Pith. sign in

REVIEW 4 major objections 5 minor 32 references

The paper claims that the top-k runner-up tokens an LLM computes while generating query keywords are already conditioned on the query and can serve as free expansion terms.

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

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

T0 review · deepseek-v4-flash

2026-08-05 11:34 UTC pith:BIJIKZ32

load-bearing objection A genuinely cheap QE trick with consistent but modest gains; the unspecified subword index and missing significance testing are the main things to fix. the 4 major comments →

arxiv 2509.02377 v1 pith:BIJIKZ32 submitted 2025-09-02 cs.IR

Upcycling Candidate Tokens of Large Language Models for Query Expansion

Jinseok Kim , Sukmin Cho , Soyeong Jeong , Sangyeop Kim , Sungzoon Cho This is my paper
classification cs.IR
keywords query expansioncandidate tokenslarge language modelsinformation retrievallexical retrievaldense retrievallearned sparse retrievalzero-shot retrieval
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Query expansion by LLMs usually trades quality for cost: diverse expansions require multiple sampling passes or long pseudo-documents. This paper claims the trade-off can be bypassed by harvesting the candidate tokens an LLM evaluates but does not output. At each decoding step the model ranks the whole vocabulary; the runner-up tokens, especially at the first position of each generated keyword, are query-conditioned and lexically diverse. After removing duplicates and tokens shorter than two characters, CTQE fuses these candidates with the original query and the generated keywords, and shows consistent gains over keyword-only expansion across ten benchmarks at the same token budget, roughly matching pseudo-document expansion at a fraction of the cost. The practical stake is that retrieval quality can improve without extra LLM inference.

Core claim

On the paper's own terms, the discovery is that unselected candidate tokens from a single LLM decoding pass carry usable expansion signal. Concretely, when the LLM is prompted to write keywords for a query, the top-k probability tokens at the first position of each keyword—tokens the model considered but did not choose, typically k=20—are collected, deduplicated, and filtered to tokens of length at least two. Treating that set as a query over a separate subword index (for BM25) or as an additional component in a fused query vector (for dense and learned-sparse retrievers) improves nDCG@10 over the keyword-only baseline Q2K on every dataset tested. The ablation shows filtering is essential: u

What carries the argument

The central object is the filtered first-position candidate-token set C = union over keywords of {w in C'_i,1 : |w| >= 2}, where C'_i,1 is the top-k set of token probabilities the LLM assigns at the first decoding position of keyword i. It carries the argument because it converts otherwise wasted probability mass into lexical diversity. The second piece is the fusion rule: for lexical retrieval, S_CTQE(d) = alpha * S_expan(d) + (1 - alpha) * S_C(d), where S_C scores documents with the candidate tokens as a query against a subword index; for dense and learned-sparse retrieval, the query, keyword, and candidate components are combined linearly in embedding or term-weight space. The paper uses

Load-bearing premise

The load-bearing premise is that the top-k unselected tokens at the first decoding position, after deduplication and length filtering, are semantically related to the query rather than irreducible sampling noise or subword fragments; if they carry no lexical utility, the CTQE advantage disappears.

What would settle it

A controlled experiment replacing CTQE's candidate tokens with random vocabulary tokens matched for length and frequency (or with candidates from an unrelated query) would settle whether the signal is query-conditioned: if random tokens give the same nDCG@10 lift, the gain is not from relevance. A second check: on a dataset with known relevant expansion terms, CTQE's candidates should overlap with those terms more than chance; if not, the method is only adding dispersion, not relevance.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Query expansion can obtain the diversity that previously required temperature sampling or multiple prompt variants within a single greedy decoding pass, at identical output-token cost.
  • Short keyword outputs become viable: even a one-token keyword plus the candidate-token harvest beats the unexpanded BM25 baseline on the low-resource average.
  • The method transfers across retriever families—BM25, SPLADE-style learned sparse, and BGE-style dense all improve from the same candidate-token signal.
  • Stronger LLMs yield larger CTQE gains, so the method's value grows as the underlying generators improve.
  • On low-resource domains, CTQE can outperform pseudo-document generation methods while consuming roughly one-eighth of the generated tokens.

