Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

RevBrowse claims that a contrastively trained retriever (PrefRAG) that fetches only the pros and cons of candidate items matching a user's extracted likes and dislikes lets an LLM reranker beat strong sequential-recommendation baselines on

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 ↗

RevBrowse retrieves preference-relevant pros and cons from reviews via a contrastively trained module, then uses an LLM to rerank candidates; experiments on four Amazon datasets show consistent improvements over baselines.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful retrieval design and clean write-up, but a possible label-leakage in the evaluation pipeline makes the headline results unverified. the 3 major comments →

arxiv 2509.00698 v1 pith:OQTLTLS6 submitted 2025-08-31 cs.CL

Learning to Shop Like Humans: A Review-driven Retrieval-Augmented Recommendation Framework with LLMs

Kaiwen Wei , Jinpeng Gao , Jiang Zhong , Yuming Yang , Fengmao Lv , Zhenyang Li This is my paper
classification cs.CL
keywords retrieval-augmented generationsequential recommendationreview-based recommendationlarge language modelscontrastive learninguser preference modelingLLM rerankinginterpretability
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

RevBrowse claims that the best way to use reviews in LLM-based recommendation is not to stuff them into the prompt, but to imitate shoppers: first extract what a user likes and dislikes, then retrieve only the pros and cons from candidate-item reviews that match those preferences, and finally let an LLM rerank the candidates from that focused context. The central claim is that this retrieval step, a contrastively trained module called PrefRAG, lets an LLM recommender outperform strong sequential-recommendation baselines on four Amazon-2014 datasets, with the largest gains at the top of the ranking. A sympathetic reader would care because the approach addresses two real limits of LLM recommenders: reviews are long and exceed context windows, and most reviews are irrelevant to a given user's current decision. If the claim holds, review-based recommendation becomes more accurate, cheaper, and transparent, since the retrieved reviews are visible and can be inspected. The paper also reports ablations showing that both the global user-preference summary and the retrieved item-specific reviews contribute, and that two retrieved reviews per feature type are better than one or three.

Core claim

The paper's discovery, on its own terms, is that a retrieval-augmented module trained to match a user's stated preferences to item features, rather than to the user's own review text, makes LLM reranking both stronger and more interpretable. PrefRAG treats Like/Dislike as queries and Pros/Cons as answers; contrastive training pulls the user's own evaluation of an item close and pushes other users' evaluations of the same item away, which the paper argues teaches discrimination between preference-aligned and preference-misaligned review content. At inference, PrefRAG retrieves the top-K pros and cons for each candidate item from peer reviews and feeds them into an LLM recommender that ranks c

What carries the argument

PrefRAG is a dual-tower, contrastively trained retrieval module. It encodes user preferences (Like/Dislike) as queries and item features (Pros/Cons) as answers using the [EOS] embedding of an LLM, scores query-answer pairs by cosine similarity, and is trained with an InfoNCE loss over sliding-window samples: the positive answer is the pros/cons of the last item the user reviewed, and the negatives are pros/cons extracted from other users' reviews of the same item. At serve time, the frozen PrefRAG embeds the user's Like/Dislike and retrieves the top-K Pros/Cons for each candidate item, which are injected into a structured prompt alongside purchase history and a global preference summary; a L

Load-bearing premise

PrefRAG is trained with positive examples drawn from the target user's own review of an item, but at serve time it retrieves features written by other users about candidate items; the whole approach assumes that contrastively pushing away same-item peer reviews during training still teaches the retriever to select preference-relevant peer reviews for new items.

What would settle it

Hold out a set of users, extract their Like/Dislike, and retrieve top-K pros/cons from candidate items' peer reviews; then measure whether those retrieved texts are judged more aligned with the user's preferences than randomly sampled review features, using held-out future reviews or an independent judge. If serve-time retrieval accuracy is at chance, the downstream ranking gains cannot be attributed to preference alignment. A sharper version: for candidate items whose peer reviews strongly contradict the target user's stated likes, a preference-aligned retriever should suppress those features

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

If this is right

  • LLM recommenders can handle long review histories without truncation: retrieving top-2 pros/cons per candidate yields the best or near-best accuracy, so the bottleneck is relevance selection, not context size.
  • The transparent retrieval gives a built-in audit trail: because the retrieved pros/cons are literally the evidence in the prompt, a wrong recommendation can be traced to the specific review features that steered the model.
  • The framework is backbone-agnostic: swapping PrefRAG's and the recommender's LLM among Llama-2, Qwen, and SFR keeps the improvements over LlamaRec, suggesting the retrieval design rather than a particular model family drives the results.
  • The ablation shows that global preferences and item-specific retrieved features are complementary: removing either degrades performance and removing both degrades it most, so both should be part of review-driven LLM recommendation systems.
  • Sliding-window contrastive training is intended to track shifting tastes: recent reviews dominate the positive samples, so the retrieved features reflect the user's evolving preferences rather than a static summary.

