Pith. sign in

REVIEW 3 major objections 4 minor 87 references

GORACS: Group-level Optimal Transport-guided Coreset Selection for LLM-based Recommender Systems

T0 review · 3 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 Selecting a small fine-tuning subset by an optimal-transport upper bound on test loss produces LLM recommenders that beat full-data training at lower cost.

desk verdict Useful, reproducible coreset-selection method for LLM recommenders, but the 'proven upper bound' claim collapses under inspection. read the letter →

arxiv 2506.04015 v1 pith:HXTPRVTE submitted 2025-06-04 cs.IR

Tiehua Mei , Hengrui Chen , Peng Yu , Jiaqing Liang , Deqing Yang This is my paper
classification cs.IR
keywords coresetselectionoptimaltransportLLM-basedrecommendationtestlossminimizationgroup-levelsubsetefficientLLMfine-tuninggradientnormsdata
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

The paper aims to make fine-tuning of large-language-model recommenders affordable by selecting a small 'coreset' of training examples that minimize the model's eventual test loss. It claims that previous selection strategies, which either match overall feature distributions or rank individual examples by importance, are misaligned with the fine-tuning objective and overlook the collective structure of recommendation data. GORACS instead chooses subsets group-wise by minimizing a Proxy Optimization Objective (POO), a bound on the intractable test loss built from an optimal-transport distance between the selected subset and the validation distribution plus per-example gradient norms. The experiments report that fine-tuning a 7-billion-parameter recommender on a GORACS-selected subset of 1,024 examples outperforms both existing selection baselines and training on the full dataset, while cutting total compute to roughly a fifth.

What carries the argument

The load-bearing object is the POO cost matrix $M = D^* - \lambda \mathbf{g}\mathbf{1}^\top$, where $D^*$ holds embedding distances between every training and validation sample and $\mathbf{g}$ holds per-sample gradient norms; the POO score is exactly the optimal-transport distance $\mathrm{OT}_M(\mu_S, \mu_V)$ with this cost. The second component is ITRA, a two-stage solver: a greedy algorithm for the relaxed p-median form of the objective produces the initial coreset, and an exchange stage uses a marginal-improvement estimator derived from the dual variables of the OT problem to prune candidate swaps and avoid exhaustive search. For classification-style recommendation tasks, the joint distribution is decomposed into class-conditional components so that selection is performed per class with preserved proportions.

What would settle it

On a dataset of choice, measure the Lipschitz constant L = sup over pairs of |L(z)-L(z')|/d(E(z),E(z')) for the fine-tuned loss; then construct two subsets with the same gradient-norm sum but very different OT distances to the validation set. If their fine-tuned test losses are equal, the OT term is not controlling the test loss as the bound claims.

Watch

Extended reading notes

Core claim

GORACS's central claim is that the POO score, $\mathcal{S}(S) = \mathrm{OT}_{D^*}(\mu_S, \nu_V) - \frac{\lambda}{|S|} \sum_{z \in S} \|\nabla_\phi L_{\phi_0}(z)\|$, is a computable upper bound on the test loss of the model fine-tuned on $S$, so minimizing it is a faithful proxy for minimizing the true objective. The bound is derived by combining Kantorovich-Rubinstein duality, which relates the OT distance to the difference between training and validation losses of a Lipschitz loss function, with a gradient-descent analysis showing that larger initial gradient norms imply larger training-loss reduction. The resulting optimization is solved by ITRA, which greedily builds a first subset under a relaxed p-median formulation and then improves it by sample exchanges whose expected benefit is estimated from optimal-transport dual variables. Experiments across three Amazon datasets and two tasks confirm that the selected coreset yields lower test loss and higher ranking metrics than all baselines and than full-data fine-tuning.

Load-bearing premise

The argument assumes that two similar-looking examples in embedding space produce similar fine-tuning losses (a Lipschitz condition) and that the validation set faithfully represents the future test distribution; neither is measured or bounded in the paper.