Where Pith is reading between the lines

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

  • The same first-position candidate-token harvest could apply to any generation task that asks an LLM to enumerate items—query clarification, tag recommendation, entity linking—wherever runner-up first tokens encode alternative category hypotheses.
  • A testable implication not developed in the paper: the gain should be largest when the lexical gap between queries and relevant documents is wide, since candidate tokens that greedy decoding suppresses are precisely the rare or domain-specific alternatives; correlating CTQE gains with a lexical-overlap measure would check this.
  • The filtering recipe (dedup, first position, length >= 2) is a proof of concept, not a tuned ceiling; combining candidates from multiple positions with different conditioning prefixes, or reconstructing subword fragments into words, could yield further gains at the same marginal cost.
  • The 'no extra inference' claim applies to decoding cost; API-level pricing or service limits on logprobs could change the practical economics of harvesting candidates.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Candidate Token Query Expansion (CTQE), which harvests the top-k alternative tokens computed by an LLM at the first decoding position of each generated keyword. After duplicate removal and a two-character length filter, these candidate tokens are used as an additional expansion signal alongside the generated keywords. The method is evaluated on TREC DL 2019/2020 and eight BEIR datasets, using BM25, SPLADE++, and BGE-base as retrievers and GPT-4.1 and Llama-3 as generators. The paper reports consistent NDCG@10 gains over the keyword-only baseline Q2K at the same LLM token budget, and competitive or better performance than pseudo-document expansion methods at lower latency.

Significance. If the empirical claims hold, the core idea is valuable: it obtains a cheap and diverse expansion signal from probability distributions that the LLM already computes, without additional decoding. The paper's main supports are the consistent (though sometimes small) improvements over Q2K across datasets, a sensible ablation separating all candidates, deduplicated candidates, and first-position candidates, and a public code release. The claims are not definitionally circular; CTQE's contribution is measured empirically. However, the paper leaves a load-bearing implementation detail—the construction and scoring of the subword index—underspecified, does not report significance tests or error bars, and gives no tuning protocol for its many hyperparameters. The evidence is therefore not yet commensurate with the strength of the claims.

