Pith. sign in

REVIEW 4 major objections 5 minor 73 references

TSPORec: Token Selection via Preference Optimization for LLM-Based Sequential Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read TSPORec claims that a learned token-selection policy can identify the informative tokens in item descriptions so that LLM-based sequential recommenders using only 64 selected tokens match or beat the same models using 256 prefix tokens…

desk verdict Genuinely new token-selection method with a consistent accuracy edge over the right baseline, but the efficiency headline and Theorem 1 both overreach. read the letter →

arxiv 2608.09605 v1 pith:WFLZR2JD submitted 2026-08-10 cs.IR cs.AI

Wenqiao Zhu , Chao Xu , Haipang Wu , Ji Liu This is my paper · ORCID
classification cs.IRcs.AI
keywords sequentialrecommendationtokenselectionpreferenceoptimizationlargelanguagemodelsefficientinferenceproxyrewarditemtextcompressionHLLM
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 sets out to show that LLM-based sequential recommenders waste their input budget when they take only the first few tokens of each item's text. It proposes TSPORec, a three-stage pipeline that learns which tokens actually carry preference-relevant information, then retrains the recommender on those selected tokens. The claimed result is that 64 selected tokens perform as well as or better than 256 prefix tokens, improving accuracy by up to 31.25% in NDCG and cutting inference cost by up to 63.4%. If true, this means the bottleneck in LLM recommenders is not input length but input relevance, and that learned token selection is a practical way to compress text while keeping the signal.

What carries the argument

The central object is a chunk-level token-selection policy head attached to a frozen LLM, trained by preference optimization. The proxy reward compares two independently sampled token-chunk subsets by their InfoNCE cross-entropy against the full item embedding; the subset with lower cross-entropy is treated as more informative and rewarded. The policy is trained to increase the probability of the lower-loss subset, then used to select high-probability chunks for the final retraining stage.

What would settle it

A controlled experiment replacing TSPORec's selected tokens with a random subset of the same size, or with a subset that simply drops the most frequent tokens, under the same retraining protocol; if either matches TSPORec's accuracy, the learned policy is not the source of the improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that in LLM-based sequential recommendation, the first-k token prefix used by existing methods is a poor choice: it wastes budget on filler words and misses informative content later in the item text. TSPORec learns a selection policy that assigns each token an importance score from its hidden state's alignment with the item-level representation, samples chunk-level subsets, and trains the policy with a preference-optimization reward that prefers whichever sampled subset yields lower InfoNCE cross-entropy. After training, the selected tokens replace the prefix and the model is retrained, achieving accuracy at 64 tokens that matches or exceeds HLLM at 256 tokens. The authors claim improvements up to 31.25% in NDCG and 63.4% inference overhead reduction compared with six baselines.

Load-bearing premise

The method assumes that a randomly sampled token subset that yields lower cross-entropy in the frozen pretrained model is genuinely more informative, and that raising the policy's probability for such subsets will produce token selections that improve the final retrained model.

Editorial extensions

If this is right

  • LLM-based recommenders can operate at a quarter of the input tokens with no accuracy loss, cutting inference cost dramatically.
  • The choice of which tokens are fed to an LLM matters as much as or more than how many; random token subsets almost wipe out the benefit.
  • The selected tokens favor content words and specific metadata over function words and high-frequency tokens, suggesting a general principle for text-based item representation.
  • The learned token sets transfer: they improve a different backbone (TinyLlama) and a different downstream model (LLMinit), so the selection captures something cross-model.

Reading between the lines

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

  • If the principle that rare, content-bearing tokens carry the signal holds broadly, a cheap frequency-based filter might recover much of the gain without any learned policy.
  • The same proxy-reward setup could be applied to other long-input LLM tasks where input must be truncated, such as document retrieval or long-context classification.
  • The method's success at 64 tokens suggests item descriptions are highly redundant for collaborative filtering; that redundancy could be quantified and exploited by other compression schemes.
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 TSPORec, a three-stage token-selection method for LLM-based sequential recommendation. Stage 1 pretrains an item LLM and a user LLM with an InfoNCE loss. Stage 2 freezes the backbone, attaches a policy head, and trains it with a proxy reward that compares the cross-entropy of two independently sampled token-subset sequences against the full item embedding. Stage 3 uses the learned policy to select informative token chunks and retrains the model on the compressed inputs. The authors report consistent gains over HLLM(first-k) on Amazon Books and Pixel with two backbones, and claim up to 31.25% improvement over SASRec and up to 63.4% inference-cost reduction, with code released at GitHub.