Editorial extensions

If this is right

  • Fine-tuning an LLM recommender on a GORACS-selected coreset of about 1,024 sequences can beat full-data training on ranking metrics while using roughly 20% of the total time and 15% of the floating-point operations.
  • GORACS's group-level selection outperforms both distribution-based and importance-based baselines, which fail when selection criteria are not aligned with the downstream fine-tuning loss.
  • Coreset selection time scales roughly linearly with dataset size, making the approach usable on datasets with close to a million sequences.
  • Using better text encoders to compute embedding distances improves the selected coreset, but the method stays effective across encoder choices and across LLaMA and Mistral backbones.
  • Incorporating label information per class further improves discriminative recommendation (CTR prediction), beyond the label-agnostic version.

Reading between the lines

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

  • Because the POO only needs embeddings and one forward-backward pass for gradient norms, the same two-term surrogate could be applied to other expensive LLM fine-tuning settings beyond recommendation, such as instruction tuning or domain adaptation, whenever a small validation set is available.
  • The proof of Theorem 4.2 assumes full-batch gradient descent and G-smoothness, while the actual pipeline uses LoRA with a different optimizer; if that mismatch makes the bound loose, the empirical success of the POO may be driven mainly by the OT distribution-matching term rather than by the gradient-norm term.
  • A natural test of the framework is to measure, per dataset, the Lipschitz constant of the fine-tuned loss in embedding space; datasets where that constant is small should show a tight correspondence between POO and test loss, and datasets where it is large should not.
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

3 major / 4 minor

Summary. The paper proposes GORACS, a coreset selection framework for fine-tuning LLM-based recommender systems. The method selects a small subset S of the training set by minimizing a Proxy Optimization Objective (POO) that combines an optimal transport distance between S and a validation set with a gradient-norm penalty derived from two theorems: Theorem 4.1 bounds the test loss by the training loss plus an OT term, and Theorem 4.2 bounds the training loss by a gradient-norm term. A two-stage Initialization-Then-Refinement Algorithm (ITRA) solves the resulting combinatorial problem via greedy search and exchange-based refinement with pruning. Experiments on sequential recommendation (SeqRec) and CTR prediction (CTRPre) tasks over three Amazon datasets, plus scalability tests, show that GORACS outperforms baseline coreset methods and full-data training while reducing computational cost.

Significance. If the central claims hold, GORACS would be a practically useful method for reducing the cost of fine-tuning LLM recommenders, and the paper's empirical results are encouraging: the authors report consistent gains over strong baselines, provide code, use chronological validation/test splits, and include ablations and a scalability study. These are genuine strengths. However, the paper's advertised theoretical foundation—that POO is an upper bound on the test loss—is not established by the current derivation, and the experiments, while suggestive, do not compensate for the missing verification of that bound. The contribution is therefore best viewed as an empirically strong heuristic with an incomplete theoretical justification.

