Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

RecLLM-R1: A Two-Stage Training Paradigm with Reinforcement Learning and Chain-of-Thought v1

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper shows that supervised fine-tuning followed by reinforcement learning with chain-of-thought reasoning lets a small language model beat six recommender baselines on every measured accuracy metric.

desk verdict A plausible industrial recipe—SFT then GRPO with CoT for LLM-based sequential recommendation—but the paper's headline claims about diversity and filter-bubble mitigation are not measured, and the accuracy evidence is thin. read the letter →

arxiv 2506.19235 v1 pith:BDC52R26 submitted 2025-06-24 cs.AI

Yu Xie , Xingkai Ren , Ying Qi , Yao Hu , Lianlei Shan This is my paper
classification cs.AI
keywords LLM-basedrecommendersystemsgrouprelativepolicyoptimizationchain-of-thoughtreasoningsequentialrecommendationtwo-stagetrainingfilterbubblerewardfunctiondesign
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

This paper tries to show that a large language model can be turned into a practical sequential recommender by training it in two stages: first supervised fine-tuning to teach the recommendation format, then group-relative policy optimization with chain-of-thought reasoning to refine its choices. The concrete quantitative claim is that the resulting model, RecLLM-R1, outperforms six established sequential recommender baselines on Recall@5/10 and NDCG@5/10 across three Amazon categories, and improves Recall@10 from 0.4053 to 0.5311 and NDCG@10 from 0.4802 to 0.5653 against an unnamed online baseline on an industrial dataset. The authors also claim the approach reduces filter bubbles and can absorb business objectives through a flexible reward function, replacing separate strategy modules. If correct, the practical payoff is that business logic and model optimization no longer need to be separate pipelines: a single LLM policy can be directly optimized for whatever reward the business defines.

What carries the argument

The engine is the GRPO objective with chain-of-thought: for each user prompt, the current policy samples a group of twelve outputs, each output first generates a CoT rationale and then a ranked item sequence, and the group reward produces a baseline for advantage estimation without a critic network. The reward signal is a normalized, position-weighted longest common subsequence (LCS) that rewards correct items ranked earlier, with penalties for common errors; this is the mechanism that lets business metrics, diversity, or novelty be folded into training by redefining the reward. The two-stage sequence (SFT then GRPO) is what activates general language knowledge and then adapts it to recommendation-specific multi-step decisions.

What would settle it

A direct test is to rerun all baselines under exactly the same protocol—same candidate pool size, same labels, same textual metadata available to the model—and check whether RecLLM-R1's Recall/NDCG margins persist; a simpler check is to report the number of candidates ranked, since a smaller or easier pool for the proposed model would invalidate the headline improvements.

Watch

Extended reading notes

Core claim

The paper's central claim is that the recent recipe for training reasoning models—supervised fine-tuning for cold start, then reinforcement learning with chain-of-thought—transfers to recommendation. RecLLM-R1 first rewrites user profiles, interaction histories, and item attributes into natural-language prompts, fine-tunes a small distilled 1.5B reasoning-oriented language model on a compact labeled set, and then runs GRPO over groups of twelve candidate outputs per prompt. Each output consists of a chain-of-thought rationale followed by a ranked sequence of item IDs, and the reward is a normalized, position-weighted longest common subsequence between the predicted sequence and the sequence of items the user later engaged with. On the evidence shown, the model beats all six baselines on every accuracy metric on all three Amazon datasets, with the largest relative gain a 34.22% NDCG@10 improvement on Sports and Outdoors; on the industrial data it also beats the online baseline. The paper additionally asserts improved diversity, novelty, and filter-bubble mitigation, though the reported tables quantify only ranking accuracy.

Load-bearing premise

The claim stands on the assumption that the baselines were evaluated against the same candidate pool and labels as RecLLM-R1; if the model ranked fewer or differently filtered candidates, or was the only one given rich item text, the accuracy gains could be an artifact of the comparison rather than of the training method.

Editorial extensions

If this is right

  • A compact SFT stage plus a GRPO stage with chain-of-thought yields accuracy gains that transfer across three Amazon categories, so the recipe does not appear to be tuned to a single domain.
  • The position-weighted LCS reward means ranking quality (NDCG) improves more than raw retrieval (Recall), so the method is especially suited to list-ranking tasks rather than pure candidate generation.
  • Business objectives can be encoded in the reward function, so the same model can be re-optimized when policy changes, without retraining separate strategy modules.
  • Because the model emits a chain of thought before each ranked list, the system produces human-readable rationales that could be used for explanation or auditing.
  • The whole pipeline runs on a 1.5B parameter model with a compact dataset on eight H800 GPUs, suggesting LLM recommenders are trainable in resource-constrained settings.

