Pith. sign in

REVIEW 3 major objections 6 minor 3 cited by

Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fine-tuning an LLM on retrieval rank produces query expansions that match or beat generate-then-filter baselines while cutting latency by roughly 70%.

desk verdict Aligned query expansion via DPO/RSFT is a neat idea, but the paper's filtering baseline is a strawman; the central claim needs a real EAR comparison. read the letter →

arxiv 2507.11042 v1 pith:IOYU7MQH submitted 2025-07-15 cs.IR

Adam Yang , Gustavo Penha , Enrico Palumbo , Hugues Bouchard This is my paper
classification cs.IR
keywords queryexpansionLLMalignmentDirectPreferenceOptimizationRejectionSamplingFine-TuningpassageretrievalBM25open-domainquestionansweringvocabularymismatch
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes Aligned Query Expansion (AQE), a method for fine-tuning a language model so that the query expansions it generates are directly the ones that help a sparse retriever find the relevant passage, removing the separate filtering step. Candidate expansions are first generated zero-shot, then ranked by the BM25 rank of the gold passage they retrieve. Rejection sampling fine-tuning keeps the best expansion per query, and direct preference optimization trains the model to prefer the best expansion over the worst. The paper reports that the combined RSFT plus DPO procedure outperforms generate-then-filter baselines in both in-domain and out-of-domain settings, while reducing inference memory and time by around 70%.

What carries the argument

The central object is the retrieval-rank preference pair: for each query, expansions are ranked by the BM25 rank of the single gold passage, the best expansion becomes an RSFT target, and the best-versus-worst pair becomes a DPO contrastive example. The DPO loss reparameterizes the reward into the policy's log-probabilities relative to a reference policy, so retrieval effectiveness itself is the alignment signal. This lets a single greedy decode at inference replace the fifty-sample generate-and-rerank pipeline.

What would settle it

Re-run AQE on a dataset where each query has multiple judged relevant passages, training one model with the single-gold-passage rank reward and another with the average rank of all relevant passages; if the single-passage variant does not match or beat the multi-relevant variant, the load-bearing assumption about the rank reward is false.

Watch

Extended reading notes

Core claim

Aligned Query Expansion (AQE) replaces the costly generate-then-filter paradigm with a single aligned model: for each training query, 50 zero-shot expansions are produced, each is scored by the rank of the gold passage under BM25, and the best and worst expansions become supervised and preference training signals. Rejection Sampling Fine-Tuning (RSFT) maximizes the likelihood of the best expansion, while Direct Preference Optimization (DPO) contrasts the best against the worst using a reference-model-regularized objective. After alignment, inference is one greedy decode, so no reranker or second generation pass is needed. The paper's empirical claim is that this alignment yields better top-N retrieval accuracy than filtering on Natural Questions and TriviaQA in-domain, transfers better out-of-distribution to WebQA and Entity Questions, and cuts computational time by 69.5% and memory by 71.1% while improving top-1 accuracy.

Load-bearing premise

The reward signal is the BM25 rank of a single gold passage, so the whole alignment depends on that gold passage being the only relevant document and on its rank being a reliable, low-noise measure of expansion quality.

Editorial extensions

If this is right

  • The filtering step becomes unnecessary: at inference, AQE performs one greedy decoding of the aligned model rather than generating many candidates and reranking them.
  • Latency and memory costs drop by roughly 70% compared with the generate-then-filter baseline, which matters for real-time retrieval systems.
  • The aligned expansions transfer across domains better than filtered ones, which the paper attributes to filtering being tied to its specific generator while alignment is more model-agnostic.
  • Combining RSFT and DPO produces expansions that are more diverse than either alone and often contain direct answers, which improves lexical match against the gold passage.
  • The method's effectiveness is demonstrated on four datasets: Natural Questions, TriviaQA, WebQA, and Entity Questions, covering both in-distribution and out-of-distribution evaluation.

