LLMs as Better Recommenders with Natural Language Collaborative Signals: A Self-Assessing Retrieval Approach

Chao Wang; Haoran Xin; Hui Xiong; Weijia Zhang; Ying Sun

REVIEW 4 major objections 4 minor 42 references

LLMs as Better Recommenders with Natural Language Collaborative Signals: A Self-Assessing Retrieval Approach

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Plain-language signals about similar users beat soft-token and identifier baselines for LLM recommenders on the datasets tested.

desk verdict A practical natural-language CI retrieval pipeline with consistent UAUC gains, but the SARE supervision is under-tested and the statistics are thin; worth refereeing, not ready as is. read the letter →

arxiv 2505.19464 v1 pith:PYXSFXKJ submitted 2025-05-26 cs.IR

Haoran Xin , Ying Sun , Chao Wang , Weijia Zhang , Hui Xiong This is my paper

classification cs.IR

keywords LLM-basedrecommendationcollaborativeinformationretrievalnaturallanguagesignalsretrieve-rerankframeworkself-assessingrerankercontrastivefine-tuningMovieLens-1MAmazon-Games

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

Collaborative information—what other users with similar tastes did—is normally fed to LLM recommenders as soft tokens or abstract identifiers, which do not match the language the model was pretrained on. This paper claims that writing those signals as plain-language behavior descriptions, and choosing them carefully, makes them useful: a collaborative retriever (CAR) finds similar users, and a self-assessing reranker (SARE) ranks candidates by how well they match the LLM's own description of what extra information would help the decision. The top-ranked behaviors are prepended to the prompt as natural-language context. On MovieLens-1M and Amazon-Games, the resulting system reports the best user-level ranking quality (UAUC) on both datasets and the best AUC on Games, beating prior CI-injection methods such as LLaRA, CoLLM, and BinLLM. The reader should care because the paper offers a practical way to inject collaborative knowledge into an LLM without changing its architecture or its vocabulary.

What carries the argument

The mechanism has three load-bearing pieces. (1) CAR: an off-the-shelf text retriever encoder, fine-tuned with a contrastive loss (Equation 6) where positive candidates are users whose CRM embeddings (Equation 3) have high inner product with the target user; this gives retrieval that is simultaneously collaborative and semantic. (2) SARE: a reranker built from a general-purpose reranker encoder, fine-tuned with InfoNCE (Equation 12) against a self-referential ranking. The ranking (Equations 7 to 9) comes from prompting the LLM to state what extra preference characteristics would improve the decision, embedding that assessment with a fixed text embedder, and sorting retrieved behavior descriptions by cosine similarity to it. (3) The prompt itself: selected behaviors are rendered as sentences—"Other users with similar preferences have given high ratings to..."—and concatenated with the target user's history and the target item, with a LoRA adapter trained for the final prediction. The self-assessment is the hinge: it converts an LLM's introspective text into a numerical ranking without any direct measurement of recommendation accuracy.

What would settle it

A direct test: for a sample of target users, rank the CAR-retrieved candidate behaviors by the measured change in the final LLM's AUC or UAUC when each candidate's behavior is added to the prompt, then compare that ranking with SARE's self-assessment ranking (Equation 9). If the two rankings show little or no positive correlation, the self-assessment mechanism is not selecting behaviors that improve recommendations, and the reported gains would need another explanation.

Watch

Extended reading notes

Core claim

The central discovery is that the semantic format of collaborative information matters as much as its content. Rather than translating user-item interactions into a latent embedding or an identifier that the LLM must learn to decode, SCORE converts them into sentences built from item metadata and treats them as retrieval-augmented context. The retrieval is two-stage. First, a collaborative retriever (CAR) is produced by fine-tuning a text retriever with a contrastive objective whose positive pairs come from collaborative similarity computed by a pretrained CRM such as SASRec; this makes the retriever joint over interaction patterns and semantics. Second, a self-assessing reranker (SARE) is trained on rankings generated by the LLM's own free-text assessment: the LLM is asked what additional movie characteristics would help decide a target item, the assessment is embedded, and retrieved behaviors are ranked by cosine similarity to that assessment. SARE is a reranker trained with a contrastive loss on those rankings. At inference, the top behaviors are verbalized and prepended to the recommendation prompt, and a lightweight low-rank adapter (LoRA) tunes the frozen LLM backbone for the final yes/no prediction. On the Games dataset SCORE reports the best AUC and UAUC; on ML-1M it reports the best UAUC, with the best AUC going to BinLLM. The authors interpret the gains as evidence that natural-language CI aligns with the LLM's pretraining, and that self-assessment identifies behaviors that are similar but not merely redundant.

