Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Boosting Parameter Efficiency in LLM-Based Recommendation through Sophisticated Pruning

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper demonstrates that most parameters in an LLM-based recommender are redundant for the recommendation task, and that a staged, fine-grained pruning pipeline can remove 95% of non-embedding parameters while keeping 88.6% of accuracy.

desk verdict A plausible and carefully staged pruning pipeline for LLM recommenders, but the 88/95 headline needs a full-pipeline random baseline and released code before I'd trust it as a general claim. read the letter →

arxiv 2507.07064 v1 pith:IX6LV2UA submitted 2025-07-09 cs.IR

Shanle Zheng , Keqin Bao , Jizhi Zhang , Yang Zhang , Fuli Feng , Xiangnan He This is my paper
classification cs.IR
keywords LLM-basedrecommendationmodelpruningparameterefficiencystructuredknowledgedistillationsequentialattentionheadMLP
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

LLM-based recommender systems carry a lot of dead weight: most of the parameters that help with general language tasks are not needed to predict the next item a user will buy. This paper proposes PruneRec, a pipeline that removes that dead weight in three stages — first attention heads and embedding dimensions, then MLP width, then whole layers — and restores performance with distillation after each cut. On three Amazon datasets, the pruned model keeps on average 88.6% of the original BIGRec model's recommendation accuracy while using only about 5% of its non-embedding parameters (17M instead of 357M). The claim matters because the largest obstacle to deploying LLM recommenders is their memory and compute cost; if this result holds broadly, much cheaper deployment is within reach.

What carries the argument

The pipeline is an iterative prune-then-restore loop over four importance estimators and one distillation objective. Head importance: suppress head $i$ in layer $l$ by scaling its attention logits with a near-zero $\epsilon^l_i$ (Eq. 3) and measure $D_{\mathrm{KL}}$ between the output distributions with and without suppression; scores are min-max normalized per layer and then propagated across layers by $\mathrm{Imp}(O^{l+1}_i) = \alpha\,\mathrm{Imp}(O^l_i) + (1-\alpha)\,\mathrm{Imp}(O^{l+1}_i)$, which is what couples shallow-layer decisions to deep-layer decisions. Embedding dimensions: average of $|E_d \odot \nabla E_d|$ over $B$ samples and $S$ positions, retaining those matching the attention input dimension. MLP width: for dimension $d$, count how often the last-token activation magnitude $|H^{(b,d)}_{\text{last}}|$ exceeds threshold $\tau$ across the $B$ samples, and keep the top $K_{\mathrm{MLP}}$ dimensions (rows/columns of $W_{\mathrm{up}}, W_{\mathrm{gate}}, W_{\mathrm{down}}$). Layer depth: mask each layer and compute $\Delta\mathrm{PPL}_l = \mathrm{PPL}_{\text{masked}(l)} - \mathrm{PPL}_{\text{original}}$; iteratively remove the least important layers down to $K_{\mathrm{Layer}}$. After each stage, restoration minimizes $\lambda\, \mathrm{KL}(p_T \parallel p_S) + (1-\lambda)\,\mathrm{CE}(y, p_S)$ with the pre-pruned model as teacher. The staged ordering is the load-bearing mechanism: each width cut changes what subsequent layer and MLP importance estimates see, and the distillation step between stages is what makes the estimates for the next stage trustworthy.

What would settle it

Recompute the Stage III layer ranking using the drop in HR@20 when each layer is masked on the same calibration set, instead of the perplexity increase. If the two rankings disagree substantially, then the layer-pruning criterion is optimizing a language-model objective rather than the recommendation objective, and the pipeline's depth cuts are not the ones a recommendation-optimal policy would choose.

Watch

Extended reading notes

Core claim

The paper's central claim is that a fine-tuned LLM recommender contains large amounts of task-irrelevant parameters, and that these can be identified and removed at multiple granularities without collapsing accuracy. The discovery is empirical: after fine-tuning Qwen2-0.5B with BIGRec, the activation distributions in self-attention and MLP layers become long-tailed — a small fraction of dimensions carries most of the signal — and layer-wise-only pruning (as in SLMRec) is too coarse. PruneRec exploits this by (i) suppressing each attention head and measuring the KL divergence of the output distribution to score head importance, with scores propagated from shallow to deep layers via a recursive weight alpha; (ii) pruning embedding dimensions by the absolute product of embedding weights and their gradients; (iii) keeping MLP intermediate dimensions whose activations frequently exceed a threshold; and (iv) dropping layers whose removal least increases perplexity. Each stage is followed by distillation (reverse KL + cross-entropy) from the original model. Across Video Games, Sports, and CDs, the resulting 17M-parameter model averages 88.6% of the full model's HR@10/20 and NDCG@10/20 — over 95% non-embedding parameter reduction. A preliminary result shows general-purpose pruning (Wanda) underperforms random pruning on this task, motivating the task-specific estimators.