Reading between the lines

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

  • A fair test of where the gain comes from is to feed the same textual item descriptions to a traditional sequential model; if that model also improves, the GRPO stage may be doing less work than the paper's attribution suggests.
  • The claim of filter-bubble mitigation is not backed by any diversity or novelty metric in the experiments, so a targeted evaluation measuring category or topic diversity of recommended lists would be needed to confirm that specific benefit.
  • Because the reward is a plug-in component, the identical training pipeline could be pointed at other objectives—say, fresh-content share or user retention—without changing the model; the paper proposes this flexibility but does not test it.
  • The use of a small 1.5B model raises the open question of whether larger models would scale the gains further, or whether the two-stage recipe saturates quickly with capacity.
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 / 6 minor

Summary. The paper proposes RecLLM-R1, a two-stage SFT+GRPO+CoT training pipeline for LLM-based sequential recommendation. It converts user profiles, interaction histories, and item metadata into natural-language prompts; fine-tunes DeepSeek-R1-Distill-Qwen-1.5B with SFT; then applies GRPO with a position-weighted LCS reward to generate a reasoning step and an item sequence. Experiments on three Amazon categories and an unnamed industrial dataset report Recall@5/10 and NDCG@5/10 improvements over six sequential recommenders and an online baseline. The abstract claims additional gains in diversity and novelty and mitigation of the filter-bubble effect, but the experimental section measures only accuracy-oriented metrics.

Significance. If the accuracy results hold under a properly controlled protocol, the paper shows a practical recipe for training a 1.5B LLM with RL for sequential recommendation, with a transparent and customizable reward. The data-construction and two-stage training design are clearly described, and the use of GRPO without a critic is well motivated. The paper's major advertised differentiators—diversity, novelty, and filter-bubble mitigation—are not measured, and the evaluation protocol is under-specified, so the current evidence supports only a narrow claim of improved Recall/NDCG on the tested benchmarks. The stated intent to open-source the code is a positive step for reproducibility, provided the release actually occurs.

major comments (4)
  1. [Abstract and Section 5] The abstract and conclusion claim that RecLLM-R1 surpasses baselines 'across ... accuracy, diversity, and novelty' and 'effectively mitigates the filter bubble effect,' but Section 4.1 defines only Recall@K and NDCG@K and Tables 1 and 2 report only these two metrics. No diversity, novelty, coverage, or consumption-entropy metric appears anywhere in the experimental section. Because these multi-objective claims are a central part of the paper's stated contribution, the current experiments do not support them.
  2. [Section 3.3 vs. Section 4.2] Section 4.2 states that the reward is 'a normalized, position-weighted Longest Common Subsequence (LCS) algorithm' with penalties for common errors, and Section 3.3 says the reward 'can combine' traditional and business-specific metrics, but no diversity or novelty term is described or tested. Without a diversity- or novelty-related component in the implemented reward, there is no mechanism by which the claimed filter-bubble mitigation could arise, and the abstract's multi-objective claim is not backed by the system as described.
  3. [Sections 4.1–4.3] The empirical comparison is not fully specified. Section 4.1 describes labels built from strong later interactions, and Section 4.2 says candidate items are drawn from 'the provided data pool,' but the paper never states the candidate-pool size, whether the same candidate pool is used for every baseline, how many candidates RecLLM-R1 ranks, or how the unnamed online baseline was exposed to items. If RecLLM-R1 ranked a smaller or leakage-prone pool, or if the baselines did not receive the same textual metadata, the gains in Tables 1 and 2 could be artifacts. In addition, Table 1 reports single runs without error bars or significance tests, while Section 4.3 uses the word 'significantly'; this should be backed by statistical evidence.
  4. [Sections 3.3 and 4.2] The reward function is closely aligned with the evaluation metrics: position-weighted LCS rewards correct items in early positions, which is essentially a smoothed version of Recall/NDCG. This train/eval alignment should be discussed and controlled for, for instance by reporting metrics that are not part of the reward (such as diversity or novelty) or by including baselines that have been tuned to optimize Recall/NDCG. As it stands, part of the observed improvement may reflect directly optimizing a surrogate of the metric rather than a general recommendation-quality gain.
minor comments (6)
  1. [Tables 1 and 2] The model name is written 'recLLM-R1' in Tables 1 and 2 but 'RecLLM-R1' elsewhere; please unify the spelling.
  2. [Equation (1), Section 3.3] The GRPO objective in Equation (1) writes the KL-penalty term in a nonstandard form; please verify it against the original GRPO formulation [29] and define all symbols (e.g., \hat{A}_{k,t}) unambiguously.
  3. [Section 3.2] Section 3.2 says the SFT stage uses 'a minimal amount of data' but gives no quantitative details (number of prompts, epochs, learning rate); a table of hyperparameters would help reproducibility.
  4. [Reference [17]] Reference [17] is cited for the filter-bubble effect, but [17] concerns matrix factorization for recommender systems; please cite a more appropriate source or clarify the connection.
  5. [Section 4.1] The phrase 'currently under review for public release' for the industrial dataset is vague; if the dataset will be released, specify the plan, and otherwise clarify the status.
  6. [Table 1] The parenthetical percentages in Table 1 are not defined; state explicitly that they are relative improvements over the best baseline.