major comments (3)
  1. [Section 4.1.3, Eq. (8) and Eq. (9)] The claim that minimizing the POO score S(S) minimizes an upper bound on the test loss is not justified. Theorem 4.1 assumes L_{\phi^*_S} is L-Lipschitz, but \phi^*_S is obtained by fine-tuning on S, so the constant is L(S), depending on the selected subset. Theorem 4.2 yields a constant C independent of S, but the overall bound in Eq. (8) is L(S)\cdot OT_{D^*}(\mu_S,\mu_V) - C\cdot G(S) + \Lambda. Replacing L(S) and C by a single validation-tuned \lambda in Eq. (9) is valid only if L(S)/C is constant across candidate subsets; the paper neither measures L(S) nor provides an argument that the fine-tuned loss is Lipschitz in the RoBERTa embedding metric with a stable constant. Figure 2(b) and Figure 3 show correlations, not a verification of the inequality in Eq. (8). Because the paper's central claim is that POO is an upper bound, this point is load-bearing and needs to be addressed, either by empirically checking Eq. (8) on random and selected subsets or by reframing POO as a heuristic proxy.
  2. [Section 4.1.2 and Appendix A.4] Theorem 4.2 analyzes full-batch gradient descent on all trainable parameters, whereas the experimental protocol (Appendix A.1) uses LoRA fine-tuning; moreover, the paper does not specify that training uses full-batch gradient descent, and in practice LLM fine-tuning uses adaptive optimizers. The gradient-norm term in Eq. (9) is therefore not the quantity that drives the actual training dynamics, so the bound in Eq. (8) does not directly apply to the trained model. In addition, the proof of Theorem 4.2 defines C via a minimum over S of ||\nabla H_S(\phi_0)||^2 / ((1/|S|)\sum_{z\in S}||\nabla L_{\phi_0}(z)||); this quantity is not guaranteed to be positive, since a subset with exactly cancelling initial gradients would make C = 0. The paper should either adapt the theory to the actual optimizer and training procedure, or state non-degeneracy conditions and verify them empirically.
  3. [Section 4.1.1, Theorem 4.1] Theorem 4.1 relies on two unquantified approximations: the validation set V is used as a proxy for the test distribution P, and the loss L_{\phi^*_S} is assumed to be L-Lipschitz with respect to the embedding metric d^*. Neither condition is measured or bounded. In high-dimensional text-embedding spaces, fine-tuned classifiers can have large local Lipschitz constants near decision boundaries, so the OT distance term may not control the test loss. Since the paper's abstract and Figure 1 describe POO as 'proven to be an upper bound on the test loss', the authors should either provide an empirical validation of Eq. (8) across a range of subsets and models, or soften the claim and present POO as a principled heuristic.
minor comments (4)
  1. [Section 4.1.3, Eq. (9)] The notation |\nabla_{\phi} L_{\phi_0}(z)| in Eq. (9) is not defined; the text and Eq. (10) use the gradient norm ||\nabla_{\phi} L_{\phi_0}(z)||. Please make the notation consistent.
  2. [Section 5.2, Tables 1 and 2] The main results are reported without standard deviations or repeated trials, so the claimed improvements over the second-best baselines lack a statistical significance assessment.
  3. [Section 4.2.3, Theorem 4.3] The proof of Theorem 4.3 invokes a Sensitivity Theorem asserting continuous differentiability of optimal dual variables with respect to the measure, but OT dual variables are not generally unique and the perturbation error is not bounded; please clarify under what conditions the MI estimator is accurate.
  4. [Figure 1 caption] The caption states that POO is 'proven to be an upper bound on the test loss (Section 4.1)', which is stronger than what the current theorems establish; please align the caption with the level of rigor actually achieved.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the POO bound is an assumption-based application of optimal transport duality, and the headline empirical claims are evaluated on a held-out test set.

full rationale

GORACS's derivation chain is not circular. Theorem 4.1 is a direct application of the Kantorovich-Rubinstein duality (Eq. 4): for any L-Lipschitz loss, the difference between expectations under two distributions is bounded by L times the OT distance, and the paper uses a validation set as an explicit surrogate for the test distribution. The Lipschitz and validation-approximation conditions are stated assumptions, not conclusions smuggled in through the proof, which simply says the theorem follows from Eq. 4 under those assumptions. Theorem 4.2 is a standard one-step gradient-descent smoothness bound, with the constant C defined independently of S by taking the minimum over all subsets; the bound is mathematically derived rather than assumed. Equation 9 is presented as a computationally tractable proxy for the right-hand side of Eq. 8, with the constants L, C, and Lambda folded into a tuned hyperparameter lambda; this is an approximation, not an identity, and the paper does not claim Eq. 9 is exactly Eq. 8. Moreover, lambda is tuned on the validation split, while the headline claims (GORACS beats baselines and full-data training) are measured on a chronologically separated test set, so no fitted quantity is renamed as a test prediction. The MI estimator in Theorem 4.3 is derived from OT duality and a perturbation argument, not from the target result. There are no load-bearing self-citations and no uniqueness theorem imported from the authors' earlier work. The unmeasured Lipschitz constant and the mismatch between full-batch GD analysis and LoRA training are validity concerns, but they are assumptions and approximations, not circular reductions of the claimed result to its own inputs.

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

No new physical or model entities are introduced; the framework reuses standard OT, embeddings, and gradient norms. The main input cost is the validation-tuned lambda and the unreported pruning parameters k and T.

