Aligning Web Query Generation with Ranking Objectives via Direct Preference Optimization

Bruno Martins; Chenyan Xiong; Jo\~ao Coelho; Jo\~ao Magalh\~aes

REVIEW 4 major objections 4 minor 40 references

Aligning Web Query Generation with Ranking Objectives via Direct Preference Optimization

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Aligning synthetic query generation with ranking feedback via Direct Preference Optimization improves downstream dense retrieval on MS MARCO and TREC-DL, and does so more data-efficiently than post-hoc filtering.

desk verdict A sensible DPO-based alternative to post-hoc filtering for synthetic query generation, with solid in-distribution gains, but the robustness claims outrun the evidence. read the letter →

arxiv 2505.19307 v1 pith:DYRZAM5E submitted 2025-05-25 cs.IR

João Coelho , Bruno Martins , João Magalhães , Chenyan Xiong This is my paper

classification cs.IR

keywords querygenerationdirectpreferenceoptimizationdenseretrievalsynthetictrainingdatarankingfeedbackMSMARCOClueWeb22efficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that synthetic query generation for dense retrieval training should not stop at filtering noisy queries after the fact. Instead, ranking feedback from a cross-encoder ranker or from an LLM list-wise judge can be turned into pairwise preferences, and Direct Preference Optimization can nudge the query generator toward queries that score higher. The authors show that a 0.5B dense retriever trained on one million queries from the aligned generator reaches 0.3795 MRR@100 on MS MARCO document retrieval, and that the same pipeline retains 92% of generated queries under consistency filtering versus 62% before alignment. The point is that aligning generation with the ranking objective is more data-efficient and produces more web-like, searchable queries.

What carries the argument

The load-bearing mechanism is a preference dataset built from ranking signals, followed by the Direct Preference Optimization (DPO) loss. For each of 100k ClueWeb22 documents, the baseline generator samples five queries; either a cross-encoder scores each query-document pair, or an LLM list-wise judge is prompted to pick the best and worst queries for the document. Pairs of preferred and dispreferred queries are then used in the DPO objective, which raises the likelihood of the ranker-preferred query without training a separate reward model. The aligned generator is then used once per document to create positive training pairs for a bi-encoder retriever, with Gecko-style relabeling and hard negative sampling.

What would settle it

Train the same dense retriever on the same number of queries from the baseline and the DPO-aligned generators, then evaluate on a web retrieval benchmark whose relevance labels come from a different annotation process than MS MARCO; if the MRR gain does not transfer, the ranker reward was overfit to the evaluation distribution.

Watch

Extended reading notes

Core claim

The central claim is that a query generator fine-tuned with DPO on preferences derived from ranking models yields synthetic query-document pairs that train better dense retrievers than pairs from the same generator without alignment, and better than pairs that only pass through filtering. The paper demonstrates this on MS MARCO document and passage retrieval and on TREC-DL 2019 and 2020, with both a point-wise ranker reward and a GPT-based list-wise reward outperforming open-source baselines trained on synthetic data, including the generator's own pre-alignment data. The authors interpret the mechanism as shifting the generator toward concise, web-like queries that preserve the document's core intent.

Load-bearing premise

The whole pipeline assumes the ranker (or LLM judge) scores are a trustworthy proxy for real retrieval quality; if those scores are biased or overfit to the evaluation distribution, the generator is aligned to that bias and the benchmark gains may not transfer.

Editorial extensions

If this is right

Retrievers trained on DPO-aligned synthetic queries beat those trained on the same volume of unaligned synthetic queries on MS MARCO document and passage benchmarks.
Alignment works with either a point-wise cross-encoder ranker or an LLM list-wise judge, so the framework does not depend on one particular reward model.
Query retention under consistency filtering rises from 62% to 92% after alignment, meaning less generated data is wasted and smaller pools can suffice.
Performance keeps improving as the number of synthetic training pairs grows from 500K to 1M, so the approach scales with available compute.

