Pith. sign in

REVIEW 4 major objections 5 minor 52 references

A pre-ranker can be trained on items users never saw, using LLM-generated pseudo-labels, and this lifted click-through rate by 3.07% in production.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 21:09 UTC pith:NG36KDVA

load-bearing objection Credible industrial LLM-pseudo-labeling for pre-ranking, but the paper's own appendix undermines its key offline proxy; the online lift is plausible and worth peer review. the 4 major comments →

arxiv 2602.20995 v2 pith:NG36KDVA submitted 2026-02-24 cs.IR cs.CL

Generative Pseudo-Labeling for Pre-Ranking with LLMs

Junyu Bi , Xinting Niu , Daixuan Cheng , Kun Yuan , Tao Wang , Binbin Cao , Jian Wu This is my paper
classification cs.IR cs.CL
keywords recommender systemssample selection biaspre-rankingpseudo-labelinglarge language modelssemantic identifiersmultimodal embeddingsuncertainty calibration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that pre-ranking models in industrial recommender systems can overcome sample selection bias by training on unexposed candidates as well as exposed ones. Its method, Generative Pseudo-Labeling (GPL), uses a large language model to generate content-based relevance pseudo-labels for unexposed items, then trains the lightweight pre-ranker on both real click labels and these pseudo-labels. If correct, the framework removes the train–serving distribution mismatch, improves long-tail and diverse recommendations, and adds zero online latency because all LLM work happens offline. The strongest evidence is a production A/B test reporting a 3.07% CTR lift and a large drop in category concentration.

Core claim

The central claim is that pseudo-labels computed from LLM-generated interest anchors provide valid supervision for unexposed items, enabling the pre-ranker to approximate the online serving distribution at training time. Specifically, each item is tokenized into hierarchical semantic identifiers (SIDs) via a frozen multimodal encoder and an RQ-VAE trained only on content, decoupling the token space from exposure patterns. A LoRA-fine-tuned LLM, trained on positive user histories converted to SIDs, predicts future SIDs via next-token prediction; beam search decodes these into a set of interest anchors per user. Relevance is scored as r = sigmoid(max_a cos(f_e(a), f_e(h))/tau), matching the be

What carries the argument

Interest anchors: per-user item embeddings generated by an LLM decoding predicted semantic identifiers, plus the matching score r = sigmoid(max_a cos(f_e(a), f_e(h))/tau) that converts content similarity into a pseudo-label. The frozen multimodal encoder f_e is the shared semantic space that grounds both anchors and candidates in content rather than interaction history; the uncertainty-weighted confidence term is what modulates the influence of noisy anchors during joint training.

Load-bearing premise

The pseudo-label r = sigmoid(max_a cos(f_e(a), f_e(h))/tau) is assumed to be a valid measure of genuine user interest for items the user never interacted with, even though the LLM that produces the anchors is trained only on positive exposed histories and the offline validation of pseudo-label quality (AUC*) is computed on exposed items.

What would settle it

Run a randomized exploration experiment: serve a cohort a set of items that were pseudo-labeled as high-relevance but never recommended, and another cohort items pseudo-labeled low-relevance; if the click-through rate does not follow the pseudo-label ordering, the signal is not genuinely predictive of reveal preference.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Pre-ranking models can be trained on the full recall set, not just exposed items, without any change to the online serving architecture.
  • Long-tail and cold-start items receive supervision from content similarity rather than exposure history, which should improve their discoverability.
  • The framework is plug-and-play for existing cascade systems; only the training data pipeline changes, not the serving infrastructure.
  • The offline, per-user anchor generation means the LLM cost is amortized and does not affect inference latency.
  • The dual-label objective with confidence weighting suggests a general recipe for mixing external semantic signals with behavioral labels in any ranking stage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If pseudo-labels truly capture latent interest, the same anchor-matching procedure could be extended to the ranking stage itself, enabling end-to-end content-aware scoring beyond pre-ranking.
  • The interest anchors could double as an interpretable user-interest summary, potentially improving recommender explainability or supporting diversity-aware candidate generation.
  • A testable extension: use anchors as actions in a reinforcement-learning formulation of recommender systems, letting the pseudo-label confidence serve as a reward-shaping signal.
  • The reported category-concentration drop hints that GPL could be used intentionally to control exposure diversity, not just correct for bias.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Generative Pseudo-Labeling (GPL), a two-stage framework for pre-ranking in industrial recommender systems. In the first stage, items are tokenized into semantic identifiers (SIDs) via a frozen multimodal encoder and a content-only RQ-VAE; a LoRA-fine-tuned LLM then predicts future SIDs from user histories, decodes them into interest anchors, and computes relevance scores for unexposed candidates by max-cosine matching in the frozen semantic space (Eq. 5). These scores are calibrated with uncertainty weights and combined with real exposed labels in a joint training objective (Eq. 13). The paper reports offline gains over several baselines on an industrial Taobao dataset and on Taobao-MM, and an online A/B test with +3.07% CTR. The core claim is that GPL generates unbiased, content-aware pseudo-labels for unexposed items at zero online latency, thereby mitigating sample-selection bias and improving long-tail discovery.