Significance. If the performance results hold, the paper makes a useful empirical contribution: it demonstrates that learned token selection can outperform naive first-k truncation for LLM-based sequential recommendation, and it provides a reusable three-stage recipe. The comparison against HLLM(first-k) is the honest baseline, and the gains there are consistent (about 5% on Amazon R@5 and about 2.9% on Pixel R@5). The release of code is a strength. However, the efficiency claim is not supported by the reported measurements, and the theoretical justification in Theorem 1 is largely definitional. The paper should be revised to correct or remove the efficiency claim and to provide statistical evidence for the modest empirical gains.

major comments (4)
  1. [4.2 (The Efficiency of TSPORec), Table 2, Eqs. (5)-(6)] The efficiency gain of 63.4%/61.3% is not supported by the reported measurements. The token-selection procedure computes info(t_i) via Eq. (5), which requires the item LLM's hidden states h_i for the full token sequence. Therefore, in any online serving scenario, TSPORec must run the item LLM over the full 256-token text to obtain the selection scores, and only then run the item LLM again on the selected 64 tokens; this full-length selection forward pass is absent from the 299 ms item-LLM time in Table 2. In the offline-precomputation scenario, the item LLM is not executed at serving time, and the user-LLM latency is constant at 26-27 ms, so a shorter item-token length does not reduce serving latency. The abstract's statement that TSPORec 'improves computational efficiency' is therefore overstated. Please report end-to-end latency that includes the selection pass, or remove the efficiency claim.
  2. [3.2, Eqs. (13)-(16), Appendix F, Theorem 1] The first claim of Theorem 1 reduces to the identity CE = KL + H and presupposes a fixed ground-truth preference distribution q that is never defined in terms of observable user preferences or the recommendation objective. The third claim defines 'informative' as 'lower cross-entropy under the frozen Stage-1 model,' so the theorem does not establish that the selected tokens improve the Stage-3 retrained model. This is a load-bearing assumption because the entire method depends on the transfer of Stage-2 selected tokens to Stage 3. To make the assumption testable, please provide a control experiment in which Stage 3 is trained on randomly selected chunks (same budget, same retraining protocol) and on chunks selected by a simple independent criterion, and report whether policy-selected chunks outperform both. Also report the variance of the three runs mentioned in Appendix A.
  3. [Appendix A, Tables 1, 3, 4, Figure 2] The paper states that each experiment is conducted three times and averages are reported, but no standard deviations, confidence intervals, or significance tests are given. The gains over the strongest baseline are modest (R@5 4.16 -> 4.37 on Amazon and 2.80 -> 2.88 on Pixel), so without variance information it is impossible to judge whether these differences are meaningful. Please add error bars or confidence intervals to the main tables and figures.
  4. [Table 1 caption and Section 4.2] The caption of Table 1 says 'text sequences truncated to 64 tokens,' but Figure 2 compares TSPORec-64 with HLLM-256, implying that TSPORec selects 64 tokens from the full text while HLLM uses the first 256 tokens. Please clarify whether TSPORec in Table 1 selects from the full description or only from a 64-token prefix, and specify the input budget for each baseline in each table. This is necessary to interpret the comparison and to reconcile Table 1 with Figure 2.
minor comments (5)
  1. [Abstract] The word 'avaliable' should be 'available'.
  2. [Section 4.2, Figure 2(a)] The text says that longer input sequences improve performance for both HLLM and TSPORec on Amazon, but Figure 2(a) shows TSPORec-64 and TSPORec-128 at the same R@5 value (4.37); please adjust the wording to reflect the flat trend.
  3. [Algorithm 1] The pseudocode uses c0 and c1 for cross-entropy values, which conflicts with the chunk-size notation c in Section 3.2; please rename the variables.
  4. [Figure 4 and Figure 6] The y-axis label 'Count' is ambiguous; specify whether the counts are across all items, across selected tokens, or across the full text corpus.
  5. [Section 4.1] Please state explicitly whether the HLLM(random) and HLLM(topk logits) baselines were trained end-to-end with those token-selection strategies or only evaluated with frozen item embeddings; this affects the interpretation of Table 1.

Circularity Check

2 steps flagged · score 4.0 of 10

Theorem 1's 'informative token' guarantees reduce to definitions of cross-entropy and the reward; empirical results remain independent.

