Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Prepending a model-generated reasoning text to a query before computing its embedding improves zero-shot dense retrieval on reasoning-intensive queries, without any training.

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 ↗

Reasoning-infused text embedding, which prepends LLM-generated reasoning to queries before embedding, improves zero-shot dense retrieval on BRIGHT.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection RITE is a plausible query-expansion trick that shows consistent gains on BRIGHT, but the test-set tuning of the generation length undermines the zero-shot claim until a held-out protocol is shown. the 3 major comments →

arxiv 2509.00276 v1 pith:ZN5BIM3U submitted 2025-08-29 cs.CL

Exploring Reasoning-Infused Text Embedding with Large Language Models for Zero-Shot Dense Retrieval

Yuxiang Liu , Tian Wang , Gourab Kundu , Tianyu Cao , Guang Cheng , Zhen Ge , Jianshu Chen , Qingjun Cui
show 1 more author
Trishul Chilimbi
This is my paper
classification cs.CL
keywords dense retrievaltext embeddingzero-shot retrievalreasoning infusionlarge language modelsBRIGHT benchmarkEcho embeddingsPromptReps
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.

The reading

This paper tries to establish that the reasoning ability of generative LLMs can be folded directly into the embedding step for zero-shot dense retrieval. The proposed approach, RITE, first prompts an LLM to produce a short reasoning text about a query, then prepends that text to the query before extracting the embedding with an existing LLM-based embedding scheme such as Echo or PromptReps. The authors claim that this reasoning-infused query embedding is more similar to genuinely relevant documents than the original query embedding alone, and they report consistent gains in nDCG@10 across all twelve domains of the BRIGHT benchmark. A sympathetic reader would care because it suggests a simple, training-free way to close the gap that encoder-only retrievers show on complex, inferential queries.

Core claim

RITE claims that injecting an explicit reasoning step into the query embedding pipeline improves retrieval for queries that require inference beyond lexical overlap. Concretely, given a query, the LLM first generates a reasoning text in response to one of three step-by-step prompts; that text is prepended to the original query; the resulting sequence is then fed into either the Echo repetition scheme or the PromptReps single-word compression scheme to produce the final query embedding. The paper reports that on BRIGHT, a benchmark of twelve reasoning-intensive retrieval datasets, RITE-Echo and RITE-PR outperform their non-reasoning counterparts Echo and PR, with the largest relative gains in

What carries the argument

The central mechanism is the prepending of an LLM-generated reasoning text to the query before embedding extraction. The reasoning text is elicited by one of three 'think step by step' prompts that ask the model to reformulate the query for a search engine or to specify what relevant documents should contain. Embedding extraction then reuses two existing zero-shot schemes: Echo, which repeats the input text and averages the hidden states of the second occurrence, and PromptReps, which extracts the hidden state before the model's first output token when asked to summarize the input in one word. The reasoning step is applied only to queries, not to documents, so document embeddings can be comp

Load-bearing premise

A short, automatically generated reasoning text placed before the query makes the query embedding land closer to genuinely relevant documents than the original query embedding does, and this advantage holds across all twelve BRIGHT domains with only a per-model choice of how long the reasoning text may be.

What would settle it

Run a control where the generated reasoning text is replaced by an irrelevant but syntactically similar sequence of the same length, generated with the same decoding settings; if retrieval gains persist, the effect is not due to reasoning content. Alternatively, evaluate RITE on a non-reasoning retrieval benchmark such as MS MARCO or BEIR; if gains disappear or reverse while BRIGHT gains remain, the claim is specific to reasoning-intensive domains rather than a general embedding improvement.

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

If this is right

  • Zero-shot retrieval on complex, inferential queries can be improved without any fine-tuning, just by adding a reasoning prompt before the embedding step.
  • RITE composes with existing LLM-based embedding schemes, so the reasoning-infusion idea may transfer to other methods that accept prompt text before embedding extraction.
  • In fact-based and structured domains, model-generated reasoning performs close to human-crafted oracle reasoning, suggesting automatic rationales can substitute for manual ones in those areas.
  • Because only query embeddings are augmented, the extra computation sits on the query side; document corpora can be embedded once and reused.
  • The paper points toward fine-tuning the reasoning step, improving prompt design, and integrating RITE with hybrid retrieval paradigms as natural follow-ups if the central claim holds.

