Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

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

T0 review · 5 major / 7 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 A trained reflection model that critiques and revises LLM-drafted knowledge improves recommendation backbones and lifts online revenue by 2.2%.

desk verdict A solid applied paper that brings trained reflection-refinement to LLM knowledge extraction for recsys, but the causal role of the reflection loop is undercut by a training-data selection confound and missing code. read the letter →

arxiv 2507.17249 v2 pith:TAKPDLX6 submitted 2025-07-23 cs.IR

Hao Gu , Rui Zhong , Yu Xia , Wei Yang , Chi Lu , Peng Jiang , Kun Gai This is my paper
classification cs.IR
keywords largelanguagemodelsrecommendationsystemsSystem-2thinkingreflectionandrefinementknowledgeaugmentationCTRpredictionself-refine
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

R4ec is a framework for making LLM-generated recommendation knowledge more reliable by adding a trained critic. An actor model writes and revises summaries of user preferences and item facts; a separate reflection model judges whether each summary is reasonable and, when it is not, supplies feedback the actor uses to rewrite the summary. The paper claims this iterative reflect-and-revise loop, a step toward the slow, deliberate mode of reasoning it calls System-2 thinking, produces knowledge that helps downstream recommendation models more than the same pipeline without reflection or a one-shot chain-of-thought knowledge baseline. Across six click-through-rate backbones on Amazon-Book and MovieLens-1M, the refined knowledge raises AUC and lowers LogLoss, and an online advertising deployment reports a 2.2% revenue gain.

What carries the argument

The load-bearing object is the trained actor-reflection loop. The actor model $\pi_\theta$ generates and refines user preference or item factual knowledge; the reflection model $\pi_\psi$ independently judges the knowledge and returns either an approval or a concrete critique. The two are trained separately with low-rank adapters on datasets built by a teacher LLM: reasoning and refinement corpora for the actor, and a reflection corpus containing both positive judgments and critiques that, when acted on, changed a wrong prediction into the correct one. At inference the loop iterates until $\pi_\psi$ approves or the retry budget is reached, and the surviving text knowledge is embedded and combined with categorical features through an MLP connector into the recommendation backbone. The design choice that carries the argument is that the critic is a separate trained model rather than the actor judging itself.

What would settle it

On a held-out slice of Amazon-Book, force the reflection model to always output 'The user preference is reasonable' while keeping everything else fixed; if AUC and LogLoss stay at R4ec levels rather than falling toward R2ec, the critic is not what drives the gain. A complementary check is to measure the reflection model's approval rate on unseen users against judgments from a stronger model or human annotators.

Watch

Extended reading notes

Core claim

The paper's central claim is that separating reasoning from judgment, and letting the two interact, is what makes LLM knowledge useful for recommendations, and it presents this as the first recommendation-system study of System-2 thinking through iterative reflection and refinement. R4ec trains an actor LLM on reasoning and refinement examples and a reflection LLM on judgments that mark knowledge as reasonable or unreasonable; the reflection model's feedback is produced by a larger teacher LLM and filtered by whether the original and refined predictions match the training label. At serving time the actor's knowledge is revised under the reflection model's feedback until approval or a retry cap, then encoded as dense vectors and fed through an MLP connector into a conventional click-through-rate backbone. Compared with R2ec, the same knowledge pipeline without reflection and refinement, R4ec reports average relative AUC improvements around 0.6% on both public datasets and larger LogLoss reductions across six backbones, and the online A/B test shows 2.2% higher revenue and 1.6% higher conversion rate.

Load-bearing premise

The load-bearing premise is that the reflection model's judgments generalize from the users and items it saw in training to unseen ones; if the critic only approves the actor's first draft at serving time, the refinement loop adds nothing and R4ec reduces to the single-pass baseline.