Significance. If the central claim is upheld, GPL is a practically interesting contribution: it uses an offline LLM pass to generate pseudo-labels and thus imposes no online latency, and it is evaluated at very large industrial scale with a consistent set of ablations. The paper also contains useful sensitivity analyses (λ, τ, λ1, λ2, beam size) and a rare, honest discussion of a failed alternative in Appendix A.3. However, the significance depends critically on whether the pseudo-labels for unexposed items are actually valid. The evidence presented for this is indirect and, in places, self-contradictory. The online A/B result is a strong systems-level signal, but it does not by itself isolate the proposed mechanism. With additional validation of the pseudo-label quality on unexposed data, the work could be a solid contribution to the pre-ranking and sample-selection-bias literature.

major comments (4)
  1. [§4.1.3, §3.3.3, Eq. (5)] The only offline pseudo-label quality metric, AUC*, is computed by generating pseudo-labels for exposed items and comparing them with real feedback. Since the LLM in §3.3.2 is trained on sequences of positively interacted (exposed) items, high agreement with exposed feedback may simply reflect content similarity to training positives. This does not validate the claim that pseudo-labels are unbiased or accurate for unexposed candidates. The paper should either provide an evaluation on unexposed items with ground truth (e.g., from randomized exploration traffic) or substantially weaken the 'unbiased' claim.
  2. [Appendix A.3] Appendix A.3 directly undermines the use of AUC* as a proxy for pseudo-label quality: adding a co-occurrence contrastive loss increases AUC* by 0.003 but decreases downstream ranking AUC by 0.007. Thus improvements in AUC* do not reliably transfer to the metric that matters. Tables 2 and 3 repeatedly use AUC* to argue for component quality and LLM scalability; this evidence is called into question by the paper's own ablation. The authors should replace or supplement AUC* with a validated measure of unexposed-label quality.
  3. [§3.4.2, Table 4] The hyperparameters λ and τ are selected jointly via online A/B testing on CTR, and the same CTR metric is then reported as the headline +3.07% gain. No confidence intervals, significance tests, or details of the number of online variants are reported. The headline estimate is therefore the maximum of a tuning procedure rather than a confirmatory estimate, which materially weakens the strength of the claimed effect. Please report the selection protocol, multiple-comparison considerations, and at least one confirmatory measurement on a held-out period.
  4. [§4.1.3, Table 1, §4.5] All offline metrics (HR@K, AUC, GAUC) are computed on exposed samples only, and the online metrics again measure feedback on exposed impressions. Consequently, the central methodological claim—that GPL explicitly aligns the training distribution with the online serving space for unexposed candidates—is not directly tested by any experiment. The improvements in long-tail buckets (Figure 4a) are also computed on exposed items. A randomized serving experiment that observes feedback on unexposed/exploration items, or a counterfactual simulation, is needed to support the stated mechanism.
minor comments (5)
  1. [Figure 3] The x-axis label says 'beam size B' while the body text refers to 'N=32 by default'. Use one symbol consistently for the number of anchors/beam width.
  2. [Eq. (9)] The final values of the historical-consistency weight λ1 and LLM-confidence weight λ2 are never stated. Sensitivity plots are useful, but reporting the deployed values would aid reproducibility.
  3. [Figure 4(a)] 'PVR' is not defined; the text uses 'page-view (PV) buckets'. Please clarify whether PVR is a typo or stands for 'page-view rate'.
  4. [Table 1] No variance or significance information is reported for the offline comparisons. At least standard deviations across multiple evaluation samples or seeds should be added, especially since the industrial dataset is a single 14-day snapshot.
  5. [§4.4] The claim that 'AUC* remains sufficiently aligned with user feedback' is not quantified beyond the raw number. Given the Appendix A.3 counterexample, a stronger justification or a different metric is needed.

