ARAG: Agentic Retrieval Augmented Generation for Personalized Recommendation

Aysenur Inan; Jason Cho; Jianpeng Xu; Kai Zhao; Kehui Yao; Pratheek Vadla; Praveen Kanumala; Priyank Gupta; Reza Yousefi Maragheh; Sushant Kumar

REVIEW 4 major objections 4 minor 3 cited by

ARAG: Agentic Retrieval Augmented Generation for Personalized Recommendation

T0 review · 4 major / 4 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 ARAG shows that four cooperating LLM agents, coordinated through a shared memory, can refine an embedding-based candidate recall set into a better-ranked, context-aware recommendation list.

desk verdict A sensible four-agent LLM ranking pipeline with large reported gains, but the evaluation is under-specified enough that the headline percentages need more support before I'd trust them. read the letter →

arxiv 2506.21931 v2 pith:MBWMF4UJ submitted 2025-06-27 cs.IR cs.AIcs.CLcs.MA

Reza Yousefi Maragheh , Pratheek Vadla , Priyank Gupta , Kai Zhao , Aysenur Inan , Kehui Yao , Jianpeng Xu , Praveen Kanumala

show 2 more authors

Jason Cho Sushant Kumar

This is my paper

classification cs.IRcs.AIcs.CLcs.MA

keywords AgenticRAGMulti-agentLLMsystemPersonalizedrecommendationNaturallanguageinferenceRetrieval-augmentedgenerationRankingAmazonReviewdatasetBlackboardarchitecture

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

ARAG is a multi-agent retrieval-augmented generation framework for personalized recommendation. The paper tries to establish that replacing the static retrieval scores of a standard RAG pipeline with a cooperative workflow of four LLM agents—a User Understanding Agent, a Natural Language Inference Agent, a Context Summary Agent, and an Item Ranker Agent—produces recommendation rankings that are substantially more accurate than recency-based prompting and vanilla embedding-based RAG. On 10,000 randomly sampled Amazon users across Clothing, Electronics, and Home, the authors report up to 42.1% relative improvement in NDCG@5 and 35.5% in Hit@5 over the best baseline. The broader claim is that agentic orchestration inside the RAG loop turns a coarse recall set into a semantically grounded, context-aware ranked list, and that this is a practical route to more personalized conversational recommendation.

What carries the argument

The load-bearing mechanism is a blackboard-style multi-agent collaboration protocol: all agents read from and write to a shared structured memory $\mathcal{B}$ with JSON messages $\{id, role, content, score, timestamp\}$. An NLI agent assigns each retrieved item an alignment score $s_{\text{NLI}}(i, u)$, items above a threshold $\theta$ form $I^{+}$, and a Context Summary Agent summarizes their metadata; in parallel, a User Understanding Agent produces a preference summary $S_{\text{user}}$. The Item Ranker Agent then fuses $S_{\text{user}}$ and $S_{\text{ctx}}$ to produce the final permutation $\pi$. These agents act as a memory moderation scheme that integrates long-term and session behavior into the final ranking.

What would settle it

Re-run the same experiment with a strictly temporal split where the ground-truth item is the next interaction after the session context, and verify that the ground-truth item lies inside the initial top-$k$ retrieval set; if the reported NDCG@5 and Hit@5 gains over vanilla RAG shrink substantially, or if the gains are reproduced with a random split, the central claim is called into question.

Watch

Extended reading notes

Core claim

The central claim is that the sequential and parallel reasoning of four specialized LLM agents, coordinated through a shared blackboard memory, can refine an embedding-based candidate recall set into a better-ranked final list than either recency-based ranking or vanilla RAG. The User Understanding Agent summarizes long-term and session preferences; the NLI Agent scores each candidate's metadata for alignment with inferred user intent; the Context Summary Agent condenses the accepted evidence; and the Item Ranker Agent fuses these signals to output a ranked list. The paper reports consistent gains across three Amazon categories, with the largest in Clothing (42.12% NDCG@5, 35.54% Hit@5), and the ablation shows that removing the NLI and Context Summary agents lowers performance, confirming that semantic reasoning adds value beyond user summarization alone.

