Bidirectional Knowledge Distillation for Enhancing Sequential Recommendation with Large Language Models

Dongsheng Li; Guangping Zhang; Hansu Gu; Jiahao Liu; Jiongran Wu; Li Shang; Mingzhe Han; Ning Gu; Peng Zhang; Tun Lu

REVIEW 3 major objections 5 minor 59 references

Bidirectional Knowledge Distillation for Enhancing Sequential Recommendation with Large Language Models

T0 review · 3 major / 5 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 LLMD4Rec claims that a lightweight SASRec model and an LLM-based recommender improve each other by alternately distilling their softened output distributions, so the small model can approach LLM-level accuracy at inference time with no…

desk verdict A plausible mutual-distillation recipe with strong reported numbers, but the semantic-knowledge attribution and the empirical reporting need hardening before the results are established. read the letter →

arxiv 2505.18120 v1 pith:VLCMU65G submitted 2025-05-23 cs.IR cs.AI

Jiongran Wu , Jiahao Liu , Dongsheng Li , Guangping Zhang , Mingzhe Han , Hansu Gu , Peng Zhang , Li Shang

show 2 more authors

Tun Lu Ning Gu

This is my paper · ORCID

classification cs.IRcs.AI

keywords sequentialrecommendationlargelanguagemodelsknowledgedistillationmutualbidirectionaltransferoutputdistributionalignmentsample-wiseadaptiveweightingrecommendersystems

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 proposes LLMD4Rec, a mutual distillation framework in which a conventional sequential recommender (SASRec) and an LLM-centric recommender (E4SRec, built on a frozen LLM) take turns teaching each other. The claim is that this bidirectional, iterative exchange of softened output distributions transfers semantic understanding from the LLM to the small model and collaborative signals from the small model to the LLM, improving both beyond any unidirectional baseline. Because the distillation operates only through loss terms on existing outputs, no extra parameters are added and inference cost is unchanged. On four real-world datasets, the framework reports consistent gains in HR@K and NDCG@K for both backbones, with the CRM variant improving the most relative to its category baselines.

What carries the argument

The engine is a mutual distillation loop built on prediction-layer output distributions: each model's softmax over all items, sharpened by a temperature below 1, is used as the distillation target for the other model via KL divergence. The two losses are $L^\downarrow_{KL}$ (CRM distribution guiding the LLM) and $L^\uparrow_{KL}$ (LLM distribution guiding the CRM), each multiplied by a sample-wise weight in $[0,2]$ derived from rank disagreement. Alternating between freezing one model and training the other transfers knowledge in a CRM-to-LLM-to-CRM cycle without modifying model architectures.

What would settle it

Run the same two-way loop but replace the teacher's softened scores with randomly shuffled scores that preserve the target item's rank; if HR@K and NDCG@K still climb, the gains come from the loop's regularization rather than from semantic or collaborative knowledge carried by the aligned distributions.

Watch

Extended reading notes

Core claim

The central discovery is that a loop of two KL-divergence distillations between a conventional recommender and an LLM-centric recommender creates a virtuous cycle: the LLM-centric model learns from the CRM's item-ID embeddings and output distribution, then becomes the teacher that refines the CRM's softened score distribution, and the process repeats. The direction from CRM to LLM is called downward enhancement and carries collaborative signals; the direction from LLM to CRM is called upward semantics distillation and carries semantic understanding. A sample-wise adaptive weight, computed from the relative ranks of the target item in the two models' predictions, scales each distillation loss so the student learns more when it is weak on that sample and less when it already outperforms the teacher. The paper reports that after two alternating rounds, both the SASRec-based variant and the E4SRec-based variant beat their respective baselines on Beauty, Sports, Toys, and Yelp, with no additional trainable parameters and no change to inference latency.

Load-bearing premise

The framework assumes that after temperature scaling, the softened item-score distributions of the SASRec-based CRM and the LLM-based recommender are comparable enough that aligning them transfers genuine knowledge rather than noise, and the paper does not calibrate the two score spaces or ablate this assumption.

