Pith. sign in

REVIEW 3 major objections 5 minor 69 references

On the Merits of LLM-Based Corpus Enrichment

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LLM-generated, query-biased documents can be ranked above the original relevant documents and improve retrieval, RAG, and attribution.

desk verdict A useful proof of concept that LLM-generated query-biased documents can be highly retrievable, but the RAG and attribution claims rest on oracle selection and lack a control that isolates LLM rewriting from simply adding relevant content. read the letter →

arxiv 2506.06015 v1 pith:SYXBLT2B submitted 2025-06-06 cs.IR

Gal Zur , Tommy Mordo , Moshe Tennenholtz , Oren Kurland This is my paper
classification cs.IR
keywords corpusenrichmentlargelanguagemodelsadhocretrievalaugmentedgenerationanswerattributionfaithfulnessquery-biasedsummarizationretrievability
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

This paper argues that a document corpus can be improved offline by asking an LLM to rewrite or summarize existing relevant documents into new, query-biased documents that are easier to retrieve. In experiments on two passage-retrieval collections, generated documents were ranked far above the median rank of existing relevant documents, and the enriched corpus improved ad hoc retrieval effectiveness. The same enrichment raised the accuracy of retrieval-augmented generation and improved the ability to find passages that support answers. The paper also proposes NLI-based faithfulness measures showing that summaries grounded in source documents stay faithful to corpus content, while zero-shot generation drifts. If the effect holds beyond these proof-of-concept settings, search engines and content owners could mitigate hard-query failures by generating more retrievable versions of content.

What carries the argument

The mechanism is prompt-based document generation: an LLM is given a query and either nothing, one relevant document, or two or three relevant documents, and asked to write a short paragraph that answers the query and would rank higher than other paragraphs of the same topic. A second component is the NLI-based faithfulness measure, which checks whether each sentence of a generated document is entailed by up to k source documents found by a greedy hill-climbing search. The generated documents are inserted into the corpus as new retrievable units, so the same retrieval function that struggled with the original query now has an easier surface to match.

What would settle it

Run the same enrichment pipeline with source documents selected by a pseudo-relevance method instead of gold judgments, and measure NDCG@10 and the median rank of the generated documents; if the generated documents no longer outrank the median relevant document, or if human annotators rate most of them non-relevant, the central claim that enrichment improves retrievability would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that relevant documents generated from existing ones can be substantially more retrievable than the documents they are based on: across three ranking functions, the median rank of one generated document per query was lower than the median rank of existing relevant documents, often dramatically so. Because most generated documents were judged relevant by human annotators, adding them to the corpus improved NDCG@10, NDCG@100, and MAP@100. For RAG, enriching the corpus changed the retrieval stage so that a correct answer appeared in the top five documents for roughly 60 to 77 percent of questions, compared with 45.9 percent without enrichment, yielding statistically significant accuracy gains. For attribution, using the enriched corpus both for RAG and for selecting supporting passages gave the highest entailment accuracy. The paper frames this as a first proof of concept for corpus enrichment as a general interface between generative models and search.

Load-bearing premise

The whole pipeline assumes you already know which existing documents are relevant before you can generate documents worth adding, using relevance judgments in the ad hoc experiments and an oracle that knows which passages contain the true answer in the RAG experiments.

Editorial extensions

If this is right

  • Hard queries with few relevant documents can be made easier by inserting one LLM-generated query-biased document per query, since generated documents are retrieved above the median rank of existing relevant documents.
  • Retrieval-augmented generation improves because the answer is more likely to appear in the top-five retrieved documents after enrichment, raising answer-in-top-five coverage from 45.9 percent to roughly 60 to 77 percent.
  • Attribution improves most when the same enriched corpus is used both to generate answers and to select supporting passages, with entailment accuracy reaching 74.5 percent in the best reported setting.
  • Corpus-grounded generation methods keep faithfulness above 90 percent by the paper's NLI measure, while zero-shot generation is less faithful, so faithfulness is a controllable property of enrichment.