free parameters (3)
  • lambda (POO balance weight) = searched over {0, 0.05, 0.1, 0.3, 0.5}, selected on validation
    Replaces the unknown constants L, C, Lambda in Eq. 8; the POO is not a fixed bound without this tuning.
  • k (exchange candidates for inner/outer pruning) = not reported in paper
    Algorithm 1 takes k as an input; no value is given, so replication requires guessing or reading the code.
  • T (max exchange iterations) = not reported in paper
    Algorithm 1 takes T as an input; no value is given, so the stopping condition is unspecified.
assumptions (5)
  • domain assumption The fine-tuned loss L_{\phi^*_S}(z) is L-Lipschitz in the embedding metric d*
    Theorem 4.1 assumes L-Lipschitz continuity of the fine-tuned loss in the embedding metric d*; no Lipschitz constant is measured or bounded.
  • domain assumption The validation set V faithfully approximates the test distribution P
    Section 4.1.1 uses E_{mu_V} approximately E_P; this is standard but introduces an approximation error that is not quantified.
  • ad hoc to paper Fine-tuning follows full-batch gradient descent with a G-smooth loss and learning rate eta0 < 2/G
    Eq. 6 and Theorem 4.2 assume full-batch GD; experiments use LoRA with a different optimizer and mini-batches, so the bound does not apply to the actual training procedure.
  • ad hoc to paper The constant C in Theorem 4.2 is positive and independent of S
    Theorem 4.2 defines C via a minimum over subsets of a ratio of gradient norms; C > 0 requires this minimum to be positive, which is not established.
  • standard math OT dual variables are continuously differentiable under small support perturbations (Sensitivity Theorem)
    Theorem 4.3 relies on the Sensitivity Theorem to approximate MI scores; no error bound is provided for the approximation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GORACS: Group-level Optimal Transport-guided Coreset Selection for LLM-based Recommender Systems." pith.science (2026). https://pith.science/paper/HXTPRVTE

@misc{pith2026250604015,
  author       = {Pith},
  title        = {Pith review of: GORACS: Group-level Optimal Transport-guided Coreset Selection for LLM-based Recommender Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HXTPRVTE}},
  note         = {Machine review of arXiv:2506.04015}
}
read the original abstract

Although large language models (LLMs) have shown great potential in recommender systems, the prohibitive computational costs for fine-tuning LLMs on entire datasets hinder their successful deployment in real-world scenarios. To develop affordable and effective LLM-based recommender systems, we focus on the task of coreset selection which identifies a small subset of fine-tuning data to optimize the test loss, thereby facilitating efficient LLMs' fine-tuning. Although there exist some intuitive solutions of subset selection, including distribution-based and importance-based approaches, they often lead to suboptimal performance due to the misalignment with downstream fine-tuning objectives or weak generalization ability caused by individual-level sample selection. To overcome these challenges, we propose GORACS, which is a novel Group-level Optimal tRAnsport-guided Coreset Selection framework for LLM-based recommender systems. GORACS is designed based on two key principles for coreset selection: 1) selecting the subsets that minimize the test loss to align with fine-tuning objectives, and 2) enhancing model generalization through group-level data selection. Corresponding to these two principles, GORACS has two key components: 1) a Proxy Optimization Objective (POO) leveraging optimal transport and gradient information to bound the intractable test loss, thus reducing computational costs by avoiding repeated LLM retraining, and 2) a two-stage Initialization-Then-Refinement Algorithm (ITRA) for efficient group-level selection. Our extensive experiments across diverse recommendation datasets and tasks validate that GORACS significantly reduces fine-tuning costs of LLMs while achieving superior performance over the state-of-the-art baselines and full data training. The source code of GORACS are available at https://github.com/Mithas-114/GORACS.

Figures

Figures reproduced from arXiv: 2506.04015 by the authors.

Figure 1
Figure 1. The overview of GORACS. It selects a representative coreset [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. (a) Distinct distributions of sample gradient norms [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Scatter Plots of Test Loss vs. Proxy Score on Movies [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Problem solving performance comparisons of ITRA [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: GORACS’s performance (HR@10) of varying selec [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

87 extracted references · 58 canonical work pages

  1. [1]

    Abhinab Acharya, Dayou Yu, Qi Yu, and Xumin Liu. 2024. Balancing Feature Similarity and Label Variability for Optimal Size-Aware One-shot Subset Selection. InICML

    2024

  2. [2]

    Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yancheng Luo, Fuli Feng, Xiangnan He, and Qi Tian. 2023. A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems. arXiv:2308.08434

    arXiv 2023

  3. [3]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He

  4. [4]

    Dimitri P Bertsekas. 1997. Nonlinear programming.Journal of the Operational Research Society48, 3 (1997), 334–334

    1997

  5. [5]