Circularity Check

2 steps flagged

Pseudo-label quality evidence is circular (AUC* on the LLM's training distribution) and headline CTR is the same metric used for online hyperparameter selection.

specific steps
  1. fitted input called prediction [§3.3.2, §4.1.3, §A.3]
    "To measure pseudo-label quality, we generate pseudo-labels for exposed items and compute the AUC between these pseudo-labels and real user feedback, denoted as AUC∗. ... The LLM is trained exclusively on sequences of positively interacted items ... this alignment improves consistency between pseudo-labels and observed user feedback on exposed items: AUC∗ increases by 0.003 ... However ... overall ranking performance ... degrades—measured by a 0.007 drop in AUC."

    AUC* is computed on exposed items, which are the same distribution as the positive exposed sequences used to train the LLM that produces the interest anchors. High agreement on exposed items is therefore expected from the training objective and does not test whether Eq. (5) labels are valid for unexposed candidates. The paper's own §A.3 shows AUC* can improve while final AUC degrades, so the only offline proxy offered for the central claim is not a reliable—and is arguably circular—measure of unexposed pseudo-label validity.

  2. fitted input called prediction [§3.4.2, §4.5.1, Table 4]
    "In practice, we select λ and τ jointly via online A/B testing to maximize click-through rate (CTR) and long-tail coverage. ... GPL achieves consistent gains across all metrics: CTR (+3.07%)"

    The headline online result is the same CTR objective used to select the hyperparameters λ and τ. Since selection is performed on the reported metric with no separate holdout or correction reported, the +3.07% CTR is the outcome of tuning on the target metric rather than an independent prediction of it. The same applies to long-tail coverage, which is also part of the selection objective.

full rationale

Most of GPL's construction—content-only SIDs from a frozen multimodal encoder, RQ-VAE trained on reconstruction, and cosine matching in semantic space—is self-contained and not derived from its own outputs. No load-bearing self-citation chain or imported uniqueness theorem was found; the related-work citations to papers with overlapping authors are not used to justify the central mechanism. However, the two evidential pillars for the central claim are circular. First, the only offline pseudo-label quality metric (AUC*) is computed on exposed items even though the LLM that generates the anchors was trained on positive exposed sequences; the §A.3 ablation explicitly shows AUC* and downstream AUC move in opposite directions, so the paper itself demonstrates that AUC* cannot support the unexposed-label claim. Second, λ and τ are explicitly selected online to maximize CTR, and the same CTR lift is then reported as the headline result; this is selection on the reported metric, not an independent prediction. The w/o Actual Labels ablation further shows pseudo-labels alone are weak (HR@3 drops from 0.5254 to 0.3687), underscoring that the unexposed supervision is not independently validated. These issues make the central evidence partially circular, though the deployed A/B framework still contains genuine external comparison, so the overall circularity score is 6 rather than higher.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The method depends on a chain of fitted hyperparameters (τ, λ, λ1, λ2, B) and on the assumption that LLM-generated anchors from exposed histories are unbiased for unexposed items. The invented entity is the anchor set, which has no external handle outside the paper's own exposed-data proxy.

free parameters (5)
  • temperature τ (Eq. 5) = log(τ) ≈ -1.2 (Fig. 3e)
    Sharpens the max-cosine pseudo-label sigmoid; tuned on validation/A-B data to optimize ranking calibration.
  • balance weight λ (Eq. 13) = near-optimal in 0.1-1.0 (Fig. 3a)
    Selected jointly with τ via online A/B testing on CTR and long-tail coverage (§3.4.2), i.e. tuned on the reported outcome.
  • historical-consistency weight λ1 (Eq. 9) = best around 0.6 (Fig. 3b)
    Weights anchor similarity to the user history; tuned via sensitivity analysis.
  • LLM-confidence weight λ2 (Eq. 9) = ≈1.1 (Fig. 3c)
    Weights the LLM's average log-probability; tuned via sensitivity analysis.
  • beam width B (Eq. 4) = 32 default
    Number of decoded interest anchors; chosen via sensitivity analysis (Fig. 3f).
axioms (4)
  • domain assumption An LLM fine-tuned on positive exposed SID sequences generates unbiased interest anchors for unexposed items
    Underlies the entire pseudo-label signal; contradicted by the exposure-dependence of the training data (§3.3.2).
  • domain assumption Cosine similarity in the frozen multimodal space is a valid relevance measure for unseen items
    Eq. (5) defines pseudo-labels; no unexposed ground truth is used to test this assumption.
  • domain assumption Content-only RQ-VAE tokenization captures sufficient item semantics for the LLM to predict future interests
    SID coverage is measured on a validation cohort, but semantic fidelity for long-tail items is not directly assessed.
  • standard math Standard ML machinery (LoRA, RQ-VAE, InfoNCE, BCE) works as intended
    No novel mathematical claims; these are established techniques.
invented entities (1)
  • interest anchors (decoded future SIDs) no independent evidence
    purpose: Intermediate latent items used to score unexposed candidates by max cosine similarity
    No direct validation on unexposed items; quality is only indirectly inferred from AUC* on exposed items.