Reading between the lines

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

  • A deployable version cannot rely on relevance judgments or true answers to choose source documents; a testable extension is to select sources by embedding similarity or pseudo-relevance feedback and check whether the rank and RAG gains survive.
  • If the effect transfers, corpus enrichment could act as an offline complement to query expansion and query reformulation, shifting effort from the query side to the document side.
  • For multi-author corpora, enrichment raises a governance question the paper flags: search engines would become content creators rather than only mediators, and the proposed faithfulness measures are a first step toward auditing generated content before it enters the corpus.
  • The same mechanism might generalize beyond text passages to other retrieval units, such as code snippets or product descriptions, wherever query-vocabulary mismatch makes existing relevant items hard to find.
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

3 major / 5 minor

Summary. The paper proposes an offline corpus-enrichment perspective: LLMs rewrite or summarize existing documents, and the newly generated documents are inserted into the corpus to improve query-based retrieval. Three families of generation methods are defined (zero-shot generation, document modification, and two/three-document query-biased summarization, plus a 2DSR variant with one random document) and evaluated in three settings: ad hoc retrieval on DL21/DL22 with BM25, E5, and Contriever; RAG question answering on a Natural Questions subset; and answer attribution via BM25 and BM25+NLI. The paper also introduces NLI-based faithfulness measures for the generated content. The main empirical findings are that generated documents are often ranked above the median existing relevant document, that enriched corpora improve NDCG, RAG accuracy, and attribution entailment in the studied configurations, and that context-based generation methods (DM, 2DS, 3DS) have high measured faithfulness.

Significance. The paper advances a genuinely useful and relatively underexplored direction: using LLMs as corpus-side content creators rather than only as query-side or ranker-side tools. The ad hoc retrieval proof of concept is credible in its basic materials: crowd workers judged the generated documents, the generation procedures are simple and reproducible, the evaluation uses standard test collections, three ranking functions, multiple LLMs, and permutation tests. The faithfulness measures are a reasonable first step toward auditing generated enrichment. The main limitations are that the RAG and attribution claims rest on oracle selection of answer-containing source documents, and that no non-LLM enrichment control isolates LLM-based generation as the active ingredient. If those gaps are addressed, the paper would be a solid proof of concept for a new research direction; as it stands, the evidence supports the weaker claim that adding oracle-selected relevant content to the corpus helps retrieval.

major comments (3)
  1. [§4–§5, Tables 2 and 4] The experimental design does not separate the effect of LLM-based generation from the effect of adding an oracle-selected relevant document to the corpus. In every enriched corpus, exactly one new document per query is inserted, and that document is generated from known relevant (ad hoc) or answer-containing (RAG) source documents; the NoEnrich baselines contain no such inserted document. The large Ans-5 gains (45.9% to 59.8–77%) and the Acc gains in Table 4 are therefore consistent with a much weaker mechanism: simply placing an answer-bearing passage into the corpus. Please add non-LLM enrichment controls under the same protocol for Tables 2, 4, and 5, for example (i) inserting the original source passage, (ii) inserting the source passage with query terms appended, and (iii) an extractive query-biased summary. Without such controls, the paper's title and abstract claim that LLM-based generation is the active ingredient is not established.
  2. [§5.1 and §5.2] The RAG and attribution experiments use an oracle for document selection: a document is defined as relevant if it contains the true answer, and questions with too few such documents are skipped. This makes the evaluation subset answerable by construction and is not a deployable procedure. The conclusion acknowledges the oracle, but the abstract and Section 5.2 present the accuracy gains as evidence for the general merits of enrichment. Please report the number of questions skipped per generation method and LLM, re-analyze the results with skipped questions counted as failures, and report both variants. This would make the proof-of-concept boundary explicit and prevent overgeneralization.
  3. [§4.3, Table 2] The wording 'substantially improved' is stronger than the statistical support shown in Table 2. Most NDCG@10, NDCG@100, and MAP differences are not marked as statistically significant, and many gains are small (e.g., DL21 Okapi NDCG@10 is 0.446 for NoEnrich and 0.470–0.485 for the enriched corpora). Please add effect sizes and significance information for all comparisons, and temper the 'substantially' phrasing in Sections 1 and 4.3 accordingly.
minor comments (5)
  1. [Table 2] Table 2 is very difficult to parse because the MG, ME, and HR columns are interleaved and line breaks split values; consider typesetting it as separate sub-tables per ranker or repeating explicit column headers.
  2. [Figure 2] The prompt template in Figure 2 contains a malformed delimiter, '< query < document >}', which appears to be missing an angle bracket before the document placeholder.
  3. [§4.1 footnote] The footnote stating that NDCG@100 was used instead of NCG@100 because it is 'correlated' should be expanded; if the numbers are compared with published DL21/DL22 results, this substitution needs more justification.
  4. [§1] There is a typo in the Introduction: 'owenrship' should be 'ownership'.
  5. [§4, references] Reference [12] is a duplicate of reference [11] and does not appear to support the sentence in Section 4 about queries having very few relevant documents; please re-check the citation.

