Pith. sign in

REVIEW 4 major objections 6 minor 39 references

GainRAG: Preference Alignment in Retrieval-Augmented Generation through Gain Signal Synthesis

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Passage selection in RAG should maximize predicted gain, not topical relevance, and GainRAG shows this improves answers across six QA benchmarks.

desk verdict A promising retriever-LLM alignment method that defines 'gain' via contrastive perplexity and distills a selector, but the teacher signal is never validated against true downstream utility and the two closest baselines are missing. read the letter →

arxiv 2505.18710 v1 pith:WZO4EEO4 submitted 2025-05-24 cs.IR cs.AI

classification cs.IRcs.AI
keywords retrieval-augmentedgenerationgainsignalpreferencealignmentcontrastivedecodingpassageselectionpseudo-passageopen-domainquestionansweringreranker
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that retrieval-augmented generation should select passages by how much they help the language model produce the correct answer, not by how topically relevant they are. To make that concrete, the authors define a 'gain' score as the perplexity of the gold answer under contrastively decoded logits, which subtracts the model's internal-knowledge prior and isolates the passage's contribution. They synthesize gain labels for a small sample of questions, distill them into a lightweight selector built on a reranker, and add a self-generated 'pseudo-passage' as a fallback candidate. On six QA benchmarks, the resulting system, GainRAG, outperforms standard RAG, reranking, and self-reflective RAG baselines. If the gain signal is a faithful measure of utility, the method shows that preference alignment between retriever and LLM can be achieved with very little training data.

What carries the argument

The load-bearing object is the gain metric $M(c,a|q)$, the perplexity of the gold answer $a$ under contrastively decoded logits: $\tilde{p}(a_t) \propto p(a_t|c,q,a_{<t})\left(p(a_t|c,q,a_{<t})/p(a_t|q,a_{<t})\right)^\alpha$, implemented as $\mathrm{softmax}((1+\alpha)\mathrm{logit}_\theta(a_t|c,q,a_{<t}) - \alpha\,\mathrm{logit}_\theta(a_t|q,a_{<t}))$. Raising the probability of the passage-conditioned distribution and lowering the query-only distribution subtracts the model's internal-knowledge prior, so the resulting score reflects what the passage itself contributes. The selector is the learned function $f(q,c;\theta)\to\hat{v}$ initialized from BGE-reranker-base, trained by KL-distilling a softmax over the transformed gain labels $v=-\log(v+1)$. A pseudo-passage $c_0=G(P_0(q))$ generated by the LLM itself is appended to the candidate list, giving the selector a way to fall back on internal knowledge when all retrieved passages are unhelpful.

What would settle it

On a held-out set of question-passage pairs, compute the correlation between the gain score $M(c,a|q)$ and whether the generator produces a correct answer when given only that passage. If the correlation is near zero or negative on a dataset where GainRAG reports a gain, the gain signal is not actually measuring utility, and the method's success must be attributed to something else (such as the pseudo-passage fallback).

Watch

Extended reading notes

Core claim

GainRAG establishes that replacing relevance-based passage selection with gain-based selection improves RAG accuracy. The paper's central empirical observation is that retrieved passages containing the gold answer still lead to incorrect generations nearly half the time, while correct generations often rely on passages that do not contain the answer at all. Gain is defined as contrastive perplexity: the perplexity of the gold answer when the LLM's logits are amplified in the presence of the passage and suppressed in its absence, removing the model's reliance on memorized knowledge. A selector trained on roughly fourteen thousand synthesized samples predicts these gain scores and, at inference, feeds the LLM only the highest-gain passage from the retrieved set plus one self-generated pseudo-passage. The authors report that this simple middleware outperforms the BGE reranker and other baselines on all six datasets, including cases where recall of the gold answer does not improve or even decreases.

Load-bearing premise

The paper assumes that the contrastive perplexity of the correct answer, with the model's internal-knowledge prior subtracted, is a faithful measure of a passage's utility for generation; if this proxy is biased or noisy, the selector learns a flawed ranking and the stated gains would not transfer.

Editorial extensions

If this is right

  • Retriever output should be re-ranked by predicted utility to the generator, not by topical relevance, so retrieval pipelines can keep their existing first-stage recall while a light middleware corrects the ranking.
  • Preference alignment can be distilled from a few thousand synthesized examples; the paper shows gains on datasets never seen in training, suggesting the gain signal transfers across domains.
  • Including a self-generated pseudo-passage as a candidate lets the system decline unhelpful external evidence, preventing the degenerate case where retrieved context is worse than the model's own knowledge.
  • Because the gain signal is defined through the generator's own logits, the same recipe applies to any instruction-tuned LLM without retraining that LLM or the retriever.