Load-bearing premise

The load-bearing premise is that the four importance scores — KL divergence from suppressing a head, gradient-times-embedding, activation frequency, and perplexity change — computed on one 100-sample batch, rank parameters correctly for the recommendation task, and that the pruning hyperparameters tuned on these three datasets transfer to other settings.

Editorial extensions

If this is right

  • LLM recommender deployments can cut non-embedding memory by roughly 20x (357M to 17M) at a cost of about 11 percentage points of relative accuracy, making on-device or edge serving feasible where it was not before.
  • Width pruning (attention heads plus embedding dimensions) is the biggest lever: it alone removes about 83% of parameters for an average 8% accuracy drop, according to the stage-wise analysis.
  • Task-specific pruning matters: general-purpose methods like Wanda are worse than random here, so recommender pruning needs its own importance estimators rather than borrowed LLM pruning tools.
  • Iterative prune-and-restore with distillation is what keeps sequential pruning stable; removing parameters at several granularities in one shot would collapse the model, so the staged ordering is part of the result.
  • Layer-only pruning (SLMRec) at the same parameter budget is markedly worse, implying fine-grained intra-layer cuts are necessary, and layer pruning alone hits a depth limit beyond which performance collapses.

Reading between the lines

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

  • The paper does not vary the order of the three stages; a natural test of the pipeline's logic is to run layer pruning first and width pruning second. The paper's claim that width cuts precondition later layer-importance estimates predicts the reversed order should be measurably worse at the same parameter budget.
  • Because Stage III uses perplexity, a language-generation metric, to rank layers for a recommendation model, swapping in a recommendation-aware score (e.g., the drop in HR@20 after masking a layer) is a plausible improvement the paper does not explore.
  • The importance estimators are computed on a single 100-sample draw; a cheap robustness check is to recompute the pruned sets across several draws and measure disagreement. If the variance is high, the stated 88/95 result is partly a statement about that particular sample.
  • The mechanism described — fine-tuned LLM recommenders concentrate signal in few dimensions — should generalize to other sequential-recommendation domains, but the specific hyperparameters alpha, lambda, K_Attn = 7, K_MLP = 896, K_Layer = 16 would likely need re-tuning per domain.
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

5 major / 6 minor

Summary. The paper describes PruneRec, a structured pruning pipeline for LLM-based sequential recommendation. It starts from the BIGRec/Qwen2-0.5B model and removes parameters in three stages: attention-head and embedding-dimension pruning (§3.1), MLP intermediate-dimension pruning (§3.2), and layer pruning (§3.3), with a distillation-based restoration after each stage (§3.4). Experiments on Amazon Video Games, Sports, and CDs report that a 17.0M non-embedding-parameter model retains 88.6% of BIGRec's average HR/NDCG across 12 metrics while pruning more than 95% of non-embedding parameters. The paper also reports observation experiments on activation sparsity and an ablation of the attention-head importance estimator.

Significance. If the central claim is accepted, PruneRec provides strong evidence that LLM-based recommenders contain exploitable structured redundancy both across layers and within attention/MLP modules, and that a staged prune-and-distill pipeline can cut non-embedding parameters by 95% while keeping about 88% of accuracy. The paper's strengths are its clear three-stage decomposition, the per-stage performance/parameter tracking in Figure 4, the head-level ablation in Table 3, and the comparisons against both traditional and LLM-based baselines. The main caveat is that the headline result currently lacks a full-pipeline random-pruning control and a documented hyperparameter-selection protocol; the load-bearing importance estimators are therefore not yet fully isolated from the effect of distillation restoration. With those controls added, the result would be a useful step toward deployable LLM-based recommendation.

