Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A New Query Expansion Approach via Agent-Mediated Dialogic Inquiry

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a three-agent LLM dialogue, AMD, which splits a query into three Socratic sub-questions, generates pseudo-answers, and filters them with a reflective feedback step, outperforms recent prompt-based query expansion…

desk verdict A plausible but under-specified query-expansion workshop paper whose three-agent Socratic loop is a new combination, yet whose headline gains are confounded by unstated baseline backbones and aggregation choices. read the letter →

arxiv 2502.08557 v3 pith:IALY6UOV submitted 2025-02-12 cs.IR cs.CLcs.LGcs.MA

Wonduk Seo , Hyunjin An , Seunghyun Lee This is my paper · ORCID
classification cs.IRcs.CLcs.LGcs.MA
keywords InformationRetrievalQueryExpansionMulti-AgentLargeLanguageModelsSocraticQuestioningFusionPseudo-Answers
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

Query expansion adds extra context to a user's search query so the retriever can find relevant documents it might otherwise miss. The paper claims that a three-agent LLM dialogue—a Socratic Questioner that splits the query into three sub-questions (clarification, assumption probing, implication probing), an Answerer that writes a pseudo-answer for each, and a Reflective Feedback Agent that rewrites and filters those answers—produces expansions that retrieve more relevant documents than recent prompt-based methods (Q2D, Q2C, GenQREnsemble, GenQRFusion). The evidence is eight standard benchmarks (six BEIR datasets and two TREC Deep Learning tracks) measured with three retrieval backends: BM25, a dense encoder, and reciprocal rank fusion. If correct, it shows that structuring LLM generation as a multi-faceted dialogue with a filtering step beats generating more text with more prompts. The paper's own ablation indicates the feedback step helps dense and TREC results but slightly hurts sparse BM25 and RRF on BEIR, which qualifies the strength of the claim.

What carries the argument

The central object is the three-role agent dialogue. The Socratic Questioning Agent prompts a non-finetuned 7-billion-parameter instruction-tuned LLM to emit exactly three sub-questions, each tied to a standard Socratic questioning dimension: clarification, which pins down what the query's key terms mean; assumption probing, which surfaces implicit premises; and implication probing, which asks about consequences. The Dialogic Answering Agent generates one short pseudo-answer per sub-question, with all three answers produced in a single parallel inference, and these pseudo-answers serve as surrogate documents. The Reflective Feedback Agent then receives the original query plus all three question-answer pairs and rewrites or filters each answer so that only the most relevant and informative content is kept. Finally, the refined answers are attached to the original query through one of three aggregation strategies: replicated concatenation for sparse BM25, a weighted embedding sum ($0.7$ initial query plus $0.3$ mean of the answer embeddings) for dense retrieval, and Reciprocal Rank Fusion over three per-answer retrieval lists. The diversity of the expansion is the intended output of the Socratic decomposition, and the filtering is the intended guard against noise.

What would settle it

Replace the three Socratic sub-questions with three generic questions (for example, 'What is this about?', 'Explain this in more detail', 'Give examples') and rerun the same eight-dataset experiment with the same aggregation methods. If retrieval scores do not drop, the Socratic labeling is not the mechanism producing the gains. A second check: compare the full AMD with the no-feedback variant; the paper already reports one regression (BM25/RRF on BEIR), so an evaluation across more datasets where the feedback agent never improves any setup would falsify the claim that reflective filtering is broadly beneficial.

