Pith. sign in

REVIEW 4 major objections 5 minor 62 references

Consensus vs. Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommenders

T0 review · 4 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Fine-tuned LLMs can pick the aggregation strategy that groups rate highest for fairness and consensus.

desk verdict Clean closed-loop demo that a fine-tuned open LLM can pick social-choice aggregation strategies and beat static baselines on human fairness/consensus, especially for minority and coalitional groups. read the letter →

arxiv 2607.10235 v1 pith:CTED7RYO submitted 2026-07-11 cs.CL cs.IR

Cedric Waterschoot , Nava Tintarev , Francesco Barile This is my paper
classification cs.CLcs.IR
keywords LargeLanguageModelsFairnessGroupRecommenderSystemsLLM-as-judgeUserstudyPreferenceaggregationSocialchoice
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

Group recommender systems usually apply one fixed rule for combining members' ratings, yet people judge fairness, satisfaction and consensus differently depending on whether the group is uniform, has a minority, or is split into coalitions. This paper shows that large language models can be fine-tuned on real human survey scores plus synthetic reasoning so they act as live judges of those three qualities. The system first computes six classic social-choice candidates, then lets the fine-tuned model score each candidate multiple times and selects the one predicted to score highest. A between-subjects user study with 284 participants found that this dynamic choice received the highest average fairness and consensus ratings overall, and statistically significant gains precisely in the minority and coalitional configurations. The result supports adapting the aggregation rule on the fly so that the final recommendation better matches how people actually feel about group outcomes.

What carries the argument

Judgmental OLMo (and Llama): 7–8 B models LoRA-fine-tuned on an augmented set of 5 318 examples that pair human ground-truth scores with synthetic reasoning; inside the pipeline they evaluate each of six social-choice candidates five times and rank by average predicted fairness/satisfaction/consensus.

What would settle it

Collect free-text justifications from a new cohort of human participants on the same scenarios, retrain the judgmental models on those genuine texts, re-run the selection pipeline, and check whether the human-rated fairness and consensus scores of the chosen recommendations drop significantly.

Watch

Extended reading notes

Core claim

Fine-tuning open LLMs on human Likert assessments of fairness, satisfaction and consensus, expanded by chain-of-thought justifications distilled from a larger teacher model, yields judgmental models that can rank social-choice aggregation strategies. When these models are placed inside a recommender pipeline that generates six candidates and selects the top-scoring one, the resulting recommendations receive the highest human-rated fairness and consensus scores in a 284-person study, with the largest advantages appearing for minority and coalitional preference distributions.

Load-bearing premise

The synthetic reasoning paths produced by the teacher model from the original human scores faithfully capture the justifications people actually use when rating fairness, satisfaction and consensus.

Editorial extensions

If this is right

  • Dynamic selection among transparent social-choice rules can raise perceived fairness and consensus without inventing new aggregation formulas.
  • Gains concentrate in minority and coalitional groups, so systems should condition strategy choice on within-group preference similarity.
  • The same fine-tuned judges can stand in for costly human studies when evaluating group recommendations at scale or in real time.
  • Averaging the three metrics already yields consistent rankings; the pipeline can later be tuned to prioritise only fairness or only consensus when the application requires it.

Reading between the lines

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

  • Distilling the same judgment capability into models under 2 B parameters would make multi-assessment committees fast enough for production recommenders.
  • The approach could transfer to other multi-stakeholder settings (content moderation, resource allocation) where fairness and consensus must be scored on the fly.
  • An ablation on committee size (number of parallel assessments) would reveal the smallest number that still preserves ranking quality and cut latency.
  • Collecting genuine free-text human reasoning instead of synthetic justifications would remove residual dependence on the teacher model’s style.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes fine-tuning Llama-3.1-8B and OLMo-3-7B on human Likert assessments of fairness, satisfaction and consensus (from Barile et al.) augmented by synthetic chain-of-thought generated by DeepSeek-V3.1, yielding Judgmental Llama/OLMo. These models are inserted into a GRS pipeline that enumerates six social-choice aggregation strategies (ADD, APP, FAI, LMS, MAJ, MPL), scores each candidate five times, and selects the strategy maximizing the average predicted scores. A preregistered between-subjects user study (n=284) finds that the dynamic LLM selection obtains the highest mean fairness (0.82) and consensus (0.81) and second-highest satisfaction, with statistically significant configuration-by-strategy interactions that favor the LLM method especially for minority and coalitional groups. Wasserstein distances on novel scenarios show the fine-tuned models closer to the human response histograms than the corresponding base models.