  1. self definitional [Section 3.2 Theorem 1, Appendix F, Eqs. (17)-(20)]
    "By definition, the Kullback–Leibler (KL) divergences are given by KL(q∥p ′) = ... = ce1 −H(q), and ... = ce2 −H(q). Since H(q) is constant with respect to the model parameters, the comparison between ce1 and ce2 is equivalent to the comparison between KL(q∥p ′) and KL(q∥p ′′). In other words, a smaller cross-entropy value corresponds to a tighter approximation of the ground-truth preference distribution in terms of KL divergence."

    The theorem's first claimed property is the identity CE = KL + H(q) for an unspecified 'ground-truth preference distribution' q. The paper never defines q or shows that the softmax distribution from e'_u approximates a real preference distribution; 'tighter approximation' is simply renamed as lower CE. The reward in Eq. (15) labels the lower-CE sample as preferred, so 'informative' is operationalized as low CE under the frozen model. The theorem's first claim reduces to the definition of cross-entropy and cannot independently justify the proxy reward.

  2. self definitional [Section 3.2 Theorem 1 claim (3), Appendix F Eqs. (22)-(23)]
    "If ce1 < ce2, indicating that the sequence associated with P1 =P(I ′M |θ) yields lower cross-entropy than P2 =P(I ′′M |θ) and is thus more informative, we set the reward r= 1. In this case, the optimization objective is: L= logP2 −logP1 (22). Minimizing Eq. (22) increases the likelihood of P1 while decreasing that of P2."

    The theorem's third property restates the definition of the training objective. Eq. (15) sets r=1 precisely when the first sample has lower CE, and Eq. (16) is a policy-gradient objective over log-probability differences; minimizing it mechanically raises the lower-CE sample's likelihood. Calling the lower-CE sample 'more informative' is a labeling choice, not a derived result. No link is established between this objective and downstream Recall/NDCG after retraining, so the theorem does not establish that the selected tokens will improve the final recommender.

full rationale

TSPORec's empirical contribution is not circular: the final Recall/NDCG numbers compare a full pipeline (Stage 1 pretraining, Stage 2 policy, Stage 3 retraining) against independently implemented baselines on held-out interactions, and the selected-token generalization (Tables 3-4) is externally measured. No load-bearing self-citation chain or fitted-parameter-as-prediction was found. The main circularity is confined to Theorem 1: its two substantive-sounding claims are definitional. Property (1) is the identity CE = KL + H(q) with q never defined, so 'lower CE approximates ground truth' is true by construction. Property (3) restates the reward definition: the objective increases the likelihood of whichever sample had lower CE. These tautologies do not prove that low-CE token subsets are informative for the retrained recommender; that remains an empirical assumption. The efficiency claim raises a non-circular measurement concern: the 63.4% figure is the item-LLM time reduction, not serving latency when item embeddings are precomputed offline, and the policy's full-length scoring pass is not counted in the online comparison. I treat that as correctness risk, not circularity. Score 4 reflects partial self-definition in the theoretical justification while the empirical pipeline retains independent content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four free parameters (chunk size, prefix size, token budget, reward threshold) and on the unproven premise that the cross-entropy-based proxy reward identifies informative tokens. No new physical or conceptual entities are introduced; the 'proxy reward' is a scoring function, not an entity.

free parameters (4)
  • chunk size c = 8 (main experiments; ablation in Appendix C)
    Chosen by hand; ablation shows smaller selection chunk sizes (S1) give better performance, so the main choice is not optimized.
  • prefix size = 16 tokens
    Retained from the start of every sequence (Appendix A); chosen by hand to preserve starting tokens, affects all experiments.
  • selected token count k = 64 (also 128, 256)
    Set per experiment; the headline efficiency claim compares TSPORec-64 with HLLM-256, so this choice drives the 63.4% figure.
  • reward threshold = ±1 in Eq. (15)
    Binary reward design choice; the magnitude is arbitrary and affects the scale of policy gradients.
assumptions (4)
  • standard math InfoNCE loss (Eq. 1) is the training objective for sequential recommendation.
    Standard contrastive objective used by the paper and its baselines.
  • ad hoc to paper Lower cross-entropy for a sampled token subset indicates closer approximation to the ground-truth user preference distribution.
    This is the core premise of the proxy reward (Eq. 15) and Theorem 1 claim (1); the paper proves only that CE = KL + H, which is a definitional identity, and never defines the ground-truth distribution q.
  • domain assumption The frozen pretrained model's full-text item embedding is a valid teacher for token informativeness.
    Eq. (12) uses f_frozen(IM) as the target for reward computation; this assumes the pretrained model, trained only on prefix tokens, produces a good full-text representation.
  • domain assumption Query-key attention scores (Eq. 5) capture token informativeness.
    The policy defines importance as alignment between token hidden states and the [ITEM] hidden state; no evidence is given that this alignment corresponds to recommendation-relevant informativeness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TSPORec: Token Selection via Preference Optimization for LLM-Based Sequential Recommendation." pith.science (2026). https://pith.science/paper/WFLZR2JD

