Pith. sign in

REVIEW 4 major objections 6 minor 50 references

K-order Ranking Preference Optimization for Large Language Models

T0 review · 4 major / 6 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 The paper claims that optimizing a K-order ranking preference—where only the top-K items are finely ordered and the rest are lumped together—gives LLM rankers better top-K accuracy than optimizing full-order or partial-order preferences…

desk verdict KPO is a sensible top-K extension of DPO with mostly convincing experiments, but the paper's headline theorem claiming strict superiority over S-DPO does not hold as written. read the letter →

arxiv 2506.00441 v1 pith:GUYL5D67 submitted 2025-05-31 cs.IR

Shihao Cai , Chongming Gao , Yang Zhang , Wentao Shi , Jizhi Zhang , Keqin Bao , Qifan Wang , Fuli Feng This is my paper
classification cs.IR
keywords K-orderrankingpreferencetop-KPlackett-LucemodeldirectoptimizationLLMlistwisecurriculumlearningquery-adaptiveK
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

This paper argues that existing list-wise preference optimization for LLM rankers targets the wrong ranking structure: partial-order methods only reward the best item, while full-order methods spend optimization effort on unreliable tail rankings. It introduces K-order Ranking Preference Optimization (KPO), which trains the model to align fine-grained order among the top-K items and to place all remaining items below them. The paper derives a Plackett-Luce-based loss for this objective, proves that its optimal policy achieves higher top-K ranking accuracy than S-DPO, and extends the method to per-query adaptive K with a curriculum over K. The reader should care because real users only inspect top-K results, so the proposed objective aligns training directly with deployment metrics like HR@K and N@K.

What carries the argument

The load-bearing object is the K-order preference model ˆp(y1≻···≻yK≻{yK+1,...,yM}|x) = ∏_{i=1}^{K} exp(r(x,yi)) / Σ_{j=i}^{M} exp(r(x,yj)), obtained by summing the Plackett-Luce likelihood over all permutations of the tail items. It unifies existing objectives as special cases: K=1 recovers S-DPO, K=M recovers DPO PL, and M=2,K=1 recovers pairwise DPO. The KPO loss in Eq. (7) trains the policy πθ against the implicit reward from DPO, and Theorem 1 compares optimal top-K accuracy of KPO versus S-DPO through the ratio wl/wk. In the query-adaptive variant, K(x) is set by counting reference-model logits above a threshold τ, and a K-aware curriculum sorts training data by increasing K so the model learns easy low-K queries first.

What would settle it

Compute the precision of the τ-thresholded reference-logit selection against ground-truth relevance per query; if selection precision is low while KPO's advantage over S-DPO persists, or if KPO trained with oracle-K does not outperform KPO trained with logit-K, the top-K explanation loses support.

Watch

Extended reading notes

Core claim

The central discovery is a new preference modeling objective for LLM ranking: instead of modeling y1≻all others or y1≻y2≻···≻yM, KPO models y1≻···≻yK≻{yK+1,...,yM}, extending the Plackett-Luce model by marginalizing over permutations of the tail items. The resulting loss, Eq. (7), increases the relative log probability of each top-K item over all of its subsequent items, preserving both fine-grained order among the top-K and the boundary between relevant and irrelevant items. Theorem 1 characterizes the optimal top-K ranking accuracy of this objective and shows it strictly exceeds that of S-DPO for l∈{2,...,K}, because the KPO weighting terms wl/wk dominate the S-DPO terms w'l/w'k. Empirically, KPO outperforms prior preference alignment methods on MovieLens, Goodreads, and Shopping Queries across backbone sizes from 135M to 3B, with query-adaptive K and K-aware curriculum learning providing additional gains.

Load-bearing premise

The central assumption is that the reference model's logits, cut off at threshold τ, correctly identify the truly relevant top-K items for every query; if those logits are miscalibrated or miss relevant tail items, the training sequence y1≻...≻yK is mislabeled and the measured gain may come from the curriculum rather than from accurate top-K alignment.