Significance. If the end-to-end claim holds, the work supplies a practical, transparent mechanism for adapting classic social-choice aggregators to within-group preference distributions using LLM-as-judge models that are explicitly anchored to human subjective metrics. Strengths that raise the contribution include the open companion repository (code, augmented CoT data, training configs), the preregistered and powered user study with mixed ANOVA and Bonferroni correction, the explicit comparison of base versus fine-tuned distributions via Wasserstein distance, and the demonstration that strategy selection varies systematically with group configuration (uniform vs. coalitional). These elements make the pipeline reproducible and falsifiable, advancing both GRS practice and the use of fine-tuned LLMs for subjective evaluation.

major comments (4)
  1. Section 3.1.4 and Table 2 report only Wasserstein distances of the marginal Likert distributions (plus generation loss for checkpoint selection) on 1 180 unlabeled novel scenarios and the 262-item stratified test split. No instance-level metrics (MAE, Pearson/Spearman correlation, or ranking agreement) are given against the actual human scores available in that held-out set. Because the pipeline’s selection decision rests on the relative ranking of the six candidates, the absence of these metrics leaves open whether Judgmental OLMo merely reproduces the overall histogram or reliably tracks human judgments on specific (matrix, recommendation) pairs; the favorable user-study outcomes on the 16 scenarios therefore provide weaker evidence of systematic human-aligned selection than claimed.
  2. Section 3.1.2: the training signal consists of synthetic CoT generated by DeepSeek-V3.1 (T=0.95) conditioned on the original human Likert scores. The paper never validates that these justifications match actual human reasoning (e.g., by collecting free-text rationales from a subset of participants or by measuring agreement with human explanations). The axiom that the distilled CoT constitutes a faithful training signal for human-like judgment is therefore untested and load-bearing for the claim that the fine-tuned models simulate nuanced group assessments.
  3. Abstract vs. Section 4.2.3 / Table 3: the abstract states that the methodology “achieved the highest scores for satisfaction and group consensus,” yet the body and Table 3 show LLM highest only for fairness (0.82) and consensus (0.81), while FAI edges satisfaction (0.70 vs. 0.68). This factual inconsistency must be corrected; it currently overstates the satisfaction result that is central to the contribution.
  4. Section 3.2.1 fixes the assessment committee size at five with no ablation or sensitivity analysis. Because the final ranking of strategies is the average of these five draws, the free parameter directly affects selection stability; without evidence that results are robust to committee size (or to the choice of averaging versus other aggregation of the three metrics), the claim that the pipeline “dynamically selects the ideal aggregation strategy” rests on an unexamined hyper-parameter.
minor comments (5)
  1. Figure 1 caption and surrounding text: the human n is given as 1 152, yet the original study is described as 288 participants × 4 scenarios; a brief clarification of any filtering would help.
  2. Table 4: after Bonferroni correction (α=0.017) the main effect of aggregation strategy on satisfaction (p=.032) is non-significant, yet the text still discusses it alongside the significant interactions; a clearer separation of main-effect versus interaction claims would improve precision.
  3. Section 4.2.2: the frequency counts of selected strategies (ADD 6/16, FAI 4/16, …) are useful; adding a small contingency table by configuration would make the reported patterns easier to verify.
  4. Throughout: occasional typographical slips (“oru user study”, “1.5.1 LLMs” in the figure legend, inconsistent capitalization of strategy acronyms) should be cleaned.
  5. Related-work discussion of LLM-as-judge biases (Shi et al., Chen et al., Malberg et al.) is thorough; a short note on whether any residual position or self-preference bias was observed after fine-tuning would strengthen the evaluation section.

Circularity Check

1 steps flagged · score 1.0 of 10

No definitional circularity; mild self-use of authors' prior survey data for training labels and scenario templates, but the central claim is independently validated by a fresh n=284 user study.