Editorial extensions

If this is right

  • Adding a single refinement step improves all six CTR backbones on both datasets, so the gain does not depend on one architecture.
  • With the actor fixed at 7B parameters, raising the reflection model from 0.5B to 72B improves AUC and LogLoss, making critic quality a separate scaling axis.
  • The Iterative Refinement strategy outperforms Reflection as a Filter, indicating that using feedback to revise knowledge adds value beyond selecting among multiple drafts.
  • Increasing the number of refinement steps yields further but diminishing gains, so inference-time computation can be traded for recommendation accuracy.
  • In the online advertising test the framework raised revenue by 2.2% overall and 4.1% on long-tail data, implying the revised knowledge helps cold-start and sparse-user cases.

Reading between the lines

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

  • The same label-filtered actor-critic loop could be applied to other LLM side-information generators in ranking, such as query rewriting or ad creative text, wherever a downstream click label can supervise which reflections are worth keeping.
  • Because the reflection model is trained to flag knowledge that correlates with wrong label predictions rather than with factual truth, its feedback may encode the backbone's existing biases; on noisy labels, refinement could reinforce rather than correct errors.
  • The paper's approximation for inactive users, substituting item knowledge from the user's history for user preference inference, suggests a cheaper serving path; one testable extension is distilling the refined outputs into a single-pass actor to remove iterative inference cost.
  • The scaling results imply that production budgets might be better spent on a larger critic than a larger generator, since a fixed 7B critic improves even a 72B actor.
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

5 major / 7 minor

Summary. The paper proposes R4ec, a framework that trains two small LLMs (an actor and a reflection model) to iteratively refine user-preference and item-factual knowledge, which is then encoded with BGE-M3 and fed as side information into conventional CTR backbones. Training data is constructed with GPT-4o: the actor is trained on reasoning and refinement corpora, and the reflection model is trained on filtered GPT-4o judgments about whether the actor's knowledge is reasonable. Offline experiments on Amazon-Book and MovieLens-1M with six backbones show consistent AUC and LogLoss gains over the base backbones, over a KAR baseline, and over R2ec (the actor without reflection). The paper also reports a 14-day online A/B test with a 2.2% revenue lift and scaling experiments for both actor and reflection model sizes.

Significance. If the reported gains genuinely come from the reflection-and-refinement mechanism, the framework is a practical, model-agnostic way to add test-time deliberation to LLM-based knowledge augmentation for recommendation, with consistent offline gains across six backbones and an industrial deployment. The scaling analyses (Sec. 4.4.2-4.4.3) are a useful contribution, and the online revenue lift, if properly benchmarked, would be practically important. However, the central attribution of the R4ec-vs-R2ec improvement to the reflection loop is not yet established because of training-set selection, a confounded comparison, and missing statistical or behavioral evidence about what the reflection model does at inference.

