularity of RTG distribution modeling.
  • Context length T = 4 in production
    Selected based on accuracy/cost tradeoff (Table 2).
  • Epsilon exploration rate for data collection
    Data is collected by an epsilon-greedy CQL baseline over one week; epsilon value not reported, affecting offline data coverage.
axioms (5)
  • domain assumption Predicted reward signals from existing deployed utility models are valid proxies for true user value and their cumulative sum is the right objective to optimize.
    Section 3 Rewards paragraph states these predicted measures are used as rewards; if biased, the trained policy optimizes the wrong signal.
  • domain assumption A finite horizon H with discount gamma captures the long-term value of notification decisions; returns beyond H are negligible or redundant.
    Section 4.1: trajectories of length T+H are sampled and RTG is computed over H steps; the policy is optimized against this truncated return.
  • domain assumption The engineered state features are sufficient statistics for the notification decision problem (adequately captures sequential user dynamics).
    Section 3 States paragraph describes engineered features; the DT is trained assuming the sequence up to T steps is sufficient.
  • standard math Minimizing pinball loss yields valid conditional quantiles of returns in the training distribution.
    Section 4.2.1 uses pinball loss; standard consistency of quantile regression is assumed without proof.
  • domain assumption The one-week epsilon-greedy data collection provides sufficient coverage of the state-action space for training.
    Section 6.1 describes data collected from CQL baseline on 2% of users for one week; no coverage analysis is provided.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Sequential Notification Optimization via Multi-Objective Decision Transformers." pith.science (2026). https://pith.science/paper/UXIIHU4P

@misc{pith2026250902458,
  author       = {Pith},
  title        = {Pith review of: Generative Sequential Notification Optimization via Multi-Objective Decision Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UXIIHU4P}},
  note         = {Machine review of arXiv:2509.02458}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Notifications are an important communication channel for delivering timely and relevant information. Optimizing their delivery involves addressing complex sequential decision-making challenges under constraints such as message utility and user fatigue. Offline reinforcement learning (RL) methods, such as Conservative Q-Learning (CQL), have been applied to this problem but face practical challenges at scale, including instability, sensitivity to distribution shifts, limited reproducibility, and difficulties with explainability in high-dimensional recommendation settings. We present a Decision Transformer (DT) based framework that reframes policy learning as return-conditioned supervised learning, improving robustness, scalability, and modeling flexibility. Our contributions include a real-world comparison with CQL, a multi-reward design suitable for non-episodic tasks, a quantile regression approach to return-to-go conditioning, and a production-ready system with circular buffer-based sequence processing for near-real-time inference. Extensive offline and online experiments in a deployed notification system show that our approach improves notification utility and overall session activity while minimizing user fatigue. Compared to a multi-objective CQL-based agent, the DT-based approach achieved a +0.72% increase in sessions for notification decision-making at LinkedIn by making notification recommendation more relevant.

Figures

Figures reproduced from arXiv: 2509.02458 by Borja Ocejo, David Liu, Fedor Borisyuk, Gokulraj Mohanasundaram, Haotian Shen, Ke Liu, Prakruthi Prabhakar, Rohit K. Patra, Ruofan Wang, Yiwen Yuan.

Figure 1
Figure 1. Figure 1: Model Architecture: During training, DT learns RTGs at a fixed set of quantile levels [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Production architecture: Notification candidates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Predicted return-to-go for CTR at different target [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of CTR Prompt Tuning on Online CTR [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. A Self-Triggered Agentic Push Recommendation System

    cs.IR 2026-08 conditional novelty 6.0

    STEPS, a self-triggered agentic push system with planning, execution, and filtering agents, improved user active days by 0.2843% and reduced push permission disablement by 1.9089% in a Douyin A/B test.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages · cited by 1 Pith paper · 8 internal anchors

  1. [1]

    Anirudhan Badrinath, Yannis Flet-Berliac, Allen Nie, and Emma Brunskill. 2023. Waypoint transformer: Reinforcement learning via supervised learning with intermediate targets. Neurips (2023)

    work page 2023

  2. [2]

    Fedor Borisyuk, Qingquan Song, Mingzhou Zhou, Ganesh Parameswaran, Madhu Arun, Siva Popuri, Tugrul Bingol, Zhuotao Pei, Kuang-Hsuan Lee, Lu Zheng, Qizhan Shao, Ali Naqvi, Sen Zhou, and Aman Gupta. 2024. LiNR: Model Based Neural Retrieval on GPUs at LinkedIn. In CIKM

    work page 2024

  3. [3]