Load-bearing premise

The evaluation assumes that the 10,000 sampled users' interactions are split into long-term context, session context, and ground truth in a temporally valid way, so that the ground-truth item is the user's next purchase and never appears in the provided contexts; the paper never states this split protocol.

Editorial extensions

If this is right

If the reported gains hold, agentic reasoning inside the RAG loop offers a practical recipe for conversational recommendation systems that already rely on embedding retrieval.
The ablation results imply that semantic alignment filtering (NLI plus context summarization) contributes the largest portion of the accuracy gain, suggesting that simple user summaries alone are not enough.
The domain-dependent gains (largest in Clothing) indicate that the value of agentic refinement varies with item attribute complexity and user preference diversity.
The blackboard collaboration protocol provides a transparent explanation trace, which the paper argues can improve user trust and interpretability of recommendations.

Reading between the lines

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

The reported improvement may partly reflect the pass-through of a more verbose, intent-focused prompt to the LLM rather than the NLI reasoning per se; a prompt-only control would separate these effects.
Because the data split protocol is not stated, the numbers should be treated as upper-bound estimates until a temporal split is confirmed; if the split is random or contains leakage, gains could shrink or vanish.
A testable extension is to vary the LLM backbone and retrieval depth $k$: the framework's sensitivity to the threshold $\theta$ and to the cost of four sequential LLM calls is not measured in the paper.
The framework could be applied to session-based or cold-start settings where the User Understanding Agent's summarization would provide an enriched profile from sparse interactions.

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.

Desk Editor's Note

read the letter

Quick take: ARAG is a clean four-agent decomposition of RAG for recommendation, and the reported gains are large enough to be worth a look. But the evaluation is under-specified to a degree that the headline numbers cannot be independently verified as written. The data split, the retrieval recall, and the exact protocol are missing. That is fixable in a revision, and the architecture is sensible, so I would still send it to review.

What is actually new is the orchestration. The authors combine four standard LLM tasks—user summarization, NLI filtering, context summarization, and ranking—into a blackboard-style protocol where each agent reads from and writes to a shared structured memory. That is a reasonable engineering contribution, especially for an industry team. The ablation study is genuinely useful: it shows the NLI and context summary agents each add measurable value, and the gains are consistent across the three Amazon categories. Using gpt-3.5-turbo with temperature 0 is also a sensible reproducibility choice.

The soft spots are real and load-bearing. Section 3.1 simply says 10,000 users were sampled; it never describes how interactions are partitioned into long-term context, session context, and ground truth. Without a temporal split, leakage cannot be ruled out, and no leakage check is offered. There is no recall@k for the initial embedding retrieval, so if the ground-truth item is not guaranteed to be inside the top-k set, the reported NDCG@5 and Hit@5 are conditional on retrieval success in ways the authors do not discuss. There are no error bars, no significance tests, and only two baselines—recency and vanilla RAG. No comparison to other agent-based recommendation systems either, which makes the novelty claim hard to calibrate. These are all fixable, but as written the 42% improvement claim is not independently verifiable.

The paper is honest in its presentation; I see no circular fitting or invented metrics. The free parameters (NLI threshold, retrieval k, ranking pool r) are not analyzed, but that is a completeness issue rather than a fatal one.

Bottom line: this is a useful industry paper for people building LLM-based recommenders. It deserves a serious referee, and the referee should demand recall@k, a full description of the data split, and ideally code or prompts before the numbers can be trusted.