Circularity Check

2 steps flagged · score 6.0 of 10

RAG and attribution gains largely reduce to oracle-injected answer-bearing documents; ad hoc retrieval remains independently grounded.

  1. self definitional [Section 5.1 (Experimental setup, Document selection) and Section 7 (Conclusion); Table 4]
    "Instead, we used an oracle method for document selection. Therefore, as a proof of concept, we defined a document as relevant to a question if it contained the true answer. ... Our evaluation of RAG with corpus enrichment was based on an oracle experiment where we used documents (passages) that contain the answer to generate new documents."

    The generated document is derived, under prompts that forbid adding new knowledge ('Do not add new knowledge not present in the document'), from source passages that contain the true answer. Once inserted into the corpus and retrieved in the top five (Gen-5 ranges from about 41% to 80% in Table 4), the answer is present in the RAG context by construction. Ans-5 rises from 45.9% to 59.8-77% and accuracy rises accordingly. Thus the RAG improvement is a direct consequence of oracle-injecting answer-bearing content, not of any demonstrated property of LLM rewriting; a baseline that adds the original answer-containing passages would likely show the same effect.

  2. self definitional [Section 6.1 (Candidate document for attribution) and Section 6.2 (Results); Table 5]
    "In cases where the candidate document is a document generated by one of our methods, we also report a modified version of the entailment accuracy, which considers for attribution only human-authored documents from the original corpus C... Recall that every generated document was generated using one or more existing documents from the original corpus C. ... The highest observed Acc value was 74.5%, obtained when both RAG and attribution were with corpus enrichment, using Llama2 with the DM generation method for BM25+NLI."

    Attribution is measured by NLI entailment of the answer by the candidate document. The generated candidates are built from source documents that contain the true answer, so they entail the answer by construction (modulo LLM fidelity), which inflates Acc in the enriched-corpus settings. The authors' own Acc-NoGen, restricted to original human-authored documents, is much lower (max 48.6% vs 74.5%), confirming that the gain comes from the oracle-selected answer content inserted into the corpus. No control injects original answer-containing passages without LLM rewriting, so the attribution advantage does not isolate the contribution of LLM-based generation.

full rationale

Score is 6 rather than lower or higher because the ad hoc retrieval study (Section 4) is grounded in external TREC relevance judgments and crowd-sourced judgments of generated documents; no parameters are fitted and no self-citation is load-bearing. The circularity is concentrated in the RAG and attribution experiments, where the oracle ('relevant if it contained the true answer') defines the input and the measured success (answer in top-5 context; NLI entailment of the answer) is the same property. Those results therefore reduce, largely by construction, to inserting answer-bearing documents into the corpus. The paper openly discloses the oracle in Section 7 and frames the work as a proof of concept, which mitigates but does not remove the circularity in the central RAG/attribution claims.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new theoretical entities are postulated; the generated documents are artifacts of the method, not entities invoked to explain the result. The free parameters listed are hand-chosen evaluation settings, not fitted model parameters.