major comments (5)
  1. [§3.4/§4.2, Table 2] The strongest evidence for the claim that the proposed importance estimators matter is missing. The only random-mask control is for Stage I attention heads (Table 3). Because every pruning stage is followed by distillation from the original model (§3.4), the final 88.6% average retention could in principle be obtained with any fixed mask of the same remaining capacity. Please add an experiment on at least one dataset (ideally all three) in which random masks are applied at the same budgets (K_Attn=7, K_MLP=896, K_Layer=16, same embedding-dimension retention) through all three stages, followed by the same restoration procedure, and report the resulting HR/NDCG. Without this control, the claim that the sophisticated scoring functions are load-bearing is not established.
  2. [§4.1.4 and Figure 5] The hyperparameters that determine the final result—alpha, lambda, K_Attn, K_MLP, K_Layer, tau, B, and the embedding retention target—are stated only as 'set/adjusted' with no validation protocol. Since these values directly set the 95% pruning budget and the 88.6% retention, please specify the validation split used to choose them, report the candidate grids and selected values, and give sensitivity curves for alpha, lambda, tau, and B. If the test benchmarks were used for selection, the comparison in Table 2 should be re-run with a proper validation-based selection.
  3. [§3.3, Eq. (12)] Using perplexity (PPL) change as the layer-importance signal for a recommendation-tuned model is a potential domain mismatch. PPL measures language-modeling quality, not next-item recommendation accuracy. Please justify that PPL increase correlates with recommendation degradation in this setting, or compute the same layer-removal importance with HR/NDCG on the calibration set and report the agreement between the two rankings. This is load-bearing because Stage III determines K_Layer=16.
  4. [Table 2] The reported numbers are single runs without error bars or multiple seeds. Several of the main comparisons are small (e.g., Games HR@10: 0.0803 vs 0.0852; HR@20: 0.1164 vs 0.1200; CDs HR@10: 0.0946 vs 0.1094). Please report mean and standard deviation over at least three seeds for the main results and for the ablations in Table 3 and Figure 5, or otherwise demonstrate that the differences are not within run-to-run noise.
  5. [§3.1.1, Eq. (6)] Equation (6) is self-referential as written: Imp(O^{l+1}_i) appears on both sides. Presumably one side is the raw score and the other the propagated score, but the notation does not say which. Since alpha is one of the tuned hyperparameters and the Stage I ablations depend on this recurrence, please rewrite the equation with distinct symbols (e.g., Imp_raw and Imp_final) and state the initialization for layer 1.
minor comments (6)
  1. [§4.2] The bullet 'an average drop of only 88%' should be 'an average retention of 88%' (or 'a drop of only 12%').
  2. [Table 2] There are typos in the table: 'Traditinoal Recommender Systems' and 'Recommender Systems Ultilizing LLMs' should be corrected.
  3. [§3.1.2] 'we randomly the sample B samples' is ungrammatical; it should read 'we randomly sample B samples'.
  4. [Figure 1] The caption of Figure 1(a) says 'Top 448 Ratio' while the text describes the top K% largest activation values; please clarify what K is and whether 448 is a fixed dimension count.
  5. [§4.3.4 and Figure 5(c)] The text says layers were pruned by removing 12, 16, and 20 layers, while Figure 5(c) labels are 12, 16, 20; state clearly whether these are numbers of remaining layers or numbers of removed layers.
  6. [§2.2] The claim that WANDA underperforms random pruning (Figure 1b) is not accompanied by the WANDA configuration; specify sparsity ratio, calibration set size, and whether the pruning is unstructured or structured, otherwise the comparison is not reproducible.

Circularity Check

1 steps flagged · score 2.0 of 10

No substantive circularity: the 88.6% retention claim is a measured test-set outcome of a prune-and-distill pipeline; the only printed self-referential artifact is Eq. (6), which does not force the central result.

  1. self definitional [Section 3.1.1, Eq. (6) (recursive attention-head importance)]
    "For the i-th head in layer l+1, its final importance score is determined by: Imp(O^{l+1}_i) = α· Imp(O^l_i)+(1−α)· Imp(O^{l+1}_i), where α∈[0,1] controls the propagation weight from preceding layers."

    The quantity being defined, Imp(O^{l+1}_i), appears on both sides of the equation. Rearranging gives α·Imp(O^{l+1}_i)=α·Imp(O^l_i), so for any α≠0 the 'final' importance is algebraically forced to equal the previous layer's importance and the current layer's own contribution cancels. As written, therefore, the claimed recursive cross-layer weighting is not a computation that combines a shallow score with a deep score; it is an identity that defines the result in terms of the previous layer only, or is vacuous if both sides denote the same final score. This is a self-definitional artifact in Stage I's importance estimator, although it does not by itself make the measured 88.6% retention a tautology.