Reading between the lines

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

  • A natural extension the paper leaves implicit is to feed the gain signal back into retriever training itself, so first-stage retrieval is optimized for utility rather than relevance; the middleware results suggest such training would improve recall of genuinely useful passages.
  • The contrastive-perplexity gain could serve as a reward or ranking signal for other RAG components, such as query rewriting or context compression, where the question is also 'does this transformation help the model answer correctly?'
  • One could test the robustness of the gain definition by replacing the gold answer with the model's own sampled answer; if the correlation with downstream accuracy persists, gain labeling would no longer require gold annotations, greatly expanding the training-data pool.
  • The pseudo-passage strategy suggests a simple diagnostic: datasets where the pseudo-passage is frequently chosen may signal that the retriever's corpus or query formulation is weak, pointing to retrieval-side improvements rather than better selection.
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 / 6 minor

Summary. The paper addresses the preference gap between retrievers and LLMs in retrieval-augmented generation by introducing a "gain" metric M(c,a|q), defined as the contrastive perplexity of the gold answer conditioned on a passage (Eq. 5). A BGE-based selector f(q,c) is trained via KL distillation (Eq. 8) on a small set of synthesized gain labels, and a pseudo-passage generated by the LLM is added to the candidate list to mitigate degenerate retrieval. Experiments on six QA datasets report EM/F1 improvements over Standard RAG, Self-RAG, Rerank, and GenRead, with ablations and a synthetic-signal analysis.

Significance. The core idea is appealing and practical: if the gain metric faithfully captures how much a passage helps the generator, then distilling this signal into a lightweight reranking middleware requires only a small training set, and the pseudo-passage strategy is a sensible safeguard against unhelpful retrieval. The paper also makes a useful empirical observation about the prevalence of the preference gap and evaluates the contribution of contrastive decoding. Credit is due for releasing the source code and for reporting ablations of the main components. However, the significance of the method is not yet established because the teacher signal is assumed rather than validated, the most direct preference-alignment baselines named in the related work (DPA-RAG, BGM) are not compared, and the training-data description contains unresolved inconsistencies.

major comments (4)
  1. [Sec. 3.2, Eq. (5); Sec. 3.5, Eq. (8); Algorithm 2] The central claim depends on the gain M(c,a|q) ranking passages in the same order as their actual contribution under the inference-time generator. M is computed under the contrastive decoding distribution of Eq. (4) with alpha=0.5, whereas Algorithm 2 (line 9) generates the final answer with standard decoding. The paper provides no experiment showing that M-ranked passages correspond to higher EM/F1 under standard decoding. A passage can score a low contrastive perplexity because it strongly shifts the LLM away from its internal prior, yet be largely ignored or harmful under ordinary likelihood sampling; conversely, a passage containing the gold answer verbatim can yield low perplexity while also introducing distracting context. I request a direct validation: sample passages, compute M and the standard-decoding EM/F1 for each passage, and report rank correlation or agreement in top-1 selection. Without this, the KL distillation in Eq. (8) may teach the selector to reproduce a flawed preference ranking.
  2. [Sec. 4.1 and Appendix C] The training-data description is inconsistent and the filtering step is under-specified. Section 4.1 says the authors randomly select 20k HotpotQA samples and "about 4k" WebQuestions samples, then filter "samples where the passage with the highest gain is incorrectly generated," yielding "about 10k samples." Appendix C reports 20,000 HotpotQA samples and all 3,778 WebQuestions samples, with 14,084 samples after filtering. The two accounts disagree, and the phrase "incorrectly generated" is ambiguous. If the authors remove training samples unless the highest-gain passage is "correctly generated" (e.g., contains the gold answer or is not a malformed pseudo-passage), the synthesized labels are no longer representative of the deployment distribution, and the generalization claims in Tables 1-2 could be partly an artifact of this filtering. Please clarify the exact filter and report results with and without it.
  3. [Tables 1-2 and Sec. 2.2] DPA-RAG (Dong et al., 2024) and BGM (Ke et al., 2024) are discussed in Related Work as the most direct preference-alignment approaches, yet neither appears in the main results. The paper claims their preference measurement is "coarse" but does not test this claim against them. Without these baselines, the state-of-the-art statement is not supported. Please add comparisons under the same retrieval and generator settings, or justify their exclusion with concrete numbers or strong reasons (e.g., unavailable code) and state the resulting limitation.
  4. [Tables 1-2, Table 4, Table 5] All reported metrics are single-run numbers with no error bars, confidence intervals, or significance tests. Some ablation differences are small (e.g., Table 4: NaturalQA Avg 41.09 for w/o pseudo vs 41.62 for GainRAG; Table 5: HotpotQA F1 drop of 0.61), so it is unclear whether these differences are stable. The claims of consistent gains across six datasets would be strengthened by running with multiple seeds and reporting means and variances, or at least paired significance tests over evaluation samples.
minor comments (6)
  1. [Sec. 4.1] There are typos in this section: "randomly selecte," "gathere," and "applie" should be corrected.
  2. [Appendix A] Appendix A says the paper uses "seven datasets on four tasks," but the paper evaluates six datasets; please correct this.
  3. [Eq. (9) and Algorithm 1] Equation (9) writes c0 = G(q), while Algorithm 1 writes c0 = G(P0(q)); please align the notation for the pseudo-passage generation prompt.
  4. [Sec. 3.2, Eq. (4)] Equation (4) uses sampling notation "a_t ~ softmax(...)" but Equation (5) then uses the probability \tilde{p}(a_j|...); clarify the relationship between the sampling statement and the probability distribution used to compute perplexity.
  5. [Table 5] The caption says "Performance degradation after removing contrastive decoding," but the table reports only two datasets; please state whether the other datasets were omitted for space and whether the trend holds there.
  6. [Sec. 4.5, Table 4] The ablation label "w/o all" is confusing because it is described as removing all modules while still using an ordinary reranker; a clearer label such as "w/o pseudo and w/o distillation" would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the gain signal is an externally computed perplexity-based label, distilled into a selector and evaluated on held-out EM/F1; no step reduces by construction to its own inputs.