pith-pipeline@v1.3.0-alltime-deepseek · 17993 in / 11739 out tokens · 111696 ms · 2026-08-02T21:09:21.928800+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Generative Pseudo-Labeling for Pre-Ranking with LLMs." pith.science (2026). https://pith.science/paper/NG36KDVA

@misc{pith2026260220995,
  author       = {Pith},
  title        = {Pith review of: Generative Pseudo-Labeling for Pre-Ranking with LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NG36KDVA}},
  note         = {Machine review of arXiv:2602.20995}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Pre-ranking is a critical stage in industrial recommendation systems, tasked with efficiently scoring thousands of recalled items for downstream ranking. A key challenge is the train-serving discrepancy: pre-ranking models are trained only on exposed interactions, yet must score all recalled candidates -- including unexposed items -- during online serving. This mismatch not only induces severe sample selection bias but also degrades generalization, especially for long-tail content. Existing debiasing approaches typically rely on heuristics (e.g., negative sampling) or distillation from biased rankers, which either mislabel plausible unexposed items as negatives or propagate exposure bias into pseudo-labels. In this work, we propose Generative Pseudo-Labeling (GPL), a framework that leverages large language models (LLMs) to generate unbiased, content-aware pseudo-labels for unexposed items, explicitly aligning the training distribution with the online serving space. By offline generating user-specific interest anchors and matching them with candidates in a frozen semantic space, GPL provides high-quality supervision without adding online latency. Deployed in a large-scale production system, GPL improves click-through rate by 3.07%, while significantly enhancing recommendation diversity and long-tail item discovery.

Figures

Figures reproduced from arXiv: 2602.20995 by Binbin Cao, Daixuan Cheng, Jian Wu, Junyu Bi, Kun Yuan, Tao Wang, Xinting Niu.

Figure 1
Figure 1. Figure 1: Illustration of the cascade-based architecture in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the GPL framework. For each user, interest anchors are generated offline via an LLM in a discrete semantic [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fine-Grained Online A/B Analysis of GPL vs. Base [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 3 linked inside Pith

  1. [1]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Win- ter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litw...

    2020

  3. [3]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. 2023. Bias and debias in recommender system: A survey and future directions. ACM Transactions on Information Systems41, 3 (2023), 1–39

    2023

  4. [4]

    Yingpeng Du, Di Luo, Rui Yan, Xiaopei Wang, Hongzhi Liu, Hengshu Zhu, Yang Song, and Jie Zhang. 2024. Enhancing job recommendation through llm-based generative adversarial networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 8363–8371

    2024

  5. [5]

    Markus Freitag and Yaser Al-Onaizan. 2017. Beam search strategies for neural machine translation.arXiv preprint arXiv:1702.01806(2017)

    Pith/arXiv arXiv 2017

  6. [6]

    Feng Gao, Xin Zhou, Yinning Shao, Yue Wu, Jiahua Gao, Yujian Ren, Fengyang Qi, Ruochen Deng, and Jie Liu. 2025. Both Supply and Precision: Sample Debias and Ranking Consistency Joint Learning for Large Scale Pre-Ranking System. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 11672–11680

    2025

  7. [7]

    Jingyue Gao, Shuguang Han, Han Zhu, Siran Yang, Yuning Jiang, Jian Xu, and Bo Zheng. 2023. Rec4ad: A free lunch to mitigate sample selection bias for ads ctr prediction in taobao. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 4574–4580

    2023

  8. [8]

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets.Advances in neural information processing systems27 (2014)

    2014

  9. [9]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9729–9738

    2020

  10. [10]

    Ruining He, Lukasz Heldt, Lichan Hong, Raghunandan Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, et al

  11. [11]

    Xiangnan He, Yang Zhang, Fuli Feng, Chonggang Song, Lingling Yi, Guohui Ling, and Yongdong Zhang. 2023. Addressing confounding feature issue for causal recommendation.ACM Transactions on Information Systems41, 3 (2023), 1–23