major comments (4)
  1. [Section 2.2, 'CTQE on Lexical Retrievers'] The 'separate subword index' used to compute S_C(d) is not specified. Please state exactly how the document corpus is tokenized and indexed, what retrieval function is used, and how S_C(d) is scored and normalized before interpolation with S_expan(d). Footnote 2 only says S_expan(d) is divided by the query repetition factor; there is no equivalent normalization described for S_C(d). This matters because candidate tokens are subword units: if the index matches substrings or BPE fragments, S_C(d) may reward documents for containing common fragments such as 'ing' or 'tion' rather than semantically related terms. Since several reported gains over Q2K are small (e.g., FiQA 28.0 vs. 27.2; Arguana 41.1 vs. 40.6), this ambiguity is load-bearing. Please provide the full indexing/scoring recipe, a control experiment replacing candidate tokens with frequency-matched random subword tokens, and quali
  2. [Section 2.2, footnote 1; Table 1] CTQE's lexical score uses the original query repeated five times before concatenation with the generated keywords. The paper does not state whether the Q2K baseline and other lexical baselines use the same query repetition. If query repetition is applied only to CTQE, the comparison is not apples-to-apples, and part of the reported gain may come from query reweighting rather than from candidate tokens. Please clarify the exact query formulation used by each baseline and, if the baselines do not already include the five-fold query repetition, add a Q2K+repeat baseline to isolate the effect of candidate tokens.
  3. [Section 4, overall results; Table 1] No significance tests or error bars are reported. The abstract and Section 4 use 'significantly outperforms' and 'consistently outperforms,' but the comparisons are presented as single runs. Several differences are small, and on a per-dataset basis some gaps may be within noise. Please report paired significance tests (e.g., bootstrap or permutation tests over queries) across the ten benchmarks, or temper the language to 'numerically outperforms' where significance is not established.
  4. [Section 3, Implementation Details; Section 2.2] The method has many free hyperparameters—alpha, alpha_q, alpha_W, alpha_C, beta_q, beta_W, beta_C, k, max token length, and the query repetition factor—but the paper states only the chosen values and does not describe a tuning protocol. If these values were selected after observing the test sets, the 'consistently outperforms' claim is weaker. Please state whether the hyperparameters were fixed a priori or chosen on a validation split, and include sensitivity analyses for at least alpha and k. This is particularly important because the interpolation weight directly controls how much of the final score comes from the candidate-token component.
minor comments (5)
  1. [Section 2.2] The notation is inconsistent: C'_{i,j} is defined as the top-k candidates at any position, while the filtered set is denoted C. Please unify the notation and make explicit that only C'_{i,1} is used after filtering.
  2. [Figure 2 (center)] The x-axis labels 'All', 'Dedup', 'Dedup + 1st Pos' are not self-explanatory. Please clarify that the reported method corresponds to the final bar, and describe the exact filtering chain in the caption.
  3. [Abstract; Section 2.2] The paper calls candidate tokens 'terms' and 'keywords,' but they are subword tokens from the LLM tokenizer. Please clarify this and avoid implying they are whole words, especially since the length filter only discards tokens shorter than two characters.
  4. [Table 2] The 'Unique Keywords' column mixes generated keywords and filtered candidate tokens with different counting rules across methods. Please define the counting rule explicitly so the diversity comparison is interpretable.
  5. [Section 3, Implementation Details] The dense retriever weights satisfy alpha_q + alpha_W + alpha_C = 0.7, not 1. If the combined vector is not normalized, the absolute scale does not affect ranking, but this should be stated for clarity.

Circularity Check

0 steps flagged

No significant circularity: CTQE's gains are empirical, not derived from its inputs.

full rationale

The paper's central claim—that first-position candidate tokens from a single LLM decoding pass provide useful query-expansion signals—is an empirical hypothesis tested on TREC DL and BEIR benchmarks, not a quantity derived from its own definitions. CTQE's candidate set C is defined as deduplicated, length-filtered top-k tokens at the first decoding position (Sec. 2.2), and its retrieval score S_CTQE is an interpolation (α=0.9) of the keyword score and a candidate-token score. Nothing in these equations presupposes that CTQE will outperform Q2K; the superiority is established by the ablation and benchmark comparisons (Table 1, Fig. 2). Hyperparameters (α, β, k) are stated as fixed constants, not fitted to the test labels in a way that would make the reported gains tautological. There are no self-citations or imported uniqueness/ansatz results; the only external method adopted (Mackie et al. [16] for neural fusion) is prior work by other authors and is not the source of the CTQE gain. The under-specified 'separate subword index' (Sec. 2.2) is a reproducibility/correctness concern—it leaves open whether S_C(d) is dominated by substring matches—but it does not make any prediction equivalent to an input. The ablations used to choose filtering are on the same benchmarks, which weakens the independence of the final configuration, but this is a standard model-selection concern rather than circularity under the defined patterns. Hence no significant circularity.

Axiom & Free-Parameter Ledger

10 free parameters · 3 axioms · 0 invented entities

The method introduces no new entities, but it rests on several hand-set numeric weights and three domain assumptions: the usefulness of first-position candidate tokens, the faithfulness of API logits, and the effectiveness of a lexical match on subword tokens. The weights are fixed rather than tuned with a documented protocol, which is a mild reproducibility and soundness concern.