major comments (5)
  1. [Sec. 3.2, Algorithm 1 (lines 5-13)] The reflection training set contains only two cells of the judge x label contingency table: reasonable-and-correct (lines 6-7) and unreasonable-and-wrong-then-corrected (lines 8-12). The discarded cells, reasonable-but-wrong and unreasonable-but-correct, are exactly the ambiguous cases the reflection model must handle for unseen users and items at inference. Because the label is absent from the reflection prompt in Fig. 3, the model may learn 'reasonable' as a proxy for 'prediction was correct'. The paper reports no statistic on how often the reflection model rejects initial outputs at inference, nor how often refinement changes the final prediction. Without such evidence, the claimed advantage of the reflection/refinement loop over R2ec is not established; the loop could be approving most outputs, and the gains could come from other components.
  2. [Sec. 3.4 and Table 2] The R4ec-vs-R2ec comparison is confounded. R2ec's actor is trained only on D_reason, whereas R4ec's actor is trained on D_reason union D_refine (Eqs. 7-9). Thus the comparison conflates the reflection/refinement mechanism with the additional refinement training data. An ablation is needed in which the actor is trained on D_reason union D_refine but at inference the reflection step is removed or replaced by non-informative feedback, and the result is compared with the full R4ec pipeline. The scaling experiments in Sec. 4.4.3 compare 'with' vs 'without a reflection model' but do not specify the actor's training data in the 'without' condition, so the same confound is present.
  3. [Sec. 4.1.4 and Sec. 4.2] The comparison with KAR is not controlled: KAR uses GPT-3.5 with direct chain-of-thought prompting, while R4ec uses fine-tuned Qwen2.5-7B with reflection and refinement. The reported 1.36% relative AUC improvement over KAR could be due to the base model, the instruction tuning, the additional training data, or the framework. A controlled comparison should run KAR with the same Qwen2.5-7B base and the same knowledge encoder, or run R2ec with GPT-3.5. Without that, the claim that the framework outperforms KAR is unsupported.
  4. [Sec. 4.1.5 and Sec. 3.5] The reflection model is trained on responses generated by GPT-4o (Sec. 3.2), but at inference the actor is Qwen2.5-7B. There is a distribution shift between the responses the reflection model saw during training and those it judges at inference. The paper provides no evidence (e.g., agreement rates or rejection rates on Qwen-generated outputs) that the trained reflection model generalizes to the actor's outputs. This is load-bearing because the framework's benefit depends on the reflection model detecting flaws in the actor's knowledge.
  5. [Sec. 4.2 and Table 2] All offline results are reported as point estimates with no standard deviations, seeds, or significance tests. The R4ec-vs-R2ec AUC differences are small (roughly 0.5-0.8% relative), and without variance estimates it is unclear whether the differences are robust. Given the confounds above, the consistency across backbones is supportive but not conclusive. Reporting at least 3-5 seeds with means and standard deviations would address this concern.
minor comments (7)
  1. [Sec. 4.1.1] The sentence 'We regard reviews with ratings greater than 5 with positive' is impossible for Amazon-Book's 1-5 rating scale and should be corrected (e.g., greater than 4, consistent with the prompt in Fig. 3).
  2. [Algorithm 1, line 4] The call to M(P_user_reflect(hist, item, upre)) is missing a closing parenthesis; the formula should be balanced.
  3. [Sec. 4.3] The online baseline is not described; please state what the 20% traffic baseline is (Base, KAR, R2ec, or the previous production model) and report confidence intervals for the 2.2% revenue and 1.6% CVR lifts.
  4. [Sec. 4.4.2 and Fig. 4] The x-axis in Fig. 4 includes 0, but no model of size 0 is defined; clarify whether 0 denotes the no-reflection condition.
  5. [Introduction and Conclusion] The claim of being 'the first study within recommendation systems to explore System-2 thinking through iterative reflection and refinement' is too strong given existing self-reflection work in NLP and recommendation, and should be softened.
  6. [Sec. 3.6] The connectors F_u and F_i are described as MLPs, but their dimensions, input formats, and training details are not specified; please add these details for reproducibility.
  7. [Sec. 1 and Implementation Details] The paper refers to 'our implementation code' but provides no code repository URL; please include one.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central results are held-out evaluations of a supervised distillation pipeline, and no derivation reduces to its own inputs.

full rationale

The paper's headline claims are empirical AUC and LogLoss gains on held-out test splits (first 80% of interactions chronologically for training, the remainder for testing), so the evaluation is against labels not used in constructing the reflection and refinement supervision. Algorithm 1 filters GPT-4o judgments using the training label (e.g., line 5 requires pred == label; line 8 requires pred != label), but this is standard supervised distillation rather than a tautology: the reflection model is trained on a subset of the training data and then applied at inference to unseen users and items, with no access to the label. The R4ec-versus-R2ec comparison is confounded because R4ec's actor is trained on D_reason ∪ D_refine while R2ec uses only D_reason, so the ablation does not perfectly isolate the reflection mechanism; however, this is an experimental design issue, not a circular derivation. The paper's self-citations ([14], [52]) are unrelated to the core framework and carry no load-bearing argument; no uniqueness theorem or ansatz is imported from prior author work. The 'System-2 thinking' language is an analogy, not a quantity derived from the framework. No step in the paper makes a predicted quantity equal to a fitted input by construction, so no circular step can be exhibited.