full rationale

The central claim of the paper is an empirical result: after a three-stage pruning pipeline with distillation restoration, PruneRec retains an average of 88.6% of BIGRec's HR@10/20 and NDCG@10/20 while cutting non-embedding parameters from 357M to 17M (Table 2). This is not definitionally tied to the pruning ratios or to any fitted parameter; it is a held-out evaluation against the unpruned model. The importance estimators in Stages I-III (KL divergence under head perturbation, gradient-times-embedding for embedding dimensions, activation-frequency for MLP units, and PPL change for layers) are proxies, but the paper does not define the final accuracy in terms of those proxies. The restoration step uses the original model as a teacher, which is standard self-distillation in pruning and does not smuggle the conclusion into the assumptions. Ablations in Table 3 and Figure 5 give at least partial independent evidence that the proposed selection criteria outperform random or simpler alternatives, though there is no full-pipeline random control; that is an experimental-design weakness, not circularity. Self-citations such as [4] and [10] are used as the base model or as related technique references, not as the load-bearing justification for the pruning result. The one genuinely self-referential passage is Eq. (6), where the final importance score appears on both sides of its own definition; this is flagged above. Overall, the headline result has independent content and the paper does not reduce to a fit or to a self-citation chain, so the circularity score is low.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The contribution is a pruning pipeline whose parameters and thresholds are fitted to the evaluation datasets, which is the main source of free parameters. The axioms are standard architectural assumptions plus domain-specific heuristic importance proxies.

free parameters (8)
  • alpha (head importance propagation weight) = tuned in [0,1] with step 0.1
    Controls the cross-layer propagation of attention-head importance in Eq. (6); selected on the validation performance of the same datasets.
  • lambda (distillation loss balance) = tuned in [0,1] with step 0.2
    Balances KL distillation and cross-entropy in Eq. (13); tuned on the same datasets.
  • K_Attn (attention heads pruned per layer) = 7
    Sets the Stage I attention pruning level; reported in Section 3.1.1 as a chosen value without a described tuning protocol.
  • K_MLP (retained MLP intermediate dimensions) = 896
    Sets the Stage II MLP pruning level, retaining 896 of 1792 dimensions (50%), despite the introduction claiming top 10% are retained.
  • K_Layer (retained transformer layers) = 16
    Sets the Stage III depth pruning level; chosen for the main result and varied in Figure 5(c).
  • tau (activation frequency threshold) = not reported
    Threshold in Eq. (10) that decides when an MLP dimension is counted as active; no value is given.
  • B (number of calibration samples) = not reported (100 in the observation study)
    Sample size for computing importance scores in Eqs. (4), (7), and (10); not specified for the main pipeline.
  • embedding dimension retention target = not specified
    The number of embedding dimensions kept to 'match the input dimension of the attention layer' in Section 3.1.2; the exact pruning ratio is not given.
assumptions (5)
  • domain assumption Transformer-based LLMs contain structurally redundant parameters for recommendation, both across layers and within attention and MLP modules.
    The entire pipeline rests on this, motivated by the small-sample observation in Section 2.2 on one dataset.
  • domain assumption Activation magnitude and gradient magnitude are valid importance proxies for pruning in this recommendation setting.
    Used in Eqs. (4), (7), and (10); no calibration against ground-truth importance is provided.
  • domain assumption Perplexity change from removing a layer is a valid measure of layer importance for a recommendation-tuned LLM.
    Stage III (Eq. 12) follows general-LLM practice, but PPL is a language metric and the model was fine-tuned with a recommendation cross-entropy objective.
  • domain assumption Distillation from the original model can sufficiently restore pruned-model performance so that later-stage importance estimates remain reliable.
    The restoration step in Section 3.4 is essential to the iterative pipeline; no evidence is given that restoration is fully sufficient.
  • standard math Standard backpropagation, softmax attention, and the Transformer forward pass are correct.
    Assumed throughout without proof; uncontroversial background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boosting Parameter Efficiency in LLM-Based Recommendation through Sophisticated Pruning." pith.science (2026). https://pith.science/paper/IX6LV2UA

@misc{pith2026250707064,
  author       = {Pith},
  title        = {Pith review of: Boosting Parameter Efficiency in LLM-Based Recommendation through Sophisticated Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IX6LV2UA}},
  note         = {Machine review of arXiv:2507.07064}
}
read the original abstract