Editorial extensions

If this is right

A lightweight CRM can approach or exceed LLM-enhanced baselines while remaining cheap to serve, since the LLM is not needed at inference time.
An LLM-centric recommender can absorb collaborative signals from a conventional recommender, improving its ranking accuracy beyond training the LLM alone.
The improvement appears after only about two alternating training rounds, so the mutual-distillation loop is a practical training-time add-on rather than a long iterative process.
Because no architectures or parameter counts change, the resulting models can be deployed in existing serving infrastructure with the same latency and storage footprint as the original backbones.

Reading between the lines

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

Beyond the paper: the same output-space mutual-distillation loop should transfer between any two recommenders that share an item vocabulary, including CRM-CRM or LLM-LLM pairs, so a control experiment with two CRMs would isolate whether the gains require an LLM at all.
Beyond the paper: the adaptive weight is based only on the ground-truth item's rank, which may distort learning on hard long-tail samples; weighting by distributional confidence or calibrated uncertainty is a testable extension.
Beyond the paper: the Yelp benchmark has a missing citation marker in the manuscript, so one of the four evaluation datasets has unverified provenance; this does not affect the mechanism claim but should be checked.
Beyond the paper: the temperatures $T_1=0.6$ and $T_2=0.2$ are fixed across datasets, and sensitivity to these values as catalog size grows is not studied, so temperature tuning may be needed for very large item sets.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Plainly: this is the first application I know of that runs mutual distillation as an alternating loop between an LLM-centric recommender and a conventional sequential recommender, and it does so with a clean output-distribution alignment plus a rank-based sample weight. That is a legitimate new combination, and the reported gains over SASRec and E4SRec are large and consistent across four datasets. The idea of letting each model teach the other in turn, with no extra inference cost, is practically attractive and clearly presented.

What the paper does well: the framework is simple to describe, the equations are complete, the baselines include both CRM-side and LLM-side distillation methods, and the ablations show that the loop and the weighting both matter. The training-efficiency comparison, whatever its labeling problems, addresses a real deployment question.

Soft spots, in rough order of severity. First, the attribution problem is real. The loop starts from a pretrained CRM, fine-tunes the LLM with KL to that CRM, then fine-tunes the CRM with KL to the LLM, twice. The teachers are not independent. With T_max=2, the gains could come mostly from extra iterative training and self-regularization rather than from the LLM's semantic knowledge. The ablations never replace the LLM teacher with a CRM-only self-distillation control, nor do they compare SASRec trained for the same number of epochs with an equivalently softened auxiliary loss. So the paper's central story — that LLM semantics drive the improvement — is not isolated.

Second, the empirical evidence is under-reported. No error bars or significance tests, no code or data release, and Table 2 appears to be single runs. Third, there are real manuscript inconsistencies: the efficiency table's rows are labeled LLMD4Rec_C while the text claims LLMD4Rec_L, and the Hit@10 values in that table do not match Table 2; the Yelp dataset citation is missing; Figure 4's caption looks like a leftover from the framework overview. These are fixable but they lower confidence.

None of this makes the method incoherent. The mutual-distillation recipe is plausible and the direction is worth pursuing. What's missing is evidence that the semantic-knowledge story is the right explanation, and that the numbers are reproducible.

Who this is for: anyone working on LLM distillation for sequential recommendation, or on efficient deployment of LLM-based recommenders. It deserves a serious referee, with the expectation of major revisions: code and data, multi-seed variance, a non-LLM self-distillation control, and a corrected efficiency table. I would not cite the results as established until that evidence appears.

Referee Report

3 major / 5 minor

Summary. The paper proposes LLMD4Rec, a mutual distillation framework that alternately trains a conventional sequential recommendation model (SASRec as the CRM) and an LLM-centric recommendation model (E4SRec with Qwen2.5-7B as the backbone). The LLM is fine-tuned with a KL-divergence loss toward the CRM's softened output distribution, and the CRM is subsequently trained with a KL-divergence loss toward the LLM's output distribution, with sample-wise adaptive weighting and up to T_max=2 iterative rounds. The authors report consistent improvements over baselines on Beauty, Sports, Toys, and Yelp in Table 2, provide ablation studies in Figure 3, and claim that the framework adds no inference-time parameters or latency.