Where Pith is reading between the lines

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

  • The same PrefRAG retriever could be repurposed as a user-facing review ranking tool: instead of 'most helpful' votes, an e-commerce platform could rank reviews by predicted match to each shopper's stated likes and dislikes, making the browse-then-decide behavior literal.
  • The reported experiments do not include an ablation that replaces PrefRAG with an untrained or random retriever; such an experiment would isolate whether the gains come from contrastive preference alignment or simply from exposing the LLM to any review-derived features.
  • Because training positives are the target user's own reviews while serve-time candidates come from peer reviews, the cleanest stress test is on idiosyncratic users whose tastes rarely appear in other people's reviews; if their retrieved features are noisy, a review-based fallback would be needed.
  • Since the gains are largest on MRR, PrefRAG mainly improves top-of-list ordering; replacing the LRU recall stage with a review-aware candidate generator could compound the benefit.
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 / 3 minor

Summary. The paper proposes RevBrowse, a review-driven retrieval-augmented framework for sequential recommendation with LLMs. It first uses a frozen LLM (Qwen2.5-72B-instruct) to extract structured user preferences (Like/Dislike) and item features (Pros/Cons) from review texts. A dual-tower retriever, PrefRAG, is trained with a contrastive InfoNCE loss to align user preference queries with item feature answers (Eq. 4). At inference, PrefRAG is frozen and retrieves the top-K pros/cons from each candidate item's review pool conditioned on the user's preferences; these features, together with the user history and preference summary, are fed into a LLaMA2-7B recommender tuned with LoRA to rank candidate items. Experiments on four Amazon-2014 datasets (Games, Food, Clothing, Sport) report consistent improvements over CF, review-based, and LLM baselines, with ablations on prompt components, retrieval top-K, retriever backbone, and feature integration strategy, plus a case study illustrating transparency.

Significance. If the reported results are valid, the paper makes a useful contribution: it offers a practical way to inject review-derived item signals into LLM recommenders under context-window constraints, and the PrefRAG retrieval process is transparent and interpretable. The framework is clearly motivated by the browse-then-decide analogy, and the authors provide a fairly comprehensive experimental suite, including multiple datasets, baselines, ablations, and backbone variants. The main value lies in demonstrating that structured, preference-conditioned retrieval of item features can improve LLM-based reranking over static summaries. However, the claimed significance depends on a data-construction audit: the paper does not state whether the user's own held-out review is excluded from the candidate item feature pool, and if it is excluded, the train/serve distribution shift of PrefRAG is not validated. The experimental comparisons also lack statistical support and contain an unresolved evaluation-protocol inconsistency.