full rationale

The derivation chain in GainRAG is self-contained. The gain metric M(c,a|q) in Eq. (5) is defined as the contrastive perplexity of the gold answer under the LLM, using Eq. (4) from CAD (Shi et al., 2023a). This label is computed for training passages and gold answers, then distilled into a selector f via the KL loss in Eq. (8). At inference, the selector predicts gain values for passages without access to the gold answer, the argmax passage is chosen by Eq. (11), and the final answer is generated by Eq. (12). Downstream EM/F1 is measured against the held-out gold answer. No parameter is fitted directly to EM/F1, no uniqueness theorem or self-citation is invoked to force the choice of contrastive perplexity, and the pseudo-passage strategy is an explicit design element rather than a renamed target. The only substantive concern is whether the perplexity-based gain proxy is monotone with true downstream utility; that is an empirical validity question, not a circularity. The paper provides ablations and cross-dataset results as independent evidence. Therefore the analysis finds no circular step and the score is 0.

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

The central claim rests on the gain metric as a faithful utility proxy, the transferability of distilled predictions, and the benefit of pseudo-passages. No invented entities are introduced; the gain is a defined metric, not a postulated hidden quantity.

free parameters (3)
  • alpha (contrastive decoding strength) = 0.5
    Set to 0.5 following CAD (Shi et al., 2023a); controls the degree of contrastive adjustment in the gain computation and thus shapes the training labels.
  • number of retrieved passages k = 100
    Chosen based on coverage analysis in Appendix B; affects the candidate pool for selection but not the gain computation itself.
  • training sample counts = 20k HotpotQA + ~4k WebQuestions, filtered to about 10k (or 14,084 per Appendix C)
    Data amount is a design choice; the discrepancy in reported counts is a consistency issue.
assumptions (3)
  • domain assumption Contrastive perplexity of the gold answer measures passage gain
    Eq. 5 defines gain as perplexity under contrastively decoded logits; the paper motivates this via CAD but does not prove it captures utility. This is the central modeling assumption.
  • domain assumption The selector can predict gain from (query, passage) without access to the gold answer
    Training distills gains computed with gold answers into f(q,c); the paper assumes this transfers to unseen queries at inference (Sec 3.3, 3.5).
  • domain assumption Pseudo-passages mitigate degradation and are beneficial when external passages fail
    The paper assumes that LLM-generated background passages are useful fallbacks and that the selector can appropriately choose them (Sec 3.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of GainRAG: Preference Alignment in Retrieval-Augmented Generation through Gain Signal Synthesis." pith.science (2026). https://pith.science/paper/WZO4EEO4

@misc{pith2026250518710,
  author       = {Pith},
  title        = {Pith review of: GainRAG: Preference Alignment in Retrieval-Augmented Generation through Gain Signal Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZO4EEO4}},
  note         = {Machine review of arXiv:2505.18710}
}
read the original abstract

The Retrieval-Augmented Generation (RAG) framework introduces a retrieval module to dynamically inject retrieved information into the input context of large language models (LLMs), and has demonstrated significant success in various NLP tasks. However, the current study points out that there is a preference gap between retrievers and LLMs in the RAG framework, which limit the further improvement of system performance. Some highly relevant passages may interfere with LLM reasoning because they contain complex or contradictory information; while some indirectly related or even inaccurate content may help LLM generate more accurate answers by providing suggestive information or logical clues. To solve this, we propose GainRAG, a novel approach that aligns the retriever's and LLM's preferences by defining a new metric, "gain", which measure how well an input passage contributes to correct outputs. Specifically, we propose a method to estimate these gain signals and train a middleware that aligns the preferences of the retriever and the LLM using only limited data. In addition, we introduce a pseudo-passage strategy to mitigate degradation. The experimental results on 6 datasets verify the effectiveness of GainRAG.