Watch

Extended reading notes

Core claim

AMD outperforms all four baselines on average across the reported metrics. On sparse retrieval, AMD's average BEIR nDCG@10 is 0.4352 versus 0.4193 for Q2D, and its TREC average score is 0.5433 versus 0.5222. On dense retrieval, AMD averages 0.4707 on BEIR versus 0.4560 for Q2D, and 0.6488 across TREC DL'19/'20 versus 0.6385. For fusion retrieval, AMD beats GenQRFusion on every average (BEIR 0.4113 vs 0.3897; TREC 0.5077 vs 0.4515). The design is that the Socratic sub-questions force the LLM to approach the query from three different facets, the answering agent turns those into pseudo-documents, and a non-finetuned feedback agent rewrites each pseudo-answer to remove vague, redundant, and irrelevant content. The three refined answers are then combined with the original query: concatenated three times for sparse search, weighted into the query embedding for dense search, and fused through reciprocal rank fusion for the RRF setting.

Load-bearing premise

The framework's advantage rests on the assumption that exactly three fixed Socratic question types, followed by reflective rewriting, produce expansions that are more diverse and more useful than the baselines' expansions; the paper does not vary the question types or their count, and its own ablation shows the reflective step slightly hurts BM25 and RRF performance on BEIR, so the filtering is not uniformly beneficial.

Editorial extensions

If this is right

  • LLM-based query expansion can be improved by structuring generation as a dialogue over query facets rather than by increasing the number of independent prompts.
  • The same three Socratic dimensions transfer across six heterogeneous BEIR datasets and two TREC collections, indicating the decomposition is not tuned to a single domain.
  • A single non-finetuned 7-billion-parameter LLM suffices for all three agent roles, so the framework requires no task-specific fine-tuning.
  • The framework plugs into sparse, dense, and fusion retrieval with only small changes to aggregation, and it uses three LLM inferences per query, fewer than the ten used by the ensemble baselines.

Reading between the lines

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

  • The paper leaves implicit that the gains might come from the sheer amount of generated text or from the reflective rewrite rather than from the Socratic framing; a natural follow-up is to swap the three Socratic prompts for three neutral 'explain more' prompts and compare.
  • Because the feedback ablation is negative for BM25/RRF on BEIR, the Reflective Feedback Agent may help mainly dense representations, a possibility the paper's conclusion does not fully accommodate.
  • An obvious extension is to make the number and content of sub-questions adaptive per query instead of fixed at three, which could improve coverage for broader or narrower information needs.
  • The Socratic decomposition plus reflective filtering is a generic recipe for generating diverse, filtered context; it could plausibly transfer to conversational search, open-domain question answering, or document expansion.
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

4 major / 5 minor

Summary. The manuscript introduces AMD, a query expansion framework in which three LLM agents interact: a Socratic Questioning Agent rewrites the original query into three sub-questions (clarification, assumption probing, implication probing); a Dialogic Answering Agent generates one pseudo-answer per sub-question; and a Reflective Feedback Agent rewrites/filters these answers. The resulting three refined answers are combined with the initial query using three retrieval strategies: sparse concatenation (Eq. 5), dense weighted aggregation (Eq. 6), and reciprocal rank fusion (Eq. 7). The paper reports experiments on six BEIR datasets and TREC DL 2019/2020, comparing against BM25, Q2C, Q2D, GenQREnsemble, GenQRFusion, and dense E5 baselines, and reports that AMD outperforms these methods across most metrics (Table 2), with an ablation of the feedback module (Table 3).

Significance. Query expansion with LLMs is an active area, and the idea of using Socratic questioning and a reflective filtering step is a reasonable extension of prior prompt-based methods. If the empirical comparisons were properly controlled, the paper would provide useful evidence that structured multi-question inquiry and feedback can improve retrieval over single-shot pseudo-document/pseudo-term expansion. The paper's strengths include evaluation on eight datasets, coverage of sparse, dense, and fusion retrieval settings, the use of an open-source generator (Qwen2.5-7B-Instruct), and an explicit ablation of the feedback module. However, the empirical claim is not yet established: the baseline generator/decoding is undisclosed, the significance tests are defined against an average baseline rather than each competitor, no variance or repeated runs are reported, and the feedback ablation is negative on one of the three retrieval settings. These are fixable within the scope of the manuscript.