major comments (3)
  1. [§4.2, §4.3] The paper does not state that the target user's held-out review is removed from the candidate item feature pool before retrieval. Section 3 defines R_i as the reviews of all users who interacted with item i; for the ground-truth next item, this includes the target user's own review. Section 4.3 builds pros/cons from the 'historical reviews of item k' for each candidate, with no exclusion rule. Under this description, the user's own held-out review is part of the feature pool for the correct item, so PrefRAG can retrieve the label's own pros/cons, directly inflating all downstream metrics. This also confounds the 'w/o Reviews' ablation (Table 7), since removing the retrieved reviews removes the leaked signal. If the user's own review is excluded instead, then PrefRAG is trained on positives that never occur at inference: the positive answers in §4.2 are the target user's own reviews, whil
  2. [§5.1.3 vs §4.3] There is an inconsistency in the evaluation protocol. Section 5.1.3 says 'During evaluation, predictions are ranked over the entire item set,' but Section 4.3 says LRU is used to obtain the initial candidate item set, and the recommendation prompt contains only those candidates. If RevBrowse ranks only the LRU candidate pool while baselines rank over the full item set, the reported R@k/N@k/M@k values are not comparable. If baselines also use LRU candidates, the sentence in §5.1.3 is inaccurate. The paper must specify the exact candidate generation procedure and verify that all methods are evaluated under the same protocol. This is load-bearing for the claimed consistent improvements in Tables 5 and 6.
  3. [Tables 5–7] The paper claims 'consistent and significant improvements' but reports only point estimates from single runs. Several improvements are very small — e.g., Games R@5 0.1219 vs Exp3rt 0.1188, Sport R@5 0.0369 vs 0.0353, Clothing R@5 0.0227 vs 0.0216 — and there are no standard deviations, confidence intervals, or paired significance tests across multiple seeds. The word 'significant' in the abstract and Section 5.2 is therefore unsupported. At minimum, the authors should add variance estimates and statistical tests (e.g., paired bootstrap or paired t-tests over user-level metrics) for the headline comparisons.
minor comments (3)
  1. [§2.1 / Appendix C] The baseline list contains an apparent duplicate: Exp3rt is cited as [16] in Related Work and as [17] in the baseline descriptions, and ZS-Ranker is listed with the same citation string as Exp3rt. Please reconcile the references and make the baseline descriptions accurate (e.g., LightGCN is a graph-based CF method, not a 'self-attentive sequential recommendation model' as written in Appendix C).
  2. [Throughout] There are several typos and inconsistent terms: 'dynamically shifts' (Abstract/Introduction), 'retrival' (near Eq. 10), 'Aamzon' in Figure 6 caption, and inconsistent use of 'pros/cons' vs 'Pros/Cons'. A careful language pass is needed.
  3. [§5.4 / Table 8] Figure 4 and Table 8 report the top-K study, but Figure 4 is not described with axis labels or error bars. Please clarify whether the plotted values are from single runs and make the figure self-contained.

Circularity Check

1 steps flagged

PrefRAG's retrieved features for the held-out item can be its own training positive: R_i includes the target user's review and Sec. 4.3 never excludes it, so the headline gains may reflect fitted-input retrieval rather than independent prediction.

specific steps
  1. fitted input called prediction [Sec. 3 (Preliminaries); Sec. 4.2 (Contrastive Learning Training Set Construction, Eq. 4); Sec. 4.3 (LLM-based Recommender, Eqs. 5-10)]
    ""each candidate item i∈I is associated with a set of reviews R_i written by users who have interacted with that item"; "the last review within the window... we extract its pros and cons to serve as the positive answer a+"; "we denote the sets of pros and cons extracted from the historical reviews of item k as pros_k and cons_k""

    By Sec. 3, R_i for the ground-truth next item includes the target user's own review. Sec. 4.2 makes that review's pros/cons the positive answer a+ for PrefRAG. Sec. 4.3 retrieves top-K pros/cons from 'historical reviews of item k' with no stated exclusion of the held-out review. Hence the retrieved features for the true item can be the exact a+ the model was trained to rank highest (Eq. 4): a fitted input, not a prediction. The paper never states the held-out review is excluded from R_i or from PrefRAG's training windows. If it is excluded, a retriever trained to prefer own-review positives over same-item peer negatives is being used to retrieve peer reviews, a distribution shift whose transfer is never validated. Either way the reported gains do not independently demonstrate preference-al

full rationale

The central derivation chain is otherwise self-contained: user/item feature extraction uses frozen LLMs, the recommender is trained on held-out next-token prediction, and comparisons are against external baselines. There are self-citations ([23],[40],[41]) but they are background NLP work and not load-bearing. The only circular step is the PrefRAG evaluation/retrieval construction: by the paper's own definitions, the positive training answer for a user is that user's own review of an item, and the item-side retrieval pool at inference is all reviews of the candidate item, which for the true next item includes that same user's review. The paper does not document excluding the held-out review; if it is present, the retriever's output for the true item is the training positive re-ranked, so the claimed 'retrieval-augmented' gains and the w/o Reviews ablation are not an independent test. If it is absent, the training objective (own review vs. peer reviews of the same item) does not match the inference task (retrieve peer reviews of unseen items), and no transfer evidence is given. This makes the headline claim partially circular/leaky, though not a mathematical derivation-equivalence. No uniqueness theorem or ansatz-smuggling via self-citation is present.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central result rests on the LLM-based extraction prompts, the contrastively trained retriever, and the LRU candidate generator. The main free parameters are the retriever's hyperparameters and the number of retrieved features. The methodological assumptions are domain-level (reviews are faithful, LLMs extract accurately, LRU covers the target) plus one ad hoc assumption about the transfer of contrastive training to the serving distribution.