  1. self citation load bearing [Section 3.1.1 Survey Data / Section 3.2.2 User Study Evaluation]
    "The starting point of obtaining fine-tuned LLMs for generating human-aligned assessments was the human ground truth presented in Barile[4], Barile et al. [5]. ... Following the code and procedure provided by Barile et al. [5] and thus, adhering to the format of the previously used survey data, we generated a set of group scenarios to be used in the user evaluation."

    Training labels and the exact scenario-generation procedure (four group configurations, restaurant matrices, fourth-ranked item) are taken from the authors' own prior study. This is load-bearing for the fine-tuning corpus and for constructing the 16 evaluation scenarios, yet the final human ratings that support the strongest claim come from a new sample; the circularity is therefore only mild and non-definitional.

full rationale

The paper's derivation chain is: (1) take human Likert scores from Barile et al. (same authors) as ground-truth labels, (2) distill synthetic CoT from DeepSeek-V3.1 conditioned on those labels, (3) fine-tune Llama/OLMo to produce score+reasoning distributions whose Wasserstein distance to the human histogram is reduced, (4) embed the best model as a selector among six social-choice candidates, and (5) measure the selected recommendations with an entirely new Prolific cohort (n=284) that never saw the training scenarios. Step (5) is an external empirical test; the high fairness/consensus means are not forced by the training labels or by any equation that equates input to output. The only mild self-reference is reuse of the prior survey's scenario-generation code and the same four group configurations; this is ordinary data reuse, not a uniqueness theorem, fitted parameter renamed as prediction, or self-definitional identity. No instance-level ranking correlation is reported on the held-out 262, but that is a validity gap, not circularity. Score remains 0-2.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on three free parameters (committee size, LoRA rank/alpha, teacher sampling temperature), a handful of domain assumptions about social-choice strategies and Likert-scale human judgments, and two invented entities (the Judgmental models themselves). No new physical or mathematical primitives are postulated; the ledger is therefore modest.

free parameters (3)
  • committee_size = 5
    Number of independent LLM assessments averaged per candidate (fixed at 5); never ablated, yet used to rank strategies.
  • LoRA_rank_and_alpha = r=16, α=32
    r=16, α=32 chosen for PEFT; standard but still free hyperparameters that affect the final judge.
  • DeepSeek_sampling_temperature = 0.95
    Temperature 0.95 and top-p 0.95 used to oversample CoT; controls diversity of synthetic reasoning.
assumptions (4)
  • domain assumption Social-choice aggregation strategies (ADD, APP, FAI, LMS, MAJ, MPL) are adequate and transparent generators of candidate recommendations.
    Stated in Section 2.1 and Table 1; the entire pipeline never leaves this closed set of six rules.
  • domain assumption Human Likert ratings of fairness, satisfaction and consensus collected under the Barile et al. protocol are a valid ground truth for training and evaluation.
    Section 3.1.1; the entire fine-tuning and user-study design inherits this instrument.
  • ad hoc to paper Synthetic chain-of-thought generated by DeepSeek-V3.1 from human scores is a useful training signal for human-like judgment.
    Section 3.1.2; no external validation that the teacher’s justifications match real human reasoning.
  • standard math Wasserstein distance on ordinal Likert histograms is an appropriate metric of distributional alignment.
    Used throughout Section 4.1; standard for ordinal data.
invented entities (1)
  • Judgmental Llama / Judgmental OLMo
    purpose: Fine-tuned open LLMs that emit human-like fairness/satisfaction/consensus scores for any group-recommendation scenario.
    Introduced in Section 3.1.3; their only evidence is the reduction in Wasserstein distance and the subsequent user-study gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Consensus vs. Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommenders." pith.science (2026). https://pith.science/paper/CTED7RYO

@misc{pith2026260710235,
  author       = {Pith},
  title        = {Pith review of: Consensus vs. Dissent: Dynamic LLM Modeling of Subjective Preferences in Group Recommenders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CTED7RYO}},
  note         = {Machine review of arXiv:2607.10235}
}
read the original abstract