Significance. If the reported results are reproducible, LLMD4Rec is a practically useful contribution to sequential recommendation: it improves both a lightweight CRM and an LLM-centric model without adding inference-time parameters, on four public datasets, with ablations isolating the loop, the sample weight, and the embedding update. The paper also makes an explicit and testable claim that knowledge can flow bidirectionally and iteratively between a small model and a large model. However, the empirical support for the central attribution claim is incomplete: all results come from single runs without error bars or significance tests, the ablations do not include a self-distillation or extra-training control, and the efficiency table is internally inconsistent. These gaps currently limit the confidence that can be placed in the specific claim that LLM semantic knowledge, rather than additional optimization steps or self-distillation, drives the gains.

major comments (3)

[§5.2, Table 2; §5.3, Fig. 3] The central claim that LLM semantic knowledge, rather than additional optimization or self-distillation, drives the reported gains is not supported by the ablations. Algorithm 1 initializes the LLM with the CRM's ID embeddings and alternately trains the two models for T_max=2, so the CRM receives extra training rounds and a teacher that is partially derived from itself. The ablations remove the loop, the sample weight, or the embedding update, but never replace the LLM teacher with the CRM's own softened distribution, nor do they compare SASRec trained for the same number of epochs with an equivalent auxiliary loss. Moreover, all metrics are single runs without error bars or significance tests, and some reported improvements are small (e.g., Table 2, Beauty NDCG@5: 0.0379 vs. 0.0360 for LLMD4Rec_L vs. E4SRec). Such differences could be within run-to-run noise, so the phrase 'significantly improves' in the abstract and Section 5.2 needs stronger statistical backing.
[§5.5, Table 3] The efficiency comparison is internally inconsistent. The text states that Table 3 compares E4SRec with LLMD4Rec_L, but the table rows are labeled 'LLMD4Rec_C'. In addition, the Hit@10 values for LLMD4Rec_C in Table 3 (Beauty 0.0776, Toys 0.0831) differ from the corresponding values in Table 2 (0.0758 and 0.0815), and E4SRec on Toys also differs (0.0766 vs. 0.0776). This undermines the RQ4 efficiency claim. The table must be corrected, and the reported numbers should be consistent with Table 2 or explicitly described as results from a different stage (e.g., after one loop rather than the final model).
[§4.2.1, Eqs. (2)-(3); §4.3.1, Eqs. (6)-(7)] The method assumes that the output distributions of a CRM and an LLM-centric model over the same item set are directly comparable after temperature scaling, and that aligning them transfers meaningful semantic knowledge. No calibration analysis, sensitivity study for T1 and T2, or diagnostics of distributional divergence are provided, even though the two models use very different architectures and training objectives. The adaptive weights depend on per-sample ranks whose stability is not reported. Given that the KL losses are load-bearing for the method, the paper should include at least a sensitivity analysis of T1, T2, gamma, and beta, and one diagnostic showing that the aligned distributions are not dominated by noise.

minor comments (5)

[§5.1.1] The Yelp dataset reference is missing; the citation appears as '[?]' in the text and should be filled in.
[Figure 4] The caption for Figure 4 duplicates the caption of Figure 2 ('Overview of the proposed LLMD4Rec framework'); it should instead describe the performance-over-training-rounds curves.
[Algorithm 1, §5.1.4] The semantics of T_max are ambiguous: the loop 'for t <- 0 to T_max' with T_max=2 could imply three iterations, while the text says the maximum number of iterative refinement steps is 2. Please clarify the loop bound.
[§5.1.4] The paper states that all models are trained on two A800 GPUs but does not report wall-clock training time for LLMD4Rec_C or the full LLMD4Rec pipeline; adding these details would strengthen the efficiency analysis.
[General] No code or trained model files are released. A public implementation would substantially improve the reproducibility of the empirical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical mutual-distillation results are measured against external baselines, not derived from the method's own definitions.