@misc{pith2026260809605,
  author       = {Pith},
  title        = {Pith review of: TSPORec: Token Selection via Preference Optimization for LLM-Based Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFLZR2JD}},
  note         = {Machine review of arXiv:2608.09605}
}
read the original abstract

Large Language Models (LLMs) have emerged as powerful tools for improving recommendation systems. The effectiveness of LLMs arises from their ability to harness rich textual information and their capacity to model heterogeneous user preferences based on users' interaction history. However, due to the large-scale and deep architectures, LLM-based sequential recommendation approaches generally incur high inference costs, resulting in a low return on investment. To mitigate this cost, many existing approaches resort to using only the first few tokens of item descriptions, which inadvertently discards valuable information contained in the full text, thereby leading to suboptimal recommendation performance. To address this limitation, we propose a novel Token Selection approach for Preference Optimization in LLM-based sequential Recommendation, i.e., TSPORec, which accurately pinpoints informative tokens throughout the entire textual content to improve recommendation performance. Specifically, we design a three-stage pipeline to select informative tokens and introduce a novel proxy reward to facilitate the implementation. TSPORec not only enhances recommendation performance but also improves computational efficiency. Extensive experiments across two models and datasets demonstrate the superb performance (up to 31.25%) and efficiency (up to 63.4%) of our approach compared with six baseline approaches. Code is available at https://github.com/WNQzhu/TSPORec.git.

Figures

Figures reproduced from arXiv: 2608.09605 by the authors.

Figure 1
Figure 1. Overview of TSPORec’s Three-Stage Training Pipeline: (1) LLM Pretraining, (2) Token Selection Policy [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Performance of HLLM and TSPORec on Amazon Books and Pixel with different token sequence lengths, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Case study on token selection for item “i9019” in the Pixel dataset, with 64 tokens selected by each [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of token distributions selected by first- [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Case study on token selection for item “i113948” in the Pixel dataset, with 64 tokens selected by each [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Comparison of token distributions selected by first- [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Performance of HLLM and TSPORec on Amazon Books with different training and selection chunk sizes, [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 35 canonical work pages

  1. [1]

    Contrastive Learning for Sequential Recommendation

    Contrastive Learning for Sequential Recommendation , author=. arXiv preprint arXiv:2010.14395 , year=

    work page Pith review arXiv 2010

  2. [2]

    arXiv preprint arXiv:2108.06479 , year=

    Contrastive Self-supervised Sequential Recommendation with Robust Augmentation , author=. arXiv preprint arXiv:2108.06479 , year=

    arXiv

  3. [3]

    2019 , booktitle =

    Sun, Fei and Liu, Jun and Wu, Jian and Pei, Changhua and Lin, Xiao and Ou, Wenwu and Jiang, Peng , title =. 2019 , booktitle =

    2019

  4. [4]

    2018 , booktitle =

    Tang, Jiaxi and Wang, Ke , title =. 2018 , booktitle =

    work page 2018

  5. [5]

    arXiv preprint arXiv:2310.20453 , year=

    Generate What You Prefer: Reshaping Sequential Recommendation via Guided Diffusion , author=. arXiv preprint arXiv:2310.20453 , year=

    arXiv

  6. [6]

    2024 , booktitle =

    Zhai, Jiaqi and Liao, Lucy and Liu, Xing and Wang, Yueming and Li, Rui and Cao, Xuan and Gao, Leon and Gong, Zhaojie and Gu, Fangda and He, Jiayuan and Lu, Yinghai and Shi, Yu , title =. 2024 , booktitle =

    work page 2024

  7. [7]