    Zalán Borsos, Mojmir Mutny, and Andreas Krause. 2020. Coresets via Bilevel Optimization for Continual Learning and Streaming. InNeurIPS

    2020

  6. [6]

    Yuwei Cao, Nikhil Mehta, Xinyang Yi, Raghunandan Hulikal Keshavan, Lukasz Heldt, Lichan Hong, Ed Chi, and Maheswaran Sathiamoorthy. 2024. Align- ing Large Language Models with Recommendation Knowledge. InFindings of NAACL

    2024

  7. [7]

    Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. HLLM: Enhancing Sequential Recommendations via Hierarchical Large Language Models for Item and User Modeling. arXiv:2409.12740

    arXiv 2024

  8. [8]

    Marco Cuturi. 2013. Sinkhorn Distances: Lightspeed Computation of Optimal Transport. InNeurIPS

    2013

Show all 87 references
  1. [9]

    Indraneel Das and J. Dennis. 1997. A Closer Look at Drawbacks of Minimizing Weighted Sums of Objectives for Pareto Set Generation in Multicriteria Opti- mization Problems.Structural Optimization14 (01 1997), 63–69

    1997

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and other. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. InNAACL

    2019

  3. [11]

    Lan Feng, Fan Nie, Yuejiang Liu, and Alexandre Alahi. 2024. TAROT: Targeted Data Selection via Optimal Transport. arXiv:2412.00420

    2024 arXiv

  4. [12]

    Rémi Flamary, Nicolas Courty, Alexandre Gramfort, et al. 2021. POT: Python optimal transport.J. Mach. Learn. Res.22, 1, Article 78 (Jan. 2021), 8 pages

    2021

  5. [13]

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-REC: Towards Interactive and Explainable LLMs-Augmented Recommender System. arXiv:2303.14524

    2023 arXiv

  6. [14]

    Guillaume Garrigos and Robert M. Gower. 2024. Handbook of Convergence Theorems for (Stochastic) Gradient Methods. arXiv:2301.11235

    2024 arXiv

  7. [15]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The Llama 3 Herd of Models. arXiv:2407.21783

    2024 arXiv

  8. [16]

    Chengcheng Guo, Bo Zhao, and Yanbing Bai. 2022. DeepCore: A Comprehensive Library for Coreset Selection in Deep Learning. InDEXA

    2022

  9. [17]

    Wenshuo Guo, Nhat Ho, and Michael I. Jordan. 2020. Fast Algorithms for Com- putational Optimal Transport and Wasserstein Barycenter. InAISTATS

    2020

  10. [18]

    Harsha Gwalani, Chetan Tiwari, and Armin R. Mikler. 2021. Evaluation of heuristics for the p-median problem: Scale and spatial demand distribution. Comput. Environ. Urban Syst.88 (2021), 101656

    2021

  11. [19]

    Hansen and N

    P. Hansen and N. Mladenović. 1997. Variable neighborhood search for the p- median.Location Science5, 4 (1997), 207–226

    1997

  12. [20]

    McAuley, and Wayne Xin Zhao

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian J. McAuley, and Wayne Xin Zhao. 2024. Large Language Models are Zero-Shot Rankers for Recommender Systems. InECIR

    2024

  13. [21]

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

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InICLR

    2022

  14. [22]

    Yuzheng Hu, Pingbang Hu, Han Zhao, and Jiaqi W. Ma. 2024. Most Influential Subset Selection: Challenges, Promises, and Beyond. arXiv:2409.18153

    2024 arXiv

  15. [23]

    Yitong Ji, Aixin Sun, Jie Zhang, and Chenliang Li. 2023. A Critical Study on Data Leakage in Recommender System Offline Evaluation.ACM Trans. Inf. Syst.41, 3 (2023), 75:1–75:27

    2023

  16. [24]

    Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Deven- dra Singh Chaplot, et al

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Deven- dra Singh Chaplot, et al. 2023. Mistral 7B. arXiv:2310.06825

    2023 arXiv

  17. [25]

    Meng Jiang, Keqin Bao, Jizhi Zhang, Wenjie Wang, Zhengyi Yang, Fuli Feng, and Xiangnan He. 2024. Item-side Fairness of Large Language Model-based Recommendation System. InWWW

    2024

  18. [26]