Load-bearing premise

The whole method depends on assuming that the more a retrieved user's behavior text resembles the LLM's own description of what would help, the more that behavior actually improves the final recommendation; the training signal for the reranker is built from this resemblance, never from measured recommendation accuracy.

Editorial extensions

If this is right

On the reported metrics, natural-language CI selected by CAR and SARE beats soft-token (LLaRA, CoLLM), binary-string (BinLLM), and identifier-based (CoRAL) CI on the Games dataset and on ML-1M UAUC, though BinLLM retains the best ML-1M AUC.
An LLM's self-assessed description of useful information is a transferable supervision signal: it can rank retrieved behaviors through embedding similarity and train a reranker without needing human utility labels.
Ablations in the paper show that neither pure collaborative retrieval nor pure semantic retrieval matches the joint CAR, and that removing SARE's rerank hurts, so similarity alone is insufficient for choosing CI.
Small final context sizes ($K_s=1$ to $2$) perform best, consistent with the lost-in-the-middle bias; injecting many similar-user behaviors degrades the LLM's use of them.
SCORE improves over LLaRA across SASRec, LightGCN, and DIN backbones in the majority of tested configurations, so the pipeline is not tied to a single CRM.

Reading between the lines

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

A testable extension the paper leaves implicit: replacing the fixed text embedder that matches assessments to behaviors with the LLM's own hidden states might sharpen SARE, since the embedder is a static proxy for what the LLM can actually exploit.
Because SARE's supervision never measures the final recommendation accuracy, an alternative explanation—that the ranking mainly encodes genre overlap or popularity—is consistent with the reported results; comparing SARE's ranking against rankings by measured per-candidate accuracy gains would settle it.
The same self-assessing retrieval recipe could be applied to other retrieval-augmented LLM uses—such as choosing in-context demonstrations or chain-of-thought examples—where the selection criterion is utility for the decision rather than topical similarity to the query.
The paper's stated limitation on prompt length suggests a compression experiment: if similar-user behavior descriptions are summarized into a few sentences before being prepended, the gains might persist at lower inference cost.

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

Here's the quick take. The paper's core move — retrieving other users' behavior sequences, writing them out in natural language, and prepending them to an LLM prompt — is a reasonable and under-explored direction. The CAR/SARE two-stage pipeline is clearly described, and the gains on UAUC are consistent across both datasets. I'd credit the authors for testing across three CRM backbones and for the -base/-cf ablations, which show that neither pure collaborative nor pure semantic retrieval alone matches the combined retriever.

The soft spots are real but mostly fixable. First, there are no error bars or significance tests anywhere. On ML-1M AUC, SCORE actually trails BinLLM, so the "best" claim depends on averaging metrics. Given ML-1M here is only 839 users, a few of those UAUC points could move a lot. Second, the paper never states whether the retrieval corpus for a test user includes validation/test interactions of other users. That's a leakage question that needs an explicit answer. Third, and this is the one I'd push hardest on: the SARE supervision is cosine similarity between an LLM-generated assessment text and each candidate user's behavior text. The prompt asks for "genres or characteristics," so the ranking is likely dominated by genre overlap with the target item. That's not necessarily useless, but the paper doesn't show it's any better than a trivial genre-based selection or that it captures "collaborative utility" beyond CAR. The -rerank ablation doesn't isolate the self-assessment signal because SARE is a cross-encoder and CAR is a bi-encoder; the gain could be model capacity. They'd need a baseline where SARE is trained on random or genre-only rankings, or on actual downstream utility.

The circularity worry from our stress-test is a bit overwrought: the assessment comes from GLM-4-Plus, not from the final Llama model, and it's a query-side text. But the genre-overlap mechanism is a live hypothesis and untested.