LLM-based recommender systems have made significant progress; however, the deployment cost associated with the large parameter volume of LLMs still hinders their real-world applications. This work explores parameter pruning to improve parameter efficiency while maintaining recommendation quality, thereby enabling easier deployment. Unlike existing approaches that focus primarily on inter-layer redundancy, we uncover intra-layer redundancy within components such as self-attention and MLP modules. Building on this analysis, we propose a more fine-grained pruning approach that integrates both intra-layer and layer-wise pruning. Specifically, we introduce a three-stage pruning strategy that progressively prunes parameters at different levels and parts of the model, moving from intra-layer to layer-wise pruning, or from width to depth. Each stage also includes a performance restoration step using distillation techniques, helping to strike a balance between performance and parameter efficiency. Empirical results demonstrate the effectiveness of our approach: across three datasets, our models achieve an average of 88% of the original model's performance while pruning more than 95% of the non-embedding parameters. This underscores the potential of our method to significantly reduce resource requirements without greatly compromising recommendation quality. Our code will be available at: https://github.com/zheng-sl/PruneRec

Figures

Figures reproduced from arXiv: 2507.07064 by the authors.

Figure 1
Figure 1. Figure (a) shows the ratio of head activations to [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. This figure shows the distribution of activation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. This figure illustrates the distribution of activation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: This figure illustrates the changes in performance and parameter count across the three stages of our method for [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: This figure presents the performance results of our analysis experiments on different modules. In each plot, the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Position-Aware Drafting for Inference Acceleration in LLM-Based Generative List-Wise Recommendation

    cs.IR 2026-04 unverdicted novelty 6.0 of 10

    PAD-Rec augments standard draft models with item-position and step-position embeddings plus learnable gates, delivering up to 3.1x wall-clock speedup and 5% average gain over strong speculative-decoding baselines on f...

Reference graph

Works this paper leans on

47 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Yongqi An, Xu Zhao, Tao Yu, Ming Tang, and Jinqiao Wang. 2024. Fluctuation- based adaptive structured pruning for large language models. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 10865–10873

    work page 2024

  3. [3]

    Sotiris Anagnostidis, Dario Pavllo, Luca Biggio, Lorenzo Noci, Aurelien Lucchi, and Thomas Hofmann. 2023. Dynamic context pruning for efficient and inter- pretable autoregressive transformers. Advances in Neural Information Processing Systems 36 (2023), 65202–65223

    work page 2023

  4. [4]

    Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yancheng Luo, Chong Chen, Fuli Feng, and Qi Tian. 2023. A bi-step grounding paradigm for large language models in recommendation systems. ACM Transactions on Recommender Systems (2023)

    2023

  5. [5]

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

  6. [6]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

    2023

  7. [7]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Se- bastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24, 240 (2023), 1–113

    2023

  8. [8]

    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. In Proceedings of the 18th ACM Conference on Recommender Systems . 507–517

    work page 2024

Show all 47 references
  1. [9]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongx- iang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recom- mender Systems. 1126–1132

    2023

  2. [10]

    Boyi Deng, Wenjie Wang, Fengbin Zhu, Qifan Wang, and Fuli Feng. 2025. Cram: Credibility-aware attention modification in llms for combating misinformation in rag. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 23760–23768

    2025

  3. [11]

    Elias Frantar and Dan Alistarh. 2023. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning . PMLR, 10323–10337

    2023

  4. [12]

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-rec: Towards interactive and explainable llms-augmented recommender system. arXiv preprint arXiv:2303.14524 (2023)

    2023 arXiv

  5. [13]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  6. [14]

    Bairu Hou, Qibin Chen, Jianyu Wang, Guoli Yin, Chong Wang, Nan Du, Ruoming Pang, Shiyu Chang, and Tao Lei. 2025. Instruction-Following Pruning for Large Language Models. arXiv preprint arXiv:2501.02086 (2025)

    2025 arXiv

  7. [15]

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large language models are zero-shot rankers for recommender systems. In European Conference on Information Retrieval. Springer, 364–381

    2024

  8. [16]

    Yitong Ji, Aixin Sun, Jie Zhang, and Chenliang Li. 2023. A critical study on data leakage in recommender system offline evaluation. ACM Transactions on Information Systems 41, 3 (2023), 1–27

    2023

  9. [17]

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

    2018

  10. [18]

    Yongqi Li, Xinyu Lin, Wenjie Wang, Fuli Feng, Liang Pang, Wenjie Li, Liqiang Nie, Xiangnan He, and Tat-Seng Chua. 2024. A survey of generative search and recom- mendation in the era of large language models. arXiv preprint arXiv:2404.16924 (2024)

    2024 arXiv

  11. [19]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al . 2025. How can recommender systems benefit from large language models: A survey. ACM Transactions on Information Systems 43, 2 (2025), 1–47