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

The paper is an empirical ML pipeline with no formal derivation; the central claim rests on supervised learning assumptions and the transferability of distilled reasoning labels. No free constants are fitted, but several labeling and training choices are ad hoc.

free parameters (4)
  • Positive-label thresholds = Amazon-Book: rating > 5 (likely >4); MovieLens-1M: rating > 3
    These thresholds convert ratings into binary labels for both LLM dataset construction and downstream evaluation; Amazon's >5 is an apparent typo, indicating at least one such rule is imprecise.
  • LoRA hyperparameters = rank r=8, alpha=16, dropout=0.05, 3 epochs
    Standard fine-tuning configuration, not core to the claim but affects results.
  • Default refinement iterations = 1
    The iterative strategy is described as central, but the default is a single refinement step; the observed gains come mostly from one round.
  • Data construction coverage = 40% of users and items
    Dataset constructed from 40% of users/items with one sample each; choice affects training distribution.
assumptions (4)
  • domain assumption Ratings are valid proxies for user preference
    The binary labels 'like/dislike' are derived solely from rating thresholds, and the entire supervised training of actor and reflection models depends on this.
  • domain assumption GPT-4o's reasonableness judgments are accurate enough to serve as training signal
    The reflection and refinement datasets are distilled from GPT-4o's own judgments (Section 3.2-3.3); any bias in GPT-4o propagates.
  • domain assumption BGE-M3 embeddings of refined text preserve the signal needed by CTR backbones
    The text knowledge is only used through dense embeddings (Eq. 11); if the encoder loses the reasoning signal, the backend cannot benefit.
  • domain assumption The six backbone models are a representative testbed
    Generalization across systems is inferred from six CTR models; the framework may not transfer to other architectures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of R4ec: A Reasoning, Reflection, and Refinement Framework for Recommendation Systems." pith.science (2026). https://pith.science/paper/TAKPDLX6

@misc{pith2026250717249,
  author       = {Pith},
  title        = {Pith review of: R4ec: A Reasoning, Reflection, and Refinement Framework for Recommendation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TAKPDLX6}},
  note         = {Machine review of arXiv:2507.17249}
}
abstract

Harnessing Large Language Models (LLMs) for recommendation systems has emerged as a prominent avenue, drawing substantial research interest. However, existing approaches primarily involve basic prompt techniques for knowledge acquisition, which resemble System-1 thinking. This makes these methods highly sensitive to errors in the reasoning path, where even a small mistake can lead to an incorrect inference. To this end, in this paper, we propose $R^{4}$ec, a reasoning, reflection and refinement framework that evolves the recommendation system into a weak System-2 model. Specifically, we introduce two models: an actor model that engages in reasoning, and a reflection model that judges these responses and provides valuable feedback. Then the actor model will refine its response based on the feedback, ultimately leading to improved responses. We employ an iterative reflection and refinement process, enabling LLMs to facilitate slow and deliberate System-2-like thinking. Ultimately, the final refined knowledge will be incorporated into a recommendation backbone for prediction. We conduct extensive experiments on Amazon-Book and MovieLens-1M datasets to demonstrate the superiority of $R^{4}$ec. We also deploy $R^{4}$ec on a large scale online advertising platform, showing 2.2\% increase of revenue. Furthermore, we investigate the scaling properties of the actor model and reflection model.

Figures

Figures reproduced from arXiv: 2507.17249 by the authors.

Figure 1
Figure 1. Our iterative reflection and refinement mechanism. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of our user preference reasoning, reflection and refinement dataset construction process. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Prompt template for constructing user preference reasoning, reflection and refinement dataset on Amazon-Book. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: We use the Qwen-2.5 7B model as the reflection [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: GRU4Rec’s AUC and LogLoss performance on [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 6
Figure 6. Figure 6: We compare the performance of AutoInt and [PITH_FULL_IMAGE:figures/full_fig_p009_6.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.