Where Pith is reading between the lines

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

  • The observed gains may partly reflect query expansion or additional context rather than logical inference per se; a length-matched control sequence of irrelevant text would separate these effects, but the authors do not run that control.
  • Since the paper augments only queries, applying the same reasoning step to document embeddings could further close the gap between RITE and oracle reasoning, or it might dilute the effect if both sides drift similarly.
  • The per-model selection of maximum reasoning-token length based on test-set performance introduces a small selection risk; the claim would be stronger if the chosen length were fixed on a development set or held-out folds.
  • RITE's reasoning prompts are generic; domain-specific prompts or externally sourced rationales could plausibly push performance closer to oracle levels in mathematical and theoretical domains.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes Reasoning-Infused Text Embedding (RITE), a zero-shot dense retrieval method that prepends an LLM-generated reasoning text to the query before applying an existing embedding extraction strategy (Echo or PromptReps). The method is evaluated on the BRIGHT benchmark with Mistral 7B and LLaMA 3 8B, reporting nDCG@10 improvements over the non-reasoning baselines across 12 domains. The paper also compares model-generated reasoning with oracle reasoning texts and includes a case study. The central claim is that explicit reasoning in the token space improves zero-shot retrieval for complex, inferential queries.

Significance. RITE is simple and directly addresses a real limitation of encoder-only embeddings on reasoning-intensive queries. The paper builds on established zero-shot embedding methods (Echo, PromptReps), uses open-source models, and evaluates across a broad set of domains. If the reported gains are reproducible under a properly controlled evaluation protocol, this would be a useful and easily adoptable contribution. However, the current evidence is weakened by test-set hyperparameter selection, the absence of a HyDE-style query-expansion baseline, and the lack of any statistical significance testing. The method itself is easy to implement and the claims are falsifiable, but the experimental support needs strengthening before the zero-shot claim is fully supported.

major comments (3)
  1. [§3.2, Table 1] The maximum generation tokens are selected from {256, 128, 64} by choosing the value that yields the best retrieval performance for each model, and BRIGHT is the only evaluation benchmark. This is test-set hyperparameter selection; the reported nDCG@10 improvements in Table 1 are therefore not a valid estimate of zero-shot behavior. Please either fix a default without looking at the test set, use a held-out validation split, or report nDCG@10 for all three lengths and show that the conclusion is unchanged. Also report the selected values for each model.
  2. [§4.1, Table 1] The contribution claim is 'reasoning-infused embeddings,' but no HyDE-style query expansion baseline is included. HyDE is the standard zero-shot LLM query-augmentation method and would isolate the effect of prepending generated reasoning text from the effect of any LLM rewrite. In addition, the abstract uses 'significantly' without any significance test; all numbers in Table 1 are single point estimates with no standard deviations, confidence intervals, or paired tests across queries. Please add HyDE and either bootstrap CIs or per-domain query-level significance tests.
  3. [§2.2.1, §3.2] Three reasoning prompts are listed in §2.2.1, but the paper does not report which prompt produced Tables 1–3, nor an ablation. If the prompt was chosen by test performance on BRIGHT, this is another instance of test-set tuning; if it was fixed in advance, that should be stated. Please report results for each prompt or at least disclose and justify the chosen prompt, and make the generation-length selection independent of the test set.
minor comments (5)
  1. [Table 3 caption] Table 3's caption says 'model-generated (RITE-Echo) vs. oracle (GT-Echo)' but the rows are RITE-PR and GT-PR; correct the caption.
  2. [§4.2] The text says oracle reasoning 'consistently achieves superior retrieval performance,' but Tables 2 and 3 contain multiple counterexamples (e.g., Mistral 7B PR Biology: RITE-PR 15.1 vs GT-PR 10.0; LLaMA 3 8B PR TheoremQA: 21.3 vs 19.9). Please qualify the claim.
  3. [Tables 2–3] 'Leet' and 'Pony' columns contain '-' for all oracle rows. Please explain whether oracle reasoning texts are unavailable for these datasets and, if so, how that affects the oracle comparison.
  4. [Table 4] The case study reports nDCG@10 of 0.2372 and 0.8503 for single queries. It is unclear whether these are per-query nDCG values (which for binary relevance would be 0 or 1) or averages over a subset. Clarify.
  5. [General] No code or reproducibility artifact is provided. Given the method's simplicity, releasing the prompts and generation-length selection scripts would strengthen the paper.