free parameters (5)
  • top-K (number of retrieved item features) = 2
    Selected by comparing top-1, top-2, top-3 on validation across datasets; Section 5.4 and Table 8.
  • sliding window size = 20
    Used to construct contrastive training samples; set in Section 5.1.4.
  • positive-to-negative sample ratio = 1:40
    Set for PrefRAG contrastive training; Section 5.1.4.
  • history length = 20
    Number of past items included in the recommender prompt; Section 5.1.4.
  • learning rate = 1e-4
    For LoRA fine-tuning of the recommender; Section 5.1.4.
axioms (4)
  • domain assumption Review text faithfully reflects user preferences and item attributes.
    The framework extracts like/dislike and pros/cons from reviews and treats them as ground truth for preferences/features without validation. Invoked throughout Section 4.1.
  • domain assumption The extraction LLM (Qwen2.5-72B) produces accurate, faithful preference and feature summaries.
    No human evaluation of the extracted text is provided; the quality of the retrieved features depends entirely on this extraction. Section 4.1.
  • domain assumption The LRU recall model provides a candidate set that contains the ground-truth next item for most test instances.
    RevBrowse only reranks the LRU candidate pool; if the true item is not in the pool it cannot be predicted. Sections 4.3 and 5.1.3.
  • ad hoc to paper Contrastive training with positives from the user's own reviews and negatives from other users' reviews of the same item transfers to retrieval over other users' reviews of candidate items.
    This is the PrefRAG training/serving distribution shift. It is not analyzed or tested in the paper. Section 4.2 vs 4.3.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Shop Like Humans: A Review-driven Retrieval-Augmented Recommendation Framework with LLMs." pith.science (2026). https://pith.science/paper/OQTLTLS6

@misc{pith2026250900698,
  author       = {Pith},
  title        = {Pith review of: Learning to Shop Like Humans: A Review-driven Retrieval-Augmented Recommendation Framework with LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQTLTLS6}},
  note         = {Machine review of arXiv:2509.00698}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) have shown strong potential in recommendation tasks due to their strengths in language understanding, reasoning and knowledge integration. These capabilities are especially beneficial for review-based recommendation, which relies on semantically rich user-generated texts to reveal fine-grained user preferences and item attributes. However, effectively incorporating reviews into LLM-based recommendation remains challenging due to (1) inefficient to dynamically utilize user reviews under LLMs' constrained context windows, and (2) lacking effective mechanisms to prioritize reviews most relevant to the user's current decision context. To address these challenges, we propose RevBrowse, a review-driven recommendation framework inspired by the "browse-then-decide" decision process commonly observed in online user behavior. RevBrowse integrates user reviews into the LLM-based reranking process to enhance its ability to distinguish between candidate items. To improve the relevance and efficiency of review usage, we introduce PrefRAG, a retrieval-augmented module that disentangles user and item representations into structured forms and adaptively retrieves preference-relevant content conditioned on the target item. Extensive experiments on four Amazon review datasets demonstrate that RevBrowse achieves consistent and significant improvements over strong baselines, highlighting its generalizability and effectiveness in modeling dynamic user preferences. Furthermore, since the retrieval-augmented process is transparent, RevBrowse offers a certain level of interpretability by making visible which reviews influence the final recommendation.

Figures

Figures reproduced from arXiv: 2509.00698 by Fengmao Lv, Jiang Zhong, Jinpeng Gao, KaiWen Wei, Yuming Yang, Zhenyang Li.

Figure 1
Figure 1. Figure 1: Analogy between the human shopping process and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The details of the proposed PrefRAG module and RevBrowse framework: (a) An LLM extracts user preferences and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of (a) the dual-tower architecture and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance of different top-K strategies in Pre [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison of different backbones of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case study of RevBrowse on the Amazon Games [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The detailed case study [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: User Preference Extraction Case [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Item Feature Extraction Case [PITH_FULL_IMAGE:figures/full_fig_p015_10.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. Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems

    cs.IR 2026-07 accept novelty 5.0

    Agentic recommender systems are organized by agent role (assisted, as-recommender, as-simulator) crossed with autonomy levels L2–L5, yielding a roadmap of architectures, evaluation limits, and open challenges.