Reading between the lines

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

  • The rank of one gold passage is a noisy reward when multiple passages are relevant or when the BM25 ranking of the gold passage is near random; using the average rank over all judged relevant passages as the reward would be a direct test of how much this noise matters.
  • The reported 70% speedup is relative to a 50-sample filtering pipeline; the practical gain over a zero-shot single-pass baseline is smaller, and the method's advantage may shift with stronger base language models.
  • The same alignment recipe could be applied to document expansion rather than query expansion, using the rank of the target query under the expanded document as the reward signal.
  • A testable extension would compare DPO trained on rank-based preferences against DPO trained on judgment-based or multi-relevant preferences to see whether the method's gains persist under richer relevance labels.
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 / 6 minor

Summary. The paper proposes Aligned Query Expansion (AQE), a method that fine-tunes a T0 3B language model to generate query expansions that are directly optimized for retrieval effectiveness. AQE first samples n=50 zero-shot expansions per training query, scores each expansion by the BM25 rank of the single gold passage, and then applies Rejection Sampling Fine-Tuning (RSFT) and/or Direct Preference Optimization (DPO) using the best and worst expansions. At inference, a single greedy expansion is generated and used with BM25. The paper reports in-domain and out-of-domain Top-N accuracy improvements over original-query, zero-shot expansion, and a filtering baseline on Natural Questions, TriviaQA, WebQA, and Entity Questions, together with roughly 70% memory and time savings. The central claim is that alignment removes the need for a separate generate-then-filter step.

Significance. If the main claim holds, AQE is a practical contribution: it reduces inference cost while improving lexical retrieval, and it demonstrates that DPO with retrieval-derived preferences is a viable alignment signal for query expansion. The paper's strengths are its clear framing of the latency problem, the use of a standard base model (T0 3B), and evaluations across four datasets with both in-domain and out-of-domain conditions. The efficiency measurements (Section 5.3) and the diversity analysis (Section 5.4) add useful context. However, the validity of the main comparison is compromised by the configuration of the filtering baseline, and the reliability of the rank-based reward signal is not analyzed. The paper does not release code, which limits reproducibility.

major comments (3)
  1. [Section 4.2 and Section 5.2 (Tables 1-4)] The filtering baseline is not the generate-then-filter approach that AQE claims to supersede. The baseline in Section 4.2 generates candidates with zero-shot T0 prompting and then applies a fine-tuned DeBERTa-V3 reranker, whereas the original Expand-and-Rerank method (Chuang et al. [2]) uses GAR-generated contexts (answers, sentences, titles) and a reranker trained on rank differences. The authors explicitly concede in Section 5.2 that the filtering approach 'relies heavily on the specific GAR generators' as noted in the original work. Because of this mismatch, the reported superiority of AQE over 'Filtering' does not establish that AQE eliminates the need for filtering in the strongest generate-then-filter system. The paper should either implement the original EAR configuration (GAR generator plus reranker) or restrict its claims to filtering with zero-shot candidates, and it should investigate the implausibly low out-of-distribution filtering numbers (e.g., 9.1 Top-1 on WebQA after training on Natural Questions in Table 3) to rule out a misconfigured baseline.
  2. [Section 3.2 and Section 3.3.2] The preference signal that drives both RSFT and DPO is the BM25 rank of a single gold passage. This assumes that the gold passage is the only relevant passage and that its rank is a low-noise measure of expansion quality. The paper provides no analysis of the reliability of this reward, such as agreement with multi-gold relevance judgments, stability across BM25 implementations, or robustness to parameter choices. Since the entire alignment is trained on this signal, the paper should analyze the signal's quality (e.g., correlation across expansions, or using multiple relevant passages) and discuss the risk of mislabeled preference pairs when the gold passage is absent or when a query has multiple relevant passages.
  3. [Section 4.1 and Section 4.3] The experimental section does not specify the retrieval corpus or index. The paper reports Top-N accuracy for open-domain QA but never states which passage collection is used (e.g., the Wikipedia split from DPR or another index), how passages are tokenized, which BM25 implementation and parameters are used, or how many candidate passages are retrieved before computing Top-N. These details are necessary for reproducibility and for interpreting the differences across datasets; without them, the empirical claims are not fully checkable.
minor comments (6)
  1. [Section 3.2] The sentence 'The higher the rank, the more effective the expansion' is ambiguous because a lower rank value (closer to 1) is better; it should say 'the lower the rank value, the more effective the expansion.'
  2. [Equation (10)] The DPO loss formula contains stray '!' characters and mismatched parentheses in the typeset version; please correct the equation.
  3. [Section 5.4 and Figure 3] The text defines diversity via average pairwise cosine similarity (lower similarity means greater diversity), while Figure 3 is labeled 'Average Pairwise Cosine Distance.' Clarify which quantity is plotted and use consistent terminology.
  4. [Section 5.2] The sentence 'filtering ... failing to surpass both the original query expansion and zero-shot query expansion' should be rephrased because the baselines are 'Original query' and 'Query expansion (zero-shot),' not 'original query expansion.'
  5. [Section 5.1 and Tables 1-2] The paper reports 'statistically significant improvements using Student's t-tests with Bonferroni correction' but does not state the unit of analysis (per query or per run) or the number of repetitions; specify this to make the significance claims interpretable.
  6. [Figure 2] The y-axis label 'GPU memeroy occupancy' contains a typo; also state the hardware used and clarify whether 'Computational Time' refers to per-query latency or end-to-end batch processing time.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: held-out test evaluation underpins the retrieval-effectiveness claim, and the efficiency gain is a measured consequence of single-pass inference rather than a fitted prediction.