Reading between the lines

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

A natural extension the authors leave implicit is to test whether the ranker reward generalizes beyond MS MARCO-like queries, for example by building preference datasets from several diverse rankers and comparing their agreement.
Because the recipe only needs paired preferences, it could transfer to other synthetic-data pipelines, such as generating questions for retrieval-augmented generation or synthetic examples for reranker training.
The observed rightward shift in ranker scores is evidence of alignment to that ranker, not necessarily to human relevance; human judgments on a sample of aligned versus baseline queries would separate the two.
If the ranker reward favors concise queries, the generator may drop useful detail on long-tail or multi-intent documents; testing such cases would reveal whether the conciseness comes at a cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a clean, useful short paper that shows DPO-aligned query generation beats post-hoc filtering for synthetic training data in dense retrieval. The core comparison—same generator before and after DPO—moves MRR@100 from 0.318 to 0.379 on MS MARCO Doc and retention from 62% to 92%. That's a real effect, and the paper does a good job of positioning it against the Doc2Query/InPars/Promptagator/Gecko family.

What's actually new: applying DPO to the query generator itself, with preferences derived from a reranker or an LLM listwise judge, rather than filtering at inference time. That is more than a trivial extension, because it changes the generator's behavior (see Table 3, verbose questions to keyword queries) and pays off in the head-to-head comparison. The paper is honest about the setup: one epoch, LoRA, 10-hour overhead, public code.

Soft spots:

Section 4.4 is partly circular: the reward model R that defines the DPO preference is the same one used to show that DPO improves 'ranker-assessed relevance.' Seeing R scores rise after training against R is not evidence of general quality. The independent evidence is the downstream retrieval numbers and the retention rate, but those numbers are all within the MS MARCO/TREC-DL distribution, which is the same distribution the reward model likely knows well. So the 'robustness' claim is untested; a transfer test to, say, BEIR or a non-web domain would be needed to rule out reward-distribution alignment. That's a real gap, but it is a limitation, not a fatal flaw—the method targets Web retrieval and works on the standard benchmark.

Two smaller issues: Table 2 compares 156K vs 231K kept queries, so the before/after comparison is confounded by dataset size; the paper should report a matched-size ablation or at least error bars. Section 3.3 says positives not in top-100 are relabeled, but Table 2's retention discards them; the relation between that protocol and the final models isn't specified. Both are fixable in a revision.

Bottom line: a serious referee should engage with this. The central claim that integrating ranking objectives into generation helps is supported by the independent retrieval metrics; the circularity weakens the internal validation but doesn't falsify the method. I'd ask for matched-size comparisons, multiple runs, and an out-of-distribution evaluation. If you work on synthetic data for dense retrieval, it's worth a read and worth citing.

Referee Report

4 major / 4 minor

Summary. The paper proposes using Direct Preference Optimization (DPO) to align an LLM-based synthetic query generator with ranking preferences obtained either from a pointwise reranker (bge-reranker-v2-m3) or from a listwise LLM (GPT-3.5). After DPO, the aligned generator produces one query per document from a ClueWeb22-B sample, and these query-document pairs are used to train a Qwen2.5-0.5B dense retriever. Experiments report strong MS MARCO document/passage and TREC-DL results, a higher retention rate (62% vs. 92%) and higher MRR in a 250K-query before/after comparison, and scaling improvements as training pairs increase.

Significance. If the causal interpretation holds, the framework is a useful step beyond post hoc query filtering, with practical value for synthetic data pipelines and for aligning generators directly to ranking objectives. The exploration of two reward families (pointwise and listwise) and the scaling analysis are strengths, and the public code/model release is a plus. However, the current evidence partially conflates DPO alignment with dataset size and partly validates the method using the same reward model that defined the DPO objective, so the significance of the paper depends on addressing the confounds and robustness concerns detailed below.

major comments (4)