major comments (4)
  1. [Section 4.1 (LLM and Retrieval Models)] The paper discloses the generator only for AMD ('For our AMD framework, we utilize ... Qwen2.5-7B-Instruct') and then states 'All baselines follow their original configurations.' This leaves open whether the baseline expansions (Q2C, Q2D, GenQREnsemble, GenQRFusion) were produced by a different, possibly older or weaker, LLM and with different decoding settings (temperature, max length). If so, the gains in Table 2 may reflect the generator or the aggregation scheme rather than the proposed dialogic inquiry. In addition, the dense aggregation in Eq. 6 (weighted mean of embeddings) differs from Q2D's original query-pseudo-document concatenation, so 'original configurations' does not isolate the dialogic mechanism. Please specify the exact backbone and decoding parameters for every baseline, and re-run or otherwise control for the generator.
  2. [Table 2, footnote] The asterisk denotes significance 'over the average baseline value for the metric' (paired t-test with Holm-Bonferroni correction). Comparing against the average of several baselines is not equivalent to comparing against the strongest competitor, and no standard deviations, confidence intervals, or number of repeated LLM runs are reported. With a stochastic generator and a single point estimate per dataset, the claim that AMD 'notably outperforms' recent methods is not statistically supported. Please report mean and variance over multiple runs and test AMD against each baseline individually, or justify the average-baseline comparison.
  3. [Section 4.4, Table 3] The ablation results are mixed and partially contradict the text. For BM25/RRF fusion on BEIR, the version without feedback achieves 0.4147 while the full model achieves 0.4113, so feedback degrades performance on that setting, yet the paragraph states that adding the Reflective Feedback Agent 'improves the overall average score' and is 'crucial.' Please address this negative result explicitly, and if the claim is about averaged scores across all settings, report the disaggregated results and discuss when feedback helps or hurts.
  4. [Section 4.2, Eqs. (5) and (7)] The construction of the individual expanded queries used in RRF is underspecified. Eq. (5) defines a single sparse query as the sum of three copies of Q_init concatenated with all refined pseudo-answers, while Eq. (7) refers to 'each refined pseudo-answer a'_i is used to form an individual expanded query Q*_i' without defining Q*_i. Is Q*_i simply Q_init concatenated with a'_i alone? Also, the set S in Eq. (5) is not defined before use. This ambiguity prevents exact replication of the RRF variant; please define the notation and provide the exact concatenation/separator scheme for each Q*_i.
minor comments (5)
  1. [Section 3.3] The sentence '... shaped by the prior Socratic and dialogic interactions, is retained' is grammatically broken; please rephrase.
  2. [Figure 1] The labels 'X3' and the flow from the question/answer sets to the three retrieval boxes are difficult to parse; please add numbered steps or a clearer legend.
  3. [Title page and references] The ACM template placeholders ('© 2018', 'Received 20 February 2007', 'Accepted 5 June 2009') should be replaced with the correct publication metadata.
  4. [Implementation Details] The exact prompts used for the three agents are not provided; please include them in an appendix to support reproducibility.
  5. [Section 4.2, Eq. (6)] The 0.7/0.3 weight is imported from reference [36] without sensitivity analysis; since this is a free parameter, a small sweep would strengthen the claim that the result is not due to this choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the AMD retrieval comparison is empirical and externally benchmarked; the one self-cited aggregation weight is a non-load-bearing design choice.

full rationale

The central claim is that AMD outperforms recent query expansion methods on BEIR and TREC (Table 2). This is an empirical result measured against external benchmarks; the framework's equations (Eqs. 1-7) generate pseudo-answers and combine them with the initial query, and none of these quantities is defined in terms of the retrieval metric or fitted to the benchmark labels. The only self-referential element is the 0.7/0.3 dense-aggregation weight in Eq. 6, which is imported from the authors' prior GenCRF paper [36] without independent justification. That is a hyperparameter choice, not a derivation, and it does not by construction force the claimed gains; the same caveat applies to the omitted baseline LLM backbone in Section 4.1, which is a confound and fairness concern rather than a circularity. Section 6 openly notes residual noise and future fine-tuning of the feedback agent. No load-bearing premise reduces to its own input, so the paper's derivation chain is self-contained for the purposes of this pass.