full rationale

AQE's derivation chain is supervised learning, not a self-referential loop. In Section 3.2, expansions are ranked by the BM25 rank of the gold passage, and Section 3.3 turns the best/worst expansions into RSFT/DPO labels. These labels are computed on training queries only, and Section 4.3.2 evaluates on held-out test splits; therefore the top-N retrieval numbers in Tables 1-4 are genuine predictions and are not determined by the training labels. The central scientific claim, that alignment improves retrieval effectiveness, is externally testable. The '~70% latency reduction' in Section 5.3 follows from the method design (single greedy decode in Section 3.4 vs 50 sampled expansions plus a DeBERTa-V3 reranker in the baseline), so it is partly by-construction, but it is reported as a measured engineering comparison and is not a fitted quantity masquerading as a prediction. The paper's self-citations (refs. 17, 23, 30-32) appear in related-work or background contexts, e.g., 'Similar approaches have shown to be able to reduce retrievability bias as well [17]'; none carries a load-bearing premise, no uniqueness theorem is imported, and no ansatz is smuggled in by citation. The skeptic's concern that the filtering baseline deviates from the original EAR/GAR setup is a comparison-validity or correctness risk, not circularity, because no equation reduces to a fitted parameter. Overall, no circular step meets the quoted-equation standard, so a low score is appropriate.

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

The central claim rests on the choice of hyperparameters (n=50, beta=0.1) and on the assumption that BM25 rank of a single gold passage is a reliable reward signal. No new entities are introduced.

free parameters (4)
  • n (number of sampled expansions per query) = 50
    Chosen for diversity; no sensitivity analysis. Directly affects the quality of the best/worst expansion labels.
  • beta (DPO scaling) = 0.1
    Hand-set; affects the strength of the preference update. No sensitivity analysis is reported.
  • Generation sampling parameters = temperature=1.0, top-k=50
    Used for zero-shot candidate generation; influences the diversity and quality of the candidate pool.
  • Fine-tuning learning rate = 5e-5
    Standard hyperparameter, but no sensitivity analysis is provided.
assumptions (4)
  • domain assumption BM25 rank of the single gold passage is a valid reward for query expansion quality.
    Used in Section 3.2 to define best and worst expansions. Assumes the gold passage is the only relevant document.
  • domain assumption The zero-shot LLM (T0 3B) generates a diverse set of candidate expansions that includes near-optimal expansions.
    Section 3.1; if the candidates are poor, alignment cannot improve over zero-shot expansion.
  • domain assumption The evaluation corpus contains the gold passage for every query and the retrieval index is fixed and realistic.
    Necessary for rank computation; the paper does not describe the exact passage index used.
  • standard math Standard DPO and RSFT objectives are appropriate for this generation task.
    Section 3.3 applies known alignment formulas without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment." pith.science (2026). https://pith.science/paper/IOYU7MQH