    Proceedings of the 19th International Conference on World Wide Web , pages =

    Rendle, Steffen and Freudenthaler, Christoph and Schmidt-Thieme, Lars , title =. Proceedings of the 19th International Conference on World Wide Web , pages =. 2010 , doi =

    work page 2010

  8. [8]

    and King, Irwin , title =

    Cheng, Chen and Yang, Haiqin and Lyu, Michael R. and King, Irwin , title =. 2013 , booktitle =

    work page 2013

Show all 73 references
  1. [9]

    2016 IEEE 16th International Conference on Data Mining (ICDM) , year=

    Fusing Similarity Models with Markov Chains for Sparse Sequential Recommendation , author=. 2016 IEEE 16th International Conference on Data Mining (ICDM) , year=

    2016

  2. [10]

    arXiv preprint arXiv:1511.06939 , year =

    Session-based Recommendations with Recurrent Neural Networks , author =. arXiv preprint arXiv:1511.06939 , year =

    arXiv

  3. [11]

    Proceedings of the 1st Workshop on Deep Learning for Recommender Systems , pages =

    Tan, Yong Kiam and Xu, Xinxing and Liu, Yong , title =. Proceedings of the 1st Workshop on Deep Learning for Recommender Systems , pages =. 2016 , doi =

    2016

  4. [12]

    2019 , doi =

    Zhou, Guorui and Mou, Na and Fan, Ying and Pi, Qi and Bian, Weijie and Zhou, Chang and Zhu, Xiaoqiang and Gai, Kun , title =. 2019 , doi =

    2019

  5. [13]

    2010 , booktitle =

    Rendle, Steffen and Freudenthaler, Christoph and Schmidt-Thieme, Lars , title =. 2010 , booktitle =

    2010

  6. [14]

    and King, Irwin , title =

    Cheng, Chen and Yang, Haiqin and Lyu, Michael R. and King, Irwin , title =. IJCAI , pages =. 2013 , isbn =

    2013

  7. [15]

    Attention is All you Need , year =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , year =

  8. [16]

    , booktitle =

    Kang, Wang-Cheng and McAuley, Julian J. , booktitle =. Self-Attentive Sequential Recommendation , year =

  9. [17]

    2020 , booktitle =

    Li, Jiacheng and Wang, Yujie and McAuley, Julian , title =. 2020 , booktitle =

    2020

  10. [18]

    2020 , booktitle =

    Wu, Liwei and Li, Shuqing and Hsieh, Cho-Jui and Sharpnack, James , title =. 2020 , booktitle =

    2020

  11. [19]

    Proceedings of the 17th ACM Conference on Recommender Systems , year=

    Leveraging Large Language Models for Sequential Recommendation , author=. Proceedings of the 17th ACM Conference on Recommender Systems , year=

  12. [20]

    arXiv preprint arXiv:2409.12740 , year=

    HLLM: Enhancing Sequential Recommendations via Hierarchical Large Language Models for Item and User Modeling , author=. arXiv preprint arXiv:2409.12740 , year=

    arXiv

  13. [21]

    arXiv preprint arXiv:2508.18118 , year=

    HLLM-Creator: Hierarchical LLM-based Personalized Creative Generation , author=. arXiv preprint arXiv:2508.18118 , year=

    arXiv

  14. [22]

    arXiv preprint arXiv:2409.19925 , year=

    Large Language Model Empowered Embedding Generator for Sequential Recommendation , author=. arXiv preprint arXiv:2409.19925 , year=

    arXiv

  15. [23]

    arXiv preprint arXiv:2504.19218 , year=

    AlphaFuse: Learn ID Embeddings for Sequential Recommendation in Null Space of Language Embeddings , author=. arXiv preprint arXiv:2504.19218 , year=

    arXiv

  16. [24]

    GPT-4 Technical Report , author=

  17. [25]

    arXiv preprint arXiv: 2401.02385 , year=

    TinyLlama: An Open-Source Small Language Model , author=. arXiv preprint arXiv: 2401.02385 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv: 2309.10305 , year=

    Baichuan 2: Open Large-scale Language Models , author=. arXiv preprint arXiv: 2309.10305 , year=

    arXiv

  19. [27]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 Technical Report , author=. arXiv preprint arXiv:2505.09388 , year=

    arXiv

  20. [28]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models , author=. arXiv preprint arXiv:2506.05176 , year=

    arXiv

  21. [29]

    Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    McAuley, Julian and Targett, Christopher and Shi, Qinfeng and van den Hengel, Anton , title =. Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2015 , doi =

    2015

  22. [30]

    arXiv preprint arXiv:2309.06789 , year=

    An Image Dataset for Benchmarking Recommender Systems with Raw Pixels , author=. arXiv preprint arXiv:2309.06789 , year=

    arXiv

  23. [31]

    arXiv preprint arXiv:1807.03748 , year=

    Representation Learning with Contrastive Predictive Coding , author=. arXiv preprint arXiv:1807.03748 , year=

    arXiv

  24. [32]

    arXiv preprint arXiv:2104.08691 , year=

    The power of scale for parameter-efficient prompt tuning , author=. arXiv preprint arXiv:2104.08691 , year=

    arXiv

  25. [33]

    arXiv preprint arXiv:2101.00190 , year=

    Prefix-tuning: Optimizing continuous prompts for generation , author=. arXiv preprint arXiv:2101.00190 , year=

    arXiv

  26. [34]

    arXiv preprint arXiv:2308.04275 , year=

    In-context alignment: Chat with vanilla language models before fine-tuning , author=. arXiv preprint arXiv:2308.04275 , year=

    arXiv

  27. [35]

    arXiv preprint arXiv:2401.16589 , year=

    Topro: Token-level prompt decomposition for cross-lingual sequence labeling tasks , author=. arXiv preprint arXiv:2401.16589 , year=

    arXiv

  28. [36]

    Advances in Neural Information Processing Systems , volume=

    Tuning multi-mode token-level prompt alignment across modalities , author=. Advances in Neural Information Processing Systems , volume=

  29. [37]

    ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models , year=

    Intent-based prompt calibration: Enhancing prompt optimization with synthetic boundary cases , author=. ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models , year=

    2024

  30. [38]

    arXiv preprint arXiv:2505.16307 , year=

    PMPO: Probabilistic Metric Prompt Optimization for Small and Large Language Models , author=. arXiv preprint arXiv:2505.16307 , year=

  31. [39]

    arXiv preprint arXiv:2205.12548 , year=

    Rlprompt: Optimizing discrete text prompts with reinforcement learning , author=. arXiv preprint arXiv:2205.12548 , year=

    arXiv

  32. [40]

    arXiv preprint arXiv:2209.11486 , year=

    MetaPrompting: Learning to learn better prompts , author=. arXiv preprint arXiv:2209.11486 , year=

    arXiv

  33. [41]

    International Journal of Pattern Recognition and Artificial Intelligence , volume=

    Few-Shot Text Classification with an Efficient Prompt Tuning Method in Meta-Learning Framework , author=. International Journal of Pattern Recognition and Artificial Intelligence , volume=. 2024 , publisher=

    2024

  34. [42]

    arXiv preprint arXiv:2505.17010 , year=

    Understanding prompt tuning and in-context learning via meta-learning , author=. arXiv preprint arXiv:2505.17010 , year=

  35. [43]

    Knowledge-Based Systems , volume=

    Deep reinforcement learning in recommender systems: A survey and new perspectives , author=. Knowledge-Based Systems , volume=. 2023 , publisher=

    2023

  36. [44]

    Advances in neural information processing systems , volume=

    Deep exploration via bootstrapped DQN , author=. Advances in neural information processing systems , volume=

  37. [45]

    International conference on machine learning , pages=

    Deterministic policy gradient algorithms , author=. International conference on machine learning , pages=. 2014 , organization=

    2014

  38. [46]

    Advances in neural information processing systems , volume=

    Actor-critic algorithms , author=. Advances in neural information processing systems , volume=

  39. [47]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Hierarchical reinforcement learning for course recommendation in MOOCs , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  40. [48]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI) , year=

    Hierarchical reinforcement learning for point of interest recommendation , author=. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI) , year=

  41. [49]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Reinforcement learning based path exploration for sequential explainable recommendation , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2023 , publisher=

    2023

  42. [50]

    2020 International Joint Conference on Neural Networks (IJCNN) , pages=

    Knowledge-guided deep reinforcement learning for interactive recommendation , author=. 2020 International Joint Conference on Neural Networks (IJCNN) , pages=. 2020 , organization=

    2020

  43. [51]

    2025 , publisher=

    Neural inductive biases for sequential recommendation , author=. 2025 , publisher=

    2025

  44. [52]

    Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

    KERL: A knowledge-guided reinforcement learning model for sequential recommendation , author=. Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