Editorial extensions

If this is right

  • If KPO's central claim is correct, list-wise preference alignment for LLM rankers should be reformulated around top-K consistency rather than exhaustive full-order ranking, since tail relations add optimization burden without matching user attention.
  • The theoretical result implies that replacing S-DPO's partial-order objective with KPO strictly increases the optimal top-K ranking accuracy under Plackett-Luce ground-truth preferences.
  • Query-adaptive K, estimated from reference-model logits and threshold τ, means the method needs no fixed list-length annotation and can vary the number of relevant items per query.
  • K-aware curriculum learning, which orders training examples by ascending K, yields smoother validation curves and higher final metrics than random or descending orderings.
  • The extra K-layer loop in the loss adds only about 2% wall-clock time over S-DPO, so the accuracy gain does not come at meaningful computational cost.

Reading between the lines

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

  • Editorial inference: the same top-K truncation idea could be applied to other listwise objectives beyond DPO-style losses, such as listwise softmax or NDCG-based losses, and the paper's experiments suggest that such truncation would concentrate capacity on head items.
  • Editorial inference: the logit-threshold heuristic for K is a calibration assumption; a learnable per-query K predictor trained on ground-truth relevance would be a natural next step and could remove the sensitivity to τ that the paper acknowledges.
  • Editorial inference: because the theorem assumes ground-truth Plackett-Luce probabilities, the most direct test of the explanation would compare KPO trained with oracle-K (from true labels) against KPO trained with logit-K; if logit-K still wins, the gain may come from the curriculum rather than accurate top-K selection.
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 / 6 minor

Summary. The paper proposes K-order Ranking Preference Optimization (KPO), a listwise preference alignment method for LLM ranking that optimizes the partial ranking y1 ≻ ... ≻ yK ≻ {yK+1, ..., yM}. The authors derive a Plackett-Luce-based loss (Eq. 7), claim a theoretical guarantee that KPO's optimal top-K ranking accuracy exceeds that of S-DPO (Theorem 1), and extend the method with query-adaptive K selection via a logit threshold and a K-aware curriculum. Experiments on MovieLens, Goodreads, and Shopping Queries with Llama-3.2 and other backbones are reported as showing consistent improvements over preference alignment baselines and other rankers.

Significance. If the theoretical comparison were sound, the paper would make a useful conceptual contribution: it identifies top-K ranking consistency as a middle ground between S-DPO's partial-order and DPO-PL's full-order objectives, and it gives a clean likelihood-based loss for that objective. The derivation of Eq. (6) as a marginalization of the Plackett-Luce model is correct, and the KPO loss is a natural DPO-style objective for top-K rankings. The paper also ships code and evaluates across multiple datasets and backbone sizes, which strengthens reproducibility. However, the central theoretical claim vis-a-vis S-DPO is not established as stated, and the empirical evidence lacks statistical grounding, so the significance is currently conditional.