[§4.4, Table 2] The before/after DPO comparison is confounded by dataset size: the before-DPO model is trained on 156K kept queries and the after-DPO model on 231K kept queries. Because the two retrieval models see different amounts of training data, the MRR@100 gain from 0.1635 to 0.2625 cannot be attributed solely to query-quality improvements from DPO. Please add a controlled comparison with matched numbers of training pairs (e.g., subsample the after-DPO kept queries to 156K, or train the before-DPO model on 231K by relaxing the consistency filter) and report multiple seeds or error bars.
[§4.4, Figure 1] This KDE plot shows the distribution of R(q,d) where R is exactly the ranker used to construct the preference dataset Drr in Eq. (2). Since DPO optimizes the generator against preferences derived from R, a rightward shift in R scores after training is an expected manipulation check rather than independent evidence of improved query quality. The non-circular evidence is the retrieval effectiveness in Table 2, but that comparison inherits the confound described above; either add a held-out ranker (not used in DPO) or relegate Figure 1 to a manipulation check.
[§3.3 and §4.4] Section 3.3 states that when the original positive is not in the top-100, the top-1 document is used as the new positive (relabeling), while Section 4.4's retention metric discards such queries. The paper does not state which protocol was used to train the Table 2 models or the Table 1 models. This ambiguity matters because the retention increase from 62% to 92% is only meaningful if the discarding protocol matches the training data construction; please clarify and align the evaluation protocol with the training protocol.
[§4.2 and §5] All downstream evaluations are on MS MARCO document/passage retrieval and TREC-DL 2019/2020, all of which draw queries from the MS MARCO distribution, and the reward model bge-reranker-v2-m3 is a web-trained reranker likely exposed to similar data. The conclusion that DPO yields 'more robust and data-efficient systems' is consequently only supported inside this distribution. Please add an out-of-distribution evaluation (e.g., a BEIR subset or another query domain) or explicitly limit the robustness claim to in-distribution settings.

minor comments (4)

[§4.1 and Eq. (4)] The DPO temperature beta in Eq. (4) and the LoRA rank/alpha are never specified; without these values the DPO training is not reproducible.
[Eq. (4)] The roles of M_g and M*_g are ambiguous; standard DPO requires a frozen reference model and a trainable policy, but the text says M_g is trained and M*_g denotes the aligned model. Clarify which distribution is the reference and which is the policy.
[§4.1] The prompt templates for the contrastive generation (Section 3.1) and for the GPT-3.5 listwise reward (Eq. (3)) are omitted; please include them or point to the repository.
[Table 1] The comparison mixes model size and number of training pairs; a same-backbone baseline trained on the raw (pre-DPO) queries with identical pair counts would isolate the effect of DPO.

Circularity Check

1 steps flagged · score 4.0 of 10

Central retrieval results are independent, but Section 4.4 validates ranker-DPO with the same reward model used to build its DPO preferences, making that internal validation circular.

fitted input called prediction [Section 4.4 (Figure 1 and surrounding text), cf. Eq. (2) and Eq. (4)]
"To further consolidate these findings, we examined the distribution of ranker reward scores R(q, d) for a sample of 10k query-document pairs. As depicted in the KDE plot in Figure 1, the post-DPO score distribution exhibits a pronounced rightward shift relative to the one from the base generator. This reflects an improvement in query–document relevance, thereby validating the efficacy of our DPO framework in steering the generator toward higher-quality outputs."

For the ranker-DPO variant, R is the very model that defines the preference data Drr in Eq. (2) (pairs with R(q+,d) > R(q-,d)) and that Eq. (4) uses to raise the likelihood of preferred queries. A rightward shift in R(q,d) after DPO is therefore a direct measure of the training reward, not an independent signal of query quality. That DPO increases scores on the reward model it was aligned to is the expected outcome of the optimization, so Figure 1 cannot be used to 'validate the efficacy' of the framework. The downstream retrieval numbers in Table 1 are the non-circular evidence; the reward-shift validation is circular by construction.

full rationale

The central derivation (synthetic queries + DPO preferences → retriever training → MS MARCO/TREC-DL retrieval numbers) is not circular: Table 1 and the MRR/nDCG numbers are measured with standard retrieval metrics and do not reuse the DPO reward model. The GPT-DPO variant's preferences come from Eq. (3) (GPT-3.5), not from the ranker R, so its retrieval gains are independent of the ranker-reward circularity. The one clear circular validation is Section 4.4's ranker-reward KDE: for ranker-DPO, R defines the preference data (Eq. 2) and the DPO loss (Eq. 4), so observing a rightward shift in R(q,d) after training is largely a check of the training objective, not independent evidence of query-quality improvement. Table 2's retention/MRR comparison is a separate, non-circular downstream signal, although the paper leaves a protocol ambiguity: Section 3.3 relabels positives not in top-100, while Section 4.4 discards them, and the relationship between the Table 2 protocol and the final models is not fully specified. No load-bearing self-citation chains were found; ClueWeb22 [27] and CMT [37] are data/method citations, not uniqueness or correctness arguments. The possible overlap of bge-reranker-v2-m3's training data with MS MARCO is a distributional/contamination concern, not a circularity of the derivation, and does not by itself raise the circularity score. Overall the central claim has independent content, so the score is 4 rather than 6+.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central contribution is empirical, so the ledger is dominated by hyperparameters and by the assumption that ranking rewards are valid proxies for relevance. No new theoretical entities are introduced. The reward model itself is an external artifact, not invented here, but its validity is the load-bearing assumption.