free parameters (3)
  • RAG context size = 5
    Section 5.1 retrieves five documents for RAG context; a hand-chosen evaluation setting rather than a fitted parameter. Results may vary with context size.
  • NLI entailment threshold = 0.5
    Used for faithfulness (Section 4.2) and attribution (Section 6.1), taken from Bohnet et al. [8]; not fitted here but influences all faithfulness and attribution numbers.
  • Relevance grade threshold = 2
    Section 4.1 follows DL21 guidelines and treats documents with grade at least two as relevant; changing this threshold changes the relevant set and all downstream effectiveness numbers.
assumptions (4)
  • domain assumption Crowd annotations of generated documents are reliable relevance judgments.
    Section 4.1 uses the median of three crowd workers to label generated documents as relevant; the reported effectiveness gains assume these labels are correct enough.
  • domain assumption NLI entailment scores quantify faithfulness and attribution.
    Sections 4.2 and 6.1 use an NLI model with a 0.5 threshold as ground truth for whether generated content is faithful to the corpus and whether a passage supports an answer; no human validation is reported for the faithfulness numbers.
  • domain assumption Relevant or answer-containing documents can be identified in practice.
    Section 4.1 assumes existing relevance judgments are available; Section 5.1 assumes an oracle can find passages containing the true answer. The paper is explicit about this, but the practical claims depend on it.
  • domain assumption LLMs prompted at temperature zero do not introduce content drift that invalidates enrichment.
    The paper sets temperature to zero to minimize hallucinations but does not measure semantic drift beyond the NLI-based faithfulness check.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Merits of LLM-Based Corpus Enrichment." pith.science (2026). https://pith.science/paper/SYXBLT2B

@misc{pith2026250606015,
  author       = {Pith},
  title        = {Pith review of: On the Merits of LLM-Based Corpus Enrichment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SYXBLT2B}},
  note         = {Machine review of arXiv:2506.06015}
}
read the original abstract

Generative AI (genAI) technologies -- specifically, large language models (LLMs) -- and search have evolving relations. We argue for a novel perspective: using genAI to enrich a document corpus so as to improve query-based retrieval effectiveness. The enrichment is based on modifying existing documents or generating new ones. As an empirical proof of concept, we use LLMs to generate documents relevant to a topic which are more retrievable than existing ones. In addition, we demonstrate the potential merits of using corpus enrichment for retrieval augmented generation (RAG) and answer attribution in question answering.

Figures

Figures reproduced from arXiv: 2506.06015 by the authors.

Figure 1
Figure 1. The prompt for zero-shot document generation (ZS). [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The prompt for document modification (DM). [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. The prompt to LLM with a question (no-RAG). [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: The prompt to LLM with a question and five docu [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 38 canonical work pages

  1. [1]

    Introducing Connect by CloudResearch: Advancing Online Participant Recruitment in the Digital Age | Request PDF

    2024. Introducing Connect by CloudResearch: Advancing Online Participant Recruitment in the Digital Age | Request PDF. doi:10.31234/osf.io/ksgyr

    work page doi:10.31234/osf.io/ksgyr 2024

  2. [2]

    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

  3. [3]

    Arian Askari, Mohammad Aliannejadi, Evangelos Kanoulas, and Suzan Verberne

    work page

  4. [4]

    Leif Azzopardi. 2015. Theory of Retrieval: The Retrievability of Information. In Proceedings of ICTIR. 3–6

    work page 2015

  5. [5]

    Peter Bailey, Alistair Moffat, Falk Scholer, and Paul Thomas. 2016. UQV100: A Test Collection with Query Variability. InProceedings of SIGIR. 725–728

    work page 2016

  6. [6]

    Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. 2018. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. arXiv:1611.09268 [cs.CL] https://arxiv.org/abs/1611.09268

    arXiv 2018

  7. [7]

    Niv Bardas, Tommy Mordo, Oren Kurland, Moshe Tennenholtz, and Gal Zur

    work page

  8. [8]

    Bernd Bohnet, Vinh Q. Tran, Pat Verga, Roee Aharoni, Daniel Andor, Livio Baldini Soares, Massimiliano Ciaramita, Jacob Eisenstein, Kuzman Ganchev, Jonathan Herzig, Kai Hui, Tom Kwiatkowski, Ji Ma, Jianmo Ni, Lierni Sestorain Saralegui, Tal Schuster, William W. Cohen, Michael Collins, Dipanjan Das, Donald Metzler, Slav Petrov, and Kellie Webster. 2023. Att...

    work page 2023

Show all 69 references
  1. [9]

    Chris Buckley, Gerard Salton, James Allan, and Amit Singhal. 1994. Automatic query expansion using SMART: TREC3. In Proceedings of the Third Text Retrieval Conference (TREC-3). 69–80

    1994

  2. [10]

    Claudio Carpineto and Giovanni Romano. 2012. A Survey of Automatic Query Expansion in Information Retrieval. ACM Comput. Surv. 44, 1 (2012), 1:1–1:50