Circularity Check

0 steps flagged · score 0.0 of 10

RecLLM-R1 shows no circular derivation: the SFT/GRPO pipeline is empirically evaluated against held-out data, and the mismatch between the implemented LCS reward and the advertised diversity/novelty gains is a missing-evidence problem, not a definitional reduction.

full rationale

The paper's central claim is an empirical one: that an SFT-then-GRPO LLM pipeline with a CoT stage outperforms baselines on Recall/NDCG on three Amazon categories and an industrial dataset. I traced the derivation chain: data construction (Section 3.1) converts user profiles, histories, and item metadata into prompts; the SFT stage (Section 3.2) fine-tunes DeepSeek-R1-Distill-Qwen-1.5B; the GRPO stage (Sections 3.3 and 4.2) optimizes a policy with a normalized, position-weighted LCS reward. Nothing in the paper defines the reported Recall@K/NDCG@K numbers in terms of the reward function or in terms of a fitted parameter; the rewards are computed on training prompts, and the metrics are measured on held-out users/items. There is no self-citation chain: the cited DeepSeek R1 and VeRL works are external infrastructure, not the authors' prior results, and no uniqueness theorem is invoked to forbid alternatives. The closest circularity-adjacent point is that the LCS reward is a ranking-matching objective similar in spirit to Recall/NDCG, but optimizing a surrogate on training data and then measuring the target metric on test data is standard practice, not a definitional identity. The paper does overclaim diversity, novelty, and filter-bubble mitigation: Section 4.1 defines only Recall@K and NDCG@K, and no diversity/novelty metric or reward term is described. That is an evidentiary gap for the advertised multi-objective contribution, not a circular step in the derivation. Accordingly, no circularity steps are flagged.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new entities are invented. The central claims rely on the correctness and applicability of GRPO from prior work, on textualization being an adequate representation, on engagement-based labels being trustworthy, and on a fair evaluation protocol. Several key hyperparameters and reward weights are free parameters that the paper does not report.

free parameters (5)
  • Reward component weights and error penalties = not reported
    The reward in Section 4.2 is a position-weighted LCS plus unspecified penalties; exact combination weights are not given.
  • GRPO KL penalty coefficient beta = not reported
    Appears in the objective in Section 3.3; no default value is stated.
  • Number of GRPO iterations = not reported
    Section 4.2 gives group size 12 but not training steps, rollout count, or convergence criterion.
  • Candidate pool size per prompt = not reported
    Section 4.1 and 4.2 refer to a 'data pool' without stating how many candidates are ranked; this directly affects Recall and NDCG values.
  • SFT training data size = not reported
    Section 4.2 mentions a compact high-quality subset but gives no sample count or data mix.
assumptions (4)
  • standard math GRPO provides a valid policy-gradient update for the stated objective.
    The objective in Section 3.3 is taken from DeepSeekMath and DeepSeek-R1 without proof; the method relies on its convergence and stability.
  • domain assumption Textual descriptions of user profiles, interactions, and item tags are sufficient representations for recommendation.
    Section 3.1 builds prompts entirely from these fields; if the text loses key signals, the whole pipeline is handicapped.
  • domain assumption Labels built from strong engagement signals reflect genuine user preference.
    Section 4.1 constructs positive sequences from dwell time and deep interactions; biased engagement would bias SFT labels and GRPO rewards.
  • domain assumption The evaluation protocol is fair and leakage-free across RecLLM-R1 and all baselines.
    Section 4.1 orders histories by timestamp and Section 4.3 compares numbers, but the candidate pool and baseline exposure are not documented.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RecLLM-R1: A Two-Stage Training Paradigm with Reinforcement Learning and Chain-of-Thought v1." pith.science (2026). https://pith.science/paper/BDC52R26