free parameters (10)
  • alpha = 0.9
    Interpolation weight between keyword-augmented and candidate-token scores for lexical retrieval. Set without a reported tuning protocol; directly controls the final CTQE score in Section 2.2.
  • alpha_q = 0.5
    Weight for the original query embedding in dense fusion (Section 2.2).
  • alpha_W = 0.1
    Weight for generated-keyword embeddings in dense fusion (Section 2.2).
  • alpha_C = 0.1
    Weight for candidate-token embeddings in dense fusion (Section 2.2). Note alpha_q+alpha_W+alpha_C=0.7, not 1.
  • beta_q = 0.5
    Weight for query term importance in learned sparse fusion (Section 2.2).
  • beta_W = 0.1
    Weight for keyword term importance in learned sparse fusion (Section 2.2).
  • beta_C = 0.1
    Weight for candidate-token term importance; set to 0 for top-20 dominant query/keyword tokens (footnote 3), a post hoc adjustment.
  • k = 20
    Number of candidate tokens kept per decoding step; chosen as the maximum provided by the OpenAI API (Section 3). Affects the candidate set size.
  • query_repetition_factor = 5
    Original query replicated 5 times in the expanded query for lexical retrievers (footnote 1), cited from prior work but a hand-set constant.
  • max_token_length = 16 (high-resource) / 32 (low-resource)
    Maximum generation length for keywords, set per dataset group (Section 3). Affects the number of generated keywords and hence candidate tokens.
axioms (3)
  • domain assumption Top-k unselected tokens at the first decoding position of each keyword, after dedup and length filtering, are semantically relevant to the user query.
    The entire CTQE benefit rests on this empirical premise; it is only validated through retrieval metrics, not directly. Introduced in Section 2.2 Token filtering.
  • domain assumption The LLM's reported top-k candidate logits (via OpenAI API) are an accurate view of the model's next-token distribution.
    Relies on the API exposing true decoding candidates; if the API truncates or distorts these, the candidate set is unrepresentative. Implicit in Section 3 implementation.
  • domain assumption Candidate tokens can be matched to documents through a separate subword index using a standard lexical retriever.
    The subword index and scoring function for S_C(d) are not specified; the paper assumes lexical overlap between single candidate tokens and document subwords is a useful signal. Section 2.2 lexical retrievers.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 10627 in / 11385 out tokens · 112886 ms · 2026-08-05T11:34:46.049366+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Upcycling Candidate Tokens of Large Language Models for Query Expansion." pith.science (2026). https://pith.science/paper/BIJIKZ32

@misc{pith2026250902377,
  author       = {Pith},
  title        = {Pith review of: Upcycling Candidate Tokens of Large Language Models for Query Expansion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIJIKZ32}},
  note         = {Machine review of arXiv:2509.02377}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Query Expansion (QE) improves retrieval performance by enriching queries with related terms. Recently, Large Language Models (LLMs) have been used for QE, but existing methods face a trade-off: generating diverse terms boosts performance but increases computational cost. To address this challenge, we propose Candidate Token Query Expansion (CTQE), which extracts diverse and relevant terms from a single LLM decoding pass by leveraging unselected candidate tokens. These tokens, though not part of the final output, are conditioned on the full query and capture useful information. By aggregating them, CTQE achieves both relevance and diversity without extra inference, reducing overhead and latency. Experiments show that CTQE delivers strong retrieval performance with significantly lower cost, outperforming or comparable to more expensive methods. Code is available at: https://github.com/bluejeans8/CTQE

Figures

Figures reproduced from arXiv: 2509.02377 by Jinseok Kim, Sangyeop Kim, Soyeong Jeong, Sukmin Cho, Sungzoon Cho.