major comments (4)
  1. [Section 4.2, Eq. (12); Appendix A.3, Eq. (30)] The proof that KPO's optimal ranking accuracy exceeds S-DPO's fixes w'_l/w'_k = 1 for all l in {2,...,K}, but this is an arbitrary tie-breaking. The S-DPO objective in Table 2 (row 4) depends on rewards only through the differences r_j - r_1 for j >= 2, so any policy with r_1 sufficiently larger than all other rewards is a loss minimizer; in particular, the full ground-truth Plackett-Luce policy pi* proportional to pi_ref exp(r*/beta) is also optimal for S-DPO and yields w'_l/w'_k equal to the KPO ratio in Eq. (10). The strict inequality in Eq. (34) is therefore an artifact of the chosen representative policy, not a consequence of the S-DPO objective, and the claim that the optimal top-K ranking accuracy of KPO exceeds that of S-DPO is unsupported as stated.
  2. [Appendix A.2, Eq. (22)] The proof of Theorem 1 assumes that the policy maximizing the KPO objective matches the ground-truth PL probability alpha(x,y_i,y_>i) for every position i, including the tail items i > K. The KPO loss in Eq. (7) depends on the tail rewards only through sums of the form sum_{j=K+1}^M exp(r_j) inside the denominator terms, so tail-internal probabilities are not identified by the objective; infinitely many optimal policies with different tail orderings satisfy the same KPO loss. Since the top-K accuracy in Eq. (9) includes indicators comparing y_l with individual tail items y_k for k > K, the theorem needs an explicit identifiability assumption on the tail or a precise statement of which policy in the optimal set is being evaluated.
  3. [Tables 1, 3, and 5] All empirical tables report single point estimates without error bars, number of random seeds, or significance tests. Several differences are small (e.g., Table 1 MovieLens HR@1: KPO 0.5579 vs. DPOPL 0.5474; Goodreads HR@10: KPO 0.9584 vs. DPOPL 0.9634, where KPO is lower). The claim that KPO 'consistently outperforms' prior methods is therefore not statistically supported. Please report means and standard deviations over multiple seeds and, where appropriate, paired significance tests.
  4. [Section 4.3.3, Eq. (14)] The query-adaptive K is obtained by thresholding reference-model logits with a fixed hyperparameter tau, and the paper's own Limitations section states that this method is heuristic, that logits may not accurately measure relevance, and that K is highly sensitive to tau. The robustness experiment in Fig. 5b only swaps pairs of logits; it does not examine miscalibrated logits, threshold shifts, or the failure mode where relevant items fall below tau. The stated robustness of KPO to noisy logits is therefore narrower than claimed. Please either add experiments that vary the threshold relative to the logit scale or soften the robustness claim.
minor comments (6)
  1. [Theorem 1 statement and Appendix A.2] The phrase 'ground-truth ranking probabilitie' contains a typo, and Eq. (30) uses alpha(x,y_l,y_>=l) while Eq. (17) defines alpha with y_>i; the notation should be made consistent.
  2. [Tables 2, 3, and 7] The model name is written inconsistently as 'KPOCUT', 'KPO CUT', and 'KPO CL' in different places; please unify the notation.
  3. [Appendix C.2.2] The text says 'The detail description of the models can be found in Appendix C.2.2 and C.2.2'; the second cross-reference should be to Appendix C.2.3.
  4. [Table 2 and Appendix A.4] The KTO row in Table 2 uses lambda_y, but Appendix A.4 defines the loss with lambda_D and lambda_U; please align these notations.
  5. [Appendix A.1] The notation 'Per(y_{K+1},...,y_M)' is used for the set of permutations of the tail items but is not defined; please define it explicitly.
  6. [Figure 4] The two panels of Figure 4 appear to use different y-axis ranges for the same metrics, which can make the stability of HR@10 look artificially different from HR@1; consider using shared axes or clearly annotating each panel.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1's KPO-over-S-DPO superiority is an artifact of Eq. (12)'s arbitrary definition of S-DPO tail ratios; empirical evaluations remain self-contained.

  1. self definitional [Section 4.2, Eq. (12); Appendix A.3, Eqs. (30)-(34)]
    "where w′l/w′k is defined as ... (α(x, yl, y>l)/α(x, yk, y>k))^{1/β} · ∏_{i=l}^{k-1}(1 − α(x, yi, y>i))^{−1/β} · I[l = 1] + I[l ̸= 1]. ... Based on Eq. (10) and Eq. (12), we can conclude that: wl/wk > w′l/w′k for all l ∈ {2, . . . , K} and k ∈ {l + 1, . . . , M}."

    Eq. (12)/(30) defines S-DPO's optimal tail ratios to be 1 for l≥2, i.e., π*(yl|x)/π*(yk|x) = πref(yl|x)/πref(yk|x). The S-DPO objective in Table 2 only constrains the first-choice probability; it does not force tail ratios to the reference level. The full PL ground-truth policy, whose ratios are KPO's Eq. (10), is also an S-DPO loss minimizer and attains R*_S-DPO = R*_KPO. Thus the strict inequality in Eq. (34) is an artifact of the arbitrary indicator definition rather than a consequence of the S-DPO objective: the claimed theoretical superiority is baked into the definition.