@misc{pith2026250619235,
  author       = {Pith},
  title        = {Pith review of: RecLLM-R1: A Two-Stage Training Paradigm with Reinforcement Learning and Chain-of-Thought v1},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BDC52R26}},
  note         = {Machine review of arXiv:2506.19235}
}
read the original abstract

Traditional recommendation systems often grapple with "filter bubbles", underutilization of external knowledge, and a disconnect between model optimization and business policy iteration. To address these limitations, this paper introduces RecLLM-R1, a novel recommendation framework leveraging Large Language Models (LLMs) and drawing inspiration from the DeepSeek R1 methodology. The framework initiates by transforming user profiles, historical interactions, and multi-faceted item attributes into LLM-interpretable natural language prompts through a carefully engineered data construction process. Subsequently, a two-stage training paradigm is employed: the initial stage involves Supervised Fine-Tuning (SFT) to imbue the LLM with fundamental recommendation capabilities. The subsequent stage utilizes Group Relative Policy Optimization (GRPO), a reinforcement learning technique, augmented with a Chain-of-Thought (CoT) mechanism. This stage guides the model through multi-step reasoning and holistic decision-making via a flexibly defined reward function, aiming to concurrently optimize recommendation accuracy, diversity, and other bespoke business objectives. Empirical evaluations on a real-world user behavior dataset from a large-scale social media platform demonstrate that RecLLM-R1 significantly surpasses existing baseline methods across a spectrum of evaluation metrics, including accuracy, diversity, and novelty. It effectively mitigates the filter bubble effect and presents a promising avenue for the integrated optimization of recommendation models and policies under intricate business goals.

Figures

Figures reproduced from arXiv: 2506.19235 by the authors.

Figure 1
Figure 1. Comparison of three mainstream recommendation frameworks, Current frameworks / [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustrates the motivation behind our research [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An overview of the modeling approach used in RecLLM-R1. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. MAViE: A Multi-scale Adaptive Vision Encoder for Fine-grained Visual Perception and Efficient Multimodal Reasoning

    cs.CV 2026-07 reject novelty 4.0 of 10

    A multi-scale gated fusion plus question-conditioned token router is specified to cut VLM visual tokens ~80% while improving accuracy, but all reported gains are simulated placeholders.

Reference graph

Works this paper leans on

39 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Reinforcement learning based recommender systems: A survey

    M Mehdi Afsar, Trafford Crump, and Behrouz Far. Reinforcement learning based recommender systems: A survey. ACM Computing Surveys, 55(7):1–38, 2022

    2022

  3. [3]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  4. [4]

    Twin: Two-stage interest network for lifelong user behavior modeling in ctr prediction at kuaishou

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. Twin: Two-stage interest network for lifelong user behavior modeling in ctr prediction at kuaishou. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3785–3794, 2023

    work page 2023

  5. [5]

    Wide & deep learning for recommender systems

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al. Wide & deep learning for recommender systems. In Proceedings of the 1st workshop on deep learning for recommender systems, pages 7–10, 2016

    work page 2016

  6. [6]

    Deep neural networks for youtube recommenda- tions

    Paul Covington, Jay Adams, and Emre Sargin. Deep neural networks for youtube recommenda- tions. In Proceedings of the 10th ACM conference on recommender systems, pages 191–198, 2016

    work page 2016

  7. [7]

    Flashattention: Fast and memory-efficient exact attention with io-awareness

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344–16359, 2022

    2022

  8. [8]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019

    2019

Show all 39 references
  1. [9]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

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

    2025 arXiv

  2. [10]

    Deepfm: a factorization-machine based neural network for ctr prediction.arXiv preprint arXiv:1703.04247, 2017

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. Deepfm: a factorization-machine based neural network for ctr prediction.arXiv preprint arXiv:1703.04247, 2017

    2017 arXiv

  3. [11]

    Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering

    Ruining He and Julian McAuley. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. Inproceedings of the 25th international conference on world wide web, pages 507–517, 2016

    2016

  4. [12]

    Neural collaborative filtering

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web, pages 173–182, 2017

    2017

  5. [13]

    Session-based recommendations with recurrent neural networks

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939, 2015

    2015 arXiv

  6. [14]

    Factual and personalized recommendations using language models and reinforcement learning

    Jihwan Jeong, Yinlam Chow, Guy Tennenholtz, Chih-Wei Hsu, Azamat Tulepbergenov, Mo- hammad Ghavamzadeh, and Craig Boutilier. Factual and personalized recommendations using language models and reinforcement learning. arXiv preprint arXiv:2310.06176, 2023

    arXiv

  7. [15]