Figure 1
Figure 1. Figure 1: An illustration of our CTQE method. short and may lack sufficient contextual information, a key chal￾lenge in IR is bridging the lexical and semantic gap between the user’s query and the relevant documents. To address this, Query Expansion (QE) has emerged as a promising solution by adding additional terms to the original query [1, 22]. Recently, the powerful generation capabilities of Large Language Model… view at source ↗
Figure 2
Figure 2. Figure 2: (Left). Retrieval performance vs. latency for CTQE (w/o PRF) and Q2D, averaged across 8 low-resource datasets. Each point corresponds to a different maximum generation length. CTQE is evaluated at token limits of [1, 2, 4, 8, 16, 32], while Q2D uses [1, 2, 4, 8, 16, 32, 64, 128]. (Center) Average NDCG@10 scores for incremental candidate token filtering on eight low-resource datasets. (Right). Retrieval per… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

32 extracted references · 8 canonical work pages

  1. [1]

    Gianni Amati and Cornelis Joost Van Rijsbergen. 2002. Probabilistic models of information retrieval based on measuring the divergence from randomness.ACM Trans. Inf. Syst. 20, 4 (Oct. 2002), 357–389. doi:10.1145/582415.582416

    arXiv 2002

  2. [2]

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

    2020

  3. [3]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the TREC 2020 deep learning track. arXiv:2102.07662 [cs.IR] https://arxiv.org/ abs/2102.07662

    Pith/arXiv arXiv 2021

  4. [4]

    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

    Pith/arXiv arXiv 2020

  5. [5]

    Dhole and Eugene Agichtein

    Kaustubh D. Dhole and Eugene Agichtein. 2024. GenQREnsemble: Zero-Shot LLM Ensemble Prompting for Generative Query Reformulation. In Advances in Information Retrieval: 46th European Conference on Information Retrieval, ECIR 2024, Glasgow, UK, March 24–28, 2024, Proceedings, Part III (Glasgow, United Kingdom). Springer-Verlag, Berlin, Heidelberg, 326–335. ...

    doi:10.1007/978-3-031- 2024

  6. [6]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ah- mad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sra- vankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurélien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozière, Bet...

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783 2024

  7. [7]

    Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stéphane Clinchant

  8. [8]

    Thibault Formal, Benjamin Piwowarski, and Stéphane Clinchant. 2021. SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (Virtual Event, Canada) (SIGIR ’21) . Association for Computing Machinery, New York, NY, USA, 2288–2292. d...

    doi:10.1145/3404835 2021

  9. [9]

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise Zero-Shot Dense Retrieval without Relevance Labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 17...

    doi:10.18653/v1/2023 2023

  10. [10]

    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). https://openreview.net/forum?id=jKN1pXi7b0

    2022

  11. [11]

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

    Pith/arXiv arXiv 2023

  12. [12]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu (Eds.). Association for C...

    doi:10.18653/v1/2020.emnlp-main.550 2020

  13. [13]

    Bruce Croft

    Victor Lavrenko and W. Bruce Croft. 2001. Relevance based language models. In Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (New Orleans, Louisiana, USA) (SIGIR ’01). Association for Computing Machinery, New York, NY, USA, 120–127. doi:10. 1145/383952.383972

    arXiv 2001

  14. [14]

    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 International ACM SIGIR Conference on Research and Development in Informa- tion Retrieval (Washington DC, USA) (SIGIR ’24). Association f...

    arXiv 2024

  15. [15]

    Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-Hong Yang, Ronak Pradeep, and Rodrigo Nogueira. 2021. Pyserini: A Python Toolkit for Reproducible Infor- mation Retrieval Research with Sparse and Dense Representations. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (Virtual Event, Canada) (S...

    work page arXiv 2021

  16. [16]

    Iain Mackie, Shubham Chatterjee, and Jeffrey Dalton. 2023. Generative and Pseudo-Relevant Feedback for Sparse, Dense and Learned Sparse Retrieval. arXiv:2305.07477 [cs.IR] https://arxiv.org/abs/2305.07477

    Pith/arXiv arXiv 2023

  17. [17]

    Iain Mackie, Shubham Chatterjee, and Jeffrey Dalton. 2023. Generative Relevance Feedback with Large Language Models. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (Taipei, Taiwan) (SIGIR ’23). Association for Computing Machinery, New York, NY, USA, 2026–2031. doi:10.1145/3539618.3591992

    arXiv 2023

  18. [18]