Referee Report

4 major / 4 minor

Summary. The paper introduces ARAG, an agentic retrieval-augmented generation framework for personalized recommendation. The system uses four LLM-based agents — a User Understanding Agent, an NLI Agent, a Context Summary Agent, and an Item Ranker Agent — to refine an embedding-based recall set into a final ranked list. The method is evaluated on three Amazon Review categories (Clothing, Electronics, Home) and compared against recency-based ranking and Vanilla RAG. The paper reports NDCG@5 and Hit@5 improvements of up to 42.1% and 35.5%, respectively, and includes an ablation study that attributes gains to the individual agents. The central claim is that multi-agent orchestration inside the RAG loop substantially improves recommendation accuracy.

Significance. If the reported results are sound, ARAG is a useful contribution to the emerging area of agentic RAG for recommendation. The framework is clearly described and the ablation study provides initial evidence of the incremental contribution of each agent. The work also offers a plausible mechanism for improving ranking by injecting semantic reasoning and user-intent summarization into the retrieval-augmented pipeline. However, the significance is currently limited by the under-specified evaluation protocol and the absence of any statistical support for the claimed improvements. The manuscript is a reasonable first step but needs stronger empirical validation before the headline numbers can be taken at face value.

major comments (4)

[§3.1, §3.3, Table 1] The evaluation protocol is under-specified: Section 3.1 states only that 10,000 users were randomly sampled from the Amazon Review dataset, but it does not describe how each user's interactions are partitioned into long-term context, session context, and ground-truth items. For the reported NDCG@5 and Hit@5 improvements to be valid, the ground-truth item must be temporally after the provided contexts and must be excluded from those contexts. The absence of a leakage check and of any statement about chronological splits makes the central claim in Section 3.3 currently unverifiable. Please provide the exact splitting procedure, including any filtering rules, and verify that no ground-truth item appears in the context.
[§3.2, §3.3, Table 1] The paper reports single-point estimates with no standard deviations, confidence intervals, or significance tests. The phrase 'significantly outperforms' in Section 3.3 is therefore not supported by any statistical evidence. The improvements of 22–42% could arise from a few favorable instances without robust underlying differences. Please report multiple runs or bootstrap replicates and a paired significance test (e.g., paired bootstrap or randomization test) for the headline improvements.
[§2.2.1, Table 1] The paper never reports recall@k for the initial embedding retrieval I0. If the ground-truth item is not guaranteed to be in I0, the reported Hit@5/NDCG@5 are conditional on retrieval success. In that case, the improvement percentages may reflect differences in what the baselines can see (i.e., retrieval conditioning) rather than the agentic ranking stage. Please report recall@k for the initial retrieval and, if the ground-truth item is not always in I0, report conditional metrics or state explicitly that the ground-truth item is guaranteed to be in the recall set.
[§2.2.5] The formal definition of the ranker output is inconsistent: the text says the Item Ranker Agent ranks 'r candidate items', but the equation defines a permutation π over the full item set of size N. The paper never defines r or its relationship to |I0| and |I+|. Please define r (the number of candidates passed to the Item Ranker Agent) and clarify how it relates to the initial recall set I0 and the NLI-filtered set I+.

minor comments (4)

[§3.3 and Table 1] The '% Improvement' row in Table 1 is computed relative to different baselines across datasets (Recency-based Ranking for Clothing, Vanilla RAG for Electronics and Home). The text in Section 3.3 mentions 'the next best performing approach' but does not make explicit that the baseline varies. Please state the reference baseline for each improvement percentage in the table caption or the main text.
[Abstract and §2] There are multiple typos and grammatical errors, including 'accross' (Abstract), 'We also, conduct' (Abstract), 'the applying a regular RAG for retrieving the an initial set of larger recall set of items' (§2), and 'Along-term context' (§2.1). These should be corrected.
[§2.2.3 and §2.2.4] The NLI acceptance threshold θ is defined in Section 2.2.3 but no value is given and no sensitivity analysis is provided. Since this is a free parameter that controls the size of I+, please report the chosen value and, if possible, the effect of varying it.
[§3.4 and Table 1] The ablation row labels 'ARAG w/o NLI & CSA' and 'ARAG w/o NLI' are not explained in the text. Please clarify which components are present in each ablation condition; for example, indicate that 'ARAG w/o NLI & CSA' still includes the User Understanding Agent and the Item Ranker Agent.