    Genrec: Large language model for generative recommendation

    Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. Genrec: Large language model for generative recommendation. InEuropean Conference on Information Retrieval, pages 494–502. Springer, 2024. 10

    2024

  8. [16]

    Self-attentive sequential recommendation

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

    2018

  9. [17]

    Matrix factorization techniques for recom- mender systems

    Yehuda Koren, Robert Bell, and Chris V olinsky. Matrix factorization techniques for recom- mender systems. Computer, 42(8):30–37, 2009

    2009

  10. [18]

    Gpt4rec: A generative framework for personalized recommendation and user interests interpretation

    Jinming Li, Wentao Zhang, Tian Wang, Guanglei Xiong, Alan Lu, and Gerard Medioni. Gpt4rec: A generative framework for personalized recommendation and user interests interpretation. arXiv preprint arXiv:2304.03879, 2023

    2023 arXiv

  11. [19]

    Llara: Large language-recommendation assistant

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. Llara: Large language-recommendation assistant. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1785– 1795, 2024

    2024

  12. [20]

    Cascade ranking for operational e-commerce search

    Shichen Liu, Fei Xiao, Wenwu Ou, and Luo Si. Cascade ranking for operational e-commerce search. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1557–1565, 2017

    2017

  13. [21]

    Integrating large language models into recommendation via mutual augmentation and adaptive aggregation

    Sichun Luo, Yuxuan Yao, Bowei He, Yinya Huang, Aojun Zhou, Xinyi Zhang, Yuanzhang Xiao, Mingjie Zhan, and Linqi Song. Integrating large language models into recommendation via mutual augmentation and adaptive aggregation. arXiv preprint arXiv:2401.13870, 2024

    2024

  14. [22]

    Human-level control through deep reinforcement learning

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529–533, 2015

    2015

  15. [23]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:2773...

    2022

  16. [24]

    Large language model based long-tail query rewriting in taobao search

    Wenjun Peng, Guiyang Li, Yue Jiang, Zilong Wang, Dan Ou, Xiaoyi Zeng, Derong Xu, Tong Xu, and Enhong Chen. Large language model based long-tail query rewriting in taobao search. In Companion Proceedings of the ACM Web Conference 2024, pages 20–28, 2024

    2024

  17. [25]

    Rankflow: Joint optimization of multi-stage cascade ranking systems as flows

    Jiarui Qin, Jiachen Zhu, Bo Chen, Zhirong Liu, Weiwen Liu, Ruiming Tang, Rui Zhang, Yong Yu, and Weinan Zhang. Rankflow: Joint optimization of multi-stage cascade ranking systems as flows. In Proceedings of the 45th International ACM SIGIR Conference on Research and Developmen...

    2022

  18. [26]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    2023

  19. [27]

    Recommender systems with generative retrieval

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al. Recommender systems with generative retrieval. Advances in Neural Information Processing Systems, 36:10299–10315, 2023

    2023

  20. [28]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  21. [29]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

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

    2024 arXiv

  22. [30]

    Hybridflow: A flexible and efficient rlhf framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv:2409.19256, 2024

    2024 arXiv

  23. [31]

    Learning to summarize with human feedback

    Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in neural information processing systems, 33:3008–3021, 2020. 11

    2020

  24. [32]

    Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, pages ...

    2019

  25. [33]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  26. [34]

    Kerl: A knowledge-guided reinforcement learning model for sequential recommendation

    Pengfei Wang, Yu Fan, Long Xia, Wayne Xin Zhao, ShaoZhang Niu, and Jimmy Huang. Kerl: A knowledge-guided reinforcement learning model for sequential recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retriev...

    2020

  27. [35]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  28. [36]

    Palr: Personalization aware llms for recommendation

    Fan Yang, Zheng Chen, Ziyan Jiang, Eunah Cho, Xiaojiang Huang, and Yanbin Lu. Palr: Personalization aware llms for recommendation. arXiv preprint arXiv:2305.07622, 2023

    2023 arXiv

  29. [37]

    Feature-level deeper self-attention network for sequential recommendation

    Tingting Zhang, Pengpeng Zhao, Yanchi Liu, Victor S Sheng, Jiajie Xu, Deqing Wang, Guan- feng Liu, Xiaofang Zhou, et al. Feature-level deeper self-attention network for sequential recommendation. In IJCAI, pages 4320–4326, 2019

    2019

  30. [38]

    Deep interest network for click-through rate prediction

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages...

    2018

  31. [39]

    S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization. In Proceedings of the 29th ACM international conference on informatio...

    1902

Pith tools

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