full rationale

The KPO loss derivation itself is a standard Plackett-Luce + DPO reparameterization and is not circular: Eq. (7) follows from Eq. (6) and the DPO reward form, with no target result assumed as input. The query-adaptive K procedure is a disclosed heuristic, and the Limitations section explicitly admits that logit-based K selection is not guaranteed to be optimal; this is a robustness caveat, not circularity. The definitional problem is isolated to the theoretical comparison in Theorem 1 and Appendix A.3. There, Eq. (12) constructs S-DPO's optimal accuracy using w′_l/w′_k = 1 for all non-first tail pairs, which is not implied by the S-DPO objective. A valid S-DPO optimum given by the ground-truth PL policy reproduces KPO's Eq. (10) ratios, so the claimed strict superiority of KPO is forced by the chosen definition rather than derived from the S-DPO objective. No load-bearing self-citation is involved: the proof's cited method (Chen et al., 2024a) is external, and the experimental comparisons are against external baselines and datasets. Because the central theoretical claim reduces by construction while the empirical KPO contribution remains independent, a score of 6 is appropriate.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; its novelty is confined to a training objective. The main free parameters are the loss temperature β and the logit threshold τ, both central to the query-adaptive variant. The theoretical result rests on PL-model assumptions about the data and the optimal policy.

free parameters (2)
  • β (loss temperature in Eq. 13) = 1.0 (tuned from {0.1, 0.5, 1.0, 3.0, 5.0})
    Scales the log-probability ratios in the KPO loss. The authors select 1.0 based on MovieLens validation results in Fig. 4.
  • τ (logit relevance threshold in Eq. 14) = 24 (tuned from {18, 20, 22, 24, 26})
    Determines which candidates count as relevant, thereby setting K(x) and the top-K training labels for the query-adaptive variant. Tuned on validation; the Limitations note that K is sensitive to this hyperparameter.
assumptions (5)
  • domain assumption Ground-truth preferences in Dp follow a Plackett-Luce model with probabilities α(x, y_i, y_{>i})
    Assumed in Theorem 1 (Section 4.2, Eq. 8) to derive the optimal top-K accuracy formula. Real user feedback may not follow PL, so the theorem's applicability to practice is conditional.
  • domain assumption After optimization, the model's induced ranking probabilities exactly match the ground-truth α
    Equation (22) in Appendix A.2 assumes the optimal policy achieves the ground-truth PL ranking probabilities. This is a standard DPO consistency assumption but is not empirically verified.
  • ad hoc to paper LLM logits are usable as relevance scores, and thresholding them with τ yields the true relevant set
    Used to construct K(x) in Eq. (14) and to select top-K training items. The authors themselves call this heuristic in the Limitations and do not prove it holds.
  • domain assumption Users care only about top-K results, and tail items lack reliable preference feedback
    This motivates the K-order objective in Section 1. It is a behavioral assumption about ranking tasks, not tested in the paper.
  • standard math The sum over all tail permutations of the Plackett-Luce probability equals 1
    Used in Appendix A.1 to derive Eq. (6). This is the normalization property of the PL distribution over permutations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of K-order Ranking Preference Optimization for Large Language Models." pith.science (2026). https://pith.science/paper/GUYL5D67

@misc{pith2026250600441,
  author       = {Pith},
  title        = {Pith review of: K-order Ranking Preference Optimization for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUYL5D67}},
  note         = {Machine review of arXiv:2506.00441}
}
read the original abstract

To adapt large language models (LLMs) to ranking tasks, existing list-wise methods, represented by list-wise Direct Preference Optimization (DPO), focus on optimizing partial-order or full-order list ranking consistency for LLMs to enhance their ranking abilities. However, we argue that optimizing top-K ranking consistency could be more appropriate for real-world applications. There are two main reasons: (1) users are typically concerned with only the top-K results, making top-K ranking more important, and (2) tail items often lack precise feedback, making top-K ranking more reliable. Based on this, we propose K-order Ranking Preference Optimization (KPO) by extending the DPO's Plackett-Luce model to accommodate top-K rankings. Additionally, recognizing that the number of important items can vary across queries, we extend KPO to dynamically determine appropriate K for different samples and introduce a curriculum learning strategy to boost training efficiency. Extensive experiments demonstrate the effectiveness of KPO, highlighting its high sample efficiency and robustness to noise. The code is available at https://github.com/Lanyu0303/KPO.

Figures

Figures reproduced from arXiv: 2506.00441 by the authors.