Circularity Check

0 steps flagged · score 0.0 of 10

No internal circularity: the agentic pipeline is an external LLM ranking evaluated on benchmarks, not a quantity fitted to its own outputs; self-citations are peripheral and non-load-bearing.

full rationale

The paper's derivation chain is a self-contained inference pipeline rather than a fitted identity. Equations (1)-(6) define an embedding recall set I0, an NLI alignment score, a filtered/summarized context, a user summary, and a final LLM ranking; no equation defines the final permutation as equal to the retrieval scores, the NLI scores, or any parameter fitted to NDCG@5/Hit@5. The headline improvements are external evaluation outcomes on Amazon datasets, not quantities reconstructed from the inputs by construction. The self-citations in the introduction (refs [9], [16], [17]) appear in a related-work sentence about ranking desiderata and are not load-bearing premises, uniqueness theorems, or ansatz-forcing citations. The main weakness is the under-specified evaluation protocol: Section 3.1 does not state the temporal split, whether the ground-truth item is excluded from the provided contexts, whether it is guaranteed to be inside the initial retrieval set I0, or whether significance testing was performed. That is a reproducibility and correctness concern, but it is not internal circularity: even a leaked or retrieval-conditional evaluation would be an experimental artifact, not an equation-level equivalence between the method's outputs and its inputs. Accordingly, no circular step is identified.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The four agents are software components, not postulated physical or mathematical entities. They are not new theoretical objects requiring independent falsifiable handles, so no invented entities are recorded.

free parameters (3)

NLI acceptance threshold (theta) = not reported
In Section 2.2.3, I+ is defined as items with s_NLI(i,u) >= theta; theta is a hand-chosen threshold that controls how many candidates reach the Context Summary Agent. Its value is not given, so the behavior of the pipeline is underspecified.
Initial recall set size (k) = not reported
In Section 2.2.1, I0 = argtop_k of embedding similarity. k determines the candidate pool for all downstream agents; the paper never states k.
Number of candidates ranked (r) = not reported
The Item Ranker returns a permutation over r candidates, but r is not specified; it may equal the size of I+ or a fixed budget.

assumptions (5)

domain assumption Cosine similarity between learned embeddings of items and user context is a meaningful relevance signal for initial retrieval.
Section 2.2.1; the whole pipeline starts from this retrieval, so if the embedding similarity is unrelated to purchase likelihood, the recall set would miss relevant items and no agent could fix it.
domain assumption Textual metadata (title, description, reviews) contains enough signal for an LLM to infer alignment with user preferences.
Section 2.2.2; NLI only sees T(i), not collaborative signals, so this assumption is load-bearing for the semantic alignment step.
domain assumption Amazon review interactions are appropriate implicit feedback for evaluating next-item prediction.
Section 3.1; users' past purchases and reviews are used as ground truth without explicit statement of how ground-truth items are defined (for example, next purchase or held-out interaction).
domain assumption A temporal split with no overlap between context and ground truth is implied but never stated.
Section 3.1 describes only random sampling of users, not how interactions are split; the evaluation only makes sense if ground truth is temporally after the provided contexts. This is an unverified assumption.
domain assumption LLM outputs at temperature 0 are deterministic enough to treat the reported numbers as stable.
Section 3.2; temperature 0 is set, but the paper reports single runs without variance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ARAG: Agentic Retrieval Augmented Generation for Personalized Recommendation." pith.science (2026). https://pith.science/paper/MBWMF4UJ