free parameters (5)

DPO temperature beta
Hyperparameter in Eq. (4) controlling preference strength; value not reported in the paper.
Queries per document during preference collection (n) = 5
Used in Eq. (1) to build the candidate query pool; chosen as a fixed number without sensitivity analysis.
DPO alignment document pool size = 100k
Section 4.1 says this was selected as a practical upper bound for training cost; not varied.
Downstream training pairs = 1M (final models)
Figure 2 scales from 500K to 1M pairs; final reported models in Table 1 use 1M pairs.
LoRA rank/alpha
LoRA is used for DPO training in Section 4.1, but rank and alpha are not specified; this affects the update and is not reported.

assumptions (4)

domain assumption Ranker scores R(q,d) from bge-reranker-v2-m3 and GPT-3.5 listwise judgments are reliable proxies for downstream retrieval relevance.
Used to build preference datasets in Eq. (2) and Eq. (3); if false, DPO aligns to a biased reward.
domain assumption MS MARCO and TREC-DL judgments adequately measure web retrieval effectiveness, so improvements on them generalize.
All headline results are evaluated on these benchmarks in Section 4.2; no out-of-domain evaluation is provided.
standard math DPO loss is an appropriate and stable way to fine-tune the generator; the reference model M_g is kept fixed.
Adopts the standard DPO formulation from Eq. (4) following Rafailov et al.; no formal proof of convergence or stability is provided, which is common in applied work.
domain assumption The 35M-document ClueWeb22-B sample is representative enough that queries generated for it transfer to MS MARCO.
DPO training and all downstream positive pairs are generated from ClueWeb22-B while evaluation is on MS MARCO/TREC-DL, as described in Sections 4.1 and 4.2.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Aligning Web Query Generation with Ranking Objectives via Direct Preference Optimization." pith.science (2026). https://pith.science/paper/DYRZAM5E

@misc{pith2026250519307,
  author       = {Pith},
  title        = {Pith review of: Aligning Web Query Generation with Ranking Objectives via Direct Preference Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYRZAM5E}},
  note         = {Machine review of arXiv:2505.19307}
}

read the original abstract

Neural retrieval models excel in Web search, but their training requires substantial amounts of labeled query-document pairs, which are costly to obtain. With the widespread availability of Web document collections like ClueWeb22, synthetic queries generated by large language models offer a scalable alternative. Still, synthetic training queries often vary in quality, which leads to suboptimal downstream retrieval performance. Existing methods typically filter out noisy query-document pairs based on signals from an external re-ranker. In contrast, we propose a framework that leverages Direct Preference Optimization (DPO) to integrate ranking signals into the query generation process, aiming to directly optimize the model towards generating high-quality queries that maximize downstream retrieval effectiveness. Experiments show higher ranker-assessed relevance between query-document pairs after DPO, leading to stronger downstream performance on the MS~MARCO benchmark when compared to baseline models trained with synthetic data.

Figures

Figures reproduced from arXiv: 2505.19307 by the authors.

**Figure 2.** Performance on MARCO Doc as synthetic training pairs increase. 4.5 Scaling Synthetic Data We also present an empirical investigation of how model performance scales with synthetic training pairs [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 27 canonical work pages

[1]

Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders.ArXivabs/2404.05961 (2024)

arXiv 2024
[2]

Luiz Henrique Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Frassetto Nogueira. 2022. InPars: Data Augmentation for Information Retrieval using Large Language Models.ArXivabs/2202.05144 (2022)

arXiv 2022
[3]

Nick Craswell. 2009. Mean Reciprocal Rank. InEncyclopedia of Database Systems

work page 2009
[4]

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the TREC 2020 Deep Learning Track. InText REtrieval Conference (TREC)

work page 2021
[5]

Voorhees

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. Overview of the TREC 2019 Deep Learning Track. InText REtrieval Conference (TREC)

work page 2020
[6]

Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B

Zhuyun Dai, Vincent Y. Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2023. Promptagator: Few- shot Dense Retrieval From 8 Examples. InInternational Conference on Learning Representations (ICLR 2023)

work page 2023
[7]

Abhimanyu Dubey et al. 2024. The Llama 3 Herd of Models.ArXivabs/2407.21783 (2024)

arXiv 2024
[8]

Yan Fang, Jingtao Zhan, Qingyao Ai, Jiaxin Mao, Weihang Su, Jia Chen, and Yiqun Liu. 2024. Scaling Laws For Dense Retrieval. InInternational Conference on Research and Development in Information Retrieval (SIGIR 2024)

work page 2024

Show all 40 references

[9]

Luyu Gao and Jamie Callan. 2022. Unsupervised Corpus Aware Language Model Pre-training for Dense Passage Retrieval. InAnnual Meeting of the Association for Computational Linguistics (ACL 2022)

2022
[10]

Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. 2023. Doc2Query-: When Less is More. InEuropean Conference on Information Retrieval (ECIR 2023)

2023
[11]

Peixuan Han, Zhenghao Liu, Zhiyuan Liu, and Chenyan Xiong. 2024. En- hancing Dense Retrievers’ Robustness with Group-level Reweighting.ArXiv abs/2310.16605 (2024)

2024 arXiv
[12]

Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations, (ICLR 2022)

2022
[13]

Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bo- janowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised Dense Infor- mation Retrieval with Contrastive Learning.Transactions on Machine Learning Research(2022)

2022
[14]

Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques.ACM Transactions on Information Systems(2002)

2002
[15]

Vitor Jeronymo, Luiz Henrique Bonifacio, Hugo Abonizio, Marzieh Fadaee, Roberto de Alencar Lotufo, Jakub Zavrel, and Rodrigo Frassetto Nogueira. 2023. InPars-v2: Large Language Models as Efficient Dataset Generators for Informa- tion Retrieval.ArXivabs/2301.01820 (2023)

2023 arXiv
[16]

Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick S. H. Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering. InConference on Empirical Methods in Natural Language Processing (EMNLP 2020)

2020
[17]

Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. 2024. NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models.ArXivabs/2405.17428 (2024)

2024 arXiv
[18]

Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R. Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, Yi Luan, Sai Meher Karthik Duddu, Gustavo Hernández Ábrego, Weiqiang Shi, Nithi Gupta, Aditya Kusupati, Prateek Jain, Siddhartha Reddy Jonnalagadda, ...

2024 arXiv
[19]

Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent Retrieval for Weakly Supervised Open Domain Question Answering. InAnnual Meeting of the Association for Computational Linguistics (ACL 2019)

2019
[20]

Shuqi Lu, Di He, Chenyan Xiong, Guolin Ke, Waleed Malik, Zhicheng Dou, Paul Bennett, Tie-Yan Liu, and Arnold Overwijk. 2021. Less is More: Pretrain a Strong Siamese Encoder for Dense Text Retrieval Using a Weak Decode. InConference on Empirical Methods in Natural Language Proc...

2021
[21]

Guangyuan Ma, Xing Wu, Zijia Lin, and Songlin Hu. 2024. Drop your Decoder: Pre-training with Bag-of-Word Prediction for Dense Passage Retrieval.ArXiv abs/2401.11248 (2024)

2024 arXiv
[22]

Xinyu Ma, Jiafeng Guo, Ruqing Zhang, Yixing Fan, and Xueqi Cheng. 2022. Pre-train a Discriminative Text Encoder for Dense Retrieval via Contrastive Span Prediction. InInternational Conference on Research and Development in Information Retrieval (SIGIR 2022)

2022
[23]

Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A Human Generated MAchine Read- ing COmprehension Dataset. InWorkshop on Cognitive Computation: Integrating Neural and Symbolic Approaches

2016
[24]

2019.From doc2query to docTTTTTquery

Rodrigo Nogueira and Jimmy Lin. 2019.From doc2query to docTTTTTquery. Technical Report. University of Waterloo

2019
[25]

Rodrigo Frassetto Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document Expansion by Query Prediction.ArXivabs/1904.08375 (2019)

2019 arXiv
[26]

Yichen Ouyang, Lu Wang, Fangkai Yang, Pu Zhao, Chenghua Huang, Jianfeng Liu, Bochen Pang, Yaming Yang, Yuefeng Zhan, Hao Sun, Qingwei Lin, Saravan Rajmohan, Weiwei Deng, Dongmei Zhang, Feng Sun, and Qi Zhang. 2024. Token- level Proximal Policy Optimization for Query Generation...