Figure 1
Figure 1. (a) Illustration of the LLM-based ranking task. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The KPO framework consists of: (1) selecting [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. N@5 on MovieLens and Goodreads validation [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: (a) Comparison of the reward of the top-1 item [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 4
Figure 4. Figure 4: Study of β, τ on MovieLens validation set. 5.4 In-depth Analysis of KPO (RQ3) In this section, we conduct an in-depth analysis of KPO from three key perspectives: (1) sample efficiency, and (2) robustness to noisy logits (3) applicability across various backbone models…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 17 canonical work pages

  1. [1]

    Gediminas Adomavicius and Jingjing Zhang. 2016. https://doi.org/10.1287/IJOC.2015.0662 Classification, ranking, and top-k stability of recommendation algorithms . INFORMS J. Comput. , 28(1):129--147

    work page arXiv 2016

  2. [2]

    Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martín Blázquez, Guilherme Penedo, Lewis Tunstall, Andrés Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémentine Fourrier, Ben Burtenshaw, Hugo Larcher, Haojun Zhao, Cyril Zakka, Mathieu Morlon, Colin Raffel, Leandro von Werr...

    arXiv 2025

  3. [3]

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41--48

    2009

  4. [4]

    Ralph Allan Bradley and Milton E Terry. 1952. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324--345

    1952

  5. [5]

    Zhang, Xinyi Chen, Qiuyi Zhang, Rajesh Ranganath, and Kyunghyun Cho

    Angelica Chen, Sadhika Malladi, Lily H. Zhang, Xinyi Chen, Qiuyi Zhang, Rajesh Ranganath, and Kyunghyun Cho. 2024 a . https://doi.org/10.48550/ARXIV.2405.19534 Preference learning algorithms do not learn preference rankings . CoRR, abs/2405.19534

    arXiv doi:10.48550/arxiv.2405.19534 2024

  6. [6]

    Yiqun Chen, Qi Liu, Yi Zhang, Weiwei Sun, Xinyu Ma, Wei Yang, Daiting Shi, Jiaxin Mao, and Dawei Yin. 2025. https://doi.org/10.1145/3696410.3714863 Tourrank: Utilizing large language models for documents ranking with a tournament-inspired strategy . In Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 20...

    arXiv 2025

  7. [7]

    Yuxin Chen, Junfei Tan, An Zhang, Zhengyi Yang, Leheng Sheng, Enzhi Zhang, Xiang Wang, and Tat - Seng Chua. 2024 b . https://doi.org/10.48550/ARXIV.2406.09215 On softmax direct preference optimization for recommendation . CoRR, abs/2406.09215

    arXiv doi:10.48550/arxiv.2406.09215 2024

  8. [8]

    Paolo Ciaccia and Davide Martinenghi. 2024. https://doi.org/10.1145/3698807 Directional queries: Making top-k queries more effective in discovering relevant results . Proc. ACM Manag. Data , 2(6):232:1--232:26

    work page doi:10.1145/3698807 2024

Show all 50 references
  1. [9]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. https://doi.org/10.48550/ARXIV.2402.01306 KTO: model alignment as prospect theoretic optimization . CoRR, abs/2402.01306

    2024 doi arXiv

  2. [10]

    Chenhao Fang, Xiaohan Li, Zezhong Fan, Jianpeng Xu, Kaushiki Nag, Evren K \" o rpeoglu, Sushant Kumar, and Kannan Achan. 2024. https://doi.org/10.1145/3626772.3661357 Llm-ensemble: Optimal large language model ensemble method for e-commerce product attribute value extraction ....