full rationale

This paper is an empirical training-method paper rather than a derivation chain, so the proposed framework's output metrics are not equivalent by construction to its inputs. The bidirectional distillation losses (Eqs. 3, 5, 7, 9) define training objectives and do not define the evaluation quantities HR@K or NDCG@K; the reported gains are measured outcomes on external benchmarks with independent baselines such as SASRec, E4SRec, DLLM2Rec, and LLM-CF. The framework does initialize the LLM-centric model from the pretrained CRM's embeddings and iteratively alternates the two models, but this defines an optimization procedure, not a prediction that reduces to a fit. Self-citations appear only in related work and in the authors' other publications; none is used as a load-bearing justification for the central claim. The concern that the improvements may come from extra iterative training or self-distillation rather than from LLM semantic knowledge is a threat to causal attribution and experimental completeness, but it is not a circularity in the defined sense: no equation equals another by construction, and no fitted parameter is renamed as a prediction. The Table 3 label mismatch is an accuracy/reporting issue, not circularity. Therefore the appropriate finding is no significant circularity.

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

The central claim depends on five hand-set hyperparameters and on the assumption that output distributions over the item catalog are directly comparable across heterogeneous architectures. The framework introduces no new model components, particles, or external knowledge sources beyond the base SASRec and E4SRec models.

free parameters (5)

Temperature T1 for LLM-side distillation = 0.6
Sharpen the CRM's softened distribution when distilling into the LLM (Section 5.1.4).
Temperature T2 for CRM-side distillation = 0.2
Sharpen the LLM's softened distribution when distilling into the CRM (Section 5.1.4).
Balancing coefficient gamma = 1
Weight on the LLM's KL divergence loss in Eq. (5).
Balancing coefficient beta = 0.5
Weight on the CRM's KL divergence loss in Eq. (9).
Maximum distillation rounds T_max = 2
Number of alternating CRM-LLM iterations in Algorithm 1; chosen as optimal in Section 5.4 without reporting a sweep.

assumptions (5)

domain assumption The CRM and the LLM-centric RM both score all items in the same output space, so their softmax distributions can be aligned directly with KL divergence without feature transformations.
Invoked in Section 4.2.1 and used in Eqs. (2), (3), (6), (7). This is load-bearing for both distillation directions.
domain assumption Qwen2.5-7B with LoRA rank 8 and linear projections can learn collaborative signals from CRM item ID embeddings.
Underlies E4SRec's design, adopted in Section 3.3 and trained in Algorithm 1.
domain assumption Leave-one-out evaluation with random negative sampling provides reliable top-K accuracy estimates.
Section 5.1.3; the paper reports no variance, and sampling-based metrics are known to be sensitive to the candidate set.
domain assumption After temporal filtering and 5-core pruning, the four datasets are representative for sequential recommendation.
Section 5.1.1; the filtering choices are reasonable but not independently justified.
standard math Standard mathematical tools (softmax, KL divergence, Adam) behave as expected.
Used throughout Sections 3 and 4.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bidirectional Knowledge Distillation for Enhancing Sequential Recommendation with Large Language Models." pith.science (2026). https://pith.science/paper/VLCMU65G

@misc{pith2026250518120,
  author       = {Pith},
  title        = {Pith review of: Bidirectional Knowledge Distillation for Enhancing Sequential Recommendation with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLCMU65G}},
  note         = {Machine review of arXiv:2505.18120}
}