@misc{pith2026250711042,
  author       = {Pith},
  title        = {Pith review of: Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOYU7MQH}},
  note         = {Machine review of arXiv:2507.11042}
}
read the original abstract

With the breakthroughs in large language models (LLMs), query generation techniques that expand documents and queries with related terms are becoming increasingly popular in the information retrieval field. Such techniques have been shown to improve the effectiveness of traditional lexical retrieval methods by dealing with the vocabulary mismatch problem. Recent work has found that generating queries with a greedy decoding strategy can produce sub-optimal queries, including hallucinations, and proposed to filter out queries before expansion. This `generate-then-filter' approach is costly, as it requires generating multiple queries and applying a relevance model to all of them and does not teach the LLM which of the generated queries is more effective for expansion. To overcome such limitations, we propose Aligned Query Expansion (AQE), a novel approach to enhance query expansion for passage retrieval in open-domain question answering. AQE leverages recent techniques in LLM alignment to fine-tune models for generating query expansions that directly optimize the effectiveness of the retrieval task, eliminating the need for additional filtering steps. This alignment ensures that queries are more relevant, reducing computational costs while improving retrieval effectiveness. Empirical evaluations show that AQE outperforms baseline models for query expansion in both in-domain and out-of-domain settings, demonstrating significant improvements in retrieval effectiveness.

Figures

Figures reproduced from arXiv: 2507.11042 by the authors.

Figure 1
Figure 1. Aligned Query Expansion (AQE) training pipeline. The pipeline begins with the generation of several query expansions [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Comparison of GPU memeroy occupancy, computational time, and top-1 retrieval accuracy of filtering and AQE when [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Diversity of the generated query expansions. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Are LLM-Based Retrievers Worth Their Cost? An Empirical Study of Efficiency, Robustness, and Reasoning Overhead

    cs.IR 2026-04 accept novelty 6.0 of 10

    Empirical comparison across 14 retrievers on the BRIGHT benchmark shows reasoning-specialized models can match strong accuracy with competitive speed while many large LLM bi-encoders add latency for small gains and co...

  2. LLM-Based Re-Ranking for Real Estate Search

    cs.IR 2026-07 conditional novelty 5.0 of 10

    An LLM pointwise re-ranker using user profiles and candidate-set statistics improved ranking quality in a real-estate marketplace, with statistically significant production gains of +5.3% CTR and +4.8% scheduled visits.

  3. Query Expansion in the Age of Pre-trained and Large Language Models: A Comprehensive Survey

    cs.IR 2025-09 unverdicted novelty 5.0 of 10

    A comprehensive survey that organizes query expansion methods in the PLM/LLM era along four design dimensions, synthesizes application patterns, and outlines future directions.

Reference graph

Works this paper leans on

34 extracted references · 12 canonical work pages · cited by 3 Pith papers

  1. [2]

    Yung-Sung Chuang, Wei Fang, Shang-Wen Li, Wen-tau Yih, and James Glass

    work page

  2. [1]

    Yingshan Chang, Mridu Narang, Hisami Suzuki, Guihong Cao, Jianfeng Gao, and Yonatan Bisk. 2022. Webqa: Multihop and multimodal qa. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 16495–16504

    work page 2022

  3. [3]

    Thomas Coste, Usman Anwar, Robert Kirk, and David Krueger. 2024. Reward model ensembles help mitigate overoptimization. In ICLR

    work page 2024

  4. [4]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

    arXiv 2024

  5. [5]

    Miles Efron, Peter Organisciak, and Katrina Fenlon. 2012. Improving retrieval of short texts through document expansion. In Proceedings of the 35th international ACM SIGIR conference on Research and development in information retrieval . 911– 920

    work page 2012

  6. [6]

    Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. 2023. Doc2query–: When less is more. In European Conference on Information Retrieval . Springer, 414–422

    work page 2023

  7. [7]

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. De- berta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654 (2020)

    arXiv 2020

  8. [8]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2023. A survey on hallucination in large language models: Principles, taxonomy, chal- lenges, and open questions. arXiv preprint arXiv:2311.05232 (2023)

    arXiv 2023

Show all 34 references
  1. [9]

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551 (2017)

    2017 arXiv

  2. [10]

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for C...