Who is this for? Researchers working on RAG-based LLM recommendation. They'll find the retrieve-rerank design and the natural-language CI formulation useful. The paper deserves a serious referee; I'd accept it for review with a request for significance testing, leakage clarification, and an ablation that controls for reranker architecture and genre-overlap.

Referee Report

4 major / 4 minor

Summary. The paper proposes SCORE, a retrieve-rerank framework that supplies collaborative information (CI) to LLM-based recommenders in natural language instead of soft tokens or abstract identifiers. A collaborative retriever (CAR) is fine-tuned with contrastive learning on embeddings derived from a pretrained collaborative model, and a self-assessing reranker (SARE) is trained on rankings produced by embedding an LLM's self-assessment text and matching it against retrieved user behavior text. At inference, the top-ranked user behaviors are prepended as natural-language CI to a LoRA-tuned Llama-3.1-8B-Instruct prompt. Experiments on ML-1M and Amazon-Games compare SCORE against conventional and LLM-based baselines, with ablations on CAR and SARE, sensitivity analyses on K_e and K_s, and a CRM-agnostic study.

Significance. If the results hold, the paper makes a useful contribution: it gives a concrete recipe for expressing collaborative signals in natural language, and its two-stage retrieval pipeline addresses a real problem of selecting informative CI under LLM context limits. The paper is generally clear, the related work is appropriate, and the experiments cover several relevant baselines. It also ships ablations for CAR and SARE and explores CRM-agnostic behavior. However, the central empirical claim depends on the validity of the self-assessing supervision signal, and on the absence of leakage from validation/test interactions; these points are not yet established. The contribution is therefore promising but currently insufficiently supported.

major comments (4)

[3.3, Eqs. (7)-(9)] The supervision ranking for SARE is generated by cosine similarity between a fixed text embedder's representation of the LLM's self-assessment text and the representation of a retrieved user's behavior text; this ranking is never validated against the actual utility of those behaviors for the downstream LLM recommendation. Because the prompt in Eq. (7) explicitly asks for 'genres or characteristics,' the assessment text is genre-focused, and cosine similarity with a semantic embedder may largely reflect genre overlap with the target item rather than collaborative value. This self-referential signal is load-bearing for the claimed benefit of SARE. Please add a control experiment that compares SARE-selected users with (a) randomly selected CAR-retrieved users and (b) users selected by genre/topic overlap with the target item, measured by final AUC/UAUC; also report the correlation between the SARE ranking and the actual change in the LLM's prediction when each candidate user is added.
[Section 2 and Section 4.1] The paper does not state explicitly that the interaction corpus H used for retrieval and for constructing the natural-language CI in the LLM prompt excludes validation and test interactions. Appendix A describes a temporal split of Y, but Eq. (2) defines R(·) over 'the interaction corpus H' without a temporal restriction, and the CI-augmented prompt in Section 3.4 does not specify that the similar-user behaviors are drawn only from the training portion. If H contains test-interval interactions of other users, the reported AUC/UAUC values would reflect leakage. Please state clearly that all retrieved behaviors are drawn exclusively from the training partition and verify that no validation or test interaction appears in any prompt.
[Table 2 and Figure 3] No error bars, confidence intervals, or significance tests are reported, although some headline differences are small (e.g., ML-1M AUC: SCORE 0.7367 vs BinLLM 0.7423, where BinLLM is actually higher on AUC; and the differences among SCORE variants in Figure 3 appear to be within a few thousandths). Without variance estimates, the claims of consistent superiority and the relative-improvement percentages are not statistically supported. Please report standard deviations over multiple seeds for the main comparisons and the key ablations, and, if feasible, paired significance tests.
[4.3, Figure 3] The '-rerank' ablation does not isolate the effect of the self-assessing ranking because SARE uses bge-reranker-large, a cross-encoder with higher capacity and a different interaction with the prompt, whereas CAR uses mpnet-base, a bi-encoder. The observed gain may come from the cross-encoder architecture rather than from the validity of the LLM self-assessment supervision. Please add an ablation that reranks CAR outputs with bge-reranker-large without SARE training, or with SARE trained on random/genre-similarity rankings, to control for architecture and training signal.

