pith. sign in

arxiv: 2606.23911 · v1 · pith:NVUUY674new · submitted 2026-06-22 · 💻 cs.IR

Scaling Dense Retrieval with LLM-Annotated Training Data: Structured Mining and Progressive Curriculum for E-Commerce Sponsored Search

Md Omar Faruk Rokon , Shasvat Desai , Jhalak Nilesh Acharya , Isha Shah , Kumar Priyam , Brahanyaa Somasundaram , Vamsee Tangirala , Minuteresa Thomas
show 4 more authors
Vivek Arora Vijay Manchi Hong Yao Kuang-chih Lee
This is my paper

Pith reviewed 2026-06-26 06:09 UTC · model grok-4.3

classification 💻 cs.IR
keywords dense retrievalLLM annotationsponsored searche-commerce retrievalcurriculum trainingtraining data miningtwo-tower modelrelevance grading
0
0 comments X

The pith

LLM cascade annotates 240M examples from retrieval disagreements to train a two-tower model that beats click-trained baselines by 5.1% NDCG@10 in sponsored search.

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

The paper establishes that disagreements among existing retrieval systems supply structured training signals that can be labeled at scale by a calibrated LLM cascade instead of clicks or humans. It combines multi-channel mining, graded annotation reaching 89.1% human agreement, and progressive curriculum training across five difficulty levels to produce a dense retrieval model. A sympathetic reader would care because click data carries position bias and tail-query sparsity while manual labeling at hundreds of millions of pairs is infeasible. The resulting model improves offline metrics most on tail queries and lifts online business metrics in a live A/B test.

Core claim

Mining easy and hard positives and negatives from three production retrieval systems, grading them with a three-model LLM cascade, and training a two-tower BERT model with three-stage progressive curriculum on over 240 million examples yields +5.1% NDCG@10 over the click-trained production baseline, drops zero-relevance results from 8.7% to 3.5%, and delivers +2.80% ad spend, +1.4% CTR, +2.8% eCPM, and +2.9% click conversion rate in a two-week online test.

What carries the argument

Multi-channel retrieval mining that extracts rank metadata from three systems, combined with graded-relevance annotation by a calibrated three-model LLM cascade and three-stage progressive curriculum training that organizes examples across five difficulty levels.

If this is right

  • The largest gains occur on tail queries where click data is sparsest.
  • Embarrassing zero-relevance retrievals fall from 8.7% to 3.5%.
  • The same pipeline produces measurable lifts in ad spend, CTR, eCPM, and conversion rate during live traffic.
  • Progressive curriculum training across five difficulty levels organizes the mined data effectively for the final model.

Where Pith is reading between the lines

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

  • The same disagreement-mining step could supply training data for dense retrieval in domains other than e-commerce sponsored search.
  • If the LLM cascade maintains its agreement rate on new query distributions, the approach could reduce dependence on click logs across additional retrieval settings.
  • Replacing the three production systems with a different set of heterogeneous retrievers would test how sensitive the structured signal is to the choice of source systems.

Load-bearing premise

The three-model LLM cascade's 89.1% agreement with trained human annotators yields labels of high enough quality to produce the observed retrieval gains.

What would settle it

Train an otherwise identical two-tower model on the same 240M examples but with human labels instead of the LLM cascade and measure whether NDCG@10 and online metrics remain at or above the reported levels.

Figures

Figures reproduced from arXiv: 2606.23911 by Brahanyaa Somasundaram, Hong Yao, Isha Shah, Jhalak Nilesh Acharya, Kuang-chih Lee, Kumar Priyam, Md Omar Faruk Rokon, Minuteresa Thomas, Shasvat Desai, Vamsee Tangirala, Vijay Manchi, Vivek Arora.