    2012

  3. [11]

    Carlos Castillo. 2018. Fairness and Transparency in Ranking. SIGIR Forum 52, 2 (2018), 64–71. doi:10.1145/3308774.3308783

    2018

  4. [12]

    Carlos Castillo. 2018. Fairness and Transparency in Ranking. SIGIR Forum 52, 2 (2018), 64–71

    2018

  5. [13]

    Charles L. A. Clarke and Laura Dietz. 2024. LLM-based relevance assessment still can’t replace human relevance assessment. CoRR abs/2412.17156 (2024)

    2024

  6. [14]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Jimmy Lin

  7. [15]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, Jimmy Lin, Ellen Voorhees, and Ian Soboroff. 2022. Overview of the TREC 2022 Deep Learning Track. Overview of the TREC 2022 Deep Learning Track . https://trec.nist.gov/ pubs/trec31/papers/Overview_deep.pdf

    2022

  8. [16]

    Voorhees

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. Overview of the TREC 2019 deep learning track. arXiv:2003.07820 [cs.IR] https://arxiv.org/abs/2003.07820

    2020 arXiv

  9. [17]

    Shane Culpepper, Guglielmo Faggioli, Nicola Ferro, and Oren Kurland

    J. Shane Culpepper, Guglielmo Faggioli, Nicola Ferro, and Oren Kurland. 2022. Topic Difficulty: Collection and Query Formulation Effects. ACM Trans. Inf. Syst. 40, 1 (2022), 19:1–19:36

    2022

  10. [18]

    Miles Efron, Peter Organisciak, and Katrina Fenlon. 2012. Improving retrieval of short texts through document expansion. In Proceedings of SIGIR. 911–920

    2012

  11. [19]

    Ekstrand, Anubrata Das, Robin Burke, and Fernando Diaz

    Michael D. Ekstrand, Anubrata Das, Robin Burke, and Fernando Diaz. 2022. Fairness in Information Access Systems. Found. Trends Inf. Retr. 16, 1-2 (2022), 1–177

    2022

  12. [20]

    Guglielmo Faggioli, Laura Dietz, Charles L. A. Clarke, Gianluca Demartini, Matthias Hagen, Claudia Hauff, Noriko Kando, Evangelos Kanoulas, Martin Potthast, Benno Stein, and Henning Wachsmuth. 2023. Perspectives on Large Language Models for Relevance Judgment. In Proceedings o...

    2023

  13. [21]

    Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu

    Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Cha- ganty, Yicheng Fan, Vincent Zhao, N. Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2022. RARR: Researching and Revising What Language Models Say, Using Language Models. ArXiv abs/2210.08726 (2022). http...

    2022 arXiv

  14. [22]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Ji- awei Sun, Meng Wang, and Haofen Wang. 2024. Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv:2312.10997 [cs.CL]

    2024 arXiv

  15. [23]

    Gregory Goren, Oren Kurland, Moshe Tennenholtz, and Fiana Raiber. 2020. Ranking-Incentivized Quality Preserving Content Modification. In Proceedings of SIGIR. 259–268

    2020

  16. [24]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, et al. 2024. The Llama 3 Herd of Models. doi:10.48550/ARXIV.2407.21783 Version Number: 3

    2024 doi arXiv

  17. [25]

    Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kuk- liansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and Yossi Matias. 2022. TRUE: Re-evaluating Factual Consistency Evaluation. In Pro- ceedings of the 2022 Conference of the North Amer...

    2022 doi

  18. [26]

    Oz Huly, Idan Pogrebinsky, David Carmel, Oren Kurland, and Yoelle Maarek. 2024. Old IR Methods Meet RAG. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (Washington DC, USA) (SIGIR ’24). Association for Computi...

    2024

  19. [27]

    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. http://arxiv.org/abs/2112.09118 arXiv:2112.09118 [cs]

    2022 arXiv

  20. [28]

    Gautier Izacard and Edouard Grave. 2021. Leveraging Passage Retrieval with Gen- erative Models for Open Domain Question Answering. arXiv:2007.01282 [cs.CL] https://arxiv.org/abs/2007.01282

    2021 arXiv

  21. [29]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research 24, 251 ...