minor comments (4)

[3.4 and Appendix B] The acronym 'RARE' appears instead of 'SARE' in the 'Step 2' description of the retrieval procedure and in the reranker fine-tuning sentence in Appendix B; please correct these typos.
[4.4] The text says 'as shown in Section 4.4' when referring to the sensitivity plots; these references should point to Figure 4, since the performance curves are figures, not a section.
[Table 3] In the LightGCN row for ML-1M, SCORE (AUC 0.7206, UAUC 0.6820) does not outperform LLaRA (AUC 0.7288, UAUC 0.6899); the statement that SCORE achieves the best performance in the 'majority of scenarios' is accurate only if this exception is acknowledged, so please make the exception explicit.
[Appendix B] The implementation details list dataset split criteria, model choices, and several hyperparameters, but do not report training steps/epochs, learning rates, or batch sizes for the LoRA tuning and the SARE fine-tuning; please add these details for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all three stages are trained on external proxies and evaluated on held-out labels.

full rationale

SCORE's derivation chain is: (i) CAR is fine-tuned with a contrastive loss whose positive labels are the top-K CRM-similar users (Eqs. 4-6); (ii) SARE is fine-tuned with InfoNCE on rankings obtained by cosine similarity between an LLM-generated assessment text and retrieved user behavior text (Eqs. 7-9, 12); (iii) the LoRA-tuned recommender is evaluated on held-out user-item labels (Eqs. 13-14). None of these stages uses the evaluation labels as supervision, and no equation defines the final prediction in terms of the training objective. The self-assessment ranking is a proxy for CI utility and is not itself the reported metric, so the claim that it improves recommendations is an empirical one, not a tautology. The paper has no load-bearing self-citations, and the cited techniques (InfoNCE, Top-k shifted by N sampling, bge-reranker) are standard external building blocks. The Limitations section raises prompt-length and representation-expressiveness issues only. The possible weakness that the SARE ranking may largely reflect genre overlap is a validity concern, not a circularity: the held-out AUC/UAUC comparison remains an independent test. Therefore score 0.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. Its empirical claims rest mainly on the pretrained CRM embeddings, the standard contrastive objectives, and the unvalidated assumption that LLM self-assessment text is a good proxy for collaborative information utility. Several hyperparameters are chosen by hand and are not justified across datasets.

free parameters (10)

Kc = 5
Number of collaborative-similar users retrieved by CRM embeddings in CAR training (Section B).
Ke = 10
Number of CAR-retrieved users passed to SARE for reranking (Section B).
Ks = 2
Number of final similar users injected into the recommendation prompt (Section B and Section 4.4).
tau_CAR = 0.1
Temperature in the contrastive loss for CAR (Eq. 6, Section B).
tau_SARE = 0.02
Temperature in the InfoNCE loss for SARE (Eq. 12, Section B).
K1 = 5
Number of positive samples in top-k shifted by N sampling for SARE training (Eq. 11, Section B).
K2 = 3
Offset for negative sampling in SARE training (Eq. 11, Section B).
LoRA rank = 8
Rank of LoRA adapter for recommendation fine-tuning (Section B).
LoRA alpha = 32
Scaling parameter for LoRA adapter (Section B).
assessed user sample size = 10000
Number of randomly sampled users used to generate LLM self-assessments for SARE training (Section B).

assumptions (5)

domain assumption SASRec embeddings provide a reliable collaborative similarity signal for retrieving similar users.
Used in Eq. 3-4; the paper relies on the pretrained CRM without auditing its quality on these datasets.
standard math InfoNCE contrastive loss and top-k shifted sampling are appropriate training objectives for retriever and reranker.
Standard practices, but the paper does not justify them for this task.
ad hoc to paper LLM self-assessment text, when embedded, is a valid proxy for collaborative information utility.
The whole SARE ranking depends on this; it is not validated against actual recommendation outcomes.
domain assumption Item metadata is available and sufficient to represent user behaviors in natural language.
The method converts histories to text using item descriptions; this is stated as feasible in Section 1 but not analyzed.
domain assumption The interaction corpus used for retrieval is temporally clean.
The paper says splits are temporal but does not state whether H excludes validation/test interactions; if it includes them, there is leakage.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LLMs as Better Recommenders with Natural Language Collaborative Signals: A Self-Assessing Retrieval Approach." pith.science (2026). https://pith.science/paper/PYXSFXKJ