Previous work in group recommender systems has demonstrated a sensitivity to the distribution of preferences within a group. Specifically, the selection of the preference aggregation strategy benefits from considering such group configurations. In this paper, we study whether LLMs are able to mimic this sensitivity and to select the ideal aggregation strategy (and corresponding recommendation) according to nuanced human perceptions of fairness, satisfaction, and consensus. We do this by fine-tuning Large Language Models (LLMs) on human survey data to serve as real-time judgmental models within the recommendation pipeline. Using a reasoning dataset distilled from DeepSeek-V3.1 and human ground truth assessments, we develop Judgmental Llama and Judgmental OLMo to simulate group assessments. Our pipeline successfully generates multiple recommendation candidates based on social choice-based aggregation strategies and dynamically selects the one that maximizes these predicted human-like evaluations. We further validate these suggestions in a user study (n=284) and find that our methodology achieved the highest scores for satisfaction and group consensus. Furthermore, we find that LLM judgments are most aligned with human perceptions of fairness, satisfaction and consensus when we also consider interaction effects between our LLM-based method and group configuration (e.g., minority or coalition). These findings give further support for dynamically adapting aggregation strategies to specific within-group preference distributions, and highlight the advantage of using LLMs for an adaptation that is aligned with subjective human judgments.

Figures

Figures reproduced from arXiv: 2607.10235 by the authors.

Figure 1
Figure 1. Human distributions of responses to rating group recommendations in terms of fairness, satisfaction and consensus [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the full pipeline. The methodology is divided into two main parts: (1) LLM fine-tuning for automated [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Distributions (across fairness, satisfaction, consensus) for both judgmental models and the human ground truth. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 9 canonical work pages

  1. [1]

    Paula Akemi Aoyagui, Kelsey Stemmler, Sharon A Ferguson, Young-Ho Kim, and Anastasia Kuzminykh. 2025. A Matter of Perspective(s): Contrasting Human and LLM Argumentation in Subjective Decision-Making on Subtle Sexism. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, N...

    doi:10.1145/3706598.3713248 2025

  2. [2]

    Bakker, Martin J

    Michiel A. Bakker, Martin J. Chadwick, Hannah R. Sheahan, Michael Henry Tessler, Lucy Campbell-Gillingham, Jan Balaguer, Nat McAleese, Amelia Glaese, John Aslanides, Matthew M. Botvinick, and Christopher Summerfield. 2022. Fine-tuning language models to find agreement among humans with diverse preferences. InProceedings of the 36th International Conferenc...

    2022

  3. [3]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Xinyue Huo, Chong Chen, and Fuli Feng

  4. [4]

    InProceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.)

    Decoding Matters: Addressing Amplification Bias and Homogeneity Issue in Recommendations for Large Language Models. InProceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguis- tics, Miami, Florida, USA, 10540–10552. doi:10.18653...

    doi:10.18653/v1/2024.emnlp-main.589 2024

  5. [5]

    Francesco Barile. 2023. UM XAI Explainable - GRS Impact of scenario complexity. doi:10.34894/8EVX4U

    work page doi:10.34894/8evx4u 2023

  6. [6]

    Francesco Barile, Tim Draws, Oana Inel, Alisa Rieger, Shabnam Najafian, Amir Ebrahimi Fard, Rishav Hada, and Nava Tintarev. 2023. Evaluating explainable social choice-based aggregation strategies for group recommendation.User Modeling and User-Adapted Interaction(2023), 1–58

    2023

  7. [7]

    Francesco Barile, Pierre Hurlin, Cedric Waterschoot, and Nava Tintarev. 2026. Critical reflections on user studies’ evaluation methods for group recommender systems.International Journal of Human-Computer Studies(2026), 103742. doi:10. 1016/j.ijhcs.2026.103742

    arXiv 2026

  8. [8]

    Francesco Barile, Shabnam Najafian, Tim Draws, Oana Inel, Alisa Rieger, Rishav Hada, and Nava Tintarev. 2021. Toward Benchmarking Group Explanations: Evaluating the Effect of Aggregation Strategies versus Explanation. (2021)

    2021

Show all 62 references
  1. [9]

    Jan Baumeister, Bernd Finkbeiner, Frederik Scheerer, Julian Siber, and Tobias Wagenpfeil. 2025. Stream-Based Monitoring of Algorithmic Fairness. InTools and Algorithms for the Construction and Analysis of Systems, Arie Gurfinkel and Marijn Heule (Eds.). Springer Nature Switzer...