    2025

  12. [20]

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. 2023. Llm-pruner: On the struc- tural pruning of large language models.Advances in neural information processing systems 36 (2023), 21702–21720

    2023

  13. [21]

    Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. 2024. Shortgpt: Layers in large language models are more redundant than you expect. arXiv preprint arXiv:2403.03853 (2024)

    2024 arXiv

  14. [22]

    Usha Ruby, Vamsidhar Yendapalli, et al. 2020. Binary cross entropy with deep learning technique for image classification. Int. J. Adv. Trends Comput. Sci. Eng 9, 10 (2020)

    2020

  15. [23]

    Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. [n. d.]. A Simple and Effec- tive Pruning Approach for Large Language Models. In The Twelfth International Conference on Learning Representations

  16. [24]

    Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. 2023. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695 (2023)

    2023 arXiv

  17. [25]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining . 565–573

    2018

  18. [26]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  19. [27]

    Hanbing Wang, Xiaorui Liu, Wenqi Fan, Xiangyu Zhao, Venkataramana Kini, Devendra Yadav, Fei Wang, Zhen Wen, Jiliang Tang, and Hui Liu. 2024. Rethinking large language model architectures for sequential recommendations. arXiv preprint arXiv:2402.09543 (2024)

    2024 arXiv

  20. [28]

    Qi Wang, Jindong Li, Shiqi Wang, Qianli Xing, Runliang Niu, He Kong, Rui Li, Guodong Long, Yi Chang, and Chengqi Zhang. 2024. Towards next- generation llm-based recommender systems: A survey and beyond.arXiv preprint Conference’17, July 2017, Washington, DC, USA Shanle Zheng, ...

    2024 arXiv

  21. [29]

    Ziheng Wang, Jeremy Wohlwend, and Tao Lei. 2019. Structured pruning of large language models. arXiv preprint arXiv:1910.04732 (2019)

    2019 arXiv

  22. [30]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Jun- feng Wang, Dawei Yin, and Chao Huang. 2024. Llmrec: Large language models with graph augmentation for recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining...

    2024

  23. [31]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2024. A survey on large language models for recommendation. World Wide Web 27, 5 (2024), 60

    2024

  24. [32]

    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

  25. [33]

    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

  26. [34]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al . 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 (2024)

    2024 arXiv

  27. [35]

    Yifei Yang, Zouying Cao, and Hai Zhao. 2024. Laco: Large language model pruning via layer collapse. arXiv preprint arXiv:2402.11187 (2024)

    2024 arXiv

  28. [36]

    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

  29. [37]

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Xin Zhao, Leyu Lin, and Ji-Rong Wen

  30. [38]

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

  31. [39]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 1, 2 (2023)

    2023 arXiv

  32. [40]

    Longguang Zhong, Fanqi Wan, Ruijun Chen, Xiaojun Quan, and Liangzhi Li. 2024. Blockpruner: Fine-grained pruning for large language models. arXiv preprint arXiv:2406.10594 (2024)

    2024 arXiv

  33. [41]

    Zhenhong Zhou, Haiyang Yu, Xinghua Zhang, Rongwu Xu, Fei Huang, Kun Wang, Yang Liu, Junfeng Fang, and Yongbin Li. 2024. On the Role of Attention Heads in Large Language Model Safety. arXiv preprint arXiv:2410.13708 (2024)

    2024 arXiv

  34. [42]

    Xunyu Zhu, Jian Li, Yong Liu, Can Ma, and Weiping Wang. 2024. A survey on model compression for large language models. Transactions of the Association for Computational Linguistics 12 (2024), 1556–1577

    2024

  35. [43]

    Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collab- orative large language model for recommender systems. In Proceedings of the ACM Web Conference 2024. 3162–3172

    2024

  36. [2015]

    arXiv preprint arXiv:1511.06939 (2015)

    2015 arXiv

  37. [2023]

    ACM Transactions on Information Systems (2023)

    Recommendation as instruction following: A large language model em- powered recommendation approach. ACM Transactions on Information Systems (2023)

    2023

  38. [2024]

    In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

    Decoding Matters: Addressing Amplification Bias and Homogeneity Issue in Recommendations for Large Language Models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing . 10540–10552

    2024

  39. [2025]

    IEEE Transactions on Knowledge and Data Engineering (2025)

    Collm: Integrating collaborative embeddings into large language models for recommendation. IEEE Transactions on Knowledge and Data Engineering (2025)

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.