@misc{pith2026250519464,
  author       = {Pith},
  title        = {Pith review of: LLMs as Better Recommenders with Natural Language Collaborative Signals: A Self-Assessing Retrieval Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PYXSFXKJ}},
  note         = {Machine review of arXiv:2505.19464}
}

read the original abstract

Incorporating collaborative information (CI) effectively is crucial for leveraging LLMs in recommendation tasks. Existing approaches often encode CI using soft tokens or abstract identifiers, which introduces a semantic misalignment with the LLM's natural language pretraining and hampers knowledge integration. To address this, we propose expressing CI directly in natural language to better align with LLMs' semantic space. We achieve this by retrieving a curated set of the most relevant user behaviors in natural language form. However, identifying informative CI is challenging due to the complexity of similarity and utility assessment. To tackle this, we introduce a Self-assessing COllaborative REtrieval framework (SCORE) following the retrieve-rerank paradigm. First, a Collaborative Retriever (CAR) is developed to consider both collaborative patterns and semantic similarity. Then, a Self-assessing Reranker (SARE) leverages LLMs' own reasoning to assess and prioritize retrieved behaviors. Finally, the selected behaviors are prepended to the LLM prompt as natural-language CI to guide recommendation. Extensive experiments on two public datasets validate the effectiveness of SCORE in improving LLM-based recommendation.

Figures

Figures reproduced from arXiv: 2505.19464 by the authors.

**Figure 2.** (a) Overview of the SCORE framework. A two-stage fine-tuning paradigm is used to develop the collaborative retriever (CAR) and the self-assessing reranker (SARE). (b) Illustration of the self-assessing ranking process. The LLM evaluates the characteristics of beneficial CI for recommendation, which guides the reranking of retrieved similar users. (c) User behaviors are retrieved and reranked by CAR and SARE, then pr… view at source ↗

**Figure 3.** Recommendation performance of different variants of SCORE. based solely on either collaborative similarity (-cf) or semantic similarity (-base) overlooks important aspects of similarity for LLM-based recommendation. Pure collaborative similarity fails to capture semantic context, while a general-purpose retriever lacks the domain-specific insights for the task. Can reranking by SARE help the recommendation? We com… view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Attention visualization of the input prompt. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Example of the user movie viewing history, the target item and the reasoning evidence generated by LLMs. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 14 canonical work pages

[1]

Michael Antonios Kruse Ayoub, Zhan Su, and Qiuchi Li. 2024. A case study of enhancing sparse retrieval using llms. In Companion Proceedings of the ACM on Web Conference 2024, pages 1609--1615

work page 2024
[2]

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. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 1007--1014

work page 2023
[3]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PMLR

2020
[4]

Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, and Dong Yu. 2023. Dense x retrieval: What retrieval granularity should we use? arXiv preprint arXiv:2312.06648

arXiv 2023
[5]

Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxiang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 1126--1132

2023
[6]

Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234

arXiv 2022
[7]

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

arXiv 2023
[8]

F Maxwell Harper and Joseph A Konstan. 2015. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis), 5(4):1--19

2015

Show all 42 references

[9]

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

2020
[10]

B Hidasi. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939

2015 arXiv
[11]

Yupeng Hou, Binbin Hu, Zhiqiang Zhang, and Wayne Xin Zhao. 2022. Core: simple and effective session-based recommendation within consistent representation space. In Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, ...

2022
[12]

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 arXiv
[13]

Pengyue Jia, Derong Xu, Xiaopeng Li, Zhaocheng Du, Xiangyang Li, Xiangyu Zhao, Yichao Wang, Yuhao Wang, Huifeng Guo, and Ruiming Tang. 2024. Bridging relevance and reasoning: Rationale distillation in retrieval-augmented generation. arXiv preprint arXiv:2412.08519

2024
[14]

Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2023. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. arXiv preprint arXiv:2310.06839

2023 arXiv
[15]

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

2018
[16]

Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large language models meet collaborative filtering: An efficient all-round llm-based recommender system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and D...