Figure 1
Figure 1. Figure 1: Overview of the training-data pipeline. Top row: Stage 1 collects retrieval results with rank positions from three heterogeneous channels; Stage 2 annotates all pairs via a calibrated model cascade (89.1% agreement with human annotators). Bottom row: Stage 3 classifies pairs into five difficulty levels using channel agreement, rank thresholds, and catalog mining; Stage 4 trains a two-tower BERT model throu… view at source ↗
Figure 2
Figure 2. Figure 2: NDCG@10 by query segment. Tail queries show the largest relative improvement (+6.8%), consistent with the hypothesis that LLM-annotated supervision addresses the data-sparsity challenge inherent to click￾based training [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
read the original abstract

How can we generate high-quality training data for dense retrieval models at production scale, without relying on click signals or manual annotation? This question is critical for e-commerce sponsored search, where click-based training suffers from position bias and tail-query sparsity, and manual labeling at the scale of hundreds of millions of query-item pairs is economically infeasible. Our work is driven by the following insight: heterogeneous retrieval systems disagree on most items they retrieve, and this disagreement creates a natural source of structured training signal -- easy positives where all systems agree, hard positives that only lexical systems find, and hard negatives that fool exactly one system. As our key novelty, we combine three ideas into an end-to-end pipeline: (a) multi-channel retrieval mining with rank metadata from three production systems, (b) graded-relevance annotation by a calibrated three-model cascade ) that reaches 89.1% agreement with trained human annotators, and (c) three-stage progressive curriculum training that organizes 240M+ training examples across five difficulty levels. We deploy the trained two-tower BERT model on Walmart's sponsored search and evaluate it against 30K queries labeled by trained third-party human annotators. First, we show that the system achieves +5.1% NDCG@10 over the click-trained production baseline, with the largest gain on tail queries . Second, we show that embarrassing retrievals (rating 0) drop from 8.7% to 3.5%. Third, a two-week online A/B test with tens of millions of ad requests per arm confirms +2.80% ad spend, +1.4% CTR, +2.8% eCPM, and +2.9% click conversion rate. Overall, our work provides a practical and scalable blueprint for replacing click-based training with structured LLM-annotated supervision in production retrieval systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that structured mining of disagreements across three production retrieval systems, combined with graded annotation from a calibrated three-model LLM cascade (89.1% human agreement) and three-stage progressive curriculum training on 240M+ examples, enables a two-tower BERT dense retriever to outperform a click-trained production baseline by +5.1% NDCG@10 on a 30K human-annotated query set (largest gains on tail queries, embarrassing retrievals reduced from 8.7% to 3.5%), with corresponding gains (+2.80% ad spend, +1.4% CTR, +2.8% eCPM, +2.9% conversion) confirmed in a two-week production A/B test.

Significance. If the end-to-end empirical results hold, the work supplies a practical, scalable alternative to click-based supervision for production sponsored-search retrieval, directly addressing position bias and tail sparsity while demonstrating measurable offline and online improvements. The structured-mining-plus-curriculum approach and the independent human-labeled test set plus live A/B validation are notable strengths.

major comments (2)
  1. [Method (mining and annotation pipeline)] The central claim rests on the quality of the LLM-generated labels, yet the manuscript provides insufficient detail on the exact mining rules, cascade calibration procedure, and data exclusion criteria used to reach the reported 89.1% human agreement; without these, it is difficult to assess whether the label-generation process is reproducible or whether the observed gains could be replicated.
  2. [Offline Evaluation] The offline evaluation reports +5.1% NDCG@10 and the drop in rating-0 retrievals on the 30K human-annotated set, but the manuscript does not report statistical significance tests, confidence intervals, or variance estimates for these metrics; this weakens the strength of the cross-system comparison.
minor comments (2)
  1. [Abstract] The abstract states the three-stage curriculum organizes examples across five difficulty levels; a one-sentence clarification of how the five levels map to the three stages would improve readability.
  2. [Training Curriculum] Table or figure presenting the per-difficulty-level breakdown of the 240M training examples would help readers understand the curriculum composition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment, recognition of the practical contributions, and recommendation for minor revision. We address the two major comments below and will incorporate the requested additions into the revised manuscript.

read point-by-point responses

  1. Referee: [Method (mining and annotation pipeline)] The central claim rests on the quality of the LLM-generated labels, yet the manuscript provides insufficient detail on the exact mining rules, cascade calibration procedure, and data exclusion criteria used to reach the reported 89.1% human agreement; without these, it is difficult to assess whether the label-generation process is reproducible or whether the observed gains could be replicated.