    Ayrton San Joaquin, Bin Wang, Zhengyuan Liu, Nicholas Asher, Brian Lim, Philippe Muller, and Nancy F. Chen. 2024. In2Core: Leveraging Influence Func- tions for Coreset Selection in Instruction Finetuning of Large Language Models. InFindings of EMNLP

    2024

  19. [27]

    Hoang Anh Just, Feiyang Kang, Tianhao Wang, Yi Zeng, Myeongseob Ko, Ming Jin, and Ruoxi Jia. 2023. LAVA: Data Valuation without Pre-Specified Learning Algorithms. InICLR

    2023

  20. [28]

    Cord, and Martin Volk

    Andrea Kaim, Anna F. Cord, and Martin Volk. 2018. A review of multi-criteria optimization techniques for agricultural land use allocation.Environ. Model. Softw.105 (2018), 79–93

    2018

  21. [29]

    Feiyang Kang, Hoang Anh Just, Yifan Sun, Himanshu Jahagirdar, Yuanzhi Zhang, Rongxing Du, Anit Kumar Sahu, and Ruoxi Jia. 2024. Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs. InICLR

    2024

  22. [30]

    Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InICDM

    2018

  23. [31]

    Chi, and Derek Zhiyuan Cheng

    Wang-Cheng Kang, Jianmo Ni, Nikhil Mehta, Maheswaran Sathiamoorthy, Lichan Hong, Ed H. Chi, and Derek Zhiyuan Cheng. 2023. Do LLMs Understand User Preferences? Evaluating LLMs On User Rating Prediction. arXiv:2305.06474

    2023 arXiv

  24. [32]

    Rubinstein

    Leonid Kantorovich and Gennady S. Rubinstein. 1958. On a space of totally additive functions.Vestnik Leningrad. Univ13 (1958), 52–59

    1958

  25. [33]

    KrishnaTeja Killamsetty, Durga Sivasubramanian, Ganesh Ramakrishnan, and Rishabh K. Iyer. 2021. GLISTER: Generalization based Data Subset Selection for Efficient and Robust Learning. InAAAI

    2021

  26. [34]

    KrishnaTeja Killamsetty, Xujiang Zhao, Feng Chen, and Rishabh K. Iyer. 2021. RETRIEVE: Coreset Selection for Efficient and Robust Semi-Supervised Learning. InNeurIPS

    2021

  27. [35]

    Kuehn and Michael J

    Alfred A. Kuehn and Michael J. Hamburger. 1963. A Heuristic Program for Locating Warehouses.Management Science9, 4 (1963), 643–666

    1963

  28. [36]

    Riwei Lai, Li Chen, Rui Chen, and Chi Zhang. 2024. A Survey on Data-Centric Recommender Systems. arXiv:2401.17878

    2024 arXiv

  29. [37]

    Lei Li, Yongfeng Zhang, and Li Chen. 2023. Prompt Distillation for Efficient LLM-based Recommendation. InCIKM

    2023

  30. [38]

    Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2024. From Quantity to Quality: Boosting LLM Performance with Self-Guided Data Selection for Instruction Tuning. In NAACL

    2024

  31. [39]

    Xinyu Lin, Wenjie Wang, Yongqi Li, Shuo Yang, Fuli Feng, Yinwei Wei, and Tat- Seng Chua. 2024. Data-efficient Fine-tuning for LLM-based Recommendation. In SIGIR

    2024

  32. [40]

    Junling Liu, Chao Liu, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is ChatGPT a Good Recommender? A Preliminary Study. arXiv:2304.10149

    2023 arXiv

  33. [41]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv:1907.11692

    2019 arXiv

  34. [42]

    Sebastian Lubos, Thi Ngoc Trang Tran, Alexander Felfernig, Seda Polat Erdeniz, and Viet-Man Le. 2024. LLM-generated Explanations for Recommender Systems. InUMAP

    2024

  35. [43]

    Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen, Christopher Leung, Jiajie Tang, and Jiebo Luo. 2024. LLM-Rec: Personalized Recommendation via Prompting Large Language Models. InFindings of NAACL

    2024

  36. [44]

    Adyasha Maharana, Prateek Yadav, and Mohit Bansal. 2024. D2 Pruning: Message Passing for Balancing Diversity & Difficulty in Data Pruning. InICLR