read the original abstract

Large language models (LLMs) have demonstrated exceptional performance in understanding and generating semantic patterns, making them promising candidates for sequential recommendation tasks. However, when combined with conventional recommendation models (CRMs), LLMs often face challenges related to high inference costs and static knowledge transfer methods. In this paper, we propose a novel mutual distillation framework, LLMD4Rec, that fosters dynamic and bidirectional knowledge exchange between LLM-centric and CRM-based recommendation systems. Unlike traditional unidirectional distillation methods, LLMD4Rec enables iterative optimization by alternately refining both models, enhancing the semantic understanding of CRMs and enriching LLMs with collaborative signals from user-item interactions. By leveraging sample-wise adaptive weighting and aligning output distributions, our approach eliminates the need for additional parameters while ensuring effective knowledge transfer. Extensive experiments on real-world datasets demonstrate that LLMD4Rec significantly improves recommendation accuracy across multiple benchmarks without increasing inference costs. This method provides a scalable and efficient solution for combining the strengths of both LLMs and CRMs in sequential recommendation systems.

Figures

Figures reproduced from arXiv: 2505.18120 by the authors.

**Figure 2.** Overview of the proposed LLMD4Rec framework, depicting the directional flow of information and the iterative [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Ablation study of the loop mechanism, distillation sample weight, and ID embedding update in LLMD4Rec. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Overview of the proposed LLMD4Rec framework [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 21 canonical work pages

[1]

Wei Chen, Shuzhe Li, Chao Huang, Yanwei Yu, Yongguo Jiang, and Junyu Dong

work page
[2]

Xuxin Cheng, Bowen Cao, Qichen Ye, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023. Ml-lmcl: Mutual learning and large-margin contrastive learning for improving asr robustness in spoken language understanding.arXiv preprint arXiv:2311.11375(2023)

work page Pith review arXiv 2023
[3]

Yu Cui, Feng Liu, Pengbo Wang, Bohao Wang, Heng Tang, Yi Wan, Jun Wang, and Jiawei Chen. 2024. Distillation matters: empowering sequential recommenders to match the performance of large language models. InProceedings of the 18th ACM Conference on Recommender Systems. 507–517

2024
[4]

Gabriel de Souza Pereira Moreira, Sara Rabhi, Jeong Min Lee, Ronay Ak, and Even Oldridge. 2021. Transformers4rec: Bridging the gap between nlp and sequential/session-based recommendation. InProceedings of the 15th ACM con- ference on recommender systems. 143–153

2021
[5]

Yingpeng Du, Zhu Sun, Ziyan Wang, Haoyan Chua, Jie Zhang, and Yew-Soon Ong. 2025. Active Large Language Model-based Knowledge Distillation for Session-based Recommendation. InProceedings of the AAAI Conference on Artifi- cial Intelligence, Vol. 39. 11607–11615

work page 2025
[6]

Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. 2021. Knowl- edge distillation: A survey.International Journal of Computer Vision129, 6 (2021), 1789–1819

2021
[7]

Shengkang Gu, Jiahao Liu, Dongsheng Li, Guangping Zhang, Mingzhe Han, Hansu Gu, Peng Zhang, Ning Gu, Li Shang, and Tun Lu. 2025. LLM-Based User Simulation for Low-Knowledge Shilling Attacks on Recommender Systems.arXiv preprint arXiv:2505.13528(2025)

arXiv 2025
[8]

Mingzhe Han, Dongsheng Li, Jiafeng Xia, Jiahao Liu, Hansu Gu, Peng Zhang, Ning Gu, and Tun Lu. 2025. FedCIA: Federated Collaborative Information Aggregation for Privacy-Preserving Recommendation.arXiv preprint arXiv:2504.14208(2025)

work page Pith review arXiv 2025

Show all 59 references

[9]

Jesse Harte, Wouter Zorgdrager, Panos Louridas, Asterios Katsifodimos, Diet- mar Jannach, and Marios Fragkoulis. 2023. Leveraging large language models for sequential recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1096–1102

2023
[10]

Ruining He and Julian McAuley. 2016. Fusing similarity models with markov chains for sparse sequential recommendation. In2016 IEEE 16th international conference on data mining (ICDM). IEEE, 191–200

2016
[11]

Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. 2019. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In Proceedings of the AAAI conference on artificial intelligence, Vol. 33. 3779–3787

2019
[12]

Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk
[13]

Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531(2015)

2015 arXiv
[14]

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.ICLR1, 2 (2022), 3

2022
[15]

Jun Hu, Wenwen Xia, Xiaolu Zhang, Chilin Fu, Weichang Wu, Zhaoxin Huan, Ang Li, Zuoli Tang, and Jun Zhou. 2024. Enhancing sequential recommendation via llm-based semantic embedding learning. InCompanion Proceedings of the ACM Web Conference 2024. 103–111

2024
[16]

Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

2018
[17]

Sein Kim, Hongseok Kang, Kibum Kim, Jiwan Kim, Donghyun Kim, Minchul Yang, Kwangjin Oh, Julian McAuley, and Chanyoung Park. 2025. Lost in Sequence: Do Large Language Models Understand Sequential Recommendation?arXiv preprint arXiv:2502.13909(2025)

2025 arXiv
[18]

Diederik P Kingma. 2014. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980(2014)

2014 arXiv
[19]

Jiacheng Li, Yujie Wang, and Julian McAuley. 2020. Time interval aware self- attention for sequential recommendation. InProceedings of the 13th international conference on web search and data mining. 322–330

2020
[20]

Lei Li, Yongfeng Zhang, and Li Chen. 2023. Prompt distillation for efficient llm- based recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1348–1357

2023
[21]

Xinhang Li, Chong Chen, Xiangyu Zhao, Yong Zhang, and Chunxiao Xing. 2023. E4srec: An elegant effective efficient extensible solution of large language models for sequential recommendation.arXiv preprint arXiv:2312.02443(2023)

2023 arXiv
[22]

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

2024
[23]

Jiahao Liu, Shengkang Gu, Dongsheng Li, Guangping Zhang, Mingzhe Han, Hansu Gu, Peng Zhang, Tun Lu, Li Shang, and Ning Gu. 2025. Enhancing Cross- Domain Recommendations with Memory-Optimized LLM-Based User Agents. arXiv preprint arXiv:2502.13843(2025)

2025 arXiv
[24]

Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Jiongran Wu, Peng Zhang, Li Shang, and Ning Gu. 2023. Recommendation unlearning via matrix correction.arXiv preprint arXiv:2307.15960(2023)

2023 arXiv
[25]

Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, and Ning Gu. 2022. Parameter-free dynamic graph embedding for link prediction.Advances in Neural Information Processing Systems35 (2022), 27623–27635

2022
[26]

Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. 2023. Personalized graph signal processing for collaborative filtering. In Proceedings of the ACM Web Conference 2023. 1264–1272

2023
[27]

Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu. 2023. Triple structural information modelling for accurate, explainable and interactive recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in In...

2023
[28]

Jiahao Liu, Dongsheng Li, Hansu Gu, Peng Zhang, Tun Lu, Li Shang, and Ning Gu. 2025. Mitigating Popularity Bias in Collaborative Filtering through Fair Sampling.arXiv preprint arXiv:2502.13840(2025)

2025 arXiv
[29]

Jiahao Liu, Yiyang Shao, Peng Zhang, Dongsheng Li, Hansu Gu, Chao Chen, Longzhi Du, Tun Lu, and Ning Gu. 2025. Filtering Discomforting Recommenda- tions with Large Language Models. InProceedings of the ACM on Web Conference

2025
[30]

Jiahao Liu, Xueshuo Yan, Dongsheng Li, Guangping Zhang, Hansu Gu, Peng Zhang, Tun Lu, Li Shang, and Ning Gu. 2025. Enhancing LLM-Based Recom- mendations Through Personalized Reasoning.arXiv preprint arXiv:2502.13845 (2025). Bidirectional Knowledge Distillation for Enhancing Se...