    2024

  3. [11]

    Chongming Gao, Ruijun Chen, Shuai Yuan, Kexin Huang, Yuanqing Yu, and Xiangnan He. 2025 a . https://doi.org/10.1145/3696410.3714524 Sprec: Self-play to debias llm-based recommendation . In Proceedings of the ACM Web Conference 2025, WWW 2025

    2025

  4. [12]

    Chongming Gao, Mengyao Gao, Chenxiao Fan, Shuai Yuan, Wentao Shi, and Xiangnan He. 2025 b . Process-supervised llm recommenders via flow-guided tuning. In Proceedings of the 48th international ACM SIGIR conference on research and development in information retrieval, SIGIR 2025

    2025

  5. [13]

    Muhammad Usman Hadi, Rizwan Qureshi, Abbas Shah, Muhammad Irfan, Anas Zafar, Muhammad Bilal Shaikh, Naveed Akhtar, Jia Wu, Seyedali Mirjalili, et al. 2023. A survey on large language models: Applications, challenges, limitations, and practical usage. Authorea Preprints

    2023

  6. [14]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. 2016. https://doi.org/10.1145/2827872 The movielens datasets: History and context . ACM Trans. Interact. Intell. Syst. , 5(4):19:1--19:19

    2016 doi

  7. [15]

    Bal \' a zs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2016. Session-based recommendations with recurrent neural networks. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Pr...

    2016

  8. [16]

    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. https://openreview.net/forum?id=nZeVKeeFYf9 Lora: Low-rank adaptation of large language models . In The Tenth International Conference on Learning Representat...

    2022

  9. [17]

    DANIEL Kai-Ineman and Amos Tversky. 1979. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):363--391

    1979

  10. [18]

    Wang - Cheng Kang and Julian J. McAuley. 2018. Self-attentive sequential recommendation. In IEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018 , pages 197--206. IEEE Computer Society

    2018

  11. [19]

    Wonbin Kweon, SeongKu Kang, Sanghwan Jang, and Hwanjo Yu. 2024. https://doi.org/10.1145/3589334.3645417 Top-personalized-k recommendation . In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024 , pages 3388--3399. ACM

    2024

  12. [20]

    Le and Hady Wirawan Lauw

    Dung D. Le and Hady Wirawan Lauw. 2021. https://doi.org/10.1613/JAIR.1.12403 Efficient retrieval of matrix factorization-based top-k recommendations: A survey of recent approaches . J. Artif. Intell. Res., 70:1441--1479

    2021 doi

  13. [21]

    Jinhyuk Lee, Zhuyun Dai, Sai Meher Karthik Duddu, Tao Lei, Iftekhar Naim, Ming - Wei Chang, and Vincent Zhao. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/31d997278ee9069d6721bc194174bb4c-Abstract-Conference.html Rethinking the role of token retrieval in multi-vect...

    2023

  14. [22]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. https://doi.org/10.1145/3626772.3657690 Llara: Large language-recommendation assistant . In Proceedings of the 47th International ACM SIGIR Conference on Research and Development...

    2024

  15. [23]

    Sichun Luo, Bowei He, Haohan Zhao, Wei Shao, Yanlin Qi, Yinya Huang, Aojun Zhou, Yuxuan Yao, Zongpeng Li, Yuanzhang Xiao, et al. 2024. Recranker: Instruction tuning large language model as ranker for top-k recommendation. ACM Transactions on Information Systems

    2024

  16. [24]

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2024. https://doi.org/10.1145/3626772.3657951 Fine-tuning llama for multi-stage text retrieval . In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGI...

    2024

  17. [25]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. https://doi.org/10.48550/ARXIV.2405.14734 Simpo: Simple preference optimization with a reference-free reward . CoRR, abs/2405.14734

    2024 doi arXiv

  18. [26]

    Meta. 2024. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices Llama 3.2: Revolutionizing edge ai and vision with open, customizable models

    2024

  19. [27]

    Eric Mitchell. 2023. https://ericmitchell.ai/cdpo.pdf A note on dpo with noisy preferences and relationship to ipo

    2023

  20. [28]

    OpenAI. 2022. Introducing chatgpt. https://openai.com/blog/chatgpt

    2022

  21. [29]

    OpenAI. 2023. Gpt-4 technical report. ArXiv, abs/2303.08774

    2023 arXiv

  22. [30]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leik...