    2024

  37. [45]

    Max Marion, Ahmet Üstün, Luiza Pozzobon, Alex Wang, Marzieh Fadaee, and Sara Hooker. 2023. When Less is More: Investigating Data Pruning for Pretraining LLMs at Scale. arXiv:2309.04564

    2023 arXiv

  38. [46]

    Bilmes, and Jure Leskovec

    Baharan Mirzasoleiman, Jeff A. Bilmes, and Jure Leskovec. 2020. Coresets for Data-efficient Training of Machine Learning Models. InICML

    2020

  39. [47]

    Sung Min Park, Kristian Georgiev, Andrew Ilyas, Guillaume Leclerc, and Alek- sander Madry. 2023. TRAK: Attributing Model Behavior at Scale. InICML

    2023

  40. [48]

    Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. 2021. Deep Learning on a Data Diet: Finding Important Examples Early in Training. In NeurIPS

    2021

  41. [49]

    Gabriel Peyré and Marco Cuturi. 2019. Computational Optimal Transport.Found. Trends Mach. Learn.11, 5-6 (2019), 355–607

    2019

  42. [50]

    Omead Pooladzandi, David Davini, and Baharan Mirzasoleiman. 2022. Adaptive Second Order Coresets for Data-efficient Machine Learning. InICML

    2022

  43. [51]

    H. S. V. N. S. Kowndinya Renduchintala, Krishnateja Killamsetty, Sumit Bhatia, Mi- lan Aggarwal, Ganesh Ramakrishnan, Rishabh K. Iyer, and Balaji Krishnamurthy

  44. [52]

    Lütfi Kerem Senel, Besnik Fetahu, Davis Yoshida, Zhiyu Chen, Giuseppe Castel- lucci, Nikhita Vedula, Jason Ingyu Choi, and Shervin Malmasi. 2024. Generative Explore-Exploit: Training-free Optimization of Generative Recommender Sys- tems using LLM Optimizers. InACL

    2024

  45. [53]

    InFindings of EMNLP

    INGENIOUS: Using Informative Data Subsets for Efficient Pre-Training of Language Models. InFindings of EMNLP

  46. [54]

    Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, and Jun Xu. 2024. Large Language Models Enhanced Collaborative Filtering. In CIKM

    2024

  47. [55]

    Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari Morcos

  48. [56]

    Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. 2019. An Empirical Study of Example Forgetting during Deep Neural Network Learning. InICLR

    2019

  49. [57]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, et al . 2023. LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971

    2023 arXiv

  50. [58]

    Michael B Teitz and Polly Bart. 1968. Heuristic methods for estimating the generalized vertex median of a weighted graph.Operations research16, 5 (1968), 955–961

    1968

  51. [59]

    Jiahao Wang, Bolin Zhang, Qianlong Du, Jiajun Zhang, and Dianhui Chu. 2024. A Survey on Data Selection for LLM Instruction Tuning. arXiv:2402.05123

    2024 arXiv

  52. [60]

    Xiao Wang, Weikang Zhou, Qi Zhang, Jie Zhou, Songyang Gao, Junzhe Wang, Menghan Zhang, Xiang Gao, Yunwen Chen, and Tao Gui. 2023. Farewell to Aimless Large-scale Pretraining: Influential Subset Selection for Language Model. InFindings of ACL

    2023

  53. [61]

    Villani and American Mathematical Society

    C. Villani and American Mathematical Society. 2003.Topics in Optimal Trans- portation. American Mathematical Society. GORACS: Group-level Optimal Transport-guided Coreset Selection for LLM-based Recommender Systems KDD ’25, August 3–7, 2025, Toronto, ON, Canada

    2003

  54. [62]

    Whitaker

    R.A. Whitaker. 1983. A Fast Algorithm For The Greedy Interchange For Large- Scale Clustering And Median Location Problems.INFOR: Information Systems and Operational Research21, 2 (1983), 95–108

    1983

  55. [63]

    Jiahao Wu, Wenqi Fan, Jingfan Chen, Shengcai Liu, Qijiong Liu, Rui He, Qing Li, and Ke Tang. 2025. Condensing Pre-Augmented Recommendation Data via Lightweight Policy Gradient Estimation.IEEE Trans. Knowl. Data Eng.37, 1 (2025), 162–173