2025 arXiv
[31]

Qidong Liu, Xian Wu, Yejing Wang, Zijian Zhang, Feng Tian, Yefeng Zheng, and Xiangyu Zhao. 2024. Llm-esr: Large language models enhancement for long- tailed sequential recommendation.Advances in Neural Information Processing Systems37 (2024), 26701–26727

2024
[32]

Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Zijian Zhang, Feng Tian, and Yefeng Zheng. 2024. Large language model distilling medication recommendation model.arXiv preprint arXiv:2402.02803(2024)

2024 arXiv
[33]

Sijia Liu, Jiahao Liu, Hansu Gu, Dongsheng Li, Tun Lu, Peng Zhang, and Ning Gu. 2023. Autoseqrec: Autoencoder for efficient sequential recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1493–1502

2023
[34]

Yuting Liu, Jinghao Zhang, Yizhou Dang, Yuliang Liang, Qiang Liu, Guibing Guo, Jianzhe Zhao, and Xingwei Wang. 2025. Cora: Collaborative information perception by large language model’s weights for recommendation. InProceedings of the AAAI Conference on Artificial Intelligence...

2025
[35]

Wensheng Lu, Jianxun Lian, Wei Zhang, Guanghua Li, Mingyang Zhou, Hao Liao, and Xing Xie. 2024. Aligning Large Language Models for Controllable Recommendations. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 8159–8172

2024
[36]

Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel
[37]

Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. 2014. Fitnets: Hints for thin deep nets.arXiv preprint arXiv:1412.6550(2014)

2014 arXiv
[38]

Guy Shani, David Heckerman, and Ronen I Brafman. 2005. An MDP-based recommender system.Journal of machine Learning research6, Sep (2005), 1265– 1295

2005
[39]

InProceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

Image-based recommendations on styles and substitutes. InProceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52
[40]

Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang
[41]

Zhu Sun, Kaidong Feng, Jie Yang, Xinghua Qu, Hui Fang, Yew-Soon Ong, and Wenyuan Liu. 2024. Adaptive In-Context Learning with Large Language Mod- els for Bundle Generation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information ...

2024
[42]

Elena Smirnova and Flavian Vasile. 2017. Contextual sequence modeling for rec- ommendation with recurrent neural networks. InProceedings of the 2nd workshop on deep learning for recommender systems. 2–9

2017
[43]

Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender ...

2024
[44]

Chengfeng Xu, Jian Feng, Pengpeng Zhao, Fuzhen Zhuang, Deqing Wang, Yanchi Liu, and Victor S Sheng. 2021. Long-and short-term self-attention network for sequential recommendation.Neurocomputing423 (2021), 580–589

2021
[45]

Sheng, Zhiming Cui, Xiaofang Zhou, and Hui Xiong

Chengfeng Xu, Pengpeng Zhao, Yanchi Liu, Jiajie Xu, Victor S Sheng S. Sheng, Zhiming Cui, Xiaofang Zhou, and Hui Xiong. 2019. Recurrent convolutional neural network for sequential recommendation. InThe world wide web conference. 3398–3404

2019
[46]

Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, and Jun Xu. 2024. Large language models enhanced collaborative filtering. InPro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2178–2188

2024
[47]

Zheng Xu, Yen-Chang Hsu, and Jiawei Huang. 2017. Training shallow and thin networks for acceleration via knowledge distillation with conditional adversarial networks.arXiv preprint arXiv:1709.00513(2017)

2017 arXiv
[48]

Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. 2017. A gift from knowl- edge distillation: Fast optimization, network minimization and transfer learning. InProceedings of the IEEE conference on computer vision and pattern recognition. 4133–4141

2017
[49]

Haoyi Zhang, Guohao Sun, Jinhu Lu, Guanfeng Liu, and Xiu Susie Fang. 2024. DELRec: Distilling Sequential Pattern to Enhance LLMs-based Sequential Rec- ommendation.arXiv e-prints(2024), arXiv–2406

2024
[50]

Wujiang Xu, Zujie Liang, Jiaojiao Han, Xuying Ning, Wenfang Lin, Linxun Chen, Feng Wei, and Yongfeng Zhang. 2024. Slmrec: empowering small language models for sequential recommendation.arXiv e-prints(2024), arXiv–2405

2024
[51]