    2022

  23. [31]

    Robin L Plackett. 1975. The analysis of permutations. Journal of the Royal Statistical Society Series C: Applied Statistics, 24(2):193--202

    1975

  24. [32]

    Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. 2023 a . https://doi.org/10.48550/ARXIV.2309.15088 Rankvicuna: Zero-shot listwise document reranking with open-source large language models . CoRR, abs/2309.15088

    2023 doi arXiv

  25. [33]

    Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. 2023 b . https://doi.org/10.48550/ARXIV.2312.02724 Rankzephyr: Effective and robust zero-shot listwise reranking is a breeze! CoRR, abs/2312.02724

    2023 doi arXiv

  26. [34]

    Zhen Qin, Rolf Jagerman, Kai Hui, Honglei Zhuang, Junru Wu, Le Yan, Jiaming Shen, Tianqi Liu, Jialu Liu, Donald Metzler, Xuanhui Wang, and Michael Bendersky. 2024. https://doi.org/10.18653/V1/2024.FINDINGS-NAACL.97 Large language models are effective text rankers with pairwise...

    2024 doi

  27. [35]

    Manning, Stefano Ermon, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.html Direct preference optimization: Your language model is sec...

    2023

  28. [36]

    Reddy, Llu \' s M \` a rquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian

    Chandan K. Reddy, Llu \' s M \` a rquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. 2022. https://doi.org/10.48550/ARXIV.2206.06588 Shopping queries dataset: A large-scale ESCI benchmark for improving product s...

    2022 doi arXiv

  29. [37]

    Rothschild, Daniel G

    Sofia Eleni Spatharioti, David M. Rothschild, Daniel G. Goldstein, and Jake M. Hofman. 2023. https://doi.org/10.48550/ARXIV.2307.03744 Comparing traditional and llm-based search for consumer choice: A randomized experiment . CoRR, abs/2307.03744

    2023 doi arXiv

  30. [38]

    Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.923 Is chatgpt good at search? investigating large language models as re-ranking agents . In Proceedings of the 2023 Co...

    2023 doi

  31. [39]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, WSDM 2018, Marina Del Rey, CA, USA, February 5-9, 2018 , pages 565--573. ACM

    2018

  32. [40]

    Qwen Team. 2024. https://qwenlm.github.io/blog/qwen2.5/ Qwen2.5: A party of foundation models

    2024

  33. [41]

    Amos Tversky and Daniel Kahneman. 1992. Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and uncertainty, 5:297--323

    1992

  34. [42]

    Mengting Wan and Julian J. McAuley. 2018. https://doi.org/10.1145/3240323.3240369 Item recommendation on monotonic behavior chains . In Proceedings of the 12th ACM Conference on Recommender Systems, RecSys 2018, Vancouver, BC, Canada, October 2-7, 2018 , pages 86--94. ACM

    2018

  35. [43]

    Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and Tat-Seng Chua. 2025. Personalized generation in large model era: A survey. arXiv preprint arXiv:2503.02614

    2025 arXiv

  36. [44]

    Zhengyi Yang, Xiangnan He, Jizhi Zhang, Jiancan Wu, Xin Xin, Jiawei Chen, and Xiang Wang. 2023. A generic learning framework for sequential recommendation with distribution shifts. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Inf...

    2023

  37. [45]

    Soyoung Yoon, Eunbi Choi, Jiyeon Kim, Hyeongu Yun, Yireun Kim, and Seung - won Hwang. 2024. https://doi.org/10.18653/V1/2024.ACL-LONG.125 Listt5: Listwise reranking with fusion-in-decoder improves zero-shot retrieval . In Proceedings of the 62nd Annual Meeting of the Associati...

    2024 doi

  38. [46]

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to go next for recommender systems? ID- vs. modality-based recommender models revisited. In Proceedings of the 46th International ACM SIGIR Conference on Research and Deve...

    2023

  39. [47]

    Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira, Dong Wang, and Even Oldridge. 2023. https://doi.org/10.48550/ARXIV.2311.02089 Llamarec: Two-stage recommendation using large language models for ranking . CoRR, abs/2311.02089

    2023 doi arXiv

  40. [48]

    Shengyao Zhuang, Honglei Zhuang, Bevan Koopman, and Guido Zuccon. 2024. https://doi.org/10.1145/3626772.3657813 A setwise approach for effective and highly efficient zero-shot ranking with large language models . In Proceedings of the 47th International ACM SIGIR Conference on...

    2024

  41. [49]

    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. [50]

    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.