    2025

  2. [10]

    Da Cao, Xiangnan He, Lianhai Miao, Yahui An, Chao Yang, and Richang Hong

  3. [11]

    InThe 41st International ACM SIGIR Conference on Research & Development in Information Retrieval(Ann Arbor, MI, USA)(SIGIR ’18)

    Attentive Group Recommendation. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retrieval(Ann Arbor, MI, USA)(SIGIR ’18). Association for Computing Machinery, New York, NY, USA, 645–654. doi:10.1145/3209978.3209998

    doi

  4. [12]

    Guiming Hardy Chen, Shunian Chen, Ziche Liu, Feng Jiang, and Benyou Wang

  5. [13]

    InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.)

    Humans or LLMs as the Judge? A Study on Judgement Bias. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 8301–8327...

    2024 doi

  6. [14]

    Lei Chen, Jie Cao, Huanhuan Chen, Weichao Liang, Haicheng Tao, and Guixiang Zhu. 2021. Attentive multi-task learning for group itinerary recommendation. Knowl. Inf. Syst.63, 7 (2021), 1687–1716. doi:10.1007/s10115-021-01567-3

    2021 doi

  7. [15]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongx- iang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering ChatGPT’s Capabilities in Recommender Systems. InProceedings of the 17th ACM Conference on Recom- mender Systems(Singapore, Singapore)(RecSys ’23). A...

    2023 doi

  8. [16]

    Yashar Deldjoo, Dietmar Jannach, Alejandro Bellogin, Alessandro Difonzo, and Dario Zanzonelli. 2024. Fairness in recommender systems: research landscape and future directions.User Model. User-adapt Interact.34, 1 (March 2024), 59–108

    2024

  9. [17]

    Amra Delić, Hanif Emamgholizadeh, Francesco Ricci, and Judith Masthoff. 2024. Supporting Group Decision-Making: Insights from a Focus Group Study. In Proceedings of the 32nd ACM Conference on User Modeling, Adaptation and Per- sonalization. 301–306. doi:10.1145/3627043.3659538

    2024 doi

  10. [18]

    Amra Delic, Judith Masthoff, Julia Neidhardt, and Hannes Werthner. 2018. How to use social relationships in group recommenders: empirical evidence. InPro- ceedings of the 26th Conference on User Modeling, Adaptation and Personalization. 121–129

    2018

  11. [19]

    Alexander Felfernig, Ludovico Boratto, Martin Stettinger, and Marko Tkalčič

  12. [20]

    InGroup Recommender Systems

    Explanations for Groups. InGroup Recommender Systems. Springer, 105– 126

  13. [21]

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-REC: Towards interactive and explainable LLMs-augmented recommender system. (2023). arXiv:2303.14524 [cs.IR]

    2023 arXiv

  14. [22]

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distill- ing Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. InFindings of the A...

    2023 doi

  15. [23]

    Zhenhua Huang, Xin Xu, Honghao Zhu, and MengChu Zhou. 2020. An Efficient Group Recommendation Model With Multiattention-Based Neural Networks. IEEE Transactions on Neural Networks and Learning Systems31, 11 (2020), 4461–

    2020

  16. [24]

    doi:10.1109/TNNLS.2019.2955567

    2019 doi

  17. [25]

    Mesut Kaya, Derek Bridge, and Nava Tintarev. 2020. Ensuring fairness in group recommendations by rank-sensitive balancing of relevance. InFourteenth ACM Conference on Recommender Systems(Virtual Event Brazil). ACM, New York, NY, USA

    2020

  18. [26]

    J.S. Kelly. 2013.Social Choice Theory: An Introduction. Springer Berlin Heidelberg

    2013

  19. [27]

    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. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Da...

    2024

  20. [28]

    Deepak Kumar, Tessa Grosz, Navid Rekabsaz, Elisabeth Greif, and Markus Schedl

  21. [29]

    doi:10.3389/fdata.2023.1245198

    Fairness of recommender systems in the recruitment domain: an analysis from technical and legal perspectives.Frontiers in Big DataVolume 6 - 2023 (2023). doi:10.3389/fdata.2023.1245198

    2023 doi

  22. [30]

    Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. 2025. From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge. InProceedings of...