Reference graph

Works this paper leans on

64 extracted references · 33 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Hong-Kyun Bae, Hae-Ri Jang, Yang-Sae Moon, and Sang-Wook Kim. 2024. Item- Ranking Promotion in Recommender Systems. In Companion Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, Singapore, May 13-17, 2024 , Tat-Seng Chua, Chong-Wah Ngo, Roy Ka-Wei Lee, Ravi Kumar, and Hady W. Lauw (Eds.). ACM, 505–508. doi:10.1145/3589335.3651529

    work page arXiv 2024

  2. [2]

    George van den Driessche

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Ruther- ford, Katie Millican, and et al. George van den Driessche. 2022. Improving Language Models by Retrieving from Trillions of Tokens. In International Con- ference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA (Proceedings of Machine Learning Research, V...

    work page 2022

  3. [3]

    Chong Chen, Min Zhang, Yiqun Liu, and Shaoping Ma. 2018. Neural Attentional Rating Regression with Review-level Explanations. In Proceedings of the 2018 World Wide Web Conference on World Wide Web, WWW 2018, Lyon, France, April 23-27, 2018, Pierre-Antoine Champin, Fabien Gandon, Mounia Lalmas, and Panagiotis G. Ipeirotis (Eds.). ACM, 1583–1592. doi:10.114...

    arXiv 2018

  4. [4]

    Honglong Chen, Jinnan Fu, Lei Zhang, Shuai Wang, Kai Lin, Leyi Shi, and Lianhai Wang. 2019. Deformable Convolutional Matrix Factorization for Document Context-Aware Recommendation in Social Networks.IEEE Access 7 (2019), 66347– 66357. doi:10.1109/ACCESS.2019.2917257

    work page arXiv 2019

  5. [5]

    Kankanhalli

    Zhiyong Cheng, Ying Ding, Xiangnan He, Lei Zhu, Xuemeng Song, and Mohan S. Kankanhalli. 2018. Aˆ3NCF: An Adaptive Aspect Attention Model for Rating Prediction. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden , Jérôme Lang (Ed.). ijcai.org, 3748–3754. doi:10.249...

    work page doi:10.24963/ijcai.2018/521 2018

  6. [6]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A Survey on RAG Meeting LLMs: To- wards Retrieval-Augmented Large Language Models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024 , Ricardo Baeza-Yates and Fran...

    arXiv 2024

  7. [7]

    Xinyu Guan, Zhiyong Cheng, Xiangnan He, Yongfeng Zhang, Zhibo Zhu, Qinke Peng, and Tat-Seng Chua. 2019. Attentive Aspect Modeling for Review-Aware Recommendation. ACM Trans. Inf. Syst. 37, 3 (2019), 28:1–28:27. doi:10.1145/ 3309546

    work page 2019

  8. [8]

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang

  9. [9]

    Emrul Hasan, Mizanur Rahman, Chen Ding, Jimmy Xiangji Huang, and Shaina Raza. 2024. Review-based Recommender Systems: A Survey of Approaches, Challenges and Future Perspectives. CoRR abs/2405.05562 (2024). doi:10.48550/ ARXIV.2405.05562 arXiv:2405.05562

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.05562 2024

  10. [10]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the 43rd International ACM SIGIR con- ference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020 , Jimmy X. Huang, Yi Chan...

    arXiv 2020

  11. [11]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net. https://openreview.net/forum?id=nZeVKeeFYf9

    work page 2022

  12. [12]

    Gautier Izacard and Edouard Grave. 2021. Distilling Knowledge from Reader to Retriever for Question Answering. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net. https://openreview.net/forum?id=NTEz-6wysdb

    work page 2021

  13. [13]

    Gautier Izacard and Edouard Grave. 2021. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021 , Paola Merlo, Jörg Tiedemann, and Reut Tsarfaty (Eds.). Associa...

    doi:10.18653/v1/2021.eacl-main.74 2021

  14. [14]

    Gawesh Jawaheer, Peter Weller, and Patty Kostkova. 2014. Modeling User Pref- erences in Recommender Systems: A Classification Framework for Explicit and Implicit User Feedback. ACM Trans. Interact. Intell. Syst. 4, 2 (2014), 8:1–8:26. doi:10.1145/2512208

    work page doi:10.1145/2512208 2014

  15. [15]

    Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Rec- ommendation. In IEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018 . IEEE Computer Society, 197–206. doi:10.1109/ Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al. ICDM.2018.00035

    arXiv 2018

  16. [16]

    Jieyong Kim, Hyunseo Kim, Hyunjin Cho, SeongKu Kang, Buru Chang, Jinyoung Yeo, and Dongha Lee. 2024. Review-driven Personalized Preference Reasoning with Large Language Models for Recommendation. CoRR abs/2408.06276 (2024). doi:10.48550/ARXIV.2408.06276 arXiv:2408.06276

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.06276 2024

  17. [17]