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

The method rests on empirical assumptions about LLM behavior rather than on mathematical axioms. The key assumptions are that Socratic question types produce diverse sub-questions, that pseudo-answers are useful synthetic documents, and that a non-finetuned feedback LLM filters noise. The paper supplies only one ablation (feedback on/off) and it is mixed, so these assumptions are only partially evidenced.

free parameters (3)
  • dense aggregation weight for the initial query embedding = 0.7
    Equation 6 sets the final dense query embedding as 0.7 * emb(initial query) + 0.3 * average(emb(pseudo-answers)). This split is adopted from the authors' prior GenCRF paper [36] and is not ablated on the BEIR/TREC benchmarks in this paper; it directly shapes all dense-retrieval results.
  • number of Socratic sub-questions / pseudo-answers = 3
    The framework fixes |S| = 3 sub-questions and answers. No ablation varies this count, so the paper does not show that 3 is the best or even a robust choice.
  • LLM sampling temperature = 0.5
    All generations use temperature 0.5 (Section 4.1). This choice affects the stochasticity of questions and pseudo-answers and is not varied or analyzed; no multiple-run statistics are reported.
assumptions (3)
  • domain assumption The three Socratic dimensions (clarification, assumption probing, implication probing) yield diverse and intent-aligned sub-questions for query expansion.
    Motivated by references on Socratic questioning [1,13,24,31,32,45], but the paper provides no ablation isolating each dimension or showing these three dimensions are sufficient.
  • domain assumption Pseudo-answers generated by Qwen2.5-7B-Instruct serve as informative surrogate documents for retrieval.
    This is the standard pseudo-relevance assumption adapted to LLM generation; the paper does not evaluate answer quality or informativeness directly.
  • domain assumption A non-finetuned Reflective Feedback Agent reliably filters redundant or irrelevant pseudo-answers.
    The ablation in Table 3 shows the feedback agent improves sparse and dense averages but decreases the BM25/RRF BEIR average (0.4113 vs 0.4147), so this assumption is not consistently supported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A New Query Expansion Approach via Agent-Mediated Dialogic Inquiry." pith.science (2026). https://pith.science/paper/IALY6UOV

@misc{pith2026250208557,
  author       = {Pith},
  title        = {Pith review of: A New Query Expansion Approach via Agent-Mediated Dialogic Inquiry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IALY6UOV}},
  note         = {Machine review of arXiv:2502.08557}
}
read the original abstract

Query expansion is widely used in Information Retrieval (IR) to improve search outcomes by supplementing initial queries with richer information. While recent Large Language Model (LLM) based methods generate pseudo-relevant content and expanded terms via multiple prompts, they often yield homogeneous, narrow expansions that lack the diverse context needed to retrieve relevant information. In this paper, we propose AMD: a new Agent-Mediated Dialogic Framework that engages in a dialogic inquiry involving three specialized roles: (1) a Socratic Questioning Agent reformulates the initial query into three sub-questions, with each question inspired by a specific Socratic questioning dimension, including clarification, assumption probing, and implication probing, (2) a Dialogic Answering Agent generates pseudo-answers, enriching the query representation with multiple perspectives aligned to the user's intent, and (3) a Reflective Feedback Agent evaluates and refines these pseudo-answers, ensuring that only the most relevant and informative content is retained. By leveraging a multi-agent process, AMD effectively crafts richer query representations through inquiry and feedback refinement. Extensive experiments on benchmarks including BEIR and TREC demonstrate that our framework outperforms previous methods, offering a robust solution for retrieval tasks.