  45. [53]

    Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , pages=

    Large language models enhanced collaborative filtering , author=. Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , pages=

  46. [54]

    Companion Proceedings of the ACM Web Conference 2024 , pages=

    Notellm: A retrievable large language model for note recommendation , author=. Companion Proceedings of the ACM Web Conference 2024 , pages=

    2024

  47. [55]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Notellm-2: Multimodal large representation models for recommendation , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  48. [56]

    arXiv preprint arXiv:2402.13840 , year=

    LLM4SBR: A lightweight and effective framework for integrating large language models in session-based recommendation , author=. arXiv preprint arXiv:2402.13840 , year=

    arXiv

  49. [57]

    Advances in Neural Information Processing Systems , volume=

    Llm-esr: Large language models enhancement for long-tailed sequential recommendation , author=. Advances in Neural Information Processing Systems , volume=

  50. [58]

    arXiv preprint arXiv:1909.05858 , year=

    Ctrl: A conditional transformer language model for controllable generation , author=. arXiv preprint arXiv:1909.05858 , year=

    1909 arXiv

  51. [59]

    arXiv preprint arXiv:2405.17890 , year=

    SLMRec: Distilling large language models into small for sequential recommendation , author=. arXiv preprint arXiv:2405.17890 , year=

    arXiv

  52. [60]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    LM-Infinite: Zero-Shot Extreme Length Generalization for Large Language Models , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2024

  53. [61]

    NeurIPS , year=

    Direct preference optimization: Your language model is secretly a reward model , author=. NeurIPS , year=

  54. [62]

    NeurIPS , year=

    SimPO: Simple Preference Optimization with a Reference-Free Reward , author=. NeurIPS , year=

  55. [63]

    AISTATS , year=

    A General Theoretical Paradigm to Understand Learning from Human Preferences , author=. AISTATS , year=

  56. [64]

    arXiv preprint arXiv:2406.10957 , year=

    Eliminating Biased Length Reliance of Direct Preference Optimization via Down-Sampled KL Divergence , author=. arXiv preprint arXiv:2406.10957 , year=

    arXiv

  57. [65]

    2024 , journal=

    KTO: Model Alignment as Prospect Theoretic Optimization , author=. 2024 , journal=

    2024

  58. [66]

    ORPO : Monolithic Preference Optimization without Reference Model

    Hong, Jiwoo and Lee, Noah and Thorne, James , booktitle =. ORPO : Monolithic Preference Optimization without Reference Model. 2024

    2024

  59. [67]

    2024 , isbn =

    Kim, Sein and Kang, Hongseok and Choi, Seungyoon and Kim, Donghyun and Yang, Minchul and Park, Chanyoung , title =. 2024 , isbn =. doi:10.1145/3637528.3671931 , pages =

    2024

  60. [68]

    2024 , isbn =

    Liao, Jiayi and Li, Sihang and Yang, Zhengyi and Wu, Jiancan and Yuan, Yancheng and Wang, Xiang and He, Xiangnan , title =. 2024 , isbn =. doi:10.1145/3626772.3657690 , pages =

    2024

  61. [69]

    Harnessing Large Language Models for Text-Rich Sequential Recommendation

    Zhi Zheng and Chao, \ Wen Shuo\ and Zhaopeng Qiu and Hengshu Zhu and Hui Xiong. Harnessing Large Language Models for Text-Rich Sequential Recommendation. WWW 2024 - Proceedings of the ACM Web Conference. 2024. doi:10.1145/3589334.3645358

    2024

  62. [70]

    2025 , isbn =

    Zhu, Jie and Fan, Zhifang and Zhu, Xiaoxie and Jiang, Yuchen and Wang, Hangyu and Han, Xintian and Ding, Haoran and Wang, Xinmin and Zhao, Wenlin and Gong, Zhen and Yang, Huizhi and Chai, Zheng and Chen, Zhe and Zheng, Yuchao and Chen, Qiwei and Zhang, Feng and Zhou, Xun and X...

    2025

  63. [71]

    Addressing Cold-start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling , author=

  64. [72]

    SGDPO: Self-Guided Direct Preference Optimization for Language Model Alignment , author =

  65. [73]

    2022 , pages =

    Zhu, Wenqiao and Xu, Yesheng and Huang, Xin and Min, Qiyang and Zhou, Xun , title =. 2022 , pages =

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.