    Authors: We agree that greater detail on the annotation pipeline is needed for reproducibility. In the revised manuscript we will expand the relevant methods section to specify: (i) the exact multi-system mining rules (items retrieved by all three systems as easy positives, by exactly one lexical system as hard positives, and by exactly one neural system as hard negatives, with rank-position thresholds); (ii) the cascade calibration procedure (three-model ensemble with majority vote, temperature scaling on a 2k human-labeled calibration set, and per-grade precision/recall); and (iii) exclusion criteria (queries with <5 candidates, LLM confidence below 0.7, and items with missing metadata). These additions will occupy approximately one page and directly address replicability concerns. revision: yes

  2. Referee: [Offline Evaluation] The offline evaluation reports +5.1% NDCG@10 and the drop in rating-0 retrievals on the 30K human-annotated set, but the manuscript does not report statistical significance tests, confidence intervals, or variance estimates for these metrics; this weakens the strength of the cross-system comparison.

    Authors: We concur that statistical reporting strengthens the claims. In the revision we will add bootstrap (10k resamples) 95% confidence intervals and paired significance tests (Wilcoxon signed-rank) for both NDCG@10 and the rating-0 rate on the 30k set. The observed +5.1% NDCG@10 and reduction from 8.7% to 3.5% remain significant (p < 0.001) under these tests; the intervals will be reported alongside the point estimates in Table 2 and the text. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper presents an empirical pipeline for generating LLM-annotated training data and evaluates the resulting two-tower BERT model directly against an independent 30K human-annotated query set (showing +5.1% NDCG@10) plus a live two-week A/B test on production traffic (showing lifts in ad spend, CTR, eCPM, and conversion). No equations, derivations, or load-bearing claims reduce to fitted parameters by construction, self-citations, or ansatzes imported from prior author work; the 89.1% cascade-human agreement is presented only as supporting context for label quality, not as the source of the reported gains. All central results are measured on external benchmarks outside the training labels themselves.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that retrieval-system disagreement supplies useful training signal and that the LLM cascade produces labels close enough to human quality to improve the downstream model. No free parameters or invented entities are explicitly introduced in the abstract.

axioms (2)
  • domain assumption Heterogeneous retrieval systems disagree on most items they retrieve, creating natural easy positives, hard positives, and hard negatives.
    This is stated as the driving insight for the mining step.
  • domain assumption An LLM cascade calibrated to 89.1% human agreement supplies training labels of adequate quality for the reported gains.
    The agreement figure is presented as evidence that the labels are usable.

pith-pipeline@v0.9.1-grok · 5937 in / 1479 out tokens · 28502 ms · 2026-06-26T06:09:04.553530+00:00 · methodology

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

42 extracted references · 10 linked inside Pith

  1. [1]

    The role of relevance in sponsored search

    Luca Maria Aiello, Ioannis Arapakis, Ricardo Baeza-Yates, Xiao Bai, Nicola Barbieri, Amin Mantrach, and Fabrizio Silvestri. The role of relevance in sponsored search. InProceedings of the 25th ACM International Conference on Information and Knowledge Management, pages 185–194, 2016

    2016

  2. [2]

    Curriculum learning

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, pages 41–48, 2009

    2009

  3. [3]

    InPars: Data augmentation for information retrieval using large language models

    Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Nogueira. InPars: Data augmentation for information retrieval using large language models. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2316–2320, 2022

    2022

  4. [4]

    FrugalGPT: How to use large language models while reducing cost and improving performance.arXiv preprint arXiv:2305.05176, 2023

    Lingjiao Chen, Matei Zaharia, and James Zou. FrugalGPT: How to use large language models while reducing cost and improving performance.arXiv preprint arXiv:2305.05176, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Deep neural networks for YouTube recommendations

    Paul Covington, Jay Adams, and Emre Sargin. Deep neural networks for YouTube recommendations. InProceedings of the 10th ACM Conference on Recommender Systems, 2016

    2016

  6. [6]

    Promptagator: Few-shot dense retrieval from 8 examples

    Zhuyun Dai, Arun Tejasvi Chaganty, Vincent Y Zhao, Aida Rashid, Mike Green, and Kelvin Guu. Promptagator: Few-shot dense retrieval from 8 examples. InInternational Conference on Learning Representations, 2023

    2023

  7. [7]