Circularity Check

1 steps flagged

RITE's zero-shot retrieval gains are partly predetermined by choosing the reasoning-text generation length on the BRIGHT test set.

specific steps
  1. fitted input called prediction [Section 3.2, Implementation Details (hyperparameter selection); Tables 1-3 (reported nDCG@10 on BRIGHT)]
    "The maximum generation tokens (for reasoning text) are selected from 256, 128, or 64, choosing the value that yields the best retrieval performance for each model."

    The only hyperparameter that controls the reasoning text is selected by maximizing retrieval performance on BRIGHT, the same benchmark whose nDCG@10 is then reported as RITE's zero-shot result. Thus each reported RITE column is the best of three test-set evaluations, not an independent prediction. Echo and PR have no analogous free length to tune, so the claimed improvement is at least partly an artifact of test-set selection rather than a consequence of reasoning infusion.

full rationale

RITE is an empirical recipe: generate a reasoning text, prepend it to the query, then apply Echo or PR. There is no formal derivation chain to inspect, and the paper does not rely on self-citations or imported uniqueness theorems. The only load-bearing circular step is the test-set selection of the reasoning-generation length in Section 3.2: because BRIGHT is both the selection criterion and the evaluation benchmark, the numbers in Tables 1-3 are maxima over 256/128/64 rather than zero-shot predictions. This does not make the entire method tautological, since the core idea (prepending generated reasoning) has independent content and is compared against oracle reasoning, but it invalidates the clean zero-shot claim and partially reduces the reported gains to a fitted choice.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The paper introduces no new entities. The only fitted quantities are the per-model maximum generation length and, implicitly, the choice of reasoning prompt. The central claim rests on the validity of BRIGHT as an evaluation set and on the usefulness of generated reasoning for embedding.

free parameters (2)
  • max_generation_tokens = unknown (chosen per model from 256, 128, or 64)
    Selected per model based on best retrieval performance on BRIGHT (Section 3.2); values not reported.
  • reasoning_prompt = unknown (one of three prompts)
    Three reasoning prompts are introduced in Section 2.2.1; the paper does not report which prompt was used for each model or whether prompt selection was made on the test set.
axioms (3)
  • domain assumption BRIGHT is a valid and representative benchmark for reasoning-intensive retrieval.
    The entire evaluation is on BRIGHT (Section 3.1); no other benchmarks are used.
  • domain assumption nDCG@10 with the provided qrels is an appropriate metric for comparing retrieval methods.
    Used as the sole evaluation metric (Section 3.1).
  • domain assumption An instruction-tuned LLM generates reasoning text that improves embedding alignment in a single generation pass.
    The method depends on the quality of the generated reasoning (Section 2.2).

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Reasoning-Infused Text Embedding with Large Language Models for Zero-Shot Dense Retrieval." pith.science (2026). https://pith.science/paper/ZN5BIM3U

@misc{pith2026250900276,
  author       = {Pith},
  title        = {Pith review of: Exploring Reasoning-Infused Text Embedding with Large Language Models for Zero-Shot Dense Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZN5BIM3U}},
  note         = {Machine review of arXiv:2509.00276}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Transformer-based models such as BERT and E5 have significantly advanced text embedding by capturing rich contextual representations. However, many complex real-world queries require sophisticated reasoning to retrieve relevant documents beyond surface-level lexical matching, where encoder-only retrievers often fall short. Decoder-only large language models (LLMs), known for their strong reasoning capabilities, offer a promising alternative. Despite this potential, existing LLM-based embedding methods primarily focus on contextual representation and do not fully exploit the reasoning strength of LLMs. To bridge this gap, we propose Reasoning-Infused Text Embedding (RITE), a simple but effective approach that integrates logical reasoning into the text embedding process using generative LLMs. RITE builds upon existing language model embedding techniques by generating intermediate reasoning texts in the token space before computing embeddings, thereby enriching representations with inferential depth. Experimental results on BRIGHT, a reasoning-intensive retrieval benchmark, demonstrate that RITE significantly enhances zero-shot retrieval performance across diverse domains, underscoring the effectiveness of incorporating reasoning into the embedding process.

Figures

Figures reproduced from arXiv: 2509.00276 by Gourab Kundu, Guang Cheng, Jianshu Chen, Qingjun Cui, Tian Wang, Tianyu Cao, Trishul Chilimbi, Yuxiang Liu, Zhen Ge.