    David Brandfonbrener, Alberto Bietti, Jacob Buckman, Romain Laroche, and Joan Bruna. 2022. When does return-conditioned supervised learning work for offline reinforcement learning? Neurips (2022)

    work page 2022

  4. [4]

    Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. 2021. Decision Transformer: Reinforcement Learning via Sequence Modeling. arXiv:2106.01345 [cs.LG] https://arxiv.org/abs/2106.01345

    Pith/arXiv arXiv 2021

  5. [5]

    Minmin Chen, Can Xu, Vince Gatto, Devanshu Jain, Aviral Kumar, and Ed Chi

    work page

  6. [6]

    Xiaocong Chen, Lina Yao, Julian McAuley, Guanglin Zhou, and Xianzhi Wang

  7. [7]

    Joseph Clinton and Robert Lieck. 2024. Planning transformer: Long- horizon offline reinforcement learning with planning tokens. arXiv preprint arXiv:2409.09513 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. 2018. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA). IEEE, 4693–4700

    work page 2018

  9. [9]

    André Correia and Luís A Alexandre. 2023. Hierarchical decision transformer. In IROS

    work page 2023

  10. [10]

    Scott Emmons, Benjamin Eysenbach, Ilya Kostrikov, and Sergey Levine. 2021. Rvs: What is essential for offline rl via supervised learning? arXiv preprint arXiv:2112.10751 (2021)

    Pith/arXiv arXiv 2021

  11. [11]

    Jingtong Gao, Yewen Li, Shuai Mao, Peng Jiang, Nan Jiang, Yejing Wang, Qing- peng Cai, Fei Pan, Peng Jiang, Kun Gai, et al . 2025. Generative Auto-Bidding with Value-Guided Explorations. In SIGIR

    work page 2025

  12. [12]

    Yan Gao, Viral Gupta, Jinyun Yan, Changji Shi, Zhongen Tao, PJ Xiao, Curtis Wang, Shipeng Yu, Romer Rosales, Ajith Muralidharan, et al. 2018. Near real-time optimization of activity-based notifications. In KDD

    work page 2018

  13. [13]

    Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Ben- jamin Eysenbach, and Sergey Levine. 2019. Learning to reach goals via iterated supervised learning. arXiv preprint arXiv:1912.06088 (2019). Ocejo, Wang, Liu, et al

    Pith/arXiv arXiv 2019

  14. [14]

    Eugene Ie, Vihan Jain, Jing Wang, Sanmit Narvekar, Ritesh Agarwal, Rui Wu, Heng-Tze Cheng, Tushar Chandra, and Craig Boutilier. 2019. SlateQ: A Tractable Decomposition for Reinforcement Learning with Recommendation Sets.. In IJCAI, Vol. 19. 2592–2599

    work page 2019

  15. [15]

    Michael Janner, Qiyang Li, and Sergey Levine. 2021. Offline reinforcement learning as one big sequence modeling problem. Neurips (2021)

    work page 2021

  16. [16]

    Huxiao Ji, Haitao Yang, Linchuan Li, Shunyu Zhang, Cunyi Zhang, Xuanping Li, and Wenwu Ou. 2024. TIM: Temporal Interaction Model in Notification System. arXiv:2406.07067 https://arxiv.org/abs/2406.07067

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Aviral Kumar, Xue Bin Peng, and Sergey Levine. 2019. Reward-conditioned policies. arXiv preprint arXiv:1912.13465 (2019)

    Pith/arXiv arXiv 2019

  18. [18]

    Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. 2020. Con- servative Q-Learning for Offline Reinforcement Learning. arXiv:2006.04779 https://arxiv.org/abs/2006.04779

    Pith/arXiv arXiv 2020

  19. [19]

    Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. 2020. Conserva- tive Q-Learning for Offline Reinforcement Learning. In Neurips

    work page 2020

  20. [20]

    Kuang-Huei Lee, Ofir Nachum, Mengjiao Sherry Yang, Lisa Lee, Daniel Freeman, Sergio Guadarrama, Ian Fischer, Winnie Xu, Eric Jang, Henryk Michalewski, et al. 2022. Multi-game decision transformers. Neurips (2022)

    work page 2022

  21. [21]

    Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. 2020. Offline rein- forcement learning: Tutorial, review, and perspectives on open problems. arXiv preprint arXiv:2005.01643 (2020)

    Pith/arXiv arXiv 2020

  22. [22]