Figures

Figures reproduced from arXiv: 2502.08557 by the authors.

Figure 1
Figure 1. The overview of the proposed approach: (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. MARIC: Multi-Agent Reasoning for Image Classification

    cs.CV 2025-09 reject novelty 4.0 of 10

    MARIC decomposes image classification into outliner, aspect, and reasoning prompts for a vision-language model, reporting accuracy gains that are inconsistent across datasets and statistically unsupported.

Reference graph

Works this paper leans on

50 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Beng Heng Ang, Sujatha Das Gollapalli, and See Kiong Ng. 2023. Socratic question generation: A novel dataset, models, and evaluation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 147–165

    work page 2023

  2. [2]

    Hiteshwar Kumar Azad and Akshay Deepak. 2019. A novel model for query expansion using pseudo-relevant web knowledge.arXiv preprint arXiv:1908.10193 (2019)

    work page Pith review arXiv 2019

  3. [3]

    Hiteshwar Kumar Azad and Akshay Deepak. 2019. Query expansion techniques for information retrieval: a survey. Information Processing & Management 56, 5 (2019), 1698–1735

    2019

  4. [4]

    Alexander Bondarenko, Maik Fröbe, Meriem Beloucif, Lukas Gienapp, Yamen Ajjour, Alexander Panchenko, Chris Biemann, Benno Stein, Henning Wachsmuth, Martin Potthast, et al . 2020. Overview of Touché 2020: argument retrieval. In Experimental IR Meets Multilinguality, Multimodality, and Interaction: 11th International Conference of the CLEF Association, CLEF...

    work page 2020

  5. [5]

    Claudio Carpineto and Giovanni Romano. 2012. A survey of automatic query expansion in information retrieval. Acm Computing Surveys (CSUR) 44, 1 (2012), 1–50

    2012

  6. [6]

    Xinran Chen, Xuanang Chen, Ben He, Tengfei Wen, and Le Sun. 2024. Analyze, generate and refine: Query expansion with LLMs for zero-shot open-domain QA. In Findings of the Association for Computational Linguistics ACL 2024 . 11908– 11922

    work page 2024

  7. [7]

    Vincent Claveau. 2020. Query expansion with artificially generated texts. arXiv preprint arXiv:2012.08787 (2020)

    work page Pith review arXiv 2020

  8. [8]

    Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel S Weld

Show all 50 references
  1. [9]

    Gordon V Cormack, Charles LA Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval. 758–759

    2009

  2. [10]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M Voorhees. 2020. Overview of the TREC 2019 deep learning track. arXiv preprint arXiv:2003.07820 (2020)

    2020 arXiv

  3. [11]

    Kaustubh D Dhole and Eugene Agichtein. 2024. Genqrensemble: Zero-shot llm ensemble prompting for generative query reformulation. In European Conference on Information Retrieval. Springer, 326–335

    2024

  4. [12]

    Kaustubh D Dhole, Ramraj Chandradevan, and Eugene Agichtein. 2024. Genera- tive Query Reformulation Using Ensemble Prompting, Document Fusion, and Relevance Feedback. arXiv preprint arXiv:2405.17658 (2024)

    2024 arXiv

  5. [13]

    Anmol Goel, Nico Daheim, Christian Montag, and Iryna Gurevych. 2024. Socratic reasoning improves positive text rewriting. arXiv preprint arXiv:2403.03029 (2024)

    2024 arXiv

  6. [14]

    Faegheh Hasibi, Fedor Nikolaev, Chenyan Xiong, Krisztian Balog, Svein Erik Bratsberg, Alexander Kotov, and Jamie Callan. 2017. DBpedia-entity v2: a test collection for entity search. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in I...

    2017

  7. [15]

    Ayyoob Imani, Amir Vakili, Ali Montazer, and Azadeh Shakery. 2019. Deep neural networks for query expansion using word embeddings. In Advances in Information Retrieval: 41st European Conference on IR Research, ECIR 2019, Cologne, Germany, April 14–18, 2019, Proceedings, Part I...