Figure 1
Figure 1. Figure 1: Model overview. extraction, enriching query representations with inferential depth. RITE improves query embeddings by introducing a reasoning stage before embedding extraction. It consists of three key steps: (1) Rea￾soning Elicitation: Generate an inferentially enriched reasoning text from the query. (2) Embedding Extraction: Apply Echo or PR on the reasoning-infused query to obtain a final embedding. (3)… 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

36 extracted references · 14 canonical work pages · 1 internal anchor

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Flo- rencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Associ- ation for Computational Linguistics: Human Language Technologies, V olume 1 (Long and Short Papers), Jill Burstein, Chri...

    doi:10.18653/v1 2019

  4. [4]

    Folasade Olubusola Isinkaye, Yetunde O Folajimi, and Bolande Adefowoke Ojokoh. 2015. Recommendation systems: Principles, methods and evaluation. Egyptian informatics journal 16, 3 (2015), 261–273

    work page 2015

  5. [5]

    Cheng Ji and Huaiying Luo. 2025. Leveraging Large Language Model for Intelli- gent Log Processing and Autonomous Debugging in Cloud AI Platforms. In 2025 8th International Conference on Advanced Electronic Materials, Computers and Software Engineering (AEMCSE). IEEE, 348–351

    work page 2025

  6. [6]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al . 2023. Mistral 7B. arXiv preprint arXiv:2310.06825 (2023)

    Pith/arXiv arXiv 2023

  7. [7]

    Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang

  8. [8]

    Yihong Jin, Ze Yang, and Xinhe Xu. 2025. Scam Detection for Ethereum Smart Contracts: Leveraging Graph Representation Learning for Secure Blockchain. In 2025 4th International Symposium on Computer Applications and Information Technology (ISCAIT). 1730–1734. https://doi.org/10.1109/ISCAIT64916.2025. 11010625

    work page arXiv 2025

  9. [9]

    Mei Kobayashi and Koichi Takeda. 2000. Information retrieval on the web. ACM computing surveys (CSUR) 32, 2 (2000), 144–173

    work page 2000

  10. [10]

    Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, et al . 2024. Gecko: Ver- satile text embeddings distilled from large language models. arXiv preprint arXiv:2403.20327 (2024)

    Pith/arXiv arXiv 2024

  11. [11]

    Yuxiang Liu and Kevin Chen-Chuan Chang. 2025. Writing Like the Best: Exemplar-Based Expository Text Generation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), ACL 2025, Vienna, Austria, July 27 - August 1, 2025 . Association for Computa- tional Linguistics, 25739–25764. https://aclantho...

    work page 2025

  12. [12]

    Yuxiang Liu, Jie Huang, and Kevin Chen-Chuan Chang. 2023. Ask To The Point: Open-Domain Entity-Centric Question Generation. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023 . Association for Computational Linguistics, 2703–2716. https://doi.org/10.18653/ v1/2023.findings-emnlp.178

    work page 2023

  13. [13]

    Huaiying Luo and Cheng Ji. 2025. Cross-Cloud Data Privacy Protection: Optimiz- ing Collaborative Mechanisms of AI Systems by Integrating Federated Learning and LLMs. arXiv preprint arXiv:2505.13292 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Huaiying Luo and Cheng Ji. 2025. Federated Learning-Based Data Collaboration Method for Enhancing Edge Cloud AI System Security Using Large Language Models. arXiv preprint arXiv:2506.18087 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query Rewriting in Retrieval-Augmented 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, 5303–5315. https://doi.org/10.18...

    doi:10.18653/v1/2023.emnlp- 2023

  16. [16]

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2023. Fine- tuning llama for multi-stage text retrieval. arXiv preprint arXiv:2310.08319 (2023)

    Pith/arXiv arXiv 2023

  17. [17]

    Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2024. Generative representational instruction tuning. arXiv preprint arXiv:2402.09906 (2024)

    Pith/arXiv arXiv 2024

  18. [18]

    Rajvardhan Patil, Sorio Boit, Venkat Gudivada, and Jagadeesh Nandigam. 2023. A survey of text representation and embedding techniques in nlp. IEEE Access 11 (2023), 36120–36146

    work page 2023

  19. [19]