    Ha-Thanh Nguyen, Manh-Kien Phi, Xuan-Bach Ngo, Vu Tran, Le-Minh Nguyen, and Minh-Phuong Tu. 2022. Attentive deep neural networks for legal document retrieval. Artif. Intell. Law 32, 1 (Dec. 2022), 57–86. doi:10.1007/s10506-022-09341- 8

    work page doi:10.1007/s10506-022-09341- 2022

  19. [19]

    Priyanka Nigam, Yiwei Song, Vijai Mohan, Vihan Lakshman, Weitian (Allen) Ding, Ankit Shingavi, Choon Hui Teo, Hao Gu, and Bing Yin. 2019. Semantic Product Search. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (Anchorage, AK, USA)(KDD ’19). Association for Computing Machinery, New York, NY, USA, 2876–28...

    doi:10.1145/3292500 2019

  20. [20]

    OpenAI. 2024. GPT-4.1. https://openai.com/index/gpt-4-1/

    work page 2024

  21. [21]

    Kirk Roberts, Tasmeer Alam, Steven Bedrick, Dina Demner-Fushman, Kyle Lo, Ian Soboroff, Ellen Voorhees, Lucy Lu Wang, and William R. Hersh. 2021. Search- ing for scientific evidence in a pandemic: An overview of TREC-COVID. J. of Biomedical Informatics 121, C (Sept. 2021), 16 pages. doi:10.1016/j.jbi.2021.103865

    arXiv 2021

  22. [22]

    S. E. Robertson. 1991. On term selection for query expansion. J. Doc. 46, 4 (Jan. 1991), 359–364. doi:10.1108/eb026866

    work page doi:10.1108/eb026866 1991

  23. [23]

    Robertson, Steve Walker, Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford

    Stephen E. Robertson, Steve Walker, Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford. 1994. Okapi at TREC-3. In Text Retrieval Conference. https: //api.semanticscholar.org/CorpusID:41563977

    work page 1994

  24. [24]

    Tao Shen, Guodong Long, Xiubo Geng, Chongyang Tao, Yibin Lei, Tianyi Zhou, Michael Blumenstein, and Daxin Jiang. 2024. Retrieval-Augmented Retrieval: Large Language Models are Strong Zero-Shot Retriever. In Findings of the As- sociation for Computational Linguistics: ACL 2024 , Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computat...

    doi:10.18653/v1/2024.findings-acl.943 2024

  25. [25]

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) . https://openreview. net/forum?id=wCu6T5xFjeJ

    work page 2021

  26. [26]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guil- laume Lample. 2023. LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971 [cs.CL] https://arxiv.org/abs/2302.13971

    Pith/arXiv arXiv 2023

  27. [27]

    Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query Expansion with Large Language Models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 9414–9423. doi:10.18653/v1/2023.emnlp-main.585 CIKM ’25, Novembe...

    doi:10.18653/v1/2023.emnlp-main.585 2023

  28. [28]

    Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini. 2024. When do Generative Query and Docu- ment Expansions Fail? A Comprehensive Study Across Methods, Retrievers, and Datasets. In Findings of the Association for Computational Linguistics: EACL 2024 , Yvette Graham and Matthew Purver (Eds.). Association f...

    work page 2024

  29. [29]

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-Pack: Packed Resources For General Chinese Embed- dings. In Proceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Retrieval (Washington DC, USA) (SI- GIR ’24). Association for Computing Machinery, New York, NY, ...

    arXiv 2024

  30. [30]

    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. In International Conference on Learning Representations. https://openreview.net/forum?id=zeFrfgyZln

    work page 2021

  31. [31]

    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 , Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 1872–1883. doi:10.18653/v1/2...

    doi:10.18653/v1/2024.findings-emnlp.103 2024

  32. [2022]

    In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (Madrid, Spain) (SIGIR ’22)

    From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (Madrid, Spain) (SIGIR ’22). Association for Computing Machinery, New York, NY, USA, 2353–2359. doi:10.1145/3477495.3531857

    arXiv