Jiawei Zhao, Wei Luo, Boxing Chen, and Andrew Gilman. 2021. Mutual-learning improves end-to-end speech translation. InProceedings of the 2021 conference on empirical methods in natural language processing. 3989–3994

2021
[52]

Zhi Zheng, Wenshuo Chao, Zhaopeng Qiu, Hengshu Zhu, and Hui Xiong. 2024. Harnessing large language models for text-rich sequential recommendation. In Proceedings of the ACM Web Conference 2024. 3207–3216

2024
[53]

Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-rec: Self-supervised learning for se- quential recommendation with mutual information maximization. InProceedings of the 29th ACM international conference on inf...

2020
[54]

Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2024. Agentcf: Collaborative learning with autonomous language agents for recommender systems. InProceedings of the ACM on Web Conference 2024. 3679–3689

2024
[55]

Zhangchi Zhu and Wei Zhang. 2025. Preference-Consistent Knowledge Dis- tillation for Recommender System.IEEE Transactions on Knowledge and Data Engineering(2025)

2025
[58]

Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. 2022. Filter-enhanced MLP is all you need for sequential recommendation. InProceedings of the ACM web conference 2022. 2388–2399

2022
[2015]

Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939(2015)

2015 arXiv
[2019]

InProceedings of the 28th ACM international conference on information and knowledge management

BERT4Rec: Sequential recommendation with bidirectional encoder rep- resentations from transformer. InProceedings of the 28th ACM international conference on information and knowledge management. 1441–1450
[2022]

Mutual distillation learning network for trajectory-user linking.arXiv preprint arXiv:2205.03773(2022)

2022 arXiv

Pith tools

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

[1] [1]

Wei Chen, Shuzhe Li, Chao Huang, Yanwei Yu, Yongguo Jiang, and Junyu Dong

work page

[2] [2]

Xuxin Cheng, Bowen Cao, Qichen Ye, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023. Ml-lmcl: Mutual learning and large-margin contrastive learning for improving asr robustness in spoken language understanding.arXiv preprint arXiv:2311.11375(2023)

work page Pith review arXiv 2023

[3] [3]

Yu Cui, Feng Liu, Pengbo Wang, Bohao Wang, Heng Tang, Yi Wan, Jun Wang, and Jiawei Chen. 2024. Distillation matters: empowering sequential recommenders to match the performance of large language models. InProceedings of the 18th ACM Conference on Recommender Systems. 507–517

2024

[4] [4]

Gabriel de Souza Pereira Moreira, Sara Rabhi, Jeong Min Lee, Ronay Ak, and Even Oldridge. 2021. Transformers4rec: Bridging the gap between nlp and sequential/session-based recommendation. InProceedings of the 15th ACM con- ference on recommender systems. 143–153

2021

[5] [5]

Yingpeng Du, Zhu Sun, Ziyan Wang, Haoyan Chua, Jie Zhang, and Yew-Soon Ong. 2025. Active Large Language Model-based Knowledge Distillation for Session-based Recommendation. InProceedings of the AAAI Conference on Artifi- cial Intelligence, Vol. 39. 11607–11615

work page 2025

[6] [6]

Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. 2021. Knowl- edge distillation: A survey.International Journal of Computer Vision129, 6 (2021), 1789–1819

2021

[7] [7]

Shengkang Gu, Jiahao Liu, Dongsheng Li, Guangping Zhang, Mingzhe Han, Hansu Gu, Peng Zhang, Ning Gu, Li Shang, and Tun Lu. 2025. LLM-Based User Simulation for Low-Knowledge Shilling Attacks on Recommender Systems.arXiv preprint arXiv:2505.13528(2025)

arXiv 2025

[8] [8]

Mingzhe Han, Dongsheng Li, Jiafeng Xia, Jiahao Liu, Hansu Gu, Peng Zhang, Ning Gu, and Tun Lu. 2025. FedCIA: Federated Collaborative Information Aggregation for Privacy-Preserving Recommendation.arXiv preprint arXiv:2504.14208(2025)

work page Pith review arXiv 2025