    2023

  22. [30]

    Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. 2023. Query Expansion by Prompting Large Language Models. arXiv:2305.03653 [cs.IR]

    2023 arXiv

  23. [31]

    Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. arXiv:2004.04906 [cs.CL] https://arxiv.org/abs/ 2004.04906

    2020 arXiv

  24. [32]

    Tibor Kiss and Jan Strunk. 2006. Unsupervised Multilingual Sentence Boundary Detection. Computational Linguistics 32, 4 (2006), 485–525. doi:10.1162/coli.2006. 32.4.485

    2006 doi

  25. [33]

    Oren Kurland and Lillian Lee. 2004. Corpus structure, language models, and ad hoc information retrieval. In Proceedings of SIGIR. 194–201

    2004

  26. [34]

    Oren Kurland and Moshe Tennenholtz. 2022. Competitive Search. In Proceedings of SIGIR. 2838–2849

    2022

  27. [35]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019 doi

  28. [36]

    Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent Retrieval for Weakly Supervised Open Domain Question Answering. arXiv:1906.00300 [cs.CL] https://arxiv.org/abs/1906.00300

    2019 arXiv

  29. [37]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing S...

    2020

  30. [38]

    Xiaoxi Li, Jiajie Jin, Yujia Zhou, Yuyao Zhang, Peitian Zhang, Yutao Zhu, and Zhicheng Dou. 2025. From Matching to Generation: A Survey on Generative Information Retrieval. arXiv:2404.14851 [cs.IR] https://arxiv.org/abs/2404.14851

    2025 arXiv

  31. [39]

    Yinheng Li. 2023. A Practical Survey on Zero-shot Prompt Design for In-context Learning. In Proceedings of the Conference Recent Advances in Natural Language Processing - Large Language Models for Natural Language Processings . INCOMA Ltd., Shoumen, BULGARIA, 641–647. doi:10.2...

    2023 doi

  32. [40]

    Manning, Christopher Ré, Diana Acosta-Navas, Drew A

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michi- hiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christian Cosgrove, Christopher D. Manning, Christopher Ré, Diana Acosta-Nav...

    2023

  33. [41]

    Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-Hong Yang, Ronak Pradeep, and Rodrigo Nogueira. 2021. Pyserini: A Python Toolkit for Reproducible Informa- tion Retrieval Research with Sparse and Dense Representations. In Proceedings of the 44th Annual International ACM SIGIR Co...

    2021

  34. [42]

    Bruce Croft

    Xiaoyong Liu and W. Bruce Croft. 2004. Cluster-Based Retrieval Using Language Models. In Proceedings of SIGIR. 186–193

    2004

  35. [43]

    Edward Loper and Steven Bird. 2002. NLTK: The Natural Language Toolkit. doi:10.48550/arXiv.cs/0205028 arXiv:cs/0205028

    2002 doi

  36. [44]

    G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher. 1978. An analysis of approxi- mations for maximizing submodular set functions–I. Math. Program. 14, 1 (Dec. 1978), 265–294. doi:10.1007/BF01588971

    1978 doi

  37. [45]

    Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document Expansion by Query Prediction. arXiv:1904.08375 [cs.IR] On the Merits of LLM-Based Corpus Enrichment Conference’17, July 2017, Washington, DC, USA

    2019 arXiv

  38. [46]

    perplexity [n. d.]. http://perplexity.ai

  39. [47]

    Zhen Qin, Rolf Jagerman, Kai Hui, Honglei Zhuang, Junru Wu, Le Yan, Jiaming Shen, Tianqi Liu, Jialu Liu, Donald Metzler, Xuanhui Wang, and Michael Ben- dersky. 2024. Large Language Models are Effective Text Rankers with Pairwise Ranking Prompting. In Findings of the Associatio...