@misc{pith2026250621931,
  author       = {Pith},
  title        = {Pith review of: ARAG: Agentic Retrieval Augmented Generation for Personalized Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBWMF4UJ}},
  note         = {Machine review of arXiv:2506.21931}
}

read the original abstract

Retrieval-Augmented Generation (RAG) has shown promise in enhancing recommendation systems by incorporating external context into large language model prompts. However, existing RAG-based approaches often rely on static retrieval heuristics and fail to capture nuanced user preferences in dynamic recommendation scenarios. In this work, we introduce ARAG, an Agentic Retrieval-Augmented Generation framework for Personalized Recommendation, which integrates a multi-agent collaboration mechanism into the RAG pipeline. To better understand the long-term and session behavior of the user, ARAG leverages four specialized LLM-based agents: a User Understanding Agent that summarizes user preferences from long-term and session contexts, a Natural Language Inference (NLI) Agent that evaluates semantic alignment between candidate items retrieved by RAG and inferred intent, a context summary agent that summarizes the findings of NLI agent, and an Item Ranker Agent that generates a ranked list of recommendations based on contextual fit. We evaluate ARAG accross three datasets. Experimental results demonstrate that ARAG significantly outperforms standard RAG and recency-based baselines, achieving up to 42.1% improvement in NDCG@5 and 35.5% in Hit@5. We also, conduct an ablation study to analyse the effect by different components of ARAG. Our findings highlight the effectiveness of integrating agentic reasoning into retrieval-augmented recommendation and provide new directions for LLM-based personalization.

Figures

Figures reproduced from arXiv: 2506.21931 by the authors.

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

The Future is Agentic: Definitions, Perspectives, and Open Challenges of Multi-Agent Recommender Systems
cs.IR 2025-07 conditional novelty 6.0 of 10

A framework for agentic recommender systems plus a pilot study showing multi-agent pipelines beat a single-shot LLM only on high-diversity user histories.
Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems
cs.IR 2026-07 accept novelty 5.0 of 10

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.
Collaborative Inference and Learning between Edge SLMs and Cloud LLMs: A Survey of Algorithms, Execution, and Open Challenges
cs.DC 2025-07 conditional novelty 4.0 of 10

A survey that builds a taxonomy of edge-cloud LLM-SLM collaboration for inference and training, claiming to be the first to unify both phases.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages · cited by 3 Pith papers

[1]

Justin Chih-Yao Chen, Archiki Prasad, Swarnadeep Saha, Elias Stengel-Eskin, and Mohit Bansal. 2024. MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning. arXiv preprint arXiv:2409.12147 (2024)

arXiv 2024
[2]

Ziqiang Cui, Haolun Wu, Bowei He, Ji Cheng, and Chen Ma. 2024. Context Matters: Enhancing Sequential Recommendation with Context-aware Diffusion- based Contrastive Learning. In Proceedings of the 33rd ACM International Confer- ence on Information and Knowledge Management . 404–414

work page 2024
[3]

Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. 2024. A review of modern recommender systems using generative models (gen-recsys). In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 6448–6458

work page 2024
[4]

Dario Di Palma. 2023. Retrieval-augmented recommender system: Enhancing recommender systems with large language models. In Proceedings of the 17th ACM Conference on Recommender Systems . 1369–1373

2023
[5]

Thennakoon Mudiyanselage Anupama Udayangani Gunathilaka, Prab- hashrini Dhanushika Manage, Jinglan Zhang, Yuefeng Li, and Wayne Kelly. 2025. Addressing sparse data challenges in recommendation systems: A Systematic review of rating estimation using sparse rating data and profile enrichment techniques. Intelligent Systems with Applications (2025), 200474

work page 2025
[6]

Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al . 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 3, 4 (2023), 6

arXiv 2023
[7]

Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. 2024. Bridging the preference gap between retrievers and llms. arXiv preprint arXiv:2401.06954 (2024)

arXiv 2024
[8]

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. arXiv preprint arXiv:2502.12586 (2025)

work page Pith review arXiv 2025

Show all 18 references

[9]

Reza Yousefi Maragheh, Ramin Giahi, Jianpeng Xu, Lalitesh Morishetti, Shanu Vashishtha, Kaushiki Nag, Jason Cho, Evren Korpeoglu, Sushant Kumar, and Kannan Achan. 2022. Prospect-net: Top-k retrieval problem using prospect theory. In 2022 IEEE International Conference on Big Da...