    Jieyong Kim, Hyunseo Kim, Hyunjin Cho, SeongKu Kang, Buru Chang, Jinyoung Yeo, and Dongha Lee. 2025. Review-driven Personalized Preference Reasoning with Large Language Models for Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2025, Padua, Italy, July 13-18, 2025 , ...

    arXiv 2025

  18. [18]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural In- formation Processing Systems 33: Annual Conference on Neural Informat...

    work page 2020

  19. [19]

    Chi, and Minmin Chen

    Pan Li, Yuyan Wang, Ed H. Chi, and Minmin Chen. 2023. Prompt Tuning Large Language Models on Personalized Aspect Extraction for Recommendations.CoRR abs/2306.01475 (2023). doi:10.48550/ARXIV.2306.01475 arXiv:2306.01475

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.01475 2023

  20. [20]

    Yuhan Li, Xinni Zhang, Linhao Luo, Heng Chang, Yuxiang Ren, Irwin King, and Jia Li. 2025. G-Refer: Graph Retrieval-Augmented Large Language Model for Explainable Recommendation. In Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 2025 , Guodong Long, Michale Blumestein, Yi Chang, Liane Lewin-Eytan, Zi H...

    arXiv 2025

  21. [21]

    Huiting Liu, Yi Chen, Pei-Pei Li, Peng Zhao, and Xindong Wu. 2023. Enhancing review-based user representation on learned social graph for recommendation. Knowl. Based Syst. 266 (2023), 110438. doi:10.1016/J.KNOSYS.2023.110438

    work page arXiv 2023

  22. [22]

    Junling Liu, Chao Liu, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is ChatGPT a Good Recommender? A Preliminary Study. CoRR abs/2304.10149 (2023). doi:10. 48550/ARXIV.2304.10149 arXiv:2304.10149

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.10149 2023

  23. [23]

    Nayu Liu, Kaiwen Wei, Yong Yang, Jianhua Tao, Xian Sun, Fanglong Yao, Hongfeng Yu, Li Jin, Zhao Lv, and Cunhang Fan. 2024. Multimodal Cross-Lingual Summarization for Videos: A Revisit in Knowledge Distillation Induced Triple- Stage Training Method. IEEE Trans. Pattern Anal. Mach. Intell. 46, 12 (2024), 10697–10714. doi:10.1109/TPAMI.2024.3447778

    work page arXiv 2024

  24. [24]

    McAuley and Jure Leskovec

    Julian J. McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. InSeventh ACM Conference on Recommender Systems, RecSys ’13, Hong Kong, China, October 12-16, 2013 , Qiang Yang, Irwin King, Qing Li, Pearl Pu, and George Karypis (Eds.). ACM, 165–172. doi:10.1145/2507157.2507163

    work page arXiv 2013

  25. [25]

    Shervin Minaee, Tomás Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. 2024. Large Language Mod- els: A Survey. CoRR abs/2402.06196 (2024). doi:10.48550/ARXIV.2402.06196 arXiv:2402.06196

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.06196 2024

  26. [26]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. In-Context Retrieval-Augmented Language Models. Trans. Assoc. Comput. Linguistics 11 (2023), 1316–1331. doi:10.1162/TACL_A_00605

    doi:10.1162/tacl_a_00605 2023

  27. [27]

    Shafiq Rayhan Joty Caiming Xiong Yingbo Zhou Semih Yavuz Rui Meng, Ye Liu

    work page

  28. [28]

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. REPLUG: Retrieval-Augmented Black-Box Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers), NAACL...

    2024

  29. [29]

    Jie Shuai, Kun Zhang, Le Wu, Peijie Sun, Richang Hong, Meng Wang, and Yong Li

    work page

  30. [30]

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

  31. [31]

    Huang, Xiaohui Yu, and Bushra Aljbawi

    Omer Tal, Yang Liu, Jimmy X. Huang, Xiaohui Yu, and Bushra Aljbawi. 2021. Neural Attention Frameworks for Explainable Recommendation. IEEE Trans. Knowl. Data Eng. 33, 5 (2021), 2137–2150. doi:10.1109/TKDE.2019.2953157

    work page arXiv 2021

  32. [32]

    Yunzhi Tan, Min Zhang, Yiqun Liu, and Shaoping Ma. 2016. Rating-Boosted Latent Topics: Understanding Users and Items with Ratings and Reviews. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intel- ligence, IJCAI 2016, New York, NY, USA, 9-15 July 2016 , Subbarao Kambhampati (Ed.). IJCAI/AAAI Press, 2640–2646. http://www.ij...