  1. TimeMM: Time-as-Operator Spectral Filtering for Dynamic Multimodal Recommendation

    cs.IR 2026-04 unverdicted novelty 6.0 of 10

    TimeMM proposes a time-as-operator spectral filtering framework with adaptive mixing and modality routing to model non-stationary multimodal user preferences in recommendation systems.

  2. TrackRec: Iterative Alternating Feedback with Chain-of-Thought via Preference Alignment for Recommendation

    cs.IR 2025-08 conditional novelty 5.0 of 10

    TrackRec trains a small LLM to generate user-preference summaries and a validator to score them, alternating the training so each improves the other, and reports gains on public and industrial recommendation benchmarks.

Reference graph

Works this paper leans on

65 extracted references · 11 canonical work pages · cited by 2 Pith papers

  1. [1]

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

    arXiv 2023

  2. [2]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023). RecSys ’25, September 22–26, 2025, Prague, Czech Republic Hao Gu et al

    arXiv 2023

  3. [3]

    Zhuoxi Bai, Ning Wu, Fengyu Cai, Xinyi Zhu, and Yun Xiong. 2024. Finetuning Large Language Model for Personalized Ranking.arXiv preprint arXiv:2405.16127 (2024)

    arXiv 2024

  4. [4]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

    2023

  5. [5]

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

    arXiv 2020

  6. [6]

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

    arXiv 2024

  7. [7]

    Xinyun Chen, Maxwell Lin, Nathanael Schärli, and Denny Zhou. 2023. Teaching large language models to self-debug.arXiv preprint arXiv:2304.05128(2023)

    arXiv 2023

  8. [8]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongx- iang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in recommender systems. InProceedings of the 17th ACM Conference on Recom- mender Systems. 1126–1132

    2023

Show all 65 references
  1. [9]

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

    2018 arXiv

  2. [10]

    Yingpeng Du, Di Luo, Rui Yan, Xiaopei Wang, Hongzhi Liu, Hengshu Zhu, Yang Song, and Jie Zhang. 2024. Enhancing job recommendation through llm-based generative adversarial networks. InProceedings of the AAAI Conference on Artifi- cial Intelligence, Vol. 38. 8363–8371

    2024

  3. [11]

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-rec: Towards interactive and explainable llms-augmented recommender system.arXiv preprint arXiv:2303.14524(2023)

    2023 arXiv

  4. [12]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InProceedings of the 16th ACM Conference on Recommender Systems. 299–315

    2022

  5. [13]

    Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. 2023. Critic: Large language models can self-correct with tool-interactive critiquing.arXiv preprint arXiv:2305.11738(2023)

    2023 arXiv

  6. [14]

    Hao Gu, Jiangyan Yi, Chenglong Wang, Jianhua Tao, Zheng Lian, Jiayi He, Yong Ren, Yujie Chen, and Zhengqi Wen. 2025. ALLM4ADD: Unlocking the Capabilities of Audio Large Language Models for Audio Deepfake Detection.arXiv preprint arXiv:2505.11079(2025)

    2025 arXiv

  7. [15]

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

    2017 arXiv

  8. [16]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639–648

    2020

  9. [17]

    1991.The Encyclopaedia Logic, with the Zus tze: Part I of the Encyclopaedia of Philosophical Sciences with the Zusätze

    Georg Wilhelm Friedrich Hegel. 1991.The Encyclopaedia Logic, with the Zus tze: Part I of the Encyclopaedia of Philosophical Sciences with the Zusätze. Vol. 1. Hackett Publishing

    1991

  10. [18]

    B Hidasi. 2015. Session-based Recommendations with Recurrent Neural Networks. arXiv preprint arXiv:1511.06939(2015)

    2015 arXiv

  11. [19]

    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

  12. [20]