    LinkedIn Engineering. 2021. Air Traffic Controller: Member-first notifications at LinkedIn. https://www.linkedin.com/blog/engineering/messaging-notifications/ air-traffic-controller-member-first-notifications-at-linkedin. Accessed: 2025- 07-22

    work page 2021

  23. [23]

    LinkedIn Engineering. 2022. Open Sourcing Venice: LinkedIn’s Derived Data Plat- form. https://www.linkedin.com/blog/engineering/open-source/open-sourcing- venice-linkedin-s-derived-data-platform

    work page 2022

  24. [24]

    Zuxin Liu, Zijian Guo, Yihang Yao, Zhepeng Cen, Wenhao Yu, Tingnan Zhang, and Ding Zhao. 2023. Constrained decision transformer for offline safe re- inforcement learning. In International conference on machine learning . PMLR, 21611–21630

    work page 2023

  25. [25]

    Yi Ma, Chenjun Xiao, Hebin Liang, and Jianye Hao. 2023. Rethinking de- cision transformer via hierarchical reinforcement learning. arXiv preprint arXiv:2311.00267 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    Conor O’Brien, Huasen Wu, Shaodan Zhai, Dalin Guo, Wenzhe Shi, and Jonathan J Hunt. 2022. Should I send this notification? Optimizing push no- tifications decision making by modeling the future. arXiv:2202.08812 https: //arxiv.org/abs/2202.08812

    Pith/arXiv arXiv 2022

  27. [27]

    Prakruthi Prabhakar, Yiping Yuan, Guangyu Yang, Wensheng Sun, and Ajith Muralidharan. 2022. Multi-objective Optimization of Notifications Using Offline Reinforcement Learning. arXiv:2207.03029 [cs.LG] https://arxiv.org/abs/2207. 03029

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  28. [28]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. 2018. Reinforcement Learning: An Intro- duction (2 ed.). MIT Press. http://incompleteideas.net/book/the-book-2nd.html

    work page 2018

  29. [29]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 https://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2023

  30. [30]

    LiMAML: Personalization of Deep Recommender Models via Meta Learning

    Ruofan Wang, Prakruthi Prabhakar, Gaurav Srivastava, Tianqi Wang, Zeinab S. Jalali, Varun Bharill, Yunbo Ouyang, Aastha Nigam, Divya Venugopalan, Aman Gupta, Fedor Borisyuk, Sathiya Keerthi, and Ajith Muralidharan. 2024. Li- MAML: Personalization of Deep Recommender Models via Meta Learning. arXiv:2403.00803 [cs.IR] https://arxiv.org/abs/2403.00803

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    Siyu Wang, Xiaocong Chen, and Lina Yao. 2024. Retentive Decision Transformer with Adaptive Masking for Reinforcement Learning based Recommendation Systems. arXiv:2403.17634 [cs.IR] https://arxiv.org/abs/2403.17634

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [32]

    Yuanfu Wang, Chao Yang, Ying Wen, Yu Liu, and Yu Qiao. 2023. Critic-Guided Decision Transformer for Offline Reinforcement Learning. arXiv:2312.13716 https://arxiv.org/abs/2312.13716

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  33. [33]

    Taku Yamagata, Ahmed Khalil, and Raul Santos-Rodriguez. 2023. Q-learning De- cision Transformer: Leveraging Dynamic Programming for Conditional Sequence Modelling in Offline RL. arXiv:2209.03993 https://arxiv.org/abs/2209.03993

    Pith/arXiv arXiv 2023

  34. [34]

    Yann Collet. 2016. Zstandard - Real-time data compression algorithm . Facebook. http://facebook.github.io/zstd/

    work page 2016

  35. [35]

    Yiping Yuan, Jing Zhang, Shaunak Chatterjee, Shipeng Yu, and Romer Rosales

    work page

  36. [36]

    Bo Zhao, Koichiro Narita, Burkay Orten, and John Egan. 2018. Notification Volume Control and Optimization System at Pinterest. In KDD

    work page 2018

  37. [38]

    A State Transition Model for Mobile Notifications via Survival Analysis

    A State Transition Model for Mobile Notifications via Survival Analysis. arXiv:2207.03099 https://arxiv.org/abs/2207.03099

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [2022]

    In RecSys

    Off-policy actor-critic for recommender systems. In RecSys

    work page

  39. [2023]

    Knowledge-Based Systems (2023)

    Deep reinforcement learning in recommender systems: A survey and new perspectives. Knowledge-Based Systems (2023)

    work page 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.