    work page 2016

  33. [33]

    Qwen Team. 2024. Qwen2.5: A Party of Foundation Models. https://qwenlm. github.io/blog/qwen2.5/

    2024

  34. [34]

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

    Pith/arXiv arXiv 2023

  35. [35]

    Yasmine Babaei

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, and et al. Yasmine Babaei. 2023. Llama 2: Open Foundation and Fine-Tuned Chat Mod- els. CoRR abs/2307.09288 (2023). doi:10.48550/ARXIV.2307.09288 arXiv:2307.09288

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.09288 2023

  36. [36]

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. CoRR abs/1807.03748 (2018). arXiv:1807.03748 http://arxiv.org/abs/1807.03748

    Pith/arXiv arXiv 2018

  37. [37]

    Shijie Wang, Wenqi Fan, Yue Feng, Xinyu Ma, Shuaiqiang Wang, and Dawei Yin. 2025. Knowledge Graph Retrieval-Augmented Generation for LLM-based Recommendation. CoRR abs/2501.02226 (2025). doi:10.48550/ARXIV.2501.02226 arXiv:2501.02226

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.02226 2025

  38. [38]

    Yile Wang, Peng Li, Maosong Sun, and Yang Liu. 2023. Self-Knowledge Guided Retrieval Augmentation for Large Language Models. In Findings of the Associa- tion for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023 , Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, 10303–10315. doi:10.18653/V1/2...

    doi:10.18653/v1/2023.findings-emnlp.691 2023

  39. [39]

    Chuyuan Wei, Ke Duan, Shengda Zhuo, Hongchun Wang, Shuqiang Huang, and Jie Liu. 2025. Enhanced Recommendation Systems with Retrieval-Augmented Large Language Model. J. Artif. Intell. Res. 82 (2025), 1147–1173. doi:10.1613/ JAIR.1.17809

    work page 2025

  40. [40]

    Kaiwen Wei, Xian Sun, Zequn Zhang, Jingyuan Zhang, Zhi Guo, and Li Jin

    work page

  41. [41]

    Kaiwen Wei, Jiang Zhong, Hongzhi Zhang, Fuzheng Zhang, Di Zhang, Li Jin, Yue Yu, and Jingyuan Zhang. 2025. Chain-of-Specificity: Enhancing Task-Specific Constraint Adherence in Large Language Models. In Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025 , Owen Rambow, Leo Wanner,...

    work page 2025

  42. [42]

    Yibiao Wei, Yang Xu, Lei Zhu, Jingwei Ma, and Chengmei Peng. 2024. Multi-level cross-modal contrastive learning for review-aware recommendation. Expert Syst. Appl. 247 (2024), 123341. doi:10.1016/J.ESWA.2024.123341

    work page arXiv 2024

  43. [43]

    Junda Wu, Cheng-Chun Chang, Tong Yu, Zhankui He, Jianing Wang, Yupeng Hou, and Julian J. McAuley. 2024. CoRAL: Collaborative Retrieval-Augmented Large Language Models Improve Long-tail Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024 , Ricardo Baeza-Yate...

    arXiv 2024

  44. [44]

    Wu-Dong Xi, Ling Huang, Chang-Dong Wang, Yin-Yu Zheng, and Jian-Huang Lai. 2022. Deep Rating and Review Neural Network for Item Recommendation. IEEE Trans. Neural Networks Learn. Syst. 33, 11 (2022), 6726–6736. doi:10.1109/ TNNLS.2021.3083264

    work page arXiv 2022

  45. [45]

    Tong Xiao and Jingbo Zhu. 2025. Foundations of Large Language Models. CoRR abs/2501.09223 (2025). doi:10.48550/ARXIV.2501.09223 arXiv:2501.09223

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.09223 2025

  46. [46]

    Mengyuan Yang, Mengying Zhu, Yan Wang, Linxun Chen, Yilei Zhao, Xiuyuan Wang, Bing Han, Xiaolin Zheng, and Jianwei Yin. 2024. Fine-Tuning Large Language Model Based Explainable Recommendation with Explainable Quality Reward. In AAAI Conference on Artificial Intelligence . AAAI Press, 9250–9259

    work page 2024

  47. [47]

    Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira, Dong Wang, and Even Oldridge. 2023. LlamaRec: Two-Stage Recommendation using Large Language Models for Ranking. CoRR abs/2311.02089 (2023). doi:10.48550/ARXIV.2311.02089 arXiv:2311.02089