2024
[17]

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

2024
[18]

Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recommendation. In Proceedings of the ACM on Web Conferen...

2024
[19]

Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is chatgpt a good recommender? a preliminary study. arXiv preprint arXiv:2304.10149

2023 arXiv
[20]

Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157--173

2024
[21]

Gabriel de Souza P Moreira, Ronay Ak, Benedikt Schifferer, Mengyao Xu, Radek Osmulski, and Even Oldridge. 2024 a . Enhancing q&a text retrieval with ranking models: Benchmarking, fine-tuning and deploying rerankers for rag. arXiv preprint arXiv:2409.07691

2024 arXiv
[22]

Gabriel de Souza P Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, and Even Oldridge. 2024 b . Nv-retriever: Improving text embedding models with effective hard-negative mining. arXiv preprint arXiv:2407.15831

2024 arXiv
[23]

Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natura...

2019
[24]

Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

2018 arXiv
[25]

Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Representation learning with large language models for recommendation. In Proceedings of the ACM on Web Conference 2024, pages 3464--3475

2024
[26]

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2012. Bpr: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618

2012 arXiv
[27]

Yuling Wang, Changxin Tian, Binbin Hu, Yanhua Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, Liang Pang, and Xiao Wang. 2024. Can small language models be good reasoners for sequential recommendation? In Proceedings of the ACM on Web Conference 2024, pages 3876--3887

2024
[28]

Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Llmrec: Large language models with graph augmentation for recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, ...

2024
[29]

Junda Wu, Cheng-Chun Chang, Tong Yu, Zhankui He, Jianing Wang, Yupeng Hou, and Julian McAuley. 2024 a . Coral: collaborative retrieval-augmented large language models improve long-tail recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and ...

2024
[30]

Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, and 1 others. 2024 b . A survey on large language models for recommendation. World Wide Web, 27(5):60

2024
[31]

Fangyuan Xu, Weijia Shi, and Eunsol Choi. 2023. Recomp: Improving retrieval-augmented lms with compression and selective augmentation. arXiv preprint arXiv:2310.04408

2023 arXiv
[32]

Shicheng Xu, Liang Pang, Jun Xu, Huawei Shen, and Xueqi Cheng. 2024. List-aware reranking-truncation joint model for search and retrieval-augmented generation. In Proceedings of the ACM on Web Conference 2024, pages 1330--1340

2024
[33]

Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. 2024. Rankrag: Unifying context ranking with retrieval-augmented generation in llms. arXiv preprint arXiv:2407.02485

2024 arXiv
[34]

Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. 2023 a . Retrieve anything to augment large language models. arXiv preprint arXiv:2310.07554

2023
[35]

Yang Zhang, Keqin Bao, Ming Yan, Wenjie Wang, Fuli Feng, and Xiangnan He. 2024. Text-like encoding of collaborative information in large language models for recommendation. arXiv preprint arXiv:2406.03210

2024 arXiv
[36]

Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He. 2023 b . Collm: Integrating collaborative embeddings into large language models for recommendation. arXiv preprint arXiv:2310.19488

2023 arXiv
[37]

Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji - Rong Wen. 2021. Recbole: Towards a unified, c...

2021
[38]

Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 1435--1448. IEEE

2024
[39]

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. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 5941--5948

2019
[40]

Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collaborative large language model for recommender systems. In Proceedings of the ACM on Web Conference 2024, pages 3162--3172

2024
[41]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[42]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Michael Antonios Kruse Ayoub, Zhan Su, and Qiuchi Li. 2024. A case study of enhancing sparse retrieval using llms. In Companion Proceedings of the ACM on Web Conference 2024, pages 1609--1615

work page 2024

[2] [2]

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. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 1007--1014

work page 2023

[3] [3]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PMLR

2020

[4] [4]

Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, and Dong Yu. 2023. Dense x retrieval: What retrieval granularity should we use? arXiv preprint arXiv:2312.06648

arXiv 2023

[5] [5]

Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxiang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 1126--1132

2023

[6] [6]

Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234

arXiv 2022

[7] [7]

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

arXiv 2023

[8] [8]

F Maxwell Harper and Joseph A Konstan. 2015. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis), 5(4):1--19

2015