    Jacob Mitchell Springer, Suhas Kotha, Daniel Fried, Graham Neubig, and Aditi Raghunathan. 2024. Repetition improves language model embeddings. arXiv preprint arXiv:2402.15449 (2024)

    Pith/arXiv arXiv 2024

  20. [20]

    Hongjin Su, Howard Yen, Mengzhou Xia, Weijia Shi, Niklas Muennighoff, Han-yu Wang, Haisu Liu, Quan Shi, Zachary S Siegel, Michael Tang, et al. 2024. BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval. arXiv preprint arXiv:2407.12883 (2024)

    Pith/arXiv arXiv 2024

  21. [21]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal

  22. [22]

    Haoyu Wang, Tuo Zhao, and Jing Gao. 2024. BlendFilter: Advancing Retrieval- Augmented Large Language Models via Query Generation Blending and Knowl- edge Filtering. arXiv preprint arXiv:2402.11129 (2024)

    Pith/arXiv arXiv 2024

  23. [23]

    In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.)

    Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge- Intensive Multi-Step Questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 10014–10037. https...

    doi:10.18653/v1/2023 2023

  24. [24]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Improving Text Embeddings with Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, B...

    doi:10.18653/v1/2024.acl-long.642 2024

  25. [25]

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

    Pith/arXiv arXiv 2022

  26. [26]

    Han Xu, Xingyuan Wang, and Haipeng Chen. 2024. Towards Real-Time and Personalized Code Generation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 5568–5569

    work page 2024

  27. [27]

    Chenghao Xiao, G Thomas Hudson, and Noura Al Moubayed. 2024. RAR-b: Reasoning as Retrieval Benchmark. arXiv preprint arXiv:2404.06347 (2024)

    Pith/arXiv arXiv 2024

  28. [28]

    Han Xu, Ruining Zhao, Jindong Wang, and Haipeng Chen. 2024. RESTful- Llama: Connecting User Queries to RESTful APIs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 1433–1443

    work page 2024

  29. [29]

    Han Xu, Jingyang Ye, Yutong Li, and Haipeng Chen. 2024. Can Speculative Sampling Accelerate ReAct Without Compromising Reasoning Quality?. In The Second Tiny Papers Track at ICLR 2024 . https://openreview.net/forum?id= 42b9hJrIpX

    work page 2024

  30. [30]

    Ze Yang, Yihong Jin, and Xinhe Xu. 2024. Hades: Hardware accelerated decoding for efficient speculation in large language models. arXiv preprint arXiv:2412.19925 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    Ze Yang, Yihong Jin, Juntian Liu, Xinhe Xu, Yihan Zhang, and Shuyang Ji. 2025. Research on Cloud Platform Network Traffic Monitoring and Anomaly Detec- tion System based on Large Language Models. In 2025 IEEE 7th International Conference on Communications, Information System and Computer Engineering (CISCE). 1029–1032. https://doi.org/10.1109/CISCE65916.2...

    work page arXiv 2025

  32. [32]

    Bowen Zhang, Kehua Chang, and Chunping Li. 2024. Simple techniques for enhancing sentence embeddings in generative language models. In International Conference on Intelligent Computing. Springer, 52–64

    work page 2024

  33. [33]

    Wenhao Yu, Hongming Zhang, Xiaoman Pan, Kaixin Ma, Hongwei Wang, and Dong Yu. 2023. Chain-of-note: Enhancing robustness in retrieval-augmented language models. arXiv preprint arXiv:2311.09210 (2023)

    Pith/arXiv arXiv 2023

  34. [34]

    Shengyao Zhuang, Xueguang Ma, Bevan Koopman, Jimmy Lin, and Guido Zuc- con. 2024. PromptReps: Prompting Large Language Models to Generate Dense and Sparse Representations for Zero-Shot Document Retrieval. arXiv preprint arXiv:2404.18424 (2024)

    Pith/arXiv arXiv 2024

  35. [35]

    Jiahao Zhang, Haiyang Zhang, Dongmei Zhang, Liu Yong, and Shen Huang. 2024. End-to-End Beam Retrieval for Multi-Hop Question Answering. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (V olume 1: Long Papers), Kevin Duh, Helena Gomez, and Steven Bethard (Eds....

    work page 2024

  36. [2023]

    arXiv preprint arXiv:2307.16645 (2023)

    Scaling sentence embeddings with large language models. arXiv preprint arXiv:2307.16645 (2023)

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.