    2024

  40. [48]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. In-Context Retrieval-Augmented Lan- guage Models. arXiv:2302.00083 [cs.CL] https://arxiv.org/abs/2302.00083

    2023 arXiv

  41. [49]

    Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter

  42. [50]

    Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Frame- work: BM25 and Beyond. Found. Trends Inf. Retr. 3, 4 (April 2009), 333–389. doi:10.1561/1500000019

    2009 doi

  43. [51]

    Rodrygo L. T. Santos, Craig MacDonald, and Iadh Ounis. 2015. Search Result Diversification. Found. Trends Inf. Retr. 9, 1 (2015), 1–90

    2015

  44. [52]

    Chirag Shah and Emily M. Bender. 2024. Envisioning Information Access Systems: What Makes for Good Tools and a Healthy Web? ACM Trans. Web 18, 3 (2024), 33:1–33:24

    2024

  45. [53]

    Computa- tional Linguistics 49, 4 (Dec

    Measuring Attribution in Natural Language Generation Models. Computa- tional Linguistics 49, 4 (Dec. 2023), 777–840. doi:10.1162/coli_a_00486

    2023 doi

  46. [54]

    Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. In Findings of the Association for Computational Linguistics: EMNLP 2021 . 3784–3803

    2021

  47. [55]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucu- rull, David Esiobu, Jude Fernandes, Jeremy...

    2023 arXiv

  48. [56]

    Shivani Upadhyay, Ronak Pradeep, Nandan Thakur, Daniel Campos, Nick Craswell, Ian Soboroff, Hoa Trang Dang, and Jimmy Lin. 2024. A Large-Scale Study of Relevance Assessments with Large Language Models: An Initial Look. CoRR abs/2411.08275 (2024)

    2024 arXiv

  49. [57]

    Eilon Sheetrit, Anna Shtok, and Oren Kurland. 2020. A passage-based approach to learning to rank documents. Inf. Retr. J. 23, 2 (2020), 159–186

    2020

  50. [58]

    Bruce Croft

    Xing Wei and W. Bruce Croft. 2006. LDA-Based document models for Ad-hoc retrieval. In Proceedings of SIGIR. 178–185

    2006

  51. [59]

    Haolun Wu, Yansen Zhang, Chen Ma, Fuyuan Lyu, Bowei He, Bhaskar Mitra, and Xue Liu. 2024. Result Diversification in Search and Recommendation: A Survey

    2024

  52. [60]

    Xuyang Wu, Shuowei Li, Hsin-Tai Wu, Zhiqiang Tao, and Yi Fang. 2025. Does RAG Introduce Unfairness in LLMs? Evaluating Fairness in Retrieval-Augmented Generation Systems. arXiv:2409.19804 [cs.CL] https://arxiv.org/abs/2409.19804

    2025 arXiv

  53. [61]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2024. Text Embeddings by Weakly-Supervised Contrastive Pre-training. http://arxiv.org/abs/2212.03533 arXiv:2212.03533 [cs]

    2024 arXiv

  54. [62]

    you [n. d.]. http://you.com

  55. [63]

    Shane Culpep- per

    Oleg Zendel, Anna Shtok, Fiana Raiber, Oren Kurland, and J. Shane Culpep- per. 2019. Information Needs, Queries, and Query Performance Prediction. In Proceedings of SIGIR. 395–404

    2019

  56. [64]

    Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, Jie Jiang, and Bin Cui. 2024. Retrieval-Augmented Generation for AI-Generated Content: A Survey. arXiv:2402.19473 [cs.CV] https://arxiv.org/abs/2402.19473

    2024 arXiv

  57. [65]

    Elad Yom-Tov, Shai Fine, David Carmel, and Adam Darlow. 2005. Learning to estimate query difficulty: including applications to missing content detection and distributed information retrieval. In Proceedings of SIGIR. 512–519

    2005

  58. [69]

    Shengyao Zhuang, Honglei Zhuang, Bevan Koopman, and Guido Zuccon. 2024. A Setwise Approach for Effective and Highly Efficient Zero-shot Ranking with Large Language Models. In Proceedings of SIGIR. 38–47

    2024

  59. [2021]

    https://trec.nist.gov/pubs/trec30/papers/Overview-DL.pdf

    Overview of the TREC 2021 Deep Learning Track.Overview of the TREC 2021 Deep Learning Track. https://trec.nist.gov/pubs/trec30/papers/Overview-DL.pdf

    2021

  60. [2023]

    Human Experts

    A Test Collection of Synthetic Documents for Training Rankers: ChatGPT vs. Human Experts. In Proceedings of CIKM. 5311–5315

  61. [2025]

    CoRR abs/2502.07315 (2025)

    White Hat Search Engine Optimization using Large Language Models. CoRR abs/2502.07315 (2025)

    2025 arXiv

Pith tools

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