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.02089 2023

  48. [48]

    Linear Recurrent Units for Sequential Recommendation

    Zhenrui Yue, Yueqi Wang, Zhankui He, Huimin Zeng, Julian J. McAuley, and Dong Wang. 2023. Linear Recurrent Units for Sequential Recommendation.CoRR abs/2310.02367 (2023). doi:10.48550/ARXIV.2310.02367 arXiv:2310.02367

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.02367 2023

  49. [49]

    Kai Zhang, Hao Qian, Qi Liu, Zhiqiang Zhang, Jun Zhou, Jianhui Ma, and Enhong Chen. 2021. SIFN: A Sentiment-aware Interactive Fusion Network for Review- based Item Recommendation. InCIKM ’21: The 30th ACM International Conference Learning to Shop Like Humans: A Review-driven Retrieval-Augmented Recommendation Framework with LLMs Conference acronym ’XX, Ju...

    2021

  50. [50]

    Lei Zheng, Vahid Noroozi, and Philip S. Yu. 2017. Joint Deep Modeling of Users and Items Using Reviews for Recommendation. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining, WSDM 2017, Cambridge, United Kingdom, February 6-10, 2017 , Maarten de Rijke, Milad Shokouhi, Andrew Tomkins, and Min Zhang (Eds.). ACM, 425–434. ...

    arXiv 2017

  51. [51]

    Zhi Zheng, Wenshuo Chao, Zhaopeng Qiu, Hengshu Zhu, and Hui Xiong. 2024. Harnessing Large Language Models for Text-Rich Sequential Recommendation. In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024, Tat-Seng Chua, Chong-Wah Ngo, Ravi Kumar, Hady W. Lauw, and Roy Ka-Wei Lee (Eds.). ACM, 3207–3216. doi:10.1145/3589334.364...

    arXiv 2024

  52. [57]

    Reflect general likes and dislikes, not specific brands or items

    work page

  53. [59]

    Exclude mentions of delivery time or pricing

    work page

  54. [60]

    Like": [

    Be concise and simple. Output format: { "Like": ["..."], "Dislike": ["..."] } Table 9: The Detailed User Preference Extraction Prompt. Item Features Extraction Prompt Template Instruction: Given user reviews of purchased items in JSON format, extract high-level item properties from the comments. Input: Item review: {review} Response: A JSON object with tw...

    work page

  55. [61]

    Summarize general strengths and weaknesses of the items

    work page

  56. [62]

    Be derived from the content of the reviews

    work page

  57. [63]

    Avoid mentioning specific brands or item names

    work page

  58. [64]

    Exclude any comments related to delivery time or pricing

    work page

  59. [65]

    Pros": [

    Be simple, short, and concise. Output format: { "Pros": ["..."], "Cons": ["..."] } Table 10: The Detailed Item Features Extraction Prompt. B Detailed Ablation Study Results We conduct extensive ablation study on the top-K strategies and dif- ferent backbones of PrefRAG cross different datasets. The detailed results are shown in Table 8 and Table 11. C Bas...

    work page 2018

  60. [2019]

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformer. In Proceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019, Wenwu Zhu, Dacheng Tao, Xueqi Cheng, Peng Cui, Elke A. Rundensteiner, David Carmel, Qi He, and Jeffrey Xu Yu (Ed...

    arXiv 2019

  61. [2020]

    CoRR abs/2002.08909 (2020)

    REALM: Retrieval-Augmented Language Model Pre-Training. CoRR abs/2002.08909 (2020). arXiv:2002.08909 https://arxiv.org/abs/2002.08909

    Pith/arXiv arXiv 2002

  62. [2021]

    Trigger is Not Sufficient: Exploiting Frame-aware Knowledge for Im- plicit Event Argument Extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021 , Chengqing Zo...

    doi:10.18653/v1/2021.acl-long.360 2021

  63. [2022]

    A Review-aware Graph Contrastive Learning Framework for Recommen- dation. In SIGIR ’22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 - 15, 2022 , Enrique Amigó, Pablo Castells, Julio Gonzalo, Ben Carterette, J. Shane Culpepper, and Gabriella Kazai (Eds.). ACM, 1283–1293. doi:10.11...

    arXiv 2022

  64. [2024]

    Salesforce AI Research Blog

    SFR-Embedding-Mistral:Enhance Text Retrieval with Transfer Learning. Salesforce AI Research Blog. https://www.salesforce.com/blog/sfr-embedding/

    work page

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