    2025 doi

  23. [31]

    Yunqi Li, Hanxiong Chen, Shuyuan Xu, Yingqiang Ge, Juntao Tan, Shuchang Liu, and Yongfeng Zhang. 2023. Fairness in Recommendation: Foundations, Methods, and Applications.ACM Trans. Intell. Syst. Technol.14, 5, Article 95 (Oct. 2023), 48 pages. doi:10.1145/3610302

    2023 doi

  24. [32]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assis- tant. InProceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Retrieval(Washington D...

    2024 doi

  25. [33]

    Sebastian Lubos, Thi Ngoc Trang Tran, Alexander Felfernig, Seda Polat Erdeniz, and Viet-Man Le. 2024. LLM-generated Explanations for Recommender Systems. InAdjunct Proceedings of the 32nd ACM Conference on User Modeling, Adaptation Conference’17, July 2017, Washington, DC, USA...

    2024 doi

  26. [34]

    Simon Malberg, Roman Poletukhin, Carolin Schuster, and Georg Groh Groh

  27. [35]

    A Comprehensive Evaluation of Cognitive Biases in LLMs. InProceedings of the 5th International Conference on Natural Language Processing for Digital Humanities, Mika Hämäläinen, Emily Öhman, Yuri Bizzoni, So Miyagawa, and Khalid Alnajjar (Eds.). Association for Computational L...

    2025 doi

  28. [36]

    Judith Masthoff. 2015. Group recommender systems: aggregation, satisfaction and group attributes. Inrecommender systems handbook. Springer, 743–776

    2015

  29. [37]

    2022.Group Recommender Systems: Beyond Preference Aggregation

    Judith Masthoff and Amra Delić. 2022.Group Recommender Systems: Beyond Preference Aggregation. Springer US, New York, NY, 381–420. doi:10.1007/978-1- 0716-2197-4_10

    2022 doi

  30. [38]

    Shabnam Najafian and Nava Tintarev. 2018. Generating Consensus Explanations for Group Recommendations: an exploratory study. InAdjunct Publication of the 26th Conference on User Modeling, Adaptation and Personalization(Singapore, Singapore)(UMAP ’18). Association for Computing...

    2018

  31. [39]

    Thuy Ngoc Nguyen, Francesco Ricci, Amra Delic, and Derek Bridge. 2019. Conflict resolution in group decision making: insights from a simulation study.User Modeling and User-Adapted Interaction29, 5 (2019), 895–941

    2019

  32. [40]

    Alessandro Petruzzelli, Cataldo Musto, Lucrezia Laraspata, Ivan Rinaldi, Marco de Gemmis, Pasquale Lops, and Giovanni Semeraro. 2024. Instructing and prompting large language models for explainable cross-domain recommendations. In18th ACM Conference on Recommender Systems(Bari...

    2024

  33. [41]

    Silvia Rossi, Francesco Cervone, and Francesco Barile. 2018. An Altruistic-Based Utility Function for Group Recommendation. InTransactions on Computational Collective Intelligence XXVIII. Springer, 25–47

    2018

  34. [42]

    Alan Said. 2025. On explaining recommendations with Large Language Models: a review.Frontiers in Big Data7 (2025). doi:10.3389/fdata.2024.1505284

    2025 doi

  35. [43]

    Scott Sanner, Krisztian Balog, Filip Radlinski, Ben Wedin, and Lucas Dixon

  36. [44]

    InProceedings of the 17th ACM Conference on Recommender Systems(Singapore Singapore), Vol

    Large language models are competitive near cold-start recommenders for language- and item-based preferences. InProceedings of the 17th ACM Conference on Recommender Systems(Singapore Singapore), Vol. 1. ACM, New York, NY, USA, 890–896

  37. [45]

    Christophe Senot, Dimitre Kostadinov, Makram Bouzid, Jérôme Picault, Armen Aghasaryan, and Cédric Bernier. 2010. Analysis of strategies for building group profiles. InInternational Conference on User Modeling, Adaptation, and Personal- ization. Springer, 40–51

    2010

  38. [46]

    Lin Shi, Chiyu Ma, Wenhua Liang, Xingjian Diao, Weicheng Ma, and Soroush Vosoughi. 2025. Judging the Judges: A Systematic Study of Position Bias in LLM- as-a-Judge. InProceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference o...