    2019

  8. [16]

    Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bender- sky. 2023. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653 (2023)

    2023 arXiv

  9. [17]

    Pengyue Jia, Yiding Liu, Xiangyu Zhao, Xiaopeng Li, Changying Hao, Shuaiqiang Wang, and Dawei Yin. 2023. MILL: Mutual Verification with Large Language Models for Zero-Shot Query Expansion. arXiv preprint arXiv:2310.19056 (2023)

    2023 arXiv

  10. [18]

    Rosie Jones, Benjamin Rey, Omid Madani, and Wiley Greiner. 2006. Generating query substitutions. In Proceedings of the 15th international conference on World Wide Web. 387–396

    2006

  11. [19]

    Ivica Kostric and Krisztian Balog. 2024. A Surprisingly Simple yet Effective Multi- Query Rewriting Method for Conversational Passage Retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2271–2275

    2024

  12. [20]

    Victor Lavrenko and W Bruce Croft. 2017. Relevance-based language models. In ACM SIGIR Forum, Vol. 51. ACM New York, NY, USA, 260–267

    2017

  13. [21]

    Yibin Lei, Yu Cao, Tianyi Zhou, Tao Shen, and Andrew Yates. 2024. Corpus- Steered Query Expansion with Large Language Models. arXiv preprint arXiv:2402.18031 (2024)

    2024 arXiv

  14. [22]

    Hang Li, Ahmed Mourad, Shengyao Zhuang, Bevan Koopman, and Guido Zuccon

  15. [23]

    Minghan Li, Honglei Zhuang, Kai Hui, Zhen Qin, Jimmy Lin, Rolf Jagerman, Xuanhui Wang, and Michael Bendersky. 2024. Can query expansion improve generalization of strong cross-encoder rankers?. In Proceedings of the 47th In- ternational ACM SIGIR Conference on Research and Deve...

    2024

  16. [24]

    Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A Smith, and Yejin Choi. 2021. DExperts: Decoding-time controlled text generation with experts and anti-experts. arXiv preprint arXiv:2105.03023 (2021)

    2021 arXiv

  17. [25]

    Jie Liu and Barzan Mozafari. 2024. Query Rewriting via Large Language Models. arXiv preprint arXiv:2403.09060 (2024)

    2024

  18. [26]

    Xing Han Lù. 2024. Bm25s: Orders of magnitude faster lexical search via eager sparse scoring. arXiv preprint arXiv:2407.03618 (2024)

    2024 arXiv

  19. [27]

    Iain Mackie, Shubham Chatterjee, and Jeffrey Dalton. 2023. Generative and pseudo-relevant feedback for sparse, dense and learned sparse retrieval. arXiv preprint arXiv:2305.07477 (2023)

    2023 arXiv

  20. [28]

    Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. 2018. Www’18 open challenge: financial opinion mining and question answering. In Companion proceedings of the the web conference 2018. 1941–1942

    2018

  21. [29]

    Shahrzad Naseri, Jeffrey Dalton, Andrew Yates, and James Allan. 2021. Ceqe: Contextualized embeddings for query expansion. In Advances in Information Retrieval: 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28–April 1, 2021, Proceedings, Part I 43 . ...

    2021

  22. [30]

    Hai-Long Nguyen, Tan-Minh Nguyen, Duc-Minh Nguyen, Thi-Hai-Yen Vuong, Ha-Thanh Nguyen, and Xuan-Hieu Phan. 2024. Exploiting LLMs’ Reasoning Capability to Infer Implicit Concepts in Legal Information Retrieval. arXiv preprint arXiv:2410.12154 (2024)

    2024 arXiv

  23. [31]

    Richard Paul and Linda Elder. 2019. The thinker’s guide to Socratic questioning . Rowman & Littlefield

    2019

  24. [32]