2022
[10]

Matin Mortaheb, Mohammad A Amir Khojastepour, Srimat T Chakradhar, and Sennur Ulukus. 2025. Re-ranking the Context for Multimodal Retrieval Aug- mented Generation. arXiv preprint arXiv:2501.04695 (2025)

2025 arXiv
[11]

H Penny Nii. 1986. The blackboard model of problem solving and the evolution of blackboard architectures. AI magazine 7, 2 (1986), 38–38

1986
[12]

Lakshmanan Rakkappan and Vaibhav Rajan. 2019. Context-aware sequential recommendations withstacked recurrent neural networks. In The world wide web conference. 3172–3178

2019
[13]

Nicholas Rossi, Juexin Lin, Feng Liu, Zhen Yang, Tony Lee, Alessandro Magnani, and Ciya Liao. 2024. Relevance filtering for embedding-based retrieval. In Pro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management. 4828–4835

2024
[14]

Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024. DRAGIN: Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models. arXiv preprint arXiv:2403.10081 (2024)

2024 arXiv
[15]

Junda Wu, Cheng-Chun Chang, Tong Yu, Zhankui He, Jianing Wang, Yupeng Hou, and Julian 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 Dat...

2024
[16]

Zikun Ye, Reza Yousefi Maragheh, Lalitesh Morishetti, Shanu Vashishtha, Jason Cho, Kaushiki Nag, Sushant Kumar, and Kannan Achan. 2023. Seller-side Outcome Fairness in Online Marketplaces. arXiv preprint arXiv:2312.03253 (2023)

2023 arXiv
[17]

Reza Yousefi Maragheh, Xin Chen, James Davis, Jason Cho, Sushant Kumar, and Kannan Achan. 2020. Choice modeling and assortment optimization in the presence of context effects. A vailable at SSRN 3747354 (2020)

2020
[18]

Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Moham- mad Shoeybi, and Bryan Catanzaro. 2024. Rankrag: Unifying context ranking with retrieval-augmented generation in llms. Advances in Neural Information Processing Systems 37 (2024), 121156–121184

2024

Pith tools

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

[1] [1]

Justin Chih-Yao Chen, Archiki Prasad, Swarnadeep Saha, Elias Stengel-Eskin, and Mohit Bansal. 2024. MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning. arXiv preprint arXiv:2409.12147 (2024)

arXiv 2024

[2] [2]

Ziqiang Cui, Haolun Wu, Bowei He, Ji Cheng, and Chen Ma. 2024. Context Matters: Enhancing Sequential Recommendation with Context-aware Diffusion- based Contrastive Learning. In Proceedings of the 33rd ACM International Confer- ence on Information and Knowledge Management . 404–414

work page 2024

[3] [3]

Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. 2024. A review of modern recommender systems using generative models (gen-recsys). In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 6448–6458

work page 2024

[4] [4]

Dario Di Palma. 2023. Retrieval-augmented recommender system: Enhancing recommender systems with large language models. In Proceedings of the 17th ACM Conference on Recommender Systems . 1369–1373

2023

[5] [5]

Thennakoon Mudiyanselage Anupama Udayangani Gunathilaka, Prab- hashrini Dhanushika Manage, Jinglan Zhang, Yuefeng Li, and Wayne Kelly. 2025. Addressing sparse data challenges in recommendation systems: A Systematic review of rating estimation using sparse rating data and profile enrichment techniques. Intelligent Systems with Applications (2025), 200474

work page 2025

[6] [6]

Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al . 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 3, 4 (2023), 6

arXiv 2023

[7] [7]

Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. 2024. Bridging the preference gap between retrievers and llms. arXiv preprint arXiv:2401.06954 (2024)

arXiv 2024

[8] [8]

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. arXiv preprint arXiv:2502.12586 (2025)

work page Pith review arXiv 2025