    2025 doi

  39. [47]

    Maria Stratigi, Evaggelia Pitoura, and Kostas Stefanidis. 2023. SQUIRREL: A framework for sequential group recommendations through reinforcement learn- ing.Information Systems112 (2023), 102128. doi:10.1016/j.is.2022.102128

    2023 doi

  40. [48]

    Antonela Tommasel. 2024. Fairness Matters: A look at LLM-generated group recommendations. In18th ACM Conference on Recommender Systems(Bari Italy). ACM, New York, NY, USA, 993–998

    2024

  41. [49]

    Thi Ngoc Trang Tran, Müslüm Atas, Alexander Felfernig, Viet Man Le, Ralph Samer, and Martin Stettinger. 2019. Towards social choice-based explanations in group recommender systems. InProceedings of the 27th ACM Conference on User Modeling, Adaptation and Personalization. 13–21

    2019

  42. [50]

    Koki Wataoka, Tsubasa Takahashi, and Ryokan Ri. 2025. Self-preference bias in LLM-as-a-judge. (June 2025). arXiv:2410.21819 [cs.CL]

    2025 arXiv

  43. [51]

    Cedric Waterschoot, Nava Tintarev, and Francesco Barile. 2025. Consistent Explainers or Unreliable Narrators? Understanding LLM-generated Group Rec- ommendations. InProceedings of the Nineteenth ACM Conference on Recommender Systems (RecSys ’25). Association for Computing Mach...

    2025 doi

  44. [52]

    Cedric Waterschoot, Nava Tintarev, and Francesco Barile. 2025. The Pitfalls of Growing Group Complexity: LLMs and Social Choice-Based Aggregation for Group Recommendations. InAdjunct Proceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization (UMAP ...

    2025 doi

  45. [53]

    Cedric Waterschoot, Nava Tintarev, and Francesco Barile. 2026. Who is the Fairest of Them All? Using Large Language Models for Fairness Assessments. InProceedings of the 34th ACM Conference on User Modeling, Adaptation and Personalization (UMAP ’26). Association for Computing ...

    2026 doi

  46. [54]

    Cedric Waterschoot and Antal van den Bosch. 2024. A time-robust group recom- mender for featured comments on news platforms.Frontiers in Big DataVolume 7 - 2024 (2024). doi:10.3389/fdata.2024.1399739

    2024 doi

  47. [55]

    Cedric Waterschoot, Raciel Yera Toledo, Nava Tintarev, and Francesco Barile

  48. [56]

    InProceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization (UMAP ’25)

    With Friends Like These, Who Needs Explanations? Evaluating User Under- standing of Group Recommendations. InProceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization (UMAP ’25). Association for Computing Machinery, New York, NY, USA. doi:10.1145/...

    doi

  49. [57]

    Xuansheng Wu, Huachi Zhou, Yucheng Shi, Wenlin Yao, Xiao Huang, and Ning- hao Liu. 2024. Could small language models serve as recommenders? Towards data-centric cold-start recommendation. InProceedings of the ACM Web Confer- ence 2024(Singapore Singapore), Vol. 55. ACM, New Yo...

    2024

  50. [58]

    Ziqing Wu. 2023. Enhanced privacy preservation for recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems(Singapore Singapore), Vol. 11. ACM, New York, NY, USA, 1364–1368

    2023

  51. [59]

    Dayu Yang, Fumian Chen, and Hui Fang. 2024. Behavior alignment: A new perspective of evaluating LLM-based conversational recommendation systems. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval(Washington DC USA)...

    2024

  52. [60]

    Junwei Zhang, Min Gao, Junliang Yu, Lei Guo, Jundong Li, and Hongzhi Yin

  53. [61]

    InProceedings of the 30th ACM International Conference on Informa- tion & Knowledge Management(Virtual Event, Queensland, Australia)(CIKM ’21)

    Double-Scale Self-Supervised Hypergraph Learning for Group Recom- mendation. InProceedings of the 30th ACM International Conference on Informa- tion & Knowledge Management(Virtual Event, Queensland, Australia)(CIKM ’21). Association for Computing Machinery, New York, NY, USA, ...

    doi

  54. [62]

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, and Qing Li. 2024. Recommender systems in the era of large language models (LLMs).IEEE Trans. Knowl. Data Eng.36, 11 (Nov. 2024), 6889–6907

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.