    Yixin Ji, Juntao Li, Hai Ye, Kaixin Wu, Jia Xu, Linjian Mo, and Min Zhang. 2025. Test-time Computing: from System-1 Thinking to System-2 Thinking.arXiv preprint arXiv:2501.02497(2025)

    2025 arXiv

  13. [21]

    Nan Jiang, Xiaopeng Li, Shiqi Wang, Qiang Zhou, Soneya Binta Hossain, Baishakhi Ray, Varun Kumar, Xiaofei Ma, and Anoop Deoras. 2024. Training LLMs to Better Self-Debug and Explain Code.arXiv preprint arXiv:2405.18649 (2024)

    2024 arXiv

  14. [22]

    Daniel Kahneman. 2011. Thinking, fast and slow.Farrar, Straus and Giroux (2011)

    2011

  15. [23]

    Pei Ke, Bosi Wen, Zhuoer Feng, Xiao Liu, Xuanyu Lei, Jiale Cheng, Shengyuan Wang, Aohan Zeng, Yuxiao Dong, Hongning Wang, et al . 2023. Critiquellm: Scaling llm-as-critic for effective and explainable evaluation of large language model generation.arXiv preprint arXiv:2311.18702(2023)

    2023 arXiv

  16. [24]

    Geunwoo Kim, Pierre Baldi, and Stephen McAleer. 2024. Language models can solve computer tasks.Advances in Neural Information Processing Systems36 (2024)

    2024

  17. [25]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems.Computer42, 8 (2009), 30–37

    2009

  18. [26]

    Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, et al . 2024. Training language models to self-correct via reinforcement learning.arXiv preprint arXiv:2409.12917(2024)

    2024 arXiv

  19. [27]

    Junlong Li, Shichao Sun, Weizhe Yuan, Run-Ze Fan, Hai Zhao, and Pengfei Liu

  20. [28]

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

    2019

  21. [29]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2023. Llara: Aligning large language models with sequential recommenders.arXiv preprint arXiv:2312.02445(2023)

    2023 arXiv

  22. [30]

    Xinyu Lin, Wenjie Wang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua

  23. [31]

    Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang

  24. [32]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al

  25. [33]

    Theo X Olausson, Jeevana Priya Inala, Chenglong Wang, Jianfeng Gao, and Armando Solar-Lezama. 2023. Is Self-Repair a Silver Bullet for Code Generation?. InThe Twelfth International Conference on Learning Representations

    2023

  26. [34]

    Is chatgpt a good recommender? a preliminary study.arXiv preprint arXiv:2304.10149(2023)

    2023 arXiv

  27. [35]

    Liangming Pan, Alon Albalak, Xinyi Wang, and William Yang Wang. 2023. Logic- lm: Empowering large language models with symbolic solvers for faithful logical reasoning.arXiv preprint arXiv:2305.12295(2023)

    2023 arXiv

  28. [36]

    Self-refine: Iterative refinement with self-feedback.Advances in Neural Information Processing Systems36 (2024)

    2024

  29. [37]

    Yuxiao Qu, Tianjun Zhang, Naman Garg, and Aviral Kumar. 2024. Recursive introspection: Teaching language model agents how to self-improve.arXiv preprint arXiv:2407.18219(2024)

    2024 arXiv

  30. [38]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in neural information processing systems35 (...

    2022

  31. [39]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test- time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314(2024)

    2024 arXiv

  32. [40]

    Debjit Paul, Mete Ismayilzada, Maxime Peyrard, Beatriz Borges, Antoine Bosselut, Robert West, and Boi Faltings. 2023. Refiner: Reasoning feedback on intermediate representations.arXiv preprint arXiv:2304.01904(2023)

    2023 arXiv

  33. [41]

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Car- los Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Stanford alpaca: an instruction-following llama model (2023).URL https://github. com/tatsu- lab/stanford_alpaca1, 9 (2023)

    2023

  34. [42]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems36 (2024)

    2024

  35. [43]

    Hanbing Wang, Xiaorui Liu, Wenqi Fan, Xiangyu Zhao, Venkataramana Kini, Devendra Yadav, Fei Wang, Zhen Wen, Jiliang Tang, and Hui Liu. 2024. Rethinking large language model architectures for sequential recommendations.arXiv preprint arXiv:2402.09543(2024)

    2024 arXiv

  36. [44]

    Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. InProceedings of the 28th ACM international conference on information and knowledge management....