    Unified supervision for walmarts sponsored search retrieval via joint semantic relevance and behavioral engagement modeling.arXiv preprint arXiv:2604.07930, 2026

    Shasvat Desai, Md Omar Faruk Rokon, Jhalak Nilesh Acharya, Isha Shah, Hong Yao, Utkarsh Porwal, and Kuang-chih Lee. Unified supervision for walmarts sponsored search retrieval via joint semantic relevance and behavioral engagement modeling.arXiv preprint arXiv:2604.07930, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    BERT: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2019

    Pith/arXiv arXiv 2019

  9. [9]

    Perspectives on large language models for relevance judgment

    Guglielmo Faggioli, Laura Dietz, Charles Clarke, Gianluca Demartini, Matthias Hagen, Claudia Hauff, Noriko Kando, Evangelos Kanoulas, Martin Potthast, Benno Stein, and Henning Wachsmuth. Perspectives on large language models for relevance judgment. InProceedings of the 2023 ACM SIGIR International Conference on the Theory of Information Retrieval, pages 3...

    2023

  10. [10]

    MOBIUS: Towards the next generation of query-ad matching in Baidu’s sponsored search

    Miao Fan, Jiacheng Guo, Shuai Zhu, Shuo Miao, Mingming Sun, and Ping Li. MOBIUS: Towards the next generation of query-ad matching in Baidu’s sponsored search. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 2509–2517, 2019

    2019

  11. [11]

    SPLADE: Sparse lexical and expansion model for first stage ranking

    Thibault Formal, Benjamin Piwowarski, and Stéphane Clinchant. SPLADE: Sparse lexical and expansion model for first stage ranking. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2288–2292, 2021

    2021

  12. [12]

    Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

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

    Pith/arXiv arXiv 2015

  13. [13]

    Improving efficient neural ranking models with cross-architecture knowledge distillation.arXiv preprint arXiv:2010.02666, 2020

    Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. Improving efficient neural ranking models with cross-architecture knowledge distillation.arXiv preprint arXiv:2010.02666, 2020

    arXiv 2010

  14. [14]

    Embedding-based retrieval in Facebook search

    Jui-Ting Huang, Ashish Sharma, Shuying Sun, Li Xia, David Zhang, Philip Pronin, Janani Padman- abhan, Giuseppe Ottaviano, and Linjun Yang. Embedding-based retrieval in Facebook search. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 2553–2561, 2020

    2020

  15. [15]

    Learning deep structured semantic models for web search using clickthrough data

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. Learning deep structured semantic models for web search using clickthrough data. InACM International Conference on Information and Knowledge Management (CIKM), 2013

    2013

  16. [16]

    Cumulated gain-based evaluation of IR techniques.ACM Transactions on Information Systems, 20(4):422–446, 2002

    Kalervo Järvelin and Jaana Kekäläinen. Cumulated gain-based evaluation of IR techniques.ACM Transactions on Information Systems, 20(4):422–446, 2002

    2002

  17. [17]

    Dense passage retrieval for open-domain question answering

    Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769–6781, 2020

    2020

  18. [18]

    Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281, 2023

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2019

    Pith/arXiv arXiv 2019

  20. [20]

    Semantic retrieval at Walmart

    Alessandro Magnani, Feng Liu, Suthee Chaidaroon, Sachin Yadav, Praveen Reddy Suram, Ajit Puthenputhussery, Sijie Chen, Min Xie, Anirudh Kashi, Tony Lee, et al. Semantic retrieval at Walmart. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3495–3503, 2022

    2022

  21. [21]

    Distant supervision for relation extraction without labeled data

    Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. Distant supervision for relation extraction without labeled data. InProceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003–1011, 2009

    2009

  22. [22]

    Learning with noisy labels

    Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. InAdvances in Neural Information Processing Systems (NeurIPS), 2013

    2013

  23. [23]

    Semantic product search

    Priyanka Nigam, Yiwei Song, Vijai Mohan, Vihan Lakshman, Weitian Ding, Ankit Shingavi, Choon Hui Teo, Hao Gu, and Bing Yin. Semantic product search. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 2876–2885, 2019

    2019

  24. [24]

    RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering

    Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics, pages 5835–5847, 2021

    2021

  25. [25]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the 38th International Conference on Machine Learning (ICML), 2021

    2021

  26. [26]

    Shopping queries dataset: A large-scale ESCI benchmark for improving product search

    Chandan K Reddy, Lluís Màrquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopad- hyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. Shopping queries dataset: A large-scale ESCI benchmark for improving product search. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4429–4439, 2022

    2022

  27. [27]

    Sentence-BERT: Sentence embeddings using siamese BERT- networks.arXiv preprint arXiv:1908.10084, 2019

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT- networks.arXiv preprint arXiv:1908.10084, 2019

    Pith/arXiv arXiv 1908

  28. [28]

    Making monolingual sentence embeddings multilingual using knowledge distillation

    Nils Reimers and Iryna Gurevych. Making monolingual sentence embeddings multilingual using knowledge distillation. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020

    2020

  29. [29]

    The probabilistic relevance framework: BM25 and beyond

    Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval, 3(4):333–389, 2009

    2009

  30. [30]

    Contrastive learning with hard negative samples

    Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. InInternational Conference on Learning Representations, 2021

    2021

  31. [31]

    Enhancement of e-commerce sponsored search relevancy with LLM

    Md Omar Faruk Rokon, Andrei Simion, Weizhi Du, Musen Wen, Hong Yao, and Kuang-chih Lee. Enhancement of e-commerce sponsored search relevancy with LLM. InProceedings of the SIGIR Workshop on eCommerce (eCom’24), 2024

    2024

  32. [32]

    Deep learning is robust to massive label noise.arXiv preprint arXiv:1705.10694, 2017

    David Rolnick, Andreas Veit, Serge Belongie, and Nir Shavit. Deep learning is robust to massive label noise.arXiv preprint arXiv:1705.10694, 2017

    Pith/arXiv arXiv 2017

  33. [33]

    User intent, behaviour, and perceived satisfaction in product search

    Ning Su, Jiyin He, Yiqun Liu, Min Zhang, and Shaoping Ma. User intent, behaviour, and perceived satisfaction in product search. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining, pages 547–555, 2018

    2018

  34. [34]

    BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021

    2021

  35. [35]

    Large language models can accurately predict searcher preferences

    Paul Thomas, Seth Spielman, Nick Craswell, and Bhaskar Mitra. Large language models can accurately predict searcher preferences. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1930–1940, 2024

    1930

  36. [36]

    Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533, 2022

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533, 2022

    Pith/arXiv arXiv 2022

  37. [37]

    Click- conversion multi-task model with position bias mitigation for sponsored search in ecommerce

    Yuanxing Wang, Yaqing Xue, Buyun Liu, Musen Wen, Wenjia Zhao, and Song Guo. Click- conversion multi-task model with position bias mitigation for sponsored search in ecommerce. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2023

    2023

  38. [38]

    Semantic ads retrieval at Walmart ecommerce with language models progressively trained on multiple knowledge domains.arXiv preprint arXiv:2502.09089, 2025

    Zhaodong Wang, Weizhi Du, Md Omar Faruk Rokon, Prabir Adhikary, Yaqing Xue, Jian Xu, Jingyi Zhou, Kuang-chih Lee, and Musen Wen. Semantic ads retrieval at Walmart ecommerce with language models progressively trained on multiple knowledge domains.arXiv preprint arXiv:2502.09089, 2025

    arXiv 2025

  39. [39]

    C-Pack: Packed resources for general Chinese embeddings.arXiv preprint arXiv:2309.07597, 2023

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-Pack: Packed resources for general Chinese embeddings.arXiv preprint arXiv:2309.07597, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Approximate nearest neighbor negative contrastive learning for dense text retrieval

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. Approximate nearest neighbor negative contrastive learning for dense text retrieval. InInternational Conference on Learning Representations, 2021

    2021

  41. [41]

    Optimizing dense retrieval model training with hard negatives

    Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. Optimizing dense retrieval model training with hard negatives. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1503–1512, 2021

    2021

  42. [42]

    Judging LLM-as-a-Judge with MT-Bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P Xing, Hao Zhang, Joseph E Gonzalez, and Ion Stoica. Judging LLM-as-a-Judge with MT-Bench and chatbot arena. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023