2024 arXiv
[27]

Arnold Overwijk, Chenyan Xiong, and Jamie Callan. 2022. ClueWeb22: 10 Billion Web Documents with Rich Information. InInternational Conference on Research and Development in Information Retrieval (SIGIR 2022)

2022
[28]

Manning, Stefano Ermon, and Chelsea Finn

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. InConference on Neural Information Processing Systems (NeurIPS 2023)

2023
[29]

Robertson and Hugo Zaragoza

Stephen E. Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond.Found. Trends Inf. Retr.(2009)

2009
[30]

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov
[31]

Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. 2023. Is ChatGPT Good at Search? Investi- gating Large Language Models as Re-Ranking Agents. InConference on Empirical Methods in Natural Language Processing (EMNLP 2023)

2023
[32]

Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding.ArXivabs/1807.03748 (2018)

2018 arXiv
[33]

Gomez, Lukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. InConference on Neural Information Processing Systems (NeurIPS 2017)

2017
[34]

Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text Embeddings by Weakly-Supervised Contrastive Pre-training.ArXivabs/2212.03533 (2022)

2022 arXiv
[35]

Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Improving Text Embeddings with Large Language Models. In Annual Meeting of the Association for Computational Linguistics (ACL 2024)

2024
[36]

Shitao Xiao, Zheng Liu, Yingxia Shao, and Zhao Cao. 2022. RetroMAE: Pre- Training Retrieval-oriented Language Models Via Masked Auto-Encoder. In Conference on Empirical Methods in Natural Language Processing (EMNLP 2022)

2022
[37]

Chenyan Xiong, Zhenghao Liu, Si Sun, Zhuyun Dai, Kaitao Zhang, Shi Yu, Zhiyuan Liu, Hoifung Poon, Jianfeng Gao, and Paul Bennett. 2020. CMT in TREC-COVID Round 2: Mitigating the Generalization Gaps from Web to Special Domain Search.ArXivabs/2011.01580 (2020)

2020 arXiv
[38]

Bennett, Junaid Ahmed, and Arnold Overwijk

Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate Nearest Neighbor Neg- ative Contrastive Learning for Dense Text Retrieval. InInternational Conference on Learning Representations (ICLR 2021)

2021
[39]

An Yang et al. 2024. Qwen2.5 Technical Report.ArXivabs/2412.15115 (2024)

2024 arXiv
[2017]

Proximal Policy Optimization Algorithms.ArXivabs/1707.06347 (2017)

2017 arXiv

Pith tools

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

[1] [1]

Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders.ArXivabs/2404.05961 (2024)

arXiv 2024

[2] [2]

Luiz Henrique Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Frassetto Nogueira. 2022. InPars: Data Augmentation for Information Retrieval using Large Language Models.ArXivabs/2202.05144 (2022)

arXiv 2022

[3] [3]

Nick Craswell. 2009. Mean Reciprocal Rank. InEncyclopedia of Database Systems

work page 2009

[4] [4]

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the TREC 2020 Deep Learning Track. InText REtrieval Conference (TREC)

work page 2021

[5] [5]

Voorhees

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. Overview of the TREC 2019 Deep Learning Track. InText REtrieval Conference (TREC)

work page 2020

[6] [6]

Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B

Zhuyun Dai, Vincent Y. Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2023. Promptagator: Few- shot Dense Retrieval From 8 Examples. InInternational Conference on Learning Representations (ICLR 2023)

work page 2023

[7] [7]

Abhimanyu Dubey et al. 2024. The Llama 3 Herd of Models.ArXivabs/2407.21783 (2024)

arXiv 2024

[8] [8]

Yan Fang, Jingtao Zhan, Qingyao Ai, Jiaxin Mao, Weihang Su, Jia Chen, and Yiqun Liu. 2024. Scaling Laws For Dense Retrieval. InInternational Conference on Research and Development in Information Retrieval (SIGIR 2024)

work page 2024