    Jingyuan Qi, Zhiyang Xu, Ying Shen, Minqian Liu, Di Jin, Qifan Wang, and Lifu Huang. 2023. The art of SOCRATIC QUESTIONING: Recursive thinking with large language models. arXiv preprint arXiv:2305.14999 (2023)

    2023 arXiv

  25. [33]

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

    2009

  26. [34]

    Stephen E Robertson. 1990. On term selection for query expansion. Journal of documentation 46, 4 (1990), 359–364

    1990

  27. [35]

    Dwaipayan Roy, Debjyoti Paul, Mandar Mitra, and Utpal Garain. 2016. Using word embeddings for automatic query expansion. arXiv preprint arXiv:1606.07608 (2016)

    2016 arXiv

  28. [36]

    Wonduk Seo, Haojie Zhang, Yueyang Zhang, Changhao Zhang, Songyao Duan, Lixin Su, Daiting Shi, Jiashu Zhao, and Dawei Yin. 2024. GenCRF: Generative Clus- tering and Reformulation Framework for Enhanced Intent-Driven Information Retrieval. arXiv preprint arXiv:2409.10909 (2024)....

    2024 arXiv

  29. [37]

    Qwen Team. 2024. Qwen2.5: A Party of Foundation Models. https://qwenlm. github.io/blog/qwen2.5/

    2024

  30. [38]

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models. arXiv preprint arXiv:2104.08663 (2021)

    2021 arXiv

  31. [39]

    Pauli Virtanen, Ralf Gommers, Travis E Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al . 2020. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature methods 17, 3 (2020), 261–272

    2020

  32. [40]

    Ellen Voorhees, Tasmeer Alam, Steven Bedrick, Dina Demner-Fushman, William R Hersh, Kyle Lo, Kirk Roberts, Ian Soboroff, and Lucy Lu Wang. 2021. TREC-COVID: constructing a pandemic information retrieval test collection. In ACM SIGIR Forum, Vol. 54. ACM New York, NY, USA, 1–12

    2021

  33. [41]

    David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or fiction: Verifying scientific claims. arXiv preprint arXiv:2004.14974 (2020)

    2020 arXiv

  34. [42]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual E5 Text Embeddings: A Technical Report. arXiv preprint arXiv:2402.05672 (2024)

    2024 arXiv

  35. [43]

    Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query expansion with large language models. arXiv preprint arXiv:2303.07678 (2023)

    2023 arXiv

  36. [44]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  37. [45]

    Yongkang Wu, Meng Han, Yutao Zhu, Lei Li, Xinyu Zhang, Ruofei Lai, Xiaoguang Li, Yuanhang Ren, Zhicheng Dou, and Zhao Cao. 2023. Hence, socrates is mor- tal: A benchmark for natural language syllogistic reasoning. In Findings of the Association for Computational Linguistics: A...

    2023

  38. [46]

    Fanghua Ye, Meng Fang, Shenghui Li, and Emine Yilmaz. 2023. Enhancing conversational search: Large language model-aided informative query rewriting. arXiv preprint arXiv:2310.09716 (2023)

    2023 arXiv

  39. [47]

    Le Zhang, Yihong Wu, Qian Yang, and Jian-Yun Nie. 2024. Exploring the Best Practices of Query Expansion with Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2024 . 1872–1883

    2024

  40. [48]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023). Received 20 February 2007; revised 12 March 2009; accepted ...

    2023 arXiv

  41. [2020]

    arXiv preprint arXiv:2004.07180 (2020)

    Specter: Document-level representation learning using citation-informed transformers. arXiv preprint arXiv:2004.07180 (2020)

    2020 arXiv

  42. [2023]

    ACM Transactions on Information Systems 41, 3 (2023), 1–40

    Pseudo relevance feedback with deep language models and dense retrievers: Successes and pitfalls. ACM Transactions on Information Systems 41, 3 (2023), 1–40

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.