    2023

  12. [12]

    Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learning vector-quantized item representation for transferable sequential recommenders. InProceedings of the ACM Web Conference 2023. 1162–1171

    2023

  13. [13]

    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

  14. [14]

    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

  15. [15]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive image generation using residual quantization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11523–11532

    2022

  16. [16]

    Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guoliang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. 2019. Multi-interest network with dynamic routing for recommendation at Tmall. InProceedings of the 28th ACM international conference on information and knowledge management. 2615–2623

    2019

  17. [17]

    Dongsheng Li, Chao Chen, Zhilin Gong, Tun Lu, Stephen M Chu, and Ning Gu

  18. [18]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning. PMLR, 19730–19742

    2023

  19. [19]

    Xuanlin Li, Xiangyu Cai, Hao Peng, Jia Duan, Wei Wang, Zehua Zhang, Chang- ping Peng, Zhangang Lin, Ching Law, and Jingping Shao. 2025. An Unbi- ased Entire-Space Causal Framework for Click-Through Rate Estimation in Pre-Ranking. InCompanion Proceedings of the ACM on Web Conference 2025. 306–314

    2025

  20. [20]

    Zhaorui Lian, Binzong Geng, Xiyu Chang, Yu Zhang, Ke Ding, Ziyu Lyu, Guanghu Yuan, Chengming Li, Min Yang, Zhaoxin Huan, et al. 2025. EGRec: Leveraging Generative Rich Intents for Enhanced Recommendation with Large Language Models. InCompanion Proceedings of the ACM on Web Conference 2025. 1113– 1117

    2025

  21. [21]

    Jiaye Lin, Qing Li, Guorui Xie, Zhongxu Guan, Yong Jiang, Ting Xu, Zhong Zhang, and Peilin Zhao. 2024. Mitigating Sample Selection Bias with Robust Domain Adaption in Multimedia Recommendation. InProceedings of the 32nd ACM International Conference on Multimedia. 7581–7590

    2024

  22. [22]

    Dugang Liu, Pengxiang Cheng, Zinan Lin, Xiaolian Zhang, Zhenhua Dong, Rui Zhang, Xiuqiang He, Weike Pan, and Zhong Ming. 2023. Bounding system- induced biases in recommender systems with a randomized dataset.ACM Trans- actions on Information Systems41, 4 (2023), 1–26

    2023

  23. [23]

    Shicong Liu, Hongtao Lu, and Junru Shao. 2015. Improved Residual Vector Quantization for High-dimensional Approximate Nearest Neighbor Search.CoRR abs/1509.05195 (2015)

    Pith/arXiv arXiv 2015

  24. [24]

    Zohreh Ovaisi, Ragib Ahsan, Yifan Zhang, Kathryn Vasilaky, and Elena Zheleva

  25. [25]

    Sinno Jialin Pan and Qiang Yang. 2010. A Survey on Transfer Learning.IEEE Transactions on Knowledge and Data Engineering22, 10 (2010), 1345–1359. doi:10. 1109/TKDE.2009.191

    2010

  26. [26]

    Jiarui Qin, Jiachen Zhu, Bo Chen, Zhirong Liu, Weiwen Liu, Ruiming Tang, Rui Zhang, Yong Yu, and Weinan Zhang. 2022. Rankflow: Joint optimization of multi- stage cascade ranking systems as flows. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 814–824

    2022

  27. [27]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learning. PmLR, 8748–8763

    2021

  28. [28]

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. 2019. Generating diverse high-fidelity images with vq-vae-2.Advances in neural information processing systems32 (2019)

    2019

  29. [29]

    Xiang-Rong Sheng, Feifan Yang, Litong Gong, Biao Wang, Zhangming Chan, Yujing Zhang, Yueyao Cheng, Yong-Nan Zhu, Tiezheng Ge, Han Zhu, et al

  30. [30]

    Haoyuan Song, Yibowen Zhao, Yixin Zhang, Hongxu Chen, and Lizhen Cui

  31. [31]

    Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning.Advances in neural information processing systems30 (2017)

    2017

  32. [32]

    Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. Neural Discrete Representation Learning. InAdvances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA. 6306–6315

    2017

  33. [33]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  34. [34]

    Yuan Wang, Peifeng Yin, Zhiqiang Tao, Hari Venkatesan, Jin Lai, Yi Fang, and PJ Xiao. 2023. An empirical study of selection bias in pinterest ads retrieval. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5174–5183