Figures

Figures reproduced from arXiv: 2505.18710 by the authors.

Figure 1
Figure 1. We analyze the preference gap between re [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the GainRAG framework. The GainRAG workflow, preference signal synthesis, and selector [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of gain. Changes in recall of the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of the pseudo-passages generated [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: As the number of passages increases, the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the change in recall as the num [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Illustration of the change in EM coverage as [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Illustration of the change in F1 coverage as [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 2 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

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

  4. [4]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511

  5. [5]

    Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on freebase from question-answer pairs. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1533--1544

  6. [6]

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. 2024. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 719--729

  7. [7]

    Guanting Dong, Yutao Zhu, Chenghao Zhang, Zechen Wang, Zhicheng Dou, and Ji-Rong Wen. 2024. Understand what llm needs: Dual preference alignment for retrieval-augmented generation. arXiv preprint arXiv:2406.18676

  8. [8]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6491--6501

Show all 39 references
  1. [9]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

  2. [10]

    Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2g: Retrieve, rerank, generate. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

  3. [11]

    Hangfeng He, Hongming Zhang, and Dan Roth. 2022. Rethinking with retrieval: Faithful large language model inference. arXiv preprint arXiv:2301.00303

  4. [13]

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020 b . Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060

  5. [14]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118

  6. [15]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299, 1(2):4

  7. [16]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1--38

  8. [17]

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p...

  9. [18]

    Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906

  10. [19]

    Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. 2024. Bridging the preference gap between retrievers and llms. arXiv preprint arXiv:2401.06954

  11. [20]

    Jaehyung Kim, Jaehyun Nam, Sangwoo Mo, Jongjin Park, Sang-Woo Lee, Minjoon Seo, Jung-Woo Ha, and Jinwoo Shin. 2024. Sure: Summarizing retrievals using answer candidates for open-domain qa of llms. arXiv preprint arXiv:2404.13081

  12. [21]

    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...

  13. [22]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

  14. [23]

    Ming Li, Yong Zhang, Shwai He, Zhitao Li, Hongyu Zhao, Jianzong Wang, Ning Cheng, and Tianyi Zhou. 2024. Superfiltering: Weak-to-strong data filtering for fast instruction-tuning. arXiv preprint arXiv:2402.00530

  15. [24]

    Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2023. From quantity to quality: Boosting llm performance with self-guided data selection for instruction tuning. arXiv preprint arXiv:2308.12032

  16. [25]

    Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2022. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097

  17. [26]

    Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Rich James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. 2023. Ra-dit: Retrieval-augmented dual instruction tuning. arXiv preprint arXiv:2310.01352

  18. [27]

    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, pages 5303--5315

  19. [28]

    Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Hannaneh Hajishirzi, and Daniel Khashabi. 2022. When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories. arXiv preprint arXiv:2212.10511, 7

  20. [29]

    Jinming Nian, Zhiyuan Peng, Qifan Wang, and Yi Fang. 2024. W-rag: Weakly supervised dense retrieval in rag for open-domain question answering. arXiv preprint arXiv:2408.08444

  21. [30]

    P Rajpurkar. 2016. Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250

  22. [31]

    Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. 2023 a . Trusting your evidence: Hallucinate less with context-aware decoding. arXiv preprint arXiv:2305.14739

  23. [32]

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2023 b . Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652

  24. [33]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

  25. [34]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. arXiv preprint arXiv:2212.10509

  26. [35]

    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, pages 9414--9423

  27. [36]

    Yuhao Wang, Ruiyang Ren, Junyi Li, Wayne Xin Zhao, Jing Liu, and Ji-Rong Wen. 2024. Rear: A relevance-aware retrieval-augmented framework for open-domain question answering. arXiv preprint arXiv:2402.17497

  28. [37]

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, pages 641--649

  29. [38]

    Fangyuan Xu, Weijia Shi, and Eunsol Choi. 2024. Recomp: Improving retrieval-augmented lms with context compression and selective augmentation. In The Twelfth International Conference on Learning Representations

  30. [39]

    Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2022. Generate rather than retrieve: Large language models are strong context generators. arXiv preprint arXiv:2209.10063

  31. [40]

    Tianjun Zhang, Shishir G Patil, Naman Jain, Sheng Shen, Matei Zaharia, Ion Stoica, and Joseph E Gonzalez. 2024. Raft: Adapting language model to domain specific rag. arXiv preprint arXiv:2403.10131

Pith tools

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