    2025

  56. [64]

    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. InWSDM

    2024

  57. [65]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A survey on large language models for recommendation.WWW(2024)

    2024

  58. [66]

    Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, et al. 2024. LESS: Selecting Influential Data for Targeted Instruction Tuning. InICML

    2024

  59. [67]

    Jiahao Wu, Wenqi Fan, Shengcai Liu, Qijiong Liu, Rui He, Qing Li, and Ke Tang

  60. [68]

    arXiv:2310.01038

    Dataset Condensation for Recommendation. arXiv:2310.01038

    arXiv

  61. [69]

    Weiwei Xiao, Yongyong Chen, Qiben Shan, Yaowei Wang, and Jingyong Su. 2024. Feature Distribution Matching by Optimal Transport for Effective and Robust Coreset Selection. InAAAI

    2024

  62. [70]

    Sang Michael Xie, Shibani Santurkar, Tengyu Ma, and Percy Liang. 2023. Data Selection for Language Models via Importance Resampling. InNeurIPS

    2023

  63. [71]

    Xiaobo Xia, Jiale Liu, Jun Yu, Xu Shen, Bo Han, and Tongliang Liu. 2023. Moderate Coreset: A Universal Method of Data Selection for Real-world Data-efficient Deep Learning. InICLR

    2023

  64. [72]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597

    2023 arXiv

  65. [73]

    Chuanyan Zhang and Xiaoguang Hong. 2021. Challenging the Long Tail Recom- mendation on Heterogeneous Information Network. InICDM

    2021

  66. [74]

    Guoxiao Zhang, Yi Wei, Yadong Zhang, Huajian Feng, and Qiang Liu. 2024. Balancing Efficiency and Effectiveness: An LLM-Infused Approach for Optimized CTR Prediction. arXiv:2412.06860

    2024 arXiv

  67. [75]

    Zhengyi Yang, Xiangnan He, Jizhi Zhang, Jiancan Wu, Xin Xin, Jiawei Chen, and Xiang Wang. 2023. A Generic Learning Framework for Sequential Recommenda- tion with Distribution Shifts. InSIGIR

    2023

  68. [76]

    Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira, Dong Wang, and Even Oldridge. 2023. LlamaRec: Two-Stage Recommendation using Large Language Models for Ranking. arXiv:2311.02089

    2023 arXiv

  69. [77]

    Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He

  70. [78]

    Aggarwal, and Tyler Derr

    Yuying Zhao, Yu Wang, Yunchao Liu, Xueqi Cheng, Charu C. Aggarwal, and Tyler Derr. 2023. Fairness and Diversity in Recommender Systems: A Survey. arXiv:2307.04644

    2023 arXiv

  71. [79]

    Xiaoyu Zhang, Juan Zhai, Shiqing Ma, Chao Shen, Tianlin Li, Weipeng Jiang, and Yang Liu. 2024. Speculative Coreset Selection for Task-Specific Fine-tuning. arXiv:2410.01296

    2024 arXiv

  72. [80]

    Yang Zhang, Keqin Bao, Ming Yan, Wenjie Wang, Fuli Feng, and Xiangnan He

  73. [81]

    Chunting Zhou, Pengfei Liu, Puxin Xu, et al . 2023. LIMA: Less Is More for Alignment. InNeurIPS. A Appendix A.1 Datasets and implementation details We conduct all experiments on four NVIDIA RTX A800 GPUs. For all the baselines and backbones, we use their open-source codes and ...

    2023

  74. [83]

    arXiv:2310.19488

    CoLLM: Integrating Collaborative Embeddings into Large Language Models for Recommendation. arXiv:2310.19488

    arXiv

  75. [85]

    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 (2024), 6889–6907

    2024

  76. [86]

    Haizhong Zheng, Rui Liu, Fan Lai, and Atul Prakash. 2023. Coverage-centric Coreset Selection for High Pruning Rates. InICLR

    2023

  77. [2022]

    InNeurIPS

    Beyond neural scaling laws: beating power law scaling via data pruning. InNeurIPS

  78. [2023]

    InRecSys

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InRecSys

  79. [2024]

    Text-like Encoding of Collaborative Information in Large Language Models for Recommendation. InACL

Pith tools

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