    2023

  35. [35]

    Knowledge Distillation Based Recommendation Systems: A Comprehensive Survey.Electronics14, 8 (2025), 1538

    2025

  36. [36]

    Bin Wu, Feifan Yang, Zhangming Chan, Yu-Ran Gu, Jiawei Feng, Chao Yi, Xiang- Rong Sheng, Han Zhu, Jian Xu, Mang Ye, and Bo Zheng. 2025. MUSE: A Simple Yet Effective Multimodal Search-Based Framework for Lifelong User Interest Modeling. arXiv:2512.07216 [cs.IR] https://arxiv.org/abs/2512.07216

    arXiv 2025

  37. [37]

    Yunjia Xi, Muyan Weng, Wen Chen, Chao Yi, Dian Chen, Gaoyang Guo, Mao Zhang, Jian Wu, Yuning Jiang, Qingwen Liu, et al. 2025. Bursting filter bubble: Enhancing serendipity recommendations with aligned large language models. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 5059–5070

    2025

  38. [38]

    Yue Xing, Qifan Song, and Guang Cheng. 2021. On the algorithmic stability of adversarial training.Advances in neural information processing systems34 (2021), 26523–26535

    2021

  39. [39]

    Xiaochuan Xu, Zeqiu Xu, Peiyang Yu, and Jiani Wang. 2025. Enhancing user intent for recommendation systems via large language models. InInternational Conference on Artificial Intelligence and Machine Learning Research (CAIMLR 2024), Vol. 13635. SPIE, 46–54

    2025

  40. [40]

    Jianping Wei, Yujie Zhou, Zhengwei Wu, and Ziqi Liu. 2024. Enhancing pre- ranking performance: Tackling intermediary challenges in multi-stage cascading recommendation systems. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5950–5958. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Junyu Bi et al

    2024

  41. [41]

    Kaike Zhang, Qi Cao, Fei Sun, Yunfan Wu, Shuchang Tao, Huawei Shen, and Xueqi Cheng. 2023. Robust recommender system: a survey and future directions. Comput. Surveys(2023)

    2023

  42. [42]

    Shuai Zhang, Lina Yao, Aixin Sun, and Yi Tay. 2019. Deep learning based recom- mender system: A survey and new perspectives.ACM computing surveys (CSUR) 52, 1 (2019), 1–38

    2019

  43. [43]

    Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, and Yongdong Zhang. 2021. Causal intervention for leveraging popularity bias in recommendation. InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 11–20

    2021

  44. [44]

    Zijian Zhang, Shuchang Liu, Ziru Liu, Rui Zhong, Qingpeng Cai, Xiangyu Zhao, Chunxu Zhang, Qidong Liu, and Peng Jiang. 2025. Llm-powered user simulator for recommender system. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 13339–13347

    2025

  45. [45]

    Zixuan Xu, Penghui Wei, Weimin Zhang, Shaoguo Liu, Liang Wang, and Bo Zheng

  46. [46]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click- Through Rate Prediction. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 1059–1068. A Details of Tokenization A.1 Multimodal Encoding To construct a uni...

    2018

  47. [51]

    Binglei Zhao, Houying Qi, Guang Xu, Mian Ma, Xiwei Zhao, Feng Mei, Sulong Xu, and Jinghe Hu. 2025. A Hybrid Cross-Stage Coordination Pre-ranking Model for Online Recommendation Systems. InCompanion Proceedings of the ACM on Web Conference 2025. 621–630

    2025

  48. [2019]

    InProceedings of the 2019 SIAM International Conference on Data Mining

    Collaborative filtering with noisy ratings. InProceedings of the 2019 SIAM International Conference on Data Mining. SIAM, 747–755

    2019

  49. [2020]

    InProceedings of the web conference 2020

    Correcting for selection bias in learning-to-rank systems. InProceedings of the web conference 2020. 1863–1873

    2020

  50. [2022]

    InProceedings of the ACM Web Conference 2022

    Ukd: Debiasing conversion rate estimation via uncertainty-regularized knowledge distillation. InProceedings of the ACM Web Conference 2022. 2078– 2087

    2022

  51. [2024]

    InProceedings of the 33rd ACM International Conference on Information and Knowledge Management

    Enhancing Taobao Display Advertising with Multimodal Representations: Challenges, Approaches and Insights. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 4858–4865

  52. [2025]

    Plum: Adapting pre-trained language models for industrial-scale generative recommendations.arXiv preprint arXiv:2510.07784(2025)

    arXiv 2025