    2019

  3. [11]

    I Loshchilov. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  4. [12]

    Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2020. Generation-augmented retrieval for open-domain question answering. arXiv preprint arXiv:2009.08553 (2020)

    2020 arXiv

  5. [13]

    Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation-Augmented Retrieval for Open-Domain Question Answering. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Inte...

    2021

  6. [14]

    Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. 2019. From doc2query to docTTTTTquery. Online preprint 6 (2019)

    2019

  7. [15]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35...

    2022

  8. [16]

    Wenjun Peng, Guiyang Li, Yue Jiang, Zilong Wang, Dan Ou, Xiaoyi Zeng, Derong Xu, Tong Xu, and Enhong Chen. 2024. Large language model based long-tail query rewriting in taobao search. In Companion Proceedings of the ACM Web Conference 2024. 20–28

    2024

  9. [17]

    Gustavo Penha, Enrico Palumbo, Maryam Aziz, Alice Wang, and Hugues Bouchard. 2023. Improving Content Retrievability in Search with Controllable Query Generation. In Proceedings of the ACM Web Conference 2023 . 3182–3192

    2023

  10. [18]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems 36 (2024)

    2024

  11. [19]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the lim- its of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683 (2019)

    2019 arXiv

  12. [20]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing . Association for Computational Linguistics. https://arxiv.org/abs/1908.10084

    2019 arXiv

  13. [21]

    Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al

  14. [22]

    Christopher Sciavolino, Zexuan Zhong, Jinhyuk Lee, and Danqi Chen. 2021. Simple entity-centric questions challenge dense retrievers. arXiv preprint arXiv:2109.08535 (2021)

    2021 arXiv

  15. [23]

    Zhengyan Shi, Adam X Yang, Bin Wu, Laurence Aitchison, Emine Yilmaz, and Aldo Lipani. 2024. Instruction Tuning With Loss Over Instructions.arXiv preprint arXiv:2405.14394 (2024)

    2024 arXiv

  16. [24]

    Amit Singhal and Fernando Pereira. 1999. Document expansion for speech retrieval. In Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval . 34–41

    1999

  17. [25]

    Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F

    Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F. Christiano. 2020. Learning to sum- marize with human feedback. In Advances in Neural Information Processing Sys- tems 33: Annual Conference on Neural In...

    2020

  18. [26]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. ArXiv preprint abs/2302.13971 (2023). https://arxi...

    2023 arXiv

  19. [27]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. ArXiv preprint abs/2307.09288 (2023). https://arxiv...

    2023 arXiv

  20. [28]

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

    2023 arXiv

  21. [29]

    Jinxi Xu and W Bruce Croft. 2000. Improving the effectiveness of information retrieval with local context analysis. ACM Transactions on Information Systems (TOIS) 18, 1 (2000), 79–112

    2000

  22. [30]

    Adam Yang, Chen Chen, and Konstantinos Pitas. 2024. Just rephrase it! Un- certainty estimation in closed-source language models via multiple rephrased queries. arXiv preprint arXiv:2405.13907 (2024)

    2024 arXiv

  23. [31]

    Adam X Yang, Maxime Robeyns, Thomas Coste, Jun Wang, Haitham Bou-Ammar, and Laurence Aitchison. 2024. Bayesian reward models for LLM alignment.arXiv preprint arXiv:2402.13210 (2024)

    2024 arXiv

  24. [32]

    Adam X Yang, Maxime Robeyns, Xi Wang, and Laurence Aitchison. 2024. Bayesian low-rank adaptation for large language models. In ICLR. 9

    2024

  25. [2021]

    arXiv preprint arXiv:2110.08207 (2021)

    Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207 (2021)

    2021 arXiv

  26. [2023]

    arXiv preprint arXiv:2305.17080 (2023)

    Expand, rerank, and retrieve: Query reranking for open-domain question answering. arXiv preprint arXiv:2305.17080 (2023)

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.