    2019

  37. [45]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17. 1–7

    2017

  38. [46]

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

    2023 arXiv

  39. [47]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171(2022)

    2022 arXiv

  40. [48]

    Lei Wang and Ee-Peng Lim. 2023. Zero-shot next-item recommendation using large pretrained language models.arXiv preprint arXiv:2304.03153(2023)

    2023 arXiv

  41. [49]

    Zhenyu Wu, Qingkai Zeng, Zhihan Zhang, Zhaoxuan Tan, Chao Shen, and Meng Jiang. 2024. Large language models can self-correct with key condition verification. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 12846–12867

    2024

  42. [50]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable Tokenizer for LLM-based Genera- tive Recommendation.arXiv preprint arXiv:2405.07314(2024)

    2024

  43. [51]

    Zhiheng Xi, Dingwen Yang, Jixuan Huang, Jiafu Tang, Guanyu Li, Yiwen Ding, Wei He, Boyang Hong, Shihan Do, Wenyu Zhan, et al. 2024. Enhancing LLM Reasoning via Critique Models with Test-Time and Training-Time Supervision. arXiv preprint arXiv:2411.16579(2024)

    2024 arXiv

  44. [52]

    Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. 2022. Generating sequences by learning to self-correct. arXiv preprint arXiv:2211.00053(2022)

    2022 arXiv

  45. [53]

    Lanling Xu, Junjie Zhang, Bingqian Li, Jinpeng Wang, Mingchen Cai, Wayne Xin Zhao, and Ji-Rong Wen. 2024. Prompting large language models for recommender systems: A comprehensive framework and empirical analysis.arXiv preprint arXiv:2401.04997(2024)

    2024 arXiv

  46. [54]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender ...

    2024

  47. [55]

    Wenhao Yu, Zhihan Zhang, Zhenwen Liang, Meng Jiang, and Ashish Sabharwal

  48. [56]

    Yu Xia, Rui Zhong, Hao Gu, Wei Yang, Chi Lu, Peng Jiang, and Kun Gai. 2025. Hier- archical Tree Search-based User Lifelong Behavior Modeling on Large Language Model.arXiv preprint arXiv:2505.19505(2025)

    2025 arXiv

  49. [57]

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, et al. 2023. Recommender systems in the era of large language models (llms).arXiv preprint arXiv:2307.02046(2023)

    2023 arXiv

  50. [58]

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

    2024 arXiv

  51. [59]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 5941–5948

    2019

  52. [60]

    Improving language models via plug-and-play retrieval feedback.arXiv preprint arXiv:2305.14002(2023)

    2023 arXiv

  53. [61]

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Xin Zhao, Leyu Lin, and Ji-Rong Wen

  54. [62]

    Recommendation as instruction following: A large language model em- powered recommendation approach.ACM Transactions on Information Systems (2023)

    2023

  55. [64]

    Xin Zheng, Jie Lou, Boxi Cao, Xueru Wen, Yuqiu Ji, Hongyu Lin, Yaojie Lu, Xianpei Han, Debing Zhang, and Le Sun. 2024. Critic-cot: Boosting the reasoning abilities of large language model via chain-of-thoughts critic.arXiv preprint arXiv:2408.16326(2024)

    2024 arXiv

  56. [2023]

    Generative judge for evaluating alignment.arXiv preprint arXiv:2310.05470 (2023)

    2023 arXiv

  57. [2024]

    InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Bridging items and language: A transition paradigm for